From 29b56903d6a22fd1f38580fd04b415c4e4829fc5 Mon Sep 17 00:00:00 2001 From: dohe0342 Date: Fri, 9 Jun 2023 13:48:04 +0900 Subject: [PATCH] from local --- .../.checkpoint.py.swp | Bin 0 -> 16384 bytes .../.data2vec_audio.py.swp | Bin 0 -> 45056 bytes .../README.md | 54 + .../__init__.py | 0 .../asr_datamodule.py | 559 ++++ .../beam_search.py | 2342 +++++++++++++++++ .../bias_compare.py | 11 + .../bitfit.py | 1827 +++++++++++++ .../checkpoint.py | 225 ++ .../convolution.py | 83 + .../ctc_decode.py | 818 ++++++ .../data2vec_audio.py | 766 ++++++ .../data2vec_encoder.py | 184 ++ .../decode.py | 887 +++++++ .../decode_new.py | 834 ++++++ .../decoder.py | 102 + .../encoder_interface.py | 43 + .../export.py | 320 +++ .../full_ft.py | 1732 ++++++++++++ .../jit_pretrained.py | 271 ++ .../jit_pretrained_ctc.py | 423 +++ .../joiner.py | 65 + .../last_layer.py | 1836 +++++++++++++ .../model.py | 241 ++ .../nets_utils.py | 503 ++++ .../optim.py | 1061 ++++++++ .../pretrained.py | 353 +++ .../pretrained_ctc.py | 441 ++++ .../prompt_tuning.py | 1825 +++++++++++++ .../pseudo.py | 904 +++++++ .../scaling.py | 1178 +++++++++ .../scaling_converter.py | 114 + .../test_model.py | 56 + .../train.py | 1719 ++++++++++++ .../train_adapter.py | 1800 +++++++++++++ .../train_lora.py | 1862 +++++++++++++ .../train_uda.py | 1960 ++++++++++++++ .../utils.py | 21 + .../zipformer.py | 1866 +++++++++++++ 39 files changed, 29286 insertions(+) create mode 100644 egs/tedlium3/ASR/pruned_transducer_stateless_d2v_v2/.checkpoint.py.swp create mode 100644 egs/tedlium3/ASR/pruned_transducer_stateless_d2v_v2/.data2vec_audio.py.swp create mode 100644 egs/tedlium3/ASR/pruned_transducer_stateless_d2v_v2/README.md create mode 100644 egs/tedlium3/ASR/pruned_transducer_stateless_d2v_v2/__init__.py create mode 100644 egs/tedlium3/ASR/pruned_transducer_stateless_d2v_v2/asr_datamodule.py create mode 100644 egs/tedlium3/ASR/pruned_transducer_stateless_d2v_v2/beam_search.py create mode 100644 egs/tedlium3/ASR/pruned_transducer_stateless_d2v_v2/bias_compare.py create mode 100755 egs/tedlium3/ASR/pruned_transducer_stateless_d2v_v2/bitfit.py create mode 100644 egs/tedlium3/ASR/pruned_transducer_stateless_d2v_v2/checkpoint.py create mode 100644 egs/tedlium3/ASR/pruned_transducer_stateless_d2v_v2/convolution.py create mode 100755 egs/tedlium3/ASR/pruned_transducer_stateless_d2v_v2/ctc_decode.py create mode 100644 egs/tedlium3/ASR/pruned_transducer_stateless_d2v_v2/data2vec_audio.py create mode 100644 egs/tedlium3/ASR/pruned_transducer_stateless_d2v_v2/data2vec_encoder.py create mode 100755 egs/tedlium3/ASR/pruned_transducer_stateless_d2v_v2/decode.py create mode 100755 egs/tedlium3/ASR/pruned_transducer_stateless_d2v_v2/decode_new.py create mode 100644 egs/tedlium3/ASR/pruned_transducer_stateless_d2v_v2/decoder.py create mode 100644 egs/tedlium3/ASR/pruned_transducer_stateless_d2v_v2/encoder_interface.py create mode 100755 egs/tedlium3/ASR/pruned_transducer_stateless_d2v_v2/export.py create mode 100755 egs/tedlium3/ASR/pruned_transducer_stateless_d2v_v2/full_ft.py create mode 100755 egs/tedlium3/ASR/pruned_transducer_stateless_d2v_v2/jit_pretrained.py create mode 100755 egs/tedlium3/ASR/pruned_transducer_stateless_d2v_v2/jit_pretrained_ctc.py create mode 100644 egs/tedlium3/ASR/pruned_transducer_stateless_d2v_v2/joiner.py create mode 100755 egs/tedlium3/ASR/pruned_transducer_stateless_d2v_v2/last_layer.py create mode 100644 egs/tedlium3/ASR/pruned_transducer_stateless_d2v_v2/model.py create mode 100644 egs/tedlium3/ASR/pruned_transducer_stateless_d2v_v2/nets_utils.py create mode 100644 egs/tedlium3/ASR/pruned_transducer_stateless_d2v_v2/optim.py create mode 100755 egs/tedlium3/ASR/pruned_transducer_stateless_d2v_v2/pretrained.py create mode 100755 egs/tedlium3/ASR/pruned_transducer_stateless_d2v_v2/pretrained_ctc.py create mode 100755 egs/tedlium3/ASR/pruned_transducer_stateless_d2v_v2/prompt_tuning.py create mode 100755 egs/tedlium3/ASR/pruned_transducer_stateless_d2v_v2/pseudo.py create mode 100644 egs/tedlium3/ASR/pruned_transducer_stateless_d2v_v2/scaling.py create mode 100644 egs/tedlium3/ASR/pruned_transducer_stateless_d2v_v2/scaling_converter.py create mode 100755 egs/tedlium3/ASR/pruned_transducer_stateless_d2v_v2/test_model.py create mode 100755 egs/tedlium3/ASR/pruned_transducer_stateless_d2v_v2/train.py create mode 100755 egs/tedlium3/ASR/pruned_transducer_stateless_d2v_v2/train_adapter.py create mode 100755 egs/tedlium3/ASR/pruned_transducer_stateless_d2v_v2/train_lora.py create mode 100755 egs/tedlium3/ASR/pruned_transducer_stateless_d2v_v2/train_uda.py create mode 100644 egs/tedlium3/ASR/pruned_transducer_stateless_d2v_v2/utils.py create mode 100644 egs/tedlium3/ASR/pruned_transducer_stateless_d2v_v2/zipformer.py diff --git a/egs/tedlium3/ASR/pruned_transducer_stateless_d2v_v2/.checkpoint.py.swp b/egs/tedlium3/ASR/pruned_transducer_stateless_d2v_v2/.checkpoint.py.swp new file mode 100644 index 0000000000000000000000000000000000000000..be516b47454af4ce9d4d51ca93715ffe491ed3be GIT binary patch literal 16384 zcmeHNO^h5z6|P`{9l#J06eND4=h;Z?O5>jKI1y-JktSXThSu}Snwj$U zbakq0cCrZu@q6P&B5{BVf+OL=DHlXUBtSw!aNrPefCCp__z?;q1YcGE%*=WfZOj3I zs)za%o?%mYni1&gri2gyVXebBoR^1Xr5%zBow-r z8E!-@vaA=i_IvxSMmQ4TUaDkbg7lyo$R|7lo`EA6IL*$t&NXetQtb>q_`uE)ZG4eu zz%$?(@CB- zsJ_Vhn`gi?;2H1?cm_NJo&nE*XTUSy8So5v20R1*1qOJ;*n2VVuk8yz{{LUT|G)fh z#-0WCfyaRJz#rem*!O|20AB*00ycsBfS;UZ>^b03;6dO4;61>-z+d0V*lplmK;{`n5Zegu3Q_!e*l_z3VY@FC!Xz)9fGZ)fb6!1sUyAOtQ0j{twa zkFh@huK_;*UIM-gyZ{(r2z(fL2>8t@#(oC82z(W215M!eNydH)d=vOO@Hp@oPy^l% zoC5xaj~%ZAuK+&=z5{#>_#(i8PXlLwSMOzP54a!r;aeFK!25vHzzN_sKAyY+DBuIY z&+#$l7LWlNcn0VKr+|~d>#&uz*RRTtG>80#iff=W?bT_YWI`#|e4GW+h;g>OvdU}| zIVg%(nakU<*@24P+E_&*?(zL$?K-Vdl?#koHcE%1U=S;A&bL+qcP^tB$s@6nT~aE=vX zNkXygn$>B2{R$$PtB=)iD3f6zlYv?u)Vh2hxw&9QM$z*F8HK)S2Dx4o zS-G}l<}M@_ac)s)W#79JTJ@d@1*e)KF_1V1nD14>2%XhQ1j7KosR~DRidB}CbItd; znFLE;hTRJNbLWCpG#?^6&}!^#0nW4`1s~O1Cf17zo4XEn|63H7%efocT+c;ebFsK| z&vN)OpLb~MA)L)=2CsyJi%MU39sFv{jZ}$co|EkwuoXJ{iDK6!T-MQ6(biyrHS^t| z*ilz$6I+o=c)WXIV!HW8Z1;S-Wfv`-`^`0+`Q+h-?yHy((Fs|p*sko&p+>GiBKIRr|Z)T5WJgQvU z>XsYEHUb!wSN)qJG;E2~L{k{^ETg$b0V*a~(>SfVXG5M1LUzf|H6ett5Nlu1$l5TH zgMrWjMoHntvM*z4q=1KKLNJ#5sC(9O43ZqD3eIZH<(?eoTVSG=&^Q)x*=`*nR?Ggx zh-jS~exhd7WKnePm0SGsi2Zozqtg0)K%0;2w3lgh|yaAE@VM&b3ToOq?PvWm?;o`e;$6 zR92>nU8BqnW7UVzm#{O9B~BW{dor3N5n_;y*_UuuIl(m0pewNYW<^z4$KmBr4rE9% zPlgkO85%0Y8yi$86?!aTVk?v<5u{x0Y=3HVXPd6Jx3=0F+ntMD+T5ZGn;Yw$?at-~ za!*iu;~G8L*;q%glNeExcsA8I08Q{JP!mfLS%spRq3_h#Zte385G8m*wRV>}-5NdK z?smF$$k<8j$!gN#nTfRVvVTB4#Rwtq#{x5&+A;Z13)CUXV5^&IIBPP5v#Noaa!|Yk zIrbVEldd)z)9EzeIOjbQfzm@9@OBJljZ2*i7dN^W&$fc5>vBy^Xv`0=x(q@v0$8Y0 z2W`xPv)eK@nks7O3s4Ow|Nr9#yZC;4ws<2};M^xMCSk~0( zX>2ku6hS{)wmQ^`PPLVfIfbnQ1VICT#Z;UNARcII{yY+EG2}Y3rXNpYBk^)Wd}2mQ zXBAD3t2goUS&=VUSA3scEADd@?sM!~{twRi;^N~g QYKqjXvib(QR;B9f-wC7{4*&oF literal 0 HcmV?d00001 diff --git a/egs/tedlium3/ASR/pruned_transducer_stateless_d2v_v2/.data2vec_audio.py.swp b/egs/tedlium3/ASR/pruned_transducer_stateless_d2v_v2/.data2vec_audio.py.swp new file mode 100644 index 0000000000000000000000000000000000000000..215c8f6718b36d6866cc18ec63b438a5f84ec002 GIT binary patch literal 45056 zcmeI536vy9dFKmBNC;hk&3ff!ajM%^SI_iRcaN}+MlCcmnz3gY9UhG^s2SH*S7mpX zrn;(BSvAu=V;Nguv9_@>2!jBdW7mko#vp{i2yE7i7bGMt2xLDH2m>}+BLr`Gi1*Cp zx4$orjEKyvo|f?LzJ0EI?_Wp8^~D!Q#&<;I;>-*7ULIXknJIBStyKEa&vqM6zVhCa z7EXLpsnJ=9XU==hb_v?(x{P@7pt$8{nRMg0zB0Yz>iyH*-k=>fs{LNAoiqmZxK~a3wSL@+lcd_% zezQbquHr+*Ry4g{*O2?;s6{N^q00>ct+{G?XzmAkB^-b{lqh_ z8L?)>fe{Br92jw6#DNh9MjRM%V8nqD2mTdtpx-^U^b+cOf~oUS^ZlgE_o?RpR`Y!< z6aG~5|M}+o?V0eC%>SnO{&pt($>x7-zOOX|SUOHN_m|A~M>FB4nEQ+7`{zuBZ2G5| z`>W>r<(cpk&HdBN_hXrGb9?kP;=qUlBMyu>Fyg?710xQMI56VChyx=Ij5sjj!2g#V zs5MF@k?qeAc^CV?mH+QMt5muN{1&(pTmdct(_kE&4gS}erP2f7v*1qfA#f{r8@L7B z2o}K(a49$i+;awTz)heBt_3dybD#{)20sP9b9$+CKln7b6}$~x54xZO=D_LTn+WW; zf%k%!f+i?~zegeQRq%RnEjS;1pNf7QyaT))B;YFWAHWtc2~Gfyp)7b9+yib0Zvw9Y zH-k2~6g&@{0ltRf;eUYpz}vygz)QgauoXOt;^Ir-UEoDv0-OyVLE-Uw@N%#R>;^vz zP6kf~kDxI5Jop@V2Y3N^DtH)0#cRRKz|~+smYT#CB2Mzs~Um-{QxCDF`W$p(!_gLb1M!Hq_8N)yN(xq3oMN&Pe*WYvu3=cE2|zuKy;$Gw8QB&rIl_TpsF>L-<2w;Q(`h0ZL!&E_SfkjDxS(l!Y5sCk@AX?%t!5R1-G6?XVrX_ zCnE|CGm@Sc#Z(}gO{rUgil$*-)NDs_d$1b!@CqusvbKH~(;gkB-%~L)mS*nGE3(X;smL-@-?sn42A_m!QA)1-kt=Lum1{4)e$edE z+DUx94A$x@-L~0n#ZytExjMgXb}BktYYpQ01HFN(e?GG+Tp;^uzmw+Rx`XgPY*3$) zg}JF{o21oq#oR%%lF~x@{S-u4tF+@I)uVB*lQ6g{N%JVv#pD$C)z8^UemDB--FUv= z>D5;%i=9r(v50Qs3RGQ$eiX1rCJjAdD9-VbXQ>%YmE-YnyXE0b$z~?an$~zD){RTN zsNaeDD=}UB`a#^T$B~&;X`#5CG&^nAd!&GLve4bM-RZ5C*MiEY1JlK*E;V~e-|O_M zA4glGwQ9Z7KI|x0f29S~4{x@nt890r=wRZhZpL)3qSI7vdRGZ1^RzeHjd-ov9rRW2 z_hzZ7XvP(w+c2HM3TFAqIhc_Wo=&gxGj1|VC-Zq}YB(BF3rM(f27uHsOg<4 zLXEw&OQS zk*L%DIB7Npwbl;3$hf@I7_{Ot&*|%`@-X=MS#NwNgDw@HtW>K!+56RMxz<}wG>O9U z^tKuxkp#{~zg#=K{RMG-`>sxVsksa#^;RuOnCx-Rg>oLkyW90nBkmaj%J~o>ztrk1 zFURswmLDuDD$8+ypSlM@3BTjRajR?t_g=B*N@kg*US~CusjSl)^qDiG=4!Xo>qih{ zZ|Mq~s6=WdHWGc;f3)e$mTF9y@%5D;W;%J5G&AqqSRRG7`G zVs~0(*<_ICfzByproJb2*`w#4-FhmbqSPvG<;&usTscC6;Q4H8S zjdhMuFJ5jYbpI-p$E`4}j1o0essVTEomJ-bSgkqfI1OC469g{Z5_%Lh$g|g6BvvD; zC9Fb{ei(s~w=`(ir44H>iLxgfE;rll;BHz7YrXEeiqJI;-C2#7Yjh0O_nMBqdrb)I z2U)b5+gEmVk-DN%LDhZZltbyQsc5Mgw;Gm!PGWzp*7_^0A9v&3E2C`o6~V?ogG#c_p{2fTYSl4SO8!Q^%V zH<7v+HImuodk@&#Oc~rarIg0WV^^oU-jk(&RIX1(d+@Ml@pNb^+S{&I)N`fTPomlq z#5GwnCzX=O{|AxFMP?WIAJJJ|{}cKCA@E6XH~2l!0ly5M15O0Dn{38yxKV&6BMJ9R@OJs(?CFC^q{WXoSM_%lGHb>PL|7s0;;-^GUi82Ao&AGiq|02hHV za1MC*RP5g1Pr++I6Rd!(;3V(`Z1xX=KLXc-7lQM^ncyUFBKQ;R_U{5>$KMYo!9QcO zzX!YlydEq9vE@IEt^VDh2d2Tdu+e`LyboLty5J&kA$SJ(47U2az$d^hpav$ux!@6O z@m~k`f=_^tf;WS8a3y#S*aDsgq`e-IHrxDV`y|~~b-()3awT4^X}fh&?X+9#&d`?} zYIduK;`M~CiB`7kx{&)N8A=9A-Sf698>3ca+HxeQ5!Y+$+SqA}Q@319i6siW<}xMi zzI=xsXedLuqlcOdh}GJmxGGWmajm|B@>*(uK3U3|n3z!C>O*d*+Mz3~4!WrI(DBVj zGtPdKA9^Sy_chybP3H8dv3- zsfBcT)v#5Z!zh>RxT(r2dRljvQaGWtvU^&cTK@og#7<9QR}2xFLa`1^RtD{+DDkUa zJFAk#zd+(aa9}jaaKzZms>^7|l?|ErQc~`ICY|*Sur@bc%w^iiK+&3w;p(U8v%$?a zs_EgWRdBW$Ro&_XNkPVV{&Q~?_#q#|gg#iUY@I1St&gTzF)N_RKjHm zk)6l170FITQ>t_1$G+pV8(R?TZ;T6uDf) z#aP*irQ#}{{4Yrqxm2_oW4*Qn3ZxpC|UH=v-bh9O%<E`-9a;FBv0Q3@d9Z@|X2*&~7G45g*0nP4iLKsCww= z1$89OYr{DC3BYz+)#h`ywUyhgqH?j;udi$>rpm#OiPMegJTALto2dma!Y2|>_sdMC zV*j6IO8knE|7j$3eF|CscfqfKgW!4K`^fiS1%CxT2;Kw!42YlqHDDhQKmQZK1IYaE z1;@Z8;GdE2?*Vs!o4~bT3wSd46tew?z$?KLI18KzzKA^k1#mOC3S0>);IUKD0e~-o zhrrFC2lj($@FQgSzXQJuUIC7Qi@`I%)4}(U=l==(1Nd8T54Z!o0UQBc@B(lRcslqT za{hk>;=_LIr3{*iGJQX~OtbaH7Ft{1K8YJKfa5nf6^8OdV2f?er z3fKk4z&DZa?*xazC146j`+WKF$@{t?qs*mNxyE~kc}Yu%;$Kx&-Zf=ET)DxmRT}YO z+)&)OwK?YRNQlb7F8S12NtRLmUH2TaImx&2#H}XyaB`bLP2pZE)u`A6n4CbS{ONAERh$A;{<-3 zr7YF*zr1rQIxv|}tt#XCO_4L0(~L?+tVn`1Ij3x^hf=YW$fI_elq9je))rK#m9S0u z;$`mqu_bQ4D^}vQE-A|{P8ybPuVMY#CN-TtpV0NnG^(tUHWP7!Zk*3`SXa)+Ebh*s)3At%%-FFD%H}1kg@3fq-y;Ql<a8TBNgNv++c{{qaO7l5(Gex- zi-AS2KPrSO3MJn~iKZP%%jBaJ4TX{(QV#XC77MB1rOFMHb?{23k)tUUEs;hVRrQQg z2LPk8toO_ACqCi5c&UkR_!jY1=4;C#e8#Oyvt=amWaJVp9CBi@#89ota*rJ$NLD@e z6UkY_swkq;jx6!R!HN5Gj1Tf5u{ycPC0m?)AEMQx6 zdw};Zjb`l#Hko{OCt-Oz^mm}>0_uLLcHCd|^tF20-A!%~`&^xSk|^ReF^mxf1DB-*S~fozlURb`6fsvvq$ zUV~Dy^x5zH9yJ?cGbtuAml18_)j48gXOJ+%xUjiwyINN0qR+`>kuu07j$1aVHTJs4 zOX1{TE2n77S;oRnAfD=&znkr)PI+mJ$W|;jhBOohcanVcUFF!J8@!f~sLUi0j{8;hbdd?nkXK35 zv=NjVdS;f5Olo_Oq_mf!$@a#q8VEwwu27NxpN*vW86*EMX{oXF-;wuU2W0<$6Rdy; zdZPz4h}^aAe$Zv^XL9$XBb2F?NxBkz9=yb~M()8HXw{V#&|gF3hd9030Y z{4H|+N5H$lHDDX~ci>cTJ97Tpz;A+I1yyhHuu> z+i&SeP3c8wGMCvA=IzNeyWOx5cuP{7F13V(l9mLghJq?OW_pqqhWgsVyl_yd4U8Qj z8LB-f*@FD!V(mll5s~3)Mi78VwyJ?8nWkXJ-nfS5Yq-rxY`9S2-#PkYU zUiI#YO}>C4t>^SHHrL&t-3)Bb5aOMSz(0h8)owq(mXieI<9d6aj?%~6&e;hzr}T&1 z@Y@Hf1Fxa{b1oCREG7#7US+A(?3Tq0G+C(+R+GUh0`a7GcT(-H_RRi5gRKADmV3D) zSH3OeE~f!tdN=#O6N(?A5z&2wWn{h+ieN+^#Bie&O~H8vw)?oLRtl9Xl$-rZFO*aS z^D&MMM46~%^OUC6?7?R9m)^z37Dpa+eL8T1tycX`COru`gP?z;6Lr`(sTQt@@&Ww9L6gv&k5VQVO_?xK&fT}>D`<@P!lVF*m^jZqm}6z0og&e(4FtN ztxkKn%tl<(Wz*C$(7Y+kA8u;3;s;i0?e$>OY=^yZ!PyxU?I&<7vpgq){h8hU?jqav zwlPVh9=_(k0 ztEhr7!F6gKZCC|UE&b!u;7g`YSc@--)j$bo-;^=RvV-KCGfLS6!!-KEv=nRgC2p~1 zUqXts`*J_1q+QhE&{%L{LD*z2t|Rj}Yy9HzW|HT| z+4ymyyH0@<+NARTKNp$*lSckuH1hvl$olUAH-a&6GPoZ(|2IGv%z>{X-+vLj1ndGI zN0yiK`F;`n0{CfgHux^GyqwV|^8GjvyZ*i49`J{t0qWp+;F;hn$m?GQ?*;DxZvq`~ z4Y&ZD4(>-@{}{Lf{2};NAa?%!U>A565MO?4>;Dq+``zF+Ab$NV; z-$Jgx7l^fhm?|{hqF9%nH^T0VE0uLhBe;#}U zG{AGg!^rdxfcJyj!HwWW-~#Yv+C$pp?&He)rol?mv9ecwDzh>AErdXvyvm%l#cLKt z9<8p%wYFZ4q_@z>n>$uFnLkX-X1S;0oL$0ECNYeAm-ceO#v&)|F^7h-nq4N!)7@nj z`q+dIBgq<}Y*igOVsu(^&%eT@iVcASCSe&9G3R{EOez_@ueK2tl8dX;ft~sQ!(9hT zHY598m}H|GT4S2;%%h!j#ZwARA*gR0W1&g%Uz;e-qQ=iE zv^{&D)FIn5#K+k&uvp+ONkX~8J~1XP;Bi7z9rIjh^%SNWs-zN`k1(y=e7t?>@%5uI z=|i5$UrIEqc-MtsULd1-%yLYTl)5LW&#(zM*(W?gkRen)cF?2^uDeP(oX6FoX~>dD z2$NNftwKp|$dw<_VQnN%uo^Q3%op7Z>zgJGlGsI8NrL^$XHPyOOL!QP5wa|Qd@l`K zgjuF~&1SQ-se+AAo&u(1vygAh{JD|p@wHM%Ul4o4xIcmyVDkxX}OkE`vcBTRco9{ zRMNz9J*qGpo}Zh@X8ZB7ncY-2Xz@#yDTUO`Y$lsr9)9Pw2Ozr9nQbvWsMW!wHfYkt z7V_e?dfe@Y4?wzEAm&KCes4V&NlWN;7N$6*h&?4x*M_G*ir$=~4Ku!b1<5E}=W7mV zR2e(tmw2H_QTC1)OGF=UlxREg;5V8Q(X81@TgyaAs$wr2n3m|(gp=jW(V`m>PD(?K zabR=>s?%|$TV@H)-=PYwnT5ELrOl%v~}yX`w$n+gXO*?2NwhsY;ff?02f%3v7&vok5V}D!NK< zd}Q-8=FF|`G2)uLkO#pqrw1_PI{DQ zF@a6#gs0_y#M9CPB4@Jyf7OVSa{m8^!JRL291Fr)w2G@X#fcOQR1|FmBRQs8>liq1B zJLPE8gxjZN!?=WFVRvcp;Ob1iT?bJFUnfo>y<;i0b27OVhnT_CR7N6;3 zLNwfYiLPWRm+wHopIqPTm)9$(EXwj&K%7U5A8$KqA59gz6`!aqb|hKCNBc2^uwzsVveIP0d6d zD$M37A_jI$k!oq4W_(0B<(sTOW|AC5^@JmE*etx3N**Z3;`2zQ=D2S z4py4|c(puv174oT)L0QYUF-rj$_?xYWz~ur&Upw+&K&=O@4;@hqQYS|)M)(QB*Zup z#=Ra!w33h}mt~merpJ-|KMzlpG2XU>ttAa{v2KWL{lh0%QN^?lE5~D_+S-`QF4Dgk zpa0L2^J8(_)i%#>7v>2}(vk&DWyXv@DA%mkIBX%C*yZkqoS%H%B|DA^#wUp#VF+c) zM60qDeT=L_*&pumqbH8}5tWyic)clAvsYP5tI_HlTo_|RIXC593@&?dJCW%@9T^Z# zFZL9HStfZo6uP1wJjqcm9dcv0f5)BD98_SGBC!^<9 zPb$6htT=)cq912U4T)kR`@$oMK?TGpXjStbLUkLq1%9)2gV&kvyb0^u&Rqx8>%$K0Dhbj4wb{nmU^EtYumHS zuaWmZ2L229&)}`#WuONdK+XavgVVr2BL9CLd>Y&V-VBa{{|I)1Q^5Do1KbDh1-F9d z134S;lgRxy133$D3Y-p31K)sVq5HSN0@w<+fbUVCPlCIF_z(O(7y#K5@LX^Xcp5kh z$ld^X*WZ`G?}2F`djXyaP62<9?Ei5f@AkU|yb|<50_2Q<%fVB?KcWx#0Qh6@YhWMv zN$_2C0e=Br0gi%oun)-j0AB?Ufe(ONz)OL==WibT40tv;1Dp;{110b`=mj1C9{_T8 z-~!kVo&`*9&Kv)>PGoG3=ERO73#2_}r0}E}&)g<7yQ6Q^IZoow%_}uJJUx4QZKKyLi*j-c zVwei4wfhx$Hi}k3Jnq=4DTz54XFZyoRx3Bw9g|Mza%Zxr%1gPn+LDJ>hNu;j%4jaK zuNGNQk8P1O3+rMt;k}5ih_ps(c_WW9z1WIcv#Kr?UVtacizFH#1-7=Un}(Z;21%DY z`^3DOr{M6Ta`@sIB(}~GFg7=DPL{M(mNTBAWmoF zOsn(|*^*GgWw%GB@7r=+pd_i$C|;r_E|;@y zQ#a;$v6pMHv^Uym)ej6B&SfAhQpHJ;Y9Oh9!ibbGb;o)h<7*=)5l0CJZ?({aGS926 zHSDbram&pFW(XL2pnfc^R7xQCYpXP^u{h8tdxs6=7daP~;gOQ%N_w8IbDAk%`0Pe?a( z$~+^dS7GDnt#Uk_-n6a@sK{}S>djuRr)?BnzeUz0NDGi5J*0%4MI~$)-xXo$Popbi ztxx~Xii=&FwDHplA1A(|UwFrJ1`V+HT0LnUpM$Eka^?gx$PGTzj~MbFOrI{VGx-(|SH!IYfwD|0g+t*#$3 zp3B*Syslh)NXlCVdhs@pz0mqW$ru+mx# zi@q{fp9?bNe6CuUIA4N!E`^ICnbp#I)==%@y5Ub~LY7)c%j=qhDk%T|ZzGRhgbXY4 z|MU5h>odsrBKLQJy!TIJ|FeMj06c~~|2N=H@Ma+A04xDH15nNb5PyKJU>tlFxxNot z;94N>11N*jz@y0a4}k~3hk%?DAie>w2A6|v;0MV0Uk4upuL3UubKnFZX9Ii`{3Uo5 zm<1C+-UaXjFyg?!0uIO{)vO{!&a Mr7cj4dTQYR4LJ#qbN~PV literal 0 HcmV?d00001 diff --git a/egs/tedlium3/ASR/pruned_transducer_stateless_d2v_v2/README.md b/egs/tedlium3/ASR/pruned_transducer_stateless_d2v_v2/README.md new file mode 100644 index 000000000..3eb99a7ba --- /dev/null +++ b/egs/tedlium3/ASR/pruned_transducer_stateless_d2v_v2/README.md @@ -0,0 +1,54 @@ +# data2vec-transducer + +| | test-clean | test-other | +| --- | --- | --- | +| greedy decoding | 2.88 | 6.69 | +| modified beam search | 2.76 | 6.37 | +| fast beam search | 2.82 | 6.59 | +- train command + +```bash +./pruned_transducer_stateless_d2v_v2/train.py \ + --wandb False \ + --input-strategy AudioSamples \ + --enable-spec-aug False \ + --multi-optim True \ + --start-epoch 1 \ + --world-size 4 \ + --num-epochs 30 \ + --full-libri 1 \ + --exp-dir ./pruned_transducer_stateless_d2v_v2/d2v-T \ + --max-duration 150 \ + --freeze-finetune-updates 3000 \ + --encoder-dim 768 \ + --decoder-dim 768 \ + --joiner-dim 768 \ + --use-fp16 1 \ + --peak-dec-lr 0.04175 \ + --peak-enc-lr 0.0003859 \ + --accum-grads 4 \ + --encoder-type d2v \ + --additional-block True \ + --prune-range 10 \ + --context-size 2 \ + --ctc-loss-scale 0.2 +``` + +- decode command + +```bash +for method in greedy_search modified_beam_search fast_beam_search; do + ./pruned_transducer_stateless_d2v_v2/decode.py \ + --input-strategy AudioSamples \ + --enable-spec-aug False \ + --additional-block True \ + --model-name epoch-27.pt \ + --exp-dir ./pruned_transducer_stateless_d2v_v2/960h_sweep_v3_388 \ + --max-duration 400 \ + --decoding-method $method \ + --max-sym-per-frame 1 \ + --encoder-type d2v \ + --encoder-dim 768 \ + --decoder-dim 768 \ + --joiner-dim 768 +``` \ No newline at end of file diff --git a/egs/tedlium3/ASR/pruned_transducer_stateless_d2v_v2/__init__.py b/egs/tedlium3/ASR/pruned_transducer_stateless_d2v_v2/__init__.py new file mode 100644 index 000000000..e69de29bb diff --git a/egs/tedlium3/ASR/pruned_transducer_stateless_d2v_v2/asr_datamodule.py b/egs/tedlium3/ASR/pruned_transducer_stateless_d2v_v2/asr_datamodule.py new file mode 100644 index 000000000..1ecda2668 --- /dev/null +++ b/egs/tedlium3/ASR/pruned_transducer_stateless_d2v_v2/asr_datamodule.py @@ -0,0 +1,559 @@ +# Copyright 2021 Piotr Żelasko +# Copyright 2022 Xiaomi Corporation (Author: Mingshuang Luo) +# +# See ../../../../LICENSE for clarification regarding multiple authors +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + + +import argparse +import inspect +import logging +from glob import glob +from functools import lru_cache +from pathlib import Path +from typing import Any, Dict, Optional + +import torch +from lhotse import CutSet, Fbank, FbankConfig, load_manifest, load_manifest_lazy +from lhotse.dataset import ( # noqa F401 for PrecomputedFeatures + CutConcatenate, + CutMix, + DynamicBucketingSampler, + K2SpeechRecognitionDataset, + PrecomputedFeatures, + SingleCutSampler, + SpecAugment, +) +from lhotse.dataset.input_strategies import ( # noqa F401 For AudioSamples + AudioSamples, + OnTheFlyFeatures, +) +from lhotse.utils import fix_random_seed +from torch.utils.data import DataLoader + +from icefall.utils import str2bool + + +class _SeedWorkers: + def __init__(self, seed: int): + self.seed = seed + + def __call__(self, worker_id: int): + fix_random_seed(self.seed + worker_id) + + +class LibriSpeechAsrDataModule: + """ + DataModule for k2 ASR experiments. + It assumes there is always one train and valid dataloader, + but there can be multiple test dataloaders (e.g. LibriSpeech test-clean + and test-other). + + It contains all the common data pipeline modules used in ASR + experiments, e.g.: + - dynamic batch size, + - bucketing samplers, + - cut concatenation, + - augmentation, + - on-the-fly feature extraction + + This class should be derived for specific corpora used in ASR tasks. + """ + + def __init__(self, args: argparse.Namespace): + self.args = args + + @classmethod + def add_arguments(cls, parser: argparse.ArgumentParser): + group = parser.add_argument_group( + title="ASR data related options", + description="These options are used for the preparation of " + "PyTorch DataLoaders from Lhotse CutSet's -- they control the " + "effective batch sizes, sampling strategies, applied data " + "augmentations, etc.", + ) + group.add_argument( + "--full-libri", + type=str2bool, + default=False, + help="When enabled, use 960h LibriSpeech. Otherwise, use 100h subset.", + ) + group.add_argument( + "--manifest-dir", + type=Path, + default=Path("data/fbank"), + help="Path to directory with train/valid/test cuts.", + ) + group.add_argument( + "--max-duration", + type=int, + default=250.0, + help="Maximum pooled recordings duration (seconds) in a " + "single batch. You can reduce it if it causes CUDA OOM.", + ) + group.add_argument( + "--bucketing-sampler", + type=str2bool, + default=True, + help="When enabled, the batches will come from buckets of " + "similar duration (saves padding frames).", + ) + group.add_argument( + "--num-buckets", + type=int, + default=30, + help="The number of buckets for the DynamicBucketingSampler" + "(you might want to increase it for larger datasets).", + ) + group.add_argument( + "--concatenate-cuts", + type=str2bool, + default=False, + help="When enabled, utterances (cuts) will be concatenated " + "to minimize the amount of padding.", + ) + group.add_argument( + "--duration-factor", + type=float, + default=1.0, + help="Determines the maximum duration of a concatenated cut " + "relative to the duration of the longest cut in a batch.", + ) + group.add_argument( + "--gap", + type=float, + default=1.0, + help="The amount of padding (in seconds) inserted between " + "concatenated cuts. This padding is filled with noise when " + "noise augmentation is used.", + ) + group.add_argument( + "--on-the-fly-feats", + type=str2bool, + default=False, + help="When enabled, use on-the-fly cut mixing and feature " + "extraction. Will drop existing precomputed feature manifests " + "if available.", + ) + group.add_argument( + "--shuffle", + type=str2bool, + default=True, + help="When enabled (=default), the examples will be " + "shuffled for each epoch.", + ) + group.add_argument( + "--drop-last", + type=str2bool, + default=True, + help="Whether to drop last batch. Used by sampler.", + ) + group.add_argument( + "--return-cuts", + type=str2bool, + default=True, + help="When enabled, each batch will have the " + "field: batch['supervisions']['cut'] with the cuts that " + "were used to construct it.", + ) + + group.add_argument( + "--num-workers", + type=int, + default=2, + help="The number of training dataloader workers that " + "collect the batches.", + ) + + group.add_argument( + "--enable-spec-aug", + type=str2bool, + default=False, + help="When enabled, use SpecAugment for training dataset.", + ) + + group.add_argument( + "--spec-aug-time-warp-factor", + type=int, + default=80, + help="Used only when --enable-spec-aug is True. " + "It specifies the factor for time warping in SpecAugment. " + "Larger values mean more warping. " + "A value less than 1 means to disable time warp.", + ) + + group.add_argument( + "--enable-musan", + type=str2bool, + default=True, + help="When enabled, select noise from MUSAN and mix it" + "with training dataset. ", + ) + + group.add_argument( + "--input-strategy", + type=str, + default="AudioSamples", + help="AudioSamples or PrecomputedFeatures", + ) + + group.add_argument( + "--spk-id", + type=int, + default=0, + ) + + group.add_argument( + "--prefix", + type=str, + default='vox', + ) + + def train_dataloaders( + self, + cuts_train: CutSet, + sampler_state_dict: Optional[Dict[str, Any]] = None, + ) -> DataLoader: + """ + Args: + cuts_train: + CutSet for training. + sampler_state_dict: + The state dict for the training sampler. + """ + transforms = [] + if self.args.enable_musan: + logging.info("Enable MUSAN") + logging.info("About to get Musan cuts") + cuts_musan = load_manifest(self.args.manifest_dir / "musan_cuts.jsonl.gz") + transforms.append( + CutMix(cuts=cuts_musan, prob=0.5, snr=(10, 20), preserve_id=True) + ) + else: + logging.info("Disable MUSAN") + + if self.args.concatenate_cuts: + logging.info( + f"Using cut concatenation with duration factor " + f"{self.args.duration_factor} and gap {self.args.gap}." + ) + # Cut concatenation should be the first transform in the list, + # so that if we e.g. mix noise in, it will fill the gaps between + # different utterances. + transforms = [ + CutConcatenate( + duration_factor=self.args.duration_factor, gap=self.args.gap + ) + ] + transforms + + input_transforms = [] + if self.args.enable_spec_aug: + logging.info("Enable SpecAugment") + logging.info(f"Time warp factor: {self.args.spec_aug_time_warp_factor}") + # Set the value of num_frame_masks according to Lhotse's version. + # In different Lhotse's versions, the default of num_frame_masks is + # different. + num_frame_masks = 10 + num_frame_masks_parameter = inspect.signature( + SpecAugment.__init__ + ).parameters["num_frame_masks"] + if num_frame_masks_parameter.default == 1: + num_frame_masks = 2 + logging.info(f"Num frame mask: {num_frame_masks}") + input_transforms.append( + SpecAugment( + time_warp_factor=self.args.spec_aug_time_warp_factor, + num_frame_masks=num_frame_masks, + features_mask_size=27, + num_feature_masks=2, + frames_mask_size=100, + ) + ) + else: + logging.info("Disable SpecAugment") + + logging.info("About to create train dataset") + train = K2SpeechRecognitionDataset( + input_strategy=eval(self.args.input_strategy)(), + cut_transforms=transforms, + input_transforms=input_transforms, + return_cuts=self.args.return_cuts, + ) + + if self.args.on_the_fly_feats: + # NOTE: the PerturbSpeed transform should be added only if we + # remove it from data prep stage. + # Add on-the-fly speed perturbation; since originally it would + # have increased epoch size by 3, we will apply prob 2/3 and use + # 3x more epochs. + # Speed perturbation probably should come first before + # concatenation, but in principle the transforms order doesn't have + # to be strict (e.g. could be randomized) + # transforms = [PerturbSpeed(factors=[0.9, 1.1], p=2/3)] + transforms # noqa + # Drop feats to be on the safe side. + train = K2SpeechRecognitionDataset( + cut_transforms=transforms, + input_strategy=OnTheFlyFeatures(Fbank(FbankConfig(num_mel_bins=80))), + input_transforms=input_transforms, + return_cuts=self.args.return_cuts, + ) + + if self.args.bucketing_sampler: + logging.info("Using DynamicBucketingSampler.") + train_sampler = DynamicBucketingSampler( + cuts_train, + max_duration=self.args.max_duration, + shuffle=self.args.shuffle, + num_buckets=self.args.num_buckets, + drop_last=self.args.drop_last, + ) + else: + logging.info("Using SingleCutSampler.") + train_sampler = SingleCutSampler( + cuts_train, + max_duration=self.args.max_duration, + shuffle=self.args.shuffle, + ) + logging.info("About to create train dataloader") + + if sampler_state_dict is not None: + logging.info("Loading sampler state dict") + train_sampler.load_state_dict(sampler_state_dict) + + # 'seed' is derived from the current random state, which will have + # previously been set in the main process. + seed = torch.randint(0, 100000, ()).item() + worker_init_fn = _SeedWorkers(seed) + + train_dl = DataLoader( + train, + sampler=train_sampler, + batch_size=None, + num_workers=self.args.num_workers, + persistent_workers=False, + worker_init_fn=worker_init_fn, + ) + + return train_dl + + def valid_dataloaders(self, cuts_valid: CutSet) -> DataLoader: + transforms = [] + if self.args.concatenate_cuts: + transforms = [ + CutConcatenate( + duration_factor=self.args.duration_factor, gap=self.args.gap + ) + ] + transforms + + logging.info("About to create dev dataset") + if self.args.on_the_fly_feats: + validate = K2SpeechRecognitionDataset( + cut_transforms=transforms, + input_strategy=eval(self.args.input_strategy)(), + #input_strategy=OnTheFlyFeatures(Fbank(FbankConfig(num_mel_bins=80))), + return_cuts=self.args.return_cuts, + ) + else: + validate = K2SpeechRecognitionDataset( + cut_transforms=transforms, + input_strategy=eval(self.args.input_strategy)(), + return_cuts=self.args.return_cuts, + ) + valid_sampler = DynamicBucketingSampler( + cuts_valid, + max_duration=self.args.max_duration, + shuffle=False, + ) + logging.info("About to create dev dataloader") + valid_dl = DataLoader( + validate, + sampler=valid_sampler, + batch_size=None, + num_workers=2, + persistent_workers=False, + ) + + return valid_dl + + def test_dataloaders(self, cuts: CutSet) -> DataLoader: + logging.debug("About to create test dataset") + test = K2SpeechRecognitionDataset( + input_strategy=OnTheFlyFeatures(Fbank(FbankConfig(num_mel_bins=80))) + if self.args.on_the_fly_feats + else eval(self.args.input_strategy)(), + return_cuts=self.args.return_cuts, + ) + sampler = DynamicBucketingSampler( + cuts, + max_duration=self.args.max_duration, + shuffle=False, + ) + logging.debug("About to create test dataloader") + test_dl = DataLoader( + test, + batch_size=None, + sampler=sampler, + num_workers=self.args.num_workers, + ) + return test_dl + + @lru_cache() + def train_clean_10_cuts(self, option=None) -> CutSet: + logging.info("About to get train-clean-10 cuts") + if option is None: + return load_manifest_lazy( + self.args.manifest_dir / f"librispeech_cuts_train-clean-100.jsonl" + ) + else: + return load_manifest_lazy( + self.args.manifest_dir / f"librispeech_cuts_train-clean-10_{option}.jsonl" + ) + + @lru_cache() + def train_clean_100_cuts(self, option=None) -> CutSet: + logging.info("About to get train-clean-100 cuts") + if option is None: + return load_manifest_lazy( + self.args.manifest_dir / f"librispeech_cuts_train-clean-100.jsonl" + ) + else: + return load_manifest_lazy( + self.args.manifest_dir / f"librispeech_cuts_train-clean-100_{option}.jsonl" + ) + + @lru_cache() + def train_clean_360_cuts(self, option=None) -> CutSet: + logging.info("About to get train-clean-360 cuts") + if option is None: + return load_manifest_lazy( + self.args.manifest_dir / f"librispeech_cuts_train-clean-360.jsonl" + ) + else: + return load_manifest_lazy( + self.args.manifest_dir / f"librispeech_cuts_train-clean-360_{option}.jsonl" + ) + + @lru_cache() + def train_other_500_cuts(self, option=None) -> CutSet: + logging.info("About to get train-other-500 cuts") + if option is None: + return load_manifest_lazy( + self.args.manifest_dir / f"librispeech_cuts_train-other-500.jsonl" + ) + else: + return load_manifest_lazy( + self.args.manifest_dir / f"librispeech_cuts_train-other-500_{option}.jsonl" + ) + + @lru_cache() + def train_all_shuf_cuts(self, option=None) -> CutSet: + logging.info( + "About to get the shuffled train-clean-100, \ + train-clean-360 and train-other-500 cuts" + ) + if option is None: + return load_manifest_lazy( + self.args.manifest_dir / f"librispeech_cuts_train-all-shuf.jsonl" + ) + else: + return load_manifest_lazy( + self.args.manifest_dir / f"librispeech_cuts_train-all-shuf_{option}.jsonl" + ) + + @lru_cache() + def dev_clean_cuts(self, option=None) -> CutSet: + logging.info("About to get dev-clean cuts") + if option is None: + return load_manifest_lazy( + self.args.manifest_dir / f"librispeech_cuts_dev-clean.jsonl" + ) + else: + return load_manifest_lazy( + self.args.manifest_dir / f"librispeech_cuts_dev-clean_{option}.jsonl" + ) + + @lru_cache() + def dev_other_cuts(self, option=None) -> CutSet: + logging.info("About to get dev-other cuts") + if option is None: + return load_manifest_lazy( + self.args.manifest_dir / f"librispeech_cuts_dev-other.jsonl" + ) + else: + return load_manifest_lazy( + self.args.manifest_dir / f"librispeech_cuts_dev-other_{option}.jsonl" + ) + + @lru_cache() + def test_clean_cuts(self, option=None) -> CutSet: + logging.info("About to get test-clean cuts") + if option is None: + return load_manifest_lazy( + self.args.manifest_dir / f"librispeech_cuts_test-clean.jsonl" + ) + elif option == 'user': + json_list = sorted(glob(str(self.args.manifest_dir) + "/userlibri/test-clean/*")) + spk_list = [json.split('/')[-1][:-6] for json in json_list] + + return [load_manifest_lazy(json) for json in json_list], spk_list + else: + return load_manifest_lazy( + self.args.manifest_dir / f"librispeech_cuts_test-clean_{option}.jsonl" + ) + + @lru_cache() + def test_other_cuts(self, option=None) -> CutSet: + logging.info("About to get test-other cuts") + if option is None: + return load_manifest_lazy( + self.args.manifest_dir / f"librispeech_cuts_test-other_{option}.jsonl" + ) + elif option == 'user': + json_list = sorted(glob(str(self.args.manifest_dir) + "/userlibri/test-other/*")) + spk_list = [json.split('/')[-1][:-6] for json in json_list] + + return [load_manifest_lazy(json) for json in json_list], spk_list + else: + return load_manifest_lazy( + self.args.manifest_dir / f"librispeech_cuts_test-other_{option}.jsonl" + ) + + @lru_cache() + def test_clean_user(self, option=None) -> CutSet: + logging.info("About to get test-clean user cuts") + return load_manifest_lazy( + self.args.manifest_dir / f"userlibri/test-clean_sampling/{option}.jsonl" + ) + + @lru_cache() + def test_other_user(self, option=None) -> CutSet: + logging.info("About to get test-other user cuts") + return load_manifest_lazy( + self.args.manifest_dir / f"userlibri/test-other_sampling/{option}.jsonl" + ) + + @lru_cache() + def vox_cuts(self, option=None) -> CutSet: + logging.info("About to get test-other user cuts") + return load_manifest_lazy( + self.args.manifest_dir / f"{self.args.prefix}_cuts_{option}.jsonl.gz" + ) + + @lru_cache() + def userlibri_cuts(self, option=None) -> CutSet: + logging.info("About to get userlibri cuts") + return load_manifest_lazy( + self.args.manifest_dir / f"{option}.jsonl" + ) + diff --git a/egs/tedlium3/ASR/pruned_transducer_stateless_d2v_v2/beam_search.py b/egs/tedlium3/ASR/pruned_transducer_stateless_d2v_v2/beam_search.py new file mode 100644 index 000000000..b324cc9b7 --- /dev/null +++ b/egs/tedlium3/ASR/pruned_transducer_stateless_d2v_v2/beam_search.py @@ -0,0 +1,2342 @@ +# Copyright 2021 Xiaomi Corp. (authors: Fangjun Kuang +# Xiaoyu Yang) +# +# See ../../../../LICENSE for clarification regarding multiple authors +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import warnings +from dataclasses import dataclass, field +from typing import Dict, List, Optional, Tuple, Union + +import k2 +import sentencepiece as spm +import torch +from model import Transducer + +from icefall import NgramLm, NgramLmStateCost +from icefall.decode import Nbest, one_best_decoding +from icefall.rnn_lm.model import RnnLmModel +from icefall.utils import ( + DecodingResults, + add_eos, + add_sos, + get_texts, + get_texts_with_timestamp, +) + + +def fast_beam_search_one_best( + model: Transducer, + decoding_graph: k2.Fsa, + encoder_out: torch.Tensor, + encoder_out_lens: torch.Tensor, + beam: float, + max_states: int, + max_contexts: int, + temperature: float = 1.0, + return_timestamps: bool = False, +) -> Union[List[List[int]], DecodingResults]: + """It limits the maximum number of symbols per frame to 1. + + A lattice is first obtained using fast beam search, and then + the shortest path within the lattice is used as the final output. + + Args: + model: + An instance of `Transducer`. + decoding_graph: + Decoding graph used for decoding, may be a TrivialGraph or a LG. + encoder_out: + A tensor of shape (N, T, C) from the encoder. + encoder_out_lens: + A tensor of shape (N,) containing the number of frames in `encoder_out` + before padding. + beam: + Beam value, similar to the beam used in Kaldi.. + max_states: + Max states per stream per frame. + max_contexts: + Max contexts pre stream per frame. + temperature: + Softmax temperature. + return_timestamps: + Whether to return timestamps. + Returns: + If return_timestamps is False, return the decoded result. + Else, return a DecodingResults object containing + decoded result and corresponding timestamps. + """ + lattice = fast_beam_search( + model=model, + decoding_graph=decoding_graph, + encoder_out=encoder_out, + encoder_out_lens=encoder_out_lens, + beam=beam, + max_states=max_states, + max_contexts=max_contexts, + temperature=temperature, + ) + + best_path = one_best_decoding(lattice) + + if not return_timestamps: + return get_texts(best_path) + else: + return get_texts_with_timestamp(best_path) + + +def fast_beam_search_nbest_LG( + model: Transducer, + decoding_graph: k2.Fsa, + encoder_out: torch.Tensor, + encoder_out_lens: torch.Tensor, + beam: float, + max_states: int, + max_contexts: int, + num_paths: int, + nbest_scale: float = 0.5, + use_double_scores: bool = True, + temperature: float = 1.0, + return_timestamps: bool = False, +) -> Union[List[List[int]], DecodingResults]: + """It limits the maximum number of symbols per frame to 1. + + The process to get the results is: + - (1) Use fast beam search to get a lattice + - (2) Select `num_paths` paths from the lattice using k2.random_paths() + - (3) Unique the selected paths + - (4) Intersect the selected paths with the lattice and compute the + shortest path from the intersection result + - (5) The path with the largest score is used as the decoding output. + + Args: + model: + An instance of `Transducer`. + decoding_graph: + Decoding graph used for decoding, may be a TrivialGraph or a LG. + encoder_out: + A tensor of shape (N, T, C) from the encoder. + encoder_out_lens: + A tensor of shape (N,) containing the number of frames in `encoder_out` + before padding. + beam: + Beam value, similar to the beam used in Kaldi.. + max_states: + Max states per stream per frame. + max_contexts: + Max contexts pre stream per frame. + num_paths: + Number of paths to extract from the decoded lattice. + nbest_scale: + It's the scale applied to the lattice.scores. A smaller value + yields more unique paths. + use_double_scores: + True to use double precision for computation. False to use + single precision. + temperature: + Softmax temperature. + return_timestamps: + Whether to return timestamps. + Returns: + If return_timestamps is False, return the decoded result. + Else, return a DecodingResults object containing + decoded result and corresponding timestamps. + """ + lattice = fast_beam_search( + model=model, + decoding_graph=decoding_graph, + encoder_out=encoder_out, + encoder_out_lens=encoder_out_lens, + beam=beam, + max_states=max_states, + max_contexts=max_contexts, + temperature=temperature, + ) + + nbest = Nbest.from_lattice( + lattice=lattice, + num_paths=num_paths, + use_double_scores=use_double_scores, + nbest_scale=nbest_scale, + ) + + # The following code is modified from nbest.intersect() + word_fsa = k2.invert(nbest.fsa) + if hasattr(lattice, "aux_labels"): + # delete token IDs as it is not needed + del word_fsa.aux_labels + word_fsa.scores.zero_() + word_fsa_with_epsilon_loops = k2.linear_fsa_with_self_loops(word_fsa) + path_to_utt_map = nbest.shape.row_ids(1) + + if hasattr(lattice, "aux_labels"): + # lattice has token IDs as labels and word IDs as aux_labels. + # inv_lattice has word IDs as labels and token IDs as aux_labels + inv_lattice = k2.invert(lattice) + inv_lattice = k2.arc_sort(inv_lattice) + else: + inv_lattice = k2.arc_sort(lattice) + + if inv_lattice.shape[0] == 1: + path_lattice = k2.intersect_device( + inv_lattice, + word_fsa_with_epsilon_loops, + b_to_a_map=torch.zeros_like(path_to_utt_map), + sorted_match_a=True, + ) + else: + path_lattice = k2.intersect_device( + inv_lattice, + word_fsa_with_epsilon_loops, + b_to_a_map=path_to_utt_map, + sorted_match_a=True, + ) + + # path_lattice has word IDs as labels and token IDs as aux_labels + path_lattice = k2.top_sort(k2.connect(path_lattice)) + tot_scores = path_lattice.get_tot_scores( + use_double_scores=use_double_scores, + log_semiring=True, # Note: we always use True + ) + # See https://github.com/k2-fsa/icefall/pull/420 for why + # we always use log_semiring=True + + ragged_tot_scores = k2.RaggedTensor(nbest.shape, tot_scores) + best_hyp_indexes = ragged_tot_scores.argmax() + best_path = k2.index_fsa(nbest.fsa, best_hyp_indexes) + + if not return_timestamps: + return get_texts(best_path) + else: + return get_texts_with_timestamp(best_path) + + +def fast_beam_search_nbest( + model: Transducer, + decoding_graph: k2.Fsa, + encoder_out: torch.Tensor, + encoder_out_lens: torch.Tensor, + beam: float, + max_states: int, + max_contexts: int, + num_paths: int, + nbest_scale: float = 0.5, + use_double_scores: bool = True, + temperature: float = 1.0, + return_timestamps: bool = False, +) -> Union[List[List[int]], DecodingResults]: + """It limits the maximum number of symbols per frame to 1. + + The process to get the results is: + - (1) Use fast beam search to get a lattice + - (2) Select `num_paths` paths from the lattice using k2.random_paths() + - (3) Unique the selected paths + - (4) Intersect the selected paths with the lattice and compute the + shortest path from the intersection result + - (5) The path with the largest score is used as the decoding output. + + Args: + model: + An instance of `Transducer`. + decoding_graph: + Decoding graph used for decoding, may be a TrivialGraph or a LG. + encoder_out: + A tensor of shape (N, T, C) from the encoder. + encoder_out_lens: + A tensor of shape (N,) containing the number of frames in `encoder_out` + before padding. + beam: + Beam value, similar to the beam used in Kaldi.. + max_states: + Max states per stream per frame. + max_contexts: + Max contexts pre stream per frame. + num_paths: + Number of paths to extract from the decoded lattice. + nbest_scale: + It's the scale applied to the lattice.scores. A smaller value + yields more unique paths. + use_double_scores: + True to use double precision for computation. False to use + single precision. + temperature: + Softmax temperature. + return_timestamps: + Whether to return timestamps. + Returns: + If return_timestamps is False, return the decoded result. + Else, return a DecodingResults object containing + decoded result and corresponding timestamps. + """ + lattice = fast_beam_search( + model=model, + decoding_graph=decoding_graph, + encoder_out=encoder_out, + encoder_out_lens=encoder_out_lens, + beam=beam, + max_states=max_states, + max_contexts=max_contexts, + temperature=temperature, + ) + + nbest = Nbest.from_lattice( + lattice=lattice, + num_paths=num_paths, + use_double_scores=use_double_scores, + nbest_scale=nbest_scale, + ) + + # at this point, nbest.fsa.scores are all zeros. + + nbest = nbest.intersect(lattice) + # Now nbest.fsa.scores contains acoustic scores + + max_indexes = nbest.tot_scores().argmax() + + best_path = k2.index_fsa(nbest.fsa, max_indexes) + + if not return_timestamps: + return get_texts(best_path) + else: + return get_texts_with_timestamp(best_path) + + +def fast_beam_search_nbest_oracle( + model: Transducer, + decoding_graph: k2.Fsa, + encoder_out: torch.Tensor, + encoder_out_lens: torch.Tensor, + beam: float, + max_states: int, + max_contexts: int, + num_paths: int, + ref_texts: List[List[int]], + use_double_scores: bool = True, + nbest_scale: float = 0.5, + temperature: float = 1.0, + return_timestamps: bool = False, +) -> Union[List[List[int]], DecodingResults]: + """It limits the maximum number of symbols per frame to 1. + + A lattice is first obtained using fast beam search, and then + we select `num_paths` linear paths from the lattice. The path + that has the minimum edit distance with the given reference transcript + is used as the output. + + This is the best result we can achieve for any nbest based rescoring + methods. + + Args: + model: + An instance of `Transducer`. + decoding_graph: + Decoding graph used for decoding, may be a TrivialGraph or a LG. + encoder_out: + A tensor of shape (N, T, C) from the encoder. + encoder_out_lens: + A tensor of shape (N,) containing the number of frames in `encoder_out` + before padding. + beam: + Beam value, similar to the beam used in Kaldi.. + max_states: + Max states per stream per frame. + max_contexts: + Max contexts pre stream per frame. + num_paths: + Number of paths to extract from the decoded lattice. + ref_texts: + A list-of-list of integers containing the reference transcripts. + If the decoding_graph is a trivial_graph, the integer ID is the + BPE token ID. + use_double_scores: + True to use double precision for computation. False to use + single precision. + nbest_scale: + It's the scale applied to the lattice.scores. A smaller value + yields more unique paths. + temperature: + Softmax temperature. + return_timestamps: + Whether to return timestamps. + Returns: + If return_timestamps is False, return the decoded result. + Else, return a DecodingResults object containing + decoded result and corresponding timestamps. + """ + lattice = fast_beam_search( + model=model, + decoding_graph=decoding_graph, + encoder_out=encoder_out, + encoder_out_lens=encoder_out_lens, + beam=beam, + max_states=max_states, + max_contexts=max_contexts, + temperature=temperature, + ) + + nbest = Nbest.from_lattice( + lattice=lattice, + num_paths=num_paths, + use_double_scores=use_double_scores, + nbest_scale=nbest_scale, + ) + + hyps = nbest.build_levenshtein_graphs() + refs = k2.levenshtein_graph(ref_texts, device=hyps.device) + + levenshtein_alignment = k2.levenshtein_alignment( + refs=refs, + hyps=hyps, + hyp_to_ref_map=nbest.shape.row_ids(1), + sorted_match_ref=True, + ) + + tot_scores = levenshtein_alignment.get_tot_scores( + use_double_scores=False, log_semiring=False + ) + ragged_tot_scores = k2.RaggedTensor(nbest.shape, tot_scores) + + max_indexes = ragged_tot_scores.argmax() + + best_path = k2.index_fsa(nbest.fsa, max_indexes) + + if not return_timestamps: + return get_texts(best_path) + else: + return get_texts_with_timestamp(best_path) + + +def fast_beam_search( + model: Transducer, + decoding_graph: k2.Fsa, + encoder_out: torch.Tensor, + encoder_out_lens: torch.Tensor, + beam: float, + max_states: int, + max_contexts: int, + temperature: float = 1.0, +) -> k2.Fsa: + """It limits the maximum number of symbols per frame to 1. + + Args: + model: + An instance of `Transducer`. + decoding_graph: + Decoding graph used for decoding, may be a TrivialGraph or a LG. + encoder_out: + A tensor of shape (N, T, C) from the encoder. + encoder_out_lens: + A tensor of shape (N,) containing the number of frames in `encoder_out` + before padding. + beam: + Beam value, similar to the beam used in Kaldi.. + max_states: + Max states per stream per frame. + max_contexts: + Max contexts pre stream per frame. + temperature: + Softmax temperature. + Returns: + Return an FsaVec with axes [utt][state][arc] containing the decoded + lattice. Note: When the input graph is a TrivialGraph, the returned + lattice is actually an acceptor. + """ + assert encoder_out.ndim == 3 + + context_size = model.decoder.context_size + vocab_size = model.decoder.vocab_size + + B, T, C = encoder_out.shape + + config = k2.RnntDecodingConfig( + vocab_size=vocab_size, + decoder_history_len=context_size, + beam=beam, + max_contexts=max_contexts, + max_states=max_states, + ) + individual_streams = [] + for i in range(B): + individual_streams.append(k2.RnntDecodingStream(decoding_graph)) + decoding_streams = k2.RnntDecodingStreams(individual_streams, config) + + encoder_out = model.joiner.encoder_proj(encoder_out) + + for t in range(T): + # shape is a RaggedShape of shape (B, context) + # contexts is a Tensor of shape (shape.NumElements(), context_size) + shape, contexts = decoding_streams.get_contexts() + # `nn.Embedding()` in torch below v1.7.1 supports only torch.int64 + contexts = contexts.to(torch.int64) + # decoder_out is of shape (shape.NumElements(), 1, decoder_out_dim) + decoder_out = model.decoder(contexts, need_pad=False) + decoder_out = model.joiner.decoder_proj(decoder_out) + # current_encoder_out is of shape + # (shape.NumElements(), 1, joiner_dim) + # fmt: off + current_encoder_out = torch.index_select( + encoder_out[:, t:t + 1, :], 0, shape.row_ids(1).to(torch.int64) + ) + # fmt: on + logits = model.joiner( + current_encoder_out.unsqueeze(2), + decoder_out.unsqueeze(1), + project_input=False, + ) + logits = logits.squeeze(1).squeeze(1) + log_probs = (logits / temperature).log_softmax(dim=-1) + decoding_streams.advance(log_probs) + decoding_streams.terminate_and_flush_to_streams() + lattice = decoding_streams.format_output(encoder_out_lens.tolist()) + + return lattice + + +def greedy_search( + model: Transducer, + encoder_out: torch.Tensor, + max_sym_per_frame: int, + return_timestamps: bool = False, +) -> Union[List[int], DecodingResults]: + """Greedy search for a single utterance. + Args: + model: + An instance of `Transducer`. + encoder_out: + A tensor of shape (N, T, C) from the encoder. Support only N==1 for now. + max_sym_per_frame: + Maximum number of symbols per frame. If it is set to 0, the WER + would be 100%. + return_timestamps: + Whether to return timestamps. + Returns: + If return_timestamps is False, return the decoded result. + Else, return a DecodingResults object containing + decoded result and corresponding timestamps. + """ + assert encoder_out.ndim == 3 + + # support only batch_size == 1 for now + assert encoder_out.size(0) == 1, encoder_out.size(0) + + blank_id = model.decoder.blank_id + context_size = model.decoder.context_size + unk_id = getattr(model, "unk_id", blank_id) + + device = next(model.parameters()).device + + decoder_input = torch.tensor( + [-1] * (context_size - 1) + [blank_id], device=device, dtype=torch.int64 + ).reshape(1, context_size) + + decoder_out = model.decoder(decoder_input, need_pad=False) + decoder_out = model.joiner.decoder_proj(decoder_out) + + encoder_out = model.joiner.encoder_proj(encoder_out) + + T = encoder_out.size(1) + t = 0 + hyp = [blank_id] * context_size + + # timestamp[i] is the frame index after subsampling + # on which hyp[i] is decoded + timestamp = [] + + # Maximum symbols per utterance. + max_sym_per_utt = 1000 + + # symbols per frame + sym_per_frame = 0 + + # symbols per utterance decoded so far + sym_per_utt = 0 + + while t < T and sym_per_utt < max_sym_per_utt: + if sym_per_frame >= max_sym_per_frame: + sym_per_frame = 0 + t += 1 + continue + + # fmt: off + current_encoder_out = encoder_out[:, t:t+1, :].unsqueeze(2) + # fmt: on + logits = model.joiner( + current_encoder_out, decoder_out.unsqueeze(1), project_input=False + ) + # logits is (1, 1, 1, vocab_size) + + y = logits.argmax().item() + if y not in (blank_id, unk_id): + hyp.append(y) + timestamp.append(t) + decoder_input = torch.tensor([hyp[-context_size:]], device=device).reshape( + 1, context_size + ) + + decoder_out = model.decoder(decoder_input, need_pad=False) + decoder_out = model.joiner.decoder_proj(decoder_out) + + sym_per_utt += 1 + sym_per_frame += 1 + else: + sym_per_frame = 0 + t += 1 + hyp = hyp[context_size:] # remove blanks + + if not return_timestamps: + return hyp + else: + return DecodingResults( + hyps=[hyp], + timestamps=[timestamp], + ) + + +def greedy_search_batch( + model: Transducer, + encoder_out: torch.Tensor, + encoder_out_lens: torch.Tensor, + return_timestamps: bool = False, +) -> Union[List[List[int]], DecodingResults]: + """Greedy search in batch mode. It hardcodes --max-sym-per-frame=1. + Args: + model: + The transducer model. + encoder_out: + Output from the encoder. Its shape is (N, T, C), where N >= 1. + encoder_out_lens: + A 1-D tensor of shape (N,), containing number of valid frames in + encoder_out before padding. + return_timestamps: + Whether to return timestamps. + Returns: + If return_timestamps is False, return the decoded result. + Else, return a DecodingResults object containing + decoded result and corresponding timestamps. + """ + assert encoder_out.ndim == 3 + assert encoder_out.size(0) >= 1, encoder_out.size(0) + + packed_encoder_out = torch.nn.utils.rnn.pack_padded_sequence( + input=encoder_out, + lengths=encoder_out_lens.cpu(), + batch_first=True, + enforce_sorted=False, + ) + + device = next(model.parameters()).device + + blank_id = model.decoder.blank_id + unk_id = getattr(model, "unk_id", blank_id) + context_size = model.decoder.context_size + + batch_size_list = packed_encoder_out.batch_sizes.tolist() + N = encoder_out.size(0) + assert torch.all(encoder_out_lens > 0), encoder_out_lens + assert N == batch_size_list[0], (N, batch_size_list) + + hyps = [[-1] * (context_size - 1) + [blank_id] for _ in range(N)] + + # timestamp[n][i] is the frame index after subsampling + # on which hyp[n][i] is decoded + timestamps = [[] for _ in range(N)] + + decoder_input = torch.tensor( + hyps, + device=device, + dtype=torch.int64, + ) # (N, context_size) + + decoder_out = model.decoder(decoder_input, need_pad=False) + decoder_out = model.joiner.decoder_proj(decoder_out) + # decoder_out: (N, 1, decoder_out_dim) + + encoder_out = model.joiner.encoder_proj(packed_encoder_out.data) + + offset = 0 + for (t, batch_size) in enumerate(batch_size_list): + start = offset + end = offset + batch_size + current_encoder_out = encoder_out.data[start:end] + current_encoder_out = current_encoder_out.unsqueeze(1).unsqueeze(1) + # current_encoder_out's shape: (batch_size, 1, 1, encoder_out_dim) + offset = end + + decoder_out = decoder_out[:batch_size] + + logits = model.joiner( + current_encoder_out, decoder_out.unsqueeze(1), project_input=False + ) + # logits'shape (batch_size, 1, 1, vocab_size) + + logits = logits.squeeze(1).squeeze(1) # (batch_size, vocab_size) + assert logits.ndim == 2, logits.shape + y = logits.argmax(dim=1).tolist() + emitted = False + for i, v in enumerate(y): + if v not in (blank_id, unk_id): + hyps[i].append(v) + timestamps[i].append(t) + emitted = True + if emitted: + # update decoder output + decoder_input = [h[-context_size:] for h in hyps[:batch_size]] + decoder_input = torch.tensor( + decoder_input, + device=device, + dtype=torch.int64, + ) + decoder_out = model.decoder(decoder_input, need_pad=False) + decoder_out = model.joiner.decoder_proj(decoder_out) + + sorted_ans = [h[context_size:] for h in hyps] + ans = [] + ans_timestamps = [] + unsorted_indices = packed_encoder_out.unsorted_indices.tolist() + for i in range(N): + ans.append(sorted_ans[unsorted_indices[i]]) + ans_timestamps.append(timestamps[unsorted_indices[i]]) + + if not return_timestamps: + return ans + else: + return DecodingResults( + hyps=ans, + timestamps=ans_timestamps, + ) + + +@dataclass +class Hypothesis: + # The predicted tokens so far. + # Newly predicted tokens are appended to `ys`. + ys: List[int] + + # The log prob of ys. + # It contains only one entry. + log_prob: torch.Tensor + + # timestamp[i] is the frame index after subsampling + # on which ys[i] is decoded + timestamp: List[int] = field(default_factory=list) + + # the lm score for next token given the current ys + lm_score: Optional[torch.Tensor] = None + + # the RNNLM states (h and c in LSTM) + state: Optional[Tuple[torch.Tensor, torch.Tensor]] = None + + # N-gram LM state + state_cost: Optional[NgramLmStateCost] = None + + @property + def key(self) -> str: + """Return a string representation of self.ys""" + return "_".join(map(str, self.ys)) + + +class HypothesisList(object): + def __init__(self, data: Optional[Dict[str, Hypothesis]] = None) -> None: + """ + Args: + data: + A dict of Hypotheses. Its key is its `value.key`. + """ + if data is None: + self._data = {} + else: + self._data = data + + @property + def data(self) -> Dict[str, Hypothesis]: + return self._data + + def add(self, hyp: Hypothesis) -> None: + """Add a Hypothesis to `self`. + + If `hyp` already exists in `self`, its probability is updated using + `log-sum-exp` with the existed one. + + Args: + hyp: + The hypothesis to be added. + """ + key = hyp.key + if key in self: + old_hyp = self._data[key] # shallow copy + torch.logaddexp(old_hyp.log_prob, hyp.log_prob, out=old_hyp.log_prob) + else: + self._data[key] = hyp + + def get_most_probable(self, length_norm: bool = False) -> Hypothesis: + """Get the most probable hypothesis, i.e., the one with + the largest `log_prob`. + + Args: + length_norm: + If True, the `log_prob` of a hypothesis is normalized by the + number of tokens in it. + Returns: + Return the hypothesis that has the largest `log_prob`. + """ + if length_norm: + return max(self._data.values(), key=lambda hyp: hyp.log_prob / len(hyp.ys)) + else: + return max(self._data.values(), key=lambda hyp: hyp.log_prob) + + def remove(self, hyp: Hypothesis) -> None: + """Remove a given hypothesis. + + Caution: + `self` is modified **in-place**. + + Args: + hyp: + The hypothesis to be removed from `self`. + Note: It must be contained in `self`. Otherwise, + an exception is raised. + """ + key = hyp.key + assert key in self, f"{key} does not exist" + del self._data[key] + + def filter(self, threshold: torch.Tensor) -> "HypothesisList": + """Remove all Hypotheses whose log_prob is less than threshold. + + Caution: + `self` is not modified. Instead, a new HypothesisList is returned. + + Returns: + Return a new HypothesisList containing all hypotheses from `self` + with `log_prob` being greater than the given `threshold`. + """ + ans = HypothesisList() + for _, hyp in self._data.items(): + if hyp.log_prob > threshold: + ans.add(hyp) # shallow copy + return ans + + def topk(self, k: int) -> "HypothesisList": + """Return the top-k hypothesis.""" + hyps = list(self._data.items()) + + hyps = sorted(hyps, key=lambda h: h[1].log_prob, reverse=True)[:k] + + ans = HypothesisList(dict(hyps)) + return ans + + def __contains__(self, key: str): + return key in self._data + + def __iter__(self): + return iter(self._data.values()) + + def __len__(self) -> int: + return len(self._data) + + def __str__(self) -> str: + s = [] + for key in self: + s.append(key) + return ", ".join(s) + + +def get_hyps_shape(hyps: List[HypothesisList]) -> k2.RaggedShape: + """Return a ragged shape with axes [utt][num_hyps]. + + Args: + hyps: + len(hyps) == batch_size. It contains the current hypothesis for + each utterance in the batch. + Returns: + Return a ragged shape with 2 axes [utt][num_hyps]. Note that + the shape is on CPU. + """ + num_hyps = [len(h) for h in hyps] + + # torch.cumsum() is inclusive sum, so we put a 0 at the beginning + # to get exclusive sum later. + num_hyps.insert(0, 0) + + num_hyps = torch.tensor(num_hyps) + row_splits = torch.cumsum(num_hyps, dim=0, dtype=torch.int32) + ans = k2.ragged.create_ragged_shape2( + row_splits=row_splits, cached_tot_size=row_splits[-1].item() + ) + return ans + + +def modified_beam_search( + model: Transducer, + encoder_out: torch.Tensor, + encoder_out_lens: torch.Tensor, + beam: int = 4, + temperature: float = 1.0, + return_timestamps: bool = False, +) -> Union[List[List[int]], DecodingResults]: + """Beam search in batch mode with --max-sym-per-frame=1 being hardcoded. + + Args: + model: + The transducer model. + encoder_out: + Output from the encoder. Its shape is (N, T, C). + encoder_out_lens: + A 1-D tensor of shape (N,), containing number of valid frames in + encoder_out before padding. + beam: + Number of active paths during the beam search. + temperature: + Softmax temperature. + return_timestamps: + Whether to return timestamps. + Returns: + If return_timestamps is False, return the decoded result. + Else, return a DecodingResults object containing + decoded result and corresponding timestamps. + """ + assert encoder_out.ndim == 3, encoder_out.shape + assert encoder_out.size(0) >= 1, encoder_out.size(0) + + packed_encoder_out = torch.nn.utils.rnn.pack_padded_sequence( + input=encoder_out, + lengths=encoder_out_lens.cpu(), + batch_first=True, + enforce_sorted=False, + ) + + blank_id = model.decoder.blank_id + unk_id = getattr(model, "unk_id", blank_id) + context_size = model.decoder.context_size + device = next(model.parameters()).device + + batch_size_list = packed_encoder_out.batch_sizes.tolist() + N = encoder_out.size(0) + assert torch.all(encoder_out_lens > 0), encoder_out_lens + assert N == batch_size_list[0], (N, batch_size_list) + + B = [HypothesisList() for _ in range(N)] + for i in range(N): + B[i].add( + Hypothesis( + ys=[blank_id] * context_size, + log_prob=torch.zeros(1, dtype=torch.float32, device=device), + timestamp=[], + ) + ) + + encoder_out = model.joiner.encoder_proj(packed_encoder_out.data) + + offset = 0 + finalized_B = [] + for (t, batch_size) in enumerate(batch_size_list): + start = offset + end = offset + batch_size + current_encoder_out = encoder_out.data[start:end] + current_encoder_out = current_encoder_out.unsqueeze(1).unsqueeze(1) + # current_encoder_out's shape is (batch_size, 1, 1, encoder_out_dim) + offset = end + + finalized_B = B[batch_size:] + finalized_B + B = B[:batch_size] + + hyps_shape = get_hyps_shape(B).to(device) + + A = [list(b) for b in B] + B = [HypothesisList() for _ in range(batch_size)] + + ys_log_probs = torch.cat( + [hyp.log_prob.reshape(1, 1) for hyps in A for hyp in hyps] + ) # (num_hyps, 1) + + decoder_input = torch.tensor( + [hyp.ys[-context_size:] for hyps in A for hyp in hyps], + device=device, + dtype=torch.int64, + ) # (num_hyps, context_size) + + decoder_out = model.decoder(decoder_input, need_pad=False).unsqueeze(1) + decoder_out = model.joiner.decoder_proj(decoder_out) + # decoder_out is of shape (num_hyps, 1, 1, joiner_dim) + + # Note: For torch 1.7.1 and below, it requires a torch.int64 tensor + # as index, so we use `to(torch.int64)` below. + current_encoder_out = torch.index_select( + current_encoder_out, + dim=0, + index=hyps_shape.row_ids(1).to(torch.int64), + ) # (num_hyps, 1, 1, encoder_out_dim) + + logits = model.joiner( + current_encoder_out, + decoder_out, + project_input=False, + ) # (num_hyps, 1, 1, vocab_size) + + logits = logits.squeeze(1).squeeze(1) # (num_hyps, vocab_size) + + log_probs = (logits / temperature).log_softmax(dim=-1) # (num_hyps, vocab_size) + + log_probs.add_(ys_log_probs) + + vocab_size = log_probs.size(-1) + + log_probs = log_probs.reshape(-1) + + row_splits = hyps_shape.row_splits(1) * vocab_size + log_probs_shape = k2.ragged.create_ragged_shape2( + row_splits=row_splits, cached_tot_size=log_probs.numel() + ) + ragged_log_probs = k2.RaggedTensor(shape=log_probs_shape, value=log_probs) + + for i in range(batch_size): + topk_log_probs, topk_indexes = ragged_log_probs[i].topk(beam) + + with warnings.catch_warnings(): + warnings.simplefilter("ignore") + topk_hyp_indexes = (topk_indexes // vocab_size).tolist() + topk_token_indexes = (topk_indexes % vocab_size).tolist() + + for k in range(len(topk_hyp_indexes)): + hyp_idx = topk_hyp_indexes[k] + hyp = A[i][hyp_idx] + + new_ys = hyp.ys[:] + new_token = topk_token_indexes[k] + new_timestamp = hyp.timestamp[:] + if new_token not in (blank_id, unk_id): + new_ys.append(new_token) + new_timestamp.append(t) + + new_log_prob = topk_log_probs[k] + new_hyp = Hypothesis( + ys=new_ys, log_prob=new_log_prob, timestamp=new_timestamp + ) + B[i].add(new_hyp) + + B = B + finalized_B + best_hyps = [b.get_most_probable(length_norm=True) for b in B] + + sorted_ans = [h.ys[context_size:] for h in best_hyps] + sorted_timestamps = [h.timestamp for h in best_hyps] + ans = [] + ans_timestamps = [] + unsorted_indices = packed_encoder_out.unsorted_indices.tolist() + for i in range(N): + ans.append(sorted_ans[unsorted_indices[i]]) + ans_timestamps.append(sorted_timestamps[unsorted_indices[i]]) + + if not return_timestamps: + return ans + else: + return DecodingResults( + hyps=ans, + timestamps=ans_timestamps, + ) + + +def _deprecated_modified_beam_search( + model: Transducer, + encoder_out: torch.Tensor, + beam: int = 4, + return_timestamps: bool = False, +) -> Union[List[int], DecodingResults]: + """It limits the maximum number of symbols per frame to 1. + + It decodes only one utterance at a time. We keep it only for reference. + The function :func:`modified_beam_search` should be preferred as it + supports batch decoding. + + + Args: + model: + An instance of `Transducer`. + encoder_out: + A tensor of shape (N, T, C) from the encoder. Support only N==1 for now. + beam: + Beam size. + return_timestamps: + Whether to return timestamps. + + Returns: + If return_timestamps is False, return the decoded result. + Else, return a DecodingResults object containing + decoded result and corresponding timestamps. + """ + + assert encoder_out.ndim == 3 + + # support only batch_size == 1 for now + assert encoder_out.size(0) == 1, encoder_out.size(0) + blank_id = model.decoder.blank_id + unk_id = getattr(model, "unk_id", blank_id) + context_size = model.decoder.context_size + + device = next(model.parameters()).device + + T = encoder_out.size(1) + + B = HypothesisList() + B.add( + Hypothesis( + ys=[blank_id] * context_size, + log_prob=torch.zeros(1, dtype=torch.float32, device=device), + timestamp=[], + ) + ) + encoder_out = model.joiner.encoder_proj(encoder_out) + + for t in range(T): + # fmt: off + current_encoder_out = encoder_out[:, t:t+1, :].unsqueeze(2) + # current_encoder_out is of shape (1, 1, 1, encoder_out_dim) + # fmt: on + A = list(B) + B = HypothesisList() + + ys_log_probs = torch.cat([hyp.log_prob.reshape(1, 1) for hyp in A]) + # ys_log_probs is of shape (num_hyps, 1) + + decoder_input = torch.tensor( + [hyp.ys[-context_size:] for hyp in A], + device=device, + dtype=torch.int64, + ) + # decoder_input is of shape (num_hyps, context_size) + + decoder_out = model.decoder(decoder_input, need_pad=False).unsqueeze(1) + decoder_out = model.joiner.decoder_proj(decoder_out) + # decoder_output is of shape (num_hyps, 1, 1, joiner_dim) + + current_encoder_out = current_encoder_out.expand( + decoder_out.size(0), 1, 1, -1 + ) # (num_hyps, 1, 1, encoder_out_dim) + + logits = model.joiner( + current_encoder_out, + decoder_out, + project_input=False, + ) + # logits is of shape (num_hyps, 1, 1, vocab_size) + logits = logits.squeeze(1).squeeze(1) + + # now logits is of shape (num_hyps, vocab_size) + log_probs = logits.log_softmax(dim=-1) + + log_probs.add_(ys_log_probs) + + log_probs = log_probs.reshape(-1) + topk_log_probs, topk_indexes = log_probs.topk(beam) + + # topk_hyp_indexes are indexes into `A` + topk_hyp_indexes = topk_indexes // logits.size(-1) + topk_token_indexes = topk_indexes % logits.size(-1) + + with warnings.catch_warnings(): + warnings.simplefilter("ignore") + topk_hyp_indexes = topk_hyp_indexes.tolist() + topk_token_indexes = topk_token_indexes.tolist() + + for i in range(len(topk_hyp_indexes)): + hyp = A[topk_hyp_indexes[i]] + new_ys = hyp.ys[:] + new_timestamp = hyp.timestamp[:] + new_token = topk_token_indexes[i] + if new_token not in (blank_id, unk_id): + new_ys.append(new_token) + new_timestamp.append(t) + new_log_prob = topk_log_probs[i] + new_hyp = Hypothesis( + ys=new_ys, log_prob=new_log_prob, timestamp=new_timestamp + ) + B.add(new_hyp) + + best_hyp = B.get_most_probable(length_norm=True) + ys = best_hyp.ys[context_size:] # [context_size:] to remove blanks + + if not return_timestamps: + return ys + else: + return DecodingResults(hyps=[ys], timestamps=[best_hyp.timestamp]) + + +def beam_search( + model: Transducer, + encoder_out: torch.Tensor, + beam: int = 4, + temperature: float = 1.0, + return_timestamps: bool = False, +) -> Union[List[int], DecodingResults]: + """ + It implements Algorithm 1 in https://arxiv.org/pdf/1211.3711.pdf + + espnet/nets/beam_search_transducer.py#L247 is used as a reference. + + Args: + model: + An instance of `Transducer`. + encoder_out: + A tensor of shape (N, T, C) from the encoder. Support only N==1 for now. + beam: + Beam size. + temperature: + Softmax temperature. + return_timestamps: + Whether to return timestamps. + + Returns: + If return_timestamps is False, return the decoded result. + Else, return a DecodingResults object containing + decoded result and corresponding timestamps. + """ + assert encoder_out.ndim == 3 + + # support only batch_size == 1 for now + assert encoder_out.size(0) == 1, encoder_out.size(0) + blank_id = model.decoder.blank_id + unk_id = getattr(model, "unk_id", blank_id) + context_size = model.decoder.context_size + + device = next(model.parameters()).device + + decoder_input = torch.tensor( + [blank_id] * context_size, + device=device, + dtype=torch.int64, + ).reshape(1, context_size) + + decoder_out = model.decoder(decoder_input, need_pad=False) + decoder_out = model.joiner.decoder_proj(decoder_out) + + encoder_out = model.joiner.encoder_proj(encoder_out) + + T = encoder_out.size(1) + t = 0 + + B = HypothesisList() + B.add(Hypothesis(ys=[blank_id] * context_size, log_prob=0.0, timestamp=[])) + + max_sym_per_utt = 20000 + + sym_per_utt = 0 + + decoder_cache: Dict[str, torch.Tensor] = {} + + while t < T and sym_per_utt < max_sym_per_utt: + # fmt: off + current_encoder_out = encoder_out[:, t:t+1, :].unsqueeze(2) + # fmt: on + A = B + B = HypothesisList() + + joint_cache: Dict[str, torch.Tensor] = {} + + # TODO(fangjun): Implement prefix search to update the `log_prob` + # of hypotheses in A + + while True: + y_star = A.get_most_probable() + A.remove(y_star) + + cached_key = y_star.key + + if cached_key not in decoder_cache: + decoder_input = torch.tensor( + [y_star.ys[-context_size:]], + device=device, + dtype=torch.int64, + ).reshape(1, context_size) + + decoder_out = model.decoder(decoder_input, need_pad=False) + decoder_out = model.joiner.decoder_proj(decoder_out) + decoder_cache[cached_key] = decoder_out + else: + decoder_out = decoder_cache[cached_key] + + cached_key += f"-t-{t}" + if cached_key not in joint_cache: + logits = model.joiner( + current_encoder_out, + decoder_out.unsqueeze(1), + project_input=False, + ) + + # TODO(fangjun): Scale the blank posterior + log_prob = (logits / temperature).log_softmax(dim=-1) + # log_prob is (1, 1, 1, vocab_size) + log_prob = log_prob.squeeze() + # Now log_prob is (vocab_size,) + joint_cache[cached_key] = log_prob + else: + log_prob = joint_cache[cached_key] + + # First, process the blank symbol + skip_log_prob = log_prob[blank_id] + new_y_star_log_prob = y_star.log_prob + skip_log_prob + + # ys[:] returns a copy of ys + B.add( + Hypothesis( + ys=y_star.ys[:], + log_prob=new_y_star_log_prob, + timestamp=y_star.timestamp[:], + ) + ) + + # Second, process other non-blank labels + values, indices = log_prob.topk(beam + 1) + for i, v in zip(indices.tolist(), values.tolist()): + if i in (blank_id, unk_id): + continue + new_ys = y_star.ys + [i] + new_log_prob = y_star.log_prob + v + new_timestamp = y_star.timestamp + [t] + A.add( + Hypothesis( + ys=new_ys, + log_prob=new_log_prob, + timestamp=new_timestamp, + ) + ) + + # Check whether B contains more than "beam" elements more probable + # than the most probable in A + A_most_probable = A.get_most_probable() + + kept_B = B.filter(A_most_probable.log_prob) + + if len(kept_B) >= beam: + B = kept_B.topk(beam) + break + + t += 1 + + best_hyp = B.get_most_probable(length_norm=True) + ys = best_hyp.ys[context_size:] # [context_size:] to remove blanks + + if not return_timestamps: + return ys + else: + return DecodingResults(hyps=[ys], timestamps=[best_hyp.timestamp]) + + +def fast_beam_search_with_nbest_rescoring( + model: Transducer, + decoding_graph: k2.Fsa, + encoder_out: torch.Tensor, + encoder_out_lens: torch.Tensor, + beam: float, + max_states: int, + max_contexts: int, + ngram_lm_scale_list: List[float], + num_paths: int, + G: k2.Fsa, + sp: spm.SentencePieceProcessor, + word_table: k2.SymbolTable, + oov_word: str = "", + use_double_scores: bool = True, + nbest_scale: float = 0.5, + temperature: float = 1.0, + return_timestamps: bool = False, +) -> Dict[str, Union[List[List[int]], DecodingResults]]: + """It limits the maximum number of symbols per frame to 1. + A lattice is first obtained using fast beam search, num_path are selected + and rescored using a given language model. The shortest path within the + lattice is used as the final output. + + Args: + model: + An instance of `Transducer`. + decoding_graph: + Decoding graph used for decoding, may be a TrivialGraph or a LG. + encoder_out: + A tensor of shape (N, T, C) from the encoder. + encoder_out_lens: + A tensor of shape (N,) containing the number of frames in `encoder_out` + before padding. + beam: + Beam value, similar to the beam used in Kaldi. + max_states: + Max states per stream per frame. + max_contexts: + Max contexts pre stream per frame. + ngram_lm_scale_list: + A list of floats representing LM score scales. + num_paths: + Number of paths to extract from the decoded lattice. + G: + An FsaVec containing only a single FSA. It is an n-gram LM. + sp: + The BPE model. + word_table: + The word symbol table. + oov_word: + OOV words are replaced with this word. + use_double_scores: + True to use double precision for computation. False to use + single precision. + nbest_scale: + It's the scale applied to the lattice.scores. A smaller value + yields more unique paths. + temperature: + Softmax temperature. + return_timestamps: + Whether to return timestamps. + Returns: + Return the decoded result in a dict, where the key has the form + 'ngram_lm_scale_xx' and the value is the decoded results + optionally with timestamps. `xx` is the ngram LM scale value + used during decoding, i.e., 0.1. + """ + lattice = fast_beam_search( + model=model, + decoding_graph=decoding_graph, + encoder_out=encoder_out, + encoder_out_lens=encoder_out_lens, + beam=beam, + max_states=max_states, + max_contexts=max_contexts, + temperature=temperature, + ) + + nbest = Nbest.from_lattice( + lattice=lattice, + num_paths=num_paths, + use_double_scores=use_double_scores, + nbest_scale=nbest_scale, + ) + # at this point, nbest.fsa.scores are all zeros. + + nbest = nbest.intersect(lattice) + # Now nbest.fsa.scores contains acoustic scores + + am_scores = nbest.tot_scores() + + # Now we need to compute the LM scores of each path. + # (1) Get the token IDs of each Path. We assume the decoding_graph + # is an acceptor, i.e., lattice is also an acceptor + tokens_shape = nbest.fsa.arcs.shape().remove_axis(1) # [path][arc] + + tokens = k2.RaggedTensor(tokens_shape, nbest.fsa.labels.contiguous()) + tokens = tokens.remove_values_leq(0) # remove -1 and 0 + + token_list: List[List[int]] = tokens.tolist() + word_list: List[List[str]] = sp.decode(token_list) + + assert isinstance(oov_word, str), oov_word + assert oov_word in word_table, oov_word + oov_word_id = word_table[oov_word] + + word_ids_list: List[List[int]] = [] + + for words in word_list: + this_word_ids = [] + for w in words.split(): + if w in word_table: + this_word_ids.append(word_table[w]) + else: + this_word_ids.append(oov_word_id) + word_ids_list.append(this_word_ids) + + word_fsas = k2.linear_fsa(word_ids_list, device=lattice.device) + word_fsas_with_self_loops = k2.add_epsilon_self_loops(word_fsas) + + num_unique_paths = len(word_ids_list) + + b_to_a_map = torch.zeros( + num_unique_paths, + dtype=torch.int32, + device=lattice.device, + ) + + rescored_word_fsas = k2.intersect_device( + a_fsas=G, + b_fsas=word_fsas_with_self_loops, + b_to_a_map=b_to_a_map, + sorted_match_a=True, + ret_arc_maps=False, + ) + + rescored_word_fsas = k2.remove_epsilon_self_loops(rescored_word_fsas) + rescored_word_fsas = k2.top_sort(k2.connect(rescored_word_fsas)) + ngram_lm_scores = rescored_word_fsas.get_tot_scores( + use_double_scores=True, + log_semiring=False, + ) + + ans: Dict[str, Union[List[List[int]], DecodingResults]] = {} + for s in ngram_lm_scale_list: + key = f"ngram_lm_scale_{s}" + tot_scores = am_scores.values + s * ngram_lm_scores + ragged_tot_scores = k2.RaggedTensor(nbest.shape, tot_scores) + max_indexes = ragged_tot_scores.argmax() + best_path = k2.index_fsa(nbest.fsa, max_indexes) + + if not return_timestamps: + ans[key] = get_texts(best_path) + else: + ans[key] = get_texts_with_timestamp(best_path) + + return ans + + +def fast_beam_search_with_nbest_rnn_rescoring( + model: Transducer, + decoding_graph: k2.Fsa, + encoder_out: torch.Tensor, + encoder_out_lens: torch.Tensor, + beam: float, + max_states: int, + max_contexts: int, + ngram_lm_scale_list: List[float], + num_paths: int, + G: k2.Fsa, + sp: spm.SentencePieceProcessor, + word_table: k2.SymbolTable, + rnn_lm_model: torch.nn.Module, + rnn_lm_scale_list: List[float], + oov_word: str = "", + use_double_scores: bool = True, + nbest_scale: float = 0.5, + temperature: float = 1.0, + return_timestamps: bool = False, +) -> Dict[str, Union[List[List[int]], DecodingResults]]: + """It limits the maximum number of symbols per frame to 1. + A lattice is first obtained using fast beam search, num_path are selected + and rescored using a given language model and a rnn-lm. + The shortest path within the lattice is used as the final output. + + Args: + model: + An instance of `Transducer`. + decoding_graph: + Decoding graph used for decoding, may be a TrivialGraph or a LG. + encoder_out: + A tensor of shape (N, T, C) from the encoder. + encoder_out_lens: + A tensor of shape (N,) containing the number of frames in `encoder_out` + before padding. + beam: + Beam value, similar to the beam used in Kaldi. + max_states: + Max states per stream per frame. + max_contexts: + Max contexts pre stream per frame. + ngram_lm_scale_list: + A list of floats representing LM score scales. + num_paths: + Number of paths to extract from the decoded lattice. + G: + An FsaVec containing only a single FSA. It is an n-gram LM. + sp: + The BPE model. + word_table: + The word symbol table. + rnn_lm_model: + A rnn-lm model used for LM rescoring + rnn_lm_scale_list: + A list of floats representing RNN score scales. + oov_word: + OOV words are replaced with this word. + use_double_scores: + True to use double precision for computation. False to use + single precision. + nbest_scale: + It's the scale applied to the lattice.scores. A smaller value + yields more unique paths. + temperature: + Softmax temperature. + return_timestamps: + Whether to return timestamps. + Returns: + Return the decoded result in a dict, where the key has the form + 'ngram_lm_scale_xx' and the value is the decoded results + optionally with timestamps. `xx` is the ngram LM scale value + used during decoding, i.e., 0.1. + """ + lattice = fast_beam_search( + model=model, + decoding_graph=decoding_graph, + encoder_out=encoder_out, + encoder_out_lens=encoder_out_lens, + beam=beam, + max_states=max_states, + max_contexts=max_contexts, + temperature=temperature, + ) + + nbest = Nbest.from_lattice( + lattice=lattice, + num_paths=num_paths, + use_double_scores=use_double_scores, + nbest_scale=nbest_scale, + ) + # at this point, nbest.fsa.scores are all zeros. + + nbest = nbest.intersect(lattice) + # Now nbest.fsa.scores contains acoustic scores + + am_scores = nbest.tot_scores() + + # Now we need to compute the LM scores of each path. + # (1) Get the token IDs of each Path. We assume the decoding_graph + # is an acceptor, i.e., lattice is also an acceptor + tokens_shape = nbest.fsa.arcs.shape().remove_axis(1) # [path][arc] + + tokens = k2.RaggedTensor(tokens_shape, nbest.fsa.labels.contiguous()) + tokens = tokens.remove_values_leq(0) # remove -1 and 0 + + token_list: List[List[int]] = tokens.tolist() + word_list: List[List[str]] = sp.decode(token_list) + + assert isinstance(oov_word, str), oov_word + assert oov_word in word_table, oov_word + oov_word_id = word_table[oov_word] + + word_ids_list: List[List[int]] = [] + + for words in word_list: + this_word_ids = [] + for w in words.split(): + if w in word_table: + this_word_ids.append(word_table[w]) + else: + this_word_ids.append(oov_word_id) + word_ids_list.append(this_word_ids) + + word_fsas = k2.linear_fsa(word_ids_list, device=lattice.device) + word_fsas_with_self_loops = k2.add_epsilon_self_loops(word_fsas) + + num_unique_paths = len(word_ids_list) + + b_to_a_map = torch.zeros( + num_unique_paths, + dtype=torch.int32, + device=lattice.device, + ) + + rescored_word_fsas = k2.intersect_device( + a_fsas=G, + b_fsas=word_fsas_with_self_loops, + b_to_a_map=b_to_a_map, + sorted_match_a=True, + ret_arc_maps=False, + ) + + rescored_word_fsas = k2.remove_epsilon_self_loops(rescored_word_fsas) + rescored_word_fsas = k2.top_sort(k2.connect(rescored_word_fsas)) + ngram_lm_scores = rescored_word_fsas.get_tot_scores( + use_double_scores=True, + log_semiring=False, + ) + + # Now RNN-LM + blank_id = model.decoder.blank_id + sos_id = sp.piece_to_id("sos_id") + eos_id = sp.piece_to_id("eos_id") + + sos_tokens = add_sos(tokens, sos_id) + tokens_eos = add_eos(tokens, eos_id) + sos_tokens_row_splits = sos_tokens.shape.row_splits(1) + sentence_lengths = sos_tokens_row_splits[1:] - sos_tokens_row_splits[:-1] + + x_tokens = sos_tokens.pad(mode="constant", padding_value=blank_id) + y_tokens = tokens_eos.pad(mode="constant", padding_value=blank_id) + + x_tokens = x_tokens.to(torch.int64) + y_tokens = y_tokens.to(torch.int64) + sentence_lengths = sentence_lengths.to(torch.int64) + + rnn_lm_nll = rnn_lm_model(x=x_tokens, y=y_tokens, lengths=sentence_lengths) + assert rnn_lm_nll.ndim == 2 + assert rnn_lm_nll.shape[0] == len(token_list) + rnn_lm_scores = -1 * rnn_lm_nll.sum(dim=1) + + ans: Dict[str, List[List[int]]] = {} + for n_scale in ngram_lm_scale_list: + for rnn_scale in rnn_lm_scale_list: + key = f"ngram_lm_scale_{n_scale}_rnn_lm_scale_{rnn_scale}" + tot_scores = ( + am_scores.values + n_scale * ngram_lm_scores + rnn_scale * rnn_lm_scores + ) + ragged_tot_scores = k2.RaggedTensor(nbest.shape, tot_scores) + max_indexes = ragged_tot_scores.argmax() + best_path = k2.index_fsa(nbest.fsa, max_indexes) + + if not return_timestamps: + ans[key] = get_texts(best_path) + else: + ans[key] = get_texts_with_timestamp(best_path) + + return ans + + +def modified_beam_search_ngram_rescoring( + model: Transducer, + encoder_out: torch.Tensor, + encoder_out_lens: torch.Tensor, + ngram_lm: NgramLm, + ngram_lm_scale: float, + beam: int = 4, + temperature: float = 1.0, +) -> List[List[int]]: + """Beam search in batch mode with --max-sym-per-frame=1 being hardcoded. + + Args: + model: + The transducer model. + encoder_out: + Output from the encoder. Its shape is (N, T, C). + encoder_out_lens: + A 1-D tensor of shape (N,), containing number of valid frames in + encoder_out before padding. + beam: + Number of active paths during the beam search. + temperature: + Softmax temperature. + Returns: + Return a list-of-list of token IDs. ans[i] is the decoding results + for the i-th utterance. + """ + assert encoder_out.ndim == 3, encoder_out.shape + assert encoder_out.size(0) >= 1, encoder_out.size(0) + + packed_encoder_out = torch.nn.utils.rnn.pack_padded_sequence( + input=encoder_out, + lengths=encoder_out_lens.cpu(), + batch_first=True, + enforce_sorted=False, + ) + + blank_id = model.decoder.blank_id + unk_id = getattr(model, "unk_id", blank_id) + context_size = model.decoder.context_size + device = next(model.parameters()).device + lm_scale = ngram_lm_scale + + batch_size_list = packed_encoder_out.batch_sizes.tolist() + N = encoder_out.size(0) + assert torch.all(encoder_out_lens > 0), encoder_out_lens + assert N == batch_size_list[0], (N, batch_size_list) + + B = [HypothesisList() for _ in range(N)] + for i in range(N): + B[i].add( + Hypothesis( + ys=[blank_id] * context_size, + log_prob=torch.zeros(1, dtype=torch.float32, device=device), + state_cost=NgramLmStateCost(ngram_lm), + ) + ) + + encoder_out = model.joiner.encoder_proj(packed_encoder_out.data) + + offset = 0 + finalized_B = [] + for batch_size in batch_size_list: + start = offset + end = offset + batch_size + current_encoder_out = encoder_out.data[start:end] + current_encoder_out = current_encoder_out.unsqueeze(1).unsqueeze(1) + # current_encoder_out's shape is (batch_size, 1, 1, encoder_out_dim) + offset = end + + finalized_B = B[batch_size:] + finalized_B + B = B[:batch_size] + + hyps_shape = get_hyps_shape(B).to(device) + + A = [list(b) for b in B] + B = [HypothesisList() for _ in range(batch_size)] + + ys_log_probs = torch.cat( + [ + hyp.log_prob.reshape(1, 1) + hyp.state_cost.lm_score * lm_scale + for hyps in A + for hyp in hyps + ] + ) # (num_hyps, 1) + + decoder_input = torch.tensor( + [hyp.ys[-context_size:] for hyps in A for hyp in hyps], + device=device, + dtype=torch.int64, + ) # (num_hyps, context_size) + + decoder_out = model.decoder(decoder_input, need_pad=False).unsqueeze(1) + decoder_out = model.joiner.decoder_proj(decoder_out) + # decoder_out is of shape (num_hyps, 1, 1, joiner_dim) + + # Note: For torch 1.7.1 and below, it requires a torch.int64 tensor + # as index, so we use `to(torch.int64)` below. + current_encoder_out = torch.index_select( + current_encoder_out, + dim=0, + index=hyps_shape.row_ids(1).to(torch.int64), + ) # (num_hyps, 1, 1, encoder_out_dim) + + logits = model.joiner( + current_encoder_out, + decoder_out, + project_input=False, + ) # (num_hyps, 1, 1, vocab_size) + + logits = logits.squeeze(1).squeeze(1) # (num_hyps, vocab_size) + + log_probs = (logits / temperature).log_softmax(dim=-1) # (num_hyps, vocab_size) + + log_probs.add_(ys_log_probs) + vocab_size = log_probs.size(-1) + log_probs = log_probs.reshape(-1) + + row_splits = hyps_shape.row_splits(1) * vocab_size + log_probs_shape = k2.ragged.create_ragged_shape2( + row_splits=row_splits, cached_tot_size=log_probs.numel() + ) + ragged_log_probs = k2.RaggedTensor(shape=log_probs_shape, value=log_probs) + + for i in range(batch_size): + topk_log_probs, topk_indexes = ragged_log_probs[i].topk(beam) + + with warnings.catch_warnings(): + warnings.simplefilter("ignore") + topk_hyp_indexes = (topk_indexes // vocab_size).tolist() + topk_token_indexes = (topk_indexes % vocab_size).tolist() + + for k in range(len(topk_hyp_indexes)): + hyp_idx = topk_hyp_indexes[k] + hyp = A[i][hyp_idx] + + new_ys = hyp.ys[:] + new_token = topk_token_indexes[k] + if new_token not in (blank_id, unk_id): + new_ys.append(new_token) + state_cost = hyp.state_cost.forward_one_step(new_token) + else: + state_cost = hyp.state_cost + + # We only keep AM scores in new_hyp.log_prob + new_log_prob = topk_log_probs[k] - hyp.state_cost.lm_score * lm_scale + + new_hyp = Hypothesis( + ys=new_ys, log_prob=new_log_prob, state_cost=state_cost + ) + B[i].add(new_hyp) + + B = B + finalized_B + best_hyps = [b.get_most_probable(length_norm=True) for b in B] + + sorted_ans = [h.ys[context_size:] for h in best_hyps] + ans = [] + unsorted_indices = packed_encoder_out.unsorted_indices.tolist() + for i in range(N): + ans.append(sorted_ans[unsorted_indices[i]]) + + return ans + + +def modified_beam_search_rnnlm_shallow_fusion( + model: Transducer, + encoder_out: torch.Tensor, + encoder_out_lens: torch.Tensor, + sp: spm.SentencePieceProcessor, + rnnlm: RnnLmModel, + rnnlm_scale: float, + beam: int = 4, + return_timestamps: bool = False, +) -> List[List[int]]: + """Modified_beam_search + RNNLM shallow fusion + + Args: + model (Transducer): + The transducer model + encoder_out (torch.Tensor): + Encoder output in (N,T,C) + encoder_out_lens (torch.Tensor): + A 1-D tensor of shape (N,), containing the number of + valid frames in encoder_out before padding. + sp: + Sentence piece generator. + rnnlm (RnnLmModel): + RNNLM + rnnlm_scale (float): + scale of RNNLM in shallow fusion + beam (int, optional): + Beam size. Defaults to 4. + + Returns: + Return a list-of-list of token IDs. ans[i] is the decoding results + for the i-th utterance. + """ + assert encoder_out.ndim == 3, encoder_out.shape + assert encoder_out.size(0) >= 1, encoder_out.size(0) + assert rnnlm is not None + lm_scale = rnnlm_scale + vocab_size = rnnlm.vocab_size + + packed_encoder_out = torch.nn.utils.rnn.pack_padded_sequence( + input=encoder_out, + lengths=encoder_out_lens.cpu(), + batch_first=True, + enforce_sorted=False, + ) + + blank_id = model.decoder.blank_id + sos_id = sp.piece_to_id("") + unk_id = getattr(model, "unk_id", blank_id) + context_size = model.decoder.context_size + device = next(model.parameters()).device + + batch_size_list = packed_encoder_out.batch_sizes.tolist() + N = encoder_out.size(0) + assert torch.all(encoder_out_lens > 0), encoder_out_lens + assert N == batch_size_list[0], (N, batch_size_list) + + # get initial lm score and lm state by scoring the "sos" token + sos_token = torch.tensor([[sos_id]]).to(torch.int64).to(device) + init_score, init_states = rnnlm.score_token(sos_token) + + B = [HypothesisList() for _ in range(N)] + for i in range(N): + B[i].add( + Hypothesis( + ys=[blank_id] * context_size, + log_prob=torch.zeros(1, dtype=torch.float32, device=device), + state=init_states, + lm_score=init_score.reshape(-1), + timestamp=[], + ) + ) + + rnnlm.clean_cache() + encoder_out = model.joiner.encoder_proj(packed_encoder_out.data) + + offset = 0 + finalized_B = [] + for (t, batch_size) in enumerate(batch_size_list): + start = offset + end = offset + batch_size + current_encoder_out = encoder_out.data[start:end] # get batch + current_encoder_out = current_encoder_out.unsqueeze(1).unsqueeze(1) + # current_encoder_out's shape is (batch_size, 1, 1, encoder_out_dim) + offset = end + + finalized_B = B[batch_size:] + finalized_B + B = B[:batch_size] + + hyps_shape = get_hyps_shape(B).to(device) + + A = [list(b) for b in B] + B = [HypothesisList() for _ in range(batch_size)] + + ys_log_probs = torch.cat( + [hyp.log_prob.reshape(1, 1) for hyps in A for hyp in hyps] + ) + + decoder_input = torch.tensor( + [hyp.ys[-context_size:] for hyps in A for hyp in hyps], + device=device, + dtype=torch.int64, + ) # (num_hyps, context_size) + + decoder_out = model.decoder(decoder_input, need_pad=False).unsqueeze(1) + decoder_out = model.joiner.decoder_proj(decoder_out) + + current_encoder_out = torch.index_select( + current_encoder_out, + dim=0, + index=hyps_shape.row_ids(1).to(torch.int64), + ) # (num_hyps, 1, 1, encoder_out_dim) + + logits = model.joiner( + current_encoder_out, + decoder_out, + project_input=False, + ) # (num_hyps, 1, 1, vocab_size) + + logits = logits.squeeze(1).squeeze(1) # (num_hyps, vocab_size) + + log_probs = logits.log_softmax(dim=-1) # (num_hyps, vocab_size) + + log_probs.add_(ys_log_probs) + + vocab_size = log_probs.size(-1) + + log_probs = log_probs.reshape(-1) + + row_splits = hyps_shape.row_splits(1) * vocab_size + log_probs_shape = k2.ragged.create_ragged_shape2( + row_splits=row_splits, cached_tot_size=log_probs.numel() + ) + ragged_log_probs = k2.RaggedTensor(shape=log_probs_shape, value=log_probs) + """ + for all hyps with a non-blank new token, score this token. + It is a little confusing here because this for-loop + looks very similar to the one below. Here, we go through all + top-k tokens and only add the non-blanks ones to the token_list. + The RNNLM will score those tokens given the LM states. Note that + the variable `scores` is the LM score after seeing the new + non-blank token. + """ + token_list = [] + hs = [] + cs = [] + for i in range(batch_size): + topk_log_probs, topk_indexes = ragged_log_probs[i].topk(beam) + + with warnings.catch_warnings(): + warnings.simplefilter("ignore") + topk_hyp_indexes = (topk_indexes // vocab_size).tolist() + topk_token_indexes = (topk_indexes % vocab_size).tolist() + for k in range(len(topk_hyp_indexes)): + hyp_idx = topk_hyp_indexes[k] + hyp = A[i][hyp_idx] + + new_token = topk_token_indexes[k] + if new_token not in (blank_id, unk_id): + assert new_token != 0, new_token + token_list.append([new_token]) + # store the LSTM states + hs.append(hyp.state[0]) + cs.append(hyp.state[1]) + + # forward RNNLM to get new states and scores + if len(token_list) != 0: + tokens_to_score = ( + torch.tensor(token_list).to(torch.int64).to(device).reshape(-1, 1) + ) + + hs = torch.cat(hs, dim=1).to(device) + cs = torch.cat(cs, dim=1).to(device) + scores, lm_states = rnnlm.score_token(tokens_to_score, (hs, cs)) + + count = 0 # index, used to locate score and lm states + for i in range(batch_size): + topk_log_probs, topk_indexes = ragged_log_probs[i].topk(beam) + + with warnings.catch_warnings(): + warnings.simplefilter("ignore") + topk_hyp_indexes = (topk_indexes // vocab_size).tolist() + topk_token_indexes = (topk_indexes % vocab_size).tolist() + + for k in range(len(topk_hyp_indexes)): + hyp_idx = topk_hyp_indexes[k] + hyp = A[i][hyp_idx] + + ys = hyp.ys[:] + + lm_score = hyp.lm_score + state = hyp.state + + hyp_log_prob = topk_log_probs[k] # get score of current hyp + new_token = topk_token_indexes[k] + new_timestamp = hyp.timestamp[:] + if new_token not in (blank_id, unk_id): + + ys.append(new_token) + new_timestamp.append(t) + hyp_log_prob += lm_score[new_token] * lm_scale # add the lm score + + lm_score = scores[count] + state = ( + lm_states[0][:, count, :].unsqueeze(1), + lm_states[1][:, count, :].unsqueeze(1), + ) + count += 1 + + new_hyp = Hypothesis( + ys=ys, + log_prob=hyp_log_prob, + state=state, + lm_score=lm_score, + timestamp=new_timestamp, + ) + B[i].add(new_hyp) + + B = B + finalized_B + best_hyps = [b.get_most_probable(length_norm=True) for b in B] + + sorted_ans = [h.ys[context_size:] for h in best_hyps] + sorted_timestamps = [h.timestamp for h in best_hyps] + ans = [] + ans_timestamps = [] + unsorted_indices = packed_encoder_out.unsorted_indices.tolist() + for i in range(N): + ans.append(sorted_ans[unsorted_indices[i]]) + ans_timestamps.append(sorted_timestamps[unsorted_indices[i]]) + + if not return_timestamps: + return ans + else: + return DecodingResults( + tokens=ans, + timestamps=ans_timestamps, + ) + + +def modified_beam_search_rnnlm_LODR( + model: Transducer, + encoder_out: torch.Tensor, + encoder_out_lens: torch.Tensor, + sp: spm.SentencePieceProcessor, + LODR_lm: NgramLm, + LODR_lm_scale: float, + rnnlm: RnnLmModel, + rnnlm_scale: float, + beam: int = 4, +) -> List[List[int]]: + """This function implements LODR (https://arxiv.org/abs/2203.16776) with + `modified_beam_search`. It uses a bi-gram language model as the estimate + of the internal language model and subtracts its score during shallow fusion + with an external language model. This implementation uses a RNNLM as the + external language model. + + Args: + model (Transducer): + The transducer model + encoder_out (torch.Tensor): + Encoder output in (N,T,C) + encoder_out_lens (torch.Tensor): + A 1-D tensor of shape (N,), containing the number of + valid frames in encoder_out before padding. + sp: + Sentence piece generator. + LODR_lm: + A low order n-gram LM + LODR_lm_scale: + The scale of the LODR_lm + rnnlm (RnnLmModel): + RNNLM, the external language model + rnnlm_scale (float): + scale of RNNLM in shallow fusion + beam (int, optional): + Beam size. Defaults to 4. + + Returns: + Return a list-of-list of token IDs. ans[i] is the decoding results + for the i-th utterance. + + """ + assert encoder_out.ndim == 3, encoder_out.shape + assert encoder_out.size(0) >= 1, encoder_out.size(0) + assert rnnlm is not None + lm_scale = rnnlm_scale + vocab_size = rnnlm.vocab_size + + packed_encoder_out = torch.nn.utils.rnn.pack_padded_sequence( + input=encoder_out, + lengths=encoder_out_lens.cpu(), + batch_first=True, + enforce_sorted=False, + ) + + blank_id = model.decoder.blank_id + sos_id = sp.piece_to_id("") + unk_id = getattr(model, "unk_id", blank_id) + context_size = model.decoder.context_size + device = next(model.parameters()).device + + batch_size_list = packed_encoder_out.batch_sizes.tolist() + N = encoder_out.size(0) + assert torch.all(encoder_out_lens > 0), encoder_out_lens + assert N == batch_size_list[0], (N, batch_size_list) + + # get initial lm score and lm state by scoring the "sos" token + sos_token = torch.tensor([[sos_id]]).to(torch.int64).to(device) + init_score, init_states = rnnlm.score_token(sos_token) + + B = [HypothesisList() for _ in range(N)] + for i in range(N): + B[i].add( + Hypothesis( + ys=[blank_id] * context_size, + log_prob=torch.zeros(1, dtype=torch.float32, device=device), + state=init_states, # state of the RNNLM + lm_score=init_score.reshape(-1), + state_cost=NgramLmStateCost( + LODR_lm + ), # state of the source domain ngram + ) + ) + + rnnlm.clean_cache() + encoder_out = model.joiner.encoder_proj(packed_encoder_out.data) + + offset = 0 + finalized_B = [] + for batch_size in batch_size_list: + start = offset + end = offset + batch_size + current_encoder_out = encoder_out.data[start:end] # get batch + current_encoder_out = current_encoder_out.unsqueeze(1).unsqueeze(1) + # current_encoder_out's shape is (batch_size, 1, 1, encoder_out_dim) + offset = end + + finalized_B = B[batch_size:] + finalized_B + B = B[:batch_size] + + hyps_shape = get_hyps_shape(B).to(device) + + A = [list(b) for b in B] + B = [HypothesisList() for _ in range(batch_size)] + + ys_log_probs = torch.cat( + [hyp.log_prob.reshape(1, 1) for hyps in A for hyp in hyps] + ) + + decoder_input = torch.tensor( + [hyp.ys[-context_size:] for hyps in A for hyp in hyps], + device=device, + dtype=torch.int64, + ) # (num_hyps, context_size) + + decoder_out = model.decoder(decoder_input, need_pad=False).unsqueeze(1) + decoder_out = model.joiner.decoder_proj(decoder_out) + + current_encoder_out = torch.index_select( + current_encoder_out, + dim=0, + index=hyps_shape.row_ids(1).to(torch.int64), + ) # (num_hyps, 1, 1, encoder_out_dim) + + logits = model.joiner( + current_encoder_out, + decoder_out, + project_input=False, + ) # (num_hyps, 1, 1, vocab_size) + + logits = logits.squeeze(1).squeeze(1) # (num_hyps, vocab_size) + + log_probs = logits.log_softmax(dim=-1) # (num_hyps, vocab_size) + + log_probs.add_(ys_log_probs) + + vocab_size = log_probs.size(-1) + + log_probs = log_probs.reshape(-1) + + row_splits = hyps_shape.row_splits(1) * vocab_size + log_probs_shape = k2.ragged.create_ragged_shape2( + row_splits=row_splits, cached_tot_size=log_probs.numel() + ) + ragged_log_probs = k2.RaggedTensor(shape=log_probs_shape, value=log_probs) + """ + for all hyps with a non-blank new token, score this token. + It is a little confusing here because this for-loop + looks very similar to the one below. Here, we go through all + top-k tokens and only add the non-blanks ones to the token_list. + The RNNLM will score those tokens given the LM states. Note that + the variable `scores` is the LM score after seeing the new + non-blank token. + """ + token_list = [] + hs = [] + cs = [] + for i in range(batch_size): + topk_log_probs, topk_indexes = ragged_log_probs[i].topk(beam) + + with warnings.catch_warnings(): + warnings.simplefilter("ignore") + topk_hyp_indexes = (topk_indexes // vocab_size).tolist() + topk_token_indexes = (topk_indexes % vocab_size).tolist() + for k in range(len(topk_hyp_indexes)): + hyp_idx = topk_hyp_indexes[k] + hyp = A[i][hyp_idx] + + new_token = topk_token_indexes[k] + if new_token not in (blank_id, unk_id): + assert new_token != 0, new_token + token_list.append([new_token]) + # store the LSTM states + hs.append(hyp.state[0]) + cs.append(hyp.state[1]) + + # forward RNNLM to get new states and scores + if len(token_list) != 0: + tokens_to_score = ( + torch.tensor(token_list).to(torch.int64).to(device).reshape(-1, 1) + ) + + hs = torch.cat(hs, dim=1).to(device) + cs = torch.cat(cs, dim=1).to(device) + scores, lm_states = rnnlm.score_token(tokens_to_score, (hs, cs)) + + count = 0 # index, used to locate score and lm states + for i in range(batch_size): + topk_log_probs, topk_indexes = ragged_log_probs[i].topk(beam) + + with warnings.catch_warnings(): + warnings.simplefilter("ignore") + topk_hyp_indexes = (topk_indexes // vocab_size).tolist() + topk_token_indexes = (topk_indexes % vocab_size).tolist() + + for k in range(len(topk_hyp_indexes)): + hyp_idx = topk_hyp_indexes[k] + hyp = A[i][hyp_idx] + + ys = hyp.ys[:] + + # current score of hyp + lm_score = hyp.lm_score + state = hyp.state + + hyp_log_prob = topk_log_probs[k] # get score of current hyp + new_token = topk_token_indexes[k] + if new_token not in (blank_id, unk_id): + + ys.append(new_token) + state_cost = hyp.state_cost.forward_one_step(new_token) + + # calculate the score of the latest token + current_ngram_score = state_cost.lm_score - hyp.state_cost.lm_score + + assert current_ngram_score <= 0.0, ( + state_cost.lm_score, + hyp.state_cost.lm_score, + ) + # score = score + RNNLM_score - LODR_score + # LODR_LM_scale is a negative number here + hyp_log_prob += ( + lm_score[new_token] * lm_scale + + LODR_lm_scale * current_ngram_score + ) # add the lm score + + lm_score = scores[count] + state = ( + lm_states[0][:, count, :].unsqueeze(1), + lm_states[1][:, count, :].unsqueeze(1), + ) + count += 1 + else: + state_cost = hyp.state_cost + + new_hyp = Hypothesis( + ys=ys, + log_prob=hyp_log_prob, + state=state, + lm_score=lm_score, + state_cost=state_cost, + ) + B[i].add(new_hyp) + + B = B + finalized_B + best_hyps = [b.get_most_probable(length_norm=True) for b in B] + + sorted_ans = [h.ys[context_size:] for h in best_hyps] + ans = [] + unsorted_indices = packed_encoder_out.unsorted_indices.tolist() + for i in range(N): + ans.append(sorted_ans[unsorted_indices[i]]) + + return ans diff --git a/egs/tedlium3/ASR/pruned_transducer_stateless_d2v_v2/bias_compare.py b/egs/tedlium3/ASR/pruned_transducer_stateless_d2v_v2/bias_compare.py new file mode 100644 index 000000000..1c18fec88 --- /dev/null +++ b/egs/tedlium3/ASR/pruned_transducer_stateless_d2v_v2/bias_compare.py @@ -0,0 +1,11 @@ +import torch + +base_model = torch.load('./d2v-base-T.pt') +bias_model = torch.load('./bitfit_533_v2/checkpoint-100.pt') + +base_model, bias_model = base_model['model'], bias_model['model'] + +for key in base_model.keys(): + if 'bias' in key: + l1_diff = torch.abs(base_model[key]-bias_model[key]).sum() / base_model[key].size(0) + print(key, l1_diff.item()) diff --git a/egs/tedlium3/ASR/pruned_transducer_stateless_d2v_v2/bitfit.py b/egs/tedlium3/ASR/pruned_transducer_stateless_d2v_v2/bitfit.py new file mode 100755 index 000000000..631f69a58 --- /dev/null +++ b/egs/tedlium3/ASR/pruned_transducer_stateless_d2v_v2/bitfit.py @@ -0,0 +1,1827 @@ +#!/usr/bin/env python3 +# Copyright 2021-2022 Xiaomi Corp. (authors: Fangjun Kuang, +# Wei Kang, +# Mingshuang Luo,) +# Zengwei Yao) +# +# See ../../../../LICENSE for clarification regarding multiple authors +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +""" +Usage: + +export CUDA_VISIBLE_DEVICES="0,1,2,3" + +./pruned_transducer_stateless7_ctc/train.py \ + --world-size 4 \ + --num-epochs 30 \ + --start-epoch 1 \ + --exp-dir pruned_transducer_stateless7_ctc/exp \ + --full-libri 1 \ + --max-duration 300 + +# For mix precision training: + +./pruned_transducer_stateless7_ctc/train.py \ + --world-size 4 \ + --num-epochs 30 \ + --start-epoch 1 \ + --use-fp16 1 \ + --exp-dir pruned_transducer_stateless7_ctc/exp \ + --full-libri 1 \ + --max-duration 550 + +# For d2v-T training: +export CUDA_VISIBLE_DEVICES="0,1,2,3,4,5,6,7" + +./pruned_transducer_stateless_d2v_v2/train.py \ + --wandb true \ + --input-strategy AudioSamples \ + --enable-spec-aug False \ + --multi-optim True \ + --world-size 8 \ + --num-epochs 30 \ + --start-epoch 1 \ + --full-libri 0 \ + --exp-dir ./pruned_transducer_stateless_d2v_v2/$1 \ + --max-duration 250 \ + --freeze-finetune-updates 2000 \ + --use-fp16 1 \ + --peak-enc-lr 0.001 \ + --peak-dec-lr 0.05 \ + --accum-grads 1 \ + --encoder-type d2v \ + --additional-block True \ + --encoder-dim 768 \ + --decoder-dim 768 \ + --joiner-dim 768 \ + --prune-range 20 \ + --context-size 2 \ + --ctc-loss-scale 0.2 + +""" + + +import random +import argparse +import copy +import logging +import warnings +from pathlib import Path +from shutil import copyfile +from typing import Any, Dict, Optional, Tuple, Union + +import k2 +import optim +import sentencepiece as spm +import torch +import torch.multiprocessing as mp +import torch.nn as nn +from asr_datamodule import LibriSpeechAsrDataModule +from decoder import Decoder +from joiner import Joiner +from lhotse.cut import Cut +from lhotse.dataset.sampling.base import CutSampler +from lhotse.utils import fix_random_seed +from model import Transducer +from optim import Eden, ScaledAdam +from torch import Tensor +from torch.cuda.amp import GradScaler +from torch.nn.parallel import DistributedDataParallel as DDP +from torch.utils.tensorboard import SummaryWriter +from zipformer import Zipformer +from data2vec_encoder import FairSeqData2VecEncoder + +from icefall import diagnostics +from icefall.checkpoint import remove_checkpoints +from icefall.checkpoint import update_averaged_model +from checkpoint import ( + save_checkpoint as save_checkpoint_impl, + save_checkpoint_with_global_batch_idx, + load_checkpoint +) +from icefall.dist import cleanup_dist, setup_dist +from icefall.env import get_env_info +from icefall.hooks import register_inf_check_hooks +from icefall.utils import ( + AttributeDict, + MetricsTracker, + encode_supervisions, + setup_logger, + str2bool, + save_args, +) + +import wandb + +#from icefall.checkpoint import save_checkpoint as save_checkpoint_impl +LRSchedulerType = Union[torch.optim.lr_scheduler._LRScheduler, optim.LRScheduler] + + +def set_batch_count(model: Union[nn.Module, DDP], batch_count: float) -> None: + if isinstance(model, DDP): + # get underlying nn.Module + model = model.module + for module in model.modules(): + if hasattr(module, "batch_count"): + module.batch_count = batch_count + model.encoder.num_updates = int(batch_count) + + +def add_adapter_arguments(parser: argparse.ArgumentParser): + parser.add_argument( + "--add-adapter", + type=str2bool, + default=False, + help="add adapter to rep model's encoder" + ) + + parser.add_argument( + "--adapter-lr", + type=float, + default=0.0001, + help="adapter learning rate" + ) + + parser.add_argument( + "--gender", + type=str, + default='male', + help="select gender" + ) + + +def add_rep_arguments(parser: argparse.ArgumentParser): + parser.add_argument( + "--wandb", + type=str2bool, + default=True, + help="Use wandb for MLOps", + ) + parser.add_argument( + "--hpo", + type=str2bool, + default=False, + help="Use small db for HPO", + ) + + parser.add_argument( + "--accum-grads", + type=int, + default=1, + help="accum-grad num.", + ) + + parser.add_argument( + "--multi-optim", + type=str2bool, + default=True, + help="use sperate optimizer (enc / dec)", + ) + + parser.add_argument( + "--peak-enc-lr", + type=float, + default=0.0001, + help="The initial learning rate. This value should not need to be changed.", + ) + + parser.add_argument( + "--peak-dec-lr", + type=float, + default=0.001, + help="The initial learning rate. This value should not need to be changed.", + ) + + parser.add_argument( + "--encoder-type", + type=str, + default='d2v', + help="Type of encoder (e.g. conformer, w2v, d2v...", + ) + + parser.add_argument( + "--encoder-dim", + type=int, + default=768, + help="encoder embedding dimension", + ) + + parser.add_argument( + "--freeze-finetune-updates", + type=int, + default=0 + ) + + parser.add_argument( + "--additional-block", + type=str2bool, + default=True, + ) + + parser.add_argument( + "--decode-interval", + type=int, + default=200, + help="decode interval", + ) + + +def add_model_arguments(parser: argparse.ArgumentParser): + parser.add_argument( + "--num-encoder-layers", + type=str, + default="2,4,3,2,4", + help="Number of zipformer encoder layers, comma separated.", + ) + + parser.add_argument( + "--feedforward-dims", + type=str, + default="1024,1024,2048,2048,1024", + help="Feedforward dimension of the zipformer encoder layers, comma separated.", + ) + + parser.add_argument( + "--nhead", + type=str, + default="8,8,8,8,8", + help="Number of attention heads in the zipformer encoder layers.", + ) + + parser.add_argument( + "--encoder-dims", + type=str, + default="384,384,384,384,384", + help="Embedding dimension in the 2 blocks of zipformer encoder layers, comma separated", + ) + + parser.add_argument( + "--attention-dims", + type=str, + default="192,192,192,192,192", + help="""Attention dimension in the 2 blocks of zipformer encoder layers, comma separated; + not the same as embedding dimension.""", + ) + + parser.add_argument( + "--encoder-unmasked-dims", + type=str, + default="256,256,256,256,256", + help="Unmasked dimensions in the encoders, relates to augmentation during training. " + "Must be <= each of encoder_dims. Empirically, less than 256 seems to make performance " + " worse.", + ) + + parser.add_argument( + "--zipformer-downsampling-factors", + type=str, + default="1,2,4,8,2", + help="Downsampling factor for each stack of encoder layers.", + ) + + parser.add_argument( + "--cnn-module-kernels", + type=str, + default="31,31,31,31,31", + help="Sizes of kernels in convolution modules", + ) + + parser.add_argument( + "--decoder-dim", + type=int, + default=768, + help="Embedding dimension in the decoder model.", + ) + + parser.add_argument( + "--joiner-dim", + type=int, + default=768, + help="""Dimension used in the joiner model. + Outputs from the encoder and decoder model are projected + to this dimension before adding. + """, + ) + + parser.add_argument( + "--prompt", + type=str2bool, + default=False, + ) + + +def get_parser(): + parser = argparse.ArgumentParser( + formatter_class=argparse.ArgumentDefaultsHelpFormatter + ) + + parser.add_argument( + "--world-size", + type=int, + default=1, + help="Number of GPUs for DDP training.", + ) + + parser.add_argument( + "--master-port", + type=int, + default=12354, + help="Master port to use for DDP training.", + ) + + parser.add_argument( + "--tensorboard", + type=str2bool, + default=True, + help="Should various information be logged in tensorboard.", + ) + + parser.add_argument( + "--num-epochs", + type=int, + default=30, + help="Number of epochs to train.", + ) + + parser.add_argument( + "--num-updates", + type=int, + default=5000, + help="Number of epochs to train.", + ) + + parser.add_argument( + "--start-epoch", + type=int, + default=1, + help="""Resume training from this epoch. It should be positive. + If larger than 1, it will load checkpoint from + exp-dir/epoch-{start_epoch-1}.pt + """, + ) + + parser.add_argument( + "--start-batch", + type=int, + default=0, + help="""If positive, --start-epoch is ignored and + it loads the checkpoint from exp-dir/checkpoint-{start_batch}.pt + """, + ) + + parser.add_argument( + "--exp-dir", + type=str, + default="pruned_transducer_stateless7_ctc/exp", + help="""The experiment dir. + It specifies the directory where all training related + files, e.g., checkpoints, log, etc, are saved + """, + ) + + parser.add_argument( + "--bpe-model", + type=str, + default="data/lang_bpe_500/bpe.model", + help="Path to the BPE model", + ) + + parser.add_argument( + "--base-lr", type=float, default=0.05, help="The base learning rate." + ) + + parser.add_argument( + "--lr-batches", + type=float, + default=5000, + help="""Number of steps that affects how rapidly the learning rate + decreases. We suggest not to change this.""", + ) + + parser.add_argument( + "--lr-epochs", + type=float, + default=3.5, + help="""Number of epochs that affects how rapidly the learning rate decreases. + """, + ) + + parser.add_argument( + "--context-size", + type=int, + default=2, + help="The context size in the decoder. 1 means bigram; 2 means tri-gram", + ) + + parser.add_argument( + "--prune-range", + type=int, + default=5, + help="The prune range for rnnt loss, it means how many symbols(context)" + "we are using to compute the loss", + ) + + parser.add_argument( + "--lm-scale", + type=float, + default=0.25, + help="The scale to smooth the loss with lm " + "(output of prediction network) part.", + ) + + parser.add_argument( + "--am-scale", + type=float, + default=0.0, + help="The scale to smooth the loss with am (output of encoder network) part.", + ) + + parser.add_argument( + "--simple-loss-scale", + type=float, + default=0.5, + help="To get pruning ranges, we will calculate a simple version" + "loss(joiner is just addition), this simple loss also uses for" + "training (as a regularization item). We will scale the simple loss" + "with this parameter before adding to the final loss.", + ) + + parser.add_argument( + "--ctc-loss-scale", + type=float, + default=0.2, + help="Scale for CTC loss.", + ) + + parser.add_argument( + "--seed", + type=int, + default=42, + help="The seed for random generators intended for reproducibility", + ) + + parser.add_argument( + "--print-diagnostics", + type=str2bool, + default=False, + help="Accumulate stats on activations, print them and exit.", + ) + + parser.add_argument( + "--inf-check", + type=str2bool, + default=False, + help="Add hooks to check for infinite module outputs and gradients.", + ) + + parser.add_argument( + "--save-every-n", + type=int, + default=200, + help="""Save checkpoint after processing this number of batches" + periodically. We save checkpoint to exp-dir/ whenever + params.batch_idx_train % save_every_n == 0. The checkpoint filename + has the form: f'exp-dir/checkpoint-{params.batch_idx_train}.pt' + Note: It also saves checkpoint to `exp-dir/epoch-xxx.pt` at the + end of each epoch where `xxx` is the epoch number counting from 0. + """, + ) + + parser.add_argument( + "--keep-last-k", + type=int, + default=30, + help="""Only keep this number of checkpoints on disk. + For instance, if it is 3, there are only 3 checkpoints + in the exp-dir with filenames `checkpoint-xxx.pt`. + It does not affect checkpoints with name `epoch-xxx.pt`. + """, + ) + + parser.add_argument( + "--average-period", + type=int, + default=10, + help="""Update the averaged model, namely `model_avg`, after processing + this number of batches. `model_avg` is a separate version of model, + in which each floating-point parameter is the average of all the + parameters from the start of training. Each time we take the average, + we do: `model_avg = model * (average_period / batch_idx_train) + + model_avg * ((batch_idx_train - average_period) / batch_idx_train)`. + """, + ) + + parser.add_argument( + "--use-fp16", + type=str2bool, + default=True, + help="Whether to use half precision training.", + ) + + add_model_arguments(parser) + add_rep_arguments(parser) + add_adapter_arguments(parser) + + return parser + + +def get_params() -> AttributeDict: + """Return a dict containing training parameters. + + All training related parameters that are not passed from the commandline + are saved in the variable `params`. + + Commandline options are merged into `params` after they are parsed, so + you can also access them via `params`. + + Explanation of options saved in `params`: + + - best_train_loss: Best training loss so far. It is used to select + the model that has the lowest training loss. It is + updated during the training. + + - best_valid_loss: Best validation loss so far. It is used to select + the model that has the lowest validation loss. It is + updated during the training. + + - best_train_epoch: It is the epoch that has the best training loss. + + - best_valid_epoch: It is the epoch that has the best validation loss. + + - batch_idx_train: Used to writing statistics to tensorboard. It + contains number of batches trained so far across + epochs. + + - log_interval: Print training loss if batch_idx % log_interval` is 0 + + - reset_interval: Reset statistics if batch_idx % reset_interval is 0 + + - valid_interval: Run validation if batch_idx % valid_interval is 0 + + - feature_dim: The model input dim. It has to match the one used + in computing features. + + - subsampling_factor: The subsampling factor for the model. + + - encoder_dim: Hidden dim for multi-head attention model. + + - num_decoder_layers: Number of decoder layer of transformer decoder. + + - warm_step: The warmup period that dictates the decay of the + scale on "simple" (un-pruned) loss. + """ + params = AttributeDict( + { + "best_train_loss": float("inf"), + "best_valid_loss": float("inf"), + "best_train_epoch": -1, + "best_valid_epoch": -1, + "batch_idx_train": 0, + "log_interval": 5, + "reset_interval": 200, + "valid_interval": 3000, # For the 100h subset, use 800 + # parameters for zipformer + "feature_dim": 80, + "subsampling_factor": 320, # not passed in, this is fixed. + # parameters for ctc loss + "beam_size": 10, + "use_double_scores": True, + "warm_step": 0, + #"warm_step": 4000, + #"warm_step": 3000, + "env_info": get_env_info(), + } + ) + + return params + + +def get_encoder_model(params: AttributeDict) -> nn.Module: + # TODO: We can add an option to switch between Zipformer and Transformer + def to_int_tuple(s: str): + return tuple(map(int, s.split(","))) + + if params.encoder_type == 'd2v': + encoder = FairSeqData2VecEncoder( + input_size=params.encoder_dim, + w2v_url='None', + output_size=params.encoder_dim, + freeze_finetune_updates=params.freeze_finetune_updates, + additional_block=params.additional_block, + ) + else: + encoder = Zipformer( + num_features=params.feature_dim, + output_downsampling_factor=2, + zipformer_downsampling_factors=to_int_tuple( + params.zipformer_downsampling_factors + ), + encoder_dims=to_int_tuple(params.encoder_dims), + attention_dim=to_int_tuple(params.attention_dims), + encoder_unmasked_dims=to_int_tuple(params.encoder_unmasked_dims), + nhead=to_int_tuple(params.nhead), + feedforward_dim=to_int_tuple(params.feedforward_dims), + cnn_module_kernels=to_int_tuple(params.cnn_module_kernels), + num_encoder_layers=to_int_tuple(params.num_encoder_layers), + ) + + return encoder + + +def get_decoder_model(params: AttributeDict) -> nn.Module: + decoder = Decoder( + vocab_size=params.vocab_size, + decoder_dim=params.decoder_dim, + blank_id=params.blank_id, + context_size=params.context_size, + ) + return decoder + + +def get_joiner_model(params: AttributeDict) -> nn.Module: + joiner = Joiner( + encoder_dim=params.encoder_dim if params.encoder_type == 'd2v' else int(params.encoder_dims.split(",")[-1]), + decoder_dim=params.decoder_dim, + joiner_dim=params.joiner_dim, + vocab_size=params.vocab_size, + ) + return joiner + + +def get_transducer_model(params: AttributeDict) -> nn.Module: + encoder = get_encoder_model(params) + decoder = get_decoder_model(params) + joiner = get_joiner_model(params) + + model = Transducer( + encoder=encoder, + decoder=decoder, + joiner=joiner, + encoder_dim=params.encoder_dim if params.encoder_type == 'd2v' else int(params.encoder_dims.split(",")[-1]), + decoder_dim=params.decoder_dim, + joiner_dim=params.joiner_dim, + vocab_size=params.vocab_size, + prompt=params.prompt, + sid=params.spk_id, + ) + return model + + +def load_checkpoint_if_available( + params: AttributeDict, + model: nn.Module, + model_avg: nn.Module = None, + optimizer: Optional[torch.optim.Optimizer] = None, + scheduler: Optional[LRSchedulerType] = None, +) -> Optional[Dict[str, Any]]: + """Load checkpoint from file. + + If params.start_batch is positive, it will load the checkpoint from + `params.exp_dir/checkpoint-{params.start_batch}.pt`. Otherwise, if + params.start_epoch is larger than 1, it will load the checkpoint from + `params.start_epoch - 1`. + + Apart from loading state dict for `model` and `optimizer` it also updates + `best_train_epoch`, `best_train_loss`, `best_valid_epoch`, + and `best_valid_loss` in `params`. + + Args: + params: + The return value of :func:`get_params`. + model: + The training model. + model_avg: + The stored model averaged from the start of training. + optimizer: + The optimizer that we are using. + scheduler: + The scheduler that we are using. + Returns: + Return a dict containing previously saved training info. + """ + if params.start_batch > 0: + filename = params.exp_dir / f"checkpoint-{params.start_batch}.pt" + elif params.start_epoch > 1: + filename = params.exp_dir / f"epoch-{params.start_epoch-1}.pt" + elif params.add_adapter: + filename = params.exp_dir / f"../d2v-base-T.pt" + else: + return None + + assert filename.is_file(), f"{filename} does not exist!" + + saved_params = load_checkpoint( + filename, + model=model, + model_avg=model_avg, + optimizer=optimizer, + scheduler=scheduler, + strict=True if not params.add_adapter else False, + ) + + keys = [ + "best_train_epoch", + "best_valid_epoch", + "batch_idx_train", + "best_train_loss", + "best_valid_loss", + ] + for k in keys: + params[k] = saved_params[k] + + if params.start_batch > 0: + if "cur_epoch" in saved_params: + params["start_epoch"] = saved_params["cur_epoch"] + + if "cur_batch_idx" in saved_params: + params["cur_batch_idx"] = saved_params["cur_batch_idx"] + + params.batch_idx_train = 0 + + return saved_params + + +def save_checkpoint( + params: AttributeDict, + model: Union[nn.Module, DDP], + model_avg: Optional[nn.Module] = None, + optimizer: Optional[torch.optim.Optimizer] = None, + scheduler: Optional[LRSchedulerType] = None, + sampler: Optional[CutSampler] = None, + scaler: Optional[GradScaler] = None, + rank: int = 0, +) -> None: + """Save model, optimizer, scheduler and training stats to file. + + Args: + params: + It is returned by :func:`get_params`. + model: + The training model. + model_avg: + The stored model averaged from the start of training. + optimizer: + The optimizer used in the training. + sampler: + The sampler for the training dataset. + scaler: + The scaler used for mix precision training. + """ + if rank != 0: + return + filename = params.exp_dir / f"epoch-{params.cur_epoch}.pt" + save_checkpoint_impl( + filename=filename, + model=model, + model_avg=model_avg, + params=params, + optimizer=optimizer, + scheduler=scheduler, + sampler=sampler, + scaler=scaler, + rank=rank, + ) + + if params.best_train_epoch == params.cur_epoch: + best_train_filename = params.exp_dir / "best-train-loss.pt" + copyfile(src=filename, dst=best_train_filename) + + if params.best_valid_epoch == params.cur_epoch: + best_valid_filename = params.exp_dir / "best-valid-loss.pt" + copyfile(src=filename, dst=best_valid_filename) + + +def compute_loss( + params: AttributeDict, + model: Union[nn.Module, DDP], + sp: spm.SentencePieceProcessor, + batch: dict, + is_training: bool, + decode: bool = False, +) -> Tuple[Tensor, MetricsTracker]: + """ + Compute transducer loss given the model and its inputs. + + Args: + params: + Parameters for training. See :func:`get_params`. + model: + The model for training. It is an instance of Zipformer in our case. + batch: + A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()` + for the content in it. + is_training: + True for training. False for validation. When it is True, this + function enables autograd during computation; when it is False, it + disables autograd. + warmup: a floating point value which increases throughout training; + values >= 1.0 are fully warmed up and have all modules present. + """ + device = model.device if isinstance(model, DDP) else next(model.parameters()).device + feature = batch["inputs"] + # at entry, feature is (N, T, C) + assert feature.ndim == 2 or feature.ndim == 3 + feature = feature.to(device) + + supervisions = batch["supervisions"] + + if feature.ndim == 2: + feature_lens = [] + for supervision in supervisions['cut']: + try: feature_lens.append(supervision.tracks[0].cut.recording.num_samples) + except: feature_lens.append(supervision.recording.num_samples) + feature_lens = torch.tensor(feature_lens) + + elif feature.ndim == 3: + feature_lens = supervisions["num_frames"].to(device) + + batch_idx_train = params.batch_idx_train + warm_step = params.warm_step + + texts = batch["supervisions"]["text"] + + token_ids = sp.encode(texts, out_type=int) + y = k2.RaggedTensor(token_ids).to(device) + + with torch.set_grad_enabled(is_training): + simple_loss, pruned_loss, ctc_output = model( + x=feature, + x_lens=feature_lens, + y=y, + prune_range=params.prune_range, + am_scale=params.am_scale, + lm_scale=params.lm_scale, + ) + + s = params.simple_loss_scale + # take down the scale on the simple loss from 1.0 at the start + # to params.simple_loss scale by warm_step. + simple_loss_scale = ( + s + if batch_idx_train >= warm_step + else 1.0 - (batch_idx_train / warm_step) * (1.0 - s) + ) + pruned_loss_scale = ( + 1.0 + if batch_idx_train >= warm_step + else 0.1 + 0.9 * (batch_idx_train / warm_step) + ) + + loss = simple_loss_scale * simple_loss + pruned_loss_scale * pruned_loss + + info = MetricsTracker() + + if params.ctc_loss_scale > 0: + # Compute ctc loss + + # NOTE: We need `encode_supervisions` to sort sequences with + # different duration in decreasing order, required by + # `k2.intersect_dense` called in `k2.ctc_loss` + with warnings.catch_warnings(): + warnings.simplefilter("ignore") + supervision_segments, token_ids = encode_supervisions( + supervisions, + subsampling_factor=params.subsampling_factor, + token_ids=token_ids, + ) + + # Works with a BPE model + decoding_graph = k2.ctc_graph(token_ids, modified=False, device=device) + dense_fsa_vec = k2.DenseFsaVec( + ctc_output, + supervision_segments, + allow_truncate=params.subsampling_factor - 1, + ) + + ctc_loss = k2.ctc_loss( + decoding_graph=decoding_graph, + dense_fsa_vec=dense_fsa_vec, + output_beam=params.beam_size, + reduction="sum", + use_double_scores=params.use_double_scores, + ) + assert ctc_loss.requires_grad == is_training + loss += params.ctc_loss_scale * ctc_loss + + info["ctc_loss"] = ctc_loss.detach().cpu().item() + + assert loss.requires_grad == is_training + + if decode: + model.eval() + with torch.no_grad(): + try: + hypos = model.module.decode( + x=feature, + x_lens=feature_lens, + y=y, + sp=sp + ) + except: + hypos = model.decode( + x=feature, + x_lens=feature_lens, + y=y, + sp=sp + ) + + logging.info(f'ref: {batch["supervisions"]["text"][0]}') + logging.info(f'hyp: {" ".join(hypos[0])}') + model.train() + + with warnings.catch_warnings(): + warnings.simplefilter("ignore") + info["frames"] = (feature_lens // params.subsampling_factor).sum().item() + + # Note: We use reduction=sum while computing the loss. + info["utterances"] = feature.size(0) + info["loss"] = loss.detach().cpu().item() + info["simple_loss"] = simple_loss.detach().cpu().item() + info["pruned_loss"] = pruned_loss.detach().cpu().item() + + return loss, info + + +def compute_validation_loss( + params: AttributeDict, + model: Union[nn.Module, DDP], + sp: spm.SentencePieceProcessor, + valid_dl: torch.utils.data.DataLoader, + world_size: int = 1, +) -> MetricsTracker: + """Run the validation process.""" + model.eval() + + tot_loss = MetricsTracker() + + for batch_idx, batch in enumerate(valid_dl): + loss, loss_info = compute_loss( + params=params, + model=model, + sp=sp, + batch=batch, + is_training=False, + ) + assert loss.requires_grad is False + tot_loss = tot_loss + loss_info + + if world_size > 1: + tot_loss.reduce(loss.device) + + loss_value = tot_loss["loss"] / tot_loss["utterances"] + if loss_value < params.best_valid_loss: + params.best_valid_epoch = params.cur_epoch + params.best_valid_loss = loss_value + + return tot_loss + + +def train_one_epoch( + params: AttributeDict, + model: Union[nn.Module, DDP], + optimizer: torch.optim.Optimizer or [torch.optim.Optimizer, torch.optim.Optimizer], + scheduler: LRSchedulerType or [LRSchedulerType, LRSchedulerType], + sp: spm.SentencePieceProcessor, + train_dl: torch.utils.data.DataLoader, + valid_dl: torch.utils.data.DataLoader, + scaler: GradScaler, + model_avg: Optional[nn.Module] = None, + tb_writer: Optional[SummaryWriter] = None, + world_size: int = 1, + rank: int = 0, + wb = None, +) -> None: + """Train the model for one epoch. + + The training loss from the mean of all frames is saved in + `params.train_loss`. It runs the validation process every + `params.valid_interval` batches. + + Args: + params: + It is returned by :func:`get_params`. + model: + The model for training. + optimizer: + The optimizer we are using. + scheduler: + The learning rate scheduler, we call step() every step. + train_dl: + Dataloader for the training dataset. + valid_dl: + Dataloader for the validation dataset. + scaler: + The scaler used for mix precision training. + model_avg: + The stored model averaged from the start of training. + tb_writer: + Writer to write log messages to tensorboard. + world_size: + Number of nodes in DDP training. If it is 1, DDP is disabled. + rank: + The rank of the node in DDP training. If no DDP is used, it should + be set to 0. + """ + model.train() + + tot_loss = MetricsTracker() + + cur_batch_idx = params.get("cur_batch_idx", 0) + + if params.multi_optim: + optimizer_enc, optimizer_dec = optimizer[0], optimizer[1] + scheduler_enc, scheduler_dec = scheduler[0], scheduler[1] + + for batch_idx, batch in enumerate(train_dl): + if params.batch_idx_train > params.num_updates: + break + if batch_idx < cur_batch_idx: + continue + cur_batch_idx = batch_idx + + if batch_idx % params.accum_grads == 0: params.batch_idx_train += 1 + batch_size = len(batch["supervisions"]["text"]) + + try: + with torch.cuda.amp.autocast(enabled=params.use_fp16): + loss, loss_info = compute_loss( + params=params, + model=model, + sp=sp, + batch=batch, + is_training=True, + decode = True if batch_idx % params.decode_interval == 0 else False, + ) + + try: loss_info.reduce(loss.device) + except: pass + + numel = params.world_size / (params.accum_grads * loss_info["utterances"]) + loss *= numel ## normalize loss over utts(batch size) + + # summary stats + tot_loss = (tot_loss * (1 - 1 / params.reset_interval)) + loss_info + + # NOTE: We use reduction==sum and loss is computed over utterances + # in the batch and there is no normalization to it so far. + scaler.scale(loss).backward() + + if params.multi_optim and (batch_idx+1) % params.accum_grads == 0: + set_batch_count(model, params.batch_idx_train) + scheduler_enc.step_batch(params.batch_idx_train) + scheduler_dec.step_batch(params.batch_idx_train) + scaler.step(optimizer_enc) + scaler.step(optimizer_dec) + scaler.update() + optimizer_enc.zero_grad() + optimizer_dec.zero_grad() + elif not params.multi_optim and (batch_idx+1) % params.accum_grads == 0: + set_batch_count(model, params.batch_idx_train) + scheduler.step_batch(params.batch_idx_train) + scaler.step(optimizer) + scaler.update() + optimizer.zero_grad() + + except: # noqa + display_and_save_batch(batch, params=params, sp=sp) + raise + + if params.print_diagnostics and batch_idx == 5: + return + + ''' + if ( + rank == 0 + and params.batch_idx_train > 0 + and params.batch_idx_train % params.average_period == 0 + ): + update_averaged_model( + params=params, + model_cur=model, + model_avg=model_avg, + ) + ''' + + if ( + params.batch_idx_train > 0 + and params.batch_idx_train % params.save_every_n == 0 + ): + params.cur_batch_idx = batch_idx + save_checkpoint_with_global_batch_idx( + out_dir=params.exp_dir, + global_batch_idx=params.batch_idx_train, + model=model, + model_avg=model_avg, + params=params, + optimizer=optimizer, + scheduler=scheduler, + sampler=train_dl.sampler, + scaler=scaler, + rank=rank, + ) + del params.cur_batch_idx + ''' + remove_checkpoints( + out_dir=params.exp_dir, + topk=params.keep_last_k, + rank=rank, + ) + ''' + + if batch_idx % 100 == 0 and params.use_fp16: + # If the grad scale was less than 1, try increasing it. The _growth_interval + # of the grad scaler is configurable, but we can't configure it to have different + # behavior depending on the current grad scale. + cur_grad_scale = scaler._scale.item() + ''' + if cur_grad_scale < 1.0 or (cur_grad_scale < 8.0 and batch_idx % 400 == 0): + scaler.update(cur_grad_scale * 2.0) + ''' + if cur_grad_scale < 0.01: + logging.warning(f"Grad scale is small: {cur_grad_scale}") + if cur_grad_scale < 1.0e-05: + wb.log({"valid/loss": 10000}) + raise RuntimeError( + f"grad_scale is too small, exiting: {cur_grad_scale}" + ) + + #if params.batch_idx_train > 4000 and loss > 300 and params.wandb: + # wb.log({"valid/loss": 10000}) + # raise RuntimeError( + # f"divergence... exiting: loss={loss}" + # ) + + if batch_idx % (params.log_interval*params.accum_grads) == 0: + #for n, p in model.named_parameters(): + # if 'adapter' in n: + # print(p) + if params.multi_optim: + cur_enc_lr = scheduler_enc.get_last_lr()[0] + cur_dec_lr = scheduler_dec.get_last_lr()[0] + cur_grad_scale = scaler._scale.item() if params.use_fp16 else 1.0 + + logging.info( + f"Epoch {params.cur_epoch}, " + f"batch {batch_idx}, loss[{loss_info}], " + f"tot_loss[{tot_loss}], batch size: {batch_size}, " + f"enc_lr: {cur_enc_lr:.2e}, " + f"dec_lr: {cur_dec_lr:.2e}, " + + (f"grad_scale: {scaler._scale.item()}" if params.use_fp16 else "") + ) + + else: + cur_lr = scheduler.get_last_lr()[0] + cur_grad_scale = scaler._scale.item() if params.use_fp16 else 1.0 + + logging.info( + f"Epoch {params.cur_epoch}, " + f"batch {batch_idx}, loss[{loss_info}], " + f"tot_loss[{tot_loss}], batch size: {batch_size}, " + f"lr: {cur_lr:.2e}, " + + (f"grad_scale: {scaler._scale.item()}" if params.use_fp16 else "") + ) + + if tb_writer is not None: + if params.multi_optim: + tb_writer.add_scalar( + "train/enc_learning_rate", cur_enc_lr, params.batch_idx_train + ) + tb_writer.add_scalar( + "train/dec_learning_rate", cur_dec_lr, params.batch_idx_train + ) + + else: + tb_writer.add_scalar( + "train/learning_rate", cur_lr, params.batch_idx_train + ) + + loss_info.write_summary( + tb_writer, "train/current_", params.batch_idx_train + ) + tot_loss.write_summary(tb_writer, "train/tot_", params.batch_idx_train) + if params.use_fp16: + tb_writer.add_scalar( + "train/grad_scale", + cur_grad_scale, + params.batch_idx_train, + ) + + if wb is not None and rank == 0: + wb.log({"train/loss": loss_info["loss"]*numel}) + wb.log({"train/simple_loss": loss_info["simple_loss"]*numel}) + wb.log({"train/pruned_loss": loss_info["pruned_loss"]*numel}) + wb.log({"train/ctc_loss": loss_info["ctc_loss"]*numel}) + + ''' + logging.info("Computing validation loss") + valid_info = compute_validation_loss( + params=params, + model=model, + sp=sp, + valid_dl=valid_dl, + world_size=world_size, + ) + model.train() + logging.info(f"Epoch {params.cur_epoch}, validation: {valid_info}") + logging.info( + f"Maximum memory allocated so far is {torch.cuda.max_memory_allocated()//1000000}MB" + ) + if tb_writer is not None: + valid_info.write_summary( + tb_writer, "train/valid_", params.batch_idx_train + ) + + if wb is not None and rank == 0: + numel = 1 / (params.accum_grads * valid_info["utterances"]) + #wb.log({"valid/loss": valid_info["loss"]*numel}) + wb.log({"valid/loss": numel*(valid_info["simple_loss"] + +valid_info["pruned_loss"] + +valid_info["ctc_loss"] + )}) + wb.log({"valid/simple_loss": valid_info["simple_loss"]*numel}) + wb.log({"valid/pruned_loss": valid_info["pruned_loss"]*numel}) + wb.log({"valid/ctc_loss": valid_info["ctc_loss"]*numel}) + ''' + loss_value = tot_loss["loss"] / tot_loss["utterances"] + params.train_loss = loss_value + if params.train_loss < params.best_train_loss: + params.best_train_epoch = params.cur_epoch + params.best_train_loss = params.train_loss + + +def run(rank, world_size, args, wb=None): + """ + Args: + rank: + It is a value between 0 and `world_size-1`, which is + passed automatically by `mp.spawn()` in :func:`main`. + The node with rank 0 is responsible for saving checkpoint. + world_size: + Number of GPUs for DDP training. + args: + The return value of get_parser().parse_args() + """ + params = get_params() + params.update(vars(args)) + #params.warm_step *= params.accum_grads + + fix_random_seed(params.seed) + if world_size > 1: + setup_dist(rank, world_size, params.master_port) + + setup_logger(f"{params.exp_dir}/log/log-train") + logging.info("Training started") + + if args.tensorboard and rank == 0: + tb_writer = SummaryWriter(log_dir=f"{params.exp_dir}/tensorboard") + else: + tb_writer = None + + device = torch.device("cpu") + if torch.cuda.is_available(): + device = torch.device("cuda", rank) + logging.info(f"Device: {device}") + + sp = spm.SentencePieceProcessor() + sp.load(params.bpe_model) + + # is defined in local/train_bpe_model.py + params.blank_id = sp.piece_to_id("") + params.vocab_size = sp.get_piece_size() + + logging.info(params) + + logging.info("About to create model") + model = get_transducer_model(params) + logging.info(model) + + num_param = sum([p.numel() for p in model.parameters()]) + logging.info(f"Number of model parameters: {num_param}") + + assert params.save_every_n >= params.average_period + model_avg: Optional[nn.Module] = None + if rank == 0: + # model_avg is only used with rank 0 + model_avg = copy.deepcopy(model).to(torch.float64) + + assert params.start_epoch > 0, params.start_epoch + checkpoints = load_checkpoint_if_available( + params=params, model=model, model_avg=model_avg + ) + + model.to(device) + if world_size > 1: + logging.info("Using DDP") + model = DDP(model, device_ids=[rank], find_unused_parameters=True) + + if params.multi_optim: + logging.info("Using seperate optimizers over encoder, decoder ...") + + enc_param = [] + enc_names = [] + + dec_names = [] + dec_param = [] + + for n, p in model.named_parameters(): + name = n.split('.')[1] + if name == 'encoder' and 'feature_extractor' not in n: + enc_names.append(n) + enc_param.append(p) + elif 'ctc_output' in n: + enc_names.append(n) + enc_param.append(p) + elif 'feature_extractor' not in n: + dec_names.append(n) + dec_param.append(p) + + optimizer_enc = ScaledAdam( + enc_param, + lr=params.peak_enc_lr, + clipping_scale=None, + parameters_names=[enc_names], + ) + optimizer_dec = ScaledAdam( + dec_param, + lr=params.peak_dec_lr, + clipping_scale=5.0, + parameters_names=[dec_names], + ) + + scheduler_enc = Eden(optimizer_enc, params.lr_batches, params.lr_epochs) + scheduler_dec = Eden(optimizer_dec, params.lr_batches, params.lr_epochs) + optimizer = [optimizer_enc, optimizer_dec] + scheduler = [scheduler_enc, scheduler_dec] + + else: + parameters_names = [] + parameters_names.append( + [name_param_pair[0] for name_param_pair in model.named_parameters()] + ) + + logging.info(f"len name = {len(parameters_names)}") + logging.info(f"len param = {len(list(model.parameters()))}") + + optimizer = ScaledAdam( + model.parameters(), + lr=params.base_lr, + clipping_scale=2.0, + parameters_names=parameters_names, + ) + + scheduler = Eden(optimizer, params.lr_batches, params.lr_epochs) + + if checkpoints and ("optimizer" in checkpoints or "optimizer_enc" in checkpoints): + if params.multi_optim: + logging.info("Loading optimizer state dict") + optimizer_enc.load_state_dict(checkpoints["optimizer_enc"]) + optimizer_dec.load_state_dict(checkpoints["optimizer_dec"]) + + else: + logging.info("Loading optimizer state dict") + optimizer.load_state_dict(checkpoints["optimizer"]) + + if checkpoints: + if ( + params.multi_optim + and "scheduler_enc" in checkpoints + and checkpoints["scheduler_enc"] is not None + ): + logging.info("Loading enc/dec scheduler state dict") + scheduler_enc.load_state_dict(checkpoints["scheduler_enc"]) + scheduler_dec.load_state_dict(checkpoints["scheduler_dec"]) + else: + logging.info("Loading scheduler state dict") + scheduler.load_state_dict(checkpoints["scheduler"]) + + if params.print_diagnostics: + opts = diagnostics.TensorDiagnosticOptions( + 2**22 + ) # allow 4 megabytes per sub-module + diagnostic = diagnostics.attach_diagnostics(model, opts) + + if params.inf_check: + register_inf_check_hooks(model) + + librispeech = LibriSpeechAsrDataModule(args) + + train_cuts = librispeech.train_clean_100_cuts() + if params.full_libri: + train_cuts += librispeech.train_clean_360_cuts() + train_cuts += librispeech.train_other_500_cuts() + + def remove_short_and_long_utt(c: Cut): + # Keep only utterances with duration between 1 second and 20 seconds + # + # Caution: There is a reason to select 20.0 here. Please see + # ../local/display_manifest_statistics.py + # + # You should use ../local/display_manifest_statistics.py to get + # an utterance duration distribution for your dataset to select + # the threshold + return 1.0 <= c.duration <= 20.0 + + train_cuts = train_cuts.filter(remove_short_and_long_utt) + + if params.start_batch > 0 and checkpoints and "sampler" in checkpoints: + # We only load the sampler's state dict when it loads a checkpoint + # saved in the middle of an epoch + sampler_state_dict = checkpoints["sampler"] + else: + sampler_state_dict = None + + train_dl = librispeech.train_dataloaders( + train_cuts, sampler_state_dict=sampler_state_dict + ) + + valid_cuts = librispeech.dev_clean_cuts() + valid_cuts += librispeech.dev_other_cuts() + valid_dl = librispeech.valid_dataloaders(valid_cuts) + + ''' + if not params.print_diagnostics: + scan_pessimistic_batches_for_oom( + model=model, + train_dl=train_dl, + optimizer=optimizer, + sp=sp, + params=params, + ) + ''' + + scaler = GradScaler(enabled=params.use_fp16, init_scale=1.0) + if checkpoints and "grad_scaler" in checkpoints: + logging.info("Loading grad scaler state dict") + scaler.load_state_dict(checkpoints["grad_scaler"]) + + for epoch in range(params.start_epoch, params.num_epochs + 1): + if params.multi_optim: + scheduler_enc.step_epoch(epoch - 1) + scheduler_dec.step_epoch(epoch - 1) + else: + scheduler.step_epoch(epoch - 1) + fix_random_seed(params.seed + epoch - 1) + train_dl.sampler.set_epoch(epoch - 1) + + if tb_writer is not None: + tb_writer.add_scalar("train/epoch", epoch, params.batch_idx_train) + + params.cur_epoch = epoch + + train_one_epoch( + params=params, + model=model, + model_avg=model_avg, + optimizer=optimizer, + scheduler=scheduler, + sp=sp, + train_dl=train_dl, + valid_dl=valid_dl, + scaler=scaler, + tb_writer=tb_writer, + world_size=world_size, + rank=rank, + wb=wb, + ) + + if params.print_diagnostics: + diagnostic.print_diagnostics() + break + + ''' + if epoch % 50 == 0: + save_checkpoint( + params=params, + model=model, + model_avg=model_avg, + optimizer=optimizer, + scheduler=scheduler, + sampler=train_dl.sampler, + scaler=scaler, + rank=rank, + ) + ''' + + logging.info("Done!") + + if world_size > 1: + torch.distributed.barrier() + cleanup_dist() + + +def run_adapter(rank, world_size, args, wb=None): + """ + Args: + rank: + It is a value between 0 and `world_size-1`, which is + passed automatically by `mp.spawn()` in :func:`main`. + The node with rank 0 is responsible for saving checkpoint. + world_size: + Number of GPUs for DDP training. + args: + The return value of get_parser().parse_args() + """ + params = get_params() + params.update(vars(args)) + + fix_random_seed(params.seed) + if world_size > 1: + setup_dist(rank, world_size, params.master_port) + + setup_logger(f"{params.exp_dir}/log/log-train") + logging.info("Training started") + + if args.tensorboard and rank == 0: + tb_writer = SummaryWriter(log_dir=f"{params.exp_dir}/tensorboard") + else: + tb_writer = None + + device = torch.device("cpu") + if torch.cuda.is_available(): + device = torch.device("cuda", rank) + logging.info(f"Device: {device}") + + sp = spm.SentencePieceProcessor() + sp.load(params.bpe_model) + + # is defined in local/train_bpe_model.py + params.blank_id = sp.piece_to_id("") + params.vocab_size = sp.get_piece_size() + + logging.info(params) + + logging.info("About to create model") + model = get_transducer_model(params) + + num_param = sum([p.numel() if p.requires_grad else 0 for p in model.parameters()]) + logging.info(f"Number of model parameters: {num_param}") + + assert params.save_every_n >= params.average_period + model_avg: Optional[nn.Module] = None + if rank == 0: + # model_avg is only used with rank 0 + #model_avg = copy.deepcopy(model).to(torch.float64) + model_avg = None + + assert params.start_epoch > 0, params.start_epoch + checkpoints = load_checkpoint_if_available( + params=params, model=model, model_avg=model_avg + ) + + model.to(device) + if world_size > 1: + logging.info("Using DDP") + model = DDP(model, device_ids=[rank], find_unused_parameters=True) + + adapter_names = [] + adapter_param = [] + for n, p in model.named_parameters(): + if 'q_proj.bias' in n or 'fc1.bias' in n: + adapter_names.append(n) + adapter_param.append(p) + else: + p.requires_grad = False + + ''' + if 'adapters' in n:# or 'joiner' in n or 'simple' in n or 'ctc' in n: + adapter_names.append(n) + adapter_param.append(p) + elif 'joiner' in n or 'simple' in n or 'ctc' in n: + p.requires_grad = True + else: + p.requires_grad = False + ''' + optimizer_adapter = ScaledAdam( + adapter_param, + lr=params.adapter_lr, + clipping_scale=5.0, + parameters_names=[adapter_names], + ) + + #for n, p in model.named_parameters(): + # p.requires_grad = False + + #prompt = torch.randn((100, 512), requires_grad=True) + #optimizer_adapter = ScaledAdam( + # [model.prompt], + # lr=params.adapter_lr, + # clipping_scale=5.0, + # parameters_names=['P'], + #) + + scheduler_adapter = Eden(optimizer_adapter, 10000, 7) #params.lr_batche, params.lr_epochs) + optimizer, scheduler = optimizer_adapter, scheduler_adapter + + librispeech = LibriSpeechAsrDataModule(args) + + ''' + if params.hpo: + train_cuts = librispeech.train_clean_10_cuts(option=params.gender) + else: + train_cuts = librispeech.train_clean_100_cuts(option=params.gender) + if params.full_libri: + train_cuts += librispeech.train_clean_360_cuts(option=params.gender) + train_cuts += librispeech.train_other_500_cuts(option=params.gender) + ''' + + #train_cuts = librispeech.train_clean_10_cuts(option='male') + #train_cuts = librispeech.test_clean_user(option='big') + train_cuts = librispeech.vox_cuts(option=params.spk_id) + + def remove_short_and_long_utt(c: Cut): + return 1.0 <= c.duration <= 20.0 + + train_cuts = train_cuts.filter(remove_short_and_long_utt) + + sampler_state_dict = None + + train_dl = librispeech.train_dataloaders( + train_cuts, sampler_state_dict=sampler_state_dict + ) + #train_dl = librispeech.test_dataloaders( + # train_cuts + #) + + ''' + print('\n'*5) + print('-'*30) + for batch in train_dl: + print(batch) + print('-'*30) + print('\n'*5) + exit() + ''' + + valid_cuts = librispeech.dev_clean_cuts(option=params.gender) + valid_cuts += librispeech.dev_other_cuts(option=params.gender) + valid_dl = librispeech.valid_dataloaders(valid_cuts) + + scaler = GradScaler(enabled=params.use_fp16, init_scale=1.0) + + for epoch in range(params.start_epoch, params.num_epochs + 1): + logging.info(f"update num : {params.batch_idx_train}") + scheduler.step_epoch(epoch - 1) + fix_random_seed(params.seed + epoch - 1) + train_dl.sampler.set_epoch(epoch - 1) + + if tb_writer is not None: + tb_writer.add_scalar("train/epoch", epoch, params.batch_idx_train) + + params.cur_epoch = epoch + + train_one_epoch( + params=params, + model=model, + model_avg=model_avg, + optimizer=optimizer, + scheduler=scheduler, + sp=sp, + train_dl=train_dl, + valid_dl=valid_dl, + scaler=scaler, + tb_writer=tb_writer, + world_size=world_size, + rank=rank, + wb=wb, + ) + + if params.print_diagnostics: + diagnostic.print_diagnostics() + break + + ''' + if epoch % 10 == 0: + save_checkpoint( + params=params, + model=model, + model_avg=model_avg, + optimizer=optimizer, + scheduler=scheduler, + sampler=train_dl.sampler, + scaler=scaler, + rank=rank, + ) + ''' + + logging.info("Done!") + + if world_size > 1: + torch.distributed.barrier() + cleanup_dist() + + +def display_and_save_batch( + batch: dict, + params: AttributeDict, + sp: spm.SentencePieceProcessor, +) -> None: + """Display the batch statistics and save the batch into disk. + + Args: + batch: + A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()` + for the content in it. + params: + Parameters for training. See :func:`get_params`. + sp: + The BPE model. + """ + from lhotse.utils import uuid4 + + filename = f"{params.exp_dir}/batch-{uuid4()}.pt" + logging.info(f"Saving batch to {filename}") + torch.save(batch, filename) + + supervisions = batch["supervisions"] + features = batch["inputs"] + + logging.info(f"features shape: {features.shape}") + + y = sp.encode(supervisions["text"], out_type=int) + num_tokens = sum(len(i) for i in y) + logging.info(f"num tokens: {num_tokens}") + + +def scan_pessimistic_batches_for_oom( + model: Union[nn.Module, DDP], + train_dl: torch.utils.data.DataLoader, + optimizer: torch.optim.Optimizer, + sp: spm.SentencePieceProcessor, + params: AttributeDict, +): + from lhotse.dataset import find_pessimistic_batches + + logging.info( + "Sanity check -- see if any of the batches in epoch 1 would cause OOM." + ) + batches, crit_values = find_pessimistic_batches(train_dl.sampler) + for criterion, cuts in batches.items(): + batch = train_dl.dataset[cuts] + try: + with torch.cuda.amp.autocast(enabled=params.use_fp16): + loss, _ = compute_loss( + params=params, + model=model, + sp=sp, + batch=batch, + is_training=True, + ) + loss.backward() + optimizer.zero_grad() + except Exception as e: + if "CUDA out of memory" in str(e): + logging.error( + "Your GPU ran out of memory with the current " + "max_duration setting. We recommend decreasing " + "max_duration and trying again.\n" + f"Failing criterion: {criterion} " + f"(={crit_values[criterion]}) ..." + ) + display_and_save_batch(batch, params=params, sp=sp) + raise + logging.info( + f"Maximum memory allocated so far is {torch.cuda.max_memory_allocated()//1000000}MB" + ) + + +def main(): + parser = get_parser() + LibriSpeechAsrDataModule.add_arguments(parser) + args = parser.parse_args() + if args.wandb: args.exp_dir = args.exp_dir + str(random.randint(0,400)) + args.exp_dir = Path(args.exp_dir) + + logging.info("save arguments to config.yaml...") + save_args(args) + + if args.wandb: wb = wandb.init(project="d2v-adapter", entity="dohe0342", config=vars(args)) + else: wb = None + + world_size = args.world_size + assert world_size >= 1 + if world_size > 1: + mp.spawn(run if not args.add_adapter else run_adapter, + args=(world_size, args, wb), + nprocs=world_size, + join=True + ) + else: + if args.add_adapter: run_adapter(rank=0, world_size=1, args=args, wb=wb) + else: run(rank=0, world_size=1, args=args, wb=wb) + +torch.set_num_threads(1) +torch.set_num_interop_threads(1) + +if __name__ == "__main__": + main() diff --git a/egs/tedlium3/ASR/pruned_transducer_stateless_d2v_v2/checkpoint.py b/egs/tedlium3/ASR/pruned_transducer_stateless_d2v_v2/checkpoint.py new file mode 100644 index 000000000..03d9632f5 --- /dev/null +++ b/egs/tedlium3/ASR/pruned_transducer_stateless_d2v_v2/checkpoint.py @@ -0,0 +1,225 @@ +# Copyright 2021-2022 Xiaomi Corporation (authors: Fangjun Kuang, +# Zengwei Yao) +# +# See ../../LICENSE for clarification regarding multiple authors +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + + +import glob +import logging +import os +import re +from pathlib import Path +from typing import Any, Dict, List, Optional, Union + +from lhotse.dataset.sampling.base import CutSampler + +import torch +import torch.nn as nn +from torch import Tensor +from torch.cuda.amp import GradScaler +from torch.nn.parallel import DistributedDataParallel as DDP +from torch.optim import Optimizer + +# use duck typing for LRScheduler since we have different possibilities, see +# our class LRScheduler. +LRSchedulerType = object + + +def save_checkpoint( + filename: Path, + model: Union[nn.Module, DDP], + model_avg: Optional[nn.Module] = None, + params: Optional[Dict[str, Any]] = None, + optimizer = None, + scheduler = None, + scaler: Optional[GradScaler] = None, + sampler: Optional[CutSampler] = None, + rank: int = 0, +) -> None: + """Save training information to a file. + + Args: + filename: + The checkpoint filename. + model: + The model to be saved. We only save its `state_dict()`. + model_avg: + The stored model averaged from the start of training. + params: + User defined parameters, e.g., epoch, loss. + optimizer: + The optimizer to be saved. We only save its `state_dict()`. + scheduler: + The scheduler to be saved. We only save its `state_dict()`. + scalar: + The GradScaler to be saved. We only save its `state_dict()`. + rank: + Used in DDP. We save checkpoint only for the node whose rank is 0. + Returns: + Return None. + """ + if rank != 0: + return + + logging.info(f"Saving checkpoint to {filename}") + + if isinstance(model, DDP): + model = model.module + + if type(optimizer) == list: + checkpoint = { + "model": model.state_dict(), + "optimizer_enc": optimizer[0].state_dict() if optimizer is not None else None, + "optimizer_dec": optimizer[1].state_dict() if optimizer is not None else None, + "scheduler_enc": scheduler[0].state_dict() if scheduler is not None else None, + "scheduler_dec": scheduler[1].state_dict() if scheduler is not None else None, + "grad_scaler": scaler.state_dict() if scaler is not None else None, + "sampler": sampler.state_dict() if sampler is not None else None, + } + else: + checkpoint = { + "model": model.state_dict(), + "optimizer": optimizer.state_dict() if optimizer is not None else None, + "scheduler": scheduler.state_dict() if scheduler is not None else None, + "grad_scaler": scaler.state_dict() if scaler is not None else None, + "sampler": sampler.state_dict() if sampler is not None else None, + } + + + if model_avg is not None: + checkpoint["model_avg"] = model_avg.to(torch.float32).state_dict() + + if params: + for k, v in params.items(): + assert k not in checkpoint + checkpoint[k] = v + + torch.save(checkpoint, filename) + + +def load_checkpoint( + filename: Path, + model: nn.Module, + model_avg: Optional[nn.Module] = None, + optimizer = None, + scheduler = None, + scaler: Optional[GradScaler] = None, + sampler: Optional[CutSampler] = None, + strict: bool = True, +) -> Dict[str, Any]: + """ + TODO: document it + """ + logging.info(f"Loading checkpoint from {filename}") + checkpoint = torch.load(filename, map_location="cpu") + + if next(iter(checkpoint["model"])).startswith("module."): + logging.info("Loading checkpoint saved by DDP") + + dst_state_dict = model.state_dict() + src_state_dict = checkpoint["model"] + for key in dst_state_dict.keys(): + src_key = "{}.{}".format("module", key) + dst_state_dict[key] = src_state_dict.pop(src_key) + assert len(src_state_dict) == 0 + model.load_state_dict(dst_state_dict, strict=strict) + else: + model.load_state_dict(checkpoint["model"], strict=strict) + + checkpoint.pop("model") + + if model_avg is not None and "model_avg" in checkpoint: + logging.info("Loading averaged model") + model_avg.load_state_dict(checkpoint["model_avg"], strict=strict) + checkpoint.pop("model_avg") + + def load(name, obj): + s = checkpoint.get(name, None) + if obj and s: + obj.load_state_dict(s) + checkpoint.pop(name) + + if type(optimizer) == list: + load("optimizer_enc", optimizer[0]) + load("optimizer_dec", optimizer[1]) + load("scheduler_enc", scheduler[0]) + load("scheduler_dec", scheduler[1]) + else: + load("optimizer", optimizer) + load("scheduler", scheduler) + + load("grad_scaler", scaler) + load("sampler", sampler) + + return checkpoint + + +def save_checkpoint_with_global_batch_idx( + out_dir: Path, + global_batch_idx: int, + model: Union[nn.Module, DDP], + model_avg: Optional[nn.Module] = None, + params: Optional[Dict[str, Any]] = None, + optimizer = None, + scheduler = None, + scaler: Optional[GradScaler] = None, + sampler: Optional[CutSampler] = None, + rank: int = 0, +): + """Save training info after processing given number of batches. + + Args: + out_dir: + The directory to save the checkpoint. + global_batch_idx: + The number of batches processed so far from the very start of the + training. The saved checkpoint will have the following filename: + + f'out_dir / checkpoint-{global_batch_idx}.pt' + model: + The neural network model whose `state_dict` will be saved in the + checkpoint. + model_avg: + The stored model averaged from the start of training. + params: + A dict of training configurations to be saved. + optimizer: + The optimizer used in the training. Its `state_dict` will be saved. + scheduler: + The learning rate scheduler used in the training. Its `state_dict` will + be saved. + scaler: + The scaler used for mix precision training. Its `state_dict` will + be saved. + sampler: + The sampler used in the training dataset. + rank: + The rank ID used in DDP training of the current node. Set it to 0 + if DDP is not used. + """ + out_dir = Path(out_dir) + out_dir.mkdir(parents=True, exist_ok=True) + filename = out_dir / f"checkpoint-{global_batch_idx}.pt" + save_checkpoint( + filename=filename, + model=model, + model_avg=model_avg, + params=params, + optimizer=optimizer, + scheduler=scheduler, + scaler=scaler, + sampler=sampler, + rank=rank, + ) diff --git a/egs/tedlium3/ASR/pruned_transducer_stateless_d2v_v2/convolution.py b/egs/tedlium3/ASR/pruned_transducer_stateless_d2v_v2/convolution.py new file mode 100644 index 000000000..e198045c3 --- /dev/null +++ b/egs/tedlium3/ASR/pruned_transducer_stateless_d2v_v2/convolution.py @@ -0,0 +1,83 @@ +#!/usr/bin/env python3 +# -*- coding: utf-8 -*- + +# Copyright 2020 Johns Hopkins University (Shinji Watanabe) +# Northwestern Polytechnical University (Pengcheng Guo) +# Apache 2.0 (http://www.apache.org/licenses/LICENSE-2.0) + +"""ConvolutionModule definition.""" + +from torch import nn + + +class ConvolutionModule(nn.Module): + """ConvolutionModule in Conformer model. + + Args: + channels (int): The number of channels of conv layers. + kernel_size (int): Kernerl size of conv layers. + + """ + + def __init__(self, channels, kernel_size, activation=nn.SiLU(), bias=True): + """Construct an ConvolutionModule object.""" + super(ConvolutionModule, self).__init__() + # kernerl_size should be a odd number for 'SAME' padding + assert (kernel_size - 1) % 2 == 0 + + self.pointwise_conv1 = nn.Conv1d( + channels, + 2 * channels, + kernel_size=1, + stride=1, + padding=0, + bias=bias, + ) + self.depthwise_conv = nn.Conv1d( + channels, + channels, + kernel_size, + stride=1, + padding=(kernel_size - 1) // 2, + groups=channels, + bias=bias, + ) + self.norm = nn.BatchNorm1d(channels) + self.layer_norm = nn.LayerNorm(channels) + + self.pointwise_conv2 = nn.Conv1d( + channels, + channels, + kernel_size=1, + stride=1, + padding=0, + bias=bias, + ) + self.activation = activation + + def forward(self, x): + """Compute convolution module. + + Args: + x (torch.Tensor): Input tensor (#batch, time, channels). + + Returns: + torch.Tensor: Output tensor (#batch, time, channels). + + """ + # exchange the temporal dimension and the feature dimension + x = x.transpose(1, 2) + + # GLU mechanism + x = self.pointwise_conv1(x) # (batch, 2*channel, dim) + x = nn.functional.glu(x, dim=1) # (batch, channel, dim) + + # 1D Depthwise Conv + x = self.depthwise_conv(x) + x = self.activation(self.norm(x)) + + x = self.pointwise_conv2(x) + x = x.transpose(1, 2) + x = self.layer_norm(x) + + return x diff --git a/egs/tedlium3/ASR/pruned_transducer_stateless_d2v_v2/ctc_decode.py b/egs/tedlium3/ASR/pruned_transducer_stateless_d2v_v2/ctc_decode.py new file mode 100755 index 000000000..9c23e7d66 --- /dev/null +++ b/egs/tedlium3/ASR/pruned_transducer_stateless_d2v_v2/ctc_decode.py @@ -0,0 +1,818 @@ +#!/usr/bin/env python3 +# +# Copyright 2021-2022 Xiaomi Corporation (Author: Fangjun Kuang, +# Liyong Guo, +# Quandong Wang, +# Zengwei Yao) +# +# See ../../../../LICENSE for clarification regarding multiple authors +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +""" +Usage: +(1) ctc-decoding +./pruned_transducer_stateless7_ctc/ctc_decode.py \ + --epoch 30 \ + --avg 15 \ + --exp-dir ./pruned_transducer_stateless7_ctc/exp \ + --max-duration 600 \ + --decoding-method ctc-decoding + +(2) 1best +./pruned_transducer_stateless7_ctc/ctc_decode.py \ + --epoch 30 \ + --avg 15 \ + --exp-dir ./pruned_transducer_stateless7_ctc/exp \ + --max-duration 600 \ + --hlg-scale 0.8 \ + --decoding-method 1best + +(3) nbest +./pruned_transducer_stateless7_ctc/ctc_decode.py \ + --epoch 30 \ + --avg 15 \ + --exp-dir ./pruned_transducer_stateless7_ctc/exp \ + --max-duration 600 \ + --hlg-scale 0.8 \ + --decoding-method 1best + +(4) nbest-rescoring +./pruned_transducer_stateless7_ctc/ctc_decode.py \ + --epoch 30 \ + --avg 15 \ + --exp-dir ./pruned_transducer_stateless7_ctc/exp \ + --max-duration 600 \ + --hlg-scale 0.8 \ + --lm-dir data/lm \ + --decoding-method nbest-rescoring + +(5) whole-lattice-rescoring +./pruned_transducer_stateless7_ctc/ctc_decode.py \ + --epoch 30 \ + --avg 15 \ + --exp-dir ./pruned_transducer_stateless7_ctc/exp \ + --max-duration 600 \ + --hlg-scale 0.8 \ + --lm-dir data/lm \ + --decoding-method whole-lattice-rescoring +""" + + +import argparse +import logging +import math +from collections import defaultdict +from pathlib import Path +from typing import Dict, List, Optional, Tuple + +import k2 +import sentencepiece as spm +import torch +import torch.nn as nn +from asr_datamodule import LibriSpeechAsrDataModule +from train import add_model_arguments, get_params, get_transducer_model + +from icefall.checkpoint import ( + average_checkpoints, + average_checkpoints_with_averaged_model, + find_checkpoints, + load_checkpoint, +) +from icefall.decode import ( + get_lattice, + nbest_decoding, + nbest_oracle, + one_best_decoding, + rescore_with_n_best_list, + rescore_with_whole_lattice, +) +from icefall.lexicon import Lexicon +from icefall.utils import ( + AttributeDict, + get_texts, + setup_logger, + store_transcripts, + str2bool, + write_error_stats, +) + +LOG_EPS = math.log(1e-10) + + +def get_parser(): + parser = argparse.ArgumentParser( + formatter_class=argparse.ArgumentDefaultsHelpFormatter + ) + + parser.add_argument( + "--epoch", + type=int, + default=30, + help="""It specifies the checkpoint to use for decoding. + Note: Epoch counts from 1. + You can specify --avg to use more checkpoints for model averaging.""", + ) + + parser.add_argument( + "--iter", + type=int, + default=0, + help="""If positive, --epoch is ignored and it + will use the checkpoint exp_dir/checkpoint-iter.pt. + You can specify --avg to use more checkpoints for model averaging. + """, + ) + + parser.add_argument( + "--avg", + type=int, + default=15, + help="Number of checkpoints to average. Automatically select " + "consecutive checkpoints before the checkpoint specified by " + "'--epoch' and '--iter'", + ) + + parser.add_argument( + "--use-averaged-model", + type=str2bool, + default=True, + help="Whether to load averaged model. Currently it only supports " + "using --epoch. If True, it would decode with the averaged model " + "over the epoch range from `epoch-avg` (excluded) to `epoch`." + "Actually only the models with epoch number of `epoch-avg` and " + "`epoch` are loaded for averaging. ", + ) + + parser.add_argument( + "--exp-dir", + type=str, + default="pruned_transducer_stateless7_ctc/exp", + help="The experiment dir", + ) + + parser.add_argument( + "--bpe-model", + type=str, + default="data/lang_bpe_500/bpe.model", + help="Path to the BPE model", + ) + + parser.add_argument( + "--lang-dir", + type=Path, + default="data/lang_bpe_500", + help="The lang dir containing word table and LG graph", + ) + + parser.add_argument( + "--context-size", + type=int, + default=2, + help="The context size in the decoder. 1 means bigram; 2 means tri-gram", + ) + + parser.add_argument( + "--decoding-method", + type=str, + default="ctc-decoding", + help="""Decoding method. + Supported values are: + - (1) ctc-decoding. Use CTC decoding. It uses a sentence piece + model, i.e., lang_dir/bpe.model, to convert word pieces to words. + It needs neither a lexicon nor an n-gram LM. + - (2) 1best. Extract the best path from the decoding lattice as the + decoding result. + - (3) nbest. Extract n paths from the decoding lattice; the path + with the highest score is the decoding result. + - (4) nbest-rescoring. Extract n paths from the decoding lattice, + rescore them with an n-gram LM (e.g., a 4-gram LM), the path with + the highest score is the decoding result. + - (5) whole-lattice-rescoring. Rescore the decoding lattice with an + n-gram LM (e.g., a 4-gram LM), the best path of rescored lattice + is the decoding result. + you have trained an RNN LM using ./rnn_lm/train.py + - (6) nbest-oracle. Its WER is the lower bound of any n-best + rescoring method can achieve. Useful for debugging n-best + rescoring method. + """, + ) + + parser.add_argument( + "--num-paths", + type=int, + default=100, + help="""Number of paths for n-best based decoding method. + Used only when "method" is one of the following values: + nbest, nbest-rescoring, and nbest-oracle + """, + ) + + parser.add_argument( + "--nbest-scale", + type=float, + default=0.5, + help="""The scale to be applied to `lattice.scores`. + It's needed if you use any kinds of n-best based rescoring. + Used only when "method" is one of the following values: + nbest, nbest-rescoring, and nbest-oracle + A smaller value results in more unique paths. + """, + ) + + parser.add_argument( + "--hlg-scale", + type=float, + default=0.8, + help="""The scale to be applied to `hlg.scores`. + """, + ) + + parser.add_argument( + "--lm-dir", + type=str, + default="data/lm", + help="""The n-gram LM dir. + It should contain either G_4_gram.pt or G_4_gram.fst.txt + """, + ) + + add_model_arguments(parser) + + return parser + + +def get_decoding_params() -> AttributeDict: + """Parameters for decoding.""" + params = AttributeDict( + { + "frame_shift_ms": 10, + "search_beam": 20, + "output_beam": 8, + "min_active_states": 30, + "max_active_states": 10000, + "use_double_scores": True, + } + ) + return params + + +def decode_one_batch( + params: AttributeDict, + model: nn.Module, + HLG: Optional[k2.Fsa], + H: Optional[k2.Fsa], + bpe_model: Optional[spm.SentencePieceProcessor], + batch: dict, + word_table: k2.SymbolTable, + G: Optional[k2.Fsa] = None, +) -> Dict[str, List[List[str]]]: + """Decode one batch and return the result in a dict. The dict has the + following format: + - key: It indicates the setting used for decoding. For example, + if no rescoring is used, the key is the string `no_rescore`. + If LM rescoring is used, the key is the string `lm_scale_xxx`, + where `xxx` is the value of `lm_scale`. An example key is + `lm_scale_0.7` + - value: It contains the decoding result. `len(value)` equals to + batch size. `value[i]` is the decoding result for the i-th + utterance in the given batch. + + Args: + params: + It's the return value of :func:`get_params`. + + - params.decoding_method is "1best", it uses 1best decoding without LM rescoring. + - params.decoding_method is "nbest", it uses nbest decoding without LM rescoring. + - params.decoding_method is "nbest-rescoring", it uses nbest LM rescoring. + - params.decoding_method is "whole-lattice-rescoring", it uses whole lattice LM + rescoring. + + model: + The neural model. + HLG: + The decoding graph. Used only when params.decoding_method is NOT ctc-decoding. + H: + The ctc topo. Used only when params.decoding_method is ctc-decoding. + bpe_model: + The BPE model. Used only when params.decoding_method is ctc-decoding. + batch: + It is the return value from iterating + `lhotse.dataset.K2SpeechRecognitionDataset`. See its documentation + for the format of the `batch`. + word_table: + The word symbol table. + G: + An LM. It is not None when params.decoding_method is "nbest-rescoring" + or "whole-lattice-rescoring". In general, the G in HLG + is a 3-gram LM, while this G is a 4-gram LM. + Returns: + Return the decoding result. See above description for the format of + the returned dict. Note: If it decodes to nothing, then return None. + """ + if HLG is not None: + device = HLG.device + else: + device = H.device + feature = batch["inputs"] + assert feature.ndim == 3 + feature = feature.to(device) + # at entry, feature is (N, T, C) + + supervisions = batch["supervisions"] + feature_lens = supervisions["num_frames"].to(device) + + encoder_out, encoder_out_lens = model.encoder(feature, feature_lens) + nnet_output = model.ctc_output(encoder_out) + # nnet_output is (N, T, C) + + supervision_segments = torch.stack( + ( + supervisions["sequence_idx"], + supervisions["start_frame"] // params.subsampling_factor, + supervisions["num_frames"] // params.subsampling_factor, + ), + 1, + ).to(torch.int32) + + if H is None: + assert HLG is not None + decoding_graph = HLG + else: + assert HLG is None + assert bpe_model is not None + decoding_graph = H + + lattice = get_lattice( + nnet_output=nnet_output, + decoding_graph=decoding_graph, + supervision_segments=supervision_segments, + search_beam=params.search_beam, + output_beam=params.output_beam, + min_active_states=params.min_active_states, + max_active_states=params.max_active_states, + subsampling_factor=params.subsampling_factor, + ) + + if params.decoding_method == "ctc-decoding": + best_path = one_best_decoding( + lattice=lattice, use_double_scores=params.use_double_scores + ) + # Note: `best_path.aux_labels` contains token IDs, not word IDs + # since we are using H, not HLG here. + # + # token_ids is a lit-of-list of IDs + token_ids = get_texts(best_path) + + # hyps is a list of str, e.g., ['xxx yyy zzz', ...] + hyps = bpe_model.decode(token_ids) + + # hyps is a list of list of str, e.g., [['xxx', 'yyy', 'zzz'], ... ] + hyps = [s.split() for s in hyps] + key = "ctc-decoding" + return {key: hyps} + + if params.decoding_method == "nbest-oracle": + # Note: You can also pass rescored lattices to it. + # We choose the HLG decoded lattice for speed reasons + # as HLG decoding is faster and the oracle WER + # is only slightly worse than that of rescored lattices. + best_path = nbest_oracle( + lattice=lattice, + num_paths=params.num_paths, + ref_texts=supervisions["text"], + word_table=word_table, + nbest_scale=params.nbest_scale, + oov="", + ) + hyps = get_texts(best_path) + hyps = [[word_table[i] for i in ids] for ids in hyps] + key = f"oracle_{params.num_paths}_nbest_scale_{params.nbest_scale}" # noqa + return {key: hyps} + + if params.decoding_method in ["1best", "nbest"]: + if params.decoding_method == "1best": + best_path = one_best_decoding( + lattice=lattice, use_double_scores=params.use_double_scores + ) + key = "no_rescore" + else: + best_path = nbest_decoding( + lattice=lattice, + num_paths=params.num_paths, + use_double_scores=params.use_double_scores, + nbest_scale=params.nbest_scale, + ) + key = f"no_rescore-nbest-scale-{params.nbest_scale}-{params.num_paths}" # noqa + + hyps = get_texts(best_path) + hyps = [[word_table[i] for i in ids] for ids in hyps] + return {key: hyps} + + assert params.decoding_method in [ + "nbest-rescoring", + "whole-lattice-rescoring", + ] + + lm_scale_list = [0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7] + lm_scale_list += [0.8, 0.9, 1.0, 1.1, 1.2, 1.3] + lm_scale_list += [1.4, 1.5, 1.6, 1.7, 1.8, 1.9, 2.0] + + if params.decoding_method == "nbest-rescoring": + best_path_dict = rescore_with_n_best_list( + lattice=lattice, + G=G, + num_paths=params.num_paths, + lm_scale_list=lm_scale_list, + nbest_scale=params.nbest_scale, + ) + elif params.decoding_method == "whole-lattice-rescoring": + best_path_dict = rescore_with_whole_lattice( + lattice=lattice, + G_with_epsilon_loops=G, + lm_scale_list=lm_scale_list, + ) + else: + assert False, f"Unsupported decoding method: {params.decoding_method}" + + ans = dict() + if best_path_dict is not None: + for lm_scale_str, best_path in best_path_dict.items(): + hyps = get_texts(best_path) + hyps = [[word_table[i] for i in ids] for ids in hyps] + ans[lm_scale_str] = hyps + else: + ans = None + return ans + + +def decode_dataset( + dl: torch.utils.data.DataLoader, + params: AttributeDict, + model: nn.Module, + HLG: Optional[k2.Fsa], + H: Optional[k2.Fsa], + bpe_model: Optional[spm.SentencePieceProcessor], + word_table: k2.SymbolTable, + G: Optional[k2.Fsa] = None, +) -> Dict[str, List[Tuple[str, List[str], List[str]]]]: + """Decode dataset. + + Args: + dl: + PyTorch's dataloader containing the dataset to decode. + params: + It is returned by :func:`get_params`. + model: + The neural model. + HLG: + The decoding graph. Used only when params.decoding_method is NOT ctc-decoding. + H: + The ctc topo. Used only when params.decoding_method is ctc-decoding. + bpe_model: + The BPE model. Used only when params.decoding_method is ctc-decoding. + word_table: + It is the word symbol table. + G: + An LM. It is not None when params.decoding_method is "nbest-rescoring" + or "whole-lattice-rescoring". In general, the G in HLG + is a 3-gram LM, while this G is a 4-gram LM. + Returns: + Return a dict, whose key may be "no-rescore" if no LM rescoring + is used, or it may be "lm_scale_0.7" if LM rescoring is used. + Its value is a list of tuples. Each tuple contains two elements: + The first is the reference transcript, and the second is the + predicted result. + """ + num_cuts = 0 + + try: + num_batches = len(dl) + except TypeError: + num_batches = "?" + + results = defaultdict(list) + for batch_idx, batch in enumerate(dl): + texts = batch["supervisions"]["text"] + cut_ids = [cut.id for cut in batch["supervisions"]["cut"]] + + hyps_dict = decode_one_batch( + params=params, + model=model, + HLG=HLG, + H=H, + bpe_model=bpe_model, + batch=batch, + word_table=word_table, + G=G, + ) + + for name, hyps in hyps_dict.items(): + this_batch = [] + assert len(hyps) == len(texts) + for cut_id, hyp_words, ref_text in zip(cut_ids, hyps, texts): + ref_words = ref_text.split() + this_batch.append((cut_id, ref_words, hyp_words)) + + results[name].extend(this_batch) + + num_cuts += len(texts) + + if batch_idx % 100 == 0: + batch_str = f"{batch_idx}/{num_batches}" + + logging.info(f"batch {batch_str}, cuts processed until now is {num_cuts}") + return results + + +def save_results( + params: AttributeDict, + test_set_name: str, + results_dict: Dict[str, List[Tuple[str, List[str], List[str]]]], +): + test_set_wers = dict() + for key, results in results_dict.items(): + recog_path = ( + params.res_dir / f"recogs-{test_set_name}-{key}-{params.suffix}.txt" + ) + results = sorted(results) + store_transcripts(filename=recog_path, texts=results) + logging.info(f"The transcripts are stored in {recog_path}") + + # The following prints out WERs, per-word error statistics and aligned + # ref/hyp pairs. + errs_filename = ( + params.res_dir / f"errs-{test_set_name}-{key}-{params.suffix}.txt" + ) + with open(errs_filename, "w") as f: + wer = write_error_stats(f, f"{test_set_name}-{key}", results) + test_set_wers[key] = wer + + logging.info("Wrote detailed error stats to {}".format(errs_filename)) + + test_set_wers = sorted(test_set_wers.items(), key=lambda x: x[1]) + errs_info = ( + params.res_dir / f"wer-summary-{test_set_name}-{key}-{params.suffix}.txt" + ) + with open(errs_info, "w") as f: + print("settings\tWER", file=f) + for key, val in test_set_wers: + print("{}\t{}".format(key, val), file=f) + + s = "\nFor {}, WER of different settings are:\n".format(test_set_name) + note = "\tbest for {}".format(test_set_name) + for key, val in test_set_wers: + s += "{}\t{}{}\n".format(key, val, note) + note = "" + logging.info(s) + + +@torch.no_grad() +def main(): + parser = get_parser() + LibriSpeechAsrDataModule.add_arguments(parser) + args = parser.parse_args() + args.exp_dir = Path(args.exp_dir) + args.lang_dir = Path(args.lang_dir) + args.lm_dir = Path(args.lm_dir) + + params = get_params() + # add decoding params + params.update(get_decoding_params()) + params.update(vars(args)) + + assert params.decoding_method in ( + "ctc-decoding", + "1best", + "nbest", + "nbest-rescoring", + "whole-lattice-rescoring", + "nbest-oracle", + ) + params.res_dir = params.exp_dir / params.decoding_method + + if params.iter > 0: + params.suffix = f"iter-{params.iter}-avg-{params.avg}" + else: + params.suffix = f"epoch-{params.epoch}-avg-{params.avg}" + + if params.use_averaged_model: + params.suffix += "-use-averaged-model" + + setup_logger(f"{params.res_dir}/log-decode-{params.suffix}") + logging.info("Decoding started") + + device = torch.device("cpu") + if torch.cuda.is_available(): + device = torch.device("cuda", 0) + + logging.info(f"Device: {device}") + logging.info(params) + + lexicon = Lexicon(params.lang_dir) + max_token_id = max(lexicon.tokens) + num_classes = max_token_id + 1 # +1 for the blank + + params.vocab_size = num_classes + # and are defined in local/train_bpe_model.py + params.blank_id = 0 + + if params.decoding_method == "ctc-decoding": + HLG = None + H = k2.ctc_topo( + max_token=max_token_id, + modified=False, + device=device, + ) + bpe_model = spm.SentencePieceProcessor() + bpe_model.load(str(params.lang_dir / "bpe.model")) + else: + H = None + bpe_model = None + HLG = k2.Fsa.from_dict( + torch.load(f"{params.lang_dir}/HLG.pt", map_location=device) + ) + assert HLG.requires_grad is False + + HLG.scores *= params.hlg_scale + if not hasattr(HLG, "lm_scores"): + HLG.lm_scores = HLG.scores.clone() + + if params.decoding_method in ( + "nbest-rescoring", + "whole-lattice-rescoring", + ): + if not (params.lm_dir / "G_4_gram.pt").is_file(): + logging.info("Loading G_4_gram.fst.txt") + logging.warning("It may take 8 minutes.") + with open(params.lm_dir / "G_4_gram.fst.txt") as f: + first_word_disambig_id = lexicon.word_table["#0"] + + G = k2.Fsa.from_openfst(f.read(), acceptor=False) + # G.aux_labels is not needed in later computations, so + # remove it here. + del G.aux_labels + # CAUTION: The following line is crucial. + # Arcs entering the back-off state have label equal to #0. + # We have to change it to 0 here. + G.labels[G.labels >= first_word_disambig_id] = 0 + # See https://github.com/k2-fsa/k2/issues/874 + # for why we need to set G.properties to None + G.__dict__["_properties"] = None + G = k2.Fsa.from_fsas([G]).to(device) + G = k2.arc_sort(G) + # Save a dummy value so that it can be loaded in C++. + # See https://github.com/pytorch/pytorch/issues/67902 + # for why we need to do this. + G.dummy = 1 + + torch.save(G.as_dict(), params.lm_dir / "G_4_gram.pt") + else: + logging.info("Loading pre-compiled G_4_gram.pt") + d = torch.load(params.lm_dir / "G_4_gram.pt", map_location=device) + G = k2.Fsa.from_dict(d) + + if params.decoding_method == "whole-lattice-rescoring": + # Add epsilon self-loops to G as we will compose + # it with the whole lattice later + G = k2.add_epsilon_self_loops(G) + G = k2.arc_sort(G) + G = G.to(device) + + # G.lm_scores is used to replace HLG.lm_scores during + # LM rescoring. + G.lm_scores = G.scores.clone() + else: + G = None + + logging.info("About to create model") + model = get_transducer_model(params) + + if not params.use_averaged_model: + if params.iter > 0: + filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[ + : params.avg + ] + if len(filenames) == 0: + raise ValueError( + f"No checkpoints found for" + f" --iter {params.iter}, --avg {params.avg}" + ) + elif len(filenames) < params.avg: + raise ValueError( + f"Not enough checkpoints ({len(filenames)}) found for" + f" --iter {params.iter}, --avg {params.avg}" + ) + logging.info(f"averaging {filenames}") + model.to(device) + model.load_state_dict(average_checkpoints(filenames, device=device)) + elif params.avg == 1: + load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model) + else: + start = params.epoch - params.avg + 1 + filenames = [] + for i in range(start, params.epoch + 1): + if i >= 1: + filenames.append(f"{params.exp_dir}/epoch-{i}.pt") + logging.info(f"averaging {filenames}") + model.to(device) + model.load_state_dict(average_checkpoints(filenames, device=device)) + else: + if params.iter > 0: + filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[ + : params.avg + 1 + ] + if len(filenames) == 0: + raise ValueError( + f"No checkpoints found for" + f" --iter {params.iter}, --avg {params.avg}" + ) + elif len(filenames) < params.avg + 1: + raise ValueError( + f"Not enough checkpoints ({len(filenames)}) found for" + f" --iter {params.iter}, --avg {params.avg}" + ) + filename_start = filenames[-1] + filename_end = filenames[0] + logging.info( + "Calculating the averaged model over iteration checkpoints" + f" from {filename_start} (excluded) to {filename_end}" + ) + model.to(device) + model.load_state_dict( + average_checkpoints_with_averaged_model( + filename_start=filename_start, + filename_end=filename_end, + device=device, + ) + ) + else: + assert params.avg > 0, params.avg + start = params.epoch - params.avg + assert start >= 1, start + filename_start = f"{params.exp_dir}/epoch-{start}.pt" + filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt" + logging.info( + f"Calculating the averaged model over epoch range from " + f"{start} (excluded) to {params.epoch}" + ) + model.to(device) + model.load_state_dict( + average_checkpoints_with_averaged_model( + filename_start=filename_start, + filename_end=filename_end, + device=device, + ) + ) + + model.to(device) + model.eval() + + num_param = sum([p.numel() for p in model.parameters()]) + logging.info(f"Number of model parameters: {num_param}") + + # we need cut ids to display recognition results. + args.return_cuts = True + librispeech = LibriSpeechAsrDataModule(args) + + test_clean_cuts = librispeech.test_clean_cuts() + test_other_cuts = librispeech.test_other_cuts() + + test_clean_dl = librispeech.test_dataloaders(test_clean_cuts) + test_other_dl = librispeech.test_dataloaders(test_other_cuts) + + test_sets = ["test-clean", "test-other"] + test_dl = [test_clean_dl, test_other_dl] + + for test_set, test_dl in zip(test_sets, test_dl): + results_dict = decode_dataset( + dl=test_dl, + params=params, + model=model, + HLG=HLG, + H=H, + bpe_model=bpe_model, + word_table=lexicon.word_table, + G=G, + ) + + save_results( + params=params, + test_set_name=test_set, + results_dict=results_dict, + ) + + logging.info("Done!") + + +if __name__ == "__main__": + main() diff --git a/egs/tedlium3/ASR/pruned_transducer_stateless_d2v_v2/data2vec_audio.py b/egs/tedlium3/ASR/pruned_transducer_stateless_d2v_v2/data2vec_audio.py new file mode 100644 index 000000000..5521c56c8 --- /dev/null +++ b/egs/tedlium3/ASR/pruned_transducer_stateless_d2v_v2/data2vec_audio.py @@ -0,0 +1,766 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import argparse +import logging +import math +import os +from dataclasses import dataclass, field +from typing import Optional + +from omegaconf import II + +import numpy as np + +import torch +import torch.nn as nn +import torch.nn.functional as F +import torch.distributed as dist + +from fairseq.data.data_utils import compute_mask_indices +from fairseq.models import BaseFairseqModel, register_model +from fairseq.models.wav2vec import ( + ConvFeatureExtractionModel, + Wav2Vec2Config, + TransformerEncoder, +) +from fairseq.modules import ( + GradMultiply, + LayerNorm, +) +from fairseq.utils import index_put +from utils import pad_to_multiple + +from convolution import ConvolutionModule + +logger = logging.getLogger().setLevel(logging.INFO) + + +class TransformerEncoderAdapter(TransformerEncoder): + def __init__(self, args: Wav2Vec2Config): + super().__init__(args) + self.adapters = ResidualAdapterModule(proj_dim=512) + + for p in self.adapters.parameters(): + p.data /= 10. + #p.data = nn.Parameter(torch.zeros(p.size()).to('cuda')) + #p.data = nn.Parameter(torch.randn(p.size()).to('cuda')/20.) + + def forward(self, x, padding_mask=None, layer=None, tgt_layer=None): + x, layer_results = self.extract_features_with_adapter( + x, + padding_mask=padding_mask, + tgt_layer=tgt_layer + ) + + if self.layer_norm_first and layer is None: + x = self.layer_norm(x) + + return x, layer_results + + def extract_features_with_adapter( + self, + x, + padding_mask=None, + tgt_layer=None, + min_layer=0, + ): + + if padding_mask is not None: + x = index_put(x, padding_mask, 0) + + x_conv = self.pos_conv(x.transpose(1, 2)) + x_conv = x_conv.transpose(1, 2) + x = x + x_conv + + if not self.layer_norm_first: + x = self.layer_norm(x) + + # pad to the sequence length dimension + x, pad_length = pad_to_multiple( + x, self.required_seq_len_multiple, dim=-2, value=0 + ) + if pad_length > 0 and padding_mask is None: + padding_mask = x.new_zeros((x.size(0), x.size(1)), dtype=torch.bool) + padding_mask[:, -pad_length:] = True + else: + padding_mask, _ = pad_to_multiple( + padding_mask, self.required_seq_len_multiple, dim=-1, value=True + ) + x = F.dropout(x, p=self.dropout, training=self.training) + + # B x T x C -> T x B x C + x = x.transpose(0, 1) + + layer_results = [] + r = None + + for i, layer in enumerate(self.layers): + dropout_probability = np.random.random() if self.layerdrop > 0 else 1 + if not self.training or (dropout_probability > self.layerdrop): + x, (z, lr) = layer( + x, self_attn_padding_mask=padding_mask, need_weights=False, + ) + x = self.adapters(x, layer_id=i) + + if i >= min_layer: + layer_results.append((x, z, lr)) + + if i == tgt_layer: + r = x + break + + if r is not None: + x = r + + # T x B x C -> B x T x C + x = x.transpose(0, 1) + + # undo paddding + if pad_length > 0: + x = x[:, :-pad_length] + + def undo_pad(a, b, c): + return ( + a[:-pad_length], + b[:-pad_length] if b is not None else b, + c[:-pad_length], + ) + + layer_results = [undo_pad(*u) for u in layer_results] + + return x, layer_results + + +class LoRAModule(nn.Module): + def __init__( + self, + embedding_dim: float = 768, + rank: int = 16, + lora_alpha: int = 1, + lora_dropout: float = 0.1, + ) -> None: + + super().__init__() + self.r = rank + self.lora_alpha = lora_alpha + #Optional dropout + if lora_dropout > 0.: + self.lora_dropout = nn.Dropout(p=lora_dropout) + else: + self.lora_dropout = lambda x: x + + #self.lora_A = nn.ModuleList( + # [nn.Linear(embedding_dim, self.r) for _ in range(layer_num)]) + #self.lora_B = nn.ModuleList( + # [nn.Linear(self.r, embedding_dim) for _ in range(layer_num)]) + self.lora_A = nn.Linear(embedding_dim, self.r) + self.lora_B = nn.Linear(self.r, embedding_dim) + self.scaling = self.lora_alpha / self.r + self.reset_parameters() + + def reset_parameters(self): + nn.init.zeros_(self.lora_B.weight.data) + nn.init.zeros_(self.lora_B.bias.data) + #nn.init.normal_(self.lora_B.weight.data) + #nn.init.normal_(self.lora_B.bias.data) + + nn.init.normal_(self.lora_A.weight.data) + nn.init.normal_(self.lora_A.bias.data) + + def forward(self, x): + #x = x.transpose(0, 1) + x = self.lora_A(x) + x = self.lora_B(x) + #x = x.transpose(0, 1) + x *= self.scaling + return x + + +class ResidualAdapterModule(nn.Module): + """ + Implements a residual adapter based on https://arxiv.org/pdf/1909.08478.pdf + modules similar to the original residual adapter except layernorm location (first -> last) + """ + def __init__( + self, + embedding_dim: float = 768, + layer_num: int = 12, + proj_dim: float = 512, + ) -> None: + + super().__init__() + + self.type = 'linear' + + def build_adapter(embedding_dim, proj_dim, type_=self.type): + if type_ == 'conv': + return ConvolutionModule(768, 31) + else: + return nn.Sequential( + #nn.LayerNorm(embedding_dim), + nn.Linear(embedding_dim, proj_dim), + nn.ReLU(), + nn.Linear(proj_dim, embedding_dim), + nn.LayerNorm(embedding_dim), + ) + + self.adapter_layers = nn.ModuleList( + [build_adapter(embedding_dim, proj_dim, type_=self.type) for _ in range(layer_num)] + ) + + def forward(self, x, layer_id=-1): + x = x.transpose(0, 1) + residual = x + x = self.adapter_layers[layer_id](x) + x = residual + x + x = x.transpose(0, 1) + + return x + + +@dataclass +class Data2VecAudioConfig(Wav2Vec2Config): + + loss_beta: float = field( + default=0, metadata={"help": "beta for smooth l1 loss. 0 means use l2 loss"} + ) + loss_scale: Optional[float] = field( + default=None, + metadata={ + "help": "scale the reconstruction loss by this constant. if None then scales by 1/sqrt(dim)" + }, + ) + average_top_k_layers: int = field( + default=8, metadata={"help": "how many layers to average"} + ) + + layer_norm_target_layer: bool = False + instance_norm_target_layer: bool = False + instance_norm_targets: bool = False + layer_norm_targets: bool = False + batch_norm_target_layer: bool = False + group_norm_target_layer: bool = False + + ema_decay: float = field(default=0.999, metadata={"help": "initial ema decay rate"}) + ema_end_decay: float = field( + default=0.9999, metadata={"help": "final ema decay rate"} + ) + + # when to finish annealing ema decay rate + ema_anneal_end_step: int = II("optimization.max_update") + + ema_transformer_only: bool = field( + default=True, + metadata={"help": "whether to momentum update only the transformer"}, + ) + ema_layers_only: bool = field( + default=True, + metadata={"help": "whether to momentum update only the transformer layers"}, + ) + + max_update: int = II("optimization.max_update") + + min_target_var: float = field( + default=0.1, metadata={"help": "stop training if target var falls below this"} + ) + min_pred_var: float = field( + default=0.01, + metadata={"help": "stop training if prediction var falls below this"}, + ) + + +def get_annealed_rate(start, end, curr_step, total_steps): + r = end - start + pct_remaining = 1 - curr_step / total_steps + return end - r * pct_remaining + + +@register_model("data2vec_audio", dataclass=Data2VecAudioConfig) +class Data2VecAudioModel(BaseFairseqModel): + def __init__(self, cfg: Data2VecAudioConfig): + super().__init__() + self.cfg = cfg + + feature_enc_layers = eval(cfg.conv_feature_layers) + self.extractor_embed = feature_enc_layers[-1][0] + + self.ema = None + self.embed = cfg.encoder_embed_dim + + self.average_top_k_layers = cfg.average_top_k_layers + self.loss_beta = cfg.loss_beta + self.loss_scale = cfg.loss_scale + + self.feature_extractor = ConvFeatureExtractionModel( + conv_layers=feature_enc_layers, + dropout=0.0, + mode=cfg.extractor_mode, + conv_bias=cfg.conv_bias, + ) + + self.post_extract_proj = nn.Linear(self.extractor_embed, cfg.encoder_embed_dim) + + self.mask_prob = cfg.mask_prob + self.mask_selection = cfg.mask_selection + self.mask_other = cfg.mask_other + self.mask_length = cfg.mask_length + self.no_mask_overlap = cfg.no_mask_overlap + self.mask_min_space = cfg.mask_min_space + + self.mask_channel_prob = cfg.mask_channel_prob + self.mask_channel_before = cfg.mask_channel_before + self.mask_channel_selection = cfg.mask_channel_selection + self.mask_channel_other = cfg.mask_channel_other + self.mask_channel_length = cfg.mask_channel_length + self.no_mask_channel_overlap = cfg.no_mask_channel_overlap + self.mask_channel_min_space = cfg.mask_channel_min_space + + self.dropout_input = nn.Dropout(cfg.dropout_input) + self.dropout_features = nn.Dropout(cfg.dropout_features) + + self.feature_grad_mult = cfg.feature_grad_mult + + self.mask_emb = nn.Parameter( + torch.FloatTensor(cfg.encoder_embed_dim).uniform_() + ) + + self.encoder = TransformerEncoder(cfg) + #self.encoder = TransformerEncoderAdapter(cfg) + self.layer_norm = LayerNorm(self.extractor_embed) + + self.final_proj = nn.Linear(self.embed, self.embed) + + self.num_updates = 0 + + ''' + def make_ema_teacher(self): + ema_config = EMAModuleConfig( + ema_decay=self.cfg.ema_decay, + ema_fp32=True, + ) + skip_keys = set() + if self.cfg.ema_layers_only: + self.cfg.ema_transformer_only = True + for k, _ in self.encoder.pos_conv.named_parameters(): + skip_keys.add(f"pos_conv.{k}") + + self.ema = EMAModule( + self.encoder if self.cfg.ema_transformer_only else self, + ema_config, + skip_keys=skip_keys, + ) + ''' + + def set_num_updates(self, num_updates): + super().set_num_updates(num_updates) + + ''' + if self.ema is None and self.final_proj is not None: + logger.info(f"making ema teacher") + self.make_ema_teacher() + elif self.training and self.ema is not None: + if self.cfg.ema_decay != self.cfg.ema_end_decay: + if num_updates >= self.cfg.ema_anneal_end_step: + decay = self.cfg.ema_end_decay + else: + decay = get_annealed_rate( + self.cfg.ema_decay, + self.cfg.ema_end_decay, + num_updates, + self.cfg.ema_anneal_end_step, + ) + self.ema.set_decay(decay) + if self.ema.get_decay() < 1: + self.ema.step(self.encoder if self.cfg.ema_transformer_only else self) + ''' + self.num_updates = num_updates + + def state_dict(self, destination=None, prefix="", keep_vars=False): + state = super().state_dict(destination, prefix, keep_vars) + + if self.ema is not None: + state[prefix + "_ema"] = self.ema.fp32_params + + return state + + def _load_from_state_dict(self, state_dict, prefix, *args, **kwargs): + if self.ema is not None: + k = prefix + "_ema" + assert k in state_dict + self.ema.restore(state_dict[k], True) + del state_dict[k] + return super()._load_from_state_dict(state_dict, prefix, *args, **kwargs) + + @classmethod + def build_model(cls, cfg: Data2VecAudioConfig, task=None): + """Build a new model instance.""" + + return cls(cfg) + + def apply_mask( + self, + x, + padding_mask, + mask_indices=None, + mask_channel_indices=None, + ): + B, T, C = x.shape + + if self.mask_channel_prob > 0 and self.mask_channel_before: + mask_channel_indices = compute_mask_indices( + (B, C), + None, + self.mask_channel_prob, + self.mask_channel_length, + self.mask_channel_selection, + self.mask_channel_other, + no_overlap=self.no_mask_channel_overlap, + min_space=self.mask_channel_min_space, + ) + mask_channel_indices = ( + torch.from_numpy(mask_channel_indices) + .to(x.device) + .unsqueeze(1) + .expand(-1, T, -1) + ) + x[mask_channel_indices] = 0 + + if self.mask_prob > 0: + if mask_indices is None: + mask_indices = compute_mask_indices( + (B, T), + padding_mask, + self.mask_prob, + self.mask_length, + self.mask_selection, + self.mask_other, + min_masks=1, + no_overlap=self.no_mask_overlap, + min_space=self.mask_min_space, + require_same_masks=self.cfg.require_same_masks, + mask_dropout=self.cfg.mask_dropout, + ) + mask_indices = torch.from_numpy(mask_indices).to(x.device) + #x = index_put(x, mask_indices, self.mask_emb) + x = index_put(x, mask_indices, 0) + else: + mask_indices = None + + if self.mask_channel_prob > 0 and not self.mask_channel_before: + if mask_channel_indices is None: + mask_channel_indices = compute_mask_indices( + (B, C), + None, + self.mask_channel_prob, + self.mask_channel_length, + self.mask_channel_selection, + self.mask_channel_other, + no_overlap=self.no_mask_channel_overlap, + min_space=self.mask_channel_min_space, + ) + mask_channel_indices = ( + torch.from_numpy(mask_channel_indices) + .to(x.device) + .unsqueeze(1) + .expand(-1, T, -1) + ) + x = index_put(x, mask_channel_indices, 0) + + return x, mask_indices + + def _get_feat_extract_output_lengths(self, input_lengths: torch.LongTensor): + """ + Computes the output length of the convolutional layers + """ + + def _conv_out_length(input_length, kernel_size, stride): + return torch.floor((input_length - kernel_size) / stride + 1) + + conv_cfg_list = eval(self.cfg.conv_feature_layers) + + for i in range(len(conv_cfg_list)): + input_lengths = _conv_out_length( + input_lengths, conv_cfg_list[i][1], conv_cfg_list[i][2] + ) + + return input_lengths.to(torch.long) + + def forward( + self, + source, + padding_mask=None, + mask=True, + features_only=False, + layer=None, + mask_indices=None, + mask_channel_indices=None, + padding_count=None, + prompt=None, + sid=None, + ): + features = source + + if self.feature_grad_mult > 0: + features = self.feature_extractor(features) + if self.feature_grad_mult != 1.0: + features = GradMultiply.apply(features, self.feature_grad_mult) + else: + with torch.no_grad(): + features = self.feature_extractor(features) + + features = features.transpose(1, 2) + + orig_padding_mask = padding_mask + + if padding_mask is not None and padding_mask.any(): + input_lengths = (1 - padding_mask.long()).sum(-1) + # apply conv formula to get real output_lengths + output_lengths = self._get_feat_extract_output_lengths(input_lengths) + + padding_mask = torch.zeros( + features.shape[:2], dtype=features.dtype, device=features.device + ) + + # these two operations makes sure that all values + # before the output lengths indices are attended to + padding_mask[ + ( + torch.arange(padding_mask.shape[0], device=padding_mask.device), + output_lengths - 1, + ) + ] = 1 + padding_mask = (1 - padding_mask.flip([-1]).cumsum(-1).flip([-1])).bool() + else: + padding_mask = None + + ## for prompt tuning + if prompt is not None: + if 0: + spk_dir = f'/home/work/workspace/icefall/egs/librispeech/ASR/conv_feat/{sid}' + if not os.path.isdir(spk_dir): + os.mkdir(spk_dir) + + conv_feat_all = torch.tensor([]).to(features.device) + length = 0 + for i in range(padding_mask.size()[0]): + nonzero = padding_mask[i].nonzero() + try: + length += nonzero[0] + conv_feat_all = torch.cat([conv_feat_all, features[i, :nonzero[0], :]]) + except: + length += features.size()[1] + conv_feat_all = torch.cat([conv_feat_all, features[i]]) + + randint = np.random.randint(10000000) + np.save(f'{spk_dir}/{randint}.npy', conv_feat_all.cpu().numpy()) + + prompt = prompt.expand((features.size()[0], prompt.size()[0], prompt.size()[1])) + features = torch.cat([prompt, features], dim=1) + prompt_padding_mask = torch.zeros(prompt.size()[0], prompt.size()[1]).type(torch.BoolTensor).to(features.device) + try: padding_mask = torch.cat([prompt_padding_mask, padding_mask], dim=1) + except: padding_mask = None + + features = self.layer_norm(features) + + #print(padding_mask.size()) + #print((padding_mask[0] == True).nonzero(as_tuple=True)[0]) + #print((padding_mask[1] == True).nonzero(as_tuple=True)[0][1]) + #print((padding_mask[2] == True).nonzero(as_tuple=True)[0][2]) + #print((padding_mask[3] == True).nonzero(as_tuple=True)[0][3]) + #exit() + + if self.post_extract_proj is not None: + features = self.post_extract_proj(features) + + pre_encoder_features = None + if self.cfg.ema_transformer_only: + pre_encoder_features = features.clone() + + features = self.dropout_input(features) + + if mask: + x, mask_indices = self.apply_mask( + features, + padding_mask, + mask_indices=mask_indices, + mask_channel_indices=mask_channel_indices, + ) + else: + x = features + mask_indices = None + + x, layer_results = self.encoder( + x, + padding_mask=padding_mask, + layer=layer, + ) + + if features_only: + return { + "x": x, + "padding_mask": padding_mask, + "layer_results": layer_results, + } + + result = { + "losses": {}, + } + + with torch.no_grad(): + self.ema.model.eval() + + if self.cfg.ema_transformer_only: + y, layer_results = self.ema.model.extract_features( + pre_encoder_features, + padding_mask=padding_mask, + min_layer=self.cfg.encoder_layers - self.average_top_k_layers, + ) + y = { + "x": y, + "padding_mask": padding_mask, + "layer_results": layer_results, + } + else: + y = self.ema.model.extract_features( + source=source, + padding_mask=orig_padding_mask, + mask=False, + ) + + target_layer_results = [l[2] for l in y["layer_results"]] + + permuted = False + if self.cfg.instance_norm_target_layer or self.cfg.batch_norm_target_layer: + target_layer_results = [ + tl.permute(1, 2, 0) for tl in target_layer_results # TBC -> BCT + ] + permuted = True + + if self.cfg.batch_norm_target_layer: + target_layer_results = [ + F.batch_norm( + tl.float(), running_mean=None, running_var=None, training=True + ) + for tl in target_layer_results + ] + + if self.cfg.instance_norm_target_layer: + target_layer_results = [ + F.instance_norm(tl.float()) for tl in target_layer_results + ] + + if permuted: + target_layer_results = [ + tl.transpose(1, 2) for tl in target_layer_results # BCT -> BTC + ] + + if self.cfg.group_norm_target_layer: + target_layer_results = [ + F.layer_norm(tl.float(), tl.shape[-2:]) + for tl in target_layer_results + ] + + if self.cfg.layer_norm_target_layer: + target_layer_results = [ + F.layer_norm(tl.float(), tl.shape[-1:]) + for tl in target_layer_results + ] + + y = sum(target_layer_results) / len(target_layer_results) + + if self.cfg.layer_norm_targets: + y = F.layer_norm(y.float(), y.shape[-1:]) + + if self.cfg.instance_norm_targets: + y = F.instance_norm(y.float().transpose(1, 2)).transpose(1, 2) + + if not permuted: + y = y.transpose(0, 1) + + y = y[mask_indices] + + x = x[mask_indices] + x = self.final_proj(x) + + sz = x.size(-1) + + if self.loss_beta == 0: + loss = F.mse_loss(x.float(), y.float(), reduction="none").sum(dim=-1) + else: + loss = F.smooth_l1_loss( + x.float(), y.float(), reduction="none", beta=self.loss_beta + ).sum(dim=-1) + + if self.loss_scale is not None: + scale = self.loss_scale + else: + scale = 1 / math.sqrt(sz) + + result["losses"]["regression"] = loss.sum() * scale + + if "sample_size" not in result: + result["sample_size"] = loss.numel() + + with torch.no_grad(): + result["target_var"] = self.compute_var(y) + result["pred_var"] = self.compute_var(x.float()) + + if self.num_updates > 5000 and result["target_var"] < self.cfg.min_target_var: + logger.error( + f"target var is {result['target_var'].item()} < {self.cfg.min_target_var}, exiting" + ) + raise Exception( + f"target var is {result['target_var'].item()} < {self.cfg.min_target_var}, exiting" + ) + if self.num_updates > 5000 and result["pred_var"] < self.cfg.min_pred_var: + logger.error( + f"pred var is {result['pred_var'].item()} < {self.cfg.min_pred_var}, exiting" + ) + raise Exception( + f"pred var is {result['pred_var'].item()} < {self.cfg.min_pred_var}, exiting" + ) + + if self.ema is not None: + result["ema_decay"] = self.ema.get_decay() * 1000 + + return result + + @staticmethod + def compute_var(y): + y = y.view(-1, y.size(-1)) + if dist.is_initialized(): + zc = torch.tensor(y.size(0)).cuda() + zs = y.sum(dim=0) + zss = (y ** 2).sum(dim=0) + + dist.all_reduce(zc) + dist.all_reduce(zs) + dist.all_reduce(zss) + + var = zss / (zc - 1) - (zs ** 2) / (zc * (zc - 1)) + return torch.sqrt(var + 1e-6).mean() + else: + return torch.sqrt(y.var(dim=0) + 1e-6).mean() + + def extract_features( + self, source, padding_mask, mask=False, layer=None + ): + res = self.forward( + source, + padding_mask, + mask=mask, + features_only=True, + layer=layer, + ) + return res + + def remove_pretraining_modules(self, last_layer=None): + self.final_proj = None + self.ema = None + if last_layer is not None: + self.encoder.layers = nn.ModuleList( + l for i, l in enumerate(self.encoder.layers) if i <= last_layer + ) diff --git a/egs/tedlium3/ASR/pruned_transducer_stateless_d2v_v2/data2vec_encoder.py b/egs/tedlium3/ASR/pruned_transducer_stateless_d2v_v2/data2vec_encoder.py new file mode 100644 index 000000000..cf9e0995f --- /dev/null +++ b/egs/tedlium3/ASR/pruned_transducer_stateless_d2v_v2/data2vec_encoder.py @@ -0,0 +1,184 @@ +# Copyright 2021 Xuankai Chang +# Apache 2.0 (http://www.apache.org/licenses/LICENSE-2.0) + +"""Encoder definition.""" +import contextlib +import time +import copy +import math +import logging +import os +from typing import List, Optional, Tuple +import warnings + +import torch +from filelock import FileLock +from typeguard import check_argument_types + +from nets_utils import make_pad_mask +from encoder_interface import EncoderInterface +from torch import Tensor, nn + +from icefall.utils import make_pad_mask, subsequent_chunk_mask +try: + import fairseq + from data2vec_audio import * +except Exception as e: + print("Error: FairSeq is not properly installed.") + print( + "Please install FairSeq: cd ${MAIN_ROOT}/tools && make fairseq.done" + ) + raise e + + +class FairSeqData2VecEncoder(EncoderInterface): + """FairSeq Wav2Vec2 encoder module. + + Args: + input_size: input dim + output_size: dimension of attention + w2v_url: url to Wav2Vec2.0 pretrained model + w2v_dir_path: directory to download the Wav2Vec2.0 pretrained model. + normalize_before: whether to use layer_norm before the first block + finetune_last_n_layers: last n layers to be finetuned in Wav2Vec2.0 + 0 means to finetune every layer if freeze_w2v=False. + """ + + def __init__( + self, + input_size: int, + w2v_url: str, + w2v_dir_path: str = "./", + output_size: int = 256, + freeze_finetune_updates: int = 0, + additional_block: bool = False, + ): + assert check_argument_types() + super().__init__() + + self.w2v_model_path = download_d2v() + self._output_size = output_size + + models, _, _ = fairseq.checkpoint_utils.load_model_ensemble_and_task( + [self.w2v_model_path], + strict=False, + ) + model = models[0] + model.feature_grad_mult = 0.0 ## for conv network freeze + ## prevent overfitting + model.mask_prob = 0.65 + model.mask_channel_prob = 0.5 + model.mask_channel_length = 64 + model.activation_dropout = 0.1 + + self.encoders = model + self.pretrained_params = copy.deepcopy(model.state_dict()) + + if model.cfg.encoder_embed_dim != output_size or additional_block: + # TODO(xkc09): try LSTM + self.output_layer = torch.nn.Sequential( + torch.nn.Linear(model.cfg.encoder_embed_dim, output_size), + torch.nn.LayerNorm(output_size), + torch.nn.GELU(), + ) + else: + self.output_layer = None + + self.freeze_finetune_updates = freeze_finetune_updates + self.num_updates = 0 + + def output_size(self) -> int: + return self._output_size + + def forward( + self, + x: torch.Tensor, + x_lens: torch.Tensor, + warmup = None, + prev_states: torch.Tensor = None, + prompt = None, + sid = None, + ) -> Tuple[torch.Tensor, torch.Tensor]: + xs_pad = x + ilens = x_lens + """Forward FairSeqWav2Vec2 Encoder. + + Args: + xs_pad: input tensor (B, L, D) + ilens: input length (B) + prev_states: Not to be used now. + Returns: + position embedded tensor and mask + """ + with torch.no_grad(): + xs_pad = torch.nn.functional.layer_norm(xs_pad, xs_pad.shape) + + masks = make_pad_mask(ilens).to(xs_pad.device) + + ft = (self.freeze_finetune_updates <= self.num_updates) and self.encoders.training + if self.num_updates <= self.freeze_finetune_updates: + self.num_updates += 1 + elif ft and self.num_updates == self.freeze_finetune_updates + 1: + self.num_updates += 1 + logging.info("Start fine-tuning data2vec parameters!") + + with torch.no_grad() if not ft else contextlib.nullcontext(): + enc_outputs = self.encoders( + xs_pad, + masks, + mask = ft, + features_only=True, + prompt=prompt, + sid=sid, + ) + + xs_pad = enc_outputs["x"] # (B,T,C), + bs = xs_pad.shape[0] + if enc_outputs["padding_mask"] is not None: + masks = enc_outputs["padding_mask"] # (B, T) + olens = (~masks).sum(dim=1) # (B) + else: + olens = torch.IntTensor([xs_pad.shape[1]]).repeat(bs).to(xs_pad.device) + + if self.output_layer is not None: + xs_pad = self.output_layer(xs_pad) + + return xs_pad, olens + + def reload_pretrained_parameters(self): + self.encoders.load_state_dict(self.pretrained_params) + logging.info("Pretrained data2vec model parameters reloaded!") + + +def download_d2v(model_url='https://dl.fbaipublicfiles.com/fairseq/data2vec/audio_base_ls.pt', dir_path='./models'): + os.makedirs(dir_path, exist_ok=True) + + model_name = model_url.split("/")[-1] + model_path = os.path.join(dir_path, model_name) + + dict_url = "https://dl.fbaipublicfiles.com/fairseq/wav2vec/dict.ltr.txt" + dict_path = os.path.join(dir_path, dict_url.split("/")[-1]) + + with FileLock(model_path + ".lock"): + if not os.path.exists(model_path): + torch.hub.download_url_to_file(model_url, model_path) + torch.hub.download_url_to_file(dict_url, dict_path) + logging.info(f"data2vec model downloaded {model_path}") + else: + logging.info(f"data2vec model {model_path} already exists.") + + return model_path + + +if __name__ == '__main__': + d2v = FairSeqData2VecEncoder(input_size=768, w2v_url='ww', output_size=768) + inputs = torch.randn([1, 211564]) + #a = torch.ones([1000] + #b = torch.ones([10000]) + #c = torch.ones([10000]) + length = torch.tensor([211564]) + outputs = d2v(inputs, length) + print(outputs[0].size()) + + #for n, p in d2v.named_parameters(): + # print(n) diff --git a/egs/tedlium3/ASR/pruned_transducer_stateless_d2v_v2/decode.py b/egs/tedlium3/ASR/pruned_transducer_stateless_d2v_v2/decode.py new file mode 100755 index 000000000..9ccc99a41 --- /dev/null +++ b/egs/tedlium3/ASR/pruned_transducer_stateless_d2v_v2/decode.py @@ -0,0 +1,887 @@ +#!/usr/bin/env python3 +# +# Copyright 2021-2022 Xiaomi Corporation (Author: Fangjun Kuang, +# Zengwei Yao) +# +# See ../../../../LICENSE for clarification regarding multiple authors +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +""" +Usage: +(0) for d2v-T decoding +for method in greedy_search modified_beam_search fast_beam_search; do + ./pruned_transducer_stateless_d2v_v2/decode.py \ + --input-strategy AudioSamples \ + --enable-spec-aug False \ + --additional-block True \ + --model-name epoc.pt \ + --exp-dir ./pruned_transducer_stateless_d2v_v2/960h_sweep_v3_388 \ + --max-duration 400 \ + --decoding-method $method \ + --max-sym-per-frame 1 \ + --encoder-type d2v \ + --encoder-dim 768 \ + --decoder-dim 768 \ + --joiner-dim 768 +done +""" + + +import argparse +import logging +import math +from collections import defaultdict +from pathlib import Path +from typing import Dict, List, Optional, Tuple + +import k2 +import sentencepiece as spm +import torch +import torch.nn as nn +from asr_datamodule import LibriSpeechAsrDataModule +from beam_search import ( + beam_search, + fast_beam_search_nbest, + fast_beam_search_nbest_LG, + fast_beam_search_nbest_oracle, + fast_beam_search_one_best, + greedy_search, + greedy_search_batch, + modified_beam_search, +) +#from train import add_model_arguments, add_rep_arguments, get_params, get_transducer_model +from prompt_tuning import add_model_arguments, add_rep_arguments, get_params, get_transducer_model + +from icefall.checkpoint import ( + average_checkpoints, + average_checkpoints_with_averaged_model, + find_checkpoints, + load_checkpoint, +) +from icefall.lexicon import Lexicon +from icefall.utils import ( + AttributeDict, + setup_logger, + store_transcripts, + str2bool, + write_error_stats, +) + +import fairseq +from data2vec_audio import LoRAModule + +LOG_EPS = math.log(1e-10) + +class LoRAHook(): + def __init__(self, module): + self.hook = module.register_forward_hook(self.hook_fn) + self.lora = LoRAModule( + embedding_dim=768, + rank=6, + lora_alpha=10000, + ) + def hook_fn(self, module, input, output): + lora_out = self.lora(input[0]) + output += lora_out + + def save_checkpoint(self, i, iter_, save_dir): + if isinstance(self.lora, DDP): + lora = self.lora.module + torch.save(lora.state_dict(), f"{save_dir}/lora_{iter_}_{i}.pt") + + +def get_parser(): + parser = argparse.ArgumentParser( + formatter_class=argparse.ArgumentDefaultsHelpFormatter + ) + parser.add_argument( + "--model-name", + type=str, + default="", + help="""It specifies the model file name to use for decoding.""", + ) + + parser.add_argument( + "--epoch", + type=int, + default=30, + help="""It specifies the checkpoint to use for decoding. + Note: Epoch counts from 1. + You can specify --avg to use more checkpoints for model averaging.""", + ) + + parser.add_argument( + "--iter", + type=int, + default=0, + help="""If positive, --epoch is ignored and it + will use the checkpoint exp_dir/checkpoint-iter.pt. + You can specify --avg to use more checkpoints for model averaging. + """, + ) + + parser.add_argument( + "--avg", + type=int, + default=9, + help="Number of checkpoints to average. Automatically select " + "consecutive checkpoints before the checkpoint specified by " + "'--epoch' and '--iter'", + ) + + parser.add_argument( + "--use-averaged-model", + type=str2bool, + default=True, + help="Whether to load averaged model. Currently it only supports " + "using --epoch. If True, it would decode with the averaged model " + "over the epoch range from `epoch-avg` (excluded) to `epoch`." + "Actually only the models with epoch number of `epoch-avg` and " + "`epoch` are loaded for averaging. ", + ) + + parser.add_argument( + "--exp-dir", + type=str, + default="pruned_transducer_stateless7_ctc/exp", + help="The experiment dir", + ) + + parser.add_argument( + "--bpe-model", + type=str, + default="data/lang_bpe_500/bpe.model", + help="Path to the BPE model", + ) + + parser.add_argument( + "--lang-dir", + type=Path, + default="data/lang_bpe_500", + help="The lang dir containing word table and LG graph", + ) + + parser.add_argument( + "--decoding-method", + type=str, + default="greedy_search", + help="""Possible values are: + - greedy_search + - beam_search + - modified_beam_search + - fast_beam_search + - fast_beam_search_nbest + - fast_beam_search_nbest_oracle + - fast_beam_search_nbest_LG + If you use fast_beam_search_nbest_LG, you have to specify + `--lang-dir`, which should contain `LG.pt`. + """, + ) + + parser.add_argument( + "--beam-size", + type=int, + default=4, + help="""An integer indicating how many candidates we will keep for each + frame. Used only when --decoding-method is beam_search or + modified_beam_search.""", + ) + + parser.add_argument( + "--beam", + type=float, + default=20.0, + help="""A floating point value to calculate the cutoff score during beam + search (i.e., `cutoff = max-score - beam`), which is the same as the + `beam` in Kaldi. + Used only when --decoding-method is fast_beam_search, + fast_beam_search_nbest, fast_beam_search_nbest_LG, + and fast_beam_search_nbest_oracle + """, + ) + + parser.add_argument( + "--ngram-lm-scale", + type=float, + default=0.01, + help=""" + Used only when --decoding_method is fast_beam_search_nbest_LG. + It specifies the scale for n-gram LM scores. + """, + ) + + parser.add_argument( + "--max-contexts", + type=int, + default=8, + help="""Used only when --decoding-method is + fast_beam_search, fast_beam_search_nbest, fast_beam_search_nbest_LG, + and fast_beam_search_nbest_oracle""", + ) + + parser.add_argument( + "--max-states", + type=int, + default=64, + help="""Used only when --decoding-method is + fast_beam_search, fast_beam_search_nbest, fast_beam_search_nbest_LG, + and fast_beam_search_nbest_oracle""", + ) + + parser.add_argument( + "--context-size", + type=int, + default=2, + help="The context size in the decoder. 1 means bigram; 2 means tri-gram", + ) + parser.add_argument( + "--max-sym-per-frame", + type=int, + default=1, + help="""Maximum number of symbols per frame. + Used only when --decoding_method is greedy_search""", + ) + + parser.add_argument( + "--num-paths", + type=int, + default=200, + help="""Number of paths for nbest decoding. + Used only when the decoding method is fast_beam_search_nbest, + fast_beam_search_nbest_LG, and fast_beam_search_nbest_oracle""", + ) + + parser.add_argument( + "--nbest-scale", + type=float, + default=0.5, + help="""Scale applied to lattice scores when computing nbest paths. + Used only when the decoding method is fast_beam_search_nbest, + fast_beam_search_nbest_LG, and fast_beam_search_nbest_oracle""", + ) + + parser.add_argument( + "--simulate-streaming", + type=str2bool, + default=False, + help="""Whether to simulate streaming in decoding, this is a good way to + test a streaming model. + """, + ) + + parser.add_argument( + "--decode-chunk-size", + type=int, + default=16, + help="The chunk size for decoding (in frames after subsampling)", + ) + + parser.add_argument( + "--left-context", + type=int, + default=64, + help="left context can be seen during decoding (in frames after subsampling)", + ) + + parser.add_argument( + "--res-name", + type=str, + ) + + add_model_arguments(parser) + add_rep_arguments(parser) + + return parser + + +def decode_one_batch( + params: AttributeDict, + model: nn.Module, + sp: spm.SentencePieceProcessor, + batch: dict, + word_table: Optional[k2.SymbolTable] = None, + decoding_graph: Optional[k2.Fsa] = None, +) -> Dict[str, List[List[str]]]: + """Decode one batch and return the result in a dict. The dict has the + following format: + + - key: It indicates the setting used for decoding. For example, + if greedy_search is used, it would be "greedy_search" + If beam search with a beam size of 7 is used, it would be + "beam_7" + - value: It contains the decoding result. `len(value)` equals to + batch size. `value[i]` is the decoding result for the i-th + utterance in the given batch. + Args: + params: + It's the return value of :func:`get_params`. + model: + The neural model. + sp: + The BPE model. + batch: + It is the return value from iterating + `lhotse.dataset.K2SpeechRecognitionDataset`. See its documentation + for the format of the `batch`. + word_table: + The word symbol table. + decoding_graph: + The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used + only when --decoding_method is fast_beam_search, fast_beam_search_nbest, + fast_beam_search_nbest_oracle, and fast_beam_search_nbest_LG. + Returns: + Return the decoding result. See above description for the format of + the returned dict. + """ + device = next(model.parameters()).device + feature = batch["inputs"] + assert feature.ndim == 2 or feature.ndim == 3 + + feature = feature.to(device) + # at entry, feature is (N, T, C) + + supervisions = batch["supervisions"] + #feature_lens = supervisions["num_frames"].to(device) + if feature.ndim == 2: + feature_lens = [] + for supervision in supervisions['cut']: + try: feature_lens.append(supervision.tracks[0].cut.recording.num_samples) + except: feature_lens.append(supervision.recording.num_samples) + feature_lens = torch.tensor(feature_lens) + + elif feature.ndim == 3: + feature_lens = supervisions["num_frames"].to(device) + + if params.simulate_streaming: + feature_lens += params.left_context + feature = torch.nn.functional.pad( + feature, + pad=(0, 0, 0, params.left_context), + value=LOG_EPS, + ) + encoder_out, encoder_out_lens, _ = model.encoder.streaming_forward( + x=feature, + x_lens=feature_lens, + chunk_size=params.decode_chunk_size, + left_context=params.left_context, + simulate_streaming=True, + ) + else: + encoder_out, encoder_out_lens = model.encoder(x=feature, x_lens=feature_lens, prompt=model.prompt) + + if 0: + encoder_out = encoder_out[:,50:,:] + encoder_out_lens -= 50 + + hyps = [] + + if params.decoding_method == "fast_beam_search": + hyp_tokens = fast_beam_search_one_best( + model=model, + decoding_graph=decoding_graph, + encoder_out=encoder_out, + encoder_out_lens=encoder_out_lens, + beam=params.beam, + max_contexts=params.max_contexts, + max_states=params.max_states, + ) + for hyp in sp.decode(hyp_tokens): + hyps.append(hyp.split()) + elif params.decoding_method == "fast_beam_search_nbest_LG": + hyp_tokens = fast_beam_search_nbest_LG( + model=model, + decoding_graph=decoding_graph, + encoder_out=encoder_out, + encoder_out_lens=encoder_out_lens, + beam=params.beam, + max_contexts=params.max_contexts, + max_states=params.max_states, + num_paths=params.num_paths, + nbest_scale=params.nbest_scale, + ) + for hyp in hyp_tokens: + hyps.append([word_table[i] for i in hyp]) + elif params.decoding_method == "fast_beam_search_nbest": + hyp_tokens = fast_beam_search_nbest( + model=model, + decoding_graph=decoding_graph, + encoder_out=encoder_out, + encoder_out_lens=encoder_out_lens, + beam=params.beam, + max_contexts=params.max_contexts, + max_states=params.max_states, + num_paths=params.num_paths, + nbest_scale=params.nbest_scale, + ) + for hyp in sp.decode(hyp_tokens): + hyps.append(hyp.split()) + elif params.decoding_method == "fast_beam_search_nbest_oracle": + hyp_tokens = fast_beam_search_nbest_oracle( + model=model, + decoding_graph=decoding_graph, + encoder_out=encoder_out, + encoder_out_lens=encoder_out_lens, + beam=params.beam, + max_contexts=params.max_contexts, + max_states=params.max_states, + num_paths=params.num_paths, + ref_texts=sp.encode(supervisions["text"]), + nbest_scale=params.nbest_scale, + ) + for hyp in sp.decode(hyp_tokens): + hyps.append(hyp.split()) + elif params.decoding_method == "greedy_search" and params.max_sym_per_frame == 1: + hyp_tokens = greedy_search_batch( + model=model, + encoder_out=encoder_out, + encoder_out_lens=encoder_out_lens, + ) + for hyp in sp.decode(hyp_tokens): + hyps.append(hyp.split()) + elif params.decoding_method == "modified_beam_search": + hyp_tokens = modified_beam_search( + model=model, + encoder_out=encoder_out, + encoder_out_lens=encoder_out_lens, + beam=params.beam_size, + ) + for hyp in sp.decode(hyp_tokens): + hyps.append(hyp.split()) + else: + batch_size = encoder_out.size(0) + + for i in range(batch_size): + # fmt: off + encoder_out_i = encoder_out[i:i+1, :encoder_out_lens[i]] + # fmt: on + if params.decoding_method == "greedy_search": + hyp = greedy_search( + model=model, + encoder_out=encoder_out_i, + max_sym_per_frame=params.max_sym_per_frame, + ) + elif params.decoding_method == "beam_search": + hyp = beam_search( + model=model, + encoder_out=encoder_out_i, + beam=params.beam_size, + ) + else: + raise ValueError( + f"Unsupported decoding method: {params.decoding_method}" + ) + hyps.append(sp.decode(hyp).split()) + + if params.decoding_method == "greedy_search": + return {"greedy_search": hyps} + elif "fast_beam_search" in params.decoding_method: + key = f"beam_{params.beam}_" + key += f"max_contexts_{params.max_contexts}_" + key += f"max_states_{params.max_states}" + if "nbest" in params.decoding_method: + key += f"_num_paths_{params.num_paths}_" + key += f"nbest_scale_{params.nbest_scale}" + if "LG" in params.decoding_method: + key += f"_ngram_lm_scale_{params.ngram_lm_scale}" + + return {key: hyps} + else: + return {f"beam_size_{params.beam_size}": hyps} + + +def decode_dataset( + dl: torch.utils.data.DataLoader, + params: AttributeDict, + model: nn.Module, + sp: spm.SentencePieceProcessor, + word_table: Optional[k2.SymbolTable] = None, + decoding_graph: Optional[k2.Fsa] = None, +) -> Dict[str, List[Tuple[str, List[str], List[str]]]]: + """Decode dataset. + + Args: + dl: + PyTorch's dataloader containing the dataset to decode. + params: + It is returned by :func:`get_params`. + model: + The neural model. + sp: + The BPE model. + word_table: + The word symbol table. + decoding_graph: + The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used + only when --decoding_method is fast_beam_search, fast_beam_search_nbest, + fast_beam_search_nbest_oracle, and fast_beam_search_nbest_LG. + Returns: + Return a dict, whose key may be "greedy_search" if greedy search + is used, or it may be "beam_7" if beam size of 7 is used. + Its value is a list of tuples. Each tuple contains two elements: + The first is the reference transcript, and the second is the + predicted result. + """ + num_cuts = 0 + + try: + num_batches = len(dl) + except TypeError: + num_batches = "?" + + if params.decoding_method == "greedy_search": + log_interval = 50 + else: + log_interval = 20 + + results = defaultdict(list) + for batch_idx, batch in enumerate(dl): + texts = batch["supervisions"]["text"] + + cut_ids = [cut.id for cut in batch["supervisions"]["cut"]] + + hyps_dict = decode_one_batch( + params=params, + model=model, + sp=sp, + decoding_graph=decoding_graph, + word_table=word_table, + batch=batch, + ) + + for name, hyps in hyps_dict.items(): + this_batch = [] + assert len(hyps) == len(texts) + for cut_id, hyp_words, ref_text in zip(cut_ids, hyps, texts): + ref_words = ref_text.split() + this_batch.append((cut_id, ref_words, hyp_words)) + + results[name].extend(this_batch) + + num_cuts += len(texts) + + if batch_idx % log_interval == 0: + batch_str = f"{batch_idx}/{num_batches}" + + logging.info(f"batch {batch_str}, cuts processed until now is {num_cuts}") + return results + + +def save_results( + params: AttributeDict, + test_set_name: str, + results_dict: Dict[str, List[Tuple[str, List[str], List[str]]]], +): + test_set_wers = dict() + for key, results in results_dict.items(): + recog_path = ( + params.res_dir / f"recogs-{test_set_name}-{key}-{params.suffix}.txt" + ) + results = sorted(results) + store_transcripts(filename=recog_path, texts=results) + logging.info(f"The transcripts are stored in {recog_path}") + + # The following prints out WERs, per-word error statistics and aligned + # ref/hyp pairs. + errs_filename = ( + params.res_dir / f"errs-{test_set_name}-{key}-{params.suffix}.txt" + ) + with open(errs_filename, "w") as f: + wer = write_error_stats( + f, f"{test_set_name}-{key}", results, enable_log=True + ) + test_set_wers[key] = wer + + logging.info("Wrote detailed error stats to {}".format(errs_filename)) + + test_set_wers = sorted(test_set_wers.items(), key=lambda x: x[1]) + errs_info = ( + params.res_dir / f"wer-summary-{test_set_name}-{key}-{params.suffix}.txt" + ) + with open(errs_info, "w") as f: + print("settings\tWER", file=f) + for key, val in test_set_wers: + print("{}\t{}".format(key, val), file=f) + + spk = None + wer = None + s = "\nFor {}, WER of different settings are:\n".format(test_set_name) + note = "\tbest for {}".format(test_set_name) + for key, val in test_set_wers: + s += "{}\t{}{}\n".format(key, val, note) + note = "" + spk = str(test_set_name) + wer = str(val) + logging.info(s) + with open(f'./{params.res_name}.txt', 'a') as f: + f.write(f"{spk} {wer}\n") + + +@torch.no_grad() +def main(): + parser = get_parser() + LibriSpeechAsrDataModule.add_arguments(parser) + args = parser.parse_args() + args.exp_dir = Path(args.exp_dir) + + params = get_params() + params.update(vars(args)) + + assert params.decoding_method in ( + "greedy_search", + "beam_search", + "fast_beam_search", + "fast_beam_search_nbest", + "fast_beam_search_nbest_LG", + "fast_beam_search_nbest_oracle", + "modified_beam_search", + ) + params.res_dir = params.exp_dir / params.decoding_method + + if params.iter > 0: + params.suffix = f"iter-{params.iter}-avg-{params.avg}" + else: + params.suffix = f"epoch-{params.epoch}-avg-{params.avg}" + + if params.simulate_streaming: + params.suffix += f"-streaming-chunk-size-{params.decode_chunk_size}" + params.suffix += f"-left-context-{params.left_context}" + + if "fast_beam_search" in params.decoding_method: + params.suffix += f"-beam-{params.beam}" + params.suffix += f"-max-contexts-{params.max_contexts}" + params.suffix += f"-max-states-{params.max_states}" + if "nbest" in params.decoding_method: + params.suffix += f"-nbest-scale-{params.nbest_scale}" + params.suffix += f"-num-paths-{params.num_paths}" + if "LG" in params.decoding_method: + params.suffix += f"-ngram-lm-scale-{params.ngram_lm_scale}" + elif "beam_search" in params.decoding_method: + params.suffix += f"-{params.decoding_method}-beam-size-{params.beam_size}" + else: + params.suffix += f"-context-{params.context_size}" + params.suffix += f"-max-sym-per-frame-{params.max_sym_per_frame}" + + if params.use_averaged_model: + params.suffix += "-use-averaged-model" + + setup_logger(f"{params.res_dir}/log-decode-{params.suffix}") + logging.info("Decoding started") + + device = torch.device("cpu") + if torch.cuda.is_available(): + device = torch.device("cuda", 0) + + logging.info(f"Device: {device}") + + sp = spm.SentencePieceProcessor() + sp.load(params.bpe_model) + + # and are defined in local/train_bpe_model.py + params.blank_id = sp.piece_to_id("") + params.unk_id = sp.piece_to_id("") + params.vocab_size = sp.get_piece_size() + + if params.simulate_streaming: + assert ( + params.causal_convolution + ), "Decoding in streaming requires causal convolution" + + logging.info(params) + + logging.info("About to create model") + model = get_transducer_model(params) + + if '.pt' in params.model_name: + load_checkpoint(f"{params.exp_dir}/{params.model_name}", model) + + elif 'lora' in params.model_name: + load_checkpoint(f"{params.exp_dir}/../d2v-base-T.pt", model) + + ## for lora hooking + lora_modules = [] + for modules in model.modules(): + if isinstance(modules, fairseq.modules.multihead_attention.MultiheadAttention): + for module in modules.modules(): + if isinstance(module, torch.nn.Linear): + lora_modules.append(LoRAHook(module)) + + for i, lora in enumerate(lora_modules): + lora_param = torch.load(f"{params.exp_dir}/lora_{params.iter}_{i}.pt") + lora.lora.load_state_dict(lora_param) + lora.lora.to(device) + logging.info("lora params load done") + else: + if not params.use_averaged_model: + if params.iter > 0: + filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[ + : params.avg + ] + if len(filenames) == 0: + raise ValueError( + f"No checkpoints found for" + f" --iter {params.iter}, --avg {params.avg}" + ) + elif len(filenames) < params.avg: + raise ValueError( + f"Not enough checkpoints ({len(filenames)}) found for" + f" --iter {params.iter}, --avg {params.avg}" + ) + logging.info(f"averaging {filenames}") + model.to(device) + model.load_state_dict(average_checkpoints(filenames, device=device)) + elif params.avg == 1: + load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model) + else: + start = params.epoch - params.avg + 1 + filenames = [] + for i in range(start, params.epoch + 1): + if i >= 1: + filenames.append(f"{params.exp_dir}/epoch-{i}.pt") + logging.info(f"averaging {filenames}") + model.to(device) + model.load_state_dict(average_checkpoints(filenames, device=device)) + else: + if params.iter > 0: + filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[ + : params.avg + 1 + ] + if len(filenames) == 0: + raise ValueError( + f"No checkpoints found for" + f" --iter {params.iter}, --avg {params.avg}" + ) + elif len(filenames) < params.avg + 1: + raise ValueError( + f"Not enough checkpoints ({len(filenames)}) found for" + f" --iter {params.iter}, --avg {params.avg}" + ) + filename_start = filenames[-1] + filename_end = filenames[0] + logging.info( + "Calculating the averaged model over iteration checkpoints" + f" from {filename_start} (excluded) to {filename_end}" + ) + model.to(device) + model.load_state_dict( + average_checkpoints_with_averaged_model( + filename_start=filename_start, + filename_end=filename_end, + device=device, + ) + ) + else: + assert params.avg > 0, params.avg + start = params.epoch - params.avg + assert start >= 1, start + filename_start = f"{params.exp_dir}/epoch-{start}.pt" + filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt" + logging.info( + f"Calculating the averaged model over epoch range from " + f"{start} (excluded) to {params.epoch}" + ) + model.to(device) + model.load_state_dict( + average_checkpoints_with_averaged_model( + filename_start=filename_start, + filename_end=filename_end, + device=device, + ) + ) + + model.to(device) + model.eval() + + if "fast_beam_search" in params.decoding_method: + if params.decoding_method == "fast_beam_search_nbest_LG": + lexicon = Lexicon(params.lang_dir) + word_table = lexicon.word_table + lg_filename = params.lang_dir / "LG.pt" + logging.info(f"Loading {lg_filename}") + decoding_graph = k2.Fsa.from_dict( + torch.load(lg_filename, map_location=device) + ) + decoding_graph.scores *= params.ngram_lm_scale + else: + word_table = None + decoding_graph = k2.trivial_graph(params.vocab_size - 1, device=device) + else: + decoding_graph = None + word_table = None + + num_param = sum([p.numel() for p in model.parameters()]) + logging.info(f"Number of model parameters: {num_param}") + + # we need cut ids to display recognition results. + args.return_cuts = True + librispeech = LibriSpeechAsrDataModule(args) + + #test_clean_cuts = librispeech.test_clean_cuts(option='male') + #test_other_cuts = librispeech.test_other_cuts(option='male') + if 0: + test_clean_cuts = librispeech.test_clean_user(option=option) + test_other_cuts = librispeech.test_other_user(option=option) + test_clean_dl = librispeech.test_dataloaders(test_clean_cuts) + test_other_dl = librispeech.test_dataloaders(test_other_cuts) + test_sets = [f"test-clean", f"test-other"] + test_dl = [test_clean_dl, test_other_dl] + + if 0: + option = 'big' + test_clean_cuts = librispeech.test_clean_user(option=option) + test_other_cuts = librispeech.test_other_user(option=option) + test_clean_dl = librispeech.test_dataloaders(test_clean_cuts) + test_other_dl = librispeech.test_dataloaders(test_other_cuts) + + test_sets = [f"test-clean_sampling"] + test_dl = [test_clean_dl] + + #test_sets = [f"test-other_sampling"] + #test_dl = [test_other_dl] + + #test_sets = [f"test-clean_sampling", f"test-other_sampling"] + #test_dl = [test_clean_dl, test_other_dl] + + if 0: + option = '6938' + test_clean_cuts = librispeech.vox_cuts(option=option) + test_clean_dl = librispeech.test_dataloaders(test_clean_cuts) + test_sets = [f"test-clean_sampling"] + test_dl = [test_clean_dl] + + if 1: + test_clean_cuts = librispeech.userlibri_cuts(option=params.spk_id) + test_clean_dl = librispeech.test_dataloaders(test_clean_cuts) + test_sets = [f"{params.spk_id}"] + test_dl = [test_clean_dl] + + for test_set, test_dl in zip(test_sets, test_dl): + results_dict = decode_dataset( + dl=test_dl, + params=params, + model=model, + sp=sp, + word_table=word_table, + decoding_graph=decoding_graph, + ) + + save_results( + params=params, + test_set_name=test_set, + results_dict=results_dict, + ) + + logging.info("Done!") + + +if __name__ == "__main__": + main() diff --git a/egs/tedlium3/ASR/pruned_transducer_stateless_d2v_v2/decode_new.py b/egs/tedlium3/ASR/pruned_transducer_stateless_d2v_v2/decode_new.py new file mode 100755 index 000000000..d245eabf5 --- /dev/null +++ b/egs/tedlium3/ASR/pruned_transducer_stateless_d2v_v2/decode_new.py @@ -0,0 +1,834 @@ +#!/usr/bin/env python3 +# +# Copyright 2021-2022 Xiaomi Corporation (Author: Fangjun Kuang, +# Zengwei Yao) +# +# See ../../../../LICENSE for clarification regarding multiple authors +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +""" +Usage: +(0) for d2v-T decoding +for method in greedy_search modified_beam_search fast_beam_search; do + ./pruned_transducer_stateless_d2v_v2/decode.py \ + --input-strategy AudioSamples \ + --enable-spec-aug False \ + --additional-block True \ + --model-name epoc.pt \ + --exp-dir ./pruned_transducer_stateless_d2v_v2/960h_sweep_v3_388 \ + --max-duration 400 \ + --decoding-method $method \ + --max-sym-per-frame 1 \ + --encoder-type d2v \ + --encoder-dim 768 \ + --decoder-dim 768 \ + --joiner-dim 768 +done +""" + + +import argparse +import logging +import math +from collections import defaultdict +from pathlib import Path +from typing import Dict, List, Optional, Tuple + +import k2 +import sentencepiece as spm +import torch +import torch.nn as nn +from asr_datamodule import LibriSpeechAsrDataModule +from beam_search import ( + beam_search, + fast_beam_search_nbest, + fast_beam_search_nbest_LG, + fast_beam_search_nbest_oracle, + fast_beam_search_one_best, + greedy_search, + greedy_search_batch, + modified_beam_search, +) +#from train import add_model_arguments, add_rep_arguments, get_params, get_transducer_model +from prompt_tuning import add_model_arguments, add_rep_arguments, get_params, get_transducer_model + +from icefall.checkpoint import ( + average_checkpoints, + average_checkpoints_with_averaged_model, + find_checkpoints, + load_checkpoint, +) +from icefall.lexicon import Lexicon +from icefall.utils import ( + AttributeDict, + setup_logger, + store_transcripts, + str2bool, + write_error_stats, +) + +from train_lora import LoRAHook + +LOG_EPS = math.log(1e-10) + + +def get_parser(): + parser = argparse.ArgumentParser( + formatter_class=argparse.ArgumentDefaultsHelpFormatter + ) + parser.add_argument( + "--model-name", + type=str, + default="", + help="""It specifies the model file name to use for decoding.""", + ) + + parser.add_argument( + "--epoch", + type=int, + default=30, + help="""It specifies the checkpoint to use for decoding. + Note: Epoch counts from 1. + You can specify --avg to use more checkpoints for model averaging.""", + ) + + parser.add_argument( + "--iter", + type=int, + default=0, + help="""If positive, --epoch is ignored and it + will use the checkpoint exp_dir/checkpoint-iter.pt. + You can specify --avg to use more checkpoints for model averaging. + """, + ) + + parser.add_argument( + "--avg", + type=int, + default=9, + help="Number of checkpoints to average. Automatically select " + "consecutive checkpoints before the checkpoint specified by " + "'--epoch' and '--iter'", + ) + + parser.add_argument( + "--use-averaged-model", + type=str2bool, + default=True, + help="Whether to load averaged model. Currently it only supports " + "using --epoch. If True, it would decode with the averaged model " + "over the epoch range from `epoch-avg` (excluded) to `epoch`." + "Actually only the models with epoch number of `epoch-avg` and " + "`epoch` are loaded for averaging. ", + ) + + parser.add_argument( + "--exp-dir", + type=str, + default="pruned_transducer_stateless7_ctc/exp", + help="The experiment dir", + ) + + parser.add_argument( + "--bpe-model", + type=str, + default="data/lang_bpe_500/bpe.model", + help="Path to the BPE model", + ) + + parser.add_argument( + "--lang-dir", + type=Path, + default="data/lang_bpe_500", + help="The lang dir containing word table and LG graph", + ) + + parser.add_argument( + "--decoding-method", + type=str, + default="greedy_search", + help="""Possible values are: + - greedy_search + - beam_search + - modified_beam_search + - fast_beam_search + - fast_beam_search_nbest + - fast_beam_search_nbest_oracle + - fast_beam_search_nbest_LG + If you use fast_beam_search_nbest_LG, you have to specify + `--lang-dir`, which should contain `LG.pt`. + """, + ) + + parser.add_argument( + "--beam-size", + type=int, + default=4, + help="""An integer indicating how many candidates we will keep for each + frame. Used only when --decoding-method is beam_search or + modified_beam_search.""", + ) + + parser.add_argument( + "--beam", + type=float, + default=20.0, + help="""A floating point value to calculate the cutoff score during beam + search (i.e., `cutoff = max-score - beam`), which is the same as the + `beam` in Kaldi. + Used only when --decoding-method is fast_beam_search, + fast_beam_search_nbest, fast_beam_search_nbest_LG, + and fast_beam_search_nbest_oracle + """, + ) + + parser.add_argument( + "--ngram-lm-scale", + type=float, + default=0.01, + help=""" + Used only when --decoding_method is fast_beam_search_nbest_LG. + It specifies the scale for n-gram LM scores. + """, + ) + + parser.add_argument( + "--max-contexts", + type=int, + default=8, + help="""Used only when --decoding-method is + fast_beam_search, fast_beam_search_nbest, fast_beam_search_nbest_LG, + and fast_beam_search_nbest_oracle""", + ) + + parser.add_argument( + "--max-states", + type=int, + default=64, + help="""Used only when --decoding-method is + fast_beam_search, fast_beam_search_nbest, fast_beam_search_nbest_LG, + and fast_beam_search_nbest_oracle""", + ) + + parser.add_argument( + "--context-size", + type=int, + default=2, + help="The context size in the decoder. 1 means bigram; 2 means tri-gram", + ) + parser.add_argument( + "--max-sym-per-frame", + type=int, + default=1, + help="""Maximum number of symbols per frame. + Used only when --decoding_method is greedy_search""", + ) + + parser.add_argument( + "--num-paths", + type=int, + default=200, + help="""Number of paths for nbest decoding. + Used only when the decoding method is fast_beam_search_nbest, + fast_beam_search_nbest_LG, and fast_beam_search_nbest_oracle""", + ) + + parser.add_argument( + "--nbest-scale", + type=float, + default=0.5, + help="""Scale applied to lattice scores when computing nbest paths. + Used only when the decoding method is fast_beam_search_nbest, + fast_beam_search_nbest_LG, and fast_beam_search_nbest_oracle""", + ) + + parser.add_argument( + "--simulate-streaming", + type=str2bool, + default=False, + help="""Whether to simulate streaming in decoding, this is a good way to + test a streaming model. + """, + ) + + parser.add_argument( + "--decode-chunk-size", + type=int, + default=16, + help="The chunk size for decoding (in frames after subsampling)", + ) + + parser.add_argument( + "--left-context", + type=int, + default=64, + help="left context can be seen during decoding (in frames after subsampling)", + ) + + parser.add_argument( + "--res-name", + type=str, + ) + + add_model_arguments(parser) + add_rep_arguments(parser) + + return parser + + +def decode_one_batch( + params: AttributeDict, + model: nn.Module, + sp: spm.SentencePieceProcessor, + batch: dict, + word_table: Optional[k2.SymbolTable] = None, + decoding_graph: Optional[k2.Fsa] = None, +) -> Dict[str, List[List[str]]]: + """Decode one batch and return the result in a dict. The dict has the + following format: + + - key: It indicates the setting used for decoding. For example, + if greedy_search is used, it would be "greedy_search" + If beam search with a beam size of 7 is used, it would be + "beam_7" + - value: It contains the decoding result. `len(value)` equals to + batch size. `value[i]` is the decoding result for the i-th + utterance in the given batch. + Args: + params: + It's the return value of :func:`get_params`. + model: + The neural model. + sp: + The BPE model. + batch: + It is the return value from iterating + `lhotse.dataset.K2SpeechRecognitionDataset`. See its documentation + for the format of the `batch`. + word_table: + The word symbol table. + decoding_graph: + The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used + only when --decoding_method is fast_beam_search, fast_beam_search_nbest, + fast_beam_search_nbest_oracle, and fast_beam_search_nbest_LG. + Returns: + Return the decoding result. See above description for the format of + the returned dict. + """ + device = next(model.parameters()).device + feature = batch["inputs"] + assert feature.ndim == 2 or feature.ndim == 3 + + feature = feature.to(device) + # at entry, feature is (N, T, C) + + supervisions = batch["supervisions"] + #feature_lens = supervisions["num_frames"].to(device) + if feature.ndim == 2: + feature_lens = [] + for supervision in supervisions['cut']: + try: feature_lens.append(supervision.tracks[0].cut.recording.num_samples) + except: feature_lens.append(supervision.recording.num_samples) + feature_lens = torch.tensor(feature_lens) + + elif feature.ndim == 3: + feature_lens = supervisions["num_frames"].to(device) + + if params.simulate_streaming: + feature_lens += params.left_context + feature = torch.nn.functional.pad( + feature, + pad=(0, 0, 0, params.left_context), + value=LOG_EPS, + ) + encoder_out, encoder_out_lens, _ = model.encoder.streaming_forward( + x=feature, + x_lens=feature_lens, + chunk_size=params.decode_chunk_size, + left_context=params.left_context, + simulate_streaming=True, + ) + else: + encoder_out, encoder_out_lens = model.encoder(x=feature, x_lens=feature_lens) + + hyps = [] + + if params.decoding_method == "fast_beam_search": + hyp_tokens = fast_beam_search_one_best( + model=model, + decoding_graph=decoding_graph, + encoder_out=encoder_out, + encoder_out_lens=encoder_out_lens, + beam=params.beam, + max_contexts=params.max_contexts, + max_states=params.max_states, + ) + for hyp in sp.decode(hyp_tokens): + hyps.append(hyp.split()) + elif params.decoding_method == "fast_beam_search_nbest_LG": + hyp_tokens = fast_beam_search_nbest_LG( + model=model, + decoding_graph=decoding_graph, + encoder_out=encoder_out, + encoder_out_lens=encoder_out_lens, + beam=params.beam, + max_contexts=params.max_contexts, + max_states=params.max_states, + num_paths=params.num_paths, + nbest_scale=params.nbest_scale, + ) + for hyp in hyp_tokens: + hyps.append([word_table[i] for i in hyp]) + elif params.decoding_method == "fast_beam_search_nbest": + hyp_tokens = fast_beam_search_nbest( + model=model, + decoding_graph=decoding_graph, + encoder_out=encoder_out, + encoder_out_lens=encoder_out_lens, + beam=params.beam, + max_contexts=params.max_contexts, + max_states=params.max_states, + num_paths=params.num_paths, + nbest_scale=params.nbest_scale, + ) + for hyp in sp.decode(hyp_tokens): + hyps.append(hyp.split()) + elif params.decoding_method == "fast_beam_search_nbest_oracle": + hyp_tokens = fast_beam_search_nbest_oracle( + model=model, + decoding_graph=decoding_graph, + encoder_out=encoder_out, + encoder_out_lens=encoder_out_lens, + beam=params.beam, + max_contexts=params.max_contexts, + max_states=params.max_states, + num_paths=params.num_paths, + ref_texts=sp.encode(supervisions["text"]), + nbest_scale=params.nbest_scale, + ) + for hyp in sp.decode(hyp_tokens): + hyps.append(hyp.split()) + elif params.decoding_method == "greedy_search" and params.max_sym_per_frame == 1: + hyp_tokens = greedy_search_batch( + model=model, + encoder_out=encoder_out, + encoder_out_lens=encoder_out_lens, + ) + for hyp in sp.decode(hyp_tokens): + hyps.append(hyp.split()) + elif params.decoding_method == "modified_beam_search": + hyp_tokens = modified_beam_search( + model=model, + encoder_out=encoder_out, + encoder_out_lens=encoder_out_lens, + beam=params.beam_size, + ) + for hyp in sp.decode(hyp_tokens): + hyps.append(hyp.split()) + else: + batch_size = encoder_out.size(0) + + for i in range(batch_size): + # fmt: off + encoder_out_i = encoder_out[i:i+1, :encoder_out_lens[i]] + # fmt: on + if params.decoding_method == "greedy_search": + hyp = greedy_search( + model=model, + encoder_out=encoder_out_i, + max_sym_per_frame=params.max_sym_per_frame, + ) + elif params.decoding_method == "beam_search": + hyp = beam_search( + model=model, + encoder_out=encoder_out_i, + beam=params.beam_size, + ) + else: + raise ValueError( + f"Unsupported decoding method: {params.decoding_method}" + ) + hyps.append(sp.decode(hyp).split()) + + if params.decoding_method == "greedy_search": + return {"greedy_search": hyps} + elif "fast_beam_search" in params.decoding_method: + key = f"beam_{params.beam}_" + key += f"max_contexts_{params.max_contexts}_" + key += f"max_states_{params.max_states}" + if "nbest" in params.decoding_method: + key += f"_num_paths_{params.num_paths}_" + key += f"nbest_scale_{params.nbest_scale}" + if "LG" in params.decoding_method: + key += f"_ngram_lm_scale_{params.ngram_lm_scale}" + + return {key: hyps} + else: + return {f"beam_size_{params.beam_size}": hyps} + + +def decode_dataset( + dl: torch.utils.data.DataLoader, + params: AttributeDict, + model: nn.Module, + sp: spm.SentencePieceProcessor, + word_table: Optional[k2.SymbolTable] = None, + decoding_graph: Optional[k2.Fsa] = None, +) -> Dict[str, List[Tuple[str, List[str], List[str]]]]: + """Decode dataset. + + Args: + dl: + PyTorch's dataloader containing the dataset to decode. + params: + It is returned by :func:`get_params`. + model: + The neural model. + sp: + The BPE model. + word_table: + The word symbol table. + decoding_graph: + The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used + only when --decoding_method is fast_beam_search, fast_beam_search_nbest, + fast_beam_search_nbest_oracle, and fast_beam_search_nbest_LG. + Returns: + Return a dict, whose key may be "greedy_search" if greedy search + is used, or it may be "beam_7" if beam size of 7 is used. + Its value is a list of tuples. Each tuple contains two elements: + The first is the reference transcript, and the second is the + predicted result. + """ + num_cuts = 0 + + try: + num_batches = len(dl) + except TypeError: + num_batches = "?" + + if params.decoding_method == "greedy_search": + log_interval = 50 + else: + log_interval = 20 + + results = defaultdict(list) + for batch_idx, batch in enumerate(dl): + texts = batch["supervisions"]["text"] + cut_ids = [cut.id for cut in batch["supervisions"]["cut"]] + + hyps_dict = decode_one_batch( + params=params, + model=model, + sp=sp, + decoding_graph=decoding_graph, + word_table=word_table, + batch=batch, + ) + + for name, hyps in hyps_dict.items(): + this_batch = [] + assert len(hyps) == len(texts) + for cut_id, hyp_words, ref_text in zip(cut_ids, hyps, texts): + ref_words = ref_text.split() + this_batch.append((cut_id, ref_words, hyp_words)) + + results[name].extend(this_batch) + + num_cuts += len(texts) + + if batch_idx % log_interval == 0: + batch_str = f"{batch_idx}/{num_batches}" + + logging.info(f"batch {batch_str}, cuts processed until now is {num_cuts}") + return results + + +def save_results( + params: AttributeDict, + test_set_name: str, + results_dict: Dict[str, List[Tuple[str, List[str], List[str]]]], +): + test_set_wers = dict() + for key, results in results_dict.items(): + recog_path = ( + params.res_dir / f"recogs-{test_set_name}-{key}-{params.suffix}.txt" + ) + results = sorted(results) + store_transcripts(filename=recog_path, texts=results) + logging.info(f"The transcripts are stored in {recog_path}") + + # The following prints out WERs, per-word error statistics and aligned + # ref/hyp pairs. + errs_filename = ( + params.res_dir / f"errs-{test_set_name}-{key}-{params.suffix}.txt" + ) + with open(errs_filename, "w") as f: + wer = write_error_stats( + f, f"{test_set_name}-{key}", results, enable_log=True + ) + test_set_wers[key] = wer + + logging.info("Wrote detailed error stats to {}".format(errs_filename)) + + test_set_wers = sorted(test_set_wers.items(), key=lambda x: x[1]) + errs_info = ( + params.res_dir / f"wer-summary-{test_set_name}-{key}-{params.suffix}.txt" + ) + with open(errs_info, "w") as f: + print("settings\tWER", file=f) + for key, val in test_set_wers: + print("{}\t{}".format(key, val), file=f) + + s = "\nFor {}, WER of different settings are:\n".format(test_set_name) + note = "\tbest for {}".format(test_set_name) + for key, val in test_set_wers: + s += "{}\t{}{}\n".format(key, val, note) + note = "" + logging.info(s) + + +@torch.no_grad() +def main(): + parser = get_parser() + LibriSpeechAsrDataModule.add_arguments(parser) + args = parser.parse_args() + args.exp_dir = Path(args.exp_dir) + + params = get_params() + params.update(vars(args)) + + assert params.decoding_method in ( + "greedy_search", + "beam_search", + "fast_beam_search", + "fast_beam_search_nbest", + "fast_beam_search_nbest_LG", + "fast_beam_search_nbest_oracle", + "modified_beam_search", + ) + params.res_dir = params.exp_dir / params.decoding_method + + if params.iter > 0: + params.suffix = f"iter-{params.iter}-avg-{params.avg}" + else: + params.suffix = f"epoch-{params.epoch}-avg-{params.avg}" + + if params.simulate_streaming: + params.suffix += f"-streaming-chunk-size-{params.decode_chunk_size}" + params.suffix += f"-left-context-{params.left_context}" + + if "fast_beam_search" in params.decoding_method: + params.suffix += f"-beam-{params.beam}" + params.suffix += f"-max-contexts-{params.max_contexts}" + params.suffix += f"-max-states-{params.max_states}" + if "nbest" in params.decoding_method: + params.suffix += f"-nbest-scale-{params.nbest_scale}" + params.suffix += f"-num-paths-{params.num_paths}" + if "LG" in params.decoding_method: + params.suffix += f"-ngram-lm-scale-{params.ngram_lm_scale}" + elif "beam_search" in params.decoding_method: + params.suffix += f"-{params.decoding_method}-beam-size-{params.beam_size}" + else: + params.suffix += f"-context-{params.context_size}" + params.suffix += f"-max-sym-per-frame-{params.max_sym_per_frame}" + + if params.use_averaged_model: + params.suffix += "-use-averaged-model" + + setup_logger(f"{params.res_dir}/log-decode-{params.suffix}") + logging.info("Decoding started") + + device = torch.device("cpu") + if torch.cuda.is_available(): + device = torch.device("cuda", 0) + + logging.info(f"Device: {device}") + + sp = spm.SentencePieceProcessor() + sp.load(params.bpe_model) + + # and are defined in local/train_bpe_model.py + params.blank_id = sp.piece_to_id("") + params.unk_id = sp.piece_to_id("") + params.vocab_size = sp.get_piece_size() + + if params.simulate_streaming: + assert ( + params.causal_convolution + ), "Decoding in streaming requires causal convolution" + + logging.info(params) + + logging.info("About to create model") + model = get_transducer_model(params) + + if '.pt' in params.model_name: + load_checkpoint(f"{params.exp_dir}/{params.model_name}", model) + elif 'lora' in params.model_name: + load_checkpoint(f"{params.exp_dir}/../d2v-base-T.pt", model) + + ## for lora hooking + lora_modules = [] + for modules in model.modules(): + if isinstance(modules, fairseq.modules.multihead_attention.MultiheadAttention): + for module in modules.modules(): + if isinstance(module, torch.nn.Linear): + lora_modules.append(LoRAHook(module)) + + for i, lora in enumerate(lora_modules): + lora_param = torch.load(f"{params.exp_dir}/lora_{params.iter}_{i}.pt") + lora.lora.load_state_dict(lora_param) + lora.lora.to(device) + logging.info("lora params load done") + else: + if not params.use_averaged_model: + if params.iter > 0: + filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[ + : params.avg + ] + if len(filenames) == 0: + raise ValueError( + f"No checkpoints found for" + f" --iter {params.iter}, --avg {params.avg}" + ) + elif len(filenames) < params.avg: + raise ValueError( + f"Not enough checkpoints ({len(filenames)}) found for" + f" --iter {params.iter}, --avg {params.avg}" + ) + logging.info(f"averaging {filenames}") + model.to(device) + model.load_state_dict(average_checkpoints(filenames, device=device)) + elif params.avg == 1: + load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model) + else: + start = params.epoch - params.avg + 1 + filenames = [] + for i in range(start, params.epoch + 1): + if i >= 1: + filenames.append(f"{params.exp_dir}/epoch-{i}.pt") + logging.info(f"averaging {filenames}") + model.to(device) + model.load_state_dict(average_checkpoints(filenames, device=device)) + else: + if params.iter > 0: + filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[ + : params.avg + 1 + ] + if len(filenames) == 0: + raise ValueError( + f"No checkpoints found for" + f" --iter {params.iter}, --avg {params.avg}" + ) + elif len(filenames) < params.avg + 1: + raise ValueError( + f"Not enough checkpoints ({len(filenames)}) found for" + f" --iter {params.iter}, --avg {params.avg}" + ) + filename_start = filenames[-1] + filename_end = filenames[0] + logging.info( + "Calculating the averaged model over iteration checkpoints" + f" from {filename_start} (excluded) to {filename_end}" + ) + model.to(device) + model.load_state_dict( + average_checkpoints_with_averaged_model( + filename_start=filename_start, + filename_end=filename_end, + device=device, + ) + ) + else: + assert params.avg > 0, params.avg + start = params.epoch - params.avg + assert start >= 1, start + filename_start = f"{params.exp_dir}/epoch-{start}.pt" + filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt" + logging.info( + f"Calculating the averaged model over epoch range from " + f"{start} (excluded) to {params.epoch}" + ) + model.to(device) + model.load_state_dict( + average_checkpoints_with_averaged_model( + filename_start=filename_start, + filename_end=filename_end, + device=device, + ) + ) + + model.to(device) + model.eval() + + if "fast_beam_search" in params.decoding_method: + if params.decoding_method == "fast_beam_search_nbest_LG": + lexicon = Lexicon(params.lang_dir) + word_table = lexicon.word_table + lg_filename = params.lang_dir / "LG.pt" + logging.info(f"Loading {lg_filename}") + decoding_graph = k2.Fsa.from_dict( + torch.load(lg_filename, map_location=device) + ) + decoding_graph.scores *= params.ngram_lm_scale + else: + word_table = None + decoding_graph = k2.trivial_graph(params.vocab_size - 1, device=device) + else: + decoding_graph = None + word_table = None + + num_param = sum([p.numel() for p in model.parameters()]) + logging.info(f"Number of model parameters: {num_param}") + + # we need cut ids to display recognition results. + args.return_cuts = True + librispeech = LibriSpeechAsrDataModule(args) + + ''' + test_clean_cuts = librispeech.test_clean_cuts() + test_other_cuts = librispeech.test_other_cuts() + + test_clean_dl = librispeech.test_dataloaders(test_clean_cuts) + test_other_dl = librispeech.test_dataloaders(test_other_cuts) + + test_sets = ["test-clean", "test-other"] + test_dl = [test_clean_dl, test_other_dl] + ''' + + test_clean_cuts = librispeech.userlibri_cuts(option=params.spk_id) + test_clean_dl = librispeech.test_dataloaders(test_clean_cuts) + test_sets = [f"{params.spk_id}"] + test_dl = [test_clean_dl] + + for test_set, test_dl in zip(test_sets, test_dl): + results_dict = decode_dataset( + dl=test_dl, + params=params, + model=model, + sp=sp, + word_table=word_table, + decoding_graph=decoding_graph, + ) + + save_results( + params=params, + test_set_name=test_set, + results_dict=results_dict, + ) + + logging.info("Done!") + + +if __name__ == "__main__": + main() diff --git a/egs/tedlium3/ASR/pruned_transducer_stateless_d2v_v2/decoder.py b/egs/tedlium3/ASR/pruned_transducer_stateless_d2v_v2/decoder.py new file mode 100644 index 000000000..5f90e6375 --- /dev/null +++ b/egs/tedlium3/ASR/pruned_transducer_stateless_d2v_v2/decoder.py @@ -0,0 +1,102 @@ +# Copyright 2021 Xiaomi Corp. (authors: Fangjun Kuang) +# +# See ../../../../LICENSE for clarification regarding multiple authors +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import torch +import torch.nn as nn +import torch.nn.functional as F + + +class Decoder(nn.Module): + """This class modifies the stateless decoder from the following paper: + + RNN-transducer with stateless prediction network + https://ieeexplore.ieee.org/stamp/stamp.jsp?arnumber=9054419 + + It removes the recurrent connection from the decoder, i.e., the prediction + network. Different from the above paper, it adds an extra Conv1d + right after the embedding layer. + + TODO: Implement https://arxiv.org/pdf/2109.07513.pdf + """ + + def __init__( + self, + vocab_size: int, + decoder_dim: int, + blank_id: int, + context_size: int, + ): + """ + Args: + vocab_size: + Number of tokens of the modeling unit including blank. + decoder_dim: + Dimension of the input embedding, and of the decoder output. + blank_id: + The ID of the blank symbol. + context_size: + Number of previous words to use to predict the next word. + 1 means bigram; 2 means trigram. n means (n+1)-gram. + """ + super().__init__() + + self.embedding = nn.Embedding( + num_embeddings=vocab_size, + embedding_dim=decoder_dim, + padding_idx=blank_id, + ) + self.blank_id = blank_id + + assert context_size >= 1, context_size + self.context_size = context_size + self.vocab_size = vocab_size + if context_size > 1: + self.conv = nn.Conv1d( + in_channels=decoder_dim, + out_channels=decoder_dim, + kernel_size=context_size, + padding=0, + groups=decoder_dim // 4, # group size == 4 + bias=False, + ) + + def forward(self, y: torch.Tensor, need_pad: bool = True) -> torch.Tensor: + """ + Args: + y: + A 2-D tensor of shape (N, U). + need_pad: + True to left pad the input. Should be True during training. + False to not pad the input. Should be False during inference. + Returns: + Return a tensor of shape (N, U, decoder_dim). + """ + y = y.to(torch.int64) + # this stuff about clamp() is a temporary fix for a mismatch + # at utterance start, we use negative ids in beam_search.py + embedding_out = self.embedding(y.clamp(min=0)) * (y >= 0).unsqueeze(-1) + if self.context_size > 1: + embedding_out = embedding_out.permute(0, 2, 1) + if need_pad is True: + embedding_out = F.pad(embedding_out, pad=(self.context_size - 1, 0)) + else: + # During inference time, there is no need to do extra padding + # as we only need one output + assert embedding_out.size(-1) == self.context_size + embedding_out = self.conv(embedding_out) + embedding_out = embedding_out.permute(0, 2, 1) + embedding_out = F.relu(embedding_out) + return embedding_out diff --git a/egs/tedlium3/ASR/pruned_transducer_stateless_d2v_v2/encoder_interface.py b/egs/tedlium3/ASR/pruned_transducer_stateless_d2v_v2/encoder_interface.py new file mode 100644 index 000000000..257facce4 --- /dev/null +++ b/egs/tedlium3/ASR/pruned_transducer_stateless_d2v_v2/encoder_interface.py @@ -0,0 +1,43 @@ +# Copyright 2021 Xiaomi Corp. (authors: Fangjun Kuang) +# +# See ../../../../LICENSE for clarification regarding multiple authors +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +from typing import Tuple + +import torch +import torch.nn as nn + + +class EncoderInterface(nn.Module): + def forward( + self, x: torch.Tensor, x_lens: torch.Tensor + ) -> Tuple[torch.Tensor, torch.Tensor]: + """ + Args: + x: + A tensor of shape (batch_size, input_seq_len, num_features) + containing the input features. + x_lens: + A tensor of shape (batch_size,) containing the number of frames + in `x` before padding. + Returns: + Return a tuple containing two tensors: + - encoder_out, a tensor of (batch_size, out_seq_len, output_dim) + containing unnormalized probabilities, i.e., the output of a + linear layer. + - encoder_out_lens, a tensor of shape (batch_size,) containing + the number of frames in `encoder_out` before padding. + """ + raise NotImplementedError("Please implement it in a subclass") diff --git a/egs/tedlium3/ASR/pruned_transducer_stateless_d2v_v2/export.py b/egs/tedlium3/ASR/pruned_transducer_stateless_d2v_v2/export.py new file mode 100755 index 000000000..59a393739 --- /dev/null +++ b/egs/tedlium3/ASR/pruned_transducer_stateless_d2v_v2/export.py @@ -0,0 +1,320 @@ +#!/usr/bin/env python3 +# +# Copyright 2021 Xiaomi Corporation (Author: Fangjun Kuang) +# +# See ../../../../LICENSE for clarification regarding multiple authors +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +# This script converts several saved checkpoints +# to a single one using model averaging. +""" + +Usage: + +(1) Export to torchscript model using torch.jit.script() + +./pruned_transducer_stateless7_ctc/export.py \ + --exp-dir ./pruned_transducer_stateless7_ctc/exp \ + --bpe-model data/lang_bpe_500/bpe.model \ + --epoch 30 \ + --avg 9 \ + --jit 1 + +It will generate a file `cpu_jit.pt` in the given `exp_dir`. You can later +load it by `torch.jit.load("cpu_jit.pt")`. + +Note `cpu` in the name `cpu_jit.pt` means the parameters when loaded into Python +are on CPU. You can use `to("cuda")` to move them to a CUDA device. + +Check +https://github.com/k2-fsa/sherpa +for how to use the exported models outside of icefall. + +(2) Export `model.state_dict()` + +./pruned_transducer_stateless7_ctc/export.py \ + --exp-dir ./pruned_transducer_stateless7_ctc/exp \ + --bpe-model data/lang_bpe_500/bpe.model \ + --epoch 20 \ + --avg 10 + +It will generate a file `pretrained.pt` in the given `exp_dir`. You can later +load it by `icefall.checkpoint.load_checkpoint()`. + +To use the generated file with `pruned_transducer_stateless7_ctc/decode.py`, +you can do: + + cd /path/to/exp_dir + ln -s pretrained.pt epoch-9999.pt + + cd /path/to/egs/librispeech/ASR + ./pruned_transducer_stateless7_ctc/decode.py \ + --exp-dir ./pruned_transducer_stateless7_ctc/exp \ + --epoch 9999 \ + --avg 1 \ + --max-duration 600 \ + --decoding-method greedy_search \ + --bpe-model data/lang_bpe_500/bpe.model + +Check ./pretrained.py for its usage. + +Note: If you don't want to train a model from scratch, we have +provided one for you. You can get it at + +https://huggingface.co/csukuangfj/icefall-asr-librispeech-pruned-transducer-stateless7-2022-11-11 + +with the following commands: + + sudo apt-get install git-lfs + git lfs install + git clone https://huggingface.co/csukuangfj/icefall-asr-librispeech-pruned-transducer-stateless7-2022-11-11 + # You will find the pre-trained model in icefall-asr-librispeech-pruned-transducer-stateless7-2022-11-11/exp +""" + +import argparse +import logging +from pathlib import Path + +import sentencepiece as spm +import torch +from scaling_converter import convert_scaled_to_non_scaled +from train import add_model_arguments, get_params, get_transducer_model + +from icefall.checkpoint import ( + average_checkpoints, + average_checkpoints_with_averaged_model, + find_checkpoints, + load_checkpoint, +) +from icefall.utils import str2bool + + +def get_parser(): + parser = argparse.ArgumentParser( + formatter_class=argparse.ArgumentDefaultsHelpFormatter + ) + + parser.add_argument( + "--epoch", + type=int, + default=30, + help="""It specifies the checkpoint to use for decoding. + Note: Epoch counts from 1. + You can specify --avg to use more checkpoints for model averaging.""", + ) + + parser.add_argument( + "--iter", + type=int, + default=0, + help="""If positive, --epoch is ignored and it + will use the checkpoint exp_dir/checkpoint-iter.pt. + You can specify --avg to use more checkpoints for model averaging. + """, + ) + + parser.add_argument( + "--avg", + type=int, + default=9, + help="Number of checkpoints to average. Automatically select " + "consecutive checkpoints before the checkpoint specified by " + "'--epoch' and '--iter'", + ) + + parser.add_argument( + "--use-averaged-model", + type=str2bool, + default=True, + help="Whether to load averaged model. Currently it only supports " + "using --epoch. If True, it would decode with the averaged model " + "over the epoch range from `epoch-avg` (excluded) to `epoch`." + "Actually only the models with epoch number of `epoch-avg` and " + "`epoch` are loaded for averaging. ", + ) + + parser.add_argument( + "--exp-dir", + type=str, + default="pruned_transducer_stateless7/exp", + help="""It specifies the directory where all training related + files, e.g., checkpoints, log, etc, are saved + """, + ) + + parser.add_argument( + "--bpe-model", + type=str, + default="data/lang_bpe_500/bpe.model", + help="Path to the BPE model", + ) + + parser.add_argument( + "--jit", + type=str2bool, + default=False, + help="""True to save a model after applying torch.jit.script. + It will generate a file named cpu_jit.pt + + Check ./jit_pretrained.py for how to use it. + """, + ) + + parser.add_argument( + "--context-size", + type=int, + default=2, + help="The context size in the decoder. 1 means bigram; 2 means tri-gram", + ) + + add_model_arguments(parser) + + return parser + + +@torch.no_grad() +def main(): + args = get_parser().parse_args() + args.exp_dir = Path(args.exp_dir) + + params = get_params() + params.update(vars(args)) + + device = torch.device("cpu") + if torch.cuda.is_available(): + device = torch.device("cuda", 0) + + logging.info(f"device: {device}") + + sp = spm.SentencePieceProcessor() + sp.load(params.bpe_model) + + # is defined in local/train_bpe_model.py + params.blank_id = sp.piece_to_id("") + params.vocab_size = sp.get_piece_size() + + logging.info(params) + + logging.info("About to create model") + model = get_transducer_model(params) + + model.to(device) + + if not params.use_averaged_model: + if params.iter > 0: + filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[ + : params.avg + ] + if len(filenames) == 0: + raise ValueError( + f"No checkpoints found for" + f" --iter {params.iter}, --avg {params.avg}" + ) + elif len(filenames) < params.avg: + raise ValueError( + f"Not enough checkpoints ({len(filenames)}) found for" + f" --iter {params.iter}, --avg {params.avg}" + ) + logging.info(f"averaging {filenames}") + model.to(device) + model.load_state_dict(average_checkpoints(filenames, device=device)) + elif params.avg == 1: + load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model) + else: + start = params.epoch - params.avg + 1 + filenames = [] + for i in range(start, params.epoch + 1): + if i >= 1: + filenames.append(f"{params.exp_dir}/epoch-{i}.pt") + logging.info(f"averaging {filenames}") + model.to(device) + model.load_state_dict(average_checkpoints(filenames, device=device)) + else: + if params.iter > 0: + filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[ + : params.avg + 1 + ] + if len(filenames) == 0: + raise ValueError( + f"No checkpoints found for" + f" --iter {params.iter}, --avg {params.avg}" + ) + elif len(filenames) < params.avg + 1: + raise ValueError( + f"Not enough checkpoints ({len(filenames)}) found for" + f" --iter {params.iter}, --avg {params.avg}" + ) + filename_start = filenames[-1] + filename_end = filenames[0] + logging.info( + "Calculating the averaged model over iteration checkpoints" + f" from {filename_start} (excluded) to {filename_end}" + ) + model.to(device) + model.load_state_dict( + average_checkpoints_with_averaged_model( + filename_start=filename_start, + filename_end=filename_end, + device=device, + ) + ) + else: + assert params.avg > 0, params.avg + start = params.epoch - params.avg + assert start >= 1, start + filename_start = f"{params.exp_dir}/epoch-{start}.pt" + filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt" + logging.info( + f"Calculating the averaged model over epoch range from " + f"{start} (excluded) to {params.epoch}" + ) + model.to(device) + model.load_state_dict( + average_checkpoints_with_averaged_model( + filename_start=filename_start, + filename_end=filename_end, + device=device, + ) + ) + + model.to("cpu") + model.eval() + + if params.jit is True: + convert_scaled_to_non_scaled(model, inplace=True) + logging.info("Using torch.jit.script()") + # We won't use the forward() method of the model in C++, so just ignore + # it here. + # Otherwise, one of its arguments is a ragged tensor and is not + # torch scriptabe. + model.__class__.forward = torch.jit.ignore(model.__class__.forward) + logging.info("Using torch.jit.script") + model = torch.jit.script(model) + filename = params.exp_dir / "cpu_jit.pt" + model.save(str(filename)) + logging.info(f"Saved to {filename}") + else: + logging.info("Not using torchscript. Export model.state_dict()") + # Save it using a format so that it can be loaded + # by :func:`load_checkpoint` + filename = params.exp_dir / "pretrained.pt" + torch.save({"model": model.state_dict()}, str(filename)) + logging.info(f"Saved to {filename}") + + +if __name__ == "__main__": + formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s" + + logging.basicConfig(format=formatter, level=logging.INFO) + main() diff --git a/egs/tedlium3/ASR/pruned_transducer_stateless_d2v_v2/full_ft.py b/egs/tedlium3/ASR/pruned_transducer_stateless_d2v_v2/full_ft.py new file mode 100755 index 000000000..edb69f684 --- /dev/null +++ b/egs/tedlium3/ASR/pruned_transducer_stateless_d2v_v2/full_ft.py @@ -0,0 +1,1732 @@ +#!/usr/bin/env python3 +# Copyright 2021-2022 Xiaomi Corp. (authors: Fangjun Kuang, +# Wei Kang, +# Mingshuang Luo,) +# Zengwei Yao) +# +# See ../../../../LICENSE for clarification regarding multiple authors +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +""" +Usage: + +export CUDA_VISIBLE_DEVICES="0,1,2,3" + +./pruned_transducer_stateless7_ctc/train.py \ + --world-size 4 \ + --num-epochs 30 \ + --start-epoch 1 \ + --exp-dir pruned_transducer_stateless7_ctc/exp \ + --full-libri 1 \ + --max-duration 300 + +# For mix precision training: + +./pruned_transducer_stateless7_ctc/train.py \ + --world-size 4 \ + --num-epochs 30 \ + --start-epoch 1 \ + --use-fp16 1 \ + --exp-dir pruned_transducer_stateless7_ctc/exp \ + --full-libri 1 \ + --max-duration 550 + +# For d2v-T training: +export CUDA_VISIBLE_DEVICES="0,1,2,3,4,5,6,7" + +./pruned_transducer_stateless_d2v_v2/train.py \ + --wandb true \ + --input-strategy AudioSamples \ + --enable-spec-aug False \ + --multi-optim True \ + --world-size 8 \ + --num-epochs 30 \ + --start-epoch 1 \ + --full-libri 0 \ + --exp-dir ./pruned_transducer_stateless_d2v_v2/$1 \ + --max-duration 250 \ + --freeze-finetune-updates 2000 \ + --use-fp16 1 \ + --peak-enc-lr 0.001 \ + --peak-dec-lr 0.05 \ + --accum-grads 1 \ + --encoder-type d2v \ + --additional-block True \ + --encoder-dim 768 \ + --decoder-dim 768 \ + --joiner-dim 768 \ + --prune-range 20 \ + --context-size 2 \ + --ctc-loss-scale 0.2 + +""" + + +import random +import argparse +import copy +import logging +import warnings +from pathlib import Path +from shutil import copyfile +from typing import Any, Dict, Optional, Tuple, Union + +import k2 +import optim +import sentencepiece as spm +import torch +import torch.multiprocessing as mp +import torch.nn as nn +from asr_datamodule import LibriSpeechAsrDataModule +from decoder import Decoder +from joiner import Joiner +from lhotse.cut import Cut +from lhotse.dataset.sampling.base import CutSampler +from lhotse.utils import fix_random_seed +from model import Transducer +from optim import Eden, ScaledAdam +from torch import Tensor +from torch.cuda.amp import GradScaler +from torch.nn.parallel import DistributedDataParallel as DDP +from torch.utils.tensorboard import SummaryWriter +from zipformer import Zipformer +from data2vec_encoder import FairSeqData2VecEncoder + +from icefall import diagnostics +from icefall.checkpoint import remove_checkpoints +from icefall.checkpoint import update_averaged_model +from checkpoint import ( + save_checkpoint as save_checkpoint_impl, + save_checkpoint_with_global_batch_idx, + load_checkpoint +) +from icefall.dist import cleanup_dist, setup_dist +from icefall.env import get_env_info +from icefall.hooks import register_inf_check_hooks +from icefall.utils import ( + AttributeDict, + MetricsTracker, + encode_supervisions, + setup_logger, + str2bool, + save_args, +) + +import wandb + +#from icefall.checkpoint import save_checkpoint as save_checkpoint_impl +LRSchedulerType = Union[torch.optim.lr_scheduler._LRScheduler, optim.LRScheduler] + + +def set_batch_count(model: Union[nn.Module, DDP], batch_count: float) -> None: + if isinstance(model, DDP): + # get underlying nn.Module + model = model.module + for module in model.modules(): + if hasattr(module, "batch_count"): + module.batch_count = batch_count + model.encoder.num_updates = int(batch_count) + + +def add_adapter_arguments(parser: argparse.ArgumentParser): + parser.add_argument( + "--add-adapter", + type=str2bool, + default=False, + help="add adapter to rep model's encoder" + ) + + parser.add_argument( + "--adapter-lr", + type=float, + default=0.0001, + help="adapter learning rate" + ) + + +def add_rep_arguments(parser: argparse.ArgumentParser): + parser.add_argument( + "--wandb", + type=str2bool, + default=True, + help="Use wandb for MLOps", + ) + parser.add_argument( + "--accum-grads", + type=int, + default=1, + help="accum-grad num.", + ) + + parser.add_argument( + "--multi-optim", + type=str2bool, + default=True, + help="use sperate optimizer (enc / dec)", + ) + + parser.add_argument( + "--peak-enc-lr", + type=float, + default=0.0001, + help="The initial learning rate. This value should not need to be changed.", + ) + + parser.add_argument( + "--peak-dec-lr", + type=float, + default=0.001, + help="The initial learning rate. This value should not need to be changed.", + ) + + parser.add_argument( + "--encoder-type", + type=str, + default='d2v', + help="Type of encoder (e.g. conformer, w2v, d2v...", + ) + + parser.add_argument( + "--encoder-dim", + type=int, + default=768, + help="encoder embedding dimension", + ) + + parser.add_argument( + "--freeze-finetune-updates", + type=int, + default=0 + ) + + parser.add_argument( + "--additional-block", + type=str2bool, + default=True, + ) + + parser.add_argument( + "--decode-interval", + type=int, + default=200, + help="decode interval", + ) + + +def add_model_arguments(parser: argparse.ArgumentParser): + parser.add_argument( + "--num-encoder-layers", + type=str, + default="2,4,3,2,4", + help="Number of zipformer encoder layers, comma separated.", + ) + + parser.add_argument( + "--feedforward-dims", + type=str, + default="1024,1024,2048,2048,1024", + help="Feedforward dimension of the zipformer encoder layers, comma separated.", + ) + + parser.add_argument( + "--nhead", + type=str, + default="8,8,8,8,8", + help="Number of attention heads in the zipformer encoder layers.", + ) + + parser.add_argument( + "--encoder-dims", + type=str, + default="384,384,384,384,384", + help="Embedding dimension in the 2 blocks of zipformer encoder layers, comma separated", + ) + + parser.add_argument( + "--attention-dims", + type=str, + default="192,192,192,192,192", + help="""Attention dimension in the 2 blocks of zipformer encoder layers, comma separated; + not the same as embedding dimension.""", + ) + + parser.add_argument( + "--encoder-unmasked-dims", + type=str, + default="256,256,256,256,256", + help="Unmasked dimensions in the encoders, relates to augmentation during training. " + "Must be <= each of encoder_dims. Empirically, less than 256 seems to make performance " + " worse.", + ) + + parser.add_argument( + "--zipformer-downsampling-factors", + type=str, + default="1,2,4,8,2", + help="Downsampling factor for each stack of encoder layers.", + ) + + parser.add_argument( + "--cnn-module-kernels", + type=str, + default="31,31,31,31,31", + help="Sizes of kernels in convolution modules", + ) + + parser.add_argument( + "--decoder-dim", + type=int, + default=512, + help="Embedding dimension in the decoder model.", + ) + + parser.add_argument( + "--joiner-dim", + type=int, + default=512, + help="""Dimension used in the joiner model. + Outputs from the encoder and decoder model are projected + to this dimension before adding. + """, + ) + + +def get_parser(): + parser = argparse.ArgumentParser( + formatter_class=argparse.ArgumentDefaultsHelpFormatter + ) + + parser.add_argument( + "--world-size", + type=int, + default=1, + help="Number of GPUs for DDP training.", + ) + + parser.add_argument( + "--master-port", + type=int, + default=12354, + help="Master port to use for DDP training.", + ) + + parser.add_argument( + "--tensorboard", + type=str2bool, + default=True, + help="Should various information be logged in tensorboard.", + ) + + parser.add_argument( + "--num-epochs", + type=int, + default=30, + help="Number of epochs to train.", + ) + + parser.add_argument( + "--num-updates", + type=int, + default=5000, + help="Number of epochs to train.", + ) + + parser.add_argument( + "--start-epoch", + type=int, + default=1, + help="""Resume training from this epoch. It should be positive. + If larger than 1, it will load checkpoint from + exp-dir/epoch-{start_epoch-1}.pt + """, + ) + + parser.add_argument( + "--start-batch", + type=int, + default=0, + help="""If positive, --start-epoch is ignored and + it loads the checkpoint from exp-dir/checkpoint-{start_batch}.pt + """, + ) + + parser.add_argument( + "--exp-dir", + type=str, + default="pruned_transducer_stateless7_ctc/exp", + help="""The experiment dir. + It specifies the directory where all training related + files, e.g., checkpoints, log, etc, are saved + """, + ) + + parser.add_argument( + "--bpe-model", + type=str, + default="data/lang_bpe_500/bpe.model", + help="Path to the BPE model", + ) + + parser.add_argument( + "--base-lr", type=float, default=0.05, help="The base learning rate." + ) + + parser.add_argument( + "--lr-batches", + type=float, + default=5000, + help="""Number of steps that affects how rapidly the learning rate + decreases. We suggest not to change this.""", + ) + + parser.add_argument( + "--lr-epochs", + type=float, + default=3.5, + help="""Number of epochs that affects how rapidly the learning rate decreases. + """, + ) + + parser.add_argument( + "--context-size", + type=int, + default=2, + help="The context size in the decoder. 1 means bigram; 2 means tri-gram", + ) + + parser.add_argument( + "--prune-range", + type=int, + default=5, + help="The prune range for rnnt loss, it means how many symbols(context)" + "we are using to compute the loss", + ) + + parser.add_argument( + "--lm-scale", + type=float, + default=0.25, + help="The scale to smooth the loss with lm " + "(output of prediction network) part.", + ) + + parser.add_argument( + "--am-scale", + type=float, + default=0.0, + help="The scale to smooth the loss with am (output of encoder network) part.", + ) + + parser.add_argument( + "--simple-loss-scale", + type=float, + default=0.5, + help="To get pruning ranges, we will calculate a simple version" + "loss(joiner is just addition), this simple loss also uses for" + "training (as a regularization item). We will scale the simple loss" + "with this parameter before adding to the final loss.", + ) + + parser.add_argument( + "--ctc-loss-scale", + type=float, + default=0.2, + help="Scale for CTC loss.", + ) + + parser.add_argument( + "--seed", + type=int, + default=42, + help="The seed for random generators intended for reproducibility", + ) + + parser.add_argument( + "--print-diagnostics", + type=str2bool, + default=False, + help="Accumulate stats on activations, print them and exit.", + ) + + parser.add_argument( + "--inf-check", + type=str2bool, + default=False, + help="Add hooks to check for infinite module outputs and gradients.", + ) + + parser.add_argument( + "--save-every-n", + type=int, + default=2000, + help="""Save checkpoint after processing this number of batches" + periodically. We save checkpoint to exp-dir/ whenever + params.batch_idx_train % save_every_n == 0. The checkpoint filename + has the form: f'exp-dir/checkpoint-{params.batch_idx_train}.pt' + Note: It also saves checkpoint to `exp-dir/epoch-xxx.pt` at the + end of each epoch where `xxx` is the epoch number counting from 0. + """, + ) + + parser.add_argument( + "--keep-last-k", + type=int, + default=30, + help="""Only keep this number of checkpoints on disk. + For instance, if it is 3, there are only 3 checkpoints + in the exp-dir with filenames `checkpoint-xxx.pt`. + It does not affect checkpoints with name `epoch-xxx.pt`. + """, + ) + + parser.add_argument( + "--average-period", + type=int, + default=200, + help="""Update the averaged model, namely `model_avg`, after processing + this number of batches. `model_avg` is a separate version of model, + in which each floating-point parameter is the average of all the + parameters from the start of training. Each time we take the average, + we do: `model_avg = model * (average_period / batch_idx_train) + + model_avg * ((batch_idx_train - average_period) / batch_idx_train)`. + """, + ) + + parser.add_argument( + "--use-fp16", + type=str2bool, + default=True, + help="Whether to use half precision training.", + ) + + add_model_arguments(parser) + add_rep_arguments(parser) + add_adapter_arguments(parser) + + return parser + + +def get_params() -> AttributeDict: + """Return a dict containing training parameters. + + All training related parameters that are not passed from the commandline + are saved in the variable `params`. + + Commandline options are merged into `params` after they are parsed, so + you can also access them via `params`. + + Explanation of options saved in `params`: + + - best_train_loss: Best training loss so far. It is used to select + the model that has the lowest training loss. It is + updated during the training. + + - best_valid_loss: Best validation loss so far. It is used to select + the model that has the lowest validation loss. It is + updated during the training. + + - best_train_epoch: It is the epoch that has the best training loss. + + - best_valid_epoch: It is the epoch that has the best validation loss. + + - batch_idx_train: Used to writing statistics to tensorboard. It + contains number of batches trained so far across + epochs. + + - log_interval: Print training loss if batch_idx % log_interval` is 0 + + - reset_interval: Reset statistics if batch_idx % reset_interval is 0 + + - valid_interval: Run validation if batch_idx % valid_interval is 0 + + - feature_dim: The model input dim. It has to match the one used + in computing features. + + - subsampling_factor: The subsampling factor for the model. + + - encoder_dim: Hidden dim for multi-head attention model. + + - num_decoder_layers: Number of decoder layer of transformer decoder. + + - warm_step: The warmup period that dictates the decay of the + scale on "simple" (un-pruned) loss. + """ + params = AttributeDict( + { + "best_train_loss": float("inf"), + "best_valid_loss": float("inf"), + "best_train_epoch": -1, + "best_valid_epoch": -1, + "batch_idx_train": 0, + "log_interval": 50, + "reset_interval": 200, + "valid_interval": 3000, # For the 100h subset, use 800 + # parameters for zipformer + "feature_dim": 80, + "subsampling_factor": 320, # not passed in, this is fixed. + # parameters for ctc loss + "beam_size": 10, + "use_double_scores": True, + "warm_step": 4000, + #"warm_step": 3000, + "env_info": get_env_info(), + } + ) + + return params + + +def get_encoder_model(params: AttributeDict) -> nn.Module: + # TODO: We can add an option to switch between Zipformer and Transformer + def to_int_tuple(s: str): + return tuple(map(int, s.split(","))) + + if params.encoder_type == 'd2v': + encoder = FairSeqData2VecEncoder( + input_size=params.encoder_dim, + w2v_url='None', + output_size=params.encoder_dim, + freeze_finetune_updates=params.freeze_finetune_updates, + additional_block=params.additional_block, + ) + else: + encoder = Zipformer( + num_features=params.feature_dim, + output_downsampling_factor=2, + zipformer_downsampling_factors=to_int_tuple( + params.zipformer_downsampling_factors + ), + encoder_dims=to_int_tuple(params.encoder_dims), + attention_dim=to_int_tuple(params.attention_dims), + encoder_unmasked_dims=to_int_tuple(params.encoder_unmasked_dims), + nhead=to_int_tuple(params.nhead), + feedforward_dim=to_int_tuple(params.feedforward_dims), + cnn_module_kernels=to_int_tuple(params.cnn_module_kernels), + num_encoder_layers=to_int_tuple(params.num_encoder_layers), + ) + + return encoder + + +def get_decoder_model(params: AttributeDict) -> nn.Module: + decoder = Decoder( + vocab_size=params.vocab_size, + decoder_dim=params.decoder_dim, + blank_id=params.blank_id, + context_size=params.context_size, + ) + return decoder + + +def get_joiner_model(params: AttributeDict) -> nn.Module: + joiner = Joiner( + encoder_dim=params.encoder_dim if params.encoder_type == 'd2v' else int(params.encoder_dims.split(",")[-1]), + decoder_dim=params.decoder_dim, + joiner_dim=params.joiner_dim, + vocab_size=params.vocab_size, + ) + return joiner + + +def get_transducer_model(params: AttributeDict) -> nn.Module: + encoder = get_encoder_model(params) + decoder = get_decoder_model(params) + joiner = get_joiner_model(params) + + model = Transducer( + encoder=encoder, + decoder=decoder, + joiner=joiner, + encoder_dim=params.encoder_dim if params.encoder_type == 'd2v' else int(params.encoder_dims.split(",")[-1]), + decoder_dim=params.decoder_dim, + joiner_dim=params.joiner_dim, + vocab_size=params.vocab_size, + ) + return model + + +def load_checkpoint_if_available( + params: AttributeDict, + model: nn.Module, + model_avg: nn.Module = None, + optimizer: Optional[torch.optim.Optimizer] = None, + scheduler: Optional[LRSchedulerType] = None, +) -> Optional[Dict[str, Any]]: + """Load checkpoint from file. + + If params.start_batch is positive, it will load the checkpoint from + `params.exp_dir/checkpoint-{params.start_batch}.pt`. Otherwise, if + params.start_epoch is larger than 1, it will load the checkpoint from + `params.start_epoch - 1`. + + Apart from loading state dict for `model` and `optimizer` it also updates + `best_train_epoch`, `best_train_loss`, `best_valid_epoch`, + and `best_valid_loss` in `params`. + + Args: + params: + The return value of :func:`get_params`. + model: + The training model. + model_avg: + The stored model averaged from the start of training. + optimizer: + The optimizer that we are using. + scheduler: + The scheduler that we are using. + Returns: + Return a dict containing previously saved training info. + """ + if params.start_batch > 0: + filename = params.exp_dir / f"checkpoint-{params.start_batch}.pt" + elif params.start_epoch > 1: + filename = params.exp_dir / f"epoch-{params.start_epoch-1}.pt" + elif params.add_adapter: + filename = params.exp_dir / f"d2v-base-T.pt" + else: + filename = params.exp_dir / f"../d2v-base-T.pt" + + assert filename.is_file(), f"{filename} does not exist!" + + saved_params = load_checkpoint( + filename, + model=model, + model_avg=model_avg, + optimizer=optimizer, + scheduler=scheduler, + strict=True if not params.add_adapter else False, + ) + + keys = [ + "best_train_epoch", + "best_valid_epoch", + "batch_idx_train", + "best_train_loss", + "best_valid_loss", + ] + for k in keys: + params[k] = saved_params[k] + + if params.start_batch > 0: + if "cur_epoch" in saved_params: + params["start_epoch"] = saved_params["cur_epoch"] + + if "cur_batch_idx" in saved_params: + params["cur_batch_idx"] = saved_params["cur_batch_idx"] + + #params.batch_idx_train = 0 + + return saved_params + + +def save_checkpoint( + params: AttributeDict, + model: Union[nn.Module, DDP], + model_avg: Optional[nn.Module] = None, + optimizer: Optional[torch.optim.Optimizer] = None, + scheduler: Optional[LRSchedulerType] = None, + sampler: Optional[CutSampler] = None, + scaler: Optional[GradScaler] = None, + rank: int = 0, +) -> None: + """Save model, optimizer, scheduler and training stats to file. + + Args: + params: + It is returned by :func:`get_params`. + model: + The training model. + model_avg: + The stored model averaged from the start of training. + optimizer: + The optimizer used in the training. + sampler: + The sampler for the training dataset. + scaler: + The scaler used for mix precision training. + """ + if rank != 0: + return + filename = params.exp_dir / f"epoch-{params.cur_epoch}.pt" + save_checkpoint_impl( + filename=filename, + model=model, + model_avg=model_avg, + params=params, + optimizer=optimizer, + scheduler=scheduler, + sampler=sampler, + scaler=scaler, + rank=rank, + ) + + if params.best_train_epoch == params.cur_epoch: + best_train_filename = params.exp_dir / "best-train-loss.pt" + copyfile(src=filename, dst=best_train_filename) + + if params.best_valid_epoch == params.cur_epoch: + best_valid_filename = params.exp_dir / "best-valid-loss.pt" + copyfile(src=filename, dst=best_valid_filename) + + +def compute_loss( + params: AttributeDict, + model: Union[nn.Module, DDP], + sp: spm.SentencePieceProcessor, + batch: dict, + is_training: bool, + decode: bool = False, +) -> Tuple[Tensor, MetricsTracker]: + """ + Compute transducer loss given the model and its inputs. + + Args: + params: + Parameters for training. See :func:`get_params`. + model: + The model for training. It is an instance of Zipformer in our case. + batch: + A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()` + for the content in it. + is_training: + True for training. False for validation. When it is True, this + function enables autograd during computation; when it is False, it + disables autograd. + warmup: a floating point value which increases throughout training; + values >= 1.0 are fully warmed up and have all modules present. + """ + device = model.device if isinstance(model, DDP) else next(model.parameters()).device + feature = batch["inputs"] + # at entry, feature is (N, T, C) + assert feature.ndim == 2 or feature.ndim == 3 + feature = feature.to(device) + + supervisions = batch["supervisions"] + + if feature.ndim == 2: + feature_lens = [] + for supervision in supervisions['cut']: + try: feature_lens.append(supervision.tracks[0].cut.recording.num_samples) + except: feature_lens.append(supervision.recording.num_samples) + feature_lens = torch.tensor(feature_lens) + + elif feature.ndim == 3: + feature_lens = supervisions["num_frames"].to(device) + + batch_idx_train = params.batch_idx_train + warm_step = params.warm_step + + texts = batch["supervisions"]["text"] + token_ids = sp.encode(texts, out_type=int) + y = k2.RaggedTensor(token_ids).to(device) + + with torch.set_grad_enabled(is_training): + simple_loss, pruned_loss, ctc_output = model( + x=feature, + x_lens=feature_lens, + y=y, + prune_range=params.prune_range, + am_scale=params.am_scale, + lm_scale=params.lm_scale, + ) + + s = params.simple_loss_scale + # take down the scale on the simple loss from 1.0 at the start + # to params.simple_loss scale by warm_step. + simple_loss_scale = ( + s + if batch_idx_train >= warm_step + else 1.0 - (batch_idx_train / warm_step) * (1.0 - s) + ) + pruned_loss_scale = ( + 1.0 + if batch_idx_train >= warm_step + else 0.1 + 0.9 * (batch_idx_train / warm_step) + ) + + loss = simple_loss_scale * simple_loss + pruned_loss_scale * pruned_loss + + info = MetricsTracker() + + if params.ctc_loss_scale > 0: + # Compute ctc loss + + # NOTE: We need `encode_supervisions` to sort sequences with + # different duration in decreasing order, required by + # `k2.intersect_dense` called in `k2.ctc_loss` + with warnings.catch_warnings(): + warnings.simplefilter("ignore") + supervision_segments, token_ids = encode_supervisions( + supervisions, + subsampling_factor=params.subsampling_factor, + token_ids=token_ids, + ) + + # Works with a BPE model + decoding_graph = k2.ctc_graph(token_ids, modified=False, device=device) + dense_fsa_vec = k2.DenseFsaVec( + ctc_output, + supervision_segments, + allow_truncate=params.subsampling_factor - 1, + ) + + ctc_loss = k2.ctc_loss( + decoding_graph=decoding_graph, + dense_fsa_vec=dense_fsa_vec, + output_beam=params.beam_size, + reduction="sum", + use_double_scores=params.use_double_scores, + ) + assert ctc_loss.requires_grad == is_training + loss += params.ctc_loss_scale * ctc_loss + + info["ctc_loss"] = ctc_loss.detach().cpu().item() + + assert loss.requires_grad == is_training + + if decode: + model.eval() + with torch.no_grad(): + hypos = model.module.decode( + x=feature, + x_lens=feature_lens, + y=y, + sp=sp + ) + logging.info(f'ref: {batch["supervisions"]["text"][0]}') + logging.info(f'hyp: {" ".join(hypos[0])}') + model.train() + + with warnings.catch_warnings(): + warnings.simplefilter("ignore") + info["frames"] = (feature_lens // params.subsampling_factor).sum().item() + + # Note: We use reduction=sum while computing the loss. + info["utterances"] = feature.size(0) + info["loss"] = loss.detach().cpu().item() + info["simple_loss"] = simple_loss.detach().cpu().item() + info["pruned_loss"] = pruned_loss.detach().cpu().item() + + return loss, info + + +def compute_validation_loss( + params: AttributeDict, + model: Union[nn.Module, DDP], + sp: spm.SentencePieceProcessor, + valid_dl: torch.utils.data.DataLoader, + world_size: int = 1, +) -> MetricsTracker: + """Run the validation process.""" + model.eval() + + tot_loss = MetricsTracker() + + for batch_idx, batch in enumerate(valid_dl): + loss, loss_info = compute_loss( + params=params, + model=model, + sp=sp, + batch=batch, + is_training=False, + ) + assert loss.requires_grad is False + tot_loss = tot_loss + loss_info + + if world_size > 1: + tot_loss.reduce(loss.device) + + loss_value = tot_loss["loss"] / tot_loss["utterances"] + if loss_value < params.best_valid_loss: + params.best_valid_epoch = params.cur_epoch + params.best_valid_loss = loss_value + + return tot_loss + + +def train_one_epoch( + params: AttributeDict, + model: Union[nn.Module, DDP], + optimizer: torch.optim.Optimizer or [torch.optim.Optimizer, torch.optim.Optimizer], + scheduler: LRSchedulerType or [LRSchedulerType, LRSchedulerType], + sp: spm.SentencePieceProcessor, + train_dl: torch.utils.data.DataLoader, + valid_dl: torch.utils.data.DataLoader, + scaler: GradScaler, + model_avg: Optional[nn.Module] = None, + tb_writer: Optional[SummaryWriter] = None, + world_size: int = 1, + rank: int = 0, + wb = None, +) -> None: + """Train the model for one epoch. + + The training loss from the mean of all frames is saved in + `params.train_loss`. It runs the validation process every + `params.valid_interval` batches. + + Args: + params: + It is returned by :func:`get_params`. + model: + The model for training. + optimizer: + The optimizer we are using. + scheduler: + The learning rate scheduler, we call step() every step. + train_dl: + Dataloader for the training dataset. + valid_dl: + Dataloader for the validation dataset. + scaler: + The scaler used for mix precision training. + model_avg: + The stored model averaged from the start of training. + tb_writer: + Writer to write log messages to tensorboard. + world_size: + Number of nodes in DDP training. If it is 1, DDP is disabled. + rank: + The rank of the node in DDP training. If no DDP is used, it should + be set to 0. + """ + model.train() + + tot_loss = MetricsTracker() + + cur_batch_idx = params.get("cur_batch_idx", 0) + + if params.multi_optim: + optimizer_enc, optimizer_dec = optimizer[0], optimizer[1] + scheduler_enc, scheduler_dec = scheduler[0], scheduler[1] + + for batch_idx, batch in enumerate(train_dl): + if params.batch_idx_train > params.num_updates: + break + if batch_idx < cur_batch_idx: + continue + cur_batch_idx = batch_idx + + if batch_idx % params.accum_grads == 0: params.batch_idx_train += 1 + batch_size = len(batch["supervisions"]["text"]) + + try: + with torch.cuda.amp.autocast(enabled=params.use_fp16): + loss, loss_info = compute_loss( + params=params, + model=model, + sp=sp, + batch=batch, + is_training=True, + decode = True if batch_idx % params.decode_interval == 0 else False, + ) + loss_info.reduce(loss.device) + + numel = params.world_size / (params.accum_grads * loss_info["utterances"]) + loss *= numel ## normalize loss over utts(batch size) + + # summary stats + tot_loss = (tot_loss * (1 - 1 / params.reset_interval)) + loss_info + + # NOTE: We use reduction==sum and loss is computed over utterances + # in the batch and there is no normalization to it so far. + scaler.scale(loss).backward() + + if params.multi_optim and (batch_idx+1) % params.accum_grads == 0: + set_batch_count(model, params.batch_idx_train) + scheduler_enc.step_batch(params.batch_idx_train) + scheduler_dec.step_batch(params.batch_idx_train) + scaler.step(optimizer_enc) + scaler.step(optimizer_dec) + scaler.update() + optimizer_enc.zero_grad() + optimizer_dec.zero_grad() + elif not params.multi_optim and (batch_idx+1) % params.accum_grads == 0: + set_batch_count(model, params.batch_idx_train) + scheduler.step_batch(params.batch_idx_train) + scaler.step(optimizer) + scaler.update() + optimizer.zero_grad() + + except: # noqa + display_and_save_batch(batch, params=params, sp=sp) + raise + + if params.print_diagnostics and batch_idx == 5: + return + + ''' + if ( + rank == 0 + and params.batch_idx_train > 0 + and params.batch_idx_train % params.average_period == 0 + ): + update_averaged_model( + params=params, + model_cur=model, + model_avg=model_avg, + ) + ''' + + if ( + params.batch_idx_train > 0 + and params.batch_idx_train % params.save_every_n == 0 + ): + params.cur_batch_idx = batch_idx + save_checkpoint_with_global_batch_idx( + out_dir=params.exp_dir, + global_batch_idx=params.batch_idx_train, + model=model, + model_avg=model_avg, + params=params, + optimizer=optimizer, + scheduler=scheduler, + sampler=train_dl.sampler, + scaler=scaler, + rank=rank, + ) + del params.cur_batch_idx + remove_checkpoints( + out_dir=params.exp_dir, + topk=params.keep_last_k, + rank=rank, + ) + + if batch_idx % 100 == 0 and params.use_fp16: + # If the grad scale was less than 1, try increasing it. The _growth_interval + # of the grad scaler is configurable, but we can't configure it to have different + # behavior depending on the current grad scale. + cur_grad_scale = scaler._scale.item() + ''' + if cur_grad_scale < 1.0 or (cur_grad_scale < 8.0 and batch_idx % 400 == 0): + scaler.update(cur_grad_scale * 2.0) + ''' + if cur_grad_scale < 0.01: + logging.warning(f"Grad scale is small: {cur_grad_scale}") + if cur_grad_scale < 1.0e-05: + wb.log({"valid/loss": 10000}) + raise RuntimeError( + f"grad_scale is too small, exiting: {cur_grad_scale}" + ) + + if params.batch_idx_train > 4000 and loss > 300 and params.wandb: + wb.log({"valid/loss": 10000}) + raise RunteimError( + f"divergence... exiting: loss={loss}" + ) + + if batch_idx % (params.log_interval*params.accum_grads) == 0: + if params.multi_optim: + cur_enc_lr = scheduler_enc.get_last_lr()[0] + cur_dec_lr = scheduler_dec.get_last_lr()[0] + cur_grad_scale = scaler._scale.item() if params.use_fp16 else 1.0 + + logging.info( + f"Epoch {params.cur_epoch}, " + f"batch {batch_idx}, loss[{loss_info}], " + f"tot_loss[{tot_loss}], batch size: {batch_size}, " + f"enc_lr: {cur_enc_lr:.2e}, " + f"dec_lr: {cur_dec_lr:.2e}, " + + (f"grad_scale: {scaler._scale.item()}" if params.use_fp16 else "") + ) + + else: + cur_lr = scheduler.get_last_lr()[0] + cur_grad_scale = scaler._scale.item() if params.use_fp16 else 1.0 + + logging.info( + f"Epoch {params.cur_epoch}, " + f"batch {batch_idx}, loss[{loss_info}], " + f"tot_loss[{tot_loss}], batch size: {batch_size}, " + f"lr: {cur_lr:.2e}, " + + (f"grad_scale: {scaler._scale.item()}" if params.use_fp16 else "") + ) + + if tb_writer is not None: + if params.multi_optim: + tb_writer.add_scalar( + "train/enc_learning_rate", cur_enc_lr, params.batch_idx_train + ) + tb_writer.add_scalar( + "train/dec_learning_rate", cur_dec_lr, params.batch_idx_train + ) + + else: + tb_writer.add_scalar( + "train/learning_rate", cur_lr, params.batch_idx_train + ) + + loss_info.write_summary( + tb_writer, "train/current_", params.batch_idx_train + ) + tot_loss.write_summary(tb_writer, "train/tot_", params.batch_idx_train) + if params.use_fp16: + tb_writer.add_scalar( + "train/grad_scale", + cur_grad_scale, + params.batch_idx_train, + ) + + if wb is not None and rank == 0: + wb.log({"train/loss": loss_info["loss"]*numel}) + wb.log({"train/simple_loss": loss_info["simple_loss"]*numel}) + wb.log({"train/pruned_loss": loss_info["pruned_loss"]*numel}) + wb.log({"train/ctc_loss": loss_info["ctc_loss"]*numel}) + + ''' + logging.info("Computing validation loss") + valid_info = compute_validation_loss( + params=params, + model=model, + sp=sp, + valid_dl=valid_dl, + world_size=world_size, + ) + model.train() + logging.info(f"Epoch {params.cur_epoch}, validation: {valid_info}") + logging.info( + f"Maximum memory allocated so far is {torch.cuda.max_memory_allocated()//1000000}MB" + ) + if tb_writer is not None: + valid_info.write_summary( + tb_writer, "train/valid_", params.batch_idx_train + ) + + if wb is not None and rank == 0: + numel = 1 / (params.accum_grads * valid_info["utterances"]) + wb.log({"valid/loss": valid_info["loss"]*numel}) + wb.log({"valid/simple_loss": valid_info["simple_loss"]*numel}) + wb.log({"valid/pruned_loss": valid_info["pruned_loss"]*numel}) + wb.log({"valid/ctc_loss": valid_info["ctc_loss"]*numel}) + ''' + + loss_value = tot_loss["loss"] / tot_loss["utterances"] + params.train_loss = loss_value + if params.train_loss < params.best_train_loss: + params.best_train_epoch = params.cur_epoch + params.best_train_loss = params.train_loss + +def run(rank, world_size, args, wb=None): + """ + Args: + rank: + It is a value between 0 and `world_size-1`, which is + passed automatically by `mp.spawn()` in :func:`main`. + The node with rank 0 is responsible for saving checkpoint. + world_size: + Number of GPUs for DDP training. + args: + The return value of get_parser().parse_args() + """ + params = get_params() + params.update(vars(args)) + #params.warm_step *= params.accum_grads + + fix_random_seed(params.seed) + if world_size > 1: + setup_dist(rank, world_size, params.master_port) + + setup_logger(f"{params.exp_dir}/log/log-train") + logging.info("Training started") + + if args.tensorboard and rank == 0: + tb_writer = SummaryWriter(log_dir=f"{params.exp_dir}/tensorboard") + else: + tb_writer = None + + device = torch.device("cpu") + if torch.cuda.is_available(): + device = torch.device("cuda", rank) + logging.info(f"Device: {device}") + + sp = spm.SentencePieceProcessor() + sp.load(params.bpe_model) + + # is defined in local/train_bpe_model.py + params.blank_id = sp.piece_to_id("") + params.vocab_size = sp.get_piece_size() + + logging.info(params) + + logging.info("About to create model") + model = get_transducer_model(params) + logging.info(model) + + num_param = sum([p.numel() for p in model.parameters()]) + logging.info(f"Number of model parameters: {num_param}") + + model_avg = None + assert params.start_epoch > 0, params.start_epoch + checkpoints = load_checkpoint_if_available( + params=params, model=model, model_avg=model_avg + ) + + model.to(device) + if world_size > 1: + logging.info("Using DDP") + model = DDP(model, device_ids=[rank], find_unused_parameters=True) + + if params.multi_optim: + logging.info("Using seperate optimizers over encoder, decoder ...") + + enc_param = [] + enc_names = [] + + dec_names = [] + dec_param = [] + + for n, p in model.named_parameters(): + name = n.split('.')[1] + if name == 'encoder' and 'feature_extractor' not in n: + enc_names.append(n) + enc_param.append(p) + elif 'ctc_output' in n: + enc_names.append(n) + enc_param.append(p) + elif 'feature_extractor' not in n: + dec_names.append(n) + dec_param.append(p) + + optimizer_enc = ScaledAdam( + enc_param, + lr=params.peak_enc_lr, + clipping_scale=None, + parameters_names=[enc_names], + ) + optimizer_dec = ScaledAdam( + dec_param, + lr=params.peak_dec_lr, + clipping_scale=5.0, + parameters_names=[dec_names], + ) + + scheduler_enc = Eden(optimizer_enc, params.lr_batches, params.lr_epochs) + scheduler_dec = Eden(optimizer_dec, params.lr_batches, params.lr_epochs) + optimizer = [optimizer_enc, optimizer_dec] + scheduler = [scheduler_enc, scheduler_dec] + + else: + parameters_names = [] + parameters_names.append( + [name_param_pair[0] for name_param_pair in model.named_parameters()] + ) + + logging.info(f"len name = {len(parameters_names)}") + logging.info(f"len param = {len(list(model.parameters()))}") + + optimizer = ScaledAdam( + model.parameters(), + lr=params.base_lr, + clipping_scale=2.0, + parameters_names=parameters_names, + ) + + scheduler = Eden(optimizer, params.lr_batches, params.lr_epochs) + + if checkpoints and ("optimizer" in checkpoints or "optimizer_enc" in checkpoints): + if params.multi_optim: + logging.info("Loading optimizer state dict") + optimizer_enc.load_state_dict(checkpoints["optimizer_enc"]) + optimizer_dec.load_state_dict(checkpoints["optimizer_dec"]) + + else: + logging.info("Loading optimizer state dict") + optimizer.load_state_dict(checkpoints["optimizer"]) + + if checkpoints: + if ( + params.multi_optim + and "scheduler_enc" in checkpoints + and checkpoints["scheduler_enc"] is not None + ): + logging.info("Loading enc/dec scheduler state dict") + scheduler_enc.load_state_dict(checkpoints["scheduler_enc"]) + scheduler_dec.load_state_dict(checkpoints["scheduler_dec"]) + else: + logging.info("Loading scheduler state dict") + scheduler.load_state_dict(checkpoints["scheduler"]) + + if params.print_diagnostics: + opts = diagnostics.TensorDiagnosticOptions( + 2**22 + ) # allow 4 megabytes per sub-module + diagnostic = diagnostics.attach_diagnostics(model, opts) + + if params.inf_check: + register_inf_check_hooks(model) + + librispeech = LibriSpeechAsrDataModule(args) + + train_cuts = librispeech.train_clean_100_cuts() + if params.full_libri: + train_cuts += librispeech.train_clean_360_cuts() + train_cuts += librispeech.train_other_500_cuts() + + def remove_short_and_long_utt(c: Cut): + # Keep only utterances with duration between 1 second and 20 seconds + # + # Caution: There is a reason to select 20.0 here. Please see + # ../local/display_manifest_statistics.py + # + # You should use ../local/display_manifest_statistics.py to get + # an utterance duration distribution for your dataset to select + # the threshold + return 1.0 <= c.duration <= 20.0 + + train_cuts = train_cuts.filter(remove_short_and_long_utt) + + if params.start_batch > 0 and checkpoints and "sampler" in checkpoints: + # We only load the sampler's state dict when it loads a checkpoint + # saved in the middle of an epoch + sampler_state_dict = checkpoints["sampler"] + else: + sampler_state_dict = None + + train_dl = librispeech.train_dataloaders( + train_cuts, sampler_state_dict=sampler_state_dict + ) + + valid_cuts = librispeech.dev_clean_cuts() + valid_cuts += librispeech.dev_other_cuts() + valid_dl = librispeech.valid_dataloaders(valid_cuts) + + ''' + if not params.print_diagnostics: + scan_pessimistic_batches_for_oom( + model=model, + train_dl=train_dl, + optimizer=optimizer, + sp=sp, + params=params, + ) + ''' + + scaler = GradScaler(enabled=params.use_fp16, init_scale=1.0) + if checkpoints and "grad_scaler" in checkpoints: + logging.info("Loading grad scaler state dict") + scaler.load_state_dict(checkpoints["grad_scaler"]) + + for epoch in range(params.start_epoch, params.num_epochs + 1): + if params.multi_optim: + scheduler_enc.step_epoch(epoch - 1) + scheduler_dec.step_epoch(epoch - 1) + else: + scheduler.step_epoch(epoch - 1) + fix_random_seed(params.seed + epoch - 1) + train_dl.sampler.set_epoch(epoch - 1) + + if tb_writer is not None: + tb_writer.add_scalar("train/epoch", epoch, params.batch_idx_train) + + params.cur_epoch = epoch + + train_one_epoch( + params=params, + model=model, + model_avg=model_avg, + optimizer=optimizer, + scheduler=scheduler, + sp=sp, + train_dl=train_dl, + valid_dl=valid_dl, + scaler=scaler, + tb_writer=tb_writer, + world_size=world_size, + rank=rank, + wb=wb, + ) + + if params.print_diagnostics: + diagnostic.print_diagnostics() + break + + ''' + save_checkpoint( + params=params, + model=model, + model_avg=model_avg, + optimizer=optimizer, + scheduler=scheduler, + sampler=train_dl.sampler, + scaler=scaler, + rank=rank, + ) + ''' + + logging.info("Done!") + + if world_size > 1: + torch.distributed.barrier() + cleanup_dist() + + +def run_adapter(rank, world_size, args, wb=None): + """ + Args: + rank: + It is a value between 0 and `world_size-1`, which is + passed automatically by `mp.spawn()` in :func:`main`. + The node with rank 0 is responsible for saving checkpoint. + world_size: + Number of GPUs for DDP training. + args: + The return value of get_parser().parse_args() + """ + params = get_params() + params.update(vars(args)) + + fix_random_seed(params.seed) + if world_size > 1: + setup_dist(rank, world_size, params.master_port) + + setup_logger(f"{params.exp_dir}/log/log-train") + logging.info("Training started") + + if args.tensorboard and rank == 0: + tb_writer = SummaryWriter(log_dir=f"{params.exp_dir}/tensorboard") + else: + tb_writer = None + + device = torch.device("cpu") + if torch.cuda.is_available(): + device = torch.device("cuda", rank) + logging.info(f"Device: {device}") + + sp = spm.SentencePieceProcessor() + sp.load(params.bpe_model) + + # is defined in local/train_bpe_model.py + params.blank_id = sp.piece_to_id("") + params.vocab_size = sp.get_piece_size() + + logging.info(params) + + logging.info("About to create model") + model = get_transducer_model(params) + + num_param = sum([p.numel() if p.requires_grad else 0 for p in model.parameters()]) + logging.info(f"Number of model parameters: {num_param}") + + assert params.save_every_n >= params.average_period + model_avg: Optional[nn.Module] = None + if rank == 0: + # model_avg is only used with rank 0 + model_avg = copy.deepcopy(model).to(torch.float64) + + assert params.start_epoch > 0, params.start_epoch + checkpoints = load_checkpoint_if_available( + params=params, model=model, model_avg=model_avg + ) + + model.to(device) + if world_size > 1: + logging.info("Using DDP") + model = DDP(model, device_ids=[rank], find_unused_parameters=True) + + adapter_names = [] + adapter_param = [] + for n, p in model.named_parameters(): + if 'adapters' in n: + adapter_names.append(n) + adapter_param.append(p) + #else: + # p.requires_grad = False + optimizer_adapter = ScaledAdam( + adapter_param, + lr=params.adapter_lr, + clipping_scale=5.0, + parameters_names=[adapter_names], + ) + scheduler_adapter = Eden(optimizer_adapter, 5000, 3.5) #params.lr_batche, params.lr_epochs) + + optimizer, scheduler = optimizer_adapter, scheduler_adapter + + librispeech = LibriSpeechAsrDataModule(args) + + ''' + train_cuts = librispeech.train_clean_100_cuts() + if params.full_libri: + train_cuts += librispeech.train_clean_360_cuts() + train_cuts += librispeech.train_other_500_cuts() + ''' + train_cuts = librispeech.vox_cuts(option=params.spk_id) + + def remove_short_and_long_utt(c: Cut): + return 1.0 <= c.duration <= 20.0 + + train_cuts = train_cuts.filter(remove_short_and_long_utt) + + sampler_state_dict = None + + train_dl = librispeech.train_dataloaders( + train_cuts, sampler_state_dict=sampler_state_dict + ) + + valid_cuts = librispeech.dev_clean_cuts() + valid_cuts += librispeech.dev_other_cuts() + valid_dl = librispeech.valid_dataloaders(valid_cuts) + + scaler = GradScaler(enabled=params.use_fp16, init_scale=1.0) + + for epoch in range(params.start_epoch, params.num_epochs + 1): + scheduler.step_epoch(epoch - 1) + fix_random_seed(params.seed + epoch - 1) + train_dl.sampler.set_epoch(epoch - 1) + + if tb_writer is not None: + tb_writer.add_scalar("train/epoch", epoch, params.batch_idx_train) + + params.cur_epoch = epoch + + train_one_epoch( + params=params, + model=model, + model_avg=model_avg, + optimizer=optimizer, + scheduler=scheduler, + sp=sp, + train_dl=train_dl, + valid_dl=valid_dl, + scaler=scaler, + tb_writer=tb_writer, + world_size=world_size, + rank=rank, + wb=wb, + ) + + if params.print_diagnostics: + diagnostic.print_diagnostics() + break + + save_checkpoint( + params=params, + model=model, + model_avg=model_avg, + optimizer=optimizer, + scheduler=scheduler, + sampler=train_dl.sampler, + scaler=scaler, + rank=rank, + ) + + logging.info("Done!") + + if world_size > 1: + torch.distributed.barrier() + cleanup_dist() + + +def display_and_save_batch( + batch: dict, + params: AttributeDict, + sp: spm.SentencePieceProcessor, +) -> None: + """Display the batch statistics and save the batch into disk. + + Args: + batch: + A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()` + for the content in it. + params: + Parameters for training. See :func:`get_params`. + sp: + The BPE model. + """ + from lhotse.utils import uuid4 + + filename = f"{params.exp_dir}/batch-{uuid4()}.pt" + logging.info(f"Saving batch to {filename}") + torch.save(batch, filename) + + supervisions = batch["supervisions"] + features = batch["inputs"] + + logging.info(f"features shape: {features.shape}") + + y = sp.encode(supervisions["text"], out_type=int) + num_tokens = sum(len(i) for i in y) + logging.info(f"num tokens: {num_tokens}") + + +def scan_pessimistic_batches_for_oom( + model: Union[nn.Module, DDP], + train_dl: torch.utils.data.DataLoader, + optimizer: torch.optim.Optimizer, + sp: spm.SentencePieceProcessor, + params: AttributeDict, +): + from lhotse.dataset import find_pessimistic_batches + + logging.info( + "Sanity check -- see if any of the batches in epoch 1 would cause OOM." + ) + batches, crit_values = find_pessimistic_batches(train_dl.sampler) + for criterion, cuts in batches.items(): + batch = train_dl.dataset[cuts] + try: + with torch.cuda.amp.autocast(enabled=params.use_fp16): + loss, _ = compute_loss( + params=params, + model=model, + sp=sp, + batch=batch, + is_training=True, + ) + loss.backward() + optimizer.zero_grad() + except Exception as e: + if "CUDA out of memory" in str(e): + logging.error( + "Your GPU ran out of memory with the current " + "max_duration setting. We recommend decreasing " + "max_duration and trying again.\n" + f"Failing criterion: {criterion} " + f"(={crit_values[criterion]}) ..." + ) + display_and_save_batch(batch, params=params, sp=sp) + raise + logging.info( + f"Maximum memory allocated so far is {torch.cuda.max_memory_allocated()//1000000}MB" + ) + + +def main(): + parser = get_parser() + LibriSpeechAsrDataModule.add_arguments(parser) + args = parser.parse_args() + #args.exp_dir = args.exp_dir + str(random.randint(0,400)) + args.exp_dir = Path(args.exp_dir) + + logging.info("save arguments to config.yaml...") + save_args(args) + + if args.wandb: wb = wandb.init(project="d2v-T", entity="dohe0342", config=vars(args)) + else: wb = None + + world_size = args.world_size + assert world_size >= 1 + if world_size > 1: + mp.spawn(run if not args.add_adapter else run_adapter, + args=(world_size, args, wb), + nprocs=world_size, + join=True + ) + else: + if not args.add_adapter: run(rank=0, world_size=1, args=args, wb=wb) + else: run(rank=0, world_size=1, args=args, wb=wb) + +torch.set_num_threads(1) +torch.set_num_interop_threads(1) + +if __name__ == "__main__": + main() diff --git a/egs/tedlium3/ASR/pruned_transducer_stateless_d2v_v2/jit_pretrained.py b/egs/tedlium3/ASR/pruned_transducer_stateless_d2v_v2/jit_pretrained.py new file mode 100755 index 000000000..280b95984 --- /dev/null +++ b/egs/tedlium3/ASR/pruned_transducer_stateless_d2v_v2/jit_pretrained.py @@ -0,0 +1,271 @@ +#!/usr/bin/env python3 +# Copyright 2022 Xiaomi Corp. (authors: Fangjun Kuang) +# +# See ../../../../LICENSE for clarification regarding multiple authors +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +""" +This script loads torchscript models, exported by `torch.jit.script()` +and uses them to decode waves. +You can use the following command to get the exported models: + +./pruned_transducer_stateless7_ctc/export.py \ + --exp-dir ./pruned_transducer_stateless7_ctc/exp \ + --bpe-model data/lang_bpe_500/bpe.model \ + --epoch 20 \ + --avg 10 \ + --jit 1 + +Usage of this script: + +./pruned_transducer_stateless7_ctc/jit_pretrained.py \ + --nn-model-filename ./pruned_transducer_stateless7_ctc/exp/cpu_jit.pt \ + /path/to/foo.wav \ + /path/to/bar.wav +""" + +import argparse +import logging +import math +from typing import List + +import kaldifeat +import sentencepiece as spm +import torch +import torchaudio +from torch.nn.utils.rnn import pad_sequence + + +def get_parser(): + parser = argparse.ArgumentParser( + formatter_class=argparse.ArgumentDefaultsHelpFormatter + ) + + parser.add_argument( + "--nn-model-filename", + type=str, + required=True, + help="Path to the torchscript model cpu_jit.pt", + ) + + parser.add_argument( + "--bpe-model", + type=str, + help="""Path to bpe.model.""", + ) + + parser.add_argument( + "sound_files", + type=str, + nargs="+", + help="The input sound file(s) to transcribe. " + "Supported formats are those supported by torchaudio.load(). " + "For example, wav and flac are supported. " + "The sample rate has to be 16kHz.", + ) + + return parser + + +def read_sound_files( + filenames: List[str], expected_sample_rate: float = 16000 +) -> List[torch.Tensor]: + """Read a list of sound files into a list 1-D float32 torch tensors. + Args: + filenames: + A list of sound filenames. + expected_sample_rate: + The expected sample rate of the sound files. + Returns: + Return a list of 1-D float32 torch tensors. + """ + ans = [] + for f in filenames: + wave, sample_rate = torchaudio.load(f) + assert ( + sample_rate == expected_sample_rate + ), f"Expected sample rate: {expected_sample_rate}. Given: {sample_rate}" + # We use only the first channel + ans.append(wave[0]) + return ans + + +def greedy_search( + model: torch.jit.ScriptModule, + encoder_out: torch.Tensor, + encoder_out_lens: torch.Tensor, +) -> List[List[int]]: + """Greedy search in batch mode. It hardcodes --max-sym-per-frame=1. + Args: + model: + The transducer model. + encoder_out: + A 3-D tensor of shape (N, T, C) + encoder_out_lens: + A 1-D tensor of shape (N,). + Returns: + Return the decoded results for each utterance. + """ + assert encoder_out.ndim == 3 + assert encoder_out.size(0) >= 1, encoder_out.size(0) + + packed_encoder_out = torch.nn.utils.rnn.pack_padded_sequence( + input=encoder_out, + lengths=encoder_out_lens.cpu(), + batch_first=True, + enforce_sorted=False, + ) + + device = encoder_out.device + blank_id = 0 # hard-code to 0 + + batch_size_list = packed_encoder_out.batch_sizes.tolist() + N = encoder_out.size(0) + + assert torch.all(encoder_out_lens > 0), encoder_out_lens + assert N == batch_size_list[0], (N, batch_size_list) + + context_size = model.decoder.context_size + hyps = [[blank_id] * context_size for _ in range(N)] + + decoder_input = torch.tensor( + hyps, + device=device, + dtype=torch.int64, + ) # (N, context_size) + + decoder_out = model.decoder( + decoder_input, + need_pad=torch.tensor([False]), + ).squeeze(1) + + offset = 0 + for batch_size in batch_size_list: + start = offset + end = offset + batch_size + current_encoder_out = packed_encoder_out.data[start:end] + current_encoder_out = current_encoder_out + # current_encoder_out's shape: (batch_size, encoder_out_dim) + offset = end + + decoder_out = decoder_out[:batch_size] + + logits = model.joiner( + current_encoder_out, + decoder_out, + ) + # logits'shape (batch_size, vocab_size) + + assert logits.ndim == 2, logits.shape + y = logits.argmax(dim=1).tolist() + emitted = False + for i, v in enumerate(y): + if v != blank_id: + hyps[i].append(v) + emitted = True + if emitted: + # update decoder output + decoder_input = [h[-context_size:] for h in hyps[:batch_size]] + decoder_input = torch.tensor( + decoder_input, + device=device, + dtype=torch.int64, + ) + decoder_out = model.decoder( + decoder_input, + need_pad=torch.tensor([False]), + ) + decoder_out = decoder_out.squeeze(1) + + sorted_ans = [h[context_size:] for h in hyps] + ans = [] + unsorted_indices = packed_encoder_out.unsorted_indices.tolist() + for i in range(N): + ans.append(sorted_ans[unsorted_indices[i]]) + + return ans + + +@torch.no_grad() +def main(): + parser = get_parser() + args = parser.parse_args() + logging.info(vars(args)) + + device = torch.device("cpu") + if torch.cuda.is_available(): + device = torch.device("cuda", 0) + + logging.info(f"device: {device}") + + model = torch.jit.load(args.nn_model_filename) + + model.eval() + + model.to(device) + + sp = spm.SentencePieceProcessor() + sp.load(args.bpe_model) + + logging.info("Constructing Fbank computer") + opts = kaldifeat.FbankOptions() + opts.device = device + opts.frame_opts.dither = 0 + opts.frame_opts.snip_edges = False + opts.frame_opts.samp_freq = 16000 + opts.mel_opts.num_bins = 80 + + fbank = kaldifeat.Fbank(opts) + + logging.info(f"Reading sound files: {args.sound_files}") + waves = read_sound_files( + filenames=args.sound_files, + ) + waves = [w.to(device) for w in waves] + + logging.info("Decoding started") + features = fbank(waves) + feature_lengths = [f.size(0) for f in features] + + features = pad_sequence( + features, + batch_first=True, + padding_value=math.log(1e-10), + ) + + feature_lengths = torch.tensor(feature_lengths, device=device) + + encoder_out, encoder_out_lens = model.encoder( + x=features, + x_lens=feature_lengths, + ) + + hyps = greedy_search( + model=model, + encoder_out=encoder_out, + encoder_out_lens=encoder_out_lens, + ) + s = "\n" + for filename, hyp in zip(args.sound_files, hyps): + words = sp.decode(hyp) + s += f"{filename}:\n{words}\n\n" + logging.info(s) + + logging.info("Decoding Done") + + +if __name__ == "__main__": + formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s" + + logging.basicConfig(format=formatter, level=logging.INFO) + main() diff --git a/egs/tedlium3/ASR/pruned_transducer_stateless_d2v_v2/jit_pretrained_ctc.py b/egs/tedlium3/ASR/pruned_transducer_stateless_d2v_v2/jit_pretrained_ctc.py new file mode 100755 index 000000000..d3343d34a --- /dev/null +++ b/egs/tedlium3/ASR/pruned_transducer_stateless_d2v_v2/jit_pretrained_ctc.py @@ -0,0 +1,423 @@ +#!/usr/bin/env python3 +# Copyright 2022 Xiaomi Corp. (authors: Fangjun Kuang, +# Zengwei Yao) +# +# See ../../../../LICENSE for clarification regarding multiple authors +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +""" +This script loads torchscript models, exported by `torch.jit.script()` +and uses them to decode waves. +You can use the following command to get the exported models: + +./pruned_transducer_stateless7_ctc/export.py \ + --exp-dir ./pruned_transducer_stateless7_ctc/exp \ + --bpe-model data/lang_bpe_500/bpe.model \ + --epoch 20 \ + --avg 10 \ + --jit 1 + +Usage of this script: + +(1) ctc-decoding +./pruned_transducer_stateless7_ctc/jit_pretrained_ctc.py \ + --nn-model-filename ./pruned_transducer_stateless7_ctc/exp/cpu_jit.pt \ + --bpe-model data/lang_bpe_500/bpe.model \ + --method ctc-decoding \ + --sample-rate 16000 \ + /path/to/foo.wav \ + /path/to/bar.wav + +(2) 1best +./pruned_transducer_stateless7_ctc/jit_pretrained_ctc.py \ + --nn-model-filename ./pruned_transducer_stateless7_ctc/exp/cpu_jit.pt \ + --HLG data/lang_bpe_500/HLG.pt \ + --words-file data/lang_bpe_500/words.txt \ + --method 1best \ + --sample-rate 16000 \ + /path/to/foo.wav \ + /path/to/bar.wav + + +(3) nbest-rescoring +./pruned_transducer_stateless7_ctc/jit_pretrained_ctc.py \ + --nn-model-filename ./pruned_transducer_stateless7_ctc/exp/cpu_jit.pt \ + --HLG data/lang_bpe_500/HLG.pt \ + --words-file data/lang_bpe_500/words.txt \ + --G data/lm/G_4_gram.pt \ + --method nbest-rescoring \ + --sample-rate 16000 \ + /path/to/foo.wav \ + /path/to/bar.wav + + +(4) whole-lattice-rescoring +./pruned_transducer_stateless7_ctc/jit_pretrained_ctc.py \ + --nn-model-filename ./pruned_transducer_stateless7_ctc/exp/cpu_jit.pt \ + --HLG data/lang_bpe_500/HLG.pt \ + --words-file data/lang_bpe_500/words.txt \ + --G data/lm/G_4_gram.pt \ + --method whole-lattice-rescoring \ + --sample-rate 16000 \ + /path/to/foo.wav \ + /path/to/bar.wav +""" + +import argparse +import logging +import math +from typing import List + +import k2 +import kaldifeat +import sentencepiece as spm +import torch +import torchaudio +from ctc_decode import get_decoding_params +from torch.nn.utils.rnn import pad_sequence +from train import get_params + +from icefall.decode import ( + get_lattice, + one_best_decoding, + rescore_with_n_best_list, + rescore_with_whole_lattice, +) +from icefall.utils import get_texts + + +def get_parser(): + parser = argparse.ArgumentParser( + formatter_class=argparse.ArgumentDefaultsHelpFormatter + ) + + parser.add_argument( + "--model-filename", + type=str, + required=True, + help="Path to the torchscript model.", + ) + + parser.add_argument( + "--words-file", + type=str, + help="""Path to words.txt. + Used only when method is not ctc-decoding. + """, + ) + + parser.add_argument( + "--HLG", + type=str, + help="""Path to HLG.pt. + Used only when method is not ctc-decoding. + """, + ) + + parser.add_argument( + "--bpe-model", + type=str, + help="""Path to bpe.model. + Used only when method is ctc-decoding. + """, + ) + + parser.add_argument( + "--method", + type=str, + default="1best", + help="""Decoding method. + Possible values are: + (0) ctc-decoding - Use CTC decoding. It uses a sentence + piece model, i.e., lang_dir/bpe.model, to convert + word pieces to words. It needs neither a lexicon + nor an n-gram LM. + (1) 1best - Use the best path as decoding output. Only + the transformer encoder output is used for decoding. + We call it HLG decoding. + (2) nbest-rescoring. Extract n paths from the decoding lattice, + rescore them with an LM, the path with + the highest score is the decoding result. + We call it HLG decoding + n-gram LM rescoring. + (3) whole-lattice-rescoring - Use an LM to rescore the + decoding lattice and then use 1best to decode the + rescored lattice. + We call it HLG decoding + n-gram LM rescoring. + """, + ) + + parser.add_argument( + "--G", + type=str, + help="""An LM for rescoring. + Used only when method is + whole-lattice-rescoring or nbest-rescoring. + It's usually a 4-gram LM. + """, + ) + + parser.add_argument( + "--num-paths", + type=int, + default=100, + help=""" + Used only when method is attention-decoder. + It specifies the size of n-best list.""", + ) + + parser.add_argument( + "--ngram-lm-scale", + type=float, + default=1.3, + help=""" + Used only when method is whole-lattice-rescoring and nbest-rescoring. + It specifies the scale for n-gram LM scores. + (Note: You need to tune it on a dataset.) + """, + ) + + parser.add_argument( + "--nbest-scale", + type=float, + default=0.5, + help=""" + Used only when method is nbest-rescoring. + It specifies the scale for lattice.scores when + extracting n-best lists. A smaller value results in + more unique number of paths with the risk of missing + the best path. + """, + ) + + parser.add_argument( + "--num-classes", + type=int, + default=500, + help=""" + Vocab size in the BPE model. + """, + ) + + parser.add_argument( + "--sample-rate", + type=int, + default=16000, + help="The sample rate of the input sound file", + ) + + parser.add_argument( + "sound_files", + type=str, + nargs="+", + help="The input sound file(s) to transcribe. " + "Supported formats are those supported by torchaudio.load(). " + "For example, wav and flac are supported. " + "The sample rate has to be 16kHz.", + ) + + return parser + + +def read_sound_files( + filenames: List[str], expected_sample_rate: float = 16000 +) -> List[torch.Tensor]: + """Read a list of sound files into a list 1-D float32 torch tensors. + Args: + filenames: + A list of sound filenames. + expected_sample_rate: + The expected sample rate of the sound files. + Returns: + Return a list of 1-D float32 torch tensors. + """ + ans = [] + for f in filenames: + wave, sample_rate = torchaudio.load(f) + assert ( + sample_rate == expected_sample_rate + ), f"Expected sample rate: {expected_sample_rate}. Given: {sample_rate}" + # We use only the first channel + ans.append(wave[0]) + return ans + + +@torch.no_grad() +def main(): + parser = get_parser() + args = parser.parse_args() + + params = get_params() + # add decoding params + params.update(get_decoding_params()) + params.update(vars(args)) + + logging.info(f"{params}") + + device = torch.device("cpu") + if torch.cuda.is_available(): + device = torch.device("cuda", 0) + + logging.info(f"device: {device}") + + model = torch.jit.load(args.model_filename) + model.to(device) + model.eval() + + logging.info("Constructing Fbank computer") + opts = kaldifeat.FbankOptions() + opts.device = device + opts.frame_opts.dither = 0 + opts.frame_opts.snip_edges = False + opts.frame_opts.samp_freq = params.sample_rate + opts.mel_opts.num_bins = params.feature_dim + + fbank = kaldifeat.Fbank(opts) + + logging.info(f"Reading sound files: {params.sound_files}") + waves = read_sound_files( + filenames=params.sound_files, expected_sample_rate=params.sample_rate + ) + waves = [w.to(device) for w in waves] + + logging.info("Decoding started") + features = fbank(waves) + feature_lengths = [f.size(0) for f in features] + + features = pad_sequence(features, batch_first=True, padding_value=math.log(1e-10)) + feature_lengths = torch.tensor(feature_lengths, device=device) + + encoder_out, encoder_out_lens = model.encoder( + x=features, + x_lens=feature_lengths, + ) + nnet_output = model.ctc_output(encoder_out) + + batch_size = nnet_output.shape[0] + supervision_segments = torch.tensor( + [[i, 0, nnet_output.shape[1]] for i in range(batch_size)], + dtype=torch.int32, + ) + + if params.method == "ctc-decoding": + logging.info("Use CTC decoding") + bpe_model = spm.SentencePieceProcessor() + bpe_model.load(params.bpe_model) + max_token_id = params.num_classes - 1 + + H = k2.ctc_topo( + max_token=max_token_id, + modified=False, + device=device, + ) + + lattice = get_lattice( + nnet_output=nnet_output, + decoding_graph=H, + supervision_segments=supervision_segments, + search_beam=params.search_beam, + output_beam=params.output_beam, + min_active_states=params.min_active_states, + max_active_states=params.max_active_states, + subsampling_factor=params.subsampling_factor, + ) + + best_path = one_best_decoding( + lattice=lattice, use_double_scores=params.use_double_scores + ) + token_ids = get_texts(best_path) + hyps = bpe_model.decode(token_ids) + hyps = [s.split() for s in hyps] + elif params.method in [ + "1best", + "nbest-rescoring", + "whole-lattice-rescoring", + ]: + logging.info(f"Loading HLG from {params.HLG}") + HLG = k2.Fsa.from_dict(torch.load(params.HLG, map_location="cpu")) + HLG = HLG.to(device) + if not hasattr(HLG, "lm_scores"): + # For whole-lattice-rescoring and attention-decoder + HLG.lm_scores = HLG.scores.clone() + + if params.method in [ + "nbest-rescoring", + "whole-lattice-rescoring", + ]: + logging.info(f"Loading G from {params.G}") + G = k2.Fsa.from_dict(torch.load(params.G, map_location="cpu")) + G = G.to(device) + if params.method == "whole-lattice-rescoring": + # Add epsilon self-loops to G as we will compose + # it with the whole lattice later + G = k2.add_epsilon_self_loops(G) + G = k2.arc_sort(G) + + # G.lm_scores is used to replace HLG.lm_scores during + # LM rescoring. + G.lm_scores = G.scores.clone() + + lattice = get_lattice( + nnet_output=nnet_output, + decoding_graph=HLG, + supervision_segments=supervision_segments, + search_beam=params.search_beam, + output_beam=params.output_beam, + min_active_states=params.min_active_states, + max_active_states=params.max_active_states, + subsampling_factor=params.subsampling_factor, + ) + + if params.method == "1best": + logging.info("Use HLG decoding") + best_path = one_best_decoding( + lattice=lattice, use_double_scores=params.use_double_scores + ) + if params.method == "nbest-rescoring": + logging.info("Use HLG decoding + LM rescoring") + best_path_dict = rescore_with_n_best_list( + lattice=lattice, + G=G, + num_paths=params.num_paths, + lm_scale_list=[params.ngram_lm_scale], + nbest_scale=params.nbest_scale, + ) + best_path = next(iter(best_path_dict.values())) + elif params.method == "whole-lattice-rescoring": + logging.info("Use HLG decoding + LM rescoring") + best_path_dict = rescore_with_whole_lattice( + lattice=lattice, + G_with_epsilon_loops=G, + lm_scale_list=[params.ngram_lm_scale], + ) + best_path = next(iter(best_path_dict.values())) + + hyps = get_texts(best_path) + word_sym_table = k2.SymbolTable.from_file(params.words_file) + hyps = [[word_sym_table[i] for i in ids] for ids in hyps] + else: + raise ValueError(f"Unsupported decoding method: {params.method}") + + s = "\n" + for filename, hyp in zip(params.sound_files, hyps): + words = " ".join(hyp) + s += f"{filename}:\n{words}\n\n" + logging.info(s) + + logging.info("Decoding Done") + + +if __name__ == "__main__": + formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s" + + logging.basicConfig(format=formatter, level=logging.INFO) + main() diff --git a/egs/tedlium3/ASR/pruned_transducer_stateless_d2v_v2/joiner.py b/egs/tedlium3/ASR/pruned_transducer_stateless_d2v_v2/joiner.py new file mode 100644 index 000000000..3ddac2cf2 --- /dev/null +++ b/egs/tedlium3/ASR/pruned_transducer_stateless_d2v_v2/joiner.py @@ -0,0 +1,65 @@ +# Copyright 2021 Xiaomi Corp. (authors: Fangjun Kuang) +# +# See ../../../../LICENSE for clarification regarding multiple authors +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import torch +import torch.nn as nn + + +class Joiner(nn.Module): + def __init__( + self, + encoder_dim: int, + decoder_dim: int, + joiner_dim: int, + vocab_size: int, + ): + super().__init__() + + self.encoder_proj = nn.Linear(encoder_dim, joiner_dim) + self.decoder_proj = nn.Linear(decoder_dim, joiner_dim) + self.output_linear = nn.Linear(joiner_dim, vocab_size) + + def forward( + self, + encoder_out: torch.Tensor, + decoder_out: torch.Tensor, + project_input: bool = True, + ) -> torch.Tensor: + """ + Args: + encoder_out: + Output from the encoder. Its shape is (N, T, s_range, C). + decoder_out: + Output from the decoder. Its shape is (N, T, s_range, C). + project_input: + If true, apply input projections encoder_proj and decoder_proj. + If this is false, it is the user's responsibility to do this + manually. + Returns: + Return a tensor of shape (N, T, s_range, C). + """ + assert encoder_out.ndim == decoder_out.ndim + assert encoder_out.ndim in (2, 4) + assert encoder_out.shape[:-1] == decoder_out.shape[:-1] + + if project_input: + logit = self.encoder_proj(encoder_out) + self.decoder_proj(decoder_out) + else: + logit = encoder_out + decoder_out + + logit = self.output_linear(torch.tanh(logit)) + + return logit diff --git a/egs/tedlium3/ASR/pruned_transducer_stateless_d2v_v2/last_layer.py b/egs/tedlium3/ASR/pruned_transducer_stateless_d2v_v2/last_layer.py new file mode 100755 index 000000000..fe04cc916 --- /dev/null +++ b/egs/tedlium3/ASR/pruned_transducer_stateless_d2v_v2/last_layer.py @@ -0,0 +1,1836 @@ +#!/usr/bin/env python3 +# Copyright 2021-2022 Xiaomi Corp. (authors: Fangjun Kuang, +# Wei Kang, +# Mingshuang Luo,) +# Zengwei Yao) +# +# See ../../../../LICENSE for clarification regarding multiple authors +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +""" +Usage: + +export CUDA_VISIBLE_DEVICES="0,1,2,3" + +./pruned_transducer_stateless7_ctc/train.py \ + --world-size 4 \ + --num-epochs 30 \ + --start-epoch 1 \ + --exp-dir pruned_transducer_stateless7_ctc/exp \ + --full-libri 1 \ + --max-duration 300 + +# For mix precision training: + +./pruned_transducer_stateless7_ctc/train.py \ + --world-size 4 \ + --num-epochs 30 \ + --start-epoch 1 \ + --use-fp16 1 \ + --exp-dir pruned_transducer_stateless7_ctc/exp \ + --full-libri 1 \ + --max-duration 550 + +# For d2v-T training: +export CUDA_VISIBLE_DEVICES="0,1,2,3,4,5,6,7" + +./pruned_transducer_stateless_d2v_v2/train.py \ + --wandb true \ + --input-strategy AudioSamples \ + --enable-spec-aug False \ + --multi-optim True \ + --world-size 8 \ + --num-epochs 30 \ + --start-epoch 1 \ + --full-libri 0 \ + --exp-dir ./pruned_transducer_stateless_d2v_v2/$1 \ + --max-duration 250 \ + --freeze-finetune-updates 2000 \ + --use-fp16 1 \ + --peak-enc-lr 0.001 \ + --peak-dec-lr 0.05 \ + --accum-grads 1 \ + --encoder-type d2v \ + --additional-block True \ + --encoder-dim 768 \ + --decoder-dim 768 \ + --joiner-dim 768 \ + --prune-range 20 \ + --context-size 2 \ + --ctc-loss-scale 0.2 + +""" + + +import random +import argparse +import copy +import logging +import warnings +from pathlib import Path +from shutil import copyfile +from typing import Any, Dict, Optional, Tuple, Union + +import k2 +import optim +import sentencepiece as spm +import torch +import torch.multiprocessing as mp +import torch.nn as nn +from asr_datamodule import LibriSpeechAsrDataModule +from decoder import Decoder +from joiner import Joiner +from lhotse.cut import Cut +from lhotse.dataset.sampling.base import CutSampler +from lhotse.utils import fix_random_seed +from model import Transducer +from optim import Eden, ScaledAdam +from torch import Tensor +from torch.cuda.amp import GradScaler +from torch.nn.parallel import DistributedDataParallel as DDP +from torch.utils.tensorboard import SummaryWriter +from zipformer import Zipformer +from data2vec_encoder import FairSeqData2VecEncoder + +from icefall import diagnostics +from icefall.checkpoint import remove_checkpoints +from icefall.checkpoint import update_averaged_model +from checkpoint import ( + save_checkpoint as save_checkpoint_impl, + save_checkpoint_with_global_batch_idx, + load_checkpoint +) +from icefall.dist import cleanup_dist, setup_dist +from icefall.env import get_env_info +from icefall.hooks import register_inf_check_hooks +from icefall.utils import ( + AttributeDict, + MetricsTracker, + encode_supervisions, + setup_logger, + str2bool, + save_args, +) + +import wandb + +#from icefall.checkpoint import save_checkpoint as save_checkpoint_impl +LRSchedulerType = Union[torch.optim.lr_scheduler._LRScheduler, optim.LRScheduler] + + +def set_batch_count(model: Union[nn.Module, DDP], batch_count: float) -> None: + if isinstance(model, DDP): + # get underlying nn.Module + model = model.module + for module in model.modules(): + if hasattr(module, "batch_count"): + module.batch_count = batch_count + model.encoder.num_updates = int(batch_count) + + +def add_adapter_arguments(parser: argparse.ArgumentParser): + parser.add_argument( + "--add-adapter", + type=str2bool, + default=False, + help="add adapter to rep model's encoder" + ) + + parser.add_argument( + "--adapter-lr", + type=float, + default=0.0001, + help="adapter learning rate" + ) + + parser.add_argument( + "--gender", + type=str, + default='male', + help="select gender" + ) + + +def add_rep_arguments(parser: argparse.ArgumentParser): + parser.add_argument( + "--wandb", + type=str2bool, + default=True, + help="Use wandb for MLOps", + ) + parser.add_argument( + "--hpo", + type=str2bool, + default=False, + help="Use small db for HPO", + ) + + parser.add_argument( + "--accum-grads", + type=int, + default=1, + help="accum-grad num.", + ) + + parser.add_argument( + "--multi-optim", + type=str2bool, + default=True, + help="use sperate optimizer (enc / dec)", + ) + + parser.add_argument( + "--peak-enc-lr", + type=float, + default=0.0001, + help="The initial learning rate. This value should not need to be changed.", + ) + + parser.add_argument( + "--peak-dec-lr", + type=float, + default=0.001, + help="The initial learning rate. This value should not need to be changed.", + ) + + parser.add_argument( + "--encoder-type", + type=str, + default='d2v', + help="Type of encoder (e.g. conformer, w2v, d2v...", + ) + + parser.add_argument( + "--encoder-dim", + type=int, + default=768, + help="encoder embedding dimension", + ) + + parser.add_argument( + "--freeze-finetune-updates", + type=int, + default=0 + ) + + parser.add_argument( + "--additional-block", + type=str2bool, + default=True, + ) + + parser.add_argument( + "--decode-interval", + type=int, + default=200, + help="decode interval", + ) + + +def add_model_arguments(parser: argparse.ArgumentParser): + parser.add_argument( + "--num-encoder-layers", + type=str, + default="2,4,3,2,4", + help="Number of zipformer encoder layers, comma separated.", + ) + + parser.add_argument( + "--feedforward-dims", + type=str, + default="1024,1024,2048,2048,1024", + help="Feedforward dimension of the zipformer encoder layers, comma separated.", + ) + + parser.add_argument( + "--nhead", + type=str, + default="8,8,8,8,8", + help="Number of attention heads in the zipformer encoder layers.", + ) + + parser.add_argument( + "--encoder-dims", + type=str, + default="384,384,384,384,384", + help="Embedding dimension in the 2 blocks of zipformer encoder layers, comma separated", + ) + + parser.add_argument( + "--attention-dims", + type=str, + default="192,192,192,192,192", + help="""Attention dimension in the 2 blocks of zipformer encoder layers, comma separated; + not the same as embedding dimension.""", + ) + + parser.add_argument( + "--encoder-unmasked-dims", + type=str, + default="256,256,256,256,256", + help="Unmasked dimensions in the encoders, relates to augmentation during training. " + "Must be <= each of encoder_dims. Empirically, less than 256 seems to make performance " + " worse.", + ) + + parser.add_argument( + "--zipformer-downsampling-factors", + type=str, + default="1,2,4,8,2", + help="Downsampling factor for each stack of encoder layers.", + ) + + parser.add_argument( + "--cnn-module-kernels", + type=str, + default="31,31,31,31,31", + help="Sizes of kernels in convolution modules", + ) + + parser.add_argument( + "--decoder-dim", + type=int, + default=768, + help="Embedding dimension in the decoder model.", + ) + + parser.add_argument( + "--joiner-dim", + type=int, + default=768, + help="""Dimension used in the joiner model. + Outputs from the encoder and decoder model are projected + to this dimension before adding. + """, + ) + + parser.add_argument( + "--prompt", + type=str2bool, + default=False, + ) + + +def get_parser(): + parser = argparse.ArgumentParser( + formatter_class=argparse.ArgumentDefaultsHelpFormatter + ) + + parser.add_argument( + "--world-size", + type=int, + default=1, + help="Number of GPUs for DDP training.", + ) + + parser.add_argument( + "--master-port", + type=int, + default=12354, + help="Master port to use for DDP training.", + ) + + parser.add_argument( + "--tensorboard", + type=str2bool, + default=True, + help="Should various information be logged in tensorboard.", + ) + + parser.add_argument( + "--num-epochs", + type=int, + default=30, + help="Number of epochs to train.", + ) + + parser.add_argument( + "--num-updates", + type=int, + default=5000, + help="Number of epochs to train.", + ) + + parser.add_argument( + "--start-epoch", + type=int, + default=1, + help="""Resume training from this epoch. It should be positive. + If larger than 1, it will load checkpoint from + exp-dir/epoch-{start_epoch-1}.pt + """, + ) + + parser.add_argument( + "--start-batch", + type=int, + default=0, + help="""If positive, --start-epoch is ignored and + it loads the checkpoint from exp-dir/checkpoint-{start_batch}.pt + """, + ) + + parser.add_argument( + "--exp-dir", + type=str, + default="pruned_transducer_stateless7_ctc/exp", + help="""The experiment dir. + It specifies the directory where all training related + files, e.g., checkpoints, log, etc, are saved + """, + ) + + parser.add_argument( + "--bpe-model", + type=str, + default="data/lang_bpe_500/bpe.model", + help="Path to the BPE model", + ) + + parser.add_argument( + "--base-lr", type=float, default=0.05, help="The base learning rate." + ) + + parser.add_argument( + "--lr-batches", + type=float, + default=5000, + help="""Number of steps that affects how rapidly the learning rate + decreases. We suggest not to change this.""", + ) + + parser.add_argument( + "--lr-epochs", + type=float, + default=3.5, + help="""Number of epochs that affects how rapidly the learning rate decreases. + """, + ) + + parser.add_argument( + "--context-size", + type=int, + default=2, + help="The context size in the decoder. 1 means bigram; 2 means tri-gram", + ) + + parser.add_argument( + "--prune-range", + type=int, + default=5, + help="The prune range for rnnt loss, it means how many symbols(context)" + "we are using to compute the loss", + ) + + parser.add_argument( + "--lm-scale", + type=float, + default=0.25, + help="The scale to smooth the loss with lm " + "(output of prediction network) part.", + ) + + parser.add_argument( + "--am-scale", + type=float, + default=0.0, + help="The scale to smooth the loss with am (output of encoder network) part.", + ) + + parser.add_argument( + "--simple-loss-scale", + type=float, + default=0.5, + help="To get pruning ranges, we will calculate a simple version" + "loss(joiner is just addition), this simple loss also uses for" + "training (as a regularization item). We will scale the simple loss" + "with this parameter before adding to the final loss.", + ) + + parser.add_argument( + "--ctc-loss-scale", + type=float, + default=0.2, + help="Scale for CTC loss.", + ) + + parser.add_argument( + "--seed", + type=int, + default=42, + help="The seed for random generators intended for reproducibility", + ) + + parser.add_argument( + "--print-diagnostics", + type=str2bool, + default=False, + help="Accumulate stats on activations, print them and exit.", + ) + + parser.add_argument( + "--inf-check", + type=str2bool, + default=False, + help="Add hooks to check for infinite module outputs and gradients.", + ) + + parser.add_argument( + "--save-every-n", + type=int, + default=200, + help="""Save checkpoint after processing this number of batches" + periodically. We save checkpoint to exp-dir/ whenever + params.batch_idx_train % save_every_n == 0. The checkpoint filename + has the form: f'exp-dir/checkpoint-{params.batch_idx_train}.pt' + Note: It also saves checkpoint to `exp-dir/epoch-xxx.pt` at the + end of each epoch where `xxx` is the epoch number counting from 0. + """, + ) + + parser.add_argument( + "--keep-last-k", + type=int, + default=30, + help="""Only keep this number of checkpoints on disk. + For instance, if it is 3, there are only 3 checkpoints + in the exp-dir with filenames `checkpoint-xxx.pt`. + It does not affect checkpoints with name `epoch-xxx.pt`. + """, + ) + + parser.add_argument( + "--average-period", + type=int, + default=10, + help="""Update the averaged model, namely `model_avg`, after processing + this number of batches. `model_avg` is a separate version of model, + in which each floating-point parameter is the average of all the + parameters from the start of training. Each time we take the average, + we do: `model_avg = model * (average_period / batch_idx_train) + + model_avg * ((batch_idx_train - average_period) / batch_idx_train)`. + """, + ) + + parser.add_argument( + "--use-fp16", + type=str2bool, + default=True, + help="Whether to use half precision training.", + ) + + add_model_arguments(parser) + add_rep_arguments(parser) + add_adapter_arguments(parser) + + return parser + + +def get_params() -> AttributeDict: + """Return a dict containing training parameters. + + All training related parameters that are not passed from the commandline + are saved in the variable `params`. + + Commandline options are merged into `params` after they are parsed, so + you can also access them via `params`. + + Explanation of options saved in `params`: + + - best_train_loss: Best training loss so far. It is used to select + the model that has the lowest training loss. It is + updated during the training. + + - best_valid_loss: Best validation loss so far. It is used to select + the model that has the lowest validation loss. It is + updated during the training. + + - best_train_epoch: It is the epoch that has the best training loss. + + - best_valid_epoch: It is the epoch that has the best validation loss. + + - batch_idx_train: Used to writing statistics to tensorboard. It + contains number of batches trained so far across + epochs. + + - log_interval: Print training loss if batch_idx % log_interval` is 0 + + - reset_interval: Reset statistics if batch_idx % reset_interval is 0 + + - valid_interval: Run validation if batch_idx % valid_interval is 0 + + - feature_dim: The model input dim. It has to match the one used + in computing features. + + - subsampling_factor: The subsampling factor for the model. + + - encoder_dim: Hidden dim for multi-head attention model. + + - num_decoder_layers: Number of decoder layer of transformer decoder. + + - warm_step: The warmup period that dictates the decay of the + scale on "simple" (un-pruned) loss. + """ + params = AttributeDict( + { + "best_train_loss": float("inf"), + "best_valid_loss": float("inf"), + "best_train_epoch": -1, + "best_valid_epoch": -1, + "batch_idx_train": 0, + "log_interval": 5, + "reset_interval": 200, + "valid_interval": 3000, # For the 100h subset, use 800 + # parameters for zipformer + "feature_dim": 80, + "subsampling_factor": 320, # not passed in, this is fixed. + # parameters for ctc loss + "beam_size": 10, + "use_double_scores": True, + "warm_step": 0, + #"warm_step": 4000, + #"warm_step": 3000, + "env_info": get_env_info(), + } + ) + + return params + + +def get_encoder_model(params: AttributeDict) -> nn.Module: + # TODO: We can add an option to switch between Zipformer and Transformer + def to_int_tuple(s: str): + return tuple(map(int, s.split(","))) + + if params.encoder_type == 'd2v': + encoder = FairSeqData2VecEncoder( + input_size=params.encoder_dim, + w2v_url='None', + output_size=params.encoder_dim, + freeze_finetune_updates=params.freeze_finetune_updates, + additional_block=params.additional_block, + ) + else: + encoder = Zipformer( + num_features=params.feature_dim, + output_downsampling_factor=2, + zipformer_downsampling_factors=to_int_tuple( + params.zipformer_downsampling_factors + ), + encoder_dims=to_int_tuple(params.encoder_dims), + attention_dim=to_int_tuple(params.attention_dims), + encoder_unmasked_dims=to_int_tuple(params.encoder_unmasked_dims), + nhead=to_int_tuple(params.nhead), + feedforward_dim=to_int_tuple(params.feedforward_dims), + cnn_module_kernels=to_int_tuple(params.cnn_module_kernels), + num_encoder_layers=to_int_tuple(params.num_encoder_layers), + ) + + return encoder + + +def get_decoder_model(params: AttributeDict) -> nn.Module: + decoder = Decoder( + vocab_size=params.vocab_size, + decoder_dim=params.decoder_dim, + blank_id=params.blank_id, + context_size=params.context_size, + ) + return decoder + + +def get_joiner_model(params: AttributeDict) -> nn.Module: + joiner = Joiner( + encoder_dim=params.encoder_dim if params.encoder_type == 'd2v' else int(params.encoder_dims.split(",")[-1]), + decoder_dim=params.decoder_dim, + joiner_dim=params.joiner_dim, + vocab_size=params.vocab_size, + ) + return joiner + + +def get_transducer_model(params: AttributeDict) -> nn.Module: + encoder = get_encoder_model(params) + decoder = get_decoder_model(params) + joiner = get_joiner_model(params) + + model = Transducer( + encoder=encoder, + decoder=decoder, + joiner=joiner, + encoder_dim=params.encoder_dim if params.encoder_type == 'd2v' else int(params.encoder_dims.split(",")[-1]), + decoder_dim=params.decoder_dim, + joiner_dim=params.joiner_dim, + vocab_size=params.vocab_size, + prompt=params.prompt, + sid=params.spk_id, + ) + return model + + +def load_checkpoint_if_available( + params: AttributeDict, + model: nn.Module, + model_avg: nn.Module = None, + optimizer: Optional[torch.optim.Optimizer] = None, + scheduler: Optional[LRSchedulerType] = None, +) -> Optional[Dict[str, Any]]: + """Load checkpoint from file. + + If params.start_batch is positive, it will load the checkpoint from + `params.exp_dir/checkpoint-{params.start_batch}.pt`. Otherwise, if + params.start_epoch is larger than 1, it will load the checkpoint from + `params.start_epoch - 1`. + + Apart from loading state dict for `model` and `optimizer` it also updates + `best_train_epoch`, `best_train_loss`, `best_valid_epoch`, + and `best_valid_loss` in `params`. + + Args: + params: + The return value of :func:`get_params`. + model: + The training model. + model_avg: + The stored model averaged from the start of training. + optimizer: + The optimizer that we are using. + scheduler: + The scheduler that we are using. + Returns: + Return a dict containing previously saved training info. + """ + if params.start_batch > 0: + filename = params.exp_dir / f"checkpoint-{params.start_batch}.pt" + elif params.start_epoch > 1: + filename = params.exp_dir / f"epoch-{params.start_epoch-1}.pt" + elif params.add_adapter: + filename = params.exp_dir / f"../d2v-base-T.pt" + else: + return None + + assert filename.is_file(), f"{filename} does not exist!" + + saved_params = load_checkpoint( + filename, + model=model, + model_avg=model_avg, + optimizer=optimizer, + scheduler=scheduler, + strict=True if not params.add_adapter else False, + ) + + keys = [ + "best_train_epoch", + "best_valid_epoch", + "batch_idx_train", + "best_train_loss", + "best_valid_loss", + ] + for k in keys: + params[k] = saved_params[k] + + if params.start_batch > 0: + if "cur_epoch" in saved_params: + params["start_epoch"] = saved_params["cur_epoch"] + + if "cur_batch_idx" in saved_params: + params["cur_batch_idx"] = saved_params["cur_batch_idx"] + + params.batch_idx_train = 0 + + return saved_params + + +def save_checkpoint( + params: AttributeDict, + model: Union[nn.Module, DDP], + model_avg: Optional[nn.Module] = None, + optimizer: Optional[torch.optim.Optimizer] = None, + scheduler: Optional[LRSchedulerType] = None, + sampler: Optional[CutSampler] = None, + scaler: Optional[GradScaler] = None, + rank: int = 0, +) -> None: + """Save model, optimizer, scheduler and training stats to file. + + Args: + params: + It is returned by :func:`get_params`. + model: + The training model. + model_avg: + The stored model averaged from the start of training. + optimizer: + The optimizer used in the training. + sampler: + The sampler for the training dataset. + scaler: + The scaler used for mix precision training. + """ + if rank != 0: + return + filename = params.exp_dir / f"epoch-{params.cur_epoch}.pt" + save_checkpoint_impl( + filename=filename, + model=model, + model_avg=model_avg, + params=params, + optimizer=optimizer, + scheduler=scheduler, + sampler=sampler, + scaler=scaler, + rank=rank, + ) + + if params.best_train_epoch == params.cur_epoch: + best_train_filename = params.exp_dir / "best-train-loss.pt" + copyfile(src=filename, dst=best_train_filename) + + if params.best_valid_epoch == params.cur_epoch: + best_valid_filename = params.exp_dir / "best-valid-loss.pt" + copyfile(src=filename, dst=best_valid_filename) + + +def compute_loss( + params: AttributeDict, + model: Union[nn.Module, DDP], + sp: spm.SentencePieceProcessor, + batch: dict, + is_training: bool, + decode: bool = False, +) -> Tuple[Tensor, MetricsTracker]: + """ + Compute transducer loss given the model and its inputs. + + Args: + params: + Parameters for training. See :func:`get_params`. + model: + The model for training. It is an instance of Zipformer in our case. + batch: + A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()` + for the content in it. + is_training: + True for training. False for validation. When it is True, this + function enables autograd during computation; when it is False, it + disables autograd. + warmup: a floating point value which increases throughout training; + values >= 1.0 are fully warmed up and have all modules present. + """ + device = model.device if isinstance(model, DDP) else next(model.parameters()).device + feature = batch["inputs"] + # at entry, feature is (N, T, C) + assert feature.ndim == 2 or feature.ndim == 3 + feature = feature.to(device) + + supervisions = batch["supervisions"] + + if feature.ndim == 2: + feature_lens = [] + for supervision in supervisions['cut']: + try: feature_lens.append(supervision.tracks[0].cut.recording.num_samples) + except: feature_lens.append(supervision.recording.num_samples) + feature_lens = torch.tensor(feature_lens) + + elif feature.ndim == 3: + feature_lens = supervisions["num_frames"].to(device) + + batch_idx_train = params.batch_idx_train + warm_step = params.warm_step + + texts = batch["supervisions"]["text"] + + token_ids = sp.encode(texts, out_type=int) + y = k2.RaggedTensor(token_ids).to(device) + + with torch.set_grad_enabled(is_training): + simple_loss, pruned_loss, ctc_output = model( + x=feature, + x_lens=feature_lens, + y=y, + prune_range=params.prune_range, + am_scale=params.am_scale, + lm_scale=params.lm_scale, + ) + + s = params.simple_loss_scale + # take down the scale on the simple loss from 1.0 at the start + # to params.simple_loss scale by warm_step. + simple_loss_scale = ( + s + if batch_idx_train >= warm_step + else 1.0 - (batch_idx_train / warm_step) * (1.0 - s) + ) + pruned_loss_scale = ( + 1.0 + if batch_idx_train >= warm_step + else 0.1 + 0.9 * (batch_idx_train / warm_step) + ) + + loss = simple_loss_scale * simple_loss + pruned_loss_scale * pruned_loss + + info = MetricsTracker() + + if params.ctc_loss_scale > 0: + # Compute ctc loss + + # NOTE: We need `encode_supervisions` to sort sequences with + # different duration in decreasing order, required by + # `k2.intersect_dense` called in `k2.ctc_loss` + with warnings.catch_warnings(): + warnings.simplefilter("ignore") + supervision_segments, token_ids = encode_supervisions( + supervisions, + subsampling_factor=params.subsampling_factor, + token_ids=token_ids, + ) + + # Works with a BPE model + decoding_graph = k2.ctc_graph(token_ids, modified=False, device=device) + dense_fsa_vec = k2.DenseFsaVec( + ctc_output, + supervision_segments, + allow_truncate=params.subsampling_factor - 1, + ) + + ctc_loss = k2.ctc_loss( + decoding_graph=decoding_graph, + dense_fsa_vec=dense_fsa_vec, + output_beam=params.beam_size, + reduction="sum", + use_double_scores=params.use_double_scores, + ) + assert ctc_loss.requires_grad == is_training + loss += params.ctc_loss_scale * ctc_loss + + info["ctc_loss"] = ctc_loss.detach().cpu().item() + + assert loss.requires_grad == is_training + + if decode: + model.eval() + with torch.no_grad(): + try: + hypos = model.module.decode( + x=feature, + x_lens=feature_lens, + y=y, + sp=sp + ) + except: + hypos = model.decode( + x=feature, + x_lens=feature_lens, + y=y, + sp=sp + ) + + logging.info(f'ref: {batch["supervisions"]["text"][0]}') + logging.info(f'hyp: {" ".join(hypos[0])}') + model.train() + + with warnings.catch_warnings(): + warnings.simplefilter("ignore") + info["frames"] = (feature_lens // params.subsampling_factor).sum().item() + + # Note: We use reduction=sum while computing the loss. + info["utterances"] = feature.size(0) + info["loss"] = loss.detach().cpu().item() + info["simple_loss"] = simple_loss.detach().cpu().item() + info["pruned_loss"] = pruned_loss.detach().cpu().item() + + return loss, info + + +def compute_validation_loss( + params: AttributeDict, + model: Union[nn.Module, DDP], + sp: spm.SentencePieceProcessor, + valid_dl: torch.utils.data.DataLoader, + world_size: int = 1, +) -> MetricsTracker: + """Run the validation process.""" + model.eval() + + tot_loss = MetricsTracker() + + for batch_idx, batch in enumerate(valid_dl): + loss, loss_info = compute_loss( + params=params, + model=model, + sp=sp, + batch=batch, + is_training=False, + ) + assert loss.requires_grad is False + tot_loss = tot_loss + loss_info + + if world_size > 1: + tot_loss.reduce(loss.device) + + loss_value = tot_loss["loss"] / tot_loss["utterances"] + if loss_value < params.best_valid_loss: + params.best_valid_epoch = params.cur_epoch + params.best_valid_loss = loss_value + + return tot_loss + + +def train_one_epoch( + params: AttributeDict, + model: Union[nn.Module, DDP], + optimizer: torch.optim.Optimizer or [torch.optim.Optimizer, torch.optim.Optimizer], + scheduler: LRSchedulerType or [LRSchedulerType, LRSchedulerType], + sp: spm.SentencePieceProcessor, + train_dl: torch.utils.data.DataLoader, + valid_dl: torch.utils.data.DataLoader, + scaler: GradScaler, + model_avg: Optional[nn.Module] = None, + tb_writer: Optional[SummaryWriter] = None, + world_size: int = 1, + rank: int = 0, + wb = None, +) -> None: + """Train the model for one epoch. + + The training loss from the mean of all frames is saved in + `params.train_loss`. It runs the validation process every + `params.valid_interval` batches. + + Args: + params: + It is returned by :func:`get_params`. + model: + The model for training. + optimizer: + The optimizer we are using. + scheduler: + The learning rate scheduler, we call step() every step. + train_dl: + Dataloader for the training dataset. + valid_dl: + Dataloader for the validation dataset. + scaler: + The scaler used for mix precision training. + model_avg: + The stored model averaged from the start of training. + tb_writer: + Writer to write log messages to tensorboard. + world_size: + Number of nodes in DDP training. If it is 1, DDP is disabled. + rank: + The rank of the node in DDP training. If no DDP is used, it should + be set to 0. + """ + model.train() + + tot_loss = MetricsTracker() + + cur_batch_idx = params.get("cur_batch_idx", 0) + + if params.multi_optim: + optimizer_enc, optimizer_dec = optimizer[0], optimizer[1] + scheduler_enc, scheduler_dec = scheduler[0], scheduler[1] + + for batch_idx, batch in enumerate(train_dl): + if params.batch_idx_train > params.num_updates: + break + if batch_idx < cur_batch_idx: + continue + cur_batch_idx = batch_idx + + if batch_idx % params.accum_grads == 0: params.batch_idx_train += 1 + batch_size = len(batch["supervisions"]["text"]) + + try: + with torch.cuda.amp.autocast(enabled=params.use_fp16): + loss, loss_info = compute_loss( + params=params, + model=model, + sp=sp, + batch=batch, + is_training=True, + decode = True if batch_idx % params.decode_interval == 0 else False, + ) + + try: loss_info.reduce(loss.device) + except: pass + + numel = params.world_size / (params.accum_grads * loss_info["utterances"]) + loss *= numel ## normalize loss over utts(batch size) + + # summary stats + tot_loss = (tot_loss * (1 - 1 / params.reset_interval)) + loss_info + + # NOTE: We use reduction==sum and loss is computed over utterances + # in the batch and there is no normalization to it so far. + scaler.scale(loss).backward() + + if params.multi_optim and (batch_idx+1) % params.accum_grads == 0: + set_batch_count(model, params.batch_idx_train) + scheduler_enc.step_batch(params.batch_idx_train) + scheduler_dec.step_batch(params.batch_idx_train) + scaler.step(optimizer_enc) + scaler.step(optimizer_dec) + scaler.update() + optimizer_enc.zero_grad() + optimizer_dec.zero_grad() + elif not params.multi_optim and (batch_idx+1) % params.accum_grads == 0: + set_batch_count(model, params.batch_idx_train) + scheduler.step_batch(params.batch_idx_train) + scaler.step(optimizer) + scaler.update() + optimizer.zero_grad() + + except: # noqa + display_and_save_batch(batch, params=params, sp=sp) + raise + + if params.print_diagnostics and batch_idx == 5: + return + + ''' + if ( + rank == 0 + and params.batch_idx_train > 0 + and params.batch_idx_train % params.average_period == 0 + ): + update_averaged_model( + params=params, + model_cur=model, + model_avg=model_avg, + ) + ''' + + if ( + params.batch_idx_train > 0 + and params.batch_idx_train % params.save_every_n == 0 + ): + params.cur_batch_idx = batch_idx + save_checkpoint_with_global_batch_idx( + out_dir=params.exp_dir, + global_batch_idx=params.batch_idx_train, + model=model, + model_avg=model_avg, + params=params, + optimizer=optimizer, + scheduler=scheduler, + sampler=train_dl.sampler, + scaler=scaler, + rank=rank, + ) + del params.cur_batch_idx + ''' + remove_checkpoints( + out_dir=params.exp_dir, + topk=params.keep_last_k, + rank=rank, + ) + ''' + + if batch_idx % 100 == 0 and params.use_fp16: + # If the grad scale was less than 1, try increasing it. The _growth_interval + # of the grad scaler is configurable, but we can't configure it to have different + # behavior depending on the current grad scale. + cur_grad_scale = scaler._scale.item() + ''' + if cur_grad_scale < 1.0 or (cur_grad_scale < 8.0 and batch_idx % 400 == 0): + scaler.update(cur_grad_scale * 2.0) + ''' + if cur_grad_scale < 0.01: + logging.warning(f"Grad scale is small: {cur_grad_scale}") + if cur_grad_scale < 1.0e-05: + wb.log({"valid/loss": 10000}) + raise RuntimeError( + f"grad_scale is too small, exiting: {cur_grad_scale}" + ) + + #if params.batch_idx_train > 4000 and loss > 300 and params.wandb: + # wb.log({"valid/loss": 10000}) + # raise RuntimeError( + # f"divergence... exiting: loss={loss}" + # ) + + if batch_idx % (params.log_interval*params.accum_grads) == 0: + #for n, p in model.named_parameters(): + # if 'adapter' in n: + # print(p) + if params.multi_optim: + cur_enc_lr = scheduler_enc.get_last_lr()[0] + cur_dec_lr = scheduler_dec.get_last_lr()[0] + cur_grad_scale = scaler._scale.item() if params.use_fp16 else 1.0 + + logging.info( + f"Epoch {params.cur_epoch}, " + f"batch {batch_idx}, loss[{loss_info}], " + f"tot_loss[{tot_loss}], batch size: {batch_size}, " + f"enc_lr: {cur_enc_lr:.2e}, " + f"dec_lr: {cur_dec_lr:.2e}, " + + (f"grad_scale: {scaler._scale.item()}" if params.use_fp16 else "") + ) + + else: + cur_lr = scheduler.get_last_lr()[0] + cur_grad_scale = scaler._scale.item() if params.use_fp16 else 1.0 + + logging.info( + f"Epoch {params.cur_epoch}, " + f"batch {batch_idx}, loss[{loss_info}], " + f"tot_loss[{tot_loss}], batch size: {batch_size}, " + f"lr: {cur_lr:.2e}, " + + (f"grad_scale: {scaler._scale.item()}" if params.use_fp16 else "") + ) + + if tb_writer is not None: + if params.multi_optim: + tb_writer.add_scalar( + "train/enc_learning_rate", cur_enc_lr, params.batch_idx_train + ) + tb_writer.add_scalar( + "train/dec_learning_rate", cur_dec_lr, params.batch_idx_train + ) + + else: + tb_writer.add_scalar( + "train/learning_rate", cur_lr, params.batch_idx_train + ) + + loss_info.write_summary( + tb_writer, "train/current_", params.batch_idx_train + ) + tot_loss.write_summary(tb_writer, "train/tot_", params.batch_idx_train) + if params.use_fp16: + tb_writer.add_scalar( + "train/grad_scale", + cur_grad_scale, + params.batch_idx_train, + ) + + if wb is not None and rank == 0: + wb.log({"train/loss": loss_info["loss"]*numel}) + wb.log({"train/simple_loss": loss_info["simple_loss"]*numel}) + wb.log({"train/pruned_loss": loss_info["pruned_loss"]*numel}) + wb.log({"train/ctc_loss": loss_info["ctc_loss"]*numel}) + + ''' + logging.info("Computing validation loss") + valid_info = compute_validation_loss( + params=params, + model=model, + sp=sp, + valid_dl=valid_dl, + world_size=world_size, + ) + model.train() + logging.info(f"Epoch {params.cur_epoch}, validation: {valid_info}") + logging.info( + f"Maximum memory allocated so far is {torch.cuda.max_memory_allocated()//1000000}MB" + ) + if tb_writer is not None: + valid_info.write_summary( + tb_writer, "train/valid_", params.batch_idx_train + ) + + if wb is not None and rank == 0: + numel = 1 / (params.accum_grads * valid_info["utterances"]) + #wb.log({"valid/loss": valid_info["loss"]*numel}) + wb.log({"valid/loss": numel*(valid_info["simple_loss"] + +valid_info["pruned_loss"] + +valid_info["ctc_loss"] + )}) + wb.log({"valid/simple_loss": valid_info["simple_loss"]*numel}) + wb.log({"valid/pruned_loss": valid_info["pruned_loss"]*numel}) + wb.log({"valid/ctc_loss": valid_info["ctc_loss"]*numel}) + ''' + loss_value = tot_loss["loss"] / tot_loss["utterances"] + params.train_loss = loss_value + if params.train_loss < params.best_train_loss: + params.best_train_epoch = params.cur_epoch + params.best_train_loss = params.train_loss + + +def run(rank, world_size, args, wb=None): + """ + Args: + rank: + It is a value between 0 and `world_size-1`, which is + passed automatically by `mp.spawn()` in :func:`main`. + The node with rank 0 is responsible for saving checkpoint. + world_size: + Number of GPUs for DDP training. + args: + The return value of get_parser().parse_args() + """ + params = get_params() + params.update(vars(args)) + #params.warm_step *= params.accum_grads + + fix_random_seed(params.seed) + if world_size > 1: + setup_dist(rank, world_size, params.master_port) + + setup_logger(f"{params.exp_dir}/log/log-train") + logging.info("Training started") + + if args.tensorboard and rank == 0: + tb_writer = SummaryWriter(log_dir=f"{params.exp_dir}/tensorboard") + else: + tb_writer = None + + device = torch.device("cpu") + if torch.cuda.is_available(): + device = torch.device("cuda", rank) + logging.info(f"Device: {device}") + + sp = spm.SentencePieceProcessor() + sp.load(params.bpe_model) + + # is defined in local/train_bpe_model.py + params.blank_id = sp.piece_to_id("") + params.vocab_size = sp.get_piece_size() + + logging.info(params) + + logging.info("About to create model") + model = get_transducer_model(params) + logging.info(model) + + num_param = sum([p.numel() for p in model.parameters()]) + logging.info(f"Number of model parameters: {num_param}") + + assert params.save_every_n >= params.average_period + model_avg: Optional[nn.Module] = None + if rank == 0: + # model_avg is only used with rank 0 + model_avg = copy.deepcopy(model).to(torch.float64) + + assert params.start_epoch > 0, params.start_epoch + checkpoints = load_checkpoint_if_available( + params=params, model=model, model_avg=model_avg + ) + + model.to(device) + if world_size > 1: + logging.info("Using DDP") + model = DDP(model, device_ids=[rank], find_unused_parameters=True) + + if params.multi_optim: + logging.info("Using seperate optimizers over encoder, decoder ...") + + enc_param = [] + enc_names = [] + + dec_names = [] + dec_param = [] + + for n, p in model.named_parameters(): + name = n.split('.')[1] + if name == 'encoder' and 'feature_extractor' not in n: + enc_names.append(n) + enc_param.append(p) + elif 'ctc_output' in n: + enc_names.append(n) + enc_param.append(p) + elif 'feature_extractor' not in n: + dec_names.append(n) + dec_param.append(p) + + optimizer_enc = ScaledAdam( + enc_param, + lr=params.peak_enc_lr, + clipping_scale=None, + parameters_names=[enc_names], + ) + optimizer_dec = ScaledAdam( + dec_param, + lr=params.peak_dec_lr, + clipping_scale=5.0, + parameters_names=[dec_names], + ) + + scheduler_enc = Eden(optimizer_enc, params.lr_batches, params.lr_epochs) + scheduler_dec = Eden(optimizer_dec, params.lr_batches, params.lr_epochs) + optimizer = [optimizer_enc, optimizer_dec] + scheduler = [scheduler_enc, scheduler_dec] + + else: + parameters_names = [] + parameters_names.append( + [name_param_pair[0] for name_param_pair in model.named_parameters()] + ) + + logging.info(f"len name = {len(parameters_names)}") + logging.info(f"len param = {len(list(model.parameters()))}") + + optimizer = ScaledAdam( + model.parameters(), + lr=params.base_lr, + clipping_scale=2.0, + parameters_names=parameters_names, + ) + + scheduler = Eden(optimizer, params.lr_batches, params.lr_epochs) + + if checkpoints and ("optimizer" in checkpoints or "optimizer_enc" in checkpoints): + if params.multi_optim: + logging.info("Loading optimizer state dict") + optimizer_enc.load_state_dict(checkpoints["optimizer_enc"]) + optimizer_dec.load_state_dict(checkpoints["optimizer_dec"]) + + else: + logging.info("Loading optimizer state dict") + optimizer.load_state_dict(checkpoints["optimizer"]) + + if checkpoints: + if ( + params.multi_optim + and "scheduler_enc" in checkpoints + and checkpoints["scheduler_enc"] is not None + ): + logging.info("Loading enc/dec scheduler state dict") + scheduler_enc.load_state_dict(checkpoints["scheduler_enc"]) + scheduler_dec.load_state_dict(checkpoints["scheduler_dec"]) + else: + logging.info("Loading scheduler state dict") + scheduler.load_state_dict(checkpoints["scheduler"]) + + if params.print_diagnostics: + opts = diagnostics.TensorDiagnosticOptions( + 2**22 + ) # allow 4 megabytes per sub-module + diagnostic = diagnostics.attach_diagnostics(model, opts) + + if params.inf_check: + register_inf_check_hooks(model) + + librispeech = LibriSpeechAsrDataModule(args) + + train_cuts = librispeech.train_clean_100_cuts() + if params.full_libri: + train_cuts += librispeech.train_clean_360_cuts() + train_cuts += librispeech.train_other_500_cuts() + + def remove_short_and_long_utt(c: Cut): + # Keep only utterances with duration between 1 second and 20 seconds + # + # Caution: There is a reason to select 20.0 here. Please see + # ../local/display_manifest_statistics.py + # + # You should use ../local/display_manifest_statistics.py to get + # an utterance duration distribution for your dataset to select + # the threshold + return 1.0 <= c.duration <= 20.0 + + train_cuts = train_cuts.filter(remove_short_and_long_utt) + + if params.start_batch > 0 and checkpoints and "sampler" in checkpoints: + # We only load the sampler's state dict when it loads a checkpoint + # saved in the middle of an epoch + sampler_state_dict = checkpoints["sampler"] + else: + sampler_state_dict = None + + train_dl = librispeech.train_dataloaders( + train_cuts, sampler_state_dict=sampler_state_dict + ) + + valid_cuts = librispeech.dev_clean_cuts() + valid_cuts += librispeech.dev_other_cuts() + valid_dl = librispeech.valid_dataloaders(valid_cuts) + + ''' + if not params.print_diagnostics: + scan_pessimistic_batches_for_oom( + model=model, + train_dl=train_dl, + optimizer=optimizer, + sp=sp, + params=params, + ) + ''' + + scaler = GradScaler(enabled=params.use_fp16, init_scale=1.0) + if checkpoints and "grad_scaler" in checkpoints: + logging.info("Loading grad scaler state dict") + scaler.load_state_dict(checkpoints["grad_scaler"]) + + for epoch in range(params.start_epoch, params.num_epochs + 1): + if params.multi_optim: + scheduler_enc.step_epoch(epoch - 1) + scheduler_dec.step_epoch(epoch - 1) + else: + scheduler.step_epoch(epoch - 1) + fix_random_seed(params.seed + epoch - 1) + train_dl.sampler.set_epoch(epoch - 1) + + if tb_writer is not None: + tb_writer.add_scalar("train/epoch", epoch, params.batch_idx_train) + + params.cur_epoch = epoch + + train_one_epoch( + params=params, + model=model, + model_avg=model_avg, + optimizer=optimizer, + scheduler=scheduler, + sp=sp, + train_dl=train_dl, + valid_dl=valid_dl, + scaler=scaler, + tb_writer=tb_writer, + world_size=world_size, + rank=rank, + wb=wb, + ) + + if params.print_diagnostics: + diagnostic.print_diagnostics() + break + + ''' + if epoch % 50 == 0: + save_checkpoint( + params=params, + model=model, + model_avg=model_avg, + optimizer=optimizer, + scheduler=scheduler, + sampler=train_dl.sampler, + scaler=scaler, + rank=rank, + ) + ''' + + logging.info("Done!") + + if world_size > 1: + torch.distributed.barrier() + cleanup_dist() + + +def run_adapter(rank, world_size, args, wb=None): + """ + Args: + rank: + It is a value between 0 and `world_size-1`, which is + passed automatically by `mp.spawn()` in :func:`main`. + The node with rank 0 is responsible for saving checkpoint. + world_size: + Number of GPUs for DDP training. + args: + The return value of get_parser().parse_args() + """ + params = get_params() + params.update(vars(args)) + + fix_random_seed(params.seed) + if world_size > 1: + setup_dist(rank, world_size, params.master_port) + + setup_logger(f"{params.exp_dir}/log/log-train") + logging.info("Training started") + + if args.tensorboard and rank == 0: + tb_writer = SummaryWriter(log_dir=f"{params.exp_dir}/tensorboard") + else: + tb_writer = None + + device = torch.device("cpu") + if torch.cuda.is_available(): + device = torch.device("cuda", rank) + logging.info(f"Device: {device}") + + sp = spm.SentencePieceProcessor() + sp.load(params.bpe_model) + + # is defined in local/train_bpe_model.py + params.blank_id = sp.piece_to_id("") + params.vocab_size = sp.get_piece_size() + + logging.info(params) + + logging.info("About to create model") + model = get_transducer_model(params) + + num_param = sum([p.numel() if p.requires_grad else 0 for p in model.parameters()]) + logging.info(f"Number of model parameters: {num_param}") + + assert params.save_every_n >= params.average_period + model_avg: Optional[nn.Module] = None + if rank == 0: + # model_avg is only used with rank 0 + #model_avg = copy.deepcopy(model).to(torch.float64) + model_avg = None + + assert params.start_epoch > 0, params.start_epoch + checkpoints = load_checkpoint_if_available( + params=params, model=model, model_avg=model_avg + ) + + model.to(device) + if world_size > 1: + logging.info("Using DDP") + model = DDP(model, device_ids=[rank], find_unused_parameters=True) + + adapter_names = [] + adapter_param = [] + for enum, (n, p) in enumerate(model.named_parameters()): + #print(enum, n, p) + if 'encoder.encoders.layer_norm' in n or 'encoder.encoders.final_proj' in n or 'encoder.output_layer' in n or 'joiner' in n or 'simple' in n or 'ctc' in n or '11' in n or '10' in n: + print(n) + adapter_names.append(n) + adapter_param.append(p) + else: + p.requires_grad = False + ''' + if 'bias' in n: + adapter_names.append(n) + adapter_param.append(p) + else: + p.requires_grad = False + ''' + + ''' + if 'adapters' in n:# or 'joiner' in n or 'simple' in n or 'ctc' in n: + adapter_names.append(n) + adapter_param.append(p) + elif 'joiner' in n or 'simple' in n or 'ctc' in n: + p.requires_grad = True + else: + p.requires_grad = False + ''' + optimizer_adapter = ScaledAdam( + adapter_param, + lr=params.adapter_lr, + clipping_scale=5.0, + parameters_names=[adapter_names], + ) + + #for n, p in model.named_parameters(): + # p.requires_grad = False + + #prompt = torch.randn((100, 512), requires_grad=True) + #optimizer_adapter = ScaledAdam( + # [model.prompt], + # lr=params.adapter_lr, + # clipping_scale=5.0, + # parameters_names=['P'], + #) + + scheduler_adapter = Eden(optimizer_adapter, 10000, 7) #params.lr_batche, params.lr_epochs) + optimizer, scheduler = optimizer_adapter, scheduler_adapter + + librispeech = LibriSpeechAsrDataModule(args) + + ''' + if params.hpo: + train_cuts = librispeech.train_clean_10_cuts(option=params.gender) + else: + train_cuts = librispeech.train_clean_100_cuts(option=params.gender) + if params.full_libri: + train_cuts += librispeech.train_clean_360_cuts(option=params.gender) + train_cuts += librispeech.train_other_500_cuts(option=params.gender) + ''' + + #train_cuts = librispeech.train_clean_10_cuts(option='male') + #train_cuts = librispeech.test_clean_user(option='big') + train_cuts = librispeech.vox_cuts(option=params.spk_id) + + def remove_short_and_long_utt(c: Cut): + return 1.0 <= c.duration <= 20.0 + + train_cuts = train_cuts.filter(remove_short_and_long_utt) + + sampler_state_dict = None + + train_dl = librispeech.train_dataloaders( + train_cuts, sampler_state_dict=sampler_state_dict + ) + #train_dl = librispeech.test_dataloaders( + # train_cuts + #) + + ''' + print('\n'*5) + print('-'*30) + for batch in train_dl: + print(batch) + print('-'*30) + print('\n'*5) + exit() + ''' + + valid_cuts = librispeech.dev_clean_cuts(option=params.gender) + valid_cuts += librispeech.dev_other_cuts(option=params.gender) + valid_dl = librispeech.valid_dataloaders(valid_cuts) + + scaler = GradScaler(enabled=params.use_fp16, init_scale=1.0) + + for epoch in range(params.start_epoch, params.num_epochs + 1): + logging.info(f"update num : {params.batch_idx_train}") + scheduler.step_epoch(epoch - 1) + fix_random_seed(params.seed + epoch - 1) + train_dl.sampler.set_epoch(epoch - 1) + + if tb_writer is not None: + tb_writer.add_scalar("train/epoch", epoch, params.batch_idx_train) + + params.cur_epoch = epoch + + train_one_epoch( + params=params, + model=model, + model_avg=model_avg, + optimizer=optimizer, + scheduler=scheduler, + sp=sp, + train_dl=train_dl, + valid_dl=valid_dl, + scaler=scaler, + tb_writer=tb_writer, + world_size=world_size, + rank=rank, + wb=wb, + ) + + if params.print_diagnostics: + diagnostic.print_diagnostics() + break + + ''' + if epoch % 10 == 0: + save_checkpoint( + params=params, + model=model, + model_avg=model_avg, + optimizer=optimizer, + scheduler=scheduler, + sampler=train_dl.sampler, + scaler=scaler, + rank=rank, + ) + ''' + + logging.info("Done!") + + if world_size > 1: + torch.distributed.barrier() + cleanup_dist() + + +def display_and_save_batch( + batch: dict, + params: AttributeDict, + sp: spm.SentencePieceProcessor, +) -> None: + """Display the batch statistics and save the batch into disk. + + Args: + batch: + A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()` + for the content in it. + params: + Parameters for training. See :func:`get_params`. + sp: + The BPE model. + """ + from lhotse.utils import uuid4 + + filename = f"{params.exp_dir}/batch-{uuid4()}.pt" + logging.info(f"Saving batch to {filename}") + torch.save(batch, filename) + + supervisions = batch["supervisions"] + features = batch["inputs"] + + logging.info(f"features shape: {features.shape}") + + y = sp.encode(supervisions["text"], out_type=int) + num_tokens = sum(len(i) for i in y) + logging.info(f"num tokens: {num_tokens}") + + +def scan_pessimistic_batches_for_oom( + model: Union[nn.Module, DDP], + train_dl: torch.utils.data.DataLoader, + optimizer: torch.optim.Optimizer, + sp: spm.SentencePieceProcessor, + params: AttributeDict, +): + from lhotse.dataset import find_pessimistic_batches + + logging.info( + "Sanity check -- see if any of the batches in epoch 1 would cause OOM." + ) + batches, crit_values = find_pessimistic_batches(train_dl.sampler) + for criterion, cuts in batches.items(): + batch = train_dl.dataset[cuts] + try: + with torch.cuda.amp.autocast(enabled=params.use_fp16): + loss, _ = compute_loss( + params=params, + model=model, + sp=sp, + batch=batch, + is_training=True, + ) + loss.backward() + optimizer.zero_grad() + except Exception as e: + if "CUDA out of memory" in str(e): + logging.error( + "Your GPU ran out of memory with the current " + "max_duration setting. We recommend decreasing " + "max_duration and trying again.\n" + f"Failing criterion: {criterion} " + f"(={crit_values[criterion]}) ..." + ) + display_and_save_batch(batch, params=params, sp=sp) + raise + logging.info( + f"Maximum memory allocated so far is {torch.cuda.max_memory_allocated()//1000000}MB" + ) + + +def main(): + parser = get_parser() + LibriSpeechAsrDataModule.add_arguments(parser) + args = parser.parse_args() + if args.wandb: args.exp_dir = args.exp_dir + str(random.randint(0,400)) + args.exp_dir = Path(args.exp_dir) + + logging.info("save arguments to config.yaml...") + save_args(args) + + if args.wandb: wb = wandb.init(project="d2v-adapter", entity="dohe0342", config=vars(args)) + else: wb = None + + world_size = args.world_size + assert world_size >= 1 + if world_size > 1: + mp.spawn(run if not args.add_adapter else run_adapter, + args=(world_size, args, wb), + nprocs=world_size, + join=True + ) + else: + if args.add_adapter: run_adapter(rank=0, world_size=1, args=args, wb=wb) + else: run(rank=0, world_size=1, args=args, wb=wb) + +torch.set_num_threads(1) +torch.set_num_interop_threads(1) + +if __name__ == "__main__": + main() diff --git a/egs/tedlium3/ASR/pruned_transducer_stateless_d2v_v2/model.py b/egs/tedlium3/ASR/pruned_transducer_stateless_d2v_v2/model.py new file mode 100644 index 000000000..048af6161 --- /dev/null +++ b/egs/tedlium3/ASR/pruned_transducer_stateless_d2v_v2/model.py @@ -0,0 +1,241 @@ +# Copyright 2021 Xiaomi Corp. (authors: Fangjun Kuang, Wei Kang) +# +# See ../../../../LICENSE for clarification regarding multiple authors +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +from typing import Tuple +import logging + +import k2 +import torch +import torch.nn as nn +from encoder_interface import EncoderInterface + +from icefall.utils import add_sos + + +class Transducer(nn.Module): + """It implements https://arxiv.org/pdf/1211.3711.pdf + "Sequence Transduction with Recurrent Neural Networks" + """ + + def __init__( + self, + encoder: EncoderInterface, + decoder: nn.Module, + joiner: nn.Module, + encoder_dim: int, + decoder_dim: int, + joiner_dim: int, + vocab_size: int, + prompt=False, + sid=None, + ): + """ + Args: + encoder: + It is the transcription network in the paper. Its accepts + two inputs: `x` of (N, T, encoder_dim) and `x_lens` of shape (N,). + It returns two tensors: `logits` of shape (N, T, encoder_dm) and + `logit_lens` of shape (N,). + decoder: + It is the prediction network in the paper. Its input shape + is (N, U) and its output shape is (N, U, decoder_dim). + It should contain one attribute: `blank_id`. + joiner: + It has two inputs with shapes: (N, T, encoder_dim) and (N, U, decoder_dim). + Its output shape is (N, T, U, vocab_size). Note that its output contains + unnormalized probs, i.e., not processed by log-softmax. + """ + super().__init__() + assert isinstance(encoder, EncoderInterface), type(encoder) + assert hasattr(decoder, "blank_id") + + self.encoder = encoder + self.decoder = decoder + self.joiner = joiner + + self.simple_am_proj = nn.Linear( + encoder_dim, + vocab_size, + ) + self.simple_lm_proj = nn.Linear(decoder_dim, vocab_size) + + self.ctc_output = nn.Sequential( + nn.Dropout(p=0.1), + nn.Linear(encoder_dim, vocab_size), + nn.LogSoftmax(dim=-1), + ) + + self.prompt = None + self.sid = sid + print('-'*20) + print(self.sid) + print('-'*20) + if prompt: + #statistic = open(f'/home/work/workspace/icefall/egs/librispeech/ASR/conv_feat/{self.sid}/{sid}_statistic.txt', 'r').readlines() + self.prompt = torch.nn.Parameter(torch.rand((50, 512))) + #print(self.prompt) + ''' + new_emb = torch.empty(512, 50) + for i in range(512): + mean, std = statistic[i].strip().split(' ') + print(new_emb[i].size()) + print(float(mean), float(std)) + new_emb[i] = torch.normal(float(mean), float(std), size=(1,50)).squeeze() + new_emb = new_emb.transpose(1,0) + self.prompt = torch.nn.Parameter(new_emb) + print(self.prompt) + ''' + + def forward( + self, + x: torch.Tensor, + x_lens: torch.Tensor, + y: k2.RaggedTensor, + prune_range: int = 5, + am_scale: float = 0.0, + lm_scale: float = 0.0, + ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]: + """ + Args: + x: + A 3-D tensor of shape (N, T, C). + x_lens: + A 1-D tensor of shape (N,). It contains the number of frames in `x` + before padding. + y: + A ragged tensor with 2 axes [utt][label]. It contains labels of each + utterance. + prune_range: + The prune range for rnnt loss, it means how many symbols(context) + we are considering for each frame to compute the loss. + am_scale: + The scale to smooth the loss with am (output of encoder network) + part + lm_scale: + The scale to smooth the loss with lm (output of predictor network) + part + Returns: + Return a tuple containing simple loss, pruned loss, and ctc-output. + + Note: + Regarding am_scale & lm_scale, it will make the loss-function one of + the form: + lm_scale * lm_probs + am_scale * am_probs + + (1-lm_scale-am_scale) * combined_probs + """ + assert x.ndim == 2 or x.ndim == 3, x.shape + assert x_lens.ndim == 1, x_lens.shape + assert y.num_axes == 2, y.num_axes + + assert x.size(0) == x_lens.size(0) == y.dim0 + + encoder_out, x_lens = self.encoder(x, x_lens, prompt=self.prompt, sid=self.sid) + assert torch.all(x_lens > 0) + + # compute ctc log-probs + ctc_output = self.ctc_output(encoder_out) + + # Now for the decoder, i.e., the prediction network + row_splits = y.shape.row_splits(1) + y_lens = row_splits[1:] - row_splits[:-1] + + blank_id = self.decoder.blank_id + sos_y = add_sos(y, sos_id=blank_id) + + # sos_y_padded: [B, S + 1], start with SOS. + sos_y_padded = sos_y.pad(mode="constant", padding_value=blank_id) + + # decoder_out: [B, S + 1, decoder_dim] + decoder_out = self.decoder(sos_y_padded) + + # Note: y does not start with SOS + # y_padded : [B, S] + y_padded = y.pad(mode="constant", padding_value=0) + + y_padded = y_padded.to(torch.int64) + boundary = torch.zeros((x.size(0), 4), dtype=torch.int64, device=x.device) + boundary[:, 2] = y_lens + boundary[:, 3] = x_lens + + lm = self.simple_lm_proj(decoder_out) + am = self.simple_am_proj(encoder_out) + + with torch.cuda.amp.autocast(enabled=False): + simple_loss, (px_grad, py_grad) = k2.rnnt_loss_smoothed( + lm=lm.float(), + am=am.float(), + symbols=y_padded, + termination_symbol=blank_id, + lm_only_scale=lm_scale, + am_only_scale=am_scale, + boundary=boundary, + reduction="sum", + return_grad=True, + ) + + # ranges : [B, T, prune_range] + ranges = k2.get_rnnt_prune_ranges( + px_grad=px_grad, + py_grad=py_grad, + boundary=boundary, + s_range=prune_range, + ) + + # am_pruned : [B, T, prune_range, encoder_dim] + # lm_pruned : [B, T, prune_range, decoder_dim] + am_pruned, lm_pruned = k2.do_rnnt_pruning( + am=self.joiner.encoder_proj(encoder_out), + lm=self.joiner.decoder_proj(decoder_out), + ranges=ranges, + ) + + # logits : [B, T, prune_range, vocab_size] + + # project_input=False since we applied the decoder's input projections + # prior to do_rnnt_pruning (this is an optimization for speed). + logits = self.joiner(am_pruned, lm_pruned, project_input=False) + + with torch.cuda.amp.autocast(enabled=False): + pruned_loss = k2.rnnt_loss_pruned( + logits=logits.float(), + symbols=y_padded, + ranges=ranges, + termination_symbol=blank_id, + boundary=boundary, + reduction="sum", + ) + + return (simple_loss, pruned_loss, ctc_output) + + def decode( + self, + x: torch.Tensor, + x_lens: torch.Tensor, + y: k2.RaggedTensor, + sp, + ): + from beam_search import greedy_search_batch + + encoder_out, x_lens = self.encoder(x, x_lens) + + hyps = [] + hyp_tokens = greedy_search_batch(self, encoder_out, x_lens) + + for hyp in sp.decode(hyp_tokens): + hyps.append(hyp.split()) + + return hyps + diff --git a/egs/tedlium3/ASR/pruned_transducer_stateless_d2v_v2/nets_utils.py b/egs/tedlium3/ASR/pruned_transducer_stateless_d2v_v2/nets_utils.py new file mode 100644 index 000000000..08e7f2098 --- /dev/null +++ b/egs/tedlium3/ASR/pruned_transducer_stateless_d2v_v2/nets_utils.py @@ -0,0 +1,503 @@ +# -*- coding: utf-8 -*- + +"""Network related utility tools.""" + +import logging +from typing import Dict + +import numpy as np +import torch + + +def to_device(m, x): + """Send tensor into the device of the module. + + Args: + m (torch.nn.Module): Torch module. + x (Tensor): Torch tensor. + + Returns: + Tensor: Torch tensor located in the same place as torch module. + + """ + if isinstance(m, torch.nn.Module): + device = next(m.parameters()).device + elif isinstance(m, torch.Tensor): + device = m.device + else: + raise TypeError( + "Expected torch.nn.Module or torch.tensor, " f"bot got: {type(m)}" + ) + return x.to(device) + + +def pad_list(xs, pad_value): + """Perform padding for the list of tensors. + + Args: + xs (List): List of Tensors [(T_1, `*`), (T_2, `*`), ..., (T_B, `*`)]. + pad_value (float): Value for padding. + + Returns: + Tensor: Padded tensor (B, Tmax, `*`). + + Examples: + >>> x = [torch.ones(4), torch.ones(2), torch.ones(1)] + >>> x + [tensor([1., 1., 1., 1.]), tensor([1., 1.]), tensor([1.])] + >>> pad_list(x, 0) + tensor([[1., 1., 1., 1.], + [1., 1., 0., 0.], + [1., 0., 0., 0.]]) + + """ + n_batch = len(xs) + max_len = max(x.size(0) for x in xs) + pad = xs[0].new(n_batch, max_len, *xs[0].size()[1:]).fill_(pad_value) + + for i in range(n_batch): + pad[i, : xs[i].size(0)] = xs[i] + + return pad + + +def make_pad_mask(lengths, xs=None, length_dim=-1, maxlen=None): + """Make mask tensor containing indices of padded part. + + Args: + lengths (LongTensor or List): Batch of lengths (B,). + xs (Tensor, optional): The reference tensor. + If set, masks will be the same shape as this tensor. + length_dim (int, optional): Dimension indicator of the above tensor. + See the example. + + Returns: + Tensor: Mask tensor containing indices of padded part. + dtype=torch.uint8 in PyTorch 1.2- + dtype=torch.bool in PyTorch 1.2+ (including 1.2) + + Examples: + With only lengths. + + >>> lengths = [5, 3, 2] + >>> make_pad_mask(lengths) + masks = [[0, 0, 0, 0 ,0], + [0, 0, 0, 1, 1], + [0, 0, 1, 1, 1]] + + With the reference tensor. + + >>> xs = torch.zeros((3, 2, 4)) + >>> make_pad_mask(lengths, xs) + tensor([[[0, 0, 0, 0], + [0, 0, 0, 0]], + [[0, 0, 0, 1], + [0, 0, 0, 1]], + [[0, 0, 1, 1], + [0, 0, 1, 1]]], dtype=torch.uint8) + >>> xs = torch.zeros((3, 2, 6)) + >>> make_pad_mask(lengths, xs) + tensor([[[0, 0, 0, 0, 0, 1], + [0, 0, 0, 0, 0, 1]], + [[0, 0, 0, 1, 1, 1], + [0, 0, 0, 1, 1, 1]], + [[0, 0, 1, 1, 1, 1], + [0, 0, 1, 1, 1, 1]]], dtype=torch.uint8) + + With the reference tensor and dimension indicator. + + >>> xs = torch.zeros((3, 6, 6)) + >>> make_pad_mask(lengths, xs, 1) + tensor([[[0, 0, 0, 0, 0, 0], + [0, 0, 0, 0, 0, 0], + [0, 0, 0, 0, 0, 0], + [0, 0, 0, 0, 0, 0], + [0, 0, 0, 0, 0, 0], + [1, 1, 1, 1, 1, 1]], + [[0, 0, 0, 0, 0, 0], + [0, 0, 0, 0, 0, 0], + [0, 0, 0, 0, 0, 0], + [1, 1, 1, 1, 1, 1], + [1, 1, 1, 1, 1, 1], + [1, 1, 1, 1, 1, 1]], + [[0, 0, 0, 0, 0, 0], + [0, 0, 0, 0, 0, 0], + [1, 1, 1, 1, 1, 1], + [1, 1, 1, 1, 1, 1], + [1, 1, 1, 1, 1, 1], + [1, 1, 1, 1, 1, 1]]], dtype=torch.uint8) + >>> make_pad_mask(lengths, xs, 2) + tensor([[[0, 0, 0, 0, 0, 1], + [0, 0, 0, 0, 0, 1], + [0, 0, 0, 0, 0, 1], + [0, 0, 0, 0, 0, 1], + [0, 0, 0, 0, 0, 1], + [0, 0, 0, 0, 0, 1]], + [[0, 0, 0, 1, 1, 1], + [0, 0, 0, 1, 1, 1], + [0, 0, 0, 1, 1, 1], + [0, 0, 0, 1, 1, 1], + [0, 0, 0, 1, 1, 1], + [0, 0, 0, 1, 1, 1]], + [[0, 0, 1, 1, 1, 1], + [0, 0, 1, 1, 1, 1], + [0, 0, 1, 1, 1, 1], + [0, 0, 1, 1, 1, 1], + [0, 0, 1, 1, 1, 1], + [0, 0, 1, 1, 1, 1]]], dtype=torch.uint8) + + """ + if length_dim == 0: + raise ValueError("length_dim cannot be 0: {}".format(length_dim)) + + if not isinstance(lengths, list): + lengths = lengths.long().tolist() + + bs = int(len(lengths)) + if maxlen is None: + if xs is None: + maxlen = int(max(lengths)) + else: + maxlen = xs.size(length_dim) + else: + assert xs is None + assert maxlen >= int(max(lengths)) + + seq_range = torch.arange(0, maxlen, dtype=torch.int64) + seq_range_expand = seq_range.unsqueeze(0).expand(bs, maxlen) + seq_length_expand = seq_range_expand.new(lengths).unsqueeze(-1) + mask = seq_range_expand >= seq_length_expand + + if xs is not None: + assert xs.size(0) == bs, (xs.size(0), bs) + + if length_dim < 0: + length_dim = xs.dim() + length_dim + # ind = (:, None, ..., None, :, , None, ..., None) + ind = tuple( + slice(None) if i in (0, length_dim) else None for i in range(xs.dim()) + ) + mask = mask[ind].expand_as(xs).to(xs.device) + return mask + + +def make_non_pad_mask(lengths, xs=None, length_dim=-1): + """Make mask tensor containing indices of non-padded part. + + Args: + lengths (LongTensor or List): Batch of lengths (B,). + xs (Tensor, optional): The reference tensor. + If set, masks will be the same shape as this tensor. + length_dim (int, optional): Dimension indicator of the above tensor. + See the example. + + Returns: + ByteTensor: mask tensor containing indices of padded part. + dtype=torch.uint8 in PyTorch 1.2- + dtype=torch.bool in PyTorch 1.2+ (including 1.2) + + Examples: + With only lengths. + + >>> lengths = [5, 3, 2] + >>> make_non_pad_mask(lengths) + masks = [[1, 1, 1, 1 ,1], + [1, 1, 1, 0, 0], + [1, 1, 0, 0, 0]] + + With the reference tensor. + + >>> xs = torch.zeros((3, 2, 4)) + >>> make_non_pad_mask(lengths, xs) + tensor([[[1, 1, 1, 1], + [1, 1, 1, 1]], + [[1, 1, 1, 0], + [1, 1, 1, 0]], + [[1, 1, 0, 0], + [1, 1, 0, 0]]], dtype=torch.uint8) + >>> xs = torch.zeros((3, 2, 6)) + >>> make_non_pad_mask(lengths, xs) + tensor([[[1, 1, 1, 1, 1, 0], + [1, 1, 1, 1, 1, 0]], + [[1, 1, 1, 0, 0, 0], + [1, 1, 1, 0, 0, 0]], + [[1, 1, 0, 0, 0, 0], + [1, 1, 0, 0, 0, 0]]], dtype=torch.uint8) + + With the reference tensor and dimension indicator. + + >>> xs = torch.zeros((3, 6, 6)) + >>> make_non_pad_mask(lengths, xs, 1) + tensor([[[1, 1, 1, 1, 1, 1], + [1, 1, 1, 1, 1, 1], + [1, 1, 1, 1, 1, 1], + [1, 1, 1, 1, 1, 1], + [1, 1, 1, 1, 1, 1], + [0, 0, 0, 0, 0, 0]], + [[1, 1, 1, 1, 1, 1], + [1, 1, 1, 1, 1, 1], + [1, 1, 1, 1, 1, 1], + [0, 0, 0, 0, 0, 0], + [0, 0, 0, 0, 0, 0], + [0, 0, 0, 0, 0, 0]], + [[1, 1, 1, 1, 1, 1], + [1, 1, 1, 1, 1, 1], + [0, 0, 0, 0, 0, 0], + [0, 0, 0, 0, 0, 0], + [0, 0, 0, 0, 0, 0], + [0, 0, 0, 0, 0, 0]]], dtype=torch.uint8) + >>> make_non_pad_mask(lengths, xs, 2) + tensor([[[1, 1, 1, 1, 1, 0], + [1, 1, 1, 1, 1, 0], + [1, 1, 1, 1, 1, 0], + [1, 1, 1, 1, 1, 0], + [1, 1, 1, 1, 1, 0], + [1, 1, 1, 1, 1, 0]], + [[1, 1, 1, 0, 0, 0], + [1, 1, 1, 0, 0, 0], + [1, 1, 1, 0, 0, 0], + [1, 1, 1, 0, 0, 0], + [1, 1, 1, 0, 0, 0], + [1, 1, 1, 0, 0, 0]], + [[1, 1, 0, 0, 0, 0], + [1, 1, 0, 0, 0, 0], + [1, 1, 0, 0, 0, 0], + [1, 1, 0, 0, 0, 0], + [1, 1, 0, 0, 0, 0], + [1, 1, 0, 0, 0, 0]]], dtype=torch.uint8) + + """ + return ~make_pad_mask(lengths, xs, length_dim) + + +def mask_by_length(xs, lengths, fill=0): + """Mask tensor according to length. + + Args: + xs (Tensor): Batch of input tensor (B, `*`). + lengths (LongTensor or List): Batch of lengths (B,). + fill (int or float): Value to fill masked part. + + Returns: + Tensor: Batch of masked input tensor (B, `*`). + + Examples: + >>> x = torch.arange(5).repeat(3, 1) + 1 + >>> x + tensor([[1, 2, 3, 4, 5], + [1, 2, 3, 4, 5], + [1, 2, 3, 4, 5]]) + >>> lengths = [5, 3, 2] + >>> mask_by_length(x, lengths) + tensor([[1, 2, 3, 4, 5], + [1, 2, 3, 0, 0], + [1, 2, 0, 0, 0]]) + + """ + assert xs.size(0) == len(lengths) + ret = xs.data.new(*xs.size()).fill_(fill) + for i, l in enumerate(lengths): + ret[i, :l] = xs[i, :l] + return ret + + +def th_accuracy(pad_outputs, pad_targets, ignore_label): + """Calculate accuracy. + + Args: + pad_outputs (Tensor): Prediction tensors (B * Lmax, D). + pad_targets (LongTensor): Target label tensors (B, Lmax, D). + ignore_label (int): Ignore label id. + + Returns: + float: Accuracy value (0.0 - 1.0). + + """ + pad_pred = pad_outputs.view( + pad_targets.size(0), pad_targets.size(1), pad_outputs.size(1) + ).argmax(2) + mask = pad_targets != ignore_label + numerator = torch.sum( + pad_pred.masked_select(mask) == pad_targets.masked_select(mask) + ) + denominator = torch.sum(mask) + return float(numerator) / float(denominator) + + +def to_torch_tensor(x): + """Change to torch.Tensor or ComplexTensor from numpy.ndarray. + + Args: + x: Inputs. It should be one of numpy.ndarray, Tensor, ComplexTensor, and dict. + + Returns: + Tensor or ComplexTensor: Type converted inputs. + + Examples: + >>> xs = np.ones(3, dtype=np.float32) + >>> xs = to_torch_tensor(xs) + tensor([1., 1., 1.]) + >>> xs = torch.ones(3, 4, 5) + >>> assert to_torch_tensor(xs) is xs + >>> xs = {'real': xs, 'imag': xs} + >>> to_torch_tensor(xs) + ComplexTensor( + Real: + tensor([1., 1., 1.]) + Imag; + tensor([1., 1., 1.]) + ) + + """ + # If numpy, change to torch tensor + if isinstance(x, np.ndarray): + if x.dtype.kind == "c": + # Dynamically importing because torch_complex requires python3 + from torch_complex.tensor import ComplexTensor + + return ComplexTensor(x) + else: + return torch.from_numpy(x) + + # If {'real': ..., 'imag': ...}, convert to ComplexTensor + elif isinstance(x, dict): + # Dynamically importing because torch_complex requires python3 + from torch_complex.tensor import ComplexTensor + + if "real" not in x or "imag" not in x: + raise ValueError("has 'real' and 'imag' keys: {}".format(list(x))) + # Relative importing because of using python3 syntax + return ComplexTensor(x["real"], x["imag"]) + + # If torch.Tensor, as it is + elif isinstance(x, torch.Tensor): + return x + + else: + error = ( + "x must be numpy.ndarray, torch.Tensor or a dict like " + "{{'real': torch.Tensor, 'imag': torch.Tensor}}, " + "but got {}".format(type(x)) + ) + try: + from torch_complex.tensor import ComplexTensor + except Exception: + # If PY2 + raise ValueError(error) + else: + # If PY3 + if isinstance(x, ComplexTensor): + return x + else: + raise ValueError(error) + + +def get_subsample(train_args, mode, arch): + """Parse the subsampling factors from the args for the specified `mode` and `arch`. + + Args: + train_args: argument Namespace containing options. + mode: one of ('asr', 'mt', 'st') + arch: one of ('rnn', 'rnn-t', 'rnn_mix', 'rnn_mulenc', 'transformer') + + Returns: + np.ndarray / List[np.ndarray]: subsampling factors. + """ + if arch == "transformer": + return np.array([1]) + + elif mode == "mt" and arch == "rnn": + # +1 means input (+1) and layers outputs (train_args.elayer) + subsample = np.ones(train_args.elayers + 1, dtype=np.int64) + logging.warning("Subsampling is not performed for machine translation.") + logging.info("subsample: " + " ".join([str(x) for x in subsample])) + return subsample + + elif ( + (mode == "asr" and arch in ("rnn", "rnn-t")) + or (mode == "mt" and arch == "rnn") + or (mode == "st" and arch == "rnn") + ): + subsample = np.ones(train_args.elayers + 1, dtype=np.int64) + if train_args.etype.endswith("p") and not train_args.etype.startswith("vgg"): + ss = train_args.subsample.split("_") + for j in range(min(train_args.elayers + 1, len(ss))): + subsample[j] = int(ss[j]) + else: + logging.warning( + "Subsampling is not performed for vgg*. " + "It is performed in max pooling layers at CNN." + ) + logging.info("subsample: " + " ".join([str(x) for x in subsample])) + return subsample + + elif mode == "asr" and arch == "rnn_mix": + subsample = np.ones( + train_args.elayers_sd + train_args.elayers + 1, dtype=np.int64 + ) + if train_args.etype.endswith("p") and not train_args.etype.startswith("vgg"): + ss = train_args.subsample.split("_") + for j in range( + min(train_args.elayers_sd + train_args.elayers + 1, len(ss)) + ): + subsample[j] = int(ss[j]) + else: + logging.warning( + "Subsampling is not performed for vgg*. " + "It is performed in max pooling layers at CNN." + ) + logging.info("subsample: " + " ".join([str(x) for x in subsample])) + return subsample + + elif mode == "asr" and arch == "rnn_mulenc": + subsample_list = [] + for idx in range(train_args.num_encs): + subsample = np.ones(train_args.elayers[idx] + 1, dtype=np.int64) + if train_args.etype[idx].endswith("p") and not train_args.etype[ + idx + ].startswith("vgg"): + ss = train_args.subsample[idx].split("_") + for j in range(min(train_args.elayers[idx] + 1, len(ss))): + subsample[j] = int(ss[j]) + else: + logging.warning( + "Encoder %d: Subsampling is not performed for vgg*. " + "It is performed in max pooling layers at CNN.", + idx + 1, + ) + logging.info("subsample: " + " ".join([str(x) for x in subsample])) + subsample_list.append(subsample) + return subsample_list + + else: + raise ValueError("Invalid options: mode={}, arch={}".format(mode, arch)) + + +def rename_state_dict( + old_prefix: str, new_prefix: str, state_dict: Dict[str, torch.Tensor] +): + """Replace keys of old prefix with new prefix in state dict.""" + # need this list not to break the dict iterator + old_keys = [k for k in state_dict if k.startswith(old_prefix)] + if len(old_keys) > 0: + logging.warning(f"Rename: {old_prefix} -> {new_prefix}") + for k in old_keys: + v = state_dict.pop(k) + new_k = k.replace(old_prefix, new_prefix) + state_dict[new_k] = v + + +def get_activation(act): + """Return activation function.""" + # Lazy load to avoid unused import + from espnet.nets.pytorch_backend.conformer.swish import Swish + + activation_funcs = { + "hardtanh": torch.nn.Hardtanh, + "tanh": torch.nn.Tanh, + "relu": torch.nn.ReLU, + "selu": torch.nn.SELU, + "swish": Swish, + } + + return activation_funcs[act]() diff --git a/egs/tedlium3/ASR/pruned_transducer_stateless_d2v_v2/optim.py b/egs/tedlium3/ASR/pruned_transducer_stateless_d2v_v2/optim.py new file mode 100644 index 000000000..4f37642cf --- /dev/null +++ b/egs/tedlium3/ASR/pruned_transducer_stateless_d2v_v2/optim.py @@ -0,0 +1,1061 @@ +# Copyright 2022 Xiaomi Corp. (authors: Daniel Povey) +# +# See ../LICENSE for clarification regarding multiple authors +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import contextlib +import logging +import random +from collections import defaultdict +from typing import List, Optional, Tuple, Union + +import torch +from lhotse.utils import fix_random_seed +from scaling import ActivationBalancer +from torch import Tensor +from torch.optim import Optimizer + + +class BatchedOptimizer(Optimizer): + """ + This class adds to class Optimizer the capability to optimize parameters in batches: + it will stack the parameters and their grads for you so the optimizer can work + on tensors with an extra leading dimension. This is intended for speed with GPUs, + as it reduces the number of kernels launched in the optimizer. + + Args: + params: + """ + + def __init__(self, params, defaults): + super(BatchedOptimizer, self).__init__(params, defaults) + + @contextlib.contextmanager + def batched_params(self, param_group, group_params_names): + """ + This function returns (technically, yields) a list of + of tuples (p, state), where + p is a `fake` parameter that is stacked (over axis 0) from real parameters + that share the same shape, and its gradient is also stacked; + `state` is the state corresponding to this batch of parameters + (it will be physically located in the "state" for one of the real + parameters, the last one that has any particular shape and dtype). + + This function is decorated as a context manager so that it can + write parameters back to their "real" locations. + + The idea is, instead of doing: + + for p in group["params"]: + state = self.state[p] + ... + + you can do: + + with self.batched_params(group["params"]) as batches: + for p, state, p_names in batches: + ... + + + Args: + group: a parameter group, which is a list of parameters; should be + one of self.param_groups. + group_params_names: name for each parameter in group, + which is List[str]. + """ + batches = defaultdict( + list + ) # `batches` maps from tuple (dtype_as_str,*shape) to list of nn.Parameter + batches_names = defaultdict( + list + ) # `batches` maps from tuple (dtype_as_str,*shape) to list of str + + assert len(param_group) == len(group_params_names) + for p, named_p in zip(param_group, group_params_names): + key = (str(p.dtype), *p.shape) + batches[key].append(p) + batches_names[key].append(named_p) + + batches_names_keys = list(batches_names.keys()) + sorted_idx = sorted( + range(len(batches_names)), key=lambda i: batches_names_keys[i] + ) + batches_names = [batches_names[batches_names_keys[idx]] for idx in sorted_idx] + batches = [batches[batches_names_keys[idx]] for idx in sorted_idx] + + stacked_params_dict = dict() + + # turn batches into a list, in deterministic order. + # tuples will contain tuples of (stacked_param, state, stacked_params_names), + # one for each batch in `batches`. + tuples = [] + + for batch, batch_names in zip(batches, batches_names): + p = batch[0] + # we arbitrarily store the state in the + # state corresponding to the 1st parameter in the + # group. class Optimizer will take care of saving/loading state. + state = self.state[p] + p_stacked = torch.stack(batch) + grad = torch.stack( + [torch.zeros_like(p) if p.grad is None else p.grad for p in batch] + ) + p_stacked.grad = grad + stacked_params_dict[key] = p_stacked + tuples.append((p_stacked, state, batch_names)) + + yield tuples # <-- calling code will do the actual optimization here! + + for ((stacked_params, _state, _names), batch) in zip(tuples, batches): + for i, p in enumerate(batch): # batch is list of Parameter + p.copy_(stacked_params[i]) + + +class ScaledAdam(BatchedOptimizer): + """ + Implements 'Scaled Adam', a variant of Adam where we scale each parameter's update + proportional to the norm of that parameter; and also learn the scale of the parameter, + in log space, subject to upper and lower limits (as if we had factored each parameter as + param = underlying_param * log_scale.exp()) + + + Args: + params: The parameters or param_groups to optimize (like other Optimizer subclasses) + lr: The learning rate. We will typically use a learning rate schedule that starts + at 0.03 and decreases over time, i.e. much higher than other common + optimizers. + clipping_scale: (e.g. 2.0) + A scale for gradient-clipping: if specified, the normalized gradients + over the whole model will be clipped to have 2-norm equal to + `clipping_scale` times the median 2-norm over the most recent period + of `clipping_update_period` minibatches. By "normalized gradients", + we mean after multiplying by the rms parameter value for this tensor + [for non-scalars]; this is appropriate because our update is scaled + by this quantity. + betas: beta1,beta2 are momentum constants for regular momentum, and moving sum-sq grad. + Must satisfy 0 < beta <= beta2 < 1. + scalar_lr_scale: A scaling factor on the learning rate, that we use to update the + scale of each parameter tensor and scalar parameters of the mode.. + If each parameter were decomposed + as p * p_scale.exp(), where (p**2).mean().sqrt() == 1.0, scalar_lr_scale + would be a the scaling factor on the learning rate of p_scale. + eps: A general-purpose epsilon to prevent division by zero + param_min_rms: Minimum root-mean-square value of parameter tensor, for purposes of + learning the scale on the parameters (we'll constrain the rms of each non-scalar + parameter tensor to be >= this value) + param_max_rms: Maximum root-mean-square value of parameter tensor, for purposes of + learning the scale on the parameters (we'll constrain the rms of each non-scalar + parameter tensor to be <= this value) + scalar_max: Maximum absolute value for scalar parameters (applicable if your + model has any parameters with numel() == 1). + size_update_period: The periodicity, in steps, with which we update the size (scale) + of the parameter tensor. This is provided to save a little time + in the update. + clipping_update_period: if clipping_scale is specified, this is the period + """ + + def __init__( + self, + params, + lr=3e-02, + clipping_scale=None, + betas=(0.9, 0.98), + scalar_lr_scale=0.1, + eps=1.0e-08, + param_min_rms=1.0e-05, + param_max_rms=3.0, + scalar_max=10.0, + size_update_period=4, + clipping_update_period=100, + parameters_names=None, + show_dominant_parameters=True, + ): + + assert parameters_names is not None, ( + "Please prepare parameters_names," + "which is a List[List[str]]. Each List[str] is for a group" + "and each str is for a parameter" + ) + defaults = dict( + lr=lr, + clipping_scale=clipping_scale, + betas=betas, + scalar_lr_scale=scalar_lr_scale, + eps=eps, + param_min_rms=param_min_rms, + param_max_rms=param_max_rms, + scalar_max=scalar_max, + size_update_period=size_update_period, + clipping_update_period=clipping_update_period, + ) + + super(ScaledAdam, self).__init__(params, defaults) + assert len(self.param_groups) == len(parameters_names) + self.parameters_names = parameters_names + self.show_dominant_parameters = show_dominant_parameters + + def __setstate__(self, state): + super(ScaledAdam, self).__setstate__(state) + + @torch.no_grad() + def step(self, closure=None): + """Performs a single optimization step. + + Arguments: + closure (callable, optional): A closure that reevaluates the model + and returns the loss. + """ + loss = None + if closure is not None: + with torch.enable_grad(): + loss = closure() + + batch = True + + for group, group_params_names in zip(self.param_groups, self.parameters_names): + + with self.batched_params(group["params"], group_params_names) as batches: + + # batches is list of pairs (stacked_param, state). stacked_param is like + # a regular parameter, and will have a .grad, but the 1st dim corresponds to + # a stacking dim, it is not a real dim. + + if ( + len(batches[0][1]) == 0 + ): # if len(first state) == 0: not yet initialized + clipping_scale = 1 + else: + clipping_scale = self._get_clipping_scale(group, batches) + + for p, state, _ in batches: + # Perform optimization step. + # grad is not going to be None, we handled that when creating the batches. + grad = p.grad + if grad.is_sparse: + raise RuntimeError( + "ScaledAdam optimizer does not support sparse gradients" + ) + # State initialization + if len(state) == 0: + self._init_state(group, p, state) + + self._step_one_batch(group, p, state, clipping_scale) + + return loss + + def _init_state(self, group: dict, p: Tensor, state: dict): + """ + Initializes state dict for parameter 'p'. Assumes that dim 0 of tensor p + is actually the batch dimension, corresponding to batched-together + parameters of a given shape. + + + Args: + group: Dict to look up configuration values. + p: The parameter that we are initializing the state for + state: Dict from string to whatever state we are initializing + """ + size_update_period = group["size_update_period"] + + state["step"] = 0 + + kwargs = {"device": p.device, "dtype": p.dtype} + + # 'delta' implements conventional momentum. There are + # several different kinds of update going on, so rather than + # compute "exp_avg" like in Adam, we store and decay a + # parameter-change "delta", which combines all forms of + # update. this is equivalent to how it's done in Adam, + # except for the first few steps. + state["delta"] = torch.zeros_like(p, memory_format=torch.preserve_format) + + batch_size = p.shape[0] + numel = p.numel() // batch_size + numel = p.numel() + + if numel > 1: + # "param_rms" just periodically records the scalar root-mean-square value of + # the parameter tensor. + # it has a shape like (batch_size, 1, 1, 1, 1) + param_rms = (p**2).mean(dim=list(range(1, p.ndim)), keepdim=True).sqrt() + state["param_rms"] = param_rms + + state["scale_exp_avg_sq"] = torch.zeros_like(param_rms) + state["scale_grads"] = torch.zeros( + size_update_period, *param_rms.shape, **kwargs + ) + + # exp_avg_sq is the weighted sum of scaled gradients. as in Adam. + state["exp_avg_sq"] = torch.zeros_like(p, memory_format=torch.preserve_format) + + def _get_clipping_scale( + self, group: dict, tuples: List[Tuple[Tensor, dict, List[str]]] + ) -> float: + """ + Returns a scalar factor <= 1.0 that dictates gradient clipping, i.e. we will scale the gradients + by this amount before applying the rest of the update. + + Args: + group: the parameter group, an item in self.param_groups + tuples: a list of tuples of (param, state, param_names) + where param is a batched set of parameters, + with a .grad (1st dim is batch dim) + and state is the state-dict where optimization parameters are kept. + param_names is a List[str] while each str is name for a parameter + in batched set of parameters "param". + """ + assert len(tuples) >= 1 + clipping_scale = group["clipping_scale"] + (first_p, first_state, _) = tuples[0] + step = first_state["step"] + if clipping_scale is None or step == 0: + # no clipping. return early on step == 0 because the other + # parameters' state won't have been initialized yet. + return 1.0 + clipping_update_period = group["clipping_update_period"] + + tot_sumsq = torch.tensor(0.0, device=first_p.device) + for (p, state, param_names) in tuples: + grad = p.grad + if grad.is_sparse: + raise RuntimeError( + "ScaledAdam optimizer does not support sparse gradients" + ) + if p.numel() == p.shape[0]: # a batch of scalars + tot_sumsq += (grad**2).sum() # sum() to change shape [1] to [] + else: + tot_sumsq += ((grad * state["param_rms"]) ** 2).sum() + + tot_norm = tot_sumsq.sqrt() + if "model_norms" not in first_state: + first_state["model_norms"] = torch.zeros( + clipping_update_period, device=p.device + ) + first_state["model_norms"][step % clipping_update_period] = tot_norm + + if step % clipping_update_period == 0: + # Print some stats. + # We don't reach here if step == 0 because we would have returned + # above. + sorted_norms = first_state["model_norms"].sort()[0].to("cpu") + quartiles = [] + for n in range(0, 5): + index = min( + clipping_update_period - 1, (clipping_update_period // 4) * n + ) + quartiles.append(sorted_norms[index].item()) + + median = quartiles[2] + threshold = clipping_scale * median + first_state["model_norm_threshold"] = threshold + percent_clipped = ( + first_state["num_clipped"] * 100.0 / clipping_update_period + if "num_clipped" in first_state + else 0.0 + ) + first_state["num_clipped"] = 0 + quartiles = " ".join(["%.3e" % x for x in quartiles]) + logging.info( + f"Clipping_scale={clipping_scale}, grad-norm quartiles {quartiles}, " + f"threshold={threshold:.3e}, percent-clipped={percent_clipped:.1f}" + ) + + if step < clipping_update_period: + return 1.0 # We have not yet estimated a norm to clip to. + else: + try: + model_norm_threshold = first_state["model_norm_threshold"] + except KeyError: + logging.info( + "Warning: model_norm_threshold not in state: possibly " + "you changed config when restarting, adding clipping_scale option?" + ) + return 1.0 + ans = min(1.0, (model_norm_threshold / (tot_norm + 1.0e-20)).item()) + if ans < 1.0: + first_state["num_clipped"] += 1 + if ans < 0.1: + logging.warn( + f"Scaling gradients by {ans}, model_norm_threshold={model_norm_threshold}" + ) + if self.show_dominant_parameters: + assert p.shape[0] == len(param_names) + self._show_gradient_dominating_parameter(tuples, tot_sumsq) + return ans + + def _show_gradient_dominating_parameter( + self, tuples: List[Tuple[Tensor, dict, List[str]]], tot_sumsq: Tensor + ): + """ + Show information of parameter wihch dominanting tot_sumsq. + + Args: + tuples: a list of tuples of (param, state, param_names) + where param is a batched set of parameters, + with a .grad (1st dim is batch dim) + and state is the state-dict where optimization parameters are kept. + param_names is a List[str] while each str is name for a parameter + in batched set of parameters "param". + tot_sumsq: sumsq of all parameters. Though it's could be calculated + from tuples, we still pass it to save some time. + """ + all_sumsq_orig = {} + for (p, state, batch_param_names) in tuples: + # p is a stacked batch parameters. + batch_grad = p.grad + if p.numel() == p.shape[0]: # a batch of scalars + batch_sumsq_orig = batch_grad**2 + # Dummpy values used by following `zip` statement. + batch_rms_orig = torch.ones(p.shape[0]) + else: + batch_rms_orig = state["param_rms"] + batch_sumsq_orig = ((batch_grad * batch_rms_orig) ** 2).sum( + dim=list(range(1, batch_grad.ndim)) + ) + + for name, sumsq_orig, rms, grad in zip( + batch_param_names, batch_sumsq_orig, batch_rms_orig, batch_grad + ): + + proportion_orig = sumsq_orig / tot_sumsq + all_sumsq_orig[name] = (proportion_orig, sumsq_orig, rms, grad) + + assert torch.isclose( + sum([value[0] for value in all_sumsq_orig.values()]).cpu(), + torch.tensor(1.0), + ) + sorted_by_proportion = { + k: v + for k, v in sorted( + all_sumsq_orig.items(), key=lambda item: item[1][0], reverse=True + ) + } + dominant_param_name = next(iter(sorted_by_proportion)) + ( + dominant_proportion, + dominant_sumsq, + dominant_rms, + dominant_grad, + ) = sorted_by_proportion[dominant_param_name] + logging.info( + f"Parameter Dominanting tot_sumsq {dominant_param_name}" + f" with proportion {dominant_proportion:.2f}," + f" where dominant_sumsq=(grad_sumsq*orig_rms_sq)" + f"={dominant_sumsq:.3e}," + f" grad_sumsq = {(dominant_grad**2).sum():.3e}," + f" orig_rms_sq={(dominant_rms**2).item():.3e}" + ) + + def _step_one_batch( + self, group: dict, p: Tensor, state: dict, clipping_scale: float + ): + """ + Do the step for one parameter, which is actually going to be a batch of + `real` parameters, with dim 0 as the batch dim. + Args: + group: dict to look up configuration values + p: parameter to update (actually multiple parameters stacked together + as a batch) + state: state-dict for p, to look up the optimizer state + """ + lr = group["lr"] + size_update_period = group["size_update_period"] + beta1 = group["betas"][0] + + grad = p.grad + if clipping_scale != 1.0: + grad = grad * clipping_scale + step = state["step"] + delta = state["delta"] + + delta.mul_(beta1) + batch_size = p.shape[0] + numel = p.numel() // batch_size + if numel > 1: + # Update the size/scale of p, and set param_rms + scale_grads = state["scale_grads"] + scale_grads[step % size_update_period] = (p * grad).sum( + dim=list(range(1, p.ndim)), keepdim=True + ) + if step % size_update_period == size_update_period - 1: + param_rms = state["param_rms"] # shape: (batch_size, 1, 1, ..) + param_rms.copy_( + (p**2).mean(dim=list(range(1, p.ndim)), keepdim=True).sqrt() + ) + if step > 0: + # self._size_update() learns the overall scale on the + # parameter, by shrinking or expanding it. + self._size_update(group, scale_grads, p, state) + + if numel == 1: + # For parameters with 1 element we just use regular Adam. + # Updates delta. + self._step_scalar(group, p, state) + else: + self._step(group, p, state) + + state["step"] = step + 1 + + def _size_update( + self, group: dict, scale_grads: Tensor, p: Tensor, state: dict + ) -> None: + """ + Called only where p.numel() > 1, this updates the scale of the parameter. + If we imagine: p = underlying_param * scale.exp(), and we are doing + gradient descent on underlying param and on scale, this function does the update + on `scale`. + + Args: + group: dict to look up configuration values + scale_grads: a tensor of shape (size_update_period, batch_size, 1, 1,...) containing + grads w.r.t. the scales. + p: The parameter to update + state: The state-dict of p + """ + + param_rms = state["param_rms"] + beta1, beta2 = group["betas"] + size_lr = group["lr"] * group["scalar_lr_scale"] + param_min_rms = group["param_min_rms"] + param_max_rms = group["param_max_rms"] + eps = group["eps"] + step = state["step"] + batch_size = p.shape[0] + + size_update_period = scale_grads.shape[0] + # correct beta2 for the size update period: we will have + # faster decay at this level. + beta2_corr = beta2**size_update_period + + scale_exp_avg_sq = state["scale_exp_avg_sq"] # shape: (batch_size, 1, 1, ..) + scale_exp_avg_sq.mul_(beta2_corr).add_( + (scale_grads**2).mean(dim=0), # mean over dim `size_update_period` + alpha=1 - beta2_corr, + ) # shape is (batch_size, 1, 1, ...) + + # The 1st time we reach here is when size_step == 1. + size_step = (step + 1) // size_update_period + bias_correction2 = 1 - beta2_corr**size_step + # we don't bother with bias_correction1; this will help prevent divergence + # at the start of training. + + denom = scale_exp_avg_sq.sqrt() + eps + + scale_step = ( + -size_lr * (bias_correction2**0.5) * scale_grads.sum(dim=0) / denom + ) + + is_too_small = param_rms < param_min_rms + is_too_large = param_rms > param_max_rms + + # when the param gets too small, just don't shrink it any further. + scale_step.masked_fill_(is_too_small, 0.0) + # when it gets too large, stop it from getting any larger. + scale_step.masked_fill_(is_too_large, -size_lr * size_update_period) + delta = state["delta"] + # the factor of (1-beta1) relates to momentum. + delta.add_(p * scale_step, alpha=(1 - beta1)) + + def _step(self, group: dict, p: Tensor, state: dict): + """ + This function does the core update of self.step(), in the case where the members of + the batch have more than 1 element. + + Args: + group: A dict which will be used to look up configuration values + p: The parameter to be updated + grad: The grad of p + state: The state-dict corresponding to parameter p + + This function modifies p. + """ + grad = p.grad + lr = group["lr"] + beta1, beta2 = group["betas"] + eps = group["eps"] + param_min_rms = group["param_min_rms"] + step = state["step"] + + exp_avg_sq = state["exp_avg_sq"] + exp_avg_sq.mul_(beta2).addcmul_(grad, grad, value=(1 - beta2)) + + this_step = state["step"] - (state["zero_step"] if "zero_step" in state else 0) + bias_correction2 = 1 - beta2 ** (this_step + 1) + if bias_correction2 < 0.99: + # note: not in-place. + exp_avg_sq = exp_avg_sq * (1.0 / bias_correction2) + + denom = exp_avg_sq.sqrt() + denom += eps + grad = grad / denom + + alpha = -lr * (1 - beta1) * state["param_rms"].clamp(min=param_min_rms) + + delta = state["delta"] + delta.add_(grad * alpha) + p.add_(delta) + + def _step_scalar(self, group: dict, p: Tensor, state: dict): + """ + A simplified form of the core update for scalar tensors, where we cannot get a good + estimate of the parameter rms. + """ + beta1, beta2 = group["betas"] + scalar_max = group["scalar_max"] + eps = group["eps"] + lr = group["lr"] * group["scalar_lr_scale"] + grad = p.grad + + exp_avg_sq = state["exp_avg_sq"] # shape: (batch_size,) + exp_avg_sq.mul_(beta2).addcmul_(grad, grad, value=1 - beta2) + + # bias_correction2 is like in Adam. Don't bother with bias_correction1; + # slower update at the start will help stability anyway. + bias_correction2 = 1 - beta2 ** (state["step"] + 1) + denom = (exp_avg_sq / bias_correction2).sqrt() + eps + + delta = state["delta"] + delta.add_(grad / denom, alpha=-lr * (1 - beta1)) + p.clamp_(min=-scalar_max, max=scalar_max) + p.add_(delta) + + +class LRScheduler(object): + """ + Base-class for learning rate schedulers where the learning-rate depends on both the + batch and the epoch. + """ + + def __init__(self, optimizer: Optimizer, verbose: bool = False): + # Attach optimizer + if not isinstance(optimizer, Optimizer): + raise TypeError("{} is not an Optimizer".format(type(optimizer).__name__)) + self.optimizer = optimizer + self.verbose = verbose + + for group in optimizer.param_groups: + group.setdefault("base_lr", group["lr"]) + + self.base_lrs = [group["base_lr"] for group in optimizer.param_groups] + + self.epoch = 0 + self.batch = 0 + + def state_dict(self): + """Returns the state of the scheduler as a :class:`dict`. + + It contains an entry for every variable in self.__dict__ which + is not the optimizer. + """ + return { + "base_lrs": self.base_lrs, + "epoch": self.epoch, + "batch": self.batch, + } + + def load_state_dict(self, state_dict): + """Loads the schedulers state. + + Args: + state_dict (dict): scheduler state. Should be an object returned + from a call to :meth:`state_dict`. + """ + self.__dict__.update(state_dict) + + def get_last_lr(self) -> List[float]: + """Return last computed learning rate by current scheduler. Will be a list of float.""" + return self._last_lr + + def get_lr(self): + # Compute list of learning rates from self.epoch and self.batch and + # self.base_lrs; this must be overloaded by the user. + # e.g. return [some_formula(self.batch, self.epoch, base_lr) for base_lr in self.base_lrs ] + raise NotImplementedError + + def step_batch(self, batch: Optional[int] = None) -> None: + # Step the batch index, or just set it. If `batch` is specified, it + # must be the batch index from the start of training, i.e. summed over + # all epochs. + # You can call this in any order; if you don't provide 'batch', it should + # of course be called once per batch. + if batch is not None: + self.batch = batch + else: + self.batch = self.batch + 1 + self._set_lrs() + + def step_epoch(self, epoch: Optional[int] = None): + # Step the epoch index, or just set it. If you provide the 'epoch' arg, + # you should call this at the start of the epoch; if you don't provide the 'epoch' + # arg, you should call it at the end of the epoch. + if epoch is not None: + self.epoch = epoch + else: + self.epoch = self.epoch + 1 + self._set_lrs() + + def _set_lrs(self): + values = self.get_lr() + assert len(values) == len(self.optimizer.param_groups) + + for i, data in enumerate(zip(self.optimizer.param_groups, values)): + param_group, lr = data + param_group["lr"] = lr + self.print_lr(self.verbose, i, lr) + self._last_lr = [group["lr"] for group in self.optimizer.param_groups] + + def print_lr(self, is_verbose, group, lr): + """Display the current learning rate.""" + if is_verbose: + logging.info( + f"Epoch={self.epoch}, batch={self.batch}: adjusting learning rate" + f" of group {group} to {lr:.4e}." + ) + + +class Eden(LRScheduler): + """ + Eden scheduler. + The basic formula (before warmup) is: + lr = base_lr * (((batch**2 + lr_batches**2) / lr_batches**2) ** -0.25 * + (((epoch**2 + lr_epochs**2) / lr_epochs**2) ** -0.25)) * warmup + where `warmup` increases from linearly 0.5 to 1 over `warmup_batches` batches + and then stays constant at 1. + + + E.g. suggest base_lr = 0.04 (passed to optimizer) if used with ScaledAdam + + Args: + optimizer: the optimizer to change the learning rates on + lr_batches: the number of batches after which we start significantly + decreasing the learning rate, suggest 5000. + lr_epochs: the number of epochs after which we start significantly + decreasing the learning rate, suggest 6 if you plan to do e.g. + 20 to 40 epochs, but may need smaller number if dataset is huge + and you will do few epochs. + """ + + def __init__( + self, + optimizer: Optimizer, + lr_batches: Union[int, float], + lr_epochs: Union[int, float], + warmup_batches: Union[int, float] = 500.0, + verbose: bool = False, + ): + super(Eden, self).__init__(optimizer, verbose) + self.lr_batches = lr_batches + self.lr_epochs = lr_epochs + self.warmup_batches = warmup_batches + + def get_lr(self): + factor = ( + (self.batch**2 + self.lr_batches**2) / self.lr_batches**2 + ) ** -0.25 * ( + ((self.epoch**2 + self.lr_epochs**2) / self.lr_epochs**2) ** -0.25 + ) + warmup_factor = ( + 1.0 + if self.batch >= self.warmup_batches + else 0.5 + 0.5 * (self.batch / self.warmup_batches) + ) + + return [x * factor * warmup_factor for x in self.base_lrs] + + +def _test_eden(): + m = torch.nn.Linear(100, 100) + optim = ScaledAdam(m.parameters(), lr=0.03) + + scheduler = Eden(optim, lr_batches=100, lr_epochs=2, verbose=True) + + for epoch in range(10): + scheduler.step_epoch(epoch) # sets epoch to `epoch` + + for step in range(20): + x = torch.randn(200, 100).detach() + x.requires_grad = True + y = m(x) + dy = torch.randn(200, 100).detach() + f = (y * dy).sum() + f.backward() + + optim.step() + scheduler.step_batch() + optim.zero_grad() + + logging.info(f"last lr = {scheduler.get_last_lr()}") + logging.info(f"state dict = {scheduler.state_dict()}") + + +# This is included mostly as a baseline for ScaledAdam. +class Eve(Optimizer): + """ + Implements Eve algorithm. This is a modified version of AdamW with a special + way of setting the weight-decay / shrinkage-factor, which is designed to make the + rms of the parameters approach a particular target_rms (default: 0.1). This is + for use with networks with 'scaled' versions of modules (see scaling.py), which + will be close to invariant to the absolute scale on the parameter matrix. + + The original Adam algorithm was proposed in `Adam: A Method for Stochastic Optimization`_. + The AdamW variant was proposed in `Decoupled Weight Decay Regularization`_. + Eve is unpublished so far. + + Arguments: + params (iterable): iterable of parameters to optimize or dicts defining + parameter groups + lr (float, optional): learning rate (default: 1e-3) + betas (Tuple[float, float], optional): coefficients used for computing + running averages of gradient and its square (default: (0.9, 0.999)) + eps (float, optional): term added to the denominator to improve + numerical stability (default: 1e-8) + weight_decay (float, optional): weight decay coefficient (default: 3e-4; + this value means that the weight would decay significantly after + about 3k minibatches. Is not multiplied by learning rate, but + is conditional on RMS-value of parameter being > target_rms. + target_rms (float, optional): target root-mean-square value of + parameters, if they fall below this we will stop applying weight decay. + + + .. _Adam: A Method for Stochastic Optimization: + https://arxiv.org/abs/1412.6980 + .. _Decoupled Weight Decay Regularization: + https://arxiv.org/abs/1711.05101 + .. _On the Convergence of Adam and Beyond: + https://openreview.net/forum?id=ryQu7f-RZ + """ + + def __init__( + self, + params, + lr=1e-3, + betas=(0.9, 0.98), + eps=1e-8, + weight_decay=1e-3, + target_rms=0.1, + ): + if not 0.0 <= lr: + raise ValueError("Invalid learning rate: {}".format(lr)) + if not 0.0 <= eps: + raise ValueError("Invalid epsilon value: {}".format(eps)) + if not 0.0 <= betas[0] < 1.0: + raise ValueError("Invalid beta parameter at index 0: {}".format(betas[0])) + if not 0.0 <= betas[1] < 1.0: + raise ValueError("Invalid beta parameter at index 1: {}".format(betas[1])) + if not 0 <= weight_decay <= 0.1: + raise ValueError("Invalid weight_decay value: {}".format(weight_decay)) + if not 0 < target_rms <= 10.0: + raise ValueError("Invalid target_rms value: {}".format(target_rms)) + defaults = dict( + lr=lr, + betas=betas, + eps=eps, + weight_decay=weight_decay, + target_rms=target_rms, + ) + super(Eve, self).__init__(params, defaults) + + def __setstate__(self, state): + super(Eve, self).__setstate__(state) + + @torch.no_grad() + def step(self, closure=None): + """Performs a single optimization step. + + Arguments: + closure (callable, optional): A closure that reevaluates the model + and returns the loss. + """ + loss = None + if closure is not None: + with torch.enable_grad(): + loss = closure() + + for group in self.param_groups: + for p in group["params"]: + if p.grad is None: + continue + + # Perform optimization step + grad = p.grad + if grad.is_sparse: + raise RuntimeError("AdamW does not support sparse gradients") + + state = self.state[p] + + # State initialization + if len(state) == 0: + state["step"] = 0 + # Exponential moving average of gradient values + state["exp_avg"] = torch.zeros_like( + p, memory_format=torch.preserve_format + ) + # Exponential moving average of squared gradient values + state["exp_avg_sq"] = torch.zeros_like( + p, memory_format=torch.preserve_format + ) + + exp_avg, exp_avg_sq = state["exp_avg"], state["exp_avg_sq"] + + beta1, beta2 = group["betas"] + + state["step"] += 1 + bias_correction1 = 1 - beta1 ** state["step"] + bias_correction2 = 1 - beta2 ** state["step"] + + # Decay the first and second moment running average coefficient + exp_avg.mul_(beta1).add_(grad, alpha=1 - beta1) + exp_avg_sq.mul_(beta2).addcmul_(grad, grad, value=1 - beta2) + denom = (exp_avg_sq.sqrt() * (bias_correction2**-0.5)).add_( + group["eps"] + ) + + step_size = group["lr"] / bias_correction1 + target_rms = group["target_rms"] + weight_decay = group["weight_decay"] + + if p.numel() > 1: + # avoid applying this weight-decay on "scaling factors" + # (which are scalar). + is_above_target_rms = p.norm() > (target_rms * (p.numel() ** 0.5)) + p.mul_(1 - (weight_decay * is_above_target_rms)) + + p.addcdiv_(exp_avg, denom, value=-step_size) + + if random.random() < 0.0005: + step = (exp_avg / denom) * step_size + logging.info( + f"Delta rms = {(step**2).mean().item()}, shape = {step.shape}" + ) + + return loss + + +def _test_scaled_adam(hidden_dim: int): + import timeit + + from scaling import ScaledLinear + + E = 100 + B = 4 + T = 2 + logging.info("in test_eve_cain") + # device = torch.device('cuda') + device = torch.device("cpu") + dtype = torch.float32 + + fix_random_seed(42) + # these input_magnitudes and output_magnitudes are to test that + # Abel is working as we expect and is able to adjust scales of + # different dims differently. + input_magnitudes = (1.0 * torch.randn(E, dtype=dtype, device=device)).exp() + output_magnitudes = (1.0 * torch.randn(E, dtype=dtype, device=device)).exp() + + for iter in [1, 0]: + fix_random_seed(42) + Linear = torch.nn.Linear if iter == 0 else ScaledLinear + + m = torch.nn.Sequential( + Linear(E, hidden_dim), + torch.nn.PReLU(), + Linear(hidden_dim, hidden_dim), + torch.nn.PReLU(), + Linear(hidden_dim, E), + ).to(device) + + train_pairs = [ + ( + 100.0 + * torch.randn(B, T, E, device=device, dtype=dtype) + * input_magnitudes, + torch.randn(B, T, E, device=device, dtype=dtype) * output_magnitudes, + ) + for _ in range(20) + ] + + if iter == 0: + optim = Eve(m.parameters(), lr=0.003) + elif iter == 1: + optim = ScaledAdam(m.parameters(), lr=0.03, clipping_scale=2.0) + scheduler = Eden(optim, lr_batches=200, lr_epochs=5, verbose=False) + + start = timeit.default_timer() + avg_loss = 0.0 + for epoch in range(180): + scheduler.step_epoch() + # if epoch == 100 and iter in [2,3]: + # optim.reset_speedup() # check it doesn't crash. + + # if epoch == 130: + # opts = diagnostics.TensorDiagnosticOptions( + # 2 ** 22 + # ) # allow 4 megabytes per sub-module + # diagnostic = diagnostics.attach_diagnostics(m, opts) + + for n, (x, y) in enumerate(train_pairs): + y_out = m(x) + loss = ((y_out - y) ** 2).mean() * 100.0 + if epoch == 0 and n == 0: + avg_loss = loss.item() + else: + avg_loss = 0.98 * avg_loss + 0.02 * loss.item() + if n == 0 and epoch % 5 == 0: + # norm1 = '%.2e' % (m[0].weight**2).mean().sqrt().item() + # norm1b = '%.2e' % (m[0].bias**2).mean().sqrt().item() + # norm2 = '%.2e' % (m[2].weight**2).mean().sqrt().item() + # norm2b = '%.2e' % (m[2].bias**2).mean().sqrt().item() + # scale1 = '%.2e' % (m[0].weight_scale.exp().item()) + # scale1b = '%.2e' % (m[0].bias_scale.exp().item()) + # scale2 = '%.2e' % (m[2].weight_scale.exp().item()) + # scale2b = '%.2e' % (m[2].bias_scale.exp().item()) + lr = scheduler.get_last_lr()[0] + logging.info( + f"Iter {iter}, epoch {epoch}, batch {n}, avg_loss {avg_loss:.4g}, lr={lr:.4e}" + ) # , norms={norm1,norm1b,norm2,norm2b}") # scales={scale1,scale1b,scale2,scale2b} + loss.log().backward() + optim.step() + optim.zero_grad() + scheduler.step_batch() + + # diagnostic.print_diagnostics() + + stop = timeit.default_timer() + logging.info(f"Iter={iter}, Time taken: {stop - start}") + + logging.info(f"last lr = {scheduler.get_last_lr()}") + # logging.info("state dict = ", scheduler.state_dict()) + # logging.info("optim state_dict = ", optim.state_dict()) + logging.info(f"input_magnitudes = {input_magnitudes}") + logging.info(f"output_magnitudes = {output_magnitudes}") + + +if __name__ == "__main__": + torch.set_num_threads(1) + torch.set_num_interop_threads(1) + logging.getLogger().setLevel(logging.INFO) + import subprocess + + s = subprocess.check_output( + "git status -uno .; git log -1; git diff HEAD .", shell=True + ) + logging.info(s) + import sys + + if len(sys.argv) > 1: + hidden_dim = int(sys.argv[1]) + else: + hidden_dim = 200 + + _test_scaled_adam(hidden_dim) + _test_eden() diff --git a/egs/tedlium3/ASR/pruned_transducer_stateless_d2v_v2/pretrained.py b/egs/tedlium3/ASR/pruned_transducer_stateless_d2v_v2/pretrained.py new file mode 100755 index 000000000..2f1b1a49f --- /dev/null +++ b/egs/tedlium3/ASR/pruned_transducer_stateless_d2v_v2/pretrained.py @@ -0,0 +1,353 @@ +#!/usr/bin/env python3 +# Copyright 2021 Xiaomi Corp. (authors: Fangjun Kuang) +# +# See ../../../../LICENSE for clarification regarding multiple authors +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +""" +This script loads a checkpoint and uses it to decode waves. +You can generate the checkpoint with the following command: + +./pruned_transducer_stateless7_ctc/export.py \ + --exp-dir ./pruned_transducer_stateless7_ctc/exp \ + --bpe-model data/lang_bpe_500/bpe.model \ + --epoch 20 \ + --avg 10 + +Usage of this script: + +(1) greedy search +./pruned_transducer_stateless7_ctc/pretrained.py \ + --checkpoint ./pruned_transducer_stateless7_ctc/exp/pretrained.pt \ + --bpe-model ./data/lang_bpe_500/bpe.model \ + --method greedy_search \ + /path/to/foo.wav \ + /path/to/bar.wav + +(2) beam search +./pruned_transducer_stateless7_ctc/pretrained.py \ + --checkpoint ./pruned_transducer_stateless7_ctc/exp/pretrained.pt \ + --bpe-model ./data/lang_bpe_500/bpe.model \ + --method beam_search \ + --beam-size 4 \ + /path/to/foo.wav \ + /path/to/bar.wav + +(3) modified beam search +./pruned_transducer_stateless7_ctc/pretrained.py \ + --checkpoint ./pruned_transducer_stateless7_ctc/exp/pretrained.pt \ + --bpe-model ./data/lang_bpe_500/bpe.model \ + --method modified_beam_search \ + --beam-size 4 \ + /path/to/foo.wav \ + /path/to/bar.wav + +(4) fast beam search +./pruned_transducer_stateless7_ctc/pretrained.py \ + --checkpoint ./pruned_transducer_stateless7_ctc/exp/pretrained.pt \ + --bpe-model ./data/lang_bpe_500/bpe.model \ + --method fast_beam_search \ + --beam-size 4 \ + /path/to/foo.wav \ + /path/to/bar.wav + +You can also use `./pruned_transducer_stateless7_ctc/exp/epoch-xx.pt`. + +Note: ./pruned_transducer_stateless7_ctc/exp/pretrained.pt is generated by +./pruned_transducer_stateless7_ctc/export.py +""" + + +import argparse +import logging +import math +from typing import List + +import k2 +import kaldifeat +import sentencepiece as spm +import torch +import torchaudio +from beam_search import ( + beam_search, + fast_beam_search_one_best, + greedy_search, + greedy_search_batch, + modified_beam_search, +) +from torch.nn.utils.rnn import pad_sequence +from train import add_model_arguments, get_params, get_transducer_model + + +def get_parser(): + parser = argparse.ArgumentParser( + formatter_class=argparse.ArgumentDefaultsHelpFormatter + ) + + parser.add_argument( + "--checkpoint", + type=str, + required=True, + help="Path to the checkpoint. " + "The checkpoint is assumed to be saved by " + "icefall.checkpoint.save_checkpoint().", + ) + + parser.add_argument( + "--bpe-model", + type=str, + help="""Path to bpe.model.""", + ) + + parser.add_argument( + "--method", + type=str, + default="greedy_search", + help="""Possible values are: + - greedy_search + - beam_search + - modified_beam_search + - fast_beam_search + """, + ) + + parser.add_argument( + "sound_files", + type=str, + nargs="+", + help="The input sound file(s) to transcribe. " + "Supported formats are those supported by torchaudio.load(). " + "For example, wav and flac are supported. " + "The sample rate has to be 16kHz.", + ) + + parser.add_argument( + "--sample-rate", + type=int, + default=16000, + help="The sample rate of the input sound file", + ) + + parser.add_argument( + "--beam-size", + type=int, + default=4, + help="""An integer indicating how many candidates we will keep for each + frame. Used only when --method is beam_search or + modified_beam_search.""", + ) + + parser.add_argument( + "--beam", + type=float, + default=4, + help="""A floating point value to calculate the cutoff score during beam + search (i.e., `cutoff = max-score - beam`), which is the same as the + `beam` in Kaldi. + Used only when --method is fast_beam_search""", + ) + + parser.add_argument( + "--max-contexts", + type=int, + default=4, + help="""Used only when --method is fast_beam_search""", + ) + + parser.add_argument( + "--max-states", + type=int, + default=8, + help="""Used only when --method is fast_beam_search""", + ) + + parser.add_argument( + "--context-size", + type=int, + default=2, + help="The context size in the decoder. 1 means bigram; 2 means tri-gram", + ) + parser.add_argument( + "--max-sym-per-frame", + type=int, + default=1, + help="""Maximum number of symbols per frame. Used only when + --method is greedy_search. + """, + ) + + add_model_arguments(parser) + + return parser + + +def read_sound_files( + filenames: List[str], expected_sample_rate: float +) -> List[torch.Tensor]: + """Read a list of sound files into a list 1-D float32 torch tensors. + Args: + filenames: + A list of sound filenames. + expected_sample_rate: + The expected sample rate of the sound files. + Returns: + Return a list of 1-D float32 torch tensors. + """ + ans = [] + for f in filenames: + wave, sample_rate = torchaudio.load(f) + assert ( + sample_rate == expected_sample_rate + ), f"Expected sample rate: {expected_sample_rate}. Given: {sample_rate}" + # We use only the first channel + ans.append(wave[0]) + return ans + + +@torch.no_grad() +def main(): + parser = get_parser() + args = parser.parse_args() + + params = get_params() + + params.update(vars(args)) + + sp = spm.SentencePieceProcessor() + sp.load(params.bpe_model) + + # is defined in local/train_bpe_model.py + params.blank_id = sp.piece_to_id("") + params.unk_id = sp.piece_to_id("") + params.vocab_size = sp.get_piece_size() + + logging.info(f"{params}") + + device = torch.device("cpu") + if torch.cuda.is_available(): + device = torch.device("cuda", 0) + + logging.info(f"device: {device}") + + logging.info("Creating model") + model = get_transducer_model(params) + + num_param = sum([p.numel() for p in model.parameters()]) + logging.info(f"Number of model parameters: {num_param}") + + checkpoint = torch.load(args.checkpoint, map_location="cpu") + model.load_state_dict(checkpoint["model"], strict=False) + model.to(device) + model.eval() + model.device = device + + logging.info("Constructing Fbank computer") + opts = kaldifeat.FbankOptions() + opts.device = device + opts.frame_opts.dither = 0 + opts.frame_opts.snip_edges = False + opts.frame_opts.samp_freq = params.sample_rate + opts.mel_opts.num_bins = params.feature_dim + + fbank = kaldifeat.Fbank(opts) + + logging.info(f"Reading sound files: {params.sound_files}") + waves = read_sound_files( + filenames=params.sound_files, expected_sample_rate=params.sample_rate + ) + waves = [w.to(device) for w in waves] + + logging.info("Decoding started") + features = fbank(waves) + feature_lengths = [f.size(0) for f in features] + + features = pad_sequence(features, batch_first=True, padding_value=math.log(1e-10)) + + feature_lengths = torch.tensor(feature_lengths, device=device) + + encoder_out, encoder_out_lens = model.encoder(x=features, x_lens=feature_lengths) + + num_waves = encoder_out.size(0) + hyps = [] + msg = f"Using {params.method}" + if params.method == "beam_search": + msg += f" with beam size {params.beam_size}" + logging.info(msg) + + if params.method == "fast_beam_search": + decoding_graph = k2.trivial_graph(params.vocab_size - 1, device=device) + hyp_tokens = fast_beam_search_one_best( + model=model, + decoding_graph=decoding_graph, + encoder_out=encoder_out, + encoder_out_lens=encoder_out_lens, + beam=params.beam, + max_contexts=params.max_contexts, + max_states=params.max_states, + ) + for hyp in sp.decode(hyp_tokens): + hyps.append(hyp.split()) + elif params.method == "modified_beam_search": + hyp_tokens = modified_beam_search( + model=model, + encoder_out=encoder_out, + encoder_out_lens=encoder_out_lens, + beam=params.beam_size, + ) + + for hyp in sp.decode(hyp_tokens): + hyps.append(hyp.split()) + elif params.method == "greedy_search" and params.max_sym_per_frame == 1: + hyp_tokens = greedy_search_batch( + model=model, + encoder_out=encoder_out, + encoder_out_lens=encoder_out_lens, + ) + for hyp in sp.decode(hyp_tokens): + hyps.append(hyp.split()) + else: + for i in range(num_waves): + # fmt: off + encoder_out_i = encoder_out[i:i+1, :encoder_out_lens[i]] + # fmt: on + if params.method == "greedy_search": + hyp = greedy_search( + model=model, + encoder_out=encoder_out_i, + max_sym_per_frame=params.max_sym_per_frame, + ) + elif params.method == "beam_search": + hyp = beam_search( + model=model, + encoder_out=encoder_out_i, + beam=params.beam_size, + ) + else: + raise ValueError(f"Unsupported method: {params.method}") + + hyps.append(sp.decode(hyp).split()) + + s = "\n" + for filename, hyp in zip(params.sound_files, hyps): + words = " ".join(hyp) + s += f"{filename}:\n{words}\n\n" + logging.info(s) + + logging.info("Decoding Done") + + +if __name__ == "__main__": + formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s" + + logging.basicConfig(format=formatter, level=logging.INFO) + main() diff --git a/egs/tedlium3/ASR/pruned_transducer_stateless_d2v_v2/pretrained_ctc.py b/egs/tedlium3/ASR/pruned_transducer_stateless_d2v_v2/pretrained_ctc.py new file mode 100755 index 000000000..74aef1bc7 --- /dev/null +++ b/egs/tedlium3/ASR/pruned_transducer_stateless_d2v_v2/pretrained_ctc.py @@ -0,0 +1,441 @@ +#!/usr/bin/env python3 +# Copyright 2022 Xiaomi Corp. (authors: Fangjun Kuang, +# Zengwei Yao) +# +# See ../../../../LICENSE for clarification regarding multiple authors +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +""" +This script loads torchscript models, exported by `torch.jit.script()` +and uses them to decode waves. +You can use the following command to get the exported models: + +./pruned_transducer_stateless7_ctc/export.py \ + --exp-dir ./pruned_transducer_stateless7_ctc/exp \ + --bpe-model data/lang_bpe_500/bpe.model \ + --epoch 20 \ + --avg 10 + +Usage of this script: + +(1) ctc-decoding +./pruned_transducer_stateless7_ctc/jit_pretrained_ctc.py \ + --checkpoint ./pruned_transducer_stateless7_ctc/exp/pretrained.pt \ + --bpe-model data/lang_bpe_500/bpe.model \ + --method ctc-decoding \ + --sample-rate 16000 \ + /path/to/foo.wav \ + /path/to/bar.wav + +(2) 1best +./pruned_transducer_stateless7_ctc/jit_pretrained_ctc.py \ + --checkpoint ./pruned_transducer_stateless7_ctc/exp/pretrained.pt \ + --HLG data/lang_bpe_500/HLG.pt \ + --words-file data/lang_bpe_500/words.txt \ + --method 1best \ + --sample-rate 16000 \ + /path/to/foo.wav \ + /path/to/bar.wav + +(3) nbest-rescoring +./bruned_transducer_stateless7_ctc/jit_pretrained_ctc.py \ + --checkpoint ./pruned_transducer_stateless7_ctc/exp/pretrained.pt \ + --HLG data/lang_bpe_500/HLG.pt \ + --words-file data/lang_bpe_500/words.txt \ + --G data/lm/G_4_gram.pt \ + --method nbest-rescoring \ + --sample-rate 16000 \ + /path/to/foo.wav \ + /path/to/bar.wav + + +(4) whole-lattice-rescoring +./pruned_transducer_stateless7_ctc/jit_pretrained_ctc.py \ + --checkpoint ./pruned_transducer_stateless7_ctc/exp/pretrained.pt \ + --HLG data/lang_bpe_500/HLG.pt \ + --words-file data/lang_bpe_500/words.txt \ + --G data/lm/G_4_gram.pt \ + --method whole-lattice-rescoring \ + --sample-rate 16000 \ + /path/to/foo.wav \ + /path/to/bar.wav +""" + +import argparse +import logging +import math +from typing import List + +import k2 +import kaldifeat +import sentencepiece as spm +import torch +import torchaudio +from ctc_decode import get_decoding_params +from torch.nn.utils.rnn import pad_sequence +from train import add_model_arguments, get_params, get_transducer_model + +from icefall.decode import ( + get_lattice, + one_best_decoding, + rescore_with_n_best_list, + rescore_with_whole_lattice, +) +from icefall.utils import get_texts + + +def get_parser(): + parser = argparse.ArgumentParser( + formatter_class=argparse.ArgumentDefaultsHelpFormatter + ) + + parser.add_argument( + "--checkpoint", + type=str, + required=True, + help="Path to the checkpoint. " + "The checkpoint is assumed to be saved by " + "icefall.checkpoint.save_checkpoint().", + ) + + parser.add_argument( + "--context-size", + type=int, + default=2, + help="The context size in the decoder. 1 means bigram; " "2 means tri-gram", + ) + + parser.add_argument( + "--words-file", + type=str, + help="""Path to words.txt. + Used only when method is not ctc-decoding. + """, + ) + + parser.add_argument( + "--HLG", + type=str, + help="""Path to HLG.pt. + Used only when method is not ctc-decoding. + """, + ) + + parser.add_argument( + "--bpe-model", + type=str, + help="""Path to bpe.model. + Used only when method is ctc-decoding. + """, + ) + + parser.add_argument( + "--method", + type=str, + default="1best", + help="""Decoding method. + Possible values are: + (0) ctc-decoding - Use CTC decoding. It uses a sentence + piece model, i.e., lang_dir/bpe.model, to convert + word pieces to words. It needs neither a lexicon + nor an n-gram LM. + (1) 1best - Use the best path as decoding output. Only + the transformer encoder output is used for decoding. + We call it HLG decoding. + (2) nbest-rescoring. Extract n paths from the decoding lattice, + rescore them with an LM, the path with + the highest score is the decoding result. + We call it HLG decoding + n-gram LM rescoring. + (3) whole-lattice-rescoring - Use an LM to rescore the + decoding lattice and then use 1best to decode the + rescored lattice. + We call it HLG decoding + n-gram LM rescoring. + """, + ) + + parser.add_argument( + "--G", + type=str, + help="""An LM for rescoring. + Used only when method is + whole-lattice-rescoring or nbest-rescoring. + It's usually a 4-gram LM. + """, + ) + + parser.add_argument( + "--num-paths", + type=int, + default=100, + help=""" + Used only when method is attention-decoder. + It specifies the size of n-best list.""", + ) + + parser.add_argument( + "--ngram-lm-scale", + type=float, + default=1.3, + help=""" + Used only when method is whole-lattice-rescoring and nbest-rescoring. + It specifies the scale for n-gram LM scores. + (Note: You need to tune it on a dataset.) + """, + ) + + parser.add_argument( + "--nbest-scale", + type=float, + default=0.5, + help=""" + Used only when method is nbest-rescoring. + It specifies the scale for lattice.scores when + extracting n-best lists. A smaller value results in + more unique number of paths with the risk of missing + the best path. + """, + ) + + parser.add_argument( + "--num-classes", + type=int, + default=500, + help=""" + Vocab size in the BPE model. + """, + ) + + parser.add_argument( + "--sample-rate", + type=int, + default=16000, + help="The sample rate of the input sound file", + ) + + parser.add_argument( + "sound_files", + type=str, + nargs="+", + help="The input sound file(s) to transcribe. " + "Supported formats are those supported by torchaudio.load(). " + "For example, wav and flac are supported. " + "The sample rate has to be 16kHz.", + ) + + add_model_arguments(parser) + + return parser + + +def read_sound_files( + filenames: List[str], expected_sample_rate: float = 16000 +) -> List[torch.Tensor]: + """Read a list of sound files into a list 1-D float32 torch tensors. + Args: + filenames: + A list of sound filenames. + expected_sample_rate: + The expected sample rate of the sound files. + Returns: + Return a list of 1-D float32 torch tensors. + """ + ans = [] + for f in filenames: + wave, sample_rate = torchaudio.load(f) + assert sample_rate == expected_sample_rate, ( + f"expected sample rate: {expected_sample_rate}. " f"Given: {sample_rate}" + ) + # We use only the first channel + ans.append(wave[0]) + return ans + + +@torch.no_grad() +def main(): + parser = get_parser() + args = parser.parse_args() + + params = get_params() + # add decoding params + params.update(get_decoding_params()) + params.update(vars(args)) + params.vocab_size = params.num_classes + params.blank_id = 0 + + logging.info(f"{params}") + + device = torch.device("cpu") + if torch.cuda.is_available(): + device = torch.device("cuda", 0) + + logging.info(f"device: {device}") + + logging.info("Creating model") + model = get_transducer_model(params) + + num_param = sum([p.numel() for p in model.parameters()]) + logging.info(f"Number of model parameters: {num_param}") + + checkpoint = torch.load(args.checkpoint, map_location="cpu") + model.load_state_dict(checkpoint["model"], strict=False) + model.to(device) + model.eval() + + logging.info("Constructing Fbank computer") + opts = kaldifeat.FbankOptions() + opts.device = device + opts.frame_opts.dither = 0 + opts.frame_opts.snip_edges = False + opts.frame_opts.samp_freq = params.sample_rate + opts.mel_opts.num_bins = params.feature_dim + + fbank = kaldifeat.Fbank(opts) + + logging.info(f"Reading sound files: {params.sound_files}") + waves = read_sound_files( + filenames=params.sound_files, expected_sample_rate=params.sample_rate + ) + waves = [w.to(device) for w in waves] + + logging.info("Decoding started") + features = fbank(waves) + feature_lengths = [f.size(0) for f in features] + + features = pad_sequence(features, batch_first=True, padding_value=math.log(1e-10)) + feature_lengths = torch.tensor(feature_lengths, device=device) + + encoder_out, encoder_out_lens = model.encoder( + x=features, + x_lens=feature_lengths, + ) + nnet_output = model.ctc_output(encoder_out) + + batch_size = nnet_output.shape[0] + supervision_segments = torch.tensor( + [[i, 0, nnet_output.shape[1]] for i in range(batch_size)], + dtype=torch.int32, + ) + + if params.method == "ctc-decoding": + logging.info("Use CTC decoding") + bpe_model = spm.SentencePieceProcessor() + bpe_model.load(params.bpe_model) + max_token_id = params.num_classes - 1 + + H = k2.ctc_topo( + max_token=max_token_id, + modified=False, + device=device, + ) + + lattice = get_lattice( + nnet_output=nnet_output, + decoding_graph=H, + supervision_segments=supervision_segments, + search_beam=params.search_beam, + output_beam=params.output_beam, + min_active_states=params.min_active_states, + max_active_states=params.max_active_states, + subsampling_factor=params.subsampling_factor, + ) + + best_path = one_best_decoding( + lattice=lattice, use_double_scores=params.use_double_scores + ) + token_ids = get_texts(best_path) + hyps = bpe_model.decode(token_ids) + hyps = [s.split() for s in hyps] + elif params.method in [ + "1best", + "nbest-rescoring", + "whole-lattice-rescoring", + ]: + logging.info(f"Loading HLG from {params.HLG}") + HLG = k2.Fsa.from_dict(torch.load(params.HLG, map_location="cpu")) + HLG = HLG.to(device) + if not hasattr(HLG, "lm_scores"): + # For whole-lattice-rescoring and attention-decoder + HLG.lm_scores = HLG.scores.clone() + + if params.method in [ + "nbest-rescoring", + "whole-lattice-rescoring", + ]: + logging.info(f"Loading G from {params.G}") + G = k2.Fsa.from_dict(torch.load(params.G, map_location="cpu")) + G = G.to(device) + if params.method == "whole-lattice-rescoring": + # Add epsilon self-loops to G as we will compose + # it with the whole lattice later + G = k2.add_epsilon_self_loops(G) + G = k2.arc_sort(G) + + # G.lm_scores is used to replace HLG.lm_scores during + # LM rescoring. + G.lm_scores = G.scores.clone() + + lattice = get_lattice( + nnet_output=nnet_output, + decoding_graph=HLG, + supervision_segments=supervision_segments, + search_beam=params.search_beam, + output_beam=params.output_beam, + min_active_states=params.min_active_states, + max_active_states=params.max_active_states, + subsampling_factor=params.subsampling_factor, + ) + + if params.method == "1best": + logging.info("Use HLG decoding") + best_path = one_best_decoding( + lattice=lattice, use_double_scores=params.use_double_scores + ) + if params.method == "nbest-rescoring": + logging.info("Use HLG decoding + LM rescoring") + best_path_dict = rescore_with_n_best_list( + lattice=lattice, + G=G, + num_paths=params.num_paths, + lm_scale_list=[params.ngram_lm_scale], + nbest_scale=params.nbest_scale, + ) + best_path = next(iter(best_path_dict.values())) + elif params.method == "whole-lattice-rescoring": + logging.info("Use HLG decoding + LM rescoring") + best_path_dict = rescore_with_whole_lattice( + lattice=lattice, + G_with_epsilon_loops=G, + lm_scale_list=[params.ngram_lm_scale], + ) + best_path = next(iter(best_path_dict.values())) + + hyps = get_texts(best_path) + word_sym_table = k2.SymbolTable.from_file(params.words_file) + hyps = [[word_sym_table[i] for i in ids] for ids in hyps] + else: + raise ValueError(f"Unsupported decoding method: {params.method}") + + s = "\n" + for filename, hyp in zip(params.sound_files, hyps): + words = " ".join(hyp) + s += f"{filename}:\n{words}\n\n" + logging.info(s) + + logging.info("Decoding Done") + + +if __name__ == "__main__": + formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s" + + logging.basicConfig(format=formatter, level=logging.INFO) + main() diff --git a/egs/tedlium3/ASR/pruned_transducer_stateless_d2v_v2/prompt_tuning.py b/egs/tedlium3/ASR/pruned_transducer_stateless_d2v_v2/prompt_tuning.py new file mode 100755 index 000000000..adc77f8f9 --- /dev/null +++ b/egs/tedlium3/ASR/pruned_transducer_stateless_d2v_v2/prompt_tuning.py @@ -0,0 +1,1825 @@ +#!/usr/bin/env python3 +# Copyright 2021-2022 Xiaomi Corp. (authors: Fangjun Kuang, +# Wei Kang, +# Mingshuang Luo,) +# Zengwei Yao) +# +# See ../../../../LICENSE for clarification regarding multiple authors +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +""" +Usage: + +export CUDA_VISIBLE_DEVICES="0,1,2,3" + +./pruned_transducer_stateless7_ctc/train.py \ + --world-size 4 \ + --num-epochs 30 \ + --start-epoch 1 \ + --exp-dir pruned_transducer_stateless7_ctc/exp \ + --full-libri 1 \ + --max-duration 300 + +# For mix precision training: + +./pruned_transducer_stateless7_ctc/train.py \ + --world-size 4 \ + --num-epochs 30 \ + --start-epoch 1 \ + --use-fp16 1 \ + --exp-dir pruned_transducer_stateless7_ctc/exp \ + --full-libri 1 \ + --max-duration 550 + +# For d2v-T training: +export CUDA_VISIBLE_DEVICES="0,1,2,3,4,5,6,7" + +./pruned_transducer_stateless_d2v_v2/train.py \ + --wandb true \ + --input-strategy AudioSamples \ + --enable-spec-aug False \ + --multi-optim True \ + --world-size 8 \ + --num-epochs 30 \ + --start-epoch 1 \ + --full-libri 0 \ + --exp-dir ./pruned_transducer_stateless_d2v_v2/$1 \ + --max-duration 250 \ + --freeze-finetune-updates 2000 \ + --use-fp16 1 \ + --peak-enc-lr 0.001 \ + --peak-dec-lr 0.05 \ + --accum-grads 1 \ + --encoder-type d2v \ + --additional-block True \ + --encoder-dim 768 \ + --decoder-dim 768 \ + --joiner-dim 768 \ + --prune-range 20 \ + --context-size 2 \ + --ctc-loss-scale 0.2 + +""" + + +import random +import argparse +import copy +import logging +import warnings +from pathlib import Path +from shutil import copyfile +from typing import Any, Dict, Optional, Tuple, Union + +import k2 +import optim +import sentencepiece as spm +import torch +import torch.multiprocessing as mp +import torch.nn as nn +from asr_datamodule import LibriSpeechAsrDataModule +from decoder import Decoder +from joiner import Joiner +from lhotse.cut import Cut +from lhotse.dataset.sampling.base import CutSampler +from lhotse.utils import fix_random_seed +from model import Transducer +from optim import Eden, ScaledAdam +from torch import Tensor +from torch.cuda.amp import GradScaler +from torch.nn.parallel import DistributedDataParallel as DDP +from torch.utils.tensorboard import SummaryWriter +from zipformer import Zipformer +from data2vec_encoder import FairSeqData2VecEncoder + +from icefall import diagnostics +from icefall.checkpoint import remove_checkpoints +from icefall.checkpoint import update_averaged_model +from checkpoint import ( + save_checkpoint as save_checkpoint_impl, + save_checkpoint_with_global_batch_idx, + load_checkpoint +) +from icefall.dist import cleanup_dist, setup_dist +from icefall.env import get_env_info +from icefall.hooks import register_inf_check_hooks +from icefall.utils import ( + AttributeDict, + MetricsTracker, + encode_supervisions, + setup_logger, + str2bool, + save_args, +) + +import wandb + +#from icefall.checkpoint import save_checkpoint as save_checkpoint_impl +LRSchedulerType = Union[torch.optim.lr_scheduler._LRScheduler, optim.LRScheduler] + + +def set_batch_count(model: Union[nn.Module, DDP], batch_count: float) -> None: + if isinstance(model, DDP): + # get underlying nn.Module + model = model.module + for module in model.modules(): + if hasattr(module, "batch_count"): + module.batch_count = batch_count + model.encoder.num_updates = int(batch_count) + + +def add_adapter_arguments(parser: argparse.ArgumentParser): + parser.add_argument( + "--add-adapter", + type=str2bool, + default=False, + help="add adapter to rep model's encoder" + ) + + parser.add_argument( + "--adapter-lr", + type=float, + default=0.0001, + help="adapter learning rate" + ) + + parser.add_argument( + "--gender", + type=str, + default='male', + help="select gender" + ) + + +def add_rep_arguments(parser: argparse.ArgumentParser): + parser.add_argument( + "--wandb", + type=str2bool, + default=True, + help="Use wandb for MLOps", + ) + parser.add_argument( + "--hpo", + type=str2bool, + default=False, + help="Use small db for HPO", + ) + + parser.add_argument( + "--accum-grads", + type=int, + default=1, + help="accum-grad num.", + ) + + parser.add_argument( + "--multi-optim", + type=str2bool, + default=True, + help="use sperate optimizer (enc / dec)", + ) + + parser.add_argument( + "--peak-enc-lr", + type=float, + default=0.0001, + help="The initial learning rate. This value should not need to be changed.", + ) + + parser.add_argument( + "--peak-dec-lr", + type=float, + default=0.001, + help="The initial learning rate. This value should not need to be changed.", + ) + + parser.add_argument( + "--encoder-type", + type=str, + default='d2v', + help="Type of encoder (e.g. conformer, w2v, d2v...", + ) + + parser.add_argument( + "--encoder-dim", + type=int, + default=768, + help="encoder embedding dimension", + ) + + parser.add_argument( + "--freeze-finetune-updates", + type=int, + default=0 + ) + + parser.add_argument( + "--additional-block", + type=str2bool, + default=True, + ) + + parser.add_argument( + "--decode-interval", + type=int, + default=200, + help="decode interval", + ) + + +def add_model_arguments(parser: argparse.ArgumentParser): + parser.add_argument( + "--num-encoder-layers", + type=str, + default="2,4,3,2,4", + help="Number of zipformer encoder layers, comma separated.", + ) + + parser.add_argument( + "--feedforward-dims", + type=str, + default="1024,1024,2048,2048,1024", + help="Feedforward dimension of the zipformer encoder layers, comma separated.", + ) + + parser.add_argument( + "--nhead", + type=str, + default="8,8,8,8,8", + help="Number of attention heads in the zipformer encoder layers.", + ) + + parser.add_argument( + "--encoder-dims", + type=str, + default="384,384,384,384,384", + help="Embedding dimension in the 2 blocks of zipformer encoder layers, comma separated", + ) + + parser.add_argument( + "--attention-dims", + type=str, + default="192,192,192,192,192", + help="""Attention dimension in the 2 blocks of zipformer encoder layers, comma separated; + not the same as embedding dimension.""", + ) + + parser.add_argument( + "--encoder-unmasked-dims", + type=str, + default="256,256,256,256,256", + help="Unmasked dimensions in the encoders, relates to augmentation during training. " + "Must be <= each of encoder_dims. Empirically, less than 256 seems to make performance " + " worse.", + ) + + parser.add_argument( + "--zipformer-downsampling-factors", + type=str, + default="1,2,4,8,2", + help="Downsampling factor for each stack of encoder layers.", + ) + + parser.add_argument( + "--cnn-module-kernels", + type=str, + default="31,31,31,31,31", + help="Sizes of kernels in convolution modules", + ) + + parser.add_argument( + "--decoder-dim", + type=int, + default=768, + help="Embedding dimension in the decoder model.", + ) + + parser.add_argument( + "--joiner-dim", + type=int, + default=768, + help="""Dimension used in the joiner model. + Outputs from the encoder and decoder model are projected + to this dimension before adding. + """, + ) + + parser.add_argument( + "--prompt", + type=str2bool, + default=False, + ) + + +def get_parser(): + parser = argparse.ArgumentParser( + formatter_class=argparse.ArgumentDefaultsHelpFormatter + ) + + parser.add_argument( + "--world-size", + type=int, + default=1, + help="Number of GPUs for DDP training.", + ) + + parser.add_argument( + "--master-port", + type=int, + default=12354, + help="Master port to use for DDP training.", + ) + + parser.add_argument( + "--tensorboard", + type=str2bool, + default=True, + help="Should various information be logged in tensorboard.", + ) + + parser.add_argument( + "--num-epochs", + type=int, + default=30, + help="Number of epochs to train.", + ) + + parser.add_argument( + "--num-updates", + type=int, + default=5000, + help="Number of epochs to train.", + ) + + parser.add_argument( + "--start-epoch", + type=int, + default=1, + help="""Resume training from this epoch. It should be positive. + If larger than 1, it will load checkpoint from + exp-dir/epoch-{start_epoch-1}.pt + """, + ) + + parser.add_argument( + "--start-batch", + type=int, + default=0, + help="""If positive, --start-epoch is ignored and + it loads the checkpoint from exp-dir/checkpoint-{start_batch}.pt + """, + ) + + parser.add_argument( + "--exp-dir", + type=str, + default="pruned_transducer_stateless7_ctc/exp", + help="""The experiment dir. + It specifies the directory where all training related + files, e.g., checkpoints, log, etc, are saved + """, + ) + + parser.add_argument( + "--bpe-model", + type=str, + default="data/lang_bpe_500/bpe.model", + help="Path to the BPE model", + ) + + parser.add_argument( + "--base-lr", type=float, default=0.05, help="The base learning rate." + ) + + parser.add_argument( + "--lr-batches", + type=float, + default=5000, + help="""Number of steps that affects how rapidly the learning rate + decreases. We suggest not to change this.""", + ) + + parser.add_argument( + "--lr-epochs", + type=float, + default=3.5, + help="""Number of epochs that affects how rapidly the learning rate decreases. + """, + ) + + parser.add_argument( + "--context-size", + type=int, + default=2, + help="The context size in the decoder. 1 means bigram; 2 means tri-gram", + ) + + parser.add_argument( + "--prune-range", + type=int, + default=5, + help="The prune range for rnnt loss, it means how many symbols(context)" + "we are using to compute the loss", + ) + + parser.add_argument( + "--lm-scale", + type=float, + default=0.25, + help="The scale to smooth the loss with lm " + "(output of prediction network) part.", + ) + + parser.add_argument( + "--am-scale", + type=float, + default=0.0, + help="The scale to smooth the loss with am (output of encoder network) part.", + ) + + parser.add_argument( + "--simple-loss-scale", + type=float, + default=0.5, + help="To get pruning ranges, we will calculate a simple version" + "loss(joiner is just addition), this simple loss also uses for" + "training (as a regularization item). We will scale the simple loss" + "with this parameter before adding to the final loss.", + ) + + parser.add_argument( + "--ctc-loss-scale", + type=float, + default=0.2, + help="Scale for CTC loss.", + ) + + parser.add_argument( + "--seed", + type=int, + default=42, + help="The seed for random generators intended for reproducibility", + ) + + parser.add_argument( + "--print-diagnostics", + type=str2bool, + default=False, + help="Accumulate stats on activations, print them and exit.", + ) + + parser.add_argument( + "--inf-check", + type=str2bool, + default=False, + help="Add hooks to check for infinite module outputs and gradients.", + ) + + parser.add_argument( + "--save-every-n", + type=int, + default=200, + help="""Save checkpoint after processing this number of batches" + periodically. We save checkpoint to exp-dir/ whenever + params.batch_idx_train % save_every_n == 0. The checkpoint filename + has the form: f'exp-dir/checkpoint-{params.batch_idx_train}.pt' + Note: It also saves checkpoint to `exp-dir/epoch-xxx.pt` at the + end of each epoch where `xxx` is the epoch number counting from 0. + """, + ) + + parser.add_argument( + "--keep-last-k", + type=int, + default=30, + help="""Only keep this number of checkpoints on disk. + For instance, if it is 3, there are only 3 checkpoints + in the exp-dir with filenames `checkpoint-xxx.pt`. + It does not affect checkpoints with name `epoch-xxx.pt`. + """, + ) + + parser.add_argument( + "--average-period", + type=int, + default=10, + help="""Update the averaged model, namely `model_avg`, after processing + this number of batches. `model_avg` is a separate version of model, + in which each floating-point parameter is the average of all the + parameters from the start of training. Each time we take the average, + we do: `model_avg = model * (average_period / batch_idx_train) + + model_avg * ((batch_idx_train - average_period) / batch_idx_train)`. + """, + ) + + parser.add_argument( + "--use-fp16", + type=str2bool, + default=True, + help="Whether to use half precision training.", + ) + + add_model_arguments(parser) + add_rep_arguments(parser) + add_adapter_arguments(parser) + + return parser + + +def get_params() -> AttributeDict: + """Return a dict containing training parameters. + + All training related parameters that are not passed from the commandline + are saved in the variable `params`. + + Commandline options are merged into `params` after they are parsed, so + you can also access them via `params`. + + Explanation of options saved in `params`: + + - best_train_loss: Best training loss so far. It is used to select + the model that has the lowest training loss. It is + updated during the training. + + - best_valid_loss: Best validation loss so far. It is used to select + the model that has the lowest validation loss. It is + updated during the training. + + - best_train_epoch: It is the epoch that has the best training loss. + + - best_valid_epoch: It is the epoch that has the best validation loss. + + - batch_idx_train: Used to writing statistics to tensorboard. It + contains number of batches trained so far across + epochs. + + - log_interval: Print training loss if batch_idx % log_interval` is 0 + + - reset_interval: Reset statistics if batch_idx % reset_interval is 0 + + - valid_interval: Run validation if batch_idx % valid_interval is 0 + + - feature_dim: The model input dim. It has to match the one used + in computing features. + + - subsampling_factor: The subsampling factor for the model. + + - encoder_dim: Hidden dim for multi-head attention model. + + - num_decoder_layers: Number of decoder layer of transformer decoder. + + - warm_step: The warmup period that dictates the decay of the + scale on "simple" (un-pruned) loss. + """ + params = AttributeDict( + { + "best_train_loss": float("inf"), + "best_valid_loss": float("inf"), + "best_train_epoch": -1, + "best_valid_epoch": -1, + "batch_idx_train": 0, + "log_interval": 5, + "reset_interval": 200, + "valid_interval": 3000, # For the 100h subset, use 800 + # parameters for zipformer + "feature_dim": 80, + "subsampling_factor": 320, # not passed in, this is fixed. + # parameters for ctc loss + "beam_size": 10, + "use_double_scores": True, + "warm_step": 0, + #"warm_step": 4000, + #"warm_step": 3000, + "env_info": get_env_info(), + } + ) + + return params + + +def get_encoder_model(params: AttributeDict) -> nn.Module: + # TODO: We can add an option to switch between Zipformer and Transformer + def to_int_tuple(s: str): + return tuple(map(int, s.split(","))) + + if params.encoder_type == 'd2v': + encoder = FairSeqData2VecEncoder( + input_size=params.encoder_dim, + w2v_url='None', + output_size=params.encoder_dim, + freeze_finetune_updates=params.freeze_finetune_updates, + additional_block=params.additional_block, + ) + else: + encoder = Zipformer( + num_features=params.feature_dim, + output_downsampling_factor=2, + zipformer_downsampling_factors=to_int_tuple( + params.zipformer_downsampling_factors + ), + encoder_dims=to_int_tuple(params.encoder_dims), + attention_dim=to_int_tuple(params.attention_dims), + encoder_unmasked_dims=to_int_tuple(params.encoder_unmasked_dims), + nhead=to_int_tuple(params.nhead), + feedforward_dim=to_int_tuple(params.feedforward_dims), + cnn_module_kernels=to_int_tuple(params.cnn_module_kernels), + num_encoder_layers=to_int_tuple(params.num_encoder_layers), + ) + + return encoder + + +def get_decoder_model(params: AttributeDict) -> nn.Module: + decoder = Decoder( + vocab_size=params.vocab_size, + decoder_dim=params.decoder_dim, + blank_id=params.blank_id, + context_size=params.context_size, + ) + return decoder + + +def get_joiner_model(params: AttributeDict) -> nn.Module: + joiner = Joiner( + encoder_dim=params.encoder_dim if params.encoder_type == 'd2v' else int(params.encoder_dims.split(",")[-1]), + decoder_dim=params.decoder_dim, + joiner_dim=params.joiner_dim, + vocab_size=params.vocab_size, + ) + return joiner + + +def get_transducer_model(params: AttributeDict) -> nn.Module: + encoder = get_encoder_model(params) + decoder = get_decoder_model(params) + joiner = get_joiner_model(params) + + model = Transducer( + encoder=encoder, + decoder=decoder, + joiner=joiner, + encoder_dim=params.encoder_dim if params.encoder_type == 'd2v' else int(params.encoder_dims.split(",")[-1]), + decoder_dim=params.decoder_dim, + joiner_dim=params.joiner_dim, + vocab_size=params.vocab_size, + prompt=params.prompt, + sid=params.spk_id, + ) + return model + + +def load_checkpoint_if_available( + params: AttributeDict, + model: nn.Module, + model_avg: nn.Module = None, + optimizer: Optional[torch.optim.Optimizer] = None, + scheduler: Optional[LRSchedulerType] = None, +) -> Optional[Dict[str, Any]]: + """Load checkpoint from file. + + If params.start_batch is positive, it will load the checkpoint from + `params.exp_dir/checkpoint-{params.start_batch}.pt`. Otherwise, if + params.start_epoch is larger than 1, it will load the checkpoint from + `params.start_epoch - 1`. + + Apart from loading state dict for `model` and `optimizer` it also updates + `best_train_epoch`, `best_train_loss`, `best_valid_epoch`, + and `best_valid_loss` in `params`. + + Args: + params: + The return value of :func:`get_params`. + model: + The training model. + model_avg: + The stored model averaged from the start of training. + optimizer: + The optimizer that we are using. + scheduler: + The scheduler that we are using. + Returns: + Return a dict containing previously saved training info. + """ + if params.start_batch > 0: + filename = params.exp_dir / f"checkpoint-{params.start_batch}.pt" + elif params.start_epoch > 1: + filename = params.exp_dir / f"epoch-{params.start_epoch-1}.pt" + elif params.add_adapter: + filename = params.exp_dir / f"../d2v-base-T.pt" + else: + return None + + assert filename.is_file(), f"{filename} does not exist!" + + saved_params = load_checkpoint( + filename, + model=model, + model_avg=model_avg, + optimizer=optimizer, + scheduler=scheduler, + strict=True if not params.add_adapter else False, + ) + + keys = [ + "best_train_epoch", + "best_valid_epoch", + "batch_idx_train", + "best_train_loss", + "best_valid_loss", + ] + for k in keys: + params[k] = saved_params[k] + + if params.start_batch > 0: + if "cur_epoch" in saved_params: + params["start_epoch"] = saved_params["cur_epoch"] + + if "cur_batch_idx" in saved_params: + params["cur_batch_idx"] = saved_params["cur_batch_idx"] + + params.batch_idx_train = 0 + + return saved_params + + +def save_checkpoint( + params: AttributeDict, + model: Union[nn.Module, DDP], + model_avg: Optional[nn.Module] = None, + optimizer: Optional[torch.optim.Optimizer] = None, + scheduler: Optional[LRSchedulerType] = None, + sampler: Optional[CutSampler] = None, + scaler: Optional[GradScaler] = None, + rank: int = 0, +) -> None: + """Save model, optimizer, scheduler and training stats to file. + + Args: + params: + It is returned by :func:`get_params`. + model: + The training model. + model_avg: + The stored model averaged from the start of training. + optimizer: + The optimizer used in the training. + sampler: + The sampler for the training dataset. + scaler: + The scaler used for mix precision training. + """ + if rank != 0: + return + filename = params.exp_dir / f"epoch-{params.cur_epoch}.pt" + save_checkpoint_impl( + filename=filename, + model=model, + model_avg=model_avg, + params=params, + optimizer=optimizer, + scheduler=scheduler, + sampler=sampler, + scaler=scaler, + rank=rank, + ) + + if params.best_train_epoch == params.cur_epoch: + best_train_filename = params.exp_dir / "best-train-loss.pt" + copyfile(src=filename, dst=best_train_filename) + + if params.best_valid_epoch == params.cur_epoch: + best_valid_filename = params.exp_dir / "best-valid-loss.pt" + copyfile(src=filename, dst=best_valid_filename) + + +def compute_loss( + params: AttributeDict, + model: Union[nn.Module, DDP], + sp: spm.SentencePieceProcessor, + batch: dict, + is_training: bool, + decode: bool = False, +) -> Tuple[Tensor, MetricsTracker]: + """ + Compute transducer loss given the model and its inputs. + + Args: + params: + Parameters for training. See :func:`get_params`. + model: + The model for training. It is an instance of Zipformer in our case. + batch: + A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()` + for the content in it. + is_training: + True for training. False for validation. When it is True, this + function enables autograd during computation; when it is False, it + disables autograd. + warmup: a floating point value which increases throughout training; + values >= 1.0 are fully warmed up and have all modules present. + """ + device = model.device if isinstance(model, DDP) else next(model.parameters()).device + feature = batch["inputs"] + # at entry, feature is (N, T, C) + assert feature.ndim == 2 or feature.ndim == 3 + feature = feature.to(device) + + supervisions = batch["supervisions"] + + if feature.ndim == 2: + feature_lens = [] + for supervision in supervisions['cut']: + try: feature_lens.append(supervision.tracks[0].cut.recording.num_samples) + except: feature_lens.append(supervision.recording.num_samples) + feature_lens = torch.tensor(feature_lens) + + elif feature.ndim == 3: + feature_lens = supervisions["num_frames"].to(device) + + batch_idx_train = params.batch_idx_train + warm_step = params.warm_step + + texts = batch["supervisions"]["text"] + + token_ids = sp.encode(texts, out_type=int) + y = k2.RaggedTensor(token_ids).to(device) + + with torch.set_grad_enabled(is_training): + simple_loss, pruned_loss, ctc_output = model( + x=feature, + x_lens=feature_lens, + y=y, + prune_range=params.prune_range, + am_scale=params.am_scale, + lm_scale=params.lm_scale, + ) + + s = params.simple_loss_scale + # take down the scale on the simple loss from 1.0 at the start + # to params.simple_loss scale by warm_step. + simple_loss_scale = ( + s + if batch_idx_train >= warm_step + else 1.0 - (batch_idx_train / warm_step) * (1.0 - s) + ) + pruned_loss_scale = ( + 1.0 + if batch_idx_train >= warm_step + else 0.1 + 0.9 * (batch_idx_train / warm_step) + ) + + loss = simple_loss_scale * simple_loss + pruned_loss_scale * pruned_loss + + info = MetricsTracker() + + if params.ctc_loss_scale > 0: + # Compute ctc loss + + # NOTE: We need `encode_supervisions` to sort sequences with + # different duration in decreasing order, required by + # `k2.intersect_dense` called in `k2.ctc_loss` + with warnings.catch_warnings(): + warnings.simplefilter("ignore") + supervision_segments, token_ids = encode_supervisions( + supervisions, + subsampling_factor=params.subsampling_factor, + token_ids=token_ids, + ) + + # Works with a BPE model + decoding_graph = k2.ctc_graph(token_ids, modified=False, device=device) + dense_fsa_vec = k2.DenseFsaVec( + ctc_output, + supervision_segments, + allow_truncate=params.subsampling_factor - 1, + ) + + ctc_loss = k2.ctc_loss( + decoding_graph=decoding_graph, + dense_fsa_vec=dense_fsa_vec, + output_beam=params.beam_size, + reduction="sum", + use_double_scores=params.use_double_scores, + ) + assert ctc_loss.requires_grad == is_training + loss += params.ctc_loss_scale * ctc_loss + + info["ctc_loss"] = ctc_loss.detach().cpu().item() + + assert loss.requires_grad == is_training + + if decode: + model.eval() + with torch.no_grad(): + try: + hypos = model.module.decode( + x=feature, + x_lens=feature_lens, + y=y, + sp=sp + ) + except: + hypos = model.decode( + x=feature, + x_lens=feature_lens, + y=y, + sp=sp + ) + + logging.info(f'ref: {batch["supervisions"]["text"][0]}') + logging.info(f'hyp: {" ".join(hypos[0])}') + model.train() + + with warnings.catch_warnings(): + warnings.simplefilter("ignore") + info["frames"] = (feature_lens // params.subsampling_factor).sum().item() + + # Note: We use reduction=sum while computing the loss. + info["utterances"] = feature.size(0) + info["loss"] = loss.detach().cpu().item() + info["simple_loss"] = simple_loss.detach().cpu().item() + info["pruned_loss"] = pruned_loss.detach().cpu().item() + + return loss, info + + +def compute_validation_loss( + params: AttributeDict, + model: Union[nn.Module, DDP], + sp: spm.SentencePieceProcessor, + valid_dl: torch.utils.data.DataLoader, + world_size: int = 1, +) -> MetricsTracker: + """Run the validation process.""" + model.eval() + + tot_loss = MetricsTracker() + + for batch_idx, batch in enumerate(valid_dl): + loss, loss_info = compute_loss( + params=params, + model=model, + sp=sp, + batch=batch, + is_training=False, + ) + assert loss.requires_grad is False + tot_loss = tot_loss + loss_info + + if world_size > 1: + tot_loss.reduce(loss.device) + + loss_value = tot_loss["loss"] / tot_loss["utterances"] + if loss_value < params.best_valid_loss: + params.best_valid_epoch = params.cur_epoch + params.best_valid_loss = loss_value + + return tot_loss + + +def train_one_epoch( + params: AttributeDict, + model: Union[nn.Module, DDP], + optimizer: torch.optim.Optimizer or [torch.optim.Optimizer, torch.optim.Optimizer], + scheduler: LRSchedulerType or [LRSchedulerType, LRSchedulerType], + sp: spm.SentencePieceProcessor, + train_dl: torch.utils.data.DataLoader, + valid_dl: torch.utils.data.DataLoader, + scaler: GradScaler, + model_avg: Optional[nn.Module] = None, + tb_writer: Optional[SummaryWriter] = None, + world_size: int = 1, + rank: int = 0, + wb = None, +) -> None: + """Train the model for one epoch. + + The training loss from the mean of all frames is saved in + `params.train_loss`. It runs the validation process every + `params.valid_interval` batches. + + Args: + params: + It is returned by :func:`get_params`. + model: + The model for training. + optimizer: + The optimizer we are using. + scheduler: + The learning rate scheduler, we call step() every step. + train_dl: + Dataloader for the training dataset. + valid_dl: + Dataloader for the validation dataset. + scaler: + The scaler used for mix precision training. + model_avg: + The stored model averaged from the start of training. + tb_writer: + Writer to write log messages to tensorboard. + world_size: + Number of nodes in DDP training. If it is 1, DDP is disabled. + rank: + The rank of the node in DDP training. If no DDP is used, it should + be set to 0. + """ + model.train() + + tot_loss = MetricsTracker() + + cur_batch_idx = params.get("cur_batch_idx", 0) + + if params.multi_optim: + optimizer_enc, optimizer_dec = optimizer[0], optimizer[1] + scheduler_enc, scheduler_dec = scheduler[0], scheduler[1] + + for batch_idx, batch in enumerate(train_dl): + if params.batch_idx_train > params.num_updates: + break + if batch_idx < cur_batch_idx: + continue + cur_batch_idx = batch_idx + + if batch_idx % params.accum_grads == 0: params.batch_idx_train += 1 + batch_size = len(batch["supervisions"]["text"]) + + try: + with torch.cuda.amp.autocast(enabled=params.use_fp16): + loss, loss_info = compute_loss( + params=params, + model=model, + sp=sp, + batch=batch, + is_training=True, + decode = True if batch_idx % params.decode_interval == 0 else False, + ) + + try: loss_info.reduce(loss.device) + except: pass + + numel = params.world_size / (params.accum_grads * loss_info["utterances"]) + loss *= numel ## normalize loss over utts(batch size) + + # summary stats + tot_loss = (tot_loss * (1 - 1 / params.reset_interval)) + loss_info + + # NOTE: We use reduction==sum and loss is computed over utterances + # in the batch and there is no normalization to it so far. + scaler.scale(loss).backward() + + if params.multi_optim and (batch_idx+1) % params.accum_grads == 0: + set_batch_count(model, params.batch_idx_train) + scheduler_enc.step_batch(params.batch_idx_train) + scheduler_dec.step_batch(params.batch_idx_train) + scaler.step(optimizer_enc) + scaler.step(optimizer_dec) + scaler.update() + optimizer_enc.zero_grad() + optimizer_dec.zero_grad() + elif not params.multi_optim and (batch_idx+1) % params.accum_grads == 0: + set_batch_count(model, params.batch_idx_train) + scheduler.step_batch(params.batch_idx_train) + scaler.step(optimizer) + scaler.update() + optimizer.zero_grad() + + except: # noqa + display_and_save_batch(batch, params=params, sp=sp) + raise + + if params.print_diagnostics and batch_idx == 5: + return + + ''' + if ( + rank == 0 + and params.batch_idx_train > 0 + and params.batch_idx_train % params.average_period == 0 + ): + update_averaged_model( + params=params, + model_cur=model, + model_avg=model_avg, + ) + ''' + + if ( + params.batch_idx_train > 0 + and params.batch_idx_train % params.save_every_n == 0 + ): + params.cur_batch_idx = batch_idx + save_checkpoint_with_global_batch_idx( + out_dir=params.exp_dir, + global_batch_idx=params.batch_idx_train, + model=model, + model_avg=model_avg, + params=params, + optimizer=optimizer, + scheduler=scheduler, + sampler=train_dl.sampler, + scaler=scaler, + rank=rank, + ) + del params.cur_batch_idx + ''' + remove_checkpoints( + out_dir=params.exp_dir, + topk=params.keep_last_k, + rank=rank, + ) + ''' + + if batch_idx % 100 == 0 and params.use_fp16: + # If the grad scale was less than 1, try increasing it. The _growth_interval + # of the grad scaler is configurable, but we can't configure it to have different + # behavior depending on the current grad scale. + cur_grad_scale = scaler._scale.item() + ''' + if cur_grad_scale < 1.0 or (cur_grad_scale < 8.0 and batch_idx % 400 == 0): + scaler.update(cur_grad_scale * 2.0) + ''' + if cur_grad_scale < 0.01: + logging.warning(f"Grad scale is small: {cur_grad_scale}") + if cur_grad_scale < 1.0e-05: + wb.log({"valid/loss": 10000}) + raise RuntimeError( + f"grad_scale is too small, exiting: {cur_grad_scale}" + ) + + #if params.batch_idx_train > 4000 and loss > 300 and params.wandb: + # wb.log({"valid/loss": 10000}) + # raise RuntimeError( + # f"divergence... exiting: loss={loss}" + # ) + + if batch_idx % (params.log_interval*params.accum_grads) == 0: + #for n, p in model.named_parameters(): + # if 'adapter' in n: + # print(p) + if params.multi_optim: + cur_enc_lr = scheduler_enc.get_last_lr()[0] + cur_dec_lr = scheduler_dec.get_last_lr()[0] + cur_grad_scale = scaler._scale.item() if params.use_fp16 else 1.0 + + logging.info( + f"Epoch {params.cur_epoch}, " + f"batch {batch_idx}, loss[{loss_info}], " + f"tot_loss[{tot_loss}], batch size: {batch_size}, " + f"enc_lr: {cur_enc_lr:.2e}, " + f"dec_lr: {cur_dec_lr:.2e}, " + + (f"grad_scale: {scaler._scale.item()}" if params.use_fp16 else "") + ) + + else: + cur_lr = scheduler.get_last_lr()[0] + cur_grad_scale = scaler._scale.item() if params.use_fp16 else 1.0 + + logging.info( + f"Epoch {params.cur_epoch}, " + f"batch {batch_idx}, loss[{loss_info}], " + f"tot_loss[{tot_loss}], batch size: {batch_size}, " + f"lr: {cur_lr:.2e}, " + + (f"grad_scale: {scaler._scale.item()}" if params.use_fp16 else "") + ) + + if tb_writer is not None: + if params.multi_optim: + tb_writer.add_scalar( + "train/enc_learning_rate", cur_enc_lr, params.batch_idx_train + ) + tb_writer.add_scalar( + "train/dec_learning_rate", cur_dec_lr, params.batch_idx_train + ) + + else: + tb_writer.add_scalar( + "train/learning_rate", cur_lr, params.batch_idx_train + ) + + loss_info.write_summary( + tb_writer, "train/current_", params.batch_idx_train + ) + tot_loss.write_summary(tb_writer, "train/tot_", params.batch_idx_train) + if params.use_fp16: + tb_writer.add_scalar( + "train/grad_scale", + cur_grad_scale, + params.batch_idx_train, + ) + + if wb is not None and rank == 0: + wb.log({"train/loss": loss_info["loss"]*numel}) + wb.log({"train/simple_loss": loss_info["simple_loss"]*numel}) + wb.log({"train/pruned_loss": loss_info["pruned_loss"]*numel}) + wb.log({"train/ctc_loss": loss_info["ctc_loss"]*numel}) + + ''' + logging.info("Computing validation loss") + valid_info = compute_validation_loss( + params=params, + model=model, + sp=sp, + valid_dl=valid_dl, + world_size=world_size, + ) + model.train() + logging.info(f"Epoch {params.cur_epoch}, validation: {valid_info}") + logging.info( + f"Maximum memory allocated so far is {torch.cuda.max_memory_allocated()//1000000}MB" + ) + if tb_writer is not None: + valid_info.write_summary( + tb_writer, "train/valid_", params.batch_idx_train + ) + + if wb is not None and rank == 0: + numel = 1 / (params.accum_grads * valid_info["utterances"]) + #wb.log({"valid/loss": valid_info["loss"]*numel}) + wb.log({"valid/loss": numel*(valid_info["simple_loss"] + +valid_info["pruned_loss"] + +valid_info["ctc_loss"] + )}) + wb.log({"valid/simple_loss": valid_info["simple_loss"]*numel}) + wb.log({"valid/pruned_loss": valid_info["pruned_loss"]*numel}) + wb.log({"valid/ctc_loss": valid_info["ctc_loss"]*numel}) + ''' + loss_value = tot_loss["loss"] / tot_loss["utterances"] + params.train_loss = loss_value + if params.train_loss < params.best_train_loss: + params.best_train_epoch = params.cur_epoch + params.best_train_loss = params.train_loss + + +def run(rank, world_size, args, wb=None): + """ + Args: + rank: + It is a value between 0 and `world_size-1`, which is + passed automatically by `mp.spawn()` in :func:`main`. + The node with rank 0 is responsible for saving checkpoint. + world_size: + Number of GPUs for DDP training. + args: + The return value of get_parser().parse_args() + """ + params = get_params() + params.update(vars(args)) + #params.warm_step *= params.accum_grads + + fix_random_seed(params.seed) + if world_size > 1: + setup_dist(rank, world_size, params.master_port) + + setup_logger(f"{params.exp_dir}/log/log-train") + logging.info("Training started") + + if args.tensorboard and rank == 0: + tb_writer = SummaryWriter(log_dir=f"{params.exp_dir}/tensorboard") + else: + tb_writer = None + + device = torch.device("cpu") + if torch.cuda.is_available(): + device = torch.device("cuda", rank) + logging.info(f"Device: {device}") + + sp = spm.SentencePieceProcessor() + sp.load(params.bpe_model) + + # is defined in local/train_bpe_model.py + params.blank_id = sp.piece_to_id("") + params.vocab_size = sp.get_piece_size() + + logging.info(params) + + logging.info("About to create model") + model = get_transducer_model(params) + logging.info(model) + + num_param = sum([p.numel() for p in model.parameters()]) + logging.info(f"Number of model parameters: {num_param}") + + assert params.save_every_n >= params.average_period + model_avg: Optional[nn.Module] = None + if rank == 0: + # model_avg is only used with rank 0 + model_avg = copy.deepcopy(model).to(torch.float64) + + assert params.start_epoch > 0, params.start_epoch + checkpoints = load_checkpoint_if_available( + params=params, model=model, model_avg=model_avg + ) + + model.to(device) + if world_size > 1: + logging.info("Using DDP") + model = DDP(model, device_ids=[rank], find_unused_parameters=True) + + if params.multi_optim: + logging.info("Using seperate optimizers over encoder, decoder ...") + + enc_param = [] + enc_names = [] + + dec_names = [] + dec_param = [] + + for n, p in model.named_parameters(): + name = n.split('.')[1] + if name == 'encoder' and 'feature_extractor' not in n: + enc_names.append(n) + enc_param.append(p) + elif 'ctc_output' in n: + enc_names.append(n) + enc_param.append(p) + elif 'feature_extractor' not in n: + dec_names.append(n) + dec_param.append(p) + + optimizer_enc = ScaledAdam( + enc_param, + lr=params.peak_enc_lr, + clipping_scale=None, + parameters_names=[enc_names], + ) + optimizer_dec = ScaledAdam( + dec_param, + lr=params.peak_dec_lr, + clipping_scale=5.0, + parameters_names=[dec_names], + ) + + scheduler_enc = Eden(optimizer_enc, params.lr_batches, params.lr_epochs) + scheduler_dec = Eden(optimizer_dec, params.lr_batches, params.lr_epochs) + optimizer = [optimizer_enc, optimizer_dec] + scheduler = [scheduler_enc, scheduler_dec] + + else: + parameters_names = [] + parameters_names.append( + [name_param_pair[0] for name_param_pair in model.named_parameters()] + ) + + logging.info(f"len name = {len(parameters_names)}") + logging.info(f"len param = {len(list(model.parameters()))}") + + optimizer = ScaledAdam( + model.parameters(), + lr=params.base_lr, + clipping_scale=2.0, + parameters_names=parameters_names, + ) + + scheduler = Eden(optimizer, params.lr_batches, params.lr_epochs) + + if checkpoints and ("optimizer" in checkpoints or "optimizer_enc" in checkpoints): + if params.multi_optim: + logging.info("Loading optimizer state dict") + optimizer_enc.load_state_dict(checkpoints["optimizer_enc"]) + optimizer_dec.load_state_dict(checkpoints["optimizer_dec"]) + + else: + logging.info("Loading optimizer state dict") + optimizer.load_state_dict(checkpoints["optimizer"]) + + if checkpoints: + if ( + params.multi_optim + and "scheduler_enc" in checkpoints + and checkpoints["scheduler_enc"] is not None + ): + logging.info("Loading enc/dec scheduler state dict") + scheduler_enc.load_state_dict(checkpoints["scheduler_enc"]) + scheduler_dec.load_state_dict(checkpoints["scheduler_dec"]) + else: + logging.info("Loading scheduler state dict") + scheduler.load_state_dict(checkpoints["scheduler"]) + + if params.print_diagnostics: + opts = diagnostics.TensorDiagnosticOptions( + 2**22 + ) # allow 4 megabytes per sub-module + diagnostic = diagnostics.attach_diagnostics(model, opts) + + if params.inf_check: + register_inf_check_hooks(model) + + librispeech = LibriSpeechAsrDataModule(args) + + train_cuts = librispeech.train_clean_100_cuts() + if params.full_libri: + train_cuts += librispeech.train_clean_360_cuts() + train_cuts += librispeech.train_other_500_cuts() + + def remove_short_and_long_utt(c: Cut): + # Keep only utterances with duration between 1 second and 20 seconds + # + # Caution: There is a reason to select 20.0 here. Please see + # ../local/display_manifest_statistics.py + # + # You should use ../local/display_manifest_statistics.py to get + # an utterance duration distribution for your dataset to select + # the threshold + return 1.0 <= c.duration <= 20.0 + + train_cuts = train_cuts.filter(remove_short_and_long_utt) + + if params.start_batch > 0 and checkpoints and "sampler" in checkpoints: + # We only load the sampler's state dict when it loads a checkpoint + # saved in the middle of an epoch + sampler_state_dict = checkpoints["sampler"] + else: + sampler_state_dict = None + + train_dl = librispeech.train_dataloaders( + train_cuts, sampler_state_dict=sampler_state_dict + ) + + valid_cuts = librispeech.dev_clean_cuts() + valid_cuts += librispeech.dev_other_cuts() + valid_dl = librispeech.valid_dataloaders(valid_cuts) + + ''' + if not params.print_diagnostics: + scan_pessimistic_batches_for_oom( + model=model, + train_dl=train_dl, + optimizer=optimizer, + sp=sp, + params=params, + ) + ''' + + scaler = GradScaler(enabled=params.use_fp16, init_scale=1.0) + if checkpoints and "grad_scaler" in checkpoints: + logging.info("Loading grad scaler state dict") + scaler.load_state_dict(checkpoints["grad_scaler"]) + + for epoch in range(params.start_epoch, params.num_epochs + 1): + if params.multi_optim: + scheduler_enc.step_epoch(epoch - 1) + scheduler_dec.step_epoch(epoch - 1) + else: + scheduler.step_epoch(epoch - 1) + fix_random_seed(params.seed + epoch - 1) + train_dl.sampler.set_epoch(epoch - 1) + + if tb_writer is not None: + tb_writer.add_scalar("train/epoch", epoch, params.batch_idx_train) + + params.cur_epoch = epoch + + train_one_epoch( + params=params, + model=model, + model_avg=model_avg, + optimizer=optimizer, + scheduler=scheduler, + sp=sp, + train_dl=train_dl, + valid_dl=valid_dl, + scaler=scaler, + tb_writer=tb_writer, + world_size=world_size, + rank=rank, + wb=wb, + ) + + if params.print_diagnostics: + diagnostic.print_diagnostics() + break + + ''' + if epoch % 50 == 0: + save_checkpoint( + params=params, + model=model, + model_avg=model_avg, + optimizer=optimizer, + scheduler=scheduler, + sampler=train_dl.sampler, + scaler=scaler, + rank=rank, + ) + ''' + + logging.info("Done!") + + if world_size > 1: + torch.distributed.barrier() + cleanup_dist() + + +def run_adapter(rank, world_size, args, wb=None): + """ + Args: + rank: + It is a value between 0 and `world_size-1`, which is + passed automatically by `mp.spawn()` in :func:`main`. + The node with rank 0 is responsible for saving checkpoint. + world_size: + Number of GPUs for DDP training. + args: + The return value of get_parser().parse_args() + """ + params = get_params() + params.update(vars(args)) + + fix_random_seed(params.seed) + if world_size > 1: + setup_dist(rank, world_size, params.master_port) + + setup_logger(f"{params.exp_dir}/log/log-train") + logging.info("Training started") + + if args.tensorboard and rank == 0: + tb_writer = SummaryWriter(log_dir=f"{params.exp_dir}/tensorboard") + else: + tb_writer = None + + device = torch.device("cpu") + if torch.cuda.is_available(): + device = torch.device("cuda", rank) + logging.info(f"Device: {device}") + + sp = spm.SentencePieceProcessor() + sp.load(params.bpe_model) + + # is defined in local/train_bpe_model.py + params.blank_id = sp.piece_to_id("") + params.vocab_size = sp.get_piece_size() + + logging.info(params) + + logging.info("About to create model") + model = get_transducer_model(params) + + num_param = sum([p.numel() if p.requires_grad else 0 for p in model.parameters()]) + logging.info(f"Number of model parameters: {num_param}") + + assert params.save_every_n >= params.average_period + model_avg: Optional[nn.Module] = None + if rank == 0: + # model_avg is only used with rank 0 + #model_avg = copy.deepcopy(model).to(torch.float64) + model_avg = None + + assert params.start_epoch > 0, params.start_epoch + checkpoints = load_checkpoint_if_available( + params=params, model=model, model_avg=model_avg + ) + + model.to(device) + if world_size > 1: + logging.info("Using DDP") + model = DDP(model, device_ids=[rank], find_unused_parameters=True) + + ''' + adapter_names = [] + adapter_param = [] + for n, p in model.named_parameters(): + if 'adapters' in n:# or 'joiner' in n or 'simple' in n or 'ctc' in n: + adapter_names.append(n) + adapter_param.append(p) + elif 'joiner' in n or 'simple' in n or 'ctc' in n: + p.requires_grad = True + else: + p.requires_grad = False + + for n, p in model.named_parameters(): + p.requires_grad = False + + optimizer_adapter = ScaledAdam( + adapter_param, + lr=params.adapter_lr, + clipping_scale=5.0, + parameters_names=[adapter_names], + ) + ''' + + #prompt = torch.randn((100, 512), requires_grad=True) + optimizer_adapter = ScaledAdam( + [model.prompt], + lr=params.adapter_lr, + clipping_scale=5.0, + parameters_names=['P'], + ) + + scheduler_adapter = Eden(optimizer_adapter, 10000, 7) #params.lr_batche, params.lr_epochs) + + optimizer, scheduler = optimizer_adapter, scheduler_adapter + + librispeech = LibriSpeechAsrDataModule(args) + + model.prompt = model.prompt.to(device) + + ''' + if params.hpo: + train_cuts = librispeech.train_clean_10_cuts(option=params.gender) + else: + train_cuts = librispeech.train_clean_100_cuts(option=params.gender) + if params.full_libri: + train_cuts += librispeech.train_clean_360_cuts(option=params.gender) + train_cuts += librispeech.train_other_500_cuts(option=params.gender) + ''' + + #train_cuts = librispeech.train_clean_10_cuts(option='male') + #train_cuts = librispeech.test_clean_user(option='big') + train_cuts = librispeech.vox_cuts(option=params.spk_id) + + def remove_short_and_long_utt(c: Cut): + return 1.0 <= c.duration <= 20.0 + + train_cuts = train_cuts.filter(remove_short_and_long_utt) + + sampler_state_dict = None + + train_dl = librispeech.train_dataloaders( + train_cuts, sampler_state_dict=sampler_state_dict + ) + #train_dl = librispeech.test_dataloaders( + # train_cuts + #) + + ''' + print('\n'*5) + print('-'*30) + for batch in train_dl: + print(batch) + print('-'*30) + print('\n'*5) + exit() + ''' + + valid_cuts = librispeech.dev_clean_cuts(option=params.gender) + valid_cuts += librispeech.dev_other_cuts(option=params.gender) + valid_dl = librispeech.valid_dataloaders(valid_cuts) + + scaler = GradScaler(enabled=params.use_fp16, init_scale=1.0) + + for epoch in range(params.start_epoch, params.num_epochs + 1): + logging.info(f"update num : {params.batch_idx_train}") + scheduler.step_epoch(epoch - 1) + fix_random_seed(params.seed + epoch - 1) + train_dl.sampler.set_epoch(epoch - 1) + + if tb_writer is not None: + tb_writer.add_scalar("train/epoch", epoch, params.batch_idx_train) + + params.cur_epoch = epoch + + train_one_epoch( + params=params, + model=model, + model_avg=model_avg, + optimizer=optimizer, + scheduler=scheduler, + sp=sp, + train_dl=train_dl, + valid_dl=valid_dl, + scaler=scaler, + tb_writer=tb_writer, + world_size=world_size, + rank=rank, + wb=wb, + ) + + if params.print_diagnostics: + diagnostic.print_diagnostics() + break + + ''' + if epoch % 10 == 0: + save_checkpoint( + params=params, + model=model, + model_avg=model_avg, + optimizer=optimizer, + scheduler=scheduler, + sampler=train_dl.sampler, + scaler=scaler, + rank=rank, + ) + ''' + + logging.info("Done!") + + if world_size > 1: + torch.distributed.barrier() + cleanup_dist() + + +def display_and_save_batch( + batch: dict, + params: AttributeDict, + sp: spm.SentencePieceProcessor, +) -> None: + """Display the batch statistics and save the batch into disk. + + Args: + batch: + A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()` + for the content in it. + params: + Parameters for training. See :func:`get_params`. + sp: + The BPE model. + """ + from lhotse.utils import uuid4 + + filename = f"{params.exp_dir}/batch-{uuid4()}.pt" + logging.info(f"Saving batch to {filename}") + torch.save(batch, filename) + + supervisions = batch["supervisions"] + features = batch["inputs"] + + logging.info(f"features shape: {features.shape}") + + y = sp.encode(supervisions["text"], out_type=int) + num_tokens = sum(len(i) for i in y) + logging.info(f"num tokens: {num_tokens}") + + +def scan_pessimistic_batches_for_oom( + model: Union[nn.Module, DDP], + train_dl: torch.utils.data.DataLoader, + optimizer: torch.optim.Optimizer, + sp: spm.SentencePieceProcessor, + params: AttributeDict, +): + from lhotse.dataset import find_pessimistic_batches + + logging.info( + "Sanity check -- see if any of the batches in epoch 1 would cause OOM." + ) + batches, crit_values = find_pessimistic_batches(train_dl.sampler) + for criterion, cuts in batches.items(): + batch = train_dl.dataset[cuts] + try: + with torch.cuda.amp.autocast(enabled=params.use_fp16): + loss, _ = compute_loss( + params=params, + model=model, + sp=sp, + batch=batch, + is_training=True, + ) + loss.backward() + optimizer.zero_grad() + except Exception as e: + if "CUDA out of memory" in str(e): + logging.error( + "Your GPU ran out of memory with the current " + "max_duration setting. We recommend decreasing " + "max_duration and trying again.\n" + f"Failing criterion: {criterion} " + f"(={crit_values[criterion]}) ..." + ) + display_and_save_batch(batch, params=params, sp=sp) + raise + logging.info( + f"Maximum memory allocated so far is {torch.cuda.max_memory_allocated()//1000000}MB" + ) + + +def main(): + parser = get_parser() + LibriSpeechAsrDataModule.add_arguments(parser) + args = parser.parse_args() + if args.wandb: args.exp_dir = args.exp_dir + str(random.randint(0,400)) + args.exp_dir = Path(args.exp_dir) + + logging.info("save arguments to config.yaml...") + save_args(args) + + if args.wandb: wb = wandb.init(project="d2v-adapter", entity="dohe0342", config=vars(args)) + else: wb = None + + world_size = args.world_size + assert world_size >= 1 + if world_size > 1: + mp.spawn(run if not args.add_adapter else run_adapter, + args=(world_size, args, wb), + nprocs=world_size, + join=True + ) + else: + if args.add_adapter: run_adapter(rank=0, world_size=1, args=args, wb=wb) + else: run(rank=0, world_size=1, args=args, wb=wb) + +torch.set_num_threads(1) +torch.set_num_interop_threads(1) + +if __name__ == "__main__": + main() diff --git a/egs/tedlium3/ASR/pruned_transducer_stateless_d2v_v2/pseudo.py b/egs/tedlium3/ASR/pruned_transducer_stateless_d2v_v2/pseudo.py new file mode 100755 index 000000000..21ca84332 --- /dev/null +++ b/egs/tedlium3/ASR/pruned_transducer_stateless_d2v_v2/pseudo.py @@ -0,0 +1,904 @@ +#!/usr/bin/env python3 +# +# Copyright 2021-2022 Xiaomi Corporation (Author: Fangjun Kuang, +# Zengwei Yao) +# +# See ../../../../LICENSE for clarification regarding multiple authors +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +""" +Usage: +(0) for d2v-T decoding +for method in greedy_search modified_beam_search fast_beam_search; do + ./pruned_transducer_stateless_d2v_v2/decode.py \ + --input-strategy AudioSamples \ + --enable-spec-aug False \ + --additional-block True \ + --model-name epoc.pt \ + --exp-dir ./pruned_transducer_stateless_d2v_v2/960h_sweep_v3_388 \ + --max-duration 400 \ + --decoding-method $method \ + --max-sym-per-frame 1 \ + --encoder-type d2v \ + --encoder-dim 768 \ + --decoder-dim 768 \ + --joiner-dim 768 +done +""" + + +import os +import argparse +import logging +import math +from collections import defaultdict +from pathlib import Path +from typing import Dict, List, Optional, Tuple + +import k2 +import sentencepiece as spm +import torch +import torch.nn as nn +from asr_datamodule import LibriSpeechAsrDataModule +from beam_search import ( + beam_search, + fast_beam_search_nbest, + fast_beam_search_nbest_LG, + fast_beam_search_nbest_oracle, + fast_beam_search_one_best, + greedy_search, + greedy_search_batch, + modified_beam_search, +) +from train import add_model_arguments, add_rep_arguments, get_params, get_transducer_model + +from icefall.checkpoint import ( + average_checkpoints, + average_checkpoints_with_averaged_model, + find_checkpoints, + load_checkpoint, +) +from icefall.lexicon import Lexicon +from icefall.utils import ( + AttributeDict, + setup_logger, + store_transcripts, + str2bool, + write_error_stats, +) + +LOG_EPS = math.log(1e-10) + + +def get_parser(): + parser = argparse.ArgumentParser( + formatter_class=argparse.ArgumentDefaultsHelpFormatter + ) + parser.add_argument( + "--model-name", + type=str, + default="", + help="""It specifies the model file name to use for decoding.""", + ) + + parser.add_argument( + "--epoch", + type=int, + default=30, + help="""It specifies the checkpoint to use for decoding. + Note: Epoch counts from 1. + You can specify --avg to use more checkpoints for model averaging.""", + ) + + parser.add_argument( + "--iter", + type=int, + default=0, + help="""If positive, --epoch is ignored and it + will use the checkpoint exp_dir/checkpoint-iter.pt. + You can specify --avg to use more checkpoints for model averaging. + """, + ) + + parser.add_argument( + "--avg", + type=int, + default=9, + help="Number of checkpoints to average. Automatically select " + "consecutive checkpoints before the checkpoint specified by " + "'--epoch' and '--iter'", + ) + + parser.add_argument( + "--use-averaged-model", + type=str2bool, + default=True, + help="Whether to load averaged model. Currently it only supports " + "using --epoch. If True, it would decode with the averaged model " + "over the epoch range from `epoch-avg` (excluded) to `epoch`." + "Actually only the models with epoch number of `epoch-avg` and " + "`epoch` are loaded for averaging. ", + ) + + parser.add_argument( + "--exp-dir", + type=str, + default="pruned_transducer_stateless7_ctc/exp", + help="The experiment dir", + ) + + parser.add_argument( + "--bpe-model", + type=str, + default="data/lang_bpe_500/bpe.model", + help="Path to the BPE model", + ) + + parser.add_argument( + "--lang-dir", + type=Path, + default="data/lang_bpe_500", + help="The lang dir containing word table and LG graph", + ) + + parser.add_argument( + "--decoding-method", + type=str, + default="greedy_search", + help="""Possible values are: + - greedy_search + - beam_search + - modified_beam_search + - fast_beam_search + - fast_beam_search_nbest + - fast_beam_search_nbest_oracle + - fast_beam_search_nbest_LG + If you use fast_beam_search_nbest_LG, you have to specify + `--lang-dir`, which should contain `LG.pt`. + """, + ) + + parser.add_argument( + "--beam-size", + type=int, + default=4, + help="""An integer indicating how many candidates we will keep for each + frame. Used only when --decoding-method is beam_search or + modified_beam_search.""", + ) + + parser.add_argument( + "--beam", + type=float, + default=20.0, + help="""A floating point value to calculate the cutoff score during beam + search (i.e., `cutoff = max-score - beam`), which is the same as the + `beam` in Kaldi. + Used only when --decoding-method is fast_beam_search, + fast_beam_search_nbest, fast_beam_search_nbest_LG, + and fast_beam_search_nbest_oracle + """, + ) + + parser.add_argument( + "--ngram-lm-scale", + type=float, + default=0.01, + help=""" + Used only when --decoding_method is fast_beam_search_nbest_LG. + It specifies the scale for n-gram LM scores. + """, + ) + + parser.add_argument( + "--max-contexts", + type=int, + default=8, + help="""Used only when --decoding-method is + fast_beam_search, fast_beam_search_nbest, fast_beam_search_nbest_LG, + and fast_beam_search_nbest_oracle""", + ) + + parser.add_argument( + "--max-states", + type=int, + default=64, + help="""Used only when --decoding-method is + fast_beam_search, fast_beam_search_nbest, fast_beam_search_nbest_LG, + and fast_beam_search_nbest_oracle""", + ) + + parser.add_argument( + "--context-size", + type=int, + default=2, + help="The context size in the decoder. 1 means bigram; 2 means tri-gram", + ) + parser.add_argument( + "--max-sym-per-frame", + type=int, + default=1, + help="""Maximum number of symbols per frame. + Used only when --decoding_method is greedy_search""", + ) + + parser.add_argument( + "--num-paths", + type=int, + default=200, + help="""Number of paths for nbest decoding. + Used only when the decoding method is fast_beam_search_nbest, + fast_beam_search_nbest_LG, and fast_beam_search_nbest_oracle""", + ) + + parser.add_argument( + "--nbest-scale", + type=float, + default=0.5, + help="""Scale applied to lattice scores when computing nbest paths. + Used only when the decoding method is fast_beam_search_nbest, + fast_beam_search_nbest_LG, and fast_beam_search_nbest_oracle""", + ) + + parser.add_argument( + "--simulate-streaming", + type=str2bool, + default=False, + help="""Whether to simulate streaming in decoding, this is a good way to + test a streaming model. + """, + ) + + parser.add_argument( + "--decode-chunk-size", + type=int, + default=16, + help="The chunk size for decoding (in frames after subsampling)", + ) + + parser.add_argument( + "--left-context", + type=int, + default=64, + help="left context can be seen during decoding (in frames after subsampling)", + ) + + add_model_arguments(parser) + add_rep_arguments(parser) + + return parser + + +def decode_one_batch( + params: AttributeDict, + model: nn.Module, + sp: spm.SentencePieceProcessor, + batch: dict, + word_table: Optional[k2.SymbolTable] = None, + decoding_graph: Optional[k2.Fsa] = None, +) -> Dict[str, List[List[str]]]: + """Decode one batch and return the result in a dict. The dict has the + following format: + + - key: It indicates the setting used for decoding. For example, + if greedy_search is used, it would be "greedy_search" + If beam search with a beam size of 7 is used, it would be + "beam_7" + - value: It contains the decoding result. `len(value)` equals to + batch size. `value[i]` is the decoding result for the i-th + utterance in the given batch. + Args: + params: + It's the return value of :func:`get_params`. + model: + The neural model. + sp: + The BPE model. + batch: + It is the return value from iterating + `lhotse.dataset.K2SpeechRecognitionDataset`. See its documentation + for the format of the `batch`. + word_table: + The word symbol table. + decoding_graph: + The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used + only when --decoding_method is fast_beam_search, fast_beam_search_nbest, + fast_beam_search_nbest_oracle, and fast_beam_search_nbest_LG. + Returns: + Return the decoding result. See above description for the format of + the returned dict. + """ + device = next(model.parameters()).device + feature = batch["inputs"] + assert feature.ndim == 2 or feature.ndim == 3 + + feature = feature.to(device) + # at entry, feature is (N, T, C) + + supervisions = batch["supervisions"] + #feature_lens = supervisions["num_frames"].to(device) + if feature.ndim == 2: + feature_lens = [] + for supervision in supervisions['cut']: + try: feature_lens.append(supervision.tracks[0].cut.recording.num_samples) + except: feature_lens.append(supervision.recording.num_samples) + feature_lens = torch.tensor(feature_lens) + + elif feature.ndim == 3: + feature_lens = supervisions["num_frames"].to(device) + + if params.simulate_streaming: + feature_lens += params.left_context + feature = torch.nn.functional.pad( + feature, + pad=(0, 0, 0, params.left_context), + value=LOG_EPS, + ) + encoder_out, encoder_out_lens, _ = model.encoder.streaming_forward( + x=feature, + x_lens=feature_lens, + chunk_size=params.decode_chunk_size, + left_context=params.left_context, + simulate_streaming=True, + ) + else: + encoder_out, encoder_out_lens = model.encoder(x=feature, x_lens=feature_lens) + + hyps = [] + + if params.decoding_method == "fast_beam_search": + hyp_tokens = fast_beam_search_one_best( + model=model, + decoding_graph=decoding_graph, + encoder_out=encoder_out, + encoder_out_lens=encoder_out_lens, + beam=params.beam, + max_contexts=params.max_contexts, + max_states=params.max_states, + ) + for hyp in sp.decode(hyp_tokens): + hyps.append(hyp.split()) + elif params.decoding_method == "fast_beam_search_nbest_LG": + hyp_tokens = fast_beam_search_nbest_LG( + model=model, + decoding_graph=decoding_graph, + encoder_out=encoder_out, + encoder_out_lens=encoder_out_lens, + beam=params.beam, + max_contexts=params.max_contexts, + max_states=params.max_states, + num_paths=params.num_paths, + nbest_scale=params.nbest_scale, + ) + for hyp in hyp_tokens: + hyps.append([word_table[i] for i in hyp]) + elif params.decoding_method == "fast_beam_search_nbest": + hyp_tokens = fast_beam_search_nbest( + model=model, + decoding_graph=decoding_graph, + encoder_out=encoder_out, + encoder_out_lens=encoder_out_lens, + beam=params.beam, + max_contexts=params.max_contexts, + max_states=params.max_states, + num_paths=params.num_paths, + nbest_scale=params.nbest_scale, + ) + for hyp in sp.decode(hyp_tokens): + hyps.append(hyp.split()) + elif params.decoding_method == "fast_beam_search_nbest_oracle": + hyp_tokens = fast_beam_search_nbest_oracle( + model=model, + decoding_graph=decoding_graph, + encoder_out=encoder_out, + encoder_out_lens=encoder_out_lens, + beam=params.beam, + max_contexts=params.max_contexts, + max_states=params.max_states, + num_paths=params.num_paths, + ref_texts=sp.encode(supervisions["text"]), + nbest_scale=params.nbest_scale, + ) + for hyp in sp.decode(hyp_tokens): + hyps.append(hyp.split()) + elif params.decoding_method == "greedy_search" and params.max_sym_per_frame == 1: + hyp_tokens = greedy_search_batch( + model=model, + encoder_out=encoder_out, + encoder_out_lens=encoder_out_lens, + ) + for hyp in sp.decode(hyp_tokens): + hyps.append(hyp.split()) + elif params.decoding_method == "modified_beam_search": + hyp_tokens = modified_beam_search( + model=model, + encoder_out=encoder_out, + encoder_out_lens=encoder_out_lens, + beam=params.beam_size, + ) + for hyp in sp.decode(hyp_tokens): + hyps.append(hyp.split()) + else: + batch_size = encoder_out.size(0) + + for i in range(batch_size): + # fmt: off + encoder_out_i = encoder_out[i:i+1, :encoder_out_lens[i]] + # fmt: on + if params.decoding_method == "greedy_search": + hyp = greedy_search( + model=model, + encoder_out=encoder_out_i, + max_sym_per_frame=params.max_sym_per_frame, + ) + elif params.decoding_method == "beam_search": + hyp = beam_search( + model=model, + encoder_out=encoder_out_i, + beam=params.beam_size, + ) + else: + raise ValueError( + f"Unsupported decoding method: {params.decoding_method}" + ) + hyps.append(sp.decode(hyp).split()) + + if params.decoding_method == "greedy_search": + return {"greedy_search": hyps} + elif "fast_beam_search" in params.decoding_method: + key = f"beam_{params.beam}_" + key += f"max_contexts_{params.max_contexts}_" + key += f"max_states_{params.max_states}" + if "nbest" in params.decoding_method: + key += f"_num_paths_{params.num_paths}_" + key += f"nbest_scale_{params.nbest_scale}" + if "LG" in params.decoding_method: + key += f"_ngram_lm_scale_{params.ngram_lm_scale}" + + return {key: hyps} + else: + return {f"beam_size_{params.beam_size}": hyps} + + +def decode_dataset( + dl: torch.utils.data.DataLoader, + params: AttributeDict, + model: nn.Module, + sp: spm.SentencePieceProcessor, + word_table: Optional[k2.SymbolTable] = None, + decoding_graph: Optional[k2.Fsa] = None, +) -> Dict[str, List[Tuple[str, List[str], List[str]]]]: + """Decode dataset. + + Args: + dl: + PyTorch's dataloader containing the dataset to decode. + params: + It is returned by :func:`get_params`. + model: + The neural model. + sp: + The BPE model. + word_table: + The word symbol table. + decoding_graph: + The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used + only when --decoding_method is fast_beam_search, fast_beam_search_nbest, + fast_beam_search_nbest_oracle, and fast_beam_search_nbest_LG. + Returns: + Return a dict, whose key may be "greedy_search" if greedy search + is used, or it may be "beam_7" if beam size of 7 is used. + Its value is a list of tuples. Each tuple contains two elements: + The first is the reference transcript, and the second is the + predicted result. + """ + num_cuts = 0 + + try: + num_batches = len(dl) + except TypeError: + num_batches = "?" + + if params.decoding_method == "greedy_search": + log_interval = 50 + else: + log_interval = 20 + + results = defaultdict(list) + for batch_idx, batch in enumerate(dl): + texts = batch["supervisions"]["text"] + + cut_ids = [cut.id for cut in batch["supervisions"]["cut"]] + + hyps_dict = decode_one_batch( + params=params, + model=model, + sp=sp, + decoding_graph=decoding_graph, + word_table=word_table, + batch=batch, + ) + + for name, hyps in hyps_dict.items(): + this_batch = [] + assert len(hyps) == len(texts) + for cut_id, hyp_words, ref_text in zip(cut_ids, hyps, texts): + ref_words = ref_text.split() + this_batch.append((cut_id, ref_words, hyp_words)) + + results[name].extend(this_batch) + + num_cuts += len(texts) + + if batch_idx % log_interval == 0: + batch_str = f"{batch_idx}/{num_batches}" + + logging.info(f"batch {batch_str}, cuts processed until now is {num_cuts}") + return results + + +def save_results( + params: AttributeDict, + test_set_name: str, + results_dict: Dict[str, List[Tuple[str, List[str], List[str]]]], +): + test_set_wers = dict() + for key, results in results_dict.items(): + recog_path = ( + params.res_dir / f"recogs-{test_set_name}-{key}-{params.suffix}.txt" + ) + results = sorted(results) + store_transcripts(filename=recog_path, texts=results) + logging.info(f"The transcripts are stored in {recog_path}") + + # The following prints out WERs, per-word error statistics and aligned + # ref/hyp pairs. + errs_filename = ( + params.res_dir / f"errs-{test_set_name}-{key}-{params.suffix}.txt" + ) + with open(errs_filename, "w") as f: + wer = write_error_stats( + f, f"{test_set_name}-{key}", results, enable_log=True + ) + test_set_wers[key] = wer + + logging.info("Wrote detailed error stats to {}".format(errs_filename)) + + test_set_wers = sorted(test_set_wers.items(), key=lambda x: x[1]) + errs_info = ( + params.res_dir / f"wer-summary-{test_set_name}-{key}-{params.suffix}.txt" + ) + with open(errs_info, "w") as f: + print("settings\tWER", file=f) + for key, val in test_set_wers: + print("{}\t{}".format(key, val), file=f) + + s = "\nFor {}, WER of different settings are:\n".format(test_set_name) + note = "\tbest for {}".format(test_set_name) + for key, val in test_set_wers: + s += "{}\t{}{}\n".format(key, val, note) + note = "" + logging.info(s) + + +@torch.no_grad() +def main(): + parser = get_parser() + LibriSpeechAsrDataModule.add_arguments(parser) + args = parser.parse_args() + args.exp_dir = Path(args.exp_dir) + + params = get_params() + params.update(vars(args)) + + assert params.decoding_method in ( + "greedy_search", + "beam_search", + "fast_beam_search", + "fast_beam_search_nbest", + "fast_beam_search_nbest_LG", + "fast_beam_search_nbest_oracle", + "modified_beam_search", + ) + params.res_dir = params.exp_dir / params.decoding_method + + if params.iter > 0: + params.suffix = f"iter-{params.iter}-avg-{params.avg}" + else: + params.suffix = f"epoch-{params.epoch}-avg-{params.avg}" + + if params.simulate_streaming: + params.suffix += f"-streaming-chunk-size-{params.decode_chunk_size}" + params.suffix += f"-left-context-{params.left_context}" + + if "fast_beam_search" in params.decoding_method: + params.suffix += f"-beam-{params.beam}" + params.suffix += f"-max-contexts-{params.max_contexts}" + params.suffix += f"-max-states-{params.max_states}" + if "nbest" in params.decoding_method: + params.suffix += f"-nbest-scale-{params.nbest_scale}" + params.suffix += f"-num-paths-{params.num_paths}" + if "LG" in params.decoding_method: + params.suffix += f"-ngram-lm-scale-{params.ngram_lm_scale}" + elif "beam_search" in params.decoding_method: + params.suffix += f"-{params.decoding_method}-beam-size-{params.beam_size}" + else: + params.suffix += f"-context-{params.context_size}" + params.suffix += f"-max-sym-per-frame-{params.max_sym_per_frame}" + + if params.use_averaged_model: + params.suffix += "-use-averaged-model" + + setup_logger(f"{params.res_dir}/log-decode-{params.suffix}") + logging.info("Decoding started") + + device = torch.device("cpu") + if torch.cuda.is_available(): + device = torch.device("cuda", 0) + + logging.info(f"Device: {device}") + + sp = spm.SentencePieceProcessor() + sp.load(params.bpe_model) + + # and are defined in local/train_bpe_model.py + params.blank_id = sp.piece_to_id("") + params.unk_id = sp.piece_to_id("") + params.vocab_size = sp.get_piece_size() + + if params.simulate_streaming: + assert ( + params.causal_convolution + ), "Decoding in streaming requires causal convolution" + + logging.info(params) + + logging.info("About to create model") + model = get_transducer_model(params) + + if params.model_name: + load_checkpoint(f"{params.exp_dir}/{params.model_name}", model) + else: + if not params.use_averaged_model: + if params.iter > 0: + filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[ + : params.avg + ] + if len(filenames) == 0: + raise ValueError( + f"No checkpoints found for" + f" --iter {params.iter}, --avg {params.avg}" + ) + elif len(filenames) < params.avg: + raise ValueError( + f"Not enough checkpoints ({len(filenames)}) found for" + f" --iter {params.iter}, --avg {params.avg}" + ) + logging.info(f"averaging {filenames}") + model.to(device) + model.load_state_dict(average_checkpoints(filenames, device=device)) + elif params.avg == 1: + load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model) + else: + start = params.epoch - params.avg + 1 + filenames = [] + for i in range(start, params.epoch + 1): + if i >= 1: + filenames.append(f"{params.exp_dir}/epoch-{i}.pt") + logging.info(f"averaging {filenames}") + model.to(device) + model.load_state_dict(average_checkpoints(filenames, device=device)) + else: + if params.iter > 0: + filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[ + : params.avg + 1 + ] + if len(filenames) == 0: + raise ValueError( + f"No checkpoints found for" + f" --iter {params.iter}, --avg {params.avg}" + ) + elif len(filenames) < params.avg + 1: + raise ValueError( + f"Not enough checkpoints ({len(filenames)}) found for" + f" --iter {params.iter}, --avg {params.avg}" + ) + filename_start = filenames[-1] + filename_end = filenames[0] + logging.info( + "Calculating the averaged model over iteration checkpoints" + f" from {filename_start} (excluded) to {filename_end}" + ) + model.to(device) + model.load_state_dict( + average_checkpoints_with_averaged_model( + filename_start=filename_start, + filename_end=filename_end, + device=device, + ) + ) + else: + assert params.avg > 0, params.avg + start = params.epoch - params.avg + assert start >= 1, start + filename_start = f"{params.exp_dir}/epoch-{start}.pt" + filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt" + logging.info( + f"Calculating the averaged model over epoch range from " + f"{start} (excluded) to {params.epoch}" + ) + model.to(device) + model.load_state_dict( + average_checkpoints_with_averaged_model( + filename_start=filename_start, + filename_end=filename_end, + device=device, + ) + ) + + model.to(device) + model.eval() + + if "fast_beam_search" in params.decoding_method: + if params.decoding_method == "fast_beam_search_nbest_LG": + lexicon = Lexicon(params.lang_dir) + word_table = lexicon.word_table + lg_filename = params.lang_dir / "LG.pt" + logging.info(f"Loading {lg_filename}") + decoding_graph = k2.Fsa.from_dict( + torch.load(lg_filename, map_location=device) + ) + decoding_graph.scores *= params.ngram_lm_scale + else: + word_table = None + decoding_graph = k2.trivial_graph(params.vocab_size - 1, device=device) + else: + decoding_graph = None + word_table = None + + num_param = sum([p.numel() for p in model.parameters()]) + logging.info(f"Number of model parameters: {num_param}") + + # we need cut ids to display recognition results. + args.return_cuts = True + librispeech = LibriSpeechAsrDataModule(args) + + #test_clean_cuts = librispeech.test_clean_cuts() + #test_clean_cuts = librispeech.test_clean_cuts(option='male') + #test_other_cuts = librispeech.test_other_cuts(option='male') + + test_clean_cuts = librispeech.vox_cuts(option=params.spk_id) + def remove_short_and_long_utt(c): + return 1.0 <= c.duration <= 20.0 + + test_clean_cuts = test_clean_cuts.filter(remove_short_and_long_utt) + #test_clean_cuts = librispeech.test_clean_user(option=option) + #test_other_cuts = librispeech.test_other_user(option=option) + #test_clean_dl = librispeech.train_dataloaders(test_clean_cuts) + test_clean_dl = librispeech.test_dataloaders(test_clean_cuts) + #test_other_dl = librispeech.test_dataloaders(test_other_cuts) + + test_sets = [f"test-clean_sampling"] + #test_sets = [f"test-clean_sampling", f"test-other_sampling"] + test_dl = [test_clean_dl] + #test_dl = [test_clean_dl, test_other_dl] + + for test_set, test_dl in zip(test_sets, test_dl): + results_dict = decode_dataset( + dl=test_dl, + params=params, + model=model, + sp=sp, + word_table=word_table, + decoding_graph=decoding_graph, + ) + #results = results_dict['greedy_search'] + print('--------------') + print(results_dict.keys()) + print('--------------') + results = results_dict['beam_size_4'] + #jsons = open(f"{params.manifest_dir}/userlibri/{test_set}/{option}.jsonl", 'r').readlines() + #new_jsons = open(f"{params.manifest_dir}/userlibri/{test_set}/{option}_p.jsonl", 'w') + + res_dict = {} + for res in results: + hypo = res + res_dict[res[0]] = ' '.join(res[2]) + + res_dict = sorted(res_dict.items(), key=lambda x:x[0]) + + try: os.makedirs(f"/DB/LibriSpeech_tar/{params.prefix}/{params.spk_id}_texts") + except: os.makedirs(f"/home/work/workspace/LibriSpeech/{params.prefix}/{params.spk_id}_texts") + + for k, v in res_dict: + #v = v.strip() + #if len(v) < 10: + # continue + utt_id = '-'.join(k.split('-')[:-1]) + #print(utt_id) + try: f = open(f'/DB/LibriSpeech_tar/{params.prefix}/{params.spk_id}_texts/{utt_id}.txt', 'w') + except: f = open(f'/home/work/workspace/LibriSpeech/{params.prefix}/{params.spk_id}_texts/{utt_id}.txt', 'w') + f.write(v) + print(k, v) + + if 0: + for line in jsons: + splited = line.split() + utt_id = splited[1][1:-2] + text_idx = splited.index('"text":') + + pseudo = f'"greedy pseudo text": "{res_dict[utt_id]}",' + #splited.insert(text_idx, pseudo) + splited.insert(len(splited)-2, pseudo) + new_line = ' '.join(splited) + new_line += '\n' + + new_jsons.write(new_line) + ''' + save_results( + params=params, + test_set_name=test_set, + results_dict=results_dict, + ) + ''' + + ''' + for test_set, test_dl in zip(test_sets, test_dl): + results_dict = decode_dataset( + dl=test_dl, + params=params, + model=model, + sp=sp, + word_table=word_table, + decoding_graph=decoding_graph, + ) + + save_results( + params=params, + test_set_name=test_set, + results_dict=results_dict, + ) + ''' + ''' + test_clean_cuts, test_clean_sets = librispeech.test_clean_cuts(option='user') + test_other_cuts, test_other_sets = librispeech.test_other_cuts(option='user') + + test_clean_dl = [librispeech.test_dataloaders(user) for user in test_clean_cuts] + test_other_dl = [librispeech.test_dataloaders(user) for user in test_other_cuts] + + test_sets = [test_clean_sets, test_other_sets] + test_dl = [test_clean_dl, test_other_dl] + + for sets, dls in zip(test_sets, test_dl): + print(len(sets), len(dls)) + for test_set, test_dl in zip(sets, dls): + results_dict = decode_dataset( + dl=test_dl, + params=params, + model=model, + sp=sp, + word_table=word_table, + decoding_graph=decoding_graph, + ) + + save_results( + params=params, + test_set_name=test_set, + results_dict=results_dict, + ) + ''' + logging.info("Done!") + + +if __name__ == "__main__": + main() diff --git a/egs/tedlium3/ASR/pruned_transducer_stateless_d2v_v2/scaling.py b/egs/tedlium3/ASR/pruned_transducer_stateless_d2v_v2/scaling.py new file mode 100644 index 000000000..6f63e0629 --- /dev/null +++ b/egs/tedlium3/ASR/pruned_transducer_stateless_d2v_v2/scaling.py @@ -0,0 +1,1178 @@ +# Copyright 2022 Xiaomi Corp. (authors: Daniel Povey) +# +# See ../../../../LICENSE for clarification regarding multiple authors +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + + +import collections +import logging +import random +from functools import reduce +from itertools import repeat +from typing import Optional, Tuple, Union + +import torch +import torch.nn as nn +import torch.nn.functional as F +from torch import Tensor +from torch.nn import Embedding as ScaledEmbedding + + +class ActivationBalancerFunction(torch.autograd.Function): + @staticmethod + def forward( + ctx, + x: Tensor, + scale_factor: Tensor, + sign_factor: Optional[Tensor], + channel_dim: int, + ) -> Tensor: + if channel_dim < 0: + channel_dim += x.ndim + ctx.channel_dim = channel_dim + xgt0 = x > 0 + if sign_factor is None: + ctx.save_for_backward(xgt0, scale_factor) + else: + ctx.save_for_backward(xgt0, scale_factor, sign_factor) + return x + + @staticmethod + def backward(ctx, x_grad: Tensor) -> Tuple[Tensor, None, None, None]: + if len(ctx.saved_tensors) == 3: + xgt0, scale_factor, sign_factor = ctx.saved_tensors + for _ in range(ctx.channel_dim, x_grad.ndim - 1): + scale_factor = scale_factor.unsqueeze(-1) + sign_factor = sign_factor.unsqueeze(-1) + factor = sign_factor + scale_factor * (xgt0.to(x_grad.dtype) - 0.5) + else: + xgt0, scale_factor = ctx.saved_tensors + for _ in range(ctx.channel_dim, x_grad.ndim - 1): + scale_factor = scale_factor.unsqueeze(-1) + factor = scale_factor * (xgt0.to(x_grad.dtype) - 0.5) + neg_delta_grad = x_grad.abs() * factor + return ( + x_grad - neg_delta_grad, + None, + None, + None, + ) + + +def _compute_scale_factor( + x: Tensor, + channel_dim: int, + min_abs: float, + max_abs: float, + gain_factor: float, + max_factor: float, +) -> Tensor: + if channel_dim < 0: + channel_dim += x.ndim + sum_dims = [d for d in range(x.ndim) if d != channel_dim] + x_abs_mean = torch.mean(x.abs(), dim=sum_dims).to(torch.float32) + + if min_abs == 0.0: + below_threshold = 0.0 + else: + # below_threshold is 0 if x_abs_mean > min_abs, can be at most max_factor if + # x_abs)_mean , min_abs. + below_threshold = ((min_abs - x_abs_mean) * (gain_factor / min_abs)).clamp( + min=0, max=max_factor + ) + + above_threshold = ((x_abs_mean - max_abs) * (gain_factor / max_abs)).clamp( + min=0, max=max_factor + ) + + return below_threshold - above_threshold + + +def _compute_sign_factor( + x: Tensor, + channel_dim: int, + min_positive: float, + max_positive: float, + gain_factor: float, + max_factor: float, +) -> Tensor: + if channel_dim < 0: + channel_dim += x.ndim + sum_dims = [d for d in range(x.ndim) if d != channel_dim] + proportion_positive = torch.mean((x > 0).to(torch.float32), dim=sum_dims) + if min_positive == 0.0: + factor1 = 0.0 + else: + # 0 if proportion_positive >= min_positive, else can be + # as large as max_factor. + factor1 = ( + (min_positive - proportion_positive) * (gain_factor / min_positive) + ).clamp_(min=0, max=max_factor) + + if max_positive == 1.0: + factor2 = 0.0 + else: + # 0 if self.proportion_positive <= max_positive, else can be + # as large as -max_factor. + factor2 = ( + (proportion_positive - max_positive) * (gain_factor / (1.0 - max_positive)) + ).clamp_(min=0, max=max_factor) + sign_factor = factor1 - factor2 + # require min_positive != 0 or max_positive != 1: + assert not isinstance(sign_factor, float) + return sign_factor + + +class ActivationScaleBalancerFunction(torch.autograd.Function): + """ + This object is used in class ActivationBalancer when the user specified + min_positive=0, max_positive=1, so there are no constraints on the signs + of the activations and only the absolute value has a constraint. + """ + + @staticmethod + def forward( + ctx, + x: Tensor, + sign_factor: Tensor, + scale_factor: Tensor, + channel_dim: int, + ) -> Tensor: + if channel_dim < 0: + channel_dim += x.ndim + ctx.channel_dim = channel_dim + xgt0 = x > 0 + ctx.save_for_backward(xgt0, sign_factor, scale_factor) + return x + + @staticmethod + def backward(ctx, x_grad: Tensor) -> Tuple[Tensor, None, None, None]: + xgt0, sign_factor, scale_factor = ctx.saved_tensors + for _ in range(ctx.channel_dim, x_grad.ndim - 1): + sign_factor = sign_factor.unsqueeze(-1) + scale_factor = scale_factor.unsqueeze(-1) + + factor = sign_factor + scale_factor * (xgt0.to(x_grad.dtype) - 0.5) + neg_delta_grad = x_grad.abs() * factor + return ( + x_grad - neg_delta_grad, + None, + None, + None, + ) + + +class RandomClampFunction(torch.autograd.Function): + @staticmethod + def forward( + ctx, + x: Tensor, + min: Optional[float], + max: Optional[float], + prob: float, + reflect: float, + ) -> Tensor: + x_clamped = torch.clamp(x, min=min, max=max) + mask = torch.rand_like(x) < prob + ans = torch.where(mask, x_clamped, x) + if x.requires_grad: + ctx.save_for_backward(ans == x) + ctx.reflect = reflect + if reflect != 0.0: + ans = ans * (1.0 + reflect) - (x * reflect) + return ans + + @staticmethod + def backward(ctx, ans_grad: Tensor) -> Tuple[Tensor, None, None, None, None]: + (is_same,) = ctx.saved_tensors + x_grad = ans_grad * is_same.to(ans_grad.dtype) + reflect = ctx.reflect + if reflect != 0.0: + x_grad = x_grad * (1.0 + reflect) - (ans_grad * reflect) + return x_grad, None, None, None, None + + +def random_clamp( + x: Tensor, + min: Optional[float] = None, + max: Optional[float] = None, + prob: float = 0.5, + reflect: float = 0.0, +): + return RandomClampFunction.apply(x, min, max, prob, reflect) + + +def random_cast_to_half(x: Tensor, min_abs: float = 5.0e-06) -> Tensor: + """ + A randomized way of casting a floating point value to half precision. + """ + if x.dtype == torch.float16: + return x + x_abs = x.abs() + is_too_small = x_abs < min_abs + # for elements where is_too_small is true, random_val will contain +-min_abs with + # probability (x.abs() / min_abs), and 0.0 otherwise. [so this preserves expectations, + # for those elements]. + random_val = min_abs * x.sign() * (torch.rand_like(x) * min_abs < x_abs) + return torch.where(is_too_small, random_val, x).to(torch.float16) + + +class RandomGradFunction(torch.autograd.Function): + """ + Does nothing in forward pass; in backward pass, gets rid of very small grads using + randomized approach that preserves expectations (intended to reduce roundoff). + """ + + @staticmethod + def forward(ctx, x: Tensor, min_abs: float) -> Tensor: + ctx.min_abs = min_abs + return x + + @staticmethod + def backward(ctx, ans_grad: Tensor) -> Tuple[Tensor, None]: + if ans_grad.dtype == torch.float16: + return ( + random_cast_to_half(ans_grad.to(torch.float32), min_abs=ctx.min_abs), + None, + ) + else: + return ans_grad, None + + +class RandomGrad(torch.nn.Module): + """ + Gets rid of very small gradients using an expectation-preserving method, intended to increase + accuracy of training when using amp (automatic mixed precision) + """ + + def __init__(self, min_abs: float = 5.0e-06): + super(RandomGrad, self).__init__() + self.min_abs = min_abs + + def forward(self, x: Tensor): + if torch.jit.is_scripting() or not self.training: + return x + else: + return RandomGradFunction.apply(x, self.min_abs) + + +class SoftmaxFunction(torch.autograd.Function): + """ + Tries to handle half-precision derivatives in a randomized way that should + be more accurate for training than the default behavior. + """ + + @staticmethod + def forward(ctx, x: Tensor, dim: int): + ans = x.softmax(dim=dim) + # if x dtype is float16, x.softmax() returns a float32 because + # (presumably) that op does not support float16, and autocast + # is enabled. + if torch.is_autocast_enabled(): + ans = ans.to(torch.float16) + ctx.save_for_backward(ans) + ctx.x_dtype = x.dtype + ctx.dim = dim + return ans + + @staticmethod + def backward(ctx, ans_grad: Tensor): + (ans,) = ctx.saved_tensors + with torch.cuda.amp.autocast(enabled=False): + ans_grad = ans_grad.to(torch.float32) + ans = ans.to(torch.float32) + x_grad = ans_grad * ans + x_grad = x_grad - ans * x_grad.sum(dim=ctx.dim, keepdim=True) + return x_grad, None + + +def softmax(x: Tensor, dim: int): + if torch.jit.is_scripting(): + return x.softmax(dim) + + return SoftmaxFunction.apply(x, dim) + + +class MaxEigLimiterFunction(torch.autograd.Function): + @staticmethod + def forward( + ctx, + x: Tensor, + coeffs: Tensor, + direction: Tensor, + channel_dim: int, + grad_scale: float, + ) -> Tensor: + ctx.channel_dim = channel_dim + ctx.grad_scale = grad_scale + ctx.save_for_backward(x.detach(), coeffs.detach(), direction.detach()) + return x + + @staticmethod + def backward(ctx, x_grad, *args): + with torch.enable_grad(): + (x_orig, coeffs, new_direction) = ctx.saved_tensors + x_orig.requires_grad = True + num_channels = x_orig.shape[ctx.channel_dim] + x = x_orig.transpose(ctx.channel_dim, -1).reshape(-1, num_channels) + new_direction.requires_grad = False + x = x - x.mean(dim=0) + x_var = (x**2).mean() + x_residual = x - coeffs * new_direction + x_residual_var = (x_residual**2).mean() + # `variance_proportion` is the proportion of the variance accounted for + # by the top eigen-direction. This is to be minimized. + variance_proportion = (x_var - x_residual_var) / (x_var + 1.0e-20) + variance_proportion.backward() + x_orig_grad = x_orig.grad + x_extra_grad = ( + x_orig.grad + * ctx.grad_scale + * x_grad.norm() + / (x_orig_grad.norm() + 1.0e-20) + ) + return x_grad + x_extra_grad.detach(), None, None, None, None + + +class BasicNorm(torch.nn.Module): + """ + This is intended to be a simpler, and hopefully cheaper, replacement for + LayerNorm. The observation this is based on, is that Transformer-type + networks, especially with pre-norm, sometimes seem to set one of the + feature dimensions to a large constant value (e.g. 50), which "defeats" + the LayerNorm because the output magnitude is then not strongly dependent + on the other (useful) features. Presumably the weight and bias of the + LayerNorm are required to allow it to do this. + + So the idea is to introduce this large constant value as an explicit + parameter, that takes the role of the "eps" in LayerNorm, so the network + doesn't have to do this trick. We make the "eps" learnable. + + Args: + num_channels: the number of channels, e.g. 512. + channel_dim: the axis/dimension corresponding to the channel, + interprted as an offset from the input's ndim if negative. + shis is NOT the num_channels; it should typically be one of + {-2, -1, 0, 1, 2, 3}. + eps: the initial "epsilon" that we add as ballast in: + scale = ((input_vec**2).mean() + epsilon)**-0.5 + Note: our epsilon is actually large, but we keep the name + to indicate the connection with conventional LayerNorm. + learn_eps: if true, we learn epsilon; if false, we keep it + at the initial value. + eps_min: float + eps_max: float + """ + + def __init__( + self, + num_channels: int, + channel_dim: int = -1, # CAUTION: see documentation. + eps: float = 0.25, + learn_eps: bool = True, + eps_min: float = -3.0, + eps_max: float = 3.0, + ) -> None: + super(BasicNorm, self).__init__() + self.num_channels = num_channels + self.channel_dim = channel_dim + if learn_eps: + self.eps = nn.Parameter(torch.tensor(eps).log().detach()) + else: + self.register_buffer("eps", torch.tensor(eps).log().detach()) + self.eps_min = eps_min + self.eps_max = eps_max + + def forward(self, x: Tensor) -> Tensor: + assert x.shape[self.channel_dim] == self.num_channels + eps = self.eps + if self.training and random.random() < 0.25: + # with probability 0.25, in training mode, clamp eps between the min + # and max; this will encourage it to learn parameters within the + # allowed range by making parameters that are outside the allowed + # range noisy. + + # gradients to allow the parameter to get back into the allowed + # region if it happens to exit it. + eps = eps.clamp(min=self.eps_min, max=self.eps_max) + scales = ( + torch.mean(x**2, dim=self.channel_dim, keepdim=True) + eps.exp() + ) ** -0.5 + return x * scales + + +def ScaledLinear(*args, initial_scale: float = 1.0, **kwargs) -> nn.Linear: + """ + Behaves like a constructor of a modified version of nn.Linear + that gives an easy way to set the default initial parameter scale. + + Args: + Accepts the standard args and kwargs that nn.Linear accepts + e.g. in_features, out_features, bias=False. + + initial_scale: you can override this if you want to increase + or decrease the initial magnitude of the module's output + (affects the initialization of weight_scale and bias_scale). + Another option, if you want to do something like this, is + to re-initialize the parameters. + """ + ans = nn.Linear(*args, **kwargs) + with torch.no_grad(): + ans.weight[:] *= initial_scale + if ans.bias is not None: + torch.nn.init.uniform_(ans.bias, -0.1 * initial_scale, 0.1 * initial_scale) + return ans + + +def ScaledConv1d(*args, initial_scale: float = 1.0, **kwargs) -> nn.Conv1d: + """ + Behaves like a constructor of a modified version of nn.Conv1d + that gives an easy way to set the default initial parameter scale. + + Args: + Accepts the standard args and kwargs that nn.Linear accepts + e.g. in_features, out_features, bias=False. + + initial_scale: you can override this if you want to increase + or decrease the initial magnitude of the module's output + (affects the initialization of weight_scale and bias_scale). + Another option, if you want to do something like this, is + to re-initialize the parameters. + """ + ans = nn.Conv1d(*args, **kwargs) + with torch.no_grad(): + ans.weight[:] *= initial_scale + if ans.bias is not None: + torch.nn.init.uniform_(ans.bias, -0.1 * initial_scale, 0.1 * initial_scale) + return ans + + +class ActivationBalancer(torch.nn.Module): + """ + Modifies the backpropped derivatives of a function to try to encourage, for + each channel, that it is positive at least a proportion `threshold` of the + time. It does this by multiplying negative derivative values by up to + (1+max_factor), and positive derivative values by up to (1-max_factor), + interpolated from 1 at the threshold to those extremal values when none + of the inputs are positive. + + Args: + num_channels: the number of channels + channel_dim: the dimension/axis corresponding to the channel, e.g. + -1, 0, 1, 2; will be interpreted as an offset from x.ndim if negative. + min_positive: the minimum, per channel, of the proportion of the time + that (x > 0), below which we start to modify the derivatives. + max_positive: the maximum, per channel, of the proportion of the time + that (x > 0), above which we start to modify the derivatives. + max_factor: the maximum factor by which we modify the derivatives for + either the sign constraint or the magnitude constraint; + e.g. with max_factor=0.02, the the derivatives would be multiplied by + values in the range [0.98..1.02]. + sign_gain_factor: determines the 'gain' with which we increase the + change in gradient once the constraints on min_positive and max_positive + are violated. + scale_gain_factor: determines the 'gain' with which we increase the + change in gradient once the constraints on min_abs and max_abs + are violated. + min_abs: the minimum average-absolute-value difference from the mean + value per channel, which we allow, before we start to modify + the derivatives to prevent this. + max_abs: the maximum average-absolute-value difference from the mean + value per channel, which we allow, before we start to modify + the derivatives to prevent this. + min_prob: determines the minimum probability with which we modify the + gradients for the {min,max}_positive and {min,max}_abs constraints, + on each forward(). This is done randomly to prevent all layers + from doing it at the same time. Early in training we may use + higher probabilities than this; it will decay to this value. + """ + + def __init__( + self, + num_channels: int, + channel_dim: int, + min_positive: float = 0.05, + max_positive: float = 0.95, + max_factor: float = 0.04, + sign_gain_factor: float = 0.01, + scale_gain_factor: float = 0.02, + min_abs: float = 0.2, + max_abs: float = 100.0, + min_prob: float = 0.1, + ): + super(ActivationBalancer, self).__init__() + self.num_channels = num_channels + self.channel_dim = channel_dim + self.min_positive = min_positive + self.max_positive = max_positive + self.max_factor = max_factor + self.min_abs = min_abs + self.max_abs = max_abs + self.min_prob = min_prob + self.sign_gain_factor = sign_gain_factor + self.scale_gain_factor = scale_gain_factor + + # count measures how many times the forward() function has been called. + # We occasionally sync this to a tensor called `count`, that exists to + # make sure it is synced to disk when we load and save the model. + self.cpu_count = 0 + self.register_buffer("count", torch.tensor(0, dtype=torch.int64)) + + def forward(self, x: Tensor) -> Tensor: + if torch.jit.is_scripting() or not x.requires_grad: + return _no_op(x) + + count = self.cpu_count + self.cpu_count += 1 + + if random.random() < 0.01: + # Occasionally sync self.cpu_count with self.count. + # count affects the decay of 'prob'. don't do this on every iter, + # because syncing with the GPU is slow. + self.cpu_count = max(self.cpu_count, self.count.item()) + self.count.fill_(self.cpu_count) + + # the prob of doing some work exponentially decreases from 0.5 till it hits + # a floor at min_prob (==0.1, by default) + prob = max(self.min_prob, 0.5 ** (1 + (count / 4000.0))) + + if random.random() < prob: + sign_gain_factor = 0.5 + if self.min_positive != 0.0 or self.max_positive != 1.0: + sign_factor = _compute_sign_factor( + x, + self.channel_dim, + self.min_positive, + self.max_positive, + gain_factor=self.sign_gain_factor / prob, + max_factor=self.max_factor, + ) + else: + sign_factor = None + + scale_factor = _compute_scale_factor( + x, + self.channel_dim, + min_abs=self.min_abs, + max_abs=self.max_abs, + gain_factor=self.scale_gain_factor / prob, + max_factor=self.max_factor, + ) + return ActivationBalancerFunction.apply( + x, + scale_factor, + sign_factor, + self.channel_dim, + ) + else: + return _no_op(x) + + +def penalize_abs_values_gt(x: Tensor, limit: float, penalty: float) -> Tensor: + """ + Returns x unmodified, but in backprop will put a penalty for the excess of + the absolute values of elements of x over the limit "limit". E.g. if + limit == 10.0, then if x has any values over 10 it will get a penalty. + + Caution: the value of this penalty will be affected by grad scaling used + in automatic mixed precision training. For this reasons we use this, + it shouldn't really matter, or may even be helpful; we just use this + to disallow really implausible values of scores to be given to softmax. + """ + x_sign = x.sign() + over_limit = (x.abs() - limit) > 0 + # The following is a memory efficient way to penalize the absolute values of + # x that's over the limit. (The memory efficiency comes when you think + # about which items torch needs to cache for the autograd, and which ones it + # can throw away). The numerical value of aux_loss as computed here will + # actually be larger than it should be, by limit * over_limit.sum(), but it + # has the same derivative as the real aux_loss which is penalty * (x.abs() - + # limit).relu(). + aux_loss = penalty * ((x_sign * over_limit).to(torch.int8) * x) + # note: we don't do sum() here on aux)_loss, but it's as if we had done + # sum() due to how with_loss() works. + x = with_loss(x, aux_loss) + # you must use x for something, or this will be ineffective. + return x + + +def _diag(x: Tensor): # like .diag(), but works for tensors with 3 dims. + if x.ndim == 2: + return x.diag() + else: + (batch, dim, dim) = x.shape + x = x.reshape(batch, dim * dim) + x = x[:, :: dim + 1] + assert x.shape == (batch, dim) + return x + + +def _whitening_metric(x: Tensor, num_groups: int): + """ + Computes the "whitening metric", a value which will be 1.0 if all the eigenvalues of + of the centered feature covariance are the same within each group's covariance matrix + and also between groups. + Args: + x: a Tensor of shape (*, num_channels) + num_groups: the number of groups of channels, a number >=1 that divides num_channels + Returns: + Returns a scalar Tensor that will be 1.0 if the data is "perfectly white" and + greater than 1.0 otherwise. + """ + assert x.dtype != torch.float16 + x = x.reshape(-1, x.shape[-1]) + (num_frames, num_channels) = x.shape + assert num_channels % num_groups == 0 + channels_per_group = num_channels // num_groups + x = x.reshape(num_frames, num_groups, channels_per_group).transpose(0, 1) + # x now has shape (num_groups, num_frames, channels_per_group) + # subtract the mean so we use the centered, not uncentered, covariance. + # My experience has been that when we "mess with the gradients" like this, + # it's better not do anything that tries to move the mean around, because + # that can easily cause instability. + x = x - x.mean(dim=1, keepdim=True) + # x_covar: (num_groups, channels_per_group, channels_per_group) + x_covar = torch.matmul(x.transpose(1, 2), x) + x_covar_mean_diag = _diag(x_covar).mean() + # the following expression is what we'd get if we took the matrix product + # of each covariance and measured the mean of its trace, i.e. + # the same as _diag(torch.matmul(x_covar, x_covar)).mean(). + x_covarsq_mean_diag = (x_covar**2).sum() / (num_groups * channels_per_group) + # this metric will be >= 1.0; the larger it is, the less 'white' the data was. + metric = x_covarsq_mean_diag / (x_covar_mean_diag**2 + 1.0e-20) + return metric + + +class WhiteningPenaltyFunction(torch.autograd.Function): + @staticmethod + def forward( + ctx, x: Tensor, num_groups: int, whitening_limit: float, grad_scale: float + ) -> Tensor: + ctx.save_for_backward(x) + ctx.num_groups = num_groups + ctx.whitening_limit = whitening_limit + ctx.grad_scale = grad_scale + return x + + @staticmethod + def backward(ctx, x_grad: Tensor): + (x_orig,) = ctx.saved_tensors + with torch.enable_grad(): + with torch.cuda.amp.autocast(enabled=False): + x_detached = x_orig.to(torch.float32).detach() + x_detached.requires_grad = True + + metric = _whitening_metric(x_detached, ctx.num_groups) + + if random.random() < 0.005 or __name__ == "__main__": + logging.info( + f"Whitening: num_groups={ctx.num_groups}, num_channels={x_orig.shape[-1]}, " + f"metric={metric.item():.2f} vs. limit={ctx.whitening_limit}" + ) + + (metric - ctx.whitening_limit).relu().backward() + penalty_grad = x_detached.grad + scale = ctx.grad_scale * ( + x_grad.to(torch.float32).norm() / (penalty_grad.norm() + 1.0e-20) + ) + penalty_grad = penalty_grad * scale + return x_grad + penalty_grad.to(x_grad.dtype), None, None, None + + +class Whiten(nn.Module): + def __init__( + self, + num_groups: int, + whitening_limit: float, + prob: Union[float, Tuple[float, float]], + grad_scale: float, + ): + """ + Args: + num_groups: the number of groups to divide the channel dim into before + whitening. We will attempt to make the feature covariance + within each group, after mean subtraction, as "white" as possible, + while having the same trace across all groups. + whitening_limit: a value greater than 1.0, that dictates how much + freedom we have to violate the constraints. 1.0 would mean perfectly + white, with exactly the same trace across groups; larger values + give more freedom. E.g. 2.0. + prob: the probability with which we apply the gradient modification + (also affects the grad scale). May be supplied as a float, + or as a pair (min_prob, max_prob) + + grad_scale: determines the scale on the gradient term from this object, + relative to the rest of the gradient on the attention weights. + E.g. 0.02 (you may want to use smaller values than this if prob is large) + """ + super(Whiten, self).__init__() + assert num_groups >= 1 + assert whitening_limit >= 1 + assert grad_scale >= 0 + self.num_groups = num_groups + self.whitening_limit = whitening_limit + if isinstance(prob, float): + assert 0 < prob <= 1 + self.prob = prob + else: + (self.min_prob, self.max_prob) = prob + assert 0 < self.min_prob < self.max_prob <= 1 + self.prob = self.max_prob + + self.grad_scale = grad_scale + + def forward(self, x: Tensor) -> Tensor: + """ + In the forward pass, this function just returns the input unmodified. + In the backward pass, it will modify the gradients to ensure that the + distribution in each group has close to (lambda times I) as the covariance + after mean subtraction, with the same lambda across groups. + For whitening_limit > 1, there will be more freedom to violate this + constraint. + + Args: + x: the input of shape (*, num_channels) + + Returns: + x, unmodified. You should make sure + you use the returned value, or the graph will be freed + and nothing will happen in backprop. + """ + if not x.requires_grad or random.random() > self.prob or self.grad_scale == 0: + return _no_op(x) + else: + if hasattr(self, "min_prob") and random.random() < 0.25: + # occasionally switch between min_prob and max_prob, based on whether + # we are above or below the threshold. + if ( + _whitening_metric(x.to(torch.float32), self.num_groups) + > self.whitening_limit + ): + # there would be a change to the grad. + self.prob = self.max_prob + else: + self.prob = self.min_prob + + return WhiteningPenaltyFunction.apply( + x, self.num_groups, self.whitening_limit, self.grad_scale + ) + + +class WithLoss(torch.autograd.Function): + @staticmethod + def forward(ctx, x: Tensor, y: Tensor): + ctx.y_shape = y.shape + return x + + @staticmethod + def backward(ctx, ans_grad: Tensor): + return ans_grad, torch.ones( + ctx.y_shape, dtype=ans_grad.dtype, device=ans_grad.device + ) + + +def with_loss(x, y): + if torch.jit.is_scripting(): + return x + # returns x but adds y.sum() to the loss function. + return WithLoss.apply(x, y) + + +def _no_op(x: Tensor) -> Tensor: + if torch.jit.is_scripting(): + return x + else: + # a no-op function that will have a node in the autograd graph, + # to avoid certain bugs relating to backward hooks + return x.chunk(1, dim=-1)[0] + + +class Identity(torch.nn.Module): + def __init__(self): + super(Identity, self).__init__() + + def forward(self, x): + return _no_op(x) + + +class MaxEig(torch.nn.Module): + """ + Modifies the backpropped derivatives of a function to try to discourage + that any given direction in activation space accounts for more than + a specified proportion of the covariance (e.g. 0.2). + + + Args: + num_channels: the number of channels + channel_dim: the dimension/axis corresponding to the channel, e.g. + -1, 0, 1, 2; will be interpreted as an offset from x.ndim if negative. + max_var_per_eig: the maximum proportion of the variance of the + features/channels, after mean subtraction, that can come from + any given eigenvalue. + min_prob: the minimum probability with which we apply this during any invocation + of forward(), assuming last time we applied the constraint it was + not active; supplied for speed. + scale: determines the scale with which we modify the gradients, relative + to the existing / unmodified gradients + """ + + def __init__( + self, + num_channels: int, + channel_dim: int, + max_var_per_eig: float = 0.2, + min_prob: float = 0.01, + scale: float = 0.01, + ): + super(MaxEig, self).__init__() + self.num_channels = num_channels + self.channel_dim = channel_dim + self.scale = scale + assert max_var_per_eig == 0.0 or max_var_per_eig > 1.0 / num_channels + self.max_var_per_eig = max_var_per_eig + + # we figure out the dominant direction using the power method: starting with + # a random vector, keep multiplying by the covariance and renormalizing. + with torch.no_grad(): + # arbitrary.. would use randn() but want to leave the rest of the model's + # random parameters unchanged for comparison + direction = torch.arange(num_channels).to(torch.float) + direction = direction / direction.norm() + self.register_buffer("max_eig_direction", direction) + + self.min_prob = min_prob + # cur_prob is the current probability we'll use to apply the ActivationBalancer. + # We'll regress this towards prob, each tiem we try to apply it and it is not + # active. + self.cur_prob = 1.0 + + def forward(self, x: Tensor) -> Tensor: + if ( + torch.jit.is_scripting() + or self.max_var_per_eig <= 0 + or random.random() > self.cur_prob + ): + return _no_op(x) + + with torch.cuda.amp.autocast(enabled=False): + eps = 1.0e-20 + orig_x = x + x = x.to(torch.float32) + with torch.no_grad(): + x = x.transpose(self.channel_dim, -1).reshape(-1, self.num_channels) + x = x - x.mean(dim=0) + new_direction, coeffs = self._find_direction_coeffs( + x, self.max_eig_direction + ) + x_var = (x**2).mean() + x_residual = x - coeffs * new_direction + x_residual_var = (x_residual**2).mean() + + # `variance_proportion` is the proportion of the variance accounted for + # by the top eigen-direction. + variance_proportion = (x_var - x_residual_var) / (x_var + 1.0e-20) + + # ensure new direction is nonzero even if x == 0, by including `direction`. + self._set_direction(0.1 * self.max_eig_direction + new_direction) + + if random.random() < 0.01 or __name__ == "__main__": + logging.info( + f"variance_proportion = {variance_proportion.item()}, shape={tuple(orig_x.shape)}, cur_prob={self.cur_prob}" + ) + + if variance_proportion >= self.max_var_per_eig: + # The constraint is active. Note, we should quite rarely + # reach here, only near the beginning of training if we are + # starting to diverge, should this constraint be active. + cur_prob = self.cur_prob + self.cur_prob = 1.0 # next time, do the update with probability 1.0. + return MaxEigLimiterFunction.apply( + orig_x, coeffs, new_direction, self.channel_dim, self.scale + ) + else: + # let self.cur_prob exponentially approach self.min_prob, as + # long as the constraint is inactive. + self.cur_prob = 0.75 * self.cur_prob + 0.25 * self.min_prob + return orig_x + + def _set_direction(self, direction: Tensor): + """ + Sets self.max_eig_direction to a normalized version of `direction` + """ + direction = direction.detach() + direction = direction / direction.norm() + direction_sum = direction.sum().item() + if direction_sum - direction_sum == 0: # no inf/nan + self.max_eig_direction[:] = direction + else: + logging.info( + f"Warning: sum of direction in MaxEig is {direction_sum}, " + "num_channels={self.num_channels}, channel_dim={self.channel_dim}" + ) + + def _find_direction_coeffs( + self, x: Tensor, prev_direction: Tensor + ) -> Tuple[Tensor, Tensor, Tensor]: + """ + Figure out (an approximation to) the proportion of the variance of a set of + feature vectors that can be attributed to the top eigen-direction. + Args: + x: a Tensor of shape (num_frames, num_channels), with num_frames > 1. + prev_direction: a Tensor of shape (num_channels,), that is our previous estimate + of the top eigen-direction, or a random direction if this is the first + iteration. Does not have to be normalized, but should be nonzero. + + Returns: (cur_direction, coeffs), where: + cur_direction: a Tensor of shape (num_channels,) that is the current + estimate of the top eigen-direction. + coeffs: a Tensor of shape (num_frames, 1) that minimizes, or + approximately minimizes, (x - coeffs * cur_direction).norm() + """ + (num_frames, num_channels) = x.shape + assert num_channels > 1 and num_frames > 1 + assert prev_direction.shape == (num_channels,) + # `coeffs` are the coefficients of `prev_direction` in x. + # actually represent the coeffs up to a constant positive factor. + coeffs = (x * prev_direction).sum(dim=1, keepdim=True) + 1.0e-10 + cur_direction = (x * coeffs).sum(dim=0) / ((coeffs**2).sum() + 1.0e-20) + return cur_direction, coeffs + + +class DoubleSwishFunction(torch.autograd.Function): + """ + double_swish(x) = x * torch.sigmoid(x-1) + This is a definition, originally motivated by its close numerical + similarity to swish(swish(x)), where swish(x) = x * sigmoid(x). + + Memory-efficient derivative computation: + double_swish(x) = x * s, where s(x) = torch.sigmoid(x-1) + double_swish'(x) = d/dx double_swish(x) = x * s'(x) + x' * s(x) = x * s'(x) + s(x). + Now, s'(x) = s(x) * (1-s(x)). + double_swish'(x) = x * s'(x) + s(x). + = x * s(x) * (1-s(x)) + s(x). + = double_swish(x) * (1-s(x)) + s(x) + ... so we just need to remember s(x) but not x itself. + """ + + @staticmethod + def forward(ctx, x: Tensor) -> Tensor: + requires_grad = x.requires_grad + x_dtype = x.dtype + if x.dtype == torch.float16: + x = x.to(torch.float32) + + s = torch.sigmoid(x - 1.0) + y = x * s + + if requires_grad: + deriv = y * (1 - s) + s + # notes on derivative of x * sigmoid(x - 1): + # https://www.wolframalpha.com/input?i=d%2Fdx+%28x+*+sigmoid%28x-1%29%29 + # min \simeq -0.043638. Take floor as -0.043637 so it's a lower bund + # max \simeq 1.1990. Take ceil to be 1.2 so it's an upper bound. + # the combination of "+ torch.rand_like(deriv)" and casting to torch.uint8 (which + # floors), should be expectation-preserving. + floor = -0.043637 + ceil = 1.2 + d_scaled = (deriv - floor) * (255.0 / (ceil - floor)) + torch.rand_like( + deriv + ) + if __name__ == "__main__": + # for self-testing only. + assert d_scaled.min() >= 0.0 + assert d_scaled.max() < 256.0 + d_int = d_scaled.to(torch.uint8) + ctx.save_for_backward(d_int) + if x.dtype == torch.float16 or torch.is_autocast_enabled(): + y = y.to(torch.float16) + return y + + @staticmethod + def backward(ctx, y_grad: Tensor) -> Tensor: + (d,) = ctx.saved_tensors + # the same constants as used in forward pass. + floor = -0.043637 + ceil = 1.2 + d = d * ((ceil - floor) / 255.0) + floor + return y_grad * d + + +class DoubleSwish(torch.nn.Module): + def forward(self, x: Tensor) -> Tensor: + """Return double-swish activation function which is an approximation to Swish(Swish(x)), + that we approximate closely with x * sigmoid(x-1). + """ + if torch.jit.is_scripting(): + return x * torch.sigmoid(x - 1.0) + return DoubleSwishFunction.apply(x) + + +def _test_max_eig(): + for proportion in [0.1, 0.5, 10.0]: + logging.info(f"proportion = {proportion}") + x = torch.randn(100, 128) + direction = torch.randn(128) + coeffs = torch.randn(100, 1) + x += proportion * direction * coeffs + + x.requires_grad = True + + num_channels = 128 + m = MaxEig( + num_channels, 1, 0.5, scale=0.1 # channel_dim # max_var_per_eig + ) # grad_scale + + for _ in range(4): + y = m(x) + + y_grad = torch.randn_like(x) + y.backward(gradient=y_grad) + + if proportion < 0.2: + assert torch.allclose(x.grad, y_grad, atol=1.0e-02) + elif proportion > 1.0: + assert not torch.allclose(x.grad, y_grad) + + +def _test_whiten(): + for proportion in [0.1, 0.5, 10.0]: + logging.info(f"_test_whiten(): proportion = {proportion}") + x = torch.randn(100, 128) + direction = torch.randn(128) + coeffs = torch.randn(100, 1) + x += proportion * direction * coeffs + + x.requires_grad = True + + num_channels = 128 + m = Whiten( + 1, 5.0, prob=1.0, grad_scale=0.1 # num_groups # whitening_limit, + ) # grad_scale + + for _ in range(4): + y = m(x) + + y_grad = torch.randn_like(x) + y.backward(gradient=y_grad) + + if proportion < 0.2: + assert torch.allclose(x.grad, y_grad) + elif proportion > 1.0: + assert not torch.allclose(x.grad, y_grad) + + +def _test_activation_balancer_sign(): + probs = torch.arange(0, 1, 0.01) + N = 1000 + x = 1.0 * ((2.0 * (torch.rand(probs.numel(), N) < probs.unsqueeze(-1))) - 1.0) + x = x.detach() + x.requires_grad = True + m = ActivationBalancer( + probs.numel(), + channel_dim=0, + min_positive=0.05, + max_positive=0.95, + max_factor=0.2, + min_abs=0.0, + ) + + y_grad = torch.sign(torch.randn(probs.numel(), N)) + + y = m(x) + y.backward(gradient=y_grad) + print("_test_activation_balancer_sign: x = ", x) + print("_test_activation_balancer_sign: y grad = ", y_grad) + print("_test_activation_balancer_sign: x grad = ", x.grad) + + +def _test_activation_balancer_magnitude(): + magnitudes = torch.arange(0, 1, 0.01) + N = 1000 + x = torch.sign(torch.randn(magnitudes.numel(), N)) * magnitudes.unsqueeze(-1) + x = x.detach() + x.requires_grad = True + m = ActivationBalancer( + magnitudes.numel(), + channel_dim=0, + min_positive=0.0, + max_positive=1.0, + max_factor=0.2, + min_abs=0.2, + max_abs=0.8, + min_prob=1.0, + ) + + y_grad = torch.sign(torch.randn(magnitudes.numel(), N)) + + y = m(x) + y.backward(gradient=y_grad) + print("_test_activation_balancer_magnitude: x = ", x) + print("_test_activation_balancer_magnitude: y grad = ", y_grad) + print("_test_activation_balancer_magnitude: x grad = ", x.grad) + + +def _test_basic_norm(): + num_channels = 128 + m = BasicNorm(num_channels=num_channels, channel_dim=1) + + x = torch.randn(500, num_channels) + + y = m(x) + + assert y.shape == x.shape + x_rms = (x**2).mean().sqrt() + y_rms = (y**2).mean().sqrt() + print("x rms = ", x_rms) + print("y rms = ", y_rms) + assert y_rms < x_rms + assert y_rms > 0.5 * x_rms + + +def _test_double_swish_deriv(): + x = torch.randn(10, 12, dtype=torch.double) * 3.0 + x.requires_grad = True + m = DoubleSwish() + + tol = (1.2 - (-0.043637)) / 255.0 + torch.autograd.gradcheck(m, x, atol=tol) + + # for self-test. + x = torch.randn(1000, 1000, dtype=torch.double) * 3.0 + x.requires_grad = True + y = m(x) + + +def _test_softmax(): + a = torch.randn(2, 10, dtype=torch.float64) + b = a.clone() + a.requires_grad = True + b.requires_grad = True + a.softmax(dim=1)[:, 0].sum().backward() + print("a grad = ", a.grad) + softmax(b, dim=1)[:, 0].sum().backward() + print("b grad = ", b.grad) + assert torch.allclose(a.grad, b.grad) + + +if __name__ == "__main__": + logging.getLogger().setLevel(logging.INFO) + torch.set_num_threads(1) + torch.set_num_interop_threads(1) + _test_softmax() + _test_whiten() + _test_max_eig() + _test_activation_balancer_sign() + _test_activation_balancer_magnitude() + _test_basic_norm() + _test_double_swish_deriv() diff --git a/egs/tedlium3/ASR/pruned_transducer_stateless_d2v_v2/scaling_converter.py b/egs/tedlium3/ASR/pruned_transducer_stateless_d2v_v2/scaling_converter.py new file mode 100644 index 000000000..56165d1f9 --- /dev/null +++ b/egs/tedlium3/ASR/pruned_transducer_stateless_d2v_v2/scaling_converter.py @@ -0,0 +1,114 @@ +# Copyright 2022 Xiaomi Corp. (authors: Fangjun Kuang) +# +# See ../../../../LICENSE for clarification regarding multiple authors +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +""" +This file replaces various modules in a model. +Specifically, ActivationBalancer is replaced with an identity operator; +Whiten is also replaced with an identity operator; +BasicNorm is replaced by a module with `exp` removed. +""" + +import copy +from typing import List + +import torch +import torch.nn as nn +from scaling import ActivationBalancer, BasicNorm, Whiten + + +class NonScaledNorm(nn.Module): + """See BasicNorm for doc""" + + def __init__( + self, + num_channels: int, + eps_exp: float, + channel_dim: int = -1, # CAUTION: see documentation. + ): + super().__init__() + self.num_channels = num_channels + self.channel_dim = channel_dim + self.eps_exp = eps_exp + + def forward(self, x: torch.Tensor) -> torch.Tensor: + if not torch.jit.is_tracing(): + assert x.shape[self.channel_dim] == self.num_channels + scales = ( + torch.mean(x * x, dim=self.channel_dim, keepdim=True) + self.eps_exp + ).pow(-0.5) + return x * scales + + +def convert_basic_norm(basic_norm: BasicNorm) -> NonScaledNorm: + assert isinstance(basic_norm, BasicNorm), type(BasicNorm) + norm = NonScaledNorm( + num_channels=basic_norm.num_channels, + eps_exp=basic_norm.eps.data.exp().item(), + channel_dim=basic_norm.channel_dim, + ) + return norm + + +# Copied from https://pytorch.org/docs/1.9.0/_modules/torch/nn/modules/module.html#Module.get_submodule # noqa +# get_submodule was added to nn.Module at v1.9.0 +def get_submodule(model, target): + if target == "": + return model + atoms: List[str] = target.split(".") + mod: torch.nn.Module = model + for item in atoms: + if not hasattr(mod, item): + raise AttributeError( + mod._get_name() + " has no " "attribute `" + item + "`" + ) + mod = getattr(mod, item) + if not isinstance(mod, torch.nn.Module): + raise AttributeError("`" + item + "` is not " "an nn.Module") + return mod + + +def convert_scaled_to_non_scaled( + model: nn.Module, + inplace: bool = False, +): + """ + Args: + model: + The model to be converted. + inplace: + If True, the input model is modified inplace. + If False, the input model is copied and we modify the copied version. + Return: + Return a model without scaled layers. + """ + if not inplace: + model = copy.deepcopy(model) + + d = {} + for name, m in model.named_modules(): + if isinstance(m, BasicNorm): + d[name] = convert_basic_norm(m) + elif isinstance(m, (ActivationBalancer, Whiten)): + d[name] = nn.Identity() + + for k, v in d.items(): + if "." in k: + parent, child = k.rsplit(".", maxsplit=1) + setattr(get_submodule(model, parent), child, v) + else: + setattr(model, k, v) + + return model diff --git a/egs/tedlium3/ASR/pruned_transducer_stateless_d2v_v2/test_model.py b/egs/tedlium3/ASR/pruned_transducer_stateless_d2v_v2/test_model.py new file mode 100755 index 000000000..e482d2040 --- /dev/null +++ b/egs/tedlium3/ASR/pruned_transducer_stateless_d2v_v2/test_model.py @@ -0,0 +1,56 @@ +#!/usr/bin/env python3 +# Copyright 2022 Xiaomi Corp. (authors: Fangjun Kuang) +# +# See ../../../../LICENSE for clarification regarding multiple authors +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + + +""" +To run this file, do: + + cd icefall/egs/librispeech/ASR + python ./pruned_transducer_stateless7_ctc/test_model.py +""" + +from train import get_params, get_transducer_model + + +def test_model_1(): + params = get_params() + params.vocab_size = 500 + params.blank_id = 0 + params.context_size = 2 + params.num_encoder_layers = "2,4,3,2,4" + # params.feedforward_dims = "1024,1024,1536,1536,1024" + params.feedforward_dims = "1024,1024,2048,2048,1024" + params.nhead = "8,8,8,8,8" + params.encoder_dims = "384,384,384,384,384" + params.attention_dims = "192,192,192,192,192" + params.encoder_unmasked_dims = "256,256,256,256,256" + params.zipformer_downsampling_factors = "1,2,4,8,2" + params.cnn_module_kernels = "31,31,31,31,31" + params.decoder_dim = 512 + params.joiner_dim = 512 + model = get_transducer_model(params) + + num_param = sum([p.numel() for p in model.parameters()]) + print(f"Number of model parameters: {num_param}") + + +def main(): + test_model_1() + + +if __name__ == "__main__": + main() diff --git a/egs/tedlium3/ASR/pruned_transducer_stateless_d2v_v2/train.py b/egs/tedlium3/ASR/pruned_transducer_stateless_d2v_v2/train.py new file mode 100755 index 000000000..f79b08d7a --- /dev/null +++ b/egs/tedlium3/ASR/pruned_transducer_stateless_d2v_v2/train.py @@ -0,0 +1,1719 @@ +#!/usr/bin/env python3 +# Copyright 2021-2022 Xiaomi Corp. (authors: Fangjun Kuang, +# Wei Kang, +# Mingshuang Luo,) +# Zengwei Yao) +# +# See ../../../../LICENSE for clarification regarding multiple authors +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +""" +Usage: + +export CUDA_VISIBLE_DEVICES="0,1,2,3" + +./pruned_transducer_stateless7_ctc/train.py \ + --world-size 4 \ + --num-epochs 30 \ + --start-epoch 1 \ + --exp-dir pruned_transducer_stateless7_ctc/exp \ + --full-libri 1 \ + --max-duration 300 + +# For mix precision training: + +./pruned_transducer_stateless7_ctc/train.py \ + --world-size 4 \ + --num-epochs 30 \ + --start-epoch 1 \ + --use-fp16 1 \ + --exp-dir pruned_transducer_stateless7_ctc/exp \ + --full-libri 1 \ + --max-duration 550 + +# For d2v-T training: +export CUDA_VISIBLE_DEVICES="0,1,2,3,4,5,6,7" + +./pruned_transducer_stateless_d2v_v2/train.py \ + --wandb true \ + --input-strategy AudioSamples \ + --enable-spec-aug False \ + --multi-optim True \ + --world-size 8 \ + --num-epochs 30 \ + --start-epoch 1 \ + --full-libri 0 \ + --exp-dir ./pruned_transducer_stateless_d2v_v2/$1 \ + --max-duration 250 \ + --freeze-finetune-updates 2000 \ + --use-fp16 1 \ + --peak-enc-lr 0.001 \ + --peak-dec-lr 0.05 \ + --accum-grads 1 \ + --encoder-type d2v \ + --additional-block True \ + --encoder-dim 768 \ + --decoder-dim 768 \ + --joiner-dim 768 \ + --prune-range 20 \ + --context-size 2 \ + --ctc-loss-scale 0.2 + +""" + + +import random +import argparse +import copy +import logging +import warnings +from pathlib import Path +from shutil import copyfile +from typing import Any, Dict, Optional, Tuple, Union + +import k2 +import optim +import sentencepiece as spm +import torch +import torch.multiprocessing as mp +import torch.nn as nn +from asr_datamodule import LibriSpeechAsrDataModule +from decoder import Decoder +from joiner import Joiner +from lhotse.cut import Cut +from lhotse.dataset.sampling.base import CutSampler +from lhotse.utils import fix_random_seed +from model import Transducer +from optim import Eden, ScaledAdam +from torch import Tensor +from torch.cuda.amp import GradScaler +from torch.nn.parallel import DistributedDataParallel as DDP +from torch.utils.tensorboard import SummaryWriter +from zipformer import Zipformer +from data2vec_encoder import FairSeqData2VecEncoder + +from icefall import diagnostics +from icefall.checkpoint import remove_checkpoints +from icefall.checkpoint import update_averaged_model +from checkpoint import ( + save_checkpoint as save_checkpoint_impl, + save_checkpoint_with_global_batch_idx, + load_checkpoint +) +from icefall.dist import cleanup_dist, setup_dist +from icefall.env import get_env_info +from icefall.hooks import register_inf_check_hooks +from icefall.utils import ( + AttributeDict, + MetricsTracker, + encode_supervisions, + setup_logger, + str2bool, + save_args, +) + +import wandb + +#from icefall.checkpoint import save_checkpoint as save_checkpoint_impl +LRSchedulerType = Union[torch.optim.lr_scheduler._LRScheduler, optim.LRScheduler] + + +def set_batch_count(model: Union[nn.Module, DDP], batch_count: float) -> None: + if isinstance(model, DDP): + # get underlying nn.Module + model = model.module + for module in model.modules(): + if hasattr(module, "batch_count"): + module.batch_count = batch_count + model.encoder.num_updates = int(batch_count) + + +def add_adapter_arguments(parser: argparse.ArgumentParser): + parser.add_argument( + "--add-adapter", + type=str2bool, + default=False, + help="add adapter to rep model's encoder" + ) + + parser.add_argument( + "--adapter-lr", + type=float, + default=0.0001, + help="adapter learning rate" + ) + + +def add_rep_arguments(parser: argparse.ArgumentParser): + parser.add_argument( + "--wandb", + type=str2bool, + default=True, + help="Use wandb for MLOps", + ) + parser.add_argument( + "--accum-grads", + type=int, + default=1, + help="accum-grad num.", + ) + + parser.add_argument( + "--multi-optim", + type=str2bool, + default=True, + help="use sperate optimizer (enc / dec)", + ) + + parser.add_argument( + "--peak-enc-lr", + type=float, + default=0.0001, + help="The initial learning rate. This value should not need to be changed.", + ) + + parser.add_argument( + "--peak-dec-lr", + type=float, + default=0.001, + help="The initial learning rate. This value should not need to be changed.", + ) + + parser.add_argument( + "--encoder-type", + type=str, + default='d2v', + help="Type of encoder (e.g. conformer, w2v, d2v...", + ) + + parser.add_argument( + "--encoder-dim", + type=int, + default=768, + help="encoder embedding dimension", + ) + + parser.add_argument( + "--freeze-finetune-updates", + type=int, + default=0 + ) + + parser.add_argument( + "--additional-block", + type=str2bool, + default=True, + ) + + parser.add_argument( + "--decode-interval", + type=int, + default=200, + help="decode interval", + ) + + +def add_model_arguments(parser: argparse.ArgumentParser): + parser.add_argument( + "--num-encoder-layers", + type=str, + default="2,4,3,2,4", + help="Number of zipformer encoder layers, comma separated.", + ) + + parser.add_argument( + "--feedforward-dims", + type=str, + default="1024,1024,2048,2048,1024", + help="Feedforward dimension of the zipformer encoder layers, comma separated.", + ) + + parser.add_argument( + "--nhead", + type=str, + default="8,8,8,8,8", + help="Number of attention heads in the zipformer encoder layers.", + ) + + parser.add_argument( + "--encoder-dims", + type=str, + default="384,384,384,384,384", + help="Embedding dimension in the 2 blocks of zipformer encoder layers, comma separated", + ) + + parser.add_argument( + "--attention-dims", + type=str, + default="192,192,192,192,192", + help="""Attention dimension in the 2 blocks of zipformer encoder layers, comma separated; + not the same as embedding dimension.""", + ) + + parser.add_argument( + "--encoder-unmasked-dims", + type=str, + default="256,256,256,256,256", + help="Unmasked dimensions in the encoders, relates to augmentation during training. " + "Must be <= each of encoder_dims. Empirically, less than 256 seems to make performance " + " worse.", + ) + + parser.add_argument( + "--zipformer-downsampling-factors", + type=str, + default="1,2,4,8,2", + help="Downsampling factor for each stack of encoder layers.", + ) + + parser.add_argument( + "--cnn-module-kernels", + type=str, + default="31,31,31,31,31", + help="Sizes of kernels in convolution modules", + ) + + parser.add_argument( + "--decoder-dim", + type=int, + default=512, + help="Embedding dimension in the decoder model.", + ) + + parser.add_argument( + "--joiner-dim", + type=int, + default=512, + help="""Dimension used in the joiner model. + Outputs from the encoder and decoder model are projected + to this dimension before adding. + """, + ) + + +def get_parser(): + parser = argparse.ArgumentParser( + formatter_class=argparse.ArgumentDefaultsHelpFormatter + ) + + parser.add_argument( + "--world-size", + type=int, + default=1, + help="Number of GPUs for DDP training.", + ) + + parser.add_argument( + "--master-port", + type=int, + default=12354, + help="Master port to use for DDP training.", + ) + + parser.add_argument( + "--tensorboard", + type=str2bool, + default=True, + help="Should various information be logged in tensorboard.", + ) + + parser.add_argument( + "--num-epochs", + type=int, + default=30, + help="Number of epochs to train.", + ) + + parser.add_argument( + "--start-epoch", + type=int, + default=1, + help="""Resume training from this epoch. It should be positive. + If larger than 1, it will load checkpoint from + exp-dir/epoch-{start_epoch-1}.pt + """, + ) + + parser.add_argument( + "--start-batch", + type=int, + default=0, + help="""If positive, --start-epoch is ignored and + it loads the checkpoint from exp-dir/checkpoint-{start_batch}.pt + """, + ) + + parser.add_argument( + "--exp-dir", + type=str, + default="pruned_transducer_stateless7_ctc/exp", + help="""The experiment dir. + It specifies the directory where all training related + files, e.g., checkpoints, log, etc, are saved + """, + ) + + parser.add_argument( + "--bpe-model", + type=str, + default="data/lang_bpe_500/bpe.model", + help="Path to the BPE model", + ) + + parser.add_argument( + "--base-lr", type=float, default=0.05, help="The base learning rate." + ) + + parser.add_argument( + "--lr-batches", + type=float, + default=5000, + help="""Number of steps that affects how rapidly the learning rate + decreases. We suggest not to change this.""", + ) + + parser.add_argument( + "--lr-epochs", + type=float, + default=3.5, + help="""Number of epochs that affects how rapidly the learning rate decreases. + """, + ) + + parser.add_argument( + "--context-size", + type=int, + default=2, + help="The context size in the decoder. 1 means bigram; 2 means tri-gram", + ) + + parser.add_argument( + "--prune-range", + type=int, + default=5, + help="The prune range for rnnt loss, it means how many symbols(context)" + "we are using to compute the loss", + ) + + parser.add_argument( + "--lm-scale", + type=float, + default=0.25, + help="The scale to smooth the loss with lm " + "(output of prediction network) part.", + ) + + parser.add_argument( + "--am-scale", + type=float, + default=0.0, + help="The scale to smooth the loss with am (output of encoder network) part.", + ) + + parser.add_argument( + "--simple-loss-scale", + type=float, + default=0.5, + help="To get pruning ranges, we will calculate a simple version" + "loss(joiner is just addition), this simple loss also uses for" + "training (as a regularization item). We will scale the simple loss" + "with this parameter before adding to the final loss.", + ) + + parser.add_argument( + "--ctc-loss-scale", + type=float, + default=0.2, + help="Scale for CTC loss.", + ) + + parser.add_argument( + "--seed", + type=int, + default=42, + help="The seed for random generators intended for reproducibility", + ) + + parser.add_argument( + "--print-diagnostics", + type=str2bool, + default=False, + help="Accumulate stats on activations, print them and exit.", + ) + + parser.add_argument( + "--inf-check", + type=str2bool, + default=False, + help="Add hooks to check for infinite module outputs and gradients.", + ) + + parser.add_argument( + "--save-every-n", + type=int, + default=2000, + help="""Save checkpoint after processing this number of batches" + periodically. We save checkpoint to exp-dir/ whenever + params.batch_idx_train % save_every_n == 0. The checkpoint filename + has the form: f'exp-dir/checkpoint-{params.batch_idx_train}.pt' + Note: It also saves checkpoint to `exp-dir/epoch-xxx.pt` at the + end of each epoch where `xxx` is the epoch number counting from 0. + """, + ) + + parser.add_argument( + "--keep-last-k", + type=int, + default=30, + help="""Only keep this number of checkpoints on disk. + For instance, if it is 3, there are only 3 checkpoints + in the exp-dir with filenames `checkpoint-xxx.pt`. + It does not affect checkpoints with name `epoch-xxx.pt`. + """, + ) + + parser.add_argument( + "--average-period", + type=int, + default=200, + help="""Update the averaged model, namely `model_avg`, after processing + this number of batches. `model_avg` is a separate version of model, + in which each floating-point parameter is the average of all the + parameters from the start of training. Each time we take the average, + we do: `model_avg = model * (average_period / batch_idx_train) + + model_avg * ((batch_idx_train - average_period) / batch_idx_train)`. + """, + ) + + parser.add_argument( + "--use-fp16", + type=str2bool, + default=True, + help="Whether to use half precision training.", + ) + + add_model_arguments(parser) + add_rep_arguments(parser) + add_adapter_arguments(parser) + + return parser + + +def get_params() -> AttributeDict: + """Return a dict containing training parameters. + + All training related parameters that are not passed from the commandline + are saved in the variable `params`. + + Commandline options are merged into `params` after they are parsed, so + you can also access them via `params`. + + Explanation of options saved in `params`: + + - best_train_loss: Best training loss so far. It is used to select + the model that has the lowest training loss. It is + updated during the training. + + - best_valid_loss: Best validation loss so far. It is used to select + the model that has the lowest validation loss. It is + updated during the training. + + - best_train_epoch: It is the epoch that has the best training loss. + + - best_valid_epoch: It is the epoch that has the best validation loss. + + - batch_idx_train: Used to writing statistics to tensorboard. It + contains number of batches trained so far across + epochs. + + - log_interval: Print training loss if batch_idx % log_interval` is 0 + + - reset_interval: Reset statistics if batch_idx % reset_interval is 0 + + - valid_interval: Run validation if batch_idx % valid_interval is 0 + + - feature_dim: The model input dim. It has to match the one used + in computing features. + + - subsampling_factor: The subsampling factor for the model. + + - encoder_dim: Hidden dim for multi-head attention model. + + - num_decoder_layers: Number of decoder layer of transformer decoder. + + - warm_step: The warmup period that dictates the decay of the + scale on "simple" (un-pruned) loss. + """ + params = AttributeDict( + { + "best_train_loss": float("inf"), + "best_valid_loss": float("inf"), + "best_train_epoch": -1, + "best_valid_epoch": -1, + "batch_idx_train": 0, + "log_interval": 50, + "reset_interval": 200, + "valid_interval": 3000, # For the 100h subset, use 800 + # parameters for zipformer + "feature_dim": 80, + "subsampling_factor": 320, # not passed in, this is fixed. + # parameters for ctc loss + "beam_size": 10, + "use_double_scores": True, + "warm_step": 4000, + #"warm_step": 3000, + "env_info": get_env_info(), + } + ) + + return params + + +def get_encoder_model(params: AttributeDict) -> nn.Module: + # TODO: We can add an option to switch between Zipformer and Transformer + def to_int_tuple(s: str): + return tuple(map(int, s.split(","))) + + if params.encoder_type == 'd2v': + encoder = FairSeqData2VecEncoder( + input_size=params.encoder_dim, + w2v_url='None', + output_size=params.encoder_dim, + freeze_finetune_updates=params.freeze_finetune_updates, + additional_block=params.additional_block, + ) + else: + encoder = Zipformer( + num_features=params.feature_dim, + output_downsampling_factor=2, + zipformer_downsampling_factors=to_int_tuple( + params.zipformer_downsampling_factors + ), + encoder_dims=to_int_tuple(params.encoder_dims), + attention_dim=to_int_tuple(params.attention_dims), + encoder_unmasked_dims=to_int_tuple(params.encoder_unmasked_dims), + nhead=to_int_tuple(params.nhead), + feedforward_dim=to_int_tuple(params.feedforward_dims), + cnn_module_kernels=to_int_tuple(params.cnn_module_kernels), + num_encoder_layers=to_int_tuple(params.num_encoder_layers), + ) + + return encoder + + +def get_decoder_model(params: AttributeDict) -> nn.Module: + decoder = Decoder( + vocab_size=params.vocab_size, + decoder_dim=params.decoder_dim, + blank_id=params.blank_id, + context_size=params.context_size, + ) + return decoder + + +def get_joiner_model(params: AttributeDict) -> nn.Module: + joiner = Joiner( + encoder_dim=params.encoder_dim if params.encoder_type == 'd2v' else int(params.encoder_dims.split(",")[-1]), + decoder_dim=params.decoder_dim, + joiner_dim=params.joiner_dim, + vocab_size=params.vocab_size, + ) + return joiner + + +def get_transducer_model(params: AttributeDict) -> nn.Module: + encoder = get_encoder_model(params) + decoder = get_decoder_model(params) + joiner = get_joiner_model(params) + + model = Transducer( + encoder=encoder, + decoder=decoder, + joiner=joiner, + encoder_dim=params.encoder_dim if params.encoder_type == 'd2v' else int(params.encoder_dims.split(",")[-1]), + decoder_dim=params.decoder_dim, + joiner_dim=params.joiner_dim, + vocab_size=params.vocab_size, + ) + return model + + +def load_checkpoint_if_available( + params: AttributeDict, + model: nn.Module, + model_avg: nn.Module = None, + optimizer: Optional[torch.optim.Optimizer] = None, + scheduler: Optional[LRSchedulerType] = None, +) -> Optional[Dict[str, Any]]: + """Load checkpoint from file. + + If params.start_batch is positive, it will load the checkpoint from + `params.exp_dir/checkpoint-{params.start_batch}.pt`. Otherwise, if + params.start_epoch is larger than 1, it will load the checkpoint from + `params.start_epoch - 1`. + + Apart from loading state dict for `model` and `optimizer` it also updates + `best_train_epoch`, `best_train_loss`, `best_valid_epoch`, + and `best_valid_loss` in `params`. + + Args: + params: + The return value of :func:`get_params`. + model: + The training model. + model_avg: + The stored model averaged from the start of training. + optimizer: + The optimizer that we are using. + scheduler: + The scheduler that we are using. + Returns: + Return a dict containing previously saved training info. + """ + if params.start_batch > 0: + filename = params.exp_dir / f"checkpoint-{params.start_batch}.pt" + elif params.start_epoch > 1: + filename = params.exp_dir / f"epoch-{params.start_epoch-1}.pt" + elif params.add_adapter: + filename = params.exp_dir / f"d2v-base-T.pt" + else: + return None + + assert filename.is_file(), f"{filename} does not exist!" + + saved_params = load_checkpoint( + filename, + model=model, + model_avg=model_avg, + optimizer=optimizer, + scheduler=scheduler, + strict=True if not params.add_adapter else False, + ) + + keys = [ + "best_train_epoch", + "best_valid_epoch", + "batch_idx_train", + "best_train_loss", + "best_valid_loss", + ] + for k in keys: + params[k] = saved_params[k] + + if params.start_batch > 0: + if "cur_epoch" in saved_params: + params["start_epoch"] = saved_params["cur_epoch"] + + if "cur_batch_idx" in saved_params: + params["cur_batch_idx"] = saved_params["cur_batch_idx"] + + return saved_params + + +def save_checkpoint( + params: AttributeDict, + model: Union[nn.Module, DDP], + model_avg: Optional[nn.Module] = None, + optimizer: Optional[torch.optim.Optimizer] = None, + scheduler: Optional[LRSchedulerType] = None, + sampler: Optional[CutSampler] = None, + scaler: Optional[GradScaler] = None, + rank: int = 0, +) -> None: + """Save model, optimizer, scheduler and training stats to file. + + Args: + params: + It is returned by :func:`get_params`. + model: + The training model. + model_avg: + The stored model averaged from the start of training. + optimizer: + The optimizer used in the training. + sampler: + The sampler for the training dataset. + scaler: + The scaler used for mix precision training. + """ + if rank != 0: + return + filename = params.exp_dir / f"epoch-{params.cur_epoch}.pt" + save_checkpoint_impl( + filename=filename, + model=model, + model_avg=model_avg, + params=params, + optimizer=optimizer, + scheduler=scheduler, + sampler=sampler, + scaler=scaler, + rank=rank, + ) + + if params.best_train_epoch == params.cur_epoch: + best_train_filename = params.exp_dir / "best-train-loss.pt" + copyfile(src=filename, dst=best_train_filename) + + if params.best_valid_epoch == params.cur_epoch: + best_valid_filename = params.exp_dir / "best-valid-loss.pt" + copyfile(src=filename, dst=best_valid_filename) + + +def compute_loss( + params: AttributeDict, + model: Union[nn.Module, DDP], + sp: spm.SentencePieceProcessor, + batch: dict, + is_training: bool, + decode: bool = False, +) -> Tuple[Tensor, MetricsTracker]: + """ + Compute transducer loss given the model and its inputs. + + Args: + params: + Parameters for training. See :func:`get_params`. + model: + The model for training. It is an instance of Zipformer in our case. + batch: + A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()` + for the content in it. + is_training: + True for training. False for validation. When it is True, this + function enables autograd during computation; when it is False, it + disables autograd. + warmup: a floating point value which increases throughout training; + values >= 1.0 are fully warmed up and have all modules present. + """ + device = model.device if isinstance(model, DDP) else next(model.parameters()).device + feature = batch["inputs"] + # at entry, feature is (N, T, C) + assert feature.ndim == 2 or feature.ndim == 3 + feature = feature.to(device) + + supervisions = batch["supervisions"] + + if feature.ndim == 2: + feature_lens = [] + for supervision in supervisions['cut']: + try: feature_lens.append(supervision.tracks[0].cut.recording.num_samples) + except: feature_lens.append(supervision.recording.num_samples) + feature_lens = torch.tensor(feature_lens) + + elif feature.ndim == 3: + feature_lens = supervisions["num_frames"].to(device) + + batch_idx_train = params.batch_idx_train + warm_step = params.warm_step + + texts = batch["supervisions"]["text"] + token_ids = sp.encode(texts, out_type=int) + y = k2.RaggedTensor(token_ids).to(device) + + with torch.set_grad_enabled(is_training): + simple_loss, pruned_loss, ctc_output = model( + x=feature, + x_lens=feature_lens, + y=y, + prune_range=params.prune_range, + am_scale=params.am_scale, + lm_scale=params.lm_scale, + ) + + s = params.simple_loss_scale + # take down the scale on the simple loss from 1.0 at the start + # to params.simple_loss scale by warm_step. + simple_loss_scale = ( + s + if batch_idx_train >= warm_step + else 1.0 - (batch_idx_train / warm_step) * (1.0 - s) + ) + pruned_loss_scale = ( + 1.0 + if batch_idx_train >= warm_step + else 0.1 + 0.9 * (batch_idx_train / warm_step) + ) + + loss = simple_loss_scale * simple_loss + pruned_loss_scale * pruned_loss + + info = MetricsTracker() + + if params.ctc_loss_scale > 0: + # Compute ctc loss + + # NOTE: We need `encode_supervisions` to sort sequences with + # different duration in decreasing order, required by + # `k2.intersect_dense` called in `k2.ctc_loss` + with warnings.catch_warnings(): + warnings.simplefilter("ignore") + supervision_segments, token_ids = encode_supervisions( + supervisions, + subsampling_factor=params.subsampling_factor, + token_ids=token_ids, + ) + + # Works with a BPE model + decoding_graph = k2.ctc_graph(token_ids, modified=False, device=device) + dense_fsa_vec = k2.DenseFsaVec( + ctc_output, + supervision_segments, + allow_truncate=params.subsampling_factor - 1, + ) + + ctc_loss = k2.ctc_loss( + decoding_graph=decoding_graph, + dense_fsa_vec=dense_fsa_vec, + output_beam=params.beam_size, + reduction="sum", + use_double_scores=params.use_double_scores, + ) + assert ctc_loss.requires_grad == is_training + loss += params.ctc_loss_scale * ctc_loss + + info["ctc_loss"] = ctc_loss.detach().cpu().item() + + assert loss.requires_grad == is_training + + if decode: + model.eval() + with torch.no_grad(): + hypos = model.module.decode( + x=feature, + x_lens=feature_lens, + y=y, + sp=sp + ) + logging.info(f'ref: {batch["supervisions"]["text"][0]}') + logging.info(f'hyp: {" ".join(hypos[0])}') + model.train() + + with warnings.catch_warnings(): + warnings.simplefilter("ignore") + info["frames"] = (feature_lens // params.subsampling_factor).sum().item() + + # Note: We use reduction=sum while computing the loss. + info["utterances"] = feature.size(0) + info["loss"] = loss.detach().cpu().item() + info["simple_loss"] = simple_loss.detach().cpu().item() + info["pruned_loss"] = pruned_loss.detach().cpu().item() + + return loss, info + + +def compute_validation_loss( + params: AttributeDict, + model: Union[nn.Module, DDP], + sp: spm.SentencePieceProcessor, + valid_dl: torch.utils.data.DataLoader, + world_size: int = 1, +) -> MetricsTracker: + """Run the validation process.""" + model.eval() + + tot_loss = MetricsTracker() + + for batch_idx, batch in enumerate(valid_dl): + loss, loss_info = compute_loss( + params=params, + model=model, + sp=sp, + batch=batch, + is_training=False, + ) + assert loss.requires_grad is False + tot_loss = tot_loss + loss_info + + if world_size > 1: + tot_loss.reduce(loss.device) + + loss_value = tot_loss["loss"] / tot_loss["utterances"] + if loss_value < params.best_valid_loss: + params.best_valid_epoch = params.cur_epoch + params.best_valid_loss = loss_value + + return tot_loss + + +def train_one_epoch( + params: AttributeDict, + model: Union[nn.Module, DDP], + optimizer: torch.optim.Optimizer or [torch.optim.Optimizer, torch.optim.Optimizer], + scheduler: LRSchedulerType or [LRSchedulerType, LRSchedulerType], + sp: spm.SentencePieceProcessor, + train_dl: torch.utils.data.DataLoader, + valid_dl: torch.utils.data.DataLoader, + scaler: GradScaler, + model_avg: Optional[nn.Module] = None, + tb_writer: Optional[SummaryWriter] = None, + world_size: int = 1, + rank: int = 0, + wb = None, +) -> None: + """Train the model for one epoch. + + The training loss from the mean of all frames is saved in + `params.train_loss`. It runs the validation process every + `params.valid_interval` batches. + + Args: + params: + It is returned by :func:`get_params`. + model: + The model for training. + optimizer: + The optimizer we are using. + scheduler: + The learning rate scheduler, we call step() every step. + train_dl: + Dataloader for the training dataset. + valid_dl: + Dataloader for the validation dataset. + scaler: + The scaler used for mix precision training. + model_avg: + The stored model averaged from the start of training. + tb_writer: + Writer to write log messages to tensorboard. + world_size: + Number of nodes in DDP training. If it is 1, DDP is disabled. + rank: + The rank of the node in DDP training. If no DDP is used, it should + be set to 0. + """ + model.train() + + tot_loss = MetricsTracker() + + cur_batch_idx = params.get("cur_batch_idx", 0) + + if params.multi_optim: + optimizer_enc, optimizer_dec = optimizer[0], optimizer[1] + scheduler_enc, scheduler_dec = scheduler[0], scheduler[1] + + for batch_idx, batch in enumerate(train_dl): + if batch_idx < cur_batch_idx: + continue + cur_batch_idx = batch_idx + + if batch_idx % params.accum_grads == 0: params.batch_idx_train += 1 + batch_size = len(batch["supervisions"]["text"]) + + try: + with torch.cuda.amp.autocast(enabled=params.use_fp16): + loss, loss_info = compute_loss( + params=params, + model=model, + sp=sp, + batch=batch, + is_training=True, + decode = True if batch_idx % params.decode_interval == 0 else False, + ) + loss_info.reduce(loss.device) + + numel = params.world_size / (params.accum_grads * loss_info["utterances"]) + loss *= numel ## normalize loss over utts(batch size) + + # summary stats + tot_loss = (tot_loss * (1 - 1 / params.reset_interval)) + loss_info + + # NOTE: We use reduction==sum and loss is computed over utterances + # in the batch and there is no normalization to it so far. + scaler.scale(loss).backward() + + if params.multi_optim and (batch_idx+1) % params.accum_grads == 0: + set_batch_count(model, params.batch_idx_train) + scheduler_enc.step_batch(params.batch_idx_train) + scheduler_dec.step_batch(params.batch_idx_train) + scaler.step(optimizer_enc) + scaler.step(optimizer_dec) + scaler.update() + optimizer_enc.zero_grad() + optimizer_dec.zero_grad() + elif not params.multi_optim and (batch_idx+1) % params.accum_grads == 0: + set_batch_count(model, params.batch_idx_train) + scheduler.step_batch(params.batch_idx_train) + scaler.step(optimizer) + scaler.update() + optimizer.zero_grad() + + except: # noqa + display_and_save_batch(batch, params=params, sp=sp) + raise + + if params.print_diagnostics and batch_idx == 5: + return + + if ( + rank == 0 + and params.batch_idx_train > 0 + and params.batch_idx_train % params.average_period == 0 + ): + update_averaged_model( + params=params, + model_cur=model, + model_avg=model_avg, + ) + + if ( + params.batch_idx_train > 0 + and params.batch_idx_train % params.save_every_n == 0 + ): + params.cur_batch_idx = batch_idx + save_checkpoint_with_global_batch_idx( + out_dir=params.exp_dir, + global_batch_idx=params.batch_idx_train, + model=model, + model_avg=model_avg, + params=params, + optimizer=optimizer, + scheduler=scheduler, + sampler=train_dl.sampler, + scaler=scaler, + rank=rank, + ) + del params.cur_batch_idx + remove_checkpoints( + out_dir=params.exp_dir, + topk=params.keep_last_k, + rank=rank, + ) + + if batch_idx % 100 == 0 and params.use_fp16: + # If the grad scale was less than 1, try increasing it. The _growth_interval + # of the grad scaler is configurable, but we can't configure it to have different + # behavior depending on the current grad scale. + cur_grad_scale = scaler._scale.item() + ''' + if cur_grad_scale < 1.0 or (cur_grad_scale < 8.0 and batch_idx % 400 == 0): + scaler.update(cur_grad_scale * 2.0) + ''' + if cur_grad_scale < 0.01: + logging.warning(f"Grad scale is small: {cur_grad_scale}") + if cur_grad_scale < 1.0e-05: + wb.log({"valid/loss": 10000}) + raise RuntimeError( + f"grad_scale is too small, exiting: {cur_grad_scale}" + ) + + if params.batch_idx_train > 4000 and loss > 300 and params.wandb: + wb.log({"valid/loss": 10000}) + raise RunteimError( + f"divergence... exiting: loss={loss}" + ) + + if batch_idx % (params.log_interval*params.accum_grads) == 0: + if params.multi_optim: + cur_enc_lr = scheduler_enc.get_last_lr()[0] + cur_dec_lr = scheduler_dec.get_last_lr()[0] + cur_grad_scale = scaler._scale.item() if params.use_fp16 else 1.0 + + logging.info( + f"Epoch {params.cur_epoch}, " + f"batch {batch_idx}, loss[{loss_info}], " + f"tot_loss[{tot_loss}], batch size: {batch_size}, " + f"enc_lr: {cur_enc_lr:.2e}, " + f"dec_lr: {cur_dec_lr:.2e}, " + + (f"grad_scale: {scaler._scale.item()}" if params.use_fp16 else "") + ) + + else: + cur_lr = scheduler.get_last_lr()[0] + cur_grad_scale = scaler._scale.item() if params.use_fp16 else 1.0 + + logging.info( + f"Epoch {params.cur_epoch}, " + f"batch {batch_idx}, loss[{loss_info}], " + f"tot_loss[{tot_loss}], batch size: {batch_size}, " + f"lr: {cur_lr:.2e}, " + + (f"grad_scale: {scaler._scale.item()}" if params.use_fp16 else "") + ) + + if tb_writer is not None: + if params.multi_optim: + tb_writer.add_scalar( + "train/enc_learning_rate", cur_enc_lr, params.batch_idx_train + ) + tb_writer.add_scalar( + "train/dec_learning_rate", cur_dec_lr, params.batch_idx_train + ) + + else: + tb_writer.add_scalar( + "train/learning_rate", cur_lr, params.batch_idx_train + ) + + loss_info.write_summary( + tb_writer, "train/current_", params.batch_idx_train + ) + tot_loss.write_summary(tb_writer, "train/tot_", params.batch_idx_train) + if params.use_fp16: + tb_writer.add_scalar( + "train/grad_scale", + cur_grad_scale, + params.batch_idx_train, + ) + + if wb is not None and rank == 0: + wb.log({"train/loss": loss_info["loss"]*numel}) + wb.log({"train/simple_loss": loss_info["simple_loss"]*numel}) + wb.log({"train/pruned_loss": loss_info["pruned_loss"]*numel}) + wb.log({"train/ctc_loss": loss_info["ctc_loss"]*numel}) + + logging.info("Computing validation loss") + valid_info = compute_validation_loss( + params=params, + model=model, + sp=sp, + valid_dl=valid_dl, + world_size=world_size, + ) + model.train() + logging.info(f"Epoch {params.cur_epoch}, validation: {valid_info}") + logging.info( + f"Maximum memory allocated so far is {torch.cuda.max_memory_allocated()//1000000}MB" + ) + if tb_writer is not None: + valid_info.write_summary( + tb_writer, "train/valid_", params.batch_idx_train + ) + + if wb is not None and rank == 0: + numel = 1 / (params.accum_grads * valid_info["utterances"]) + wb.log({"valid/loss": valid_info["loss"]*numel}) + wb.log({"valid/simple_loss": valid_info["simple_loss"]*numel}) + wb.log({"valid/pruned_loss": valid_info["pruned_loss"]*numel}) + wb.log({"valid/ctc_loss": valid_info["ctc_loss"]*numel}) + + loss_value = tot_loss["loss"] / tot_loss["utterances"] + params.train_loss = loss_value + if params.train_loss < params.best_train_loss: + params.best_train_epoch = params.cur_epoch + params.best_train_loss = params.train_loss + + +def run(rank, world_size, args, wb=None): + """ + Args: + rank: + It is a value between 0 and `world_size-1`, which is + passed automatically by `mp.spawn()` in :func:`main`. + The node with rank 0 is responsible for saving checkpoint. + world_size: + Number of GPUs for DDP training. + args: + The return value of get_parser().parse_args() + """ + params = get_params() + params.update(vars(args)) + #params.warm_step *= params.accum_grads + + fix_random_seed(params.seed) + if world_size > 1: + setup_dist(rank, world_size, params.master_port) + + setup_logger(f"{params.exp_dir}/log/log-train") + logging.info("Training started") + + if args.tensorboard and rank == 0: + tb_writer = SummaryWriter(log_dir=f"{params.exp_dir}/tensorboard") + else: + tb_writer = None + + device = torch.device("cpu") + if torch.cuda.is_available(): + device = torch.device("cuda", rank) + logging.info(f"Device: {device}") + + sp = spm.SentencePieceProcessor() + sp.load(params.bpe_model) + + # is defined in local/train_bpe_model.py + params.blank_id = sp.piece_to_id("") + params.vocab_size = sp.get_piece_size() + + logging.info(params) + + logging.info("About to create model") + model = get_transducer_model(params) + logging.info(model) + exit() + + num_param = sum([p.numel() for p in model.parameters()]) + logging.info(f"Number of model parameters: {num_param}") + + assert params.save_every_n >= params.average_period + model_avg: Optional[nn.Module] = None + if rank == 0: + # model_avg is only used with rank 0 + model_avg = copy.deepcopy(model).to(torch.float64) + + assert params.start_epoch > 0, params.start_epoch + checkpoints = load_checkpoint_if_available( + params=params, model=model, model_avg=model_avg + ) + + model.to(device) + if world_size > 1: + logging.info("Using DDP") + model = DDP(model, device_ids=[rank], find_unused_parameters=True) + + if params.multi_optim: + logging.info("Using seperate optimizers over encoder, decoder ...") + + enc_param = [] + enc_names = [] + + dec_names = [] + dec_param = [] + + for n, p in model.named_parameters(): + name = n.split('.')[1] + if name == 'encoder' and 'feature_extractor' not in n: + enc_names.append(n) + enc_param.append(p) + elif 'ctc_output' in n: + enc_names.append(n) + enc_param.append(p) + elif 'feature_extractor' not in n: + dec_names.append(n) + dec_param.append(p) + + optimizer_enc = ScaledAdam( + enc_param, + lr=params.peak_enc_lr, + clipping_scale=None, + parameters_names=[enc_names], + ) + optimizer_dec = ScaledAdam( + dec_param, + lr=params.peak_dec_lr, + clipping_scale=5.0, + parameters_names=[dec_names], + ) + + scheduler_enc = Eden(optimizer_enc, params.lr_batches, params.lr_epochs) + scheduler_dec = Eden(optimizer_dec, params.lr_batches, params.lr_epochs) + optimizer = [optimizer_enc, optimizer_dec] + scheduler = [scheduler_enc, scheduler_dec] + + else: + parameters_names = [] + parameters_names.append( + [name_param_pair[0] for name_param_pair in model.named_parameters()] + ) + + logging.info(f"len name = {len(parameters_names)}") + logging.info(f"len param = {len(list(model.parameters()))}") + + optimizer = ScaledAdam( + model.parameters(), + lr=params.base_lr, + clipping_scale=2.0, + parameters_names=parameters_names, + ) + + scheduler = Eden(optimizer, params.lr_batches, params.lr_epochs) + + if checkpoints and ("optimizer" in checkpoints or "optimizer_enc" in checkpoints): + if params.multi_optim: + logging.info("Loading optimizer state dict") + optimizer_enc.load_state_dict(checkpoints["optimizer_enc"]) + optimizer_dec.load_state_dict(checkpoints["optimizer_dec"]) + + else: + logging.info("Loading optimizer state dict") + optimizer.load_state_dict(checkpoints["optimizer"]) + + if checkpoints: + if ( + params.multi_optim + and "scheduler_enc" in checkpoints + and checkpoints["scheduler_enc"] is not None + ): + logging.info("Loading enc/dec scheduler state dict") + scheduler_enc.load_state_dict(checkpoints["scheduler_enc"]) + scheduler_dec.load_state_dict(checkpoints["scheduler_dec"]) + else: + logging.info("Loading scheduler state dict") + scheduler.load_state_dict(checkpoints["scheduler"]) + + if params.print_diagnostics: + opts = diagnostics.TensorDiagnosticOptions( + 2**22 + ) # allow 4 megabytes per sub-module + diagnostic = diagnostics.attach_diagnostics(model, opts) + + if params.inf_check: + register_inf_check_hooks(model) + + librispeech = LibriSpeechAsrDataModule(args) + + train_cuts = librispeech.train_clean_100_cuts() + if params.full_libri: + train_cuts += librispeech.train_clean_360_cuts() + train_cuts += librispeech.train_other_500_cuts() + + def remove_short_and_long_utt(c: Cut): + # Keep only utterances with duration between 1 second and 20 seconds + # + # Caution: There is a reason to select 20.0 here. Please see + # ../local/display_manifest_statistics.py + # + # You should use ../local/display_manifest_statistics.py to get + # an utterance duration distribution for your dataset to select + # the threshold + return 1.0 <= c.duration <= 20.0 + + train_cuts = train_cuts.filter(remove_short_and_long_utt) + + if params.start_batch > 0 and checkpoints and "sampler" in checkpoints: + # We only load the sampler's state dict when it loads a checkpoint + # saved in the middle of an epoch + sampler_state_dict = checkpoints["sampler"] + else: + sampler_state_dict = None + + train_dl = librispeech.train_dataloaders( + train_cuts, sampler_state_dict=sampler_state_dict + ) + + valid_cuts = librispeech.dev_clean_cuts() + valid_cuts += librispeech.dev_other_cuts() + valid_dl = librispeech.valid_dataloaders(valid_cuts) + + ''' + if not params.print_diagnostics: + scan_pessimistic_batches_for_oom( + model=model, + train_dl=train_dl, + optimizer=optimizer, + sp=sp, + params=params, + ) + ''' + + scaler = GradScaler(enabled=params.use_fp16, init_scale=1.0) + if checkpoints and "grad_scaler" in checkpoints: + logging.info("Loading grad scaler state dict") + scaler.load_state_dict(checkpoints["grad_scaler"]) + + for epoch in range(params.start_epoch, params.num_epochs + 1): + if params.multi_optim: + scheduler_enc.step_epoch(epoch - 1) + scheduler_dec.step_epoch(epoch - 1) + else: + scheduler.step_epoch(epoch - 1) + fix_random_seed(params.seed + epoch - 1) + train_dl.sampler.set_epoch(epoch - 1) + + if tb_writer is not None: + tb_writer.add_scalar("train/epoch", epoch, params.batch_idx_train) + + params.cur_epoch = epoch + + train_one_epoch( + params=params, + model=model, + model_avg=model_avg, + optimizer=optimizer, + scheduler=scheduler, + sp=sp, + train_dl=train_dl, + valid_dl=valid_dl, + scaler=scaler, + tb_writer=tb_writer, + world_size=world_size, + rank=rank, + wb=wb, + ) + + if params.print_diagnostics: + diagnostic.print_diagnostics() + break + + save_checkpoint( + params=params, + model=model, + model_avg=model_avg, + optimizer=optimizer, + scheduler=scheduler, + sampler=train_dl.sampler, + scaler=scaler, + rank=rank, + ) + + logging.info("Done!") + + if world_size > 1: + torch.distributed.barrier() + cleanup_dist() + + +def run_adapter(rank, world_size, args, wb=None): + """ + Args: + rank: + It is a value between 0 and `world_size-1`, which is + passed automatically by `mp.spawn()` in :func:`main`. + The node with rank 0 is responsible for saving checkpoint. + world_size: + Number of GPUs for DDP training. + args: + The return value of get_parser().parse_args() + """ + params = get_params() + params.update(vars(args)) + + fix_random_seed(params.seed) + if world_size > 1: + setup_dist(rank, world_size, params.master_port) + + setup_logger(f"{params.exp_dir}/log/log-train") + logging.info("Training started") + + if args.tensorboard and rank == 0: + tb_writer = SummaryWriter(log_dir=f"{params.exp_dir}/tensorboard") + else: + tb_writer = None + + device = torch.device("cpu") + if torch.cuda.is_available(): + device = torch.device("cuda", rank) + logging.info(f"Device: {device}") + + sp = spm.SentencePieceProcessor() + sp.load(params.bpe_model) + + # is defined in local/train_bpe_model.py + params.blank_id = sp.piece_to_id("") + params.vocab_size = sp.get_piece_size() + + logging.info(params) + + logging.info("About to create model") + model = get_transducer_model(params) + + num_param = sum([p.numel() if p.requires_grad else 0 for p in model.parameters()]) + logging.info(f"Number of model parameters: {num_param}") + + assert params.save_every_n >= params.average_period + model_avg: Optional[nn.Module] = None + if rank == 0: + # model_avg is only used with rank 0 + model_avg = copy.deepcopy(model).to(torch.float64) + + assert params.start_epoch > 0, params.start_epoch + checkpoints = load_checkpoint_if_available( + params=params, model=model, model_avg=model_avg + ) + + model.to(device) + if world_size > 1: + logging.info("Using DDP") + model = DDP(model, device_ids=[rank], find_unused_parameters=True) + + adapter_names = [] + adapter_param = [] + for n, p in model.named_parameters(): + if 'adapters' in n: + adapter_names.append(n) + adapter_param.append(p) + #else: + # p.requires_grad = False + optimizer_adapter = ScaledAdam( + adapter_param, + lr=params.adapter_lr, + clipping_scale=5.0, + parameters_names=[adapter_names], + ) + scheduler_adapter = Eden(optimizer_adapter, 5000, 3.5) #params.lr_batche, params.lr_epochs) + + optimizer, scheduler = optimizer_adapter, scheduler_adapter + + librispeech = LibriSpeechAsrDataModule(args) + + train_cuts = librispeech.train_clean_100_cuts() + if params.full_libri: + train_cuts += librispeech.train_clean_360_cuts() + train_cuts += librispeech.train_other_500_cuts() + + def remove_short_and_long_utt(c: Cut): + return 1.0 <= c.duration <= 20.0 + + train_cuts = train_cuts.filter(remove_short_and_long_utt) + + sampler_state_dict = None + + train_dl = librispeech.train_dataloaders( + train_cuts, sampler_state_dict=sampler_state_dict + ) + + valid_cuts = librispeech.dev_clean_cuts() + valid_cuts += librispeech.dev_other_cuts() + valid_dl = librispeech.valid_dataloaders(valid_cuts) + + scaler = GradScaler(enabled=params.use_fp16, init_scale=1.0) + + for epoch in range(params.start_epoch, params.num_epochs + 1): + scheduler.step_epoch(epoch - 1) + fix_random_seed(params.seed + epoch - 1) + train_dl.sampler.set_epoch(epoch - 1) + + if tb_writer is not None: + tb_writer.add_scalar("train/epoch", epoch, params.batch_idx_train) + + params.cur_epoch = epoch + + train_one_epoch( + params=params, + model=model, + model_avg=model_avg, + optimizer=optimizer, + scheduler=scheduler, + sp=sp, + train_dl=train_dl, + valid_dl=valid_dl, + scaler=scaler, + tb_writer=tb_writer, + world_size=world_size, + rank=rank, + wb=wb, + ) + + if params.print_diagnostics: + diagnostic.print_diagnostics() + break + + save_checkpoint( + params=params, + model=model, + model_avg=model_avg, + optimizer=optimizer, + scheduler=scheduler, + sampler=train_dl.sampler, + scaler=scaler, + rank=rank, + ) + + logging.info("Done!") + + if world_size > 1: + torch.distributed.barrier() + cleanup_dist() + + +def display_and_save_batch( + batch: dict, + params: AttributeDict, + sp: spm.SentencePieceProcessor, +) -> None: + """Display the batch statistics and save the batch into disk. + + Args: + batch: + A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()` + for the content in it. + params: + Parameters for training. See :func:`get_params`. + sp: + The BPE model. + """ + from lhotse.utils import uuid4 + + filename = f"{params.exp_dir}/batch-{uuid4()}.pt" + logging.info(f"Saving batch to {filename}") + torch.save(batch, filename) + + supervisions = batch["supervisions"] + features = batch["inputs"] + + logging.info(f"features shape: {features.shape}") + + y = sp.encode(supervisions["text"], out_type=int) + num_tokens = sum(len(i) for i in y) + logging.info(f"num tokens: {num_tokens}") + + +def scan_pessimistic_batches_for_oom( + model: Union[nn.Module, DDP], + train_dl: torch.utils.data.DataLoader, + optimizer: torch.optim.Optimizer, + sp: spm.SentencePieceProcessor, + params: AttributeDict, +): + from lhotse.dataset import find_pessimistic_batches + + logging.info( + "Sanity check -- see if any of the batches in epoch 1 would cause OOM." + ) + batches, crit_values = find_pessimistic_batches(train_dl.sampler) + for criterion, cuts in batches.items(): + batch = train_dl.dataset[cuts] + try: + with torch.cuda.amp.autocast(enabled=params.use_fp16): + loss, _ = compute_loss( + params=params, + model=model, + sp=sp, + batch=batch, + is_training=True, + ) + loss.backward() + optimizer.zero_grad() + except Exception as e: + if "CUDA out of memory" in str(e): + logging.error( + "Your GPU ran out of memory with the current " + "max_duration setting. We recommend decreasing " + "max_duration and trying again.\n" + f"Failing criterion: {criterion} " + f"(={crit_values[criterion]}) ..." + ) + display_and_save_batch(batch, params=params, sp=sp) + raise + logging.info( + f"Maximum memory allocated so far is {torch.cuda.max_memory_allocated()//1000000}MB" + ) + + +def main(): + parser = get_parser() + LibriSpeechAsrDataModule.add_arguments(parser) + args = parser.parse_args() + #args.exp_dir = args.exp_dir + str(random.randint(0,400)) + args.exp_dir = Path(args.exp_dir) + + logging.info("save arguments to config.yaml...") + save_args(args) + + if args.wandb: wb = wandb.init(project="d2v-T", entity="dohe0342", config=vars(args)) + else: wb = None + + world_size = args.world_size + assert world_size >= 1 + if world_size > 1: + mp.spawn(run if not args.add_adapter else run_adapter, + args=(world_size, args, wb), + nprocs=world_size, + join=True + ) + else: + if not args.add_adapter: run(rank=0, world_size=1, args=args, wb=wb) + else: run(rank=0, world_size=1, args=args, wb=wb) + +torch.set_num_threads(1) +torch.set_num_interop_threads(1) + +if __name__ == "__main__": + main() diff --git a/egs/tedlium3/ASR/pruned_transducer_stateless_d2v_v2/train_adapter.py b/egs/tedlium3/ASR/pruned_transducer_stateless_d2v_v2/train_adapter.py new file mode 100755 index 000000000..173404859 --- /dev/null +++ b/egs/tedlium3/ASR/pruned_transducer_stateless_d2v_v2/train_adapter.py @@ -0,0 +1,1800 @@ +#!/usr/bin/env python3 +# Copyright 2021-2022 Xiaomi Corp. (authors: Fangjun Kuang, +# Wei Kang, +# Mingshuang Luo,) +# Zengwei Yao) +# +# See ../../../../LICENSE for clarification regarding multiple authors +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +""" +Usage: + +export CUDA_VISIBLE_DEVICES="0,1,2,3" + +./pruned_transducer_stateless7_ctc/train.py \ + --world-size 4 \ + --num-epochs 30 \ + --start-epoch 1 \ + --exp-dir pruned_transducer_stateless7_ctc/exp \ + --full-libri 1 \ + --max-duration 300 + +# For mix precision training: + +./pruned_transducer_stateless7_ctc/train.py \ + --world-size 4 \ + --num-epochs 30 \ + --start-epoch 1 \ + --use-fp16 1 \ + --exp-dir pruned_transducer_stateless7_ctc/exp \ + --full-libri 1 \ + --max-duration 550 + +# For d2v-T training: +export CUDA_VISIBLE_DEVICES="0,1,2,3,4,5,6,7" + +./pruned_transducer_stateless_d2v_v2/train.py \ + --wandb true \ + --input-strategy AudioSamples \ + --enable-spec-aug False \ + --multi-optim True \ + --world-size 8 \ + --num-epochs 30 \ + --start-epoch 1 \ + --full-libri 0 \ + --exp-dir ./pruned_transducer_stateless_d2v_v2/$1 \ + --max-duration 250 \ + --freeze-finetune-updates 2000 \ + --use-fp16 1 \ + --peak-enc-lr 0.001 \ + --peak-dec-lr 0.05 \ + --accum-grads 1 \ + --encoder-type d2v \ + --additional-block True \ + --encoder-dim 768 \ + --decoder-dim 768 \ + --joiner-dim 768 \ + --prune-range 20 \ + --context-size 2 \ + --ctc-loss-scale 0.2 + +""" + + +import random +import argparse +import copy +import logging +import warnings +from pathlib import Path +from shutil import copyfile +from typing import Any, Dict, Optional, Tuple, Union + +import k2 +import optim +import sentencepiece as spm +import torch +import torch.multiprocessing as mp +import torch.nn as nn +from asr_datamodule import LibriSpeechAsrDataModule +from decoder import Decoder +from joiner import Joiner +from lhotse.cut import Cut +from lhotse.dataset.sampling.base import CutSampler +from lhotse.utils import fix_random_seed +from model import Transducer +from optim import Eden, ScaledAdam +from torch import Tensor +from torch.cuda.amp import GradScaler +from torch.nn.parallel import DistributedDataParallel as DDP +from torch.utils.tensorboard import SummaryWriter +from zipformer import Zipformer +from data2vec_encoder import FairSeqData2VecEncoder + +from icefall import diagnostics +from icefall.checkpoint import remove_checkpoints +from icefall.checkpoint import update_averaged_model +from checkpoint import ( + save_checkpoint as save_checkpoint_impl, + save_checkpoint_with_global_batch_idx, + load_checkpoint +) +from icefall.dist import cleanup_dist, setup_dist +from icefall.env import get_env_info +from icefall.hooks import register_inf_check_hooks +from icefall.utils import ( + AttributeDict, + MetricsTracker, + encode_supervisions, + setup_logger, + str2bool, + save_args, +) + +import wandb + +#from icefall.checkpoint import save_checkpoint as save_checkpoint_impl +LRSchedulerType = Union[torch.optim.lr_scheduler._LRScheduler, optim.LRScheduler] + + +def set_batch_count(model: Union[nn.Module, DDP], batch_count: float) -> None: + if isinstance(model, DDP): + # get underlying nn.Module + model = model.module + for module in model.modules(): + if hasattr(module, "batch_count"): + module.batch_count = batch_count + model.encoder.num_updates = int(batch_count) + + +def add_adapter_arguments(parser: argparse.ArgumentParser): + parser.add_argument( + "--add-adapter", + type=str2bool, + default=False, + help="add adapter to rep model's encoder" + ) + + parser.add_argument( + "--adapter-lr", + type=float, + default=0.0001, + help="adapter learning rate" + ) + + parser.add_argument( + "--gender", + type=str, + default='male', + help="select gender" + ) + + +def add_rep_arguments(parser: argparse.ArgumentParser): + parser.add_argument( + "--wandb", + type=str2bool, + default=True, + help="Use wandb for MLOps", + ) + parser.add_argument( + "--hpo", + type=str2bool, + default=False, + help="Use small db for HPO", + ) + + parser.add_argument( + "--accum-grads", + type=int, + default=1, + help="accum-grad num.", + ) + + parser.add_argument( + "--multi-optim", + type=str2bool, + default=True, + help="use sperate optimizer (enc / dec)", + ) + + parser.add_argument( + "--peak-enc-lr", + type=float, + default=0.0001, + help="The initial learning rate. This value should not need to be changed.", + ) + + parser.add_argument( + "--peak-dec-lr", + type=float, + default=0.001, + help="The initial learning rate. This value should not need to be changed.", + ) + + parser.add_argument( + "--encoder-type", + type=str, + default='d2v', + help="Type of encoder (e.g. conformer, w2v, d2v...", + ) + + parser.add_argument( + "--encoder-dim", + type=int, + default=768, + help="encoder embedding dimension", + ) + + parser.add_argument( + "--freeze-finetune-updates", + type=int, + default=0 + ) + + parser.add_argument( + "--additional-block", + type=str2bool, + default=True, + ) + + parser.add_argument( + "--decode-interval", + type=int, + default=200, + help="decode interval", + ) + + +def add_model_arguments(parser: argparse.ArgumentParser): + parser.add_argument( + "--num-encoder-layers", + type=str, + default="2,4,3,2,4", + help="Number of zipformer encoder layers, comma separated.", + ) + + parser.add_argument( + "--feedforward-dims", + type=str, + default="1024,1024,2048,2048,1024", + help="Feedforward dimension of the zipformer encoder layers, comma separated.", + ) + + parser.add_argument( + "--nhead", + type=str, + default="8,8,8,8,8", + help="Number of attention heads in the zipformer encoder layers.", + ) + + parser.add_argument( + "--encoder-dims", + type=str, + default="384,384,384,384,384", + help="Embedding dimension in the 2 blocks of zipformer encoder layers, comma separated", + ) + + parser.add_argument( + "--attention-dims", + type=str, + default="192,192,192,192,192", + help="""Attention dimension in the 2 blocks of zipformer encoder layers, comma separated; + not the same as embedding dimension.""", + ) + + parser.add_argument( + "--encoder-unmasked-dims", + type=str, + default="256,256,256,256,256", + help="Unmasked dimensions in the encoders, relates to augmentation during training. " + "Must be <= each of encoder_dims. Empirically, less than 256 seems to make performance " + " worse.", + ) + + parser.add_argument( + "--zipformer-downsampling-factors", + type=str, + default="1,2,4,8,2", + help="Downsampling factor for each stack of encoder layers.", + ) + + parser.add_argument( + "--cnn-module-kernels", + type=str, + default="31,31,31,31,31", + help="Sizes of kernels in convolution modules", + ) + + parser.add_argument( + "--decoder-dim", + type=int, + default=768, + help="Embedding dimension in the decoder model.", + ) + + parser.add_argument( + "--joiner-dim", + type=int, + default=768, + help="""Dimension used in the joiner model. + Outputs from the encoder and decoder model are projected + to this dimension before adding. + """, + ) + + +def get_parser(): + parser = argparse.ArgumentParser( + formatter_class=argparse.ArgumentDefaultsHelpFormatter + ) + + parser.add_argument( + "--world-size", + type=int, + default=1, + help="Number of GPUs for DDP training.", + ) + + parser.add_argument( + "--master-port", + type=int, + default=12354, + help="Master port to use for DDP training.", + ) + + parser.add_argument( + "--tensorboard", + type=str2bool, + default=True, + help="Should various information be logged in tensorboard.", + ) + + parser.add_argument( + "--num-epochs", + type=int, + default=30, + help="Number of epochs to train.", + ) + + parser.add_argument( + "--num-updates", + type=int, + default=5000, + help="Number of epochs to train.", + ) + + parser.add_argument( + "--start-epoch", + type=int, + default=1, + help="""Resume training from this epoch. It should be positive. + If larger than 1, it will load checkpoint from + exp-dir/epoch-{start_epoch-1}.pt + """, + ) + + parser.add_argument( + "--start-batch", + type=int, + default=0, + help="""If positive, --start-epoch is ignored and + it loads the checkpoint from exp-dir/checkpoint-{start_batch}.pt + """, + ) + + parser.add_argument( + "--exp-dir", + type=str, + default="pruned_transducer_stateless7_ctc/exp", + help="""The experiment dir. + It specifies the directory where all training related + files, e.g., checkpoints, log, etc, are saved + """, + ) + + parser.add_argument( + "--bpe-model", + type=str, + default="data/lang_bpe_500/bpe.model", + help="Path to the BPE model", + ) + + parser.add_argument( + "--base-lr", type=float, default=0.05, help="The base learning rate." + ) + + parser.add_argument( + "--lr-batches", + type=float, + default=5000, + help="""Number of steps that affects how rapidly the learning rate + decreases. We suggest not to change this.""", + ) + + parser.add_argument( + "--lr-epochs", + type=float, + default=3.5, + help="""Number of epochs that affects how rapidly the learning rate decreases. + """, + ) + + parser.add_argument( + "--context-size", + type=int, + default=2, + help="The context size in the decoder. 1 means bigram; 2 means tri-gram", + ) + + parser.add_argument( + "--prune-range", + type=int, + default=5, + help="The prune range for rnnt loss, it means how many symbols(context)" + "we are using to compute the loss", + ) + + parser.add_argument( + "--lm-scale", + type=float, + default=0.25, + help="The scale to smooth the loss with lm " + "(output of prediction network) part.", + ) + + parser.add_argument( + "--am-scale", + type=float, + default=0.0, + help="The scale to smooth the loss with am (output of encoder network) part.", + ) + + parser.add_argument( + "--simple-loss-scale", + type=float, + default=0.5, + help="To get pruning ranges, we will calculate a simple version" + "loss(joiner is just addition), this simple loss also uses for" + "training (as a regularization item). We will scale the simple loss" + "with this parameter before adding to the final loss.", + ) + + parser.add_argument( + "--ctc-loss-scale", + type=float, + default=0.2, + help="Scale for CTC loss.", + ) + + parser.add_argument( + "--seed", + type=int, + default=42, + help="The seed for random generators intended for reproducibility", + ) + + parser.add_argument( + "--print-diagnostics", + type=str2bool, + default=False, + help="Accumulate stats on activations, print them and exit.", + ) + + parser.add_argument( + "--inf-check", + type=str2bool, + default=False, + help="Add hooks to check for infinite module outputs and gradients.", + ) + + parser.add_argument( + "--save-every-n", + type=int, + default=200, + help="""Save checkpoint after processing this number of batches" + periodically. We save checkpoint to exp-dir/ whenever + params.batch_idx_train % save_every_n == 0. The checkpoint filename + has the form: f'exp-dir/checkpoint-{params.batch_idx_train}.pt' + Note: It also saves checkpoint to `exp-dir/epoch-xxx.pt` at the + end of each epoch where `xxx` is the epoch number counting from 0. + """, + ) + + parser.add_argument( + "--keep-last-k", + type=int, + default=30, + help="""Only keep this number of checkpoints on disk. + For instance, if it is 3, there are only 3 checkpoints + in the exp-dir with filenames `checkpoint-xxx.pt`. + It does not affect checkpoints with name `epoch-xxx.pt`. + """, + ) + + parser.add_argument( + "--average-period", + type=int, + default=10, + help="""Update the averaged model, namely `model_avg`, after processing + this number of batches. `model_avg` is a separate version of model, + in which each floating-point parameter is the average of all the + parameters from the start of training. Each time we take the average, + we do: `model_avg = model * (average_period / batch_idx_train) + + model_avg * ((batch_idx_train - average_period) / batch_idx_train)`. + """, + ) + + parser.add_argument( + "--use-fp16", + type=str2bool, + default=True, + help="Whether to use half precision training.", + ) + + add_model_arguments(parser) + add_rep_arguments(parser) + add_adapter_arguments(parser) + + return parser + + +def get_params() -> AttributeDict: + """Return a dict containing training parameters. + + All training related parameters that are not passed from the commandline + are saved in the variable `params`. + + Commandline options are merged into `params` after they are parsed, so + you can also access them via `params`. + + Explanation of options saved in `params`: + + - best_train_loss: Best training loss so far. It is used to select + the model that has the lowest training loss. It is + updated during the training. + + - best_valid_loss: Best validation loss so far. It is used to select + the model that has the lowest validation loss. It is + updated during the training. + + - best_train_epoch: It is the epoch that has the best training loss. + + - best_valid_epoch: It is the epoch that has the best validation loss. + + - batch_idx_train: Used to writing statistics to tensorboard. It + contains number of batches trained so far across + epochs. + + - log_interval: Print training loss if batch_idx % log_interval` is 0 + + - reset_interval: Reset statistics if batch_idx % reset_interval is 0 + + - valid_interval: Run validation if batch_idx % valid_interval is 0 + + - feature_dim: The model input dim. It has to match the one used + in computing features. + + - subsampling_factor: The subsampling factor for the model. + + - encoder_dim: Hidden dim for multi-head attention model. + + - num_decoder_layers: Number of decoder layer of transformer decoder. + + - warm_step: The warmup period that dictates the decay of the + scale on "simple" (un-pruned) loss. + """ + params = AttributeDict( + { + "best_train_loss": float("inf"), + "best_valid_loss": float("inf"), + "best_train_epoch": -1, + "best_valid_epoch": -1, + "batch_idx_train": 0, + "log_interval": 20, + "reset_interval": 200, + "valid_interval": 3000, # For the 100h subset, use 800 + # parameters for zipformer + "feature_dim": 80, + "subsampling_factor": 320, # not passed in, this is fixed. + # parameters for ctc loss + "beam_size": 10, + "use_double_scores": True, + "warm_step": 0, + #"warm_step": 4000, + #"warm_step": 3000, + "env_info": get_env_info(), + } + ) + + return params + + +def get_encoder_model(params: AttributeDict) -> nn.Module: + # TODO: We can add an option to switch between Zipformer and Transformer + def to_int_tuple(s: str): + return tuple(map(int, s.split(","))) + + if params.encoder_type == 'd2v': + encoder = FairSeqData2VecEncoder( + input_size=params.encoder_dim, + w2v_url='None', + output_size=params.encoder_dim, + freeze_finetune_updates=params.freeze_finetune_updates, + additional_block=params.additional_block, + ) + else: + encoder = Zipformer( + num_features=params.feature_dim, + output_downsampling_factor=2, + zipformer_downsampling_factors=to_int_tuple( + params.zipformer_downsampling_factors + ), + encoder_dims=to_int_tuple(params.encoder_dims), + attention_dim=to_int_tuple(params.attention_dims), + encoder_unmasked_dims=to_int_tuple(params.encoder_unmasked_dims), + nhead=to_int_tuple(params.nhead), + feedforward_dim=to_int_tuple(params.feedforward_dims), + cnn_module_kernels=to_int_tuple(params.cnn_module_kernels), + num_encoder_layers=to_int_tuple(params.num_encoder_layers), + ) + + return encoder + + +def get_decoder_model(params: AttributeDict) -> nn.Module: + decoder = Decoder( + vocab_size=params.vocab_size, + decoder_dim=params.decoder_dim, + blank_id=params.blank_id, + context_size=params.context_size, + ) + return decoder + + +def get_joiner_model(params: AttributeDict) -> nn.Module: + joiner = Joiner( + encoder_dim=params.encoder_dim if params.encoder_type == 'd2v' else int(params.encoder_dims.split(",")[-1]), + decoder_dim=params.decoder_dim, + joiner_dim=params.joiner_dim, + vocab_size=params.vocab_size, + ) + return joiner + + +def get_transducer_model(params: AttributeDict) -> nn.Module: + encoder = get_encoder_model(params) + decoder = get_decoder_model(params) + joiner = get_joiner_model(params) + + model = Transducer( + encoder=encoder, + decoder=decoder, + joiner=joiner, + encoder_dim=params.encoder_dim if params.encoder_type == 'd2v' else int(params.encoder_dims.split(",")[-1]), + decoder_dim=params.decoder_dim, + joiner_dim=params.joiner_dim, + vocab_size=params.vocab_size, + ) + return model + + +def load_checkpoint_if_available( + params: AttributeDict, + model: nn.Module, + model_avg: nn.Module = None, + optimizer: Optional[torch.optim.Optimizer] = None, + scheduler: Optional[LRSchedulerType] = None, +) -> Optional[Dict[str, Any]]: + """Load checkpoint from file. + + If params.start_batch is positive, it will load the checkpoint from + `params.exp_dir/checkpoint-{params.start_batch}.pt`. Otherwise, if + params.start_epoch is larger than 1, it will load the checkpoint from + `params.start_epoch - 1`. + + Apart from loading state dict for `model` and `optimizer` it also updates + `best_train_epoch`, `best_train_loss`, `best_valid_epoch`, + and `best_valid_loss` in `params`. + + Args: + params: + The return value of :func:`get_params`. + model: + The training model. + model_avg: + The stored model averaged from the start of training. + optimizer: + The optimizer that we are using. + scheduler: + The scheduler that we are using. + Returns: + Return a dict containing previously saved training info. + """ + if params.start_batch > 0: + filename = params.exp_dir / f"checkpoint-{params.start_batch}.pt" + elif params.start_epoch > 1: + filename = params.exp_dir / f"epoch-{params.start_epoch-1}.pt" + elif params.add_adapter: + filename = params.exp_dir / f"../d2v-base-T.pt" + else: + return None + + assert filename.is_file(), f"{filename} does not exist!" + + saved_params = load_checkpoint( + filename, + model=model, + model_avg=model_avg, + optimizer=optimizer, + scheduler=scheduler, + strict=True if not params.add_adapter else False, + ) + + keys = [ + "best_train_epoch", + "best_valid_epoch", + "batch_idx_train", + "best_train_loss", + "best_valid_loss", + ] + for k in keys: + params[k] = saved_params[k] + + if params.start_batch > 0: + if "cur_epoch" in saved_params: + params["start_epoch"] = saved_params["cur_epoch"] + + if "cur_batch_idx" in saved_params: + params["cur_batch_idx"] = saved_params["cur_batch_idx"] + + params.batch_idx_train = 0 + + return saved_params + + +def save_checkpoint( + params: AttributeDict, + model: Union[nn.Module, DDP], + model_avg: Optional[nn.Module] = None, + optimizer: Optional[torch.optim.Optimizer] = None, + scheduler: Optional[LRSchedulerType] = None, + sampler: Optional[CutSampler] = None, + scaler: Optional[GradScaler] = None, + rank: int = 0, +) -> None: + """Save model, optimizer, scheduler and training stats to file. + + Args: + params: + It is returned by :func:`get_params`. + model: + The training model. + model_avg: + The stored model averaged from the start of training. + optimizer: + The optimizer used in the training. + sampler: + The sampler for the training dataset. + scaler: + The scaler used for mix precision training. + """ + if rank != 0: + return + filename = params.exp_dir / f"epoch-{params.cur_epoch}.pt" + save_checkpoint_impl( + filename=filename, + model=model, + model_avg=model_avg, + params=params, + optimizer=optimizer, + scheduler=scheduler, + sampler=sampler, + scaler=scaler, + rank=rank, + ) + + if params.best_train_epoch == params.cur_epoch: + best_train_filename = params.exp_dir / "best-train-loss.pt" + copyfile(src=filename, dst=best_train_filename) + + if params.best_valid_epoch == params.cur_epoch: + best_valid_filename = params.exp_dir / "best-valid-loss.pt" + copyfile(src=filename, dst=best_valid_filename) + + +def compute_loss( + params: AttributeDict, + model: Union[nn.Module, DDP], + sp: spm.SentencePieceProcessor, + batch: dict, + is_training: bool, + decode: bool = False, +) -> Tuple[Tensor, MetricsTracker]: + """ + Compute transducer loss given the model and its inputs. + + Args: + params: + Parameters for training. See :func:`get_params`. + model: + The model for training. It is an instance of Zipformer in our case. + batch: + A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()` + for the content in it. + is_training: + True for training. False for validation. When it is True, this + function enables autograd during computation; when it is False, it + disables autograd. + warmup: a floating point value which increases throughout training; + values >= 1.0 are fully warmed up and have all modules present. + """ + device = model.device if isinstance(model, DDP) else next(model.parameters()).device + feature = batch["inputs"] + # at entry, feature is (N, T, C) + assert feature.ndim == 2 or feature.ndim == 3 + feature = feature.to(device) + + supervisions = batch["supervisions"] + + if feature.ndim == 2: + feature_lens = [] + for supervision in supervisions['cut']: + try: feature_lens.append(supervision.tracks[0].cut.recording.num_samples) + except: feature_lens.append(supervision.recording.num_samples) + feature_lens = torch.tensor(feature_lens) + + elif feature.ndim == 3: + feature_lens = supervisions["num_frames"].to(device) + + batch_idx_train = params.batch_idx_train + warm_step = params.warm_step + + texts = batch["supervisions"]["text"] + + token_ids = sp.encode(texts, out_type=int) + y = k2.RaggedTensor(token_ids).to(device) + + with torch.set_grad_enabled(is_training): + simple_loss, pruned_loss, ctc_output = model( + x=feature, + x_lens=feature_lens, + y=y, + prune_range=params.prune_range, + am_scale=params.am_scale, + lm_scale=params.lm_scale, + ) + + s = params.simple_loss_scale + # take down the scale on the simple loss from 1.0 at the start + # to params.simple_loss scale by warm_step. + simple_loss_scale = ( + s + if batch_idx_train >= warm_step + else 1.0 - (batch_idx_train / warm_step) * (1.0 - s) + ) + pruned_loss_scale = ( + 1.0 + if batch_idx_train >= warm_step + else 0.1 + 0.9 * (batch_idx_train / warm_step) + ) + + loss = simple_loss_scale * simple_loss + pruned_loss_scale * pruned_loss + + info = MetricsTracker() + + if params.ctc_loss_scale > 0: + # Compute ctc loss + + # NOTE: We need `encode_supervisions` to sort sequences with + # different duration in decreasing order, required by + # `k2.intersect_dense` called in `k2.ctc_loss` + with warnings.catch_warnings(): + warnings.simplefilter("ignore") + supervision_segments, token_ids = encode_supervisions( + supervisions, + subsampling_factor=params.subsampling_factor, + token_ids=token_ids, + ) + + # Works with a BPE model + decoding_graph = k2.ctc_graph(token_ids, modified=False, device=device) + dense_fsa_vec = k2.DenseFsaVec( + ctc_output, + supervision_segments, + allow_truncate=params.subsampling_factor - 1, + ) + + ctc_loss = k2.ctc_loss( + decoding_graph=decoding_graph, + dense_fsa_vec=dense_fsa_vec, + output_beam=params.beam_size, + reduction="sum", + use_double_scores=params.use_double_scores, + ) + assert ctc_loss.requires_grad == is_training + loss += params.ctc_loss_scale * ctc_loss + + info["ctc_loss"] = ctc_loss.detach().cpu().item() + + assert loss.requires_grad == is_training + + if decode: + model.eval() + with torch.no_grad(): + try: + hypos = model.module.decode( + x=feature, + x_lens=feature_lens, + y=y, + sp=sp + ) + except: + hypos = model.decode( + x=feature, + x_lens=feature_lens, + y=y, + sp=sp + ) + + logging.info(f'ref: {batch["supervisions"]["text"][0]}') + logging.info(f'hyp: {" ".join(hypos[0])}') + model.train() + + with warnings.catch_warnings(): + warnings.simplefilter("ignore") + info["frames"] = (feature_lens // params.subsampling_factor).sum().item() + + # Note: We use reduction=sum while computing the loss. + info["utterances"] = feature.size(0) + info["loss"] = loss.detach().cpu().item() + info["simple_loss"] = simple_loss.detach().cpu().item() + info["pruned_loss"] = pruned_loss.detach().cpu().item() + + return loss, info + + +def compute_validation_loss( + params: AttributeDict, + model: Union[nn.Module, DDP], + sp: spm.SentencePieceProcessor, + valid_dl: torch.utils.data.DataLoader, + world_size: int = 1, +) -> MetricsTracker: + """Run the validation process.""" + model.eval() + + tot_loss = MetricsTracker() + + for batch_idx, batch in enumerate(valid_dl): + loss, loss_info = compute_loss( + params=params, + model=model, + sp=sp, + batch=batch, + is_training=False, + ) + assert loss.requires_grad is False + tot_loss = tot_loss + loss_info + + if world_size > 1: + tot_loss.reduce(loss.device) + + loss_value = tot_loss["loss"] / tot_loss["utterances"] + if loss_value < params.best_valid_loss: + params.best_valid_epoch = params.cur_epoch + params.best_valid_loss = loss_value + + return tot_loss + + +def train_one_epoch( + params: AttributeDict, + model: Union[nn.Module, DDP], + optimizer: torch.optim.Optimizer or [torch.optim.Optimizer, torch.optim.Optimizer], + scheduler: LRSchedulerType or [LRSchedulerType, LRSchedulerType], + sp: spm.SentencePieceProcessor, + train_dl: torch.utils.data.DataLoader, + valid_dl: torch.utils.data.DataLoader, + scaler: GradScaler, + model_avg: Optional[nn.Module] = None, + tb_writer: Optional[SummaryWriter] = None, + world_size: int = 1, + rank: int = 0, + wb = None, +) -> None: + """Train the model for one epoch. + + The training loss from the mean of all frames is saved in + `params.train_loss`. It runs the validation process every + `params.valid_interval` batches. + + Args: + params: + It is returned by :func:`get_params`. + model: + The model for training. + optimizer: + The optimizer we are using. + scheduler: + The learning rate scheduler, we call step() every step. + train_dl: + Dataloader for the training dataset. + valid_dl: + Dataloader for the validation dataset. + scaler: + The scaler used for mix precision training. + model_avg: + The stored model averaged from the start of training. + tb_writer: + Writer to write log messages to tensorboard. + world_size: + Number of nodes in DDP training. If it is 1, DDP is disabled. + rank: + The rank of the node in DDP training. If no DDP is used, it should + be set to 0. + """ + model.train() + + tot_loss = MetricsTracker() + + cur_batch_idx = params.get("cur_batch_idx", 0) + + if params.multi_optim: + optimizer_enc, optimizer_dec = optimizer[0], optimizer[1] + scheduler_enc, scheduler_dec = scheduler[0], scheduler[1] + + for batch_idx, batch in enumerate(train_dl): + if params.batch_idx_train > params.num_updates: + break + if batch_idx < cur_batch_idx: + continue + cur_batch_idx = batch_idx + + if batch_idx % params.accum_grads == 0: params.batch_idx_train += 1 + batch_size = len(batch["supervisions"]["text"]) + + try: + with torch.cuda.amp.autocast(enabled=params.use_fp16): + loss, loss_info = compute_loss( + params=params, + model=model, + sp=sp, + batch=batch, + is_training=True, + decode = True if batch_idx % params.decode_interval == 0 else False, + ) + + try: loss_info.reduce(loss.device) + except: pass + + numel = params.world_size / (params.accum_grads * loss_info["utterances"]) + loss *= numel ## normalize loss over utts(batch size) + + # summary stats + tot_loss = (tot_loss * (1 - 1 / params.reset_interval)) + loss_info + + # NOTE: We use reduction==sum and loss is computed over utterances + # in the batch and there is no normalization to it so far. + scaler.scale(loss).backward() + + if params.multi_optim and (batch_idx+1) % params.accum_grads == 0: + set_batch_count(model, params.batch_idx_train) + scheduler_enc.step_batch(params.batch_idx_train) + scheduler_dec.step_batch(params.batch_idx_train) + scaler.step(optimizer_enc) + scaler.step(optimizer_dec) + scaler.update() + optimizer_enc.zero_grad() + optimizer_dec.zero_grad() + elif not params.multi_optim and (batch_idx+1) % params.accum_grads == 0: + set_batch_count(model, params.batch_idx_train) + scheduler.step_batch(params.batch_idx_train) + scaler.step(optimizer) + scaler.update() + optimizer.zero_grad() + + except: # noqa + display_and_save_batch(batch, params=params, sp=sp) + raise + + if params.print_diagnostics and batch_idx == 5: + return + + ''' + if ( + rank == 0 + and params.batch_idx_train > 0 + and params.batch_idx_train % params.average_period == 0 + ): + update_averaged_model( + params=params, + model_cur=model, + model_avg=model_avg, + ) + ''' + + if ( + params.batch_idx_train > 0 + and params.batch_idx_train % params.save_every_n == 0 + ): + params.cur_batch_idx = batch_idx + save_checkpoint_with_global_batch_idx( + out_dir=params.exp_dir, + global_batch_idx=params.batch_idx_train, + model=model, + model_avg=model_avg, + params=params, + optimizer=optimizer, + scheduler=scheduler, + sampler=train_dl.sampler, + scaler=scaler, + rank=rank, + ) + del params.cur_batch_idx + ''' + remove_checkpoints( + out_dir=params.exp_dir, + topk=params.keep_last_k, + rank=rank, + ) + ''' + + if batch_idx % 100 == 0 and params.use_fp16: + # If the grad scale was less than 1, try increasing it. The _growth_interval + # of the grad scaler is configurable, but we can't configure it to have different + # behavior depending on the current grad scale. + cur_grad_scale = scaler._scale.item() + ''' + if cur_grad_scale < 1.0 or (cur_grad_scale < 8.0 and batch_idx % 400 == 0): + scaler.update(cur_grad_scale * 2.0) + ''' + if cur_grad_scale < 0.01: + logging.warning(f"Grad scale is small: {cur_grad_scale}") + if cur_grad_scale < 1.0e-05: + wb.log({"valid/loss": 10000}) + raise RuntimeError( + f"grad_scale is too small, exiting: {cur_grad_scale}" + ) + + #if params.batch_idx_train > 4000 and loss > 300 and params.wandb: + # wb.log({"valid/loss": 10000}) + # raise RuntimeError( + # f"divergence... exiting: loss={loss}" + # ) + + if batch_idx % (params.log_interval*params.accum_grads) == 0: + #for n, p in model.named_parameters(): + # if 'adapter' in n: + # print(p) + if params.multi_optim: + cur_enc_lr = scheduler_enc.get_last_lr()[0] + cur_dec_lr = scheduler_dec.get_last_lr()[0] + cur_grad_scale = scaler._scale.item() if params.use_fp16 else 1.0 + + logging.info( + f"Epoch {params.cur_epoch}, " + f"batch {batch_idx}, loss[{loss_info}], " + f"tot_loss[{tot_loss}], batch size: {batch_size}, " + f"enc_lr: {cur_enc_lr:.2e}, " + f"dec_lr: {cur_dec_lr:.2e}, " + + (f"grad_scale: {scaler._scale.item()}" if params.use_fp16 else "") + ) + + else: + cur_lr = scheduler.get_last_lr()[0] + cur_grad_scale = scaler._scale.item() if params.use_fp16 else 1.0 + + logging.info( + f"Epoch {params.cur_epoch}, " + f"batch {batch_idx}, loss[{loss_info}], " + f"tot_loss[{tot_loss}], batch size: {batch_size}, " + f"lr: {cur_lr:.2e}, " + + (f"grad_scale: {scaler._scale.item()}" if params.use_fp16 else "") + ) + + if tb_writer is not None: + if params.multi_optim: + tb_writer.add_scalar( + "train/enc_learning_rate", cur_enc_lr, params.batch_idx_train + ) + tb_writer.add_scalar( + "train/dec_learning_rate", cur_dec_lr, params.batch_idx_train + ) + + else: + tb_writer.add_scalar( + "train/learning_rate", cur_lr, params.batch_idx_train + ) + + loss_info.write_summary( + tb_writer, "train/current_", params.batch_idx_train + ) + tot_loss.write_summary(tb_writer, "train/tot_", params.batch_idx_train) + if params.use_fp16: + tb_writer.add_scalar( + "train/grad_scale", + cur_grad_scale, + params.batch_idx_train, + ) + + if wb is not None and rank == 0: + wb.log({"train/loss": loss_info["loss"]*numel}) + wb.log({"train/simple_loss": loss_info["simple_loss"]*numel}) + wb.log({"train/pruned_loss": loss_info["pruned_loss"]*numel}) + wb.log({"train/ctc_loss": loss_info["ctc_loss"]*numel}) + + ''' + logging.info("Computing validation loss") + valid_info = compute_validation_loss( + params=params, + model=model, + sp=sp, + valid_dl=valid_dl, + world_size=world_size, + ) + model.train() + logging.info(f"Epoch {params.cur_epoch}, validation: {valid_info}") + logging.info( + f"Maximum memory allocated so far is {torch.cuda.max_memory_allocated()//1000000}MB" + ) + if tb_writer is not None: + valid_info.write_summary( + tb_writer, "train/valid_", params.batch_idx_train + ) + + if wb is not None and rank == 0: + numel = 1 / (params.accum_grads * valid_info["utterances"]) + #wb.log({"valid/loss": valid_info["loss"]*numel}) + wb.log({"valid/loss": numel*(valid_info["simple_loss"] + +valid_info["pruned_loss"] + +valid_info["ctc_loss"] + )}) + wb.log({"valid/simple_loss": valid_info["simple_loss"]*numel}) + wb.log({"valid/pruned_loss": valid_info["pruned_loss"]*numel}) + wb.log({"valid/ctc_loss": valid_info["ctc_loss"]*numel}) + ''' + loss_value = tot_loss["loss"] / tot_loss["utterances"] + params.train_loss = loss_value + if params.train_loss < params.best_train_loss: + params.best_train_epoch = params.cur_epoch + params.best_train_loss = params.train_loss + + +def run(rank, world_size, args, wb=None): + """ + Args: + rank: + It is a value between 0 and `world_size-1`, which is + passed automatically by `mp.spawn()` in :func:`main`. + The node with rank 0 is responsible for saving checkpoint. + world_size: + Number of GPUs for DDP training. + args: + The return value of get_parser().parse_args() + """ + params = get_params() + params.update(vars(args)) + #params.warm_step *= params.accum_grads + + fix_random_seed(params.seed) + if world_size > 1: + setup_dist(rank, world_size, params.master_port) + + setup_logger(f"{params.exp_dir}/log/log-train") + logging.info("Training started") + + if args.tensorboard and rank == 0: + tb_writer = SummaryWriter(log_dir=f"{params.exp_dir}/tensorboard") + else: + tb_writer = None + + device = torch.device("cpu") + if torch.cuda.is_available(): + device = torch.device("cuda", rank) + logging.info(f"Device: {device}") + + sp = spm.SentencePieceProcessor() + sp.load(params.bpe_model) + + # is defined in local/train_bpe_model.py + params.blank_id = sp.piece_to_id("") + params.vocab_size = sp.get_piece_size() + + logging.info(params) + + logging.info("About to create model") + model = get_transducer_model(params) + logging.info(model) + + num_param = sum([p.numel() for p in model.parameters()]) + logging.info(f"Number of model parameters: {num_param}") + + assert params.save_every_n >= params.average_period + model_avg: Optional[nn.Module] = None + if rank == 0: + # model_avg is only used with rank 0 + model_avg = copy.deepcopy(model).to(torch.float64) + + assert params.start_epoch > 0, params.start_epoch + checkpoints = load_checkpoint_if_available( + params=params, model=model, model_avg=model_avg + ) + + model.to(device) + if world_size > 1: + logging.info("Using DDP") + model = DDP(model, device_ids=[rank], find_unused_parameters=True) + + if params.multi_optim: + logging.info("Using seperate optimizers over encoder, decoder ...") + + enc_param = [] + enc_names = [] + + dec_names = [] + dec_param = [] + + for n, p in model.named_parameters(): + name = n.split('.')[1] + if name == 'encoder' and 'feature_extractor' not in n: + enc_names.append(n) + enc_param.append(p) + elif 'ctc_output' in n: + enc_names.append(n) + enc_param.append(p) + elif 'feature_extractor' not in n: + dec_names.append(n) + dec_param.append(p) + + optimizer_enc = ScaledAdam( + enc_param, + lr=params.peak_enc_lr, + clipping_scale=None, + parameters_names=[enc_names], + ) + optimizer_dec = ScaledAdam( + dec_param, + lr=params.peak_dec_lr, + clipping_scale=5.0, + parameters_names=[dec_names], + ) + + scheduler_enc = Eden(optimizer_enc, params.lr_batches, params.lr_epochs) + scheduler_dec = Eden(optimizer_dec, params.lr_batches, params.lr_epochs) + optimizer = [optimizer_enc, optimizer_dec] + scheduler = [scheduler_enc, scheduler_dec] + + else: + parameters_names = [] + parameters_names.append( + [name_param_pair[0] for name_param_pair in model.named_parameters()] + ) + + logging.info(f"len name = {len(parameters_names)}") + logging.info(f"len param = {len(list(model.parameters()))}") + + optimizer = ScaledAdam( + model.parameters(), + lr=params.base_lr, + clipping_scale=2.0, + parameters_names=parameters_names, + ) + + scheduler = Eden(optimizer, params.lr_batches, params.lr_epochs) + + if checkpoints and ("optimizer" in checkpoints or "optimizer_enc" in checkpoints): + if params.multi_optim: + logging.info("Loading optimizer state dict") + optimizer_enc.load_state_dict(checkpoints["optimizer_enc"]) + optimizer_dec.load_state_dict(checkpoints["optimizer_dec"]) + + else: + logging.info("Loading optimizer state dict") + optimizer.load_state_dict(checkpoints["optimizer"]) + + if checkpoints: + if ( + params.multi_optim + and "scheduler_enc" in checkpoints + and checkpoints["scheduler_enc"] is not None + ): + logging.info("Loading enc/dec scheduler state dict") + scheduler_enc.load_state_dict(checkpoints["scheduler_enc"]) + scheduler_dec.load_state_dict(checkpoints["scheduler_dec"]) + else: + logging.info("Loading scheduler state dict") + scheduler.load_state_dict(checkpoints["scheduler"]) + + if params.print_diagnostics: + opts = diagnostics.TensorDiagnosticOptions( + 2**22 + ) # allow 4 megabytes per sub-module + diagnostic = diagnostics.attach_diagnostics(model, opts) + + if params.inf_check: + register_inf_check_hooks(model) + + librispeech = LibriSpeechAsrDataModule(args) + + train_cuts = librispeech.train_clean_100_cuts() + if params.full_libri: + train_cuts += librispeech.train_clean_360_cuts() + train_cuts += librispeech.train_other_500_cuts() + + def remove_short_and_long_utt(c: Cut): + # Keep only utterances with duration between 1 second and 20 seconds + # + # Caution: There is a reason to select 20.0 here. Please see + # ../local/display_manifest_statistics.py + # + # You should use ../local/display_manifest_statistics.py to get + # an utterance duration distribution for your dataset to select + # the threshold + return 1.0 <= c.duration <= 20.0 + + train_cuts = train_cuts.filter(remove_short_and_long_utt) + + if params.start_batch > 0 and checkpoints and "sampler" in checkpoints: + # We only load the sampler's state dict when it loads a checkpoint + # saved in the middle of an epoch + sampler_state_dict = checkpoints["sampler"] + else: + sampler_state_dict = None + + train_dl = librispeech.train_dataloaders( + train_cuts, sampler_state_dict=sampler_state_dict + ) + + valid_cuts = librispeech.dev_clean_cuts() + valid_cuts += librispeech.dev_other_cuts() + valid_dl = librispeech.valid_dataloaders(valid_cuts) + + ''' + if not params.print_diagnostics: + scan_pessimistic_batches_for_oom( + model=model, + train_dl=train_dl, + optimizer=optimizer, + sp=sp, + params=params, + ) + ''' + + scaler = GradScaler(enabled=params.use_fp16, init_scale=1.0) + if checkpoints and "grad_scaler" in checkpoints: + logging.info("Loading grad scaler state dict") + scaler.load_state_dict(checkpoints["grad_scaler"]) + + for epoch in range(params.start_epoch, params.num_epochs + 1): + if params.multi_optim: + scheduler_enc.step_epoch(epoch - 1) + scheduler_dec.step_epoch(epoch - 1) + else: + scheduler.step_epoch(epoch - 1) + fix_random_seed(params.seed + epoch - 1) + train_dl.sampler.set_epoch(epoch - 1) + + if tb_writer is not None: + tb_writer.add_scalar("train/epoch", epoch, params.batch_idx_train) + + params.cur_epoch = epoch + + train_one_epoch( + params=params, + model=model, + model_avg=model_avg, + optimizer=optimizer, + scheduler=scheduler, + sp=sp, + train_dl=train_dl, + valid_dl=valid_dl, + scaler=scaler, + tb_writer=tb_writer, + world_size=world_size, + rank=rank, + wb=wb, + ) + + if params.print_diagnostics: + diagnostic.print_diagnostics() + break + + ''' + if epoch % 50 == 0: + save_checkpoint( + params=params, + model=model, + model_avg=model_avg, + optimizer=optimizer, + scheduler=scheduler, + sampler=train_dl.sampler, + scaler=scaler, + rank=rank, + ) + ''' + + logging.info("Done!") + + if world_size > 1: + torch.distributed.barrier() + cleanup_dist() + + +def run_adapter(rank, world_size, args, wb=None): + """ + Args: + rank: + It is a value between 0 and `world_size-1`, which is + passed automatically by `mp.spawn()` in :func:`main`. + The node with rank 0 is responsible for saving checkpoint. + world_size: + Number of GPUs for DDP training. + args: + The return value of get_parser().parse_args() + """ + params = get_params() + params.update(vars(args)) + + fix_random_seed(params.seed) + if world_size > 1: + setup_dist(rank, world_size, params.master_port) + + setup_logger(f"{params.exp_dir}/log/log-train") + logging.info("Training started") + + if args.tensorboard and rank == 0: + tb_writer = SummaryWriter(log_dir=f"{params.exp_dir}/tensorboard") + else: + tb_writer = None + + device = torch.device("cpu") + if torch.cuda.is_available(): + device = torch.device("cuda", rank) + logging.info(f"Device: {device}") + + sp = spm.SentencePieceProcessor() + sp.load(params.bpe_model) + + # is defined in local/train_bpe_model.py + params.blank_id = sp.piece_to_id("") + params.vocab_size = sp.get_piece_size() + + logging.info(params) + + logging.info("About to create model") + model = get_transducer_model(params) + + num_param = sum([p.numel() if p.requires_grad else 0 for p in model.parameters()]) + logging.info(f"Number of model parameters: {num_param}") + + assert params.save_every_n >= params.average_period + model_avg: Optional[nn.Module] = None + if rank == 0: + # model_avg is only used with rank 0 + model_avg = copy.deepcopy(model).to(torch.float64) + + assert params.start_epoch > 0, params.start_epoch + checkpoints = load_checkpoint_if_available( + params=params, model=model, model_avg=model_avg + ) + + model.to(device) + if world_size > 1: + logging.info("Using DDP") + model = DDP(model, device_ids=[rank], find_unused_parameters=True) + + adapter_names = [] + adapter_param = [] + for n, p in model.named_parameters(): + if 'adapters' in n:# or 'joiner' in n or 'simple' in n or 'ctc' in n: + adapter_names.append(n) + adapter_param.append(p) + elif 'joiner' in n or 'simple' in n or 'ctc' in n: + p.requires_grad = True + else: + p.requires_grad = False + + optimizer_adapter = ScaledAdam( + adapter_param, + lr=params.adapter_lr, + clipping_scale=5.0, + parameters_names=[adapter_names], + ) + scheduler_adapter = Eden(optimizer_adapter, 10000, 7) #params.lr_batche, params.lr_epochs) + + optimizer, scheduler = optimizer_adapter, scheduler_adapter + + librispeech = LibriSpeechAsrDataModule(args) + + ''' + if params.hpo: + train_cuts = librispeech.train_clean_10_cuts(option=params.gender) + else: + train_cuts = librispeech.train_clean_100_cuts(option=params.gender) + if params.full_libri: + train_cuts += librispeech.train_clean_360_cuts(option=params.gender) + train_cuts += librispeech.train_other_500_cuts(option=params.gender) + ''' + + #train_cuts = librispeech.train_clean_10_cuts(option='male') + #train_cuts = librispeech.test_clean_user(option='big') + train_cuts = librispeech.vox_cuts(option=params.spk_id) + + def remove_short_and_long_utt(c: Cut): + return 1.0 <= c.duration <= 20.0 + + train_cuts = train_cuts.filter(remove_short_and_long_utt) + + sampler_state_dict = None + + train_dl = librispeech.train_dataloaders( + train_cuts, sampler_state_dict=sampler_state_dict + ) + #train_dl = librispeech.test_dataloaders( + # train_cuts + #) + + ''' + print('\n'*5) + print('-'*30) + for batch in train_dl: + print(batch) + print('-'*30) + print('\n'*5) + exit() + ''' + + valid_cuts = librispeech.dev_clean_cuts(option=params.gender) + valid_cuts += librispeech.dev_other_cuts(option=params.gender) + valid_dl = librispeech.valid_dataloaders(valid_cuts) + + scaler = GradScaler(enabled=params.use_fp16, init_scale=1.0) + + for epoch in range(params.start_epoch, params.num_epochs + 1): + logging.info(f"update num : {params.batch_idx_train}") + scheduler.step_epoch(epoch - 1) + fix_random_seed(params.seed + epoch - 1) + train_dl.sampler.set_epoch(epoch - 1) + + if tb_writer is not None: + tb_writer.add_scalar("train/epoch", epoch, params.batch_idx_train) + + params.cur_epoch = epoch + + train_one_epoch( + params=params, + model=model, + model_avg=model_avg, + optimizer=optimizer, + scheduler=scheduler, + sp=sp, + train_dl=train_dl, + valid_dl=valid_dl, + scaler=scaler, + tb_writer=tb_writer, + world_size=world_size, + rank=rank, + wb=wb, + ) + + if params.print_diagnostics: + diagnostic.print_diagnostics() + break + + ''' + if epoch % 10 == 0: + save_checkpoint( + params=params, + model=model, + model_avg=model_avg, + optimizer=optimizer, + scheduler=scheduler, + sampler=train_dl.sampler, + scaler=scaler, + rank=rank, + ) + ''' + + logging.info("Done!") + + if world_size > 1: + torch.distributed.barrier() + cleanup_dist() + + +def display_and_save_batch( + batch: dict, + params: AttributeDict, + sp: spm.SentencePieceProcessor, +) -> None: + """Display the batch statistics and save the batch into disk. + + Args: + batch: + A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()` + for the content in it. + params: + Parameters for training. See :func:`get_params`. + sp: + The BPE model. + """ + from lhotse.utils import uuid4 + + filename = f"{params.exp_dir}/batch-{uuid4()}.pt" + logging.info(f"Saving batch to {filename}") + torch.save(batch, filename) + + supervisions = batch["supervisions"] + features = batch["inputs"] + + logging.info(f"features shape: {features.shape}") + + y = sp.encode(supervisions["text"], out_type=int) + num_tokens = sum(len(i) for i in y) + logging.info(f"num tokens: {num_tokens}") + + +def scan_pessimistic_batches_for_oom( + model: Union[nn.Module, DDP], + train_dl: torch.utils.data.DataLoader, + optimizer: torch.optim.Optimizer, + sp: spm.SentencePieceProcessor, + params: AttributeDict, +): + from lhotse.dataset import find_pessimistic_batches + + logging.info( + "Sanity check -- see if any of the batches in epoch 1 would cause OOM." + ) + batches, crit_values = find_pessimistic_batches(train_dl.sampler) + for criterion, cuts in batches.items(): + batch = train_dl.dataset[cuts] + try: + with torch.cuda.amp.autocast(enabled=params.use_fp16): + loss, _ = compute_loss( + params=params, + model=model, + sp=sp, + batch=batch, + is_training=True, + ) + loss.backward() + optimizer.zero_grad() + except Exception as e: + if "CUDA out of memory" in str(e): + logging.error( + "Your GPU ran out of memory with the current " + "max_duration setting. We recommend decreasing " + "max_duration and trying again.\n" + f"Failing criterion: {criterion} " + f"(={crit_values[criterion]}) ..." + ) + display_and_save_batch(batch, params=params, sp=sp) + raise + logging.info( + f"Maximum memory allocated so far is {torch.cuda.max_memory_allocated()//1000000}MB" + ) + + +def main(): + parser = get_parser() + LibriSpeechAsrDataModule.add_arguments(parser) + args = parser.parse_args() + if args.wandb: args.exp_dir = args.exp_dir + str(random.randint(0,400)) + args.exp_dir = Path(args.exp_dir) + + logging.info("save arguments to config.yaml...") + save_args(args) + + if args.wandb: wb = wandb.init(project="d2v-adapter", entity="dohe0342", config=vars(args)) + else: wb = None + + world_size = args.world_size + assert world_size >= 1 + if world_size > 1: + mp.spawn(run if not args.add_adapter else run_adapter, + args=(world_size, args, wb), + nprocs=world_size, + join=True + ) + else: + if args.add_adapter: run_adapter(rank=0, world_size=1, args=args, wb=wb) + else: run(rank=0, world_size=1, args=args, wb=wb) + +torch.set_num_threads(1) +torch.set_num_interop_threads(1) + +if __name__ == "__main__": + main() diff --git a/egs/tedlium3/ASR/pruned_transducer_stateless_d2v_v2/train_lora.py b/egs/tedlium3/ASR/pruned_transducer_stateless_d2v_v2/train_lora.py new file mode 100755 index 000000000..ecc1fad06 --- /dev/null +++ b/egs/tedlium3/ASR/pruned_transducer_stateless_d2v_v2/train_lora.py @@ -0,0 +1,1862 @@ +#!/usr/bin/env python3 +# Copyright 2021-2022 Xiaomi Corp. (authors: Fangjun Kuang, +# Wei Kang, +# Mingshuang Luo,) +# Zengwei Yao) +# +# See ../../../../LICENSE for clarification regarding multiple authors +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +""" +Usage: + +export CUDA_VISIBLE_DEVICES="0,1,2,3" + +./pruned_transducer_stateless7_ctc/train.py \ + --world-size 4 \ + --num-epochs 30 \ + --start-epoch 1 \ + --exp-dir pruned_transducer_stateless7_ctc/exp \ + --full-libri 1 \ + --max-duration 300 + +# For mix precision training: + +./pruned_transducer_stateless7_ctc/train.py \ + --world-size 4 \ + --num-epochs 30 \ + --start-epoch 1 \ + --use-fp16 1 \ + --exp-dir pruned_transducer_stateless7_ctc/exp \ + --full-libri 1 \ + --max-duration 550 + +# For d2v-T training: +export CUDA_VISIBLE_DEVICES="0,1,2,3,4,5,6,7" + +./pruned_transducer_stateless_d2v_v2/train.py \ + --wandb true \ + --input-strategy AudioSamples \ + --enable-spec-aug False \ + --multi-optim True \ + --world-size 8 \ + --num-epochs 30 \ + --start-epoch 1 \ + --full-libri 0 \ + --exp-dir ./pruned_transducer_stateless_d2v_v2/$1 \ + --max-duration 250 \ + --freeze-finetune-updates 2000 \ + --use-fp16 1 \ + --peak-enc-lr 0.001 \ + --peak-dec-lr 0.05 \ + --accum-grads 1 \ + --encoder-type d2v \ + --additional-block True \ + --encoder-dim 768 \ + --decoder-dim 768 \ + --joiner-dim 768 \ + --prune-range 20 \ + --context-size 2 \ + --ctc-loss-scale 0.2 + +""" + + +import random +import argparse +import copy +import logging +import warnings +from pathlib import Path +from shutil import copyfile +from typing import Any, Dict, Optional, Tuple, Union + +import k2 +import optim +import sentencepiece as spm +import torch +import torch.multiprocessing as mp +import torch.nn as nn +from asr_datamodule import LibriSpeechAsrDataModule +from decoder import Decoder +from joiner import Joiner +from lhotse.cut import Cut +from lhotse.dataset.sampling.base import CutSampler +from lhotse.utils import fix_random_seed +from model import Transducer +from optim import Eden, ScaledAdam +from torch import Tensor +from torch.cuda.amp import GradScaler +from torch.nn.parallel import DistributedDataParallel as DDP +from torch.utils.tensorboard import SummaryWriter +from zipformer import Zipformer +from data2vec_encoder import FairSeqData2VecEncoder +from data2vec_audio import LoRAModule + +from icefall import diagnostics +from icefall.checkpoint import remove_checkpoints +from icefall.checkpoint import update_averaged_model +from checkpoint import ( + save_checkpoint as save_checkpoint_impl, + save_checkpoint_with_global_batch_idx, + load_checkpoint +) +from icefall.dist import cleanup_dist, setup_dist +from icefall.env import get_env_info +from icefall.hooks import register_inf_check_hooks +from icefall.utils import ( + AttributeDict, + MetricsTracker, + encode_supervisions, + setup_logger, + str2bool, + save_args, +) + +import wandb +import fairseq + +LRSchedulerType = Union[torch.optim.lr_scheduler._LRScheduler, optim.LRScheduler] + + +class LoRAHook(): + def __init__(self, module): + self.hook = module.register_forward_hook(self.hook_fn) + self.lora = LoRAModule( + embedding_dim=768, + rank=6, + lora_alpha=10000., + ) + def hook_fn(self, module, input, output): + lora_out = self.lora(input[0]) + output += lora_out + + def save_checkpoint(self, i, iter_, save_dir): + if isinstance(self.lora, DDP): + lora = self.lora.module + torch.save(lora.state_dict(), f"{save_dir}/lora_{iter_}_{i}.pt") + + +def set_batch_count(model: Union[nn.Module, DDP], batch_count: float) -> None: + if isinstance(model, DDP): + # get underlying nn.Module + model = model.module + for module in model.modules(): + if hasattr(module, "batch_count"): + module.batch_count = batch_count + model.encoder.num_updates = int(batch_count) + + +def add_adapter_arguments(parser: argparse.ArgumentParser): + parser.add_argument( + "--add-adapter", + type=str2bool, + default=False, + help="add adapter to rep model's encoder" + ) + + parser.add_argument( + "--adapter-lr", + type=float, + default=0.0001, + help="adapter learning rate" + ) + + parser.add_argument( + "--gender", + type=str, + default='male', + help="select gender" + ) + + +def add_rep_arguments(parser: argparse.ArgumentParser): + parser.add_argument( + "--wandb", + type=str2bool, + default=True, + help="Use wandb for MLOps", + ) + parser.add_argument( + "--hpo", + type=str2bool, + default=False, + help="Use small db for HPO", + ) + + parser.add_argument( + "--accum-grads", + type=int, + default=1, + help="accum-grad num.", + ) + + parser.add_argument( + "--multi-optim", + type=str2bool, + default=True, + help="use sperate optimizer (enc / dec)", + ) + + parser.add_argument( + "--peak-enc-lr", + type=float, + default=0.0001, + help="The initial learning rate. This value should not need to be changed.", + ) + + parser.add_argument( + "--peak-dec-lr", + type=float, + default=0.001, + help="The initial learning rate. This value should not need to be changed.", + ) + + parser.add_argument( + "--encoder-type", + type=str, + default='d2v', + help="Type of encoder (e.g. conformer, w2v, d2v...", + ) + + parser.add_argument( + "--encoder-dim", + type=int, + default=768, + help="encoder embedding dimension", + ) + + parser.add_argument( + "--freeze-finetune-updates", + type=int, + default=0 + ) + + parser.add_argument( + "--additional-block", + type=str2bool, + default=True, + ) + + parser.add_argument( + "--decode-interval", + type=int, + default=200, + help="decode interval", + ) + + +def add_model_arguments(parser: argparse.ArgumentParser): + parser.add_argument( + "--num-encoder-layers", + type=str, + default="2,4,3,2,4", + help="Number of zipformer encoder layers, comma separated.", + ) + + parser.add_argument( + "--feedforward-dims", + type=str, + default="1024,1024,2048,2048,1024", + help="Feedforward dimension of the zipformer encoder layers, comma separated.", + ) + + parser.add_argument( + "--nhead", + type=str, + default="8,8,8,8,8", + help="Number of attention heads in the zipformer encoder layers.", + ) + + parser.add_argument( + "--encoder-dims", + type=str, + default="384,384,384,384,384", + help="Embedding dimension in the 2 blocks of zipformer encoder layers, comma separated", + ) + + parser.add_argument( + "--attention-dims", + type=str, + default="192,192,192,192,192", + help="""Attention dimension in the 2 blocks of zipformer encoder layers, comma separated; + not the same as embedding dimension.""", + ) + + parser.add_argument( + "--encoder-unmasked-dims", + type=str, + default="256,256,256,256,256", + help="Unmasked dimensions in the encoders, relates to augmentation during training. " + "Must be <= each of encoder_dims. Empirically, less than 256 seems to make performance " + " worse.", + ) + + parser.add_argument( + "--zipformer-downsampling-factors", + type=str, + default="1,2,4,8,2", + help="Downsampling factor for each stack of encoder layers.", + ) + + parser.add_argument( + "--cnn-module-kernels", + type=str, + default="31,31,31,31,31", + help="Sizes of kernels in convolution modules", + ) + + parser.add_argument( + "--decoder-dim", + type=int, + default=768, + help="Embedding dimension in the decoder model.", + ) + + parser.add_argument( + "--joiner-dim", + type=int, + default=768, + help="""Dimension used in the joiner model. + Outputs from the encoder and decoder model are projected + to this dimension before adding. + """, + ) + + +def get_parser(): + parser = argparse.ArgumentParser( + formatter_class=argparse.ArgumentDefaultsHelpFormatter + ) + + parser.add_argument( + "--world-size", + type=int, + default=1, + help="Number of GPUs for DDP training.", + ) + + parser.add_argument( + "--master-port", + type=int, + default=12354, + help="Master port to use for DDP training.", + ) + + parser.add_argument( + "--tensorboard", + type=str2bool, + default=True, + help="Should various information be logged in tensorboard.", + ) + + parser.add_argument( + "--num-epochs", + type=int, + default=30, + help="Number of epochs to train.", + ) + + parser.add_argument( + "--num-updates", + type=int, + default=5000, + help="Number of epochs to train.", + ) + + parser.add_argument( + "--start-epoch", + type=int, + default=1, + help="""Resume training from this epoch. It should be positive. + If larger than 1, it will load checkpoint from + exp-dir/epoch-{start_epoch-1}.pt + """, + ) + + parser.add_argument( + "--start-batch", + type=int, + default=0, + help="""If positive, --start-epoch is ignored and + it loads the checkpoint from exp-dir/checkpoint-{start_batch}.pt + """, + ) + + parser.add_argument( + "--exp-dir", + type=str, + default="pruned_transducer_stateless7_ctc/exp", + help="""The experiment dir. + It specifies the directory where all training related + files, e.g., checkpoints, log, etc, are saved + """, + ) + + parser.add_argument( + "--bpe-model", + type=str, + default="data/lang_bpe_500/bpe.model", + help="Path to the BPE model", + ) + + parser.add_argument( + "--base-lr", type=float, default=0.05, help="The base learning rate." + ) + + parser.add_argument( + "--lr-batches", + type=float, + default=5000, + help="""Number of steps that affects how rapidly the learning rate + decreases. We suggest not to change this.""", + ) + + parser.add_argument( + "--lr-epochs", + type=float, + default=3.5, + help="""Number of epochs that affects how rapidly the learning rate decreases. + """, + ) + + parser.add_argument( + "--context-size", + type=int, + default=2, + help="The context size in the decoder. 1 means bigram; 2 means tri-gram", + ) + + parser.add_argument( + "--prune-range", + type=int, + default=5, + help="The prune range for rnnt loss, it means how many symbols(context)" + "we are using to compute the loss", + ) + + parser.add_argument( + "--lm-scale", + type=float, + default=0.25, + help="The scale to smooth the loss with lm " + "(output of prediction network) part.", + ) + + parser.add_argument( + "--am-scale", + type=float, + default=0.0, + help="The scale to smooth the loss with am (output of encoder network) part.", + ) + + parser.add_argument( + "--simple-loss-scale", + type=float, + default=0.5, + help="To get pruning ranges, we will calculate a simple version" + "loss(joiner is just addition), this simple loss also uses for" + "training (as a regularization item). We will scale the simple loss" + "with this parameter before adding to the final loss.", + ) + + parser.add_argument( + "--ctc-loss-scale", + type=float, + default=0.2, + help="Scale for CTC loss.", + ) + + parser.add_argument( + "--seed", + type=int, + default=42, + help="The seed for random generators intended for reproducibility", + ) + + parser.add_argument( + "--print-diagnostics", + type=str2bool, + default=False, + help="Accumulate stats on activations, print them and exit.", + ) + + parser.add_argument( + "--inf-check", + type=str2bool, + default=False, + help="Add hooks to check for infinite module outputs and gradients.", + ) + + parser.add_argument( + "--save-every-n", + type=int, + default=200, + help="""Save checkpoint after processing this number of batches" + periodically. We save checkpoint to exp-dir/ whenever + params.batch_idx_train % save_every_n == 0. The checkpoint filename + has the form: f'exp-dir/checkpoint-{params.batch_idx_train}.pt' + Note: It also saves checkpoint to `exp-dir/epoch-xxx.pt` at the + end of each epoch where `xxx` is the epoch number counting from 0. + """, + ) + + parser.add_argument( + "--keep-last-k", + type=int, + default=30, + help="""Only keep this number of checkpoints on disk. + For instance, if it is 3, there are only 3 checkpoints + in the exp-dir with filenames `checkpoint-xxx.pt`. + It does not affect checkpoints with name `epoch-xxx.pt`. + """, + ) + + parser.add_argument( + "--average-period", + type=int, + default=10, + help="""Update the averaged model, namely `model_avg`, after processing + this number of batches. `model_avg` is a separate version of model, + in which each floating-point parameter is the average of all the + parameters from the start of training. Each time we take the average, + we do: `model_avg = model * (average_period / batch_idx_train) + + model_avg * ((batch_idx_train - average_period) / batch_idx_train)`. + """, + ) + + parser.add_argument( + "--use-fp16", + type=str2bool, + default=True, + help="Whether to use half precision training.", + ) + + add_model_arguments(parser) + add_rep_arguments(parser) + add_adapter_arguments(parser) + + return parser + + +def get_params() -> AttributeDict: + """Return a dict containing training parameters. + + All training related parameters that are not passed from the commandline + are saved in the variable `params`. + + Commandline options are merged into `params` after they are parsed, so + you can also access them via `params`. + + Explanation of options saved in `params`: + + - best_train_loss: Best training loss so far. It is used to select + the model that has the lowest training loss. It is + updated during the training. + + - best_valid_loss: Best validation loss so far. It is used to select + the model that has the lowest validation loss. It is + updated during the training. + + - best_train_epoch: It is the epoch that has the best training loss. + + - best_valid_epoch: It is the epoch that has the best validation loss. + + - batch_idx_train: Used to writing statistics to tensorboard. It + contains number of batches trained so far across + epochs. + + - log_interval: Print training loss if batch_idx % log_interval` is 0 + + - reset_interval: Reset statistics if batch_idx % reset_interval is 0 + + - valid_interval: Run validation if batch_idx % valid_interval is 0 + + - feature_dim: The model input dim. It has to match the one used + in computing features. + + - subsampling_factor: The subsampling factor for the model. + + - encoder_dim: Hidden dim for multi-head attention model. + + - num_decoder_layers: Number of decoder layer of transformer decoder. + + - warm_step: The warmup period that dictates the decay of the + scale on "simple" (un-pruned) loss. + """ + params = AttributeDict( + { + "best_train_loss": float("inf"), + "best_valid_loss": float("inf"), + "best_train_epoch": -1, + "best_valid_epoch": -1, + "batch_idx_train": 0, + "log_interval": 20, + "reset_interval": 200, + "valid_interval": 3000, # For the 100h subset, use 800 + # parameters for zipformer + "feature_dim": 80, + "subsampling_factor": 320, # not passed in, this is fixed. + # parameters for ctc loss + "beam_size": 10, + "use_double_scores": True, + "warm_step": 0, + #"warm_step": 4000, + #"warm_step": 3000, + "env_info": get_env_info(), + } + ) + + return params + + +def get_encoder_model(params: AttributeDict) -> nn.Module: + # TODO: We can add an option to switch between Zipformer and Transformer + def to_int_tuple(s: str): + return tuple(map(int, s.split(","))) + + if params.encoder_type == 'd2v': + encoder = FairSeqData2VecEncoder( + input_size=params.encoder_dim, + w2v_url='None', + output_size=params.encoder_dim, + freeze_finetune_updates=params.freeze_finetune_updates, + additional_block=params.additional_block, + ) + else: + encoder = Zipformer( + num_features=params.feature_dim, + output_downsampling_factor=2, + zipformer_downsampling_factors=to_int_tuple( + params.zipformer_downsampling_factors + ), + encoder_dims=to_int_tuple(params.encoder_dims), + attention_dim=to_int_tuple(params.attention_dims), + encoder_unmasked_dims=to_int_tuple(params.encoder_unmasked_dims), + nhead=to_int_tuple(params.nhead), + feedforward_dim=to_int_tuple(params.feedforward_dims), + cnn_module_kernels=to_int_tuple(params.cnn_module_kernels), + num_encoder_layers=to_int_tuple(params.num_encoder_layers), + ) + + return encoder + + +def get_decoder_model(params: AttributeDict) -> nn.Module: + decoder = Decoder( + vocab_size=params.vocab_size, + decoder_dim=params.decoder_dim, + blank_id=params.blank_id, + context_size=params.context_size, + ) + return decoder + + +def get_joiner_model(params: AttributeDict) -> nn.Module: + joiner = Joiner( + encoder_dim=params.encoder_dim if params.encoder_type == 'd2v' else int(params.encoder_dims.split(",")[-1]), + decoder_dim=params.decoder_dim, + joiner_dim=params.joiner_dim, + vocab_size=params.vocab_size, + ) + return joiner + + +def get_transducer_model(params: AttributeDict) -> nn.Module: + encoder = get_encoder_model(params) + decoder = get_decoder_model(params) + joiner = get_joiner_model(params) + + model = Transducer( + encoder=encoder, + decoder=decoder, + joiner=joiner, + encoder_dim=params.encoder_dim if params.encoder_type == 'd2v' else int(params.encoder_dims.split(",")[-1]), + decoder_dim=params.decoder_dim, + joiner_dim=params.joiner_dim, + vocab_size=params.vocab_size, + ) + return model + + +def load_checkpoint_if_available( + params: AttributeDict, + model: nn.Module, + model_avg: nn.Module = None, + optimizer: Optional[torch.optim.Optimizer] = None, + scheduler: Optional[LRSchedulerType] = None, +) -> Optional[Dict[str, Any]]: + """Load checkpoint from file. + + If params.start_batch is positive, it will load the checkpoint from + `params.exp_dir/checkpoint-{params.start_batch}.pt`. Otherwise, if + params.start_epoch is larger than 1, it will load the checkpoint from + `params.start_epoch - 1`. + + Apart from loading state dict for `model` and `optimizer` it also updates + `best_train_epoch`, `best_train_loss`, `best_valid_epoch`, + and `best_valid_loss` in `params`. + + Args: + params: + The return value of :func:`get_params`. + model: + The training model. + model_avg: + The stored model averaged from the start of training. + optimizer: + The optimizer that we are using. + scheduler: + The scheduler that we are using. + Returns: + Return a dict containing previously saved training info. + """ + if params.start_batch > 0: + filename = params.exp_dir / f"checkpoint-{params.start_batch}.pt" + elif params.start_epoch > 1: + filename = params.exp_dir / f"epoch-{params.start_epoch-1}.pt" + elif params.add_adapter: + filename = params.exp_dir / f"../d2v-base-T.pt" + else: + return None + + assert filename.is_file(), f"{filename} does not exist!" + + saved_params = load_checkpoint( + filename, + model=model, + model_avg=model_avg, + optimizer=optimizer, + scheduler=scheduler, + strict=True if not params.add_adapter else False, + ) + + keys = [ + "best_train_epoch", + "best_valid_epoch", + "batch_idx_train", + "best_train_loss", + "best_valid_loss", + ] + for k in keys: + params[k] = saved_params[k] + + if params.start_batch > 0: + if "cur_epoch" in saved_params: + params["start_epoch"] = saved_params["cur_epoch"] + + if "cur_batch_idx" in saved_params: + params["cur_batch_idx"] = saved_params["cur_batch_idx"] + + params.batch_idx_train = 0 + + return saved_params + + +def save_checkpoint( + params: AttributeDict, + model: Union[nn.Module, DDP], + model_avg: Optional[nn.Module] = None, + optimizer: Optional[torch.optim.Optimizer] = None, + scheduler: Optional[LRSchedulerType] = None, + sampler: Optional[CutSampler] = None, + scaler: Optional[GradScaler] = None, + rank: int = 0, +) -> None: + """Save model, optimizer, scheduler and training stats to file. + + Args: + params: + It is returned by :func:`get_params`. + model: + The training model. + model_avg: + The stored model averaged from the start of training. + optimizer: + The optimizer used in the training. + sampler: + The sampler for the training dataset. + scaler: + The scaler used for mix precision training. + """ + if rank != 0: + return + filename = params.exp_dir / f"epoch-{params.cur_epoch}.pt" + save_checkpoint_impl( + filename=filename, + model=model, + model_avg=model_avg, + params=params, + optimizer=optimizer, + scheduler=scheduler, + sampler=sampler, + scaler=scaler, + rank=rank, + ) + + if params.best_train_epoch == params.cur_epoch: + best_train_filename = params.exp_dir / "best-train-loss.pt" + copyfile(src=filename, dst=best_train_filename) + + if params.best_valid_epoch == params.cur_epoch: + best_valid_filename = params.exp_dir / "best-valid-loss.pt" + copyfile(src=filename, dst=best_valid_filename) + + +def compute_loss( + params: AttributeDict, + model: Union[nn.Module, DDP], + sp: spm.SentencePieceProcessor, + batch: dict, + is_training: bool, + decode: bool = False, +) -> Tuple[Tensor, MetricsTracker]: + """ + Compute transducer loss given the model and its inputs. + + Args: + params: + Parameters for training. See :func:`get_params`. + model: + The model for training. It is an instance of Zipformer in our case. + batch: + A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()` + for the content in it. + is_training: + True for training. False for validation. When it is True, this + function enables autograd during computation; when it is False, it + disables autograd. + warmup: a floating point value which increases throughout training; + values >= 1.0 are fully warmed up and have all modules present. + """ + device = model.device if isinstance(model, DDP) else next(model.parameters()).device + feature = batch["inputs"] + # at entry, feature is (N, T, C) + assert feature.ndim == 2 or feature.ndim == 3 + feature = feature.to(device) + + supervisions = batch["supervisions"] + + if feature.ndim == 2: + feature_lens = [] + for supervision in supervisions['cut']: + try: feature_lens.append(supervision.tracks[0].cut.recording.num_samples) + except: feature_lens.append(supervision.recording.num_samples) + feature_lens = torch.tensor(feature_lens) + + elif feature.ndim == 3: + feature_lens = supervisions["num_frames"].to(device) + + batch_idx_train = params.batch_idx_train + warm_step = params.warm_step + + texts = batch["supervisions"]["text"] + + token_ids = sp.encode(texts, out_type=int) + y = k2.RaggedTensor(token_ids).to(device) + + with torch.set_grad_enabled(is_training): + simple_loss, pruned_loss, ctc_output = model( + x=feature, + x_lens=feature_lens, + y=y, + prune_range=params.prune_range, + am_scale=params.am_scale, + lm_scale=params.lm_scale, + ) + + s = params.simple_loss_scale + # take down the scale on the simple loss from 1.0 at the start + # to params.simple_loss scale by warm_step. + simple_loss_scale = ( + s + if batch_idx_train >= warm_step + else 1.0 - (batch_idx_train / warm_step) * (1.0 - s) + ) + pruned_loss_scale = ( + 1.0 + if batch_idx_train >= warm_step + else 0.1 + 0.9 * (batch_idx_train / warm_step) + ) + + loss = simple_loss_scale * simple_loss + pruned_loss_scale * pruned_loss + + info = MetricsTracker() + + if params.ctc_loss_scale > 0: + # Compute ctc loss + + # NOTE: We need `encode_supervisions` to sort sequences with + # different duration in decreasing order, required by + # `k2.intersect_dense` called in `k2.ctc_loss` + with warnings.catch_warnings(): + warnings.simplefilter("ignore") + supervision_segments, token_ids = encode_supervisions( + supervisions, + subsampling_factor=params.subsampling_factor, + token_ids=token_ids, + ) + + # Works with a BPE model + decoding_graph = k2.ctc_graph(token_ids, modified=False, device=device) + dense_fsa_vec = k2.DenseFsaVec( + ctc_output, + supervision_segments, + allow_truncate=params.subsampling_factor - 1, + ) + + ctc_loss = k2.ctc_loss( + decoding_graph=decoding_graph, + dense_fsa_vec=dense_fsa_vec, + output_beam=params.beam_size, + reduction="sum", + use_double_scores=params.use_double_scores, + ) + assert ctc_loss.requires_grad == is_training + loss += params.ctc_loss_scale * ctc_loss + + info["ctc_loss"] = ctc_loss.detach().cpu().item() + + assert loss.requires_grad == is_training + + if decode: + model.eval() + with torch.no_grad(): + try: + hypos = model.module.decode( + x=feature, + x_lens=feature_lens, + y=y, + sp=sp + ) + except: + hypos = model.decode( + x=feature, + x_lens=feature_lens, + y=y, + sp=sp + ) + + logging.info(f'ref: {batch["supervisions"]["text"][0]}') + logging.info(f'hyp: {" ".join(hypos[0])}') + model.train() + + with warnings.catch_warnings(): + warnings.simplefilter("ignore") + info["frames"] = (feature_lens // params.subsampling_factor).sum().item() + + # Note: We use reduction=sum while computing the loss. + info["utterances"] = feature.size(0) + info["loss"] = loss.detach().cpu().item() + info["simple_loss"] = simple_loss.detach().cpu().item() + info["pruned_loss"] = pruned_loss.detach().cpu().item() + + return loss, info + + +def compute_validation_loss( + params: AttributeDict, + model: Union[nn.Module, DDP], + sp: spm.SentencePieceProcessor, + valid_dl: torch.utils.data.DataLoader, + world_size: int = 1, +) -> MetricsTracker: + """Run the validation process.""" + model.eval() + + tot_loss = MetricsTracker() + + for batch_idx, batch in enumerate(valid_dl): + loss, loss_info = compute_loss( + params=params, + model=model, + sp=sp, + batch=batch, + is_training=False, + ) + assert loss.requires_grad is False + tot_loss = tot_loss + loss_info + + if world_size > 1: + tot_loss.reduce(loss.device) + + loss_value = tot_loss["loss"] / tot_loss["utterances"] + if loss_value < params.best_valid_loss: + params.best_valid_epoch = params.cur_epoch + params.best_valid_loss = loss_value + + return tot_loss + + +def train_one_epoch( + params: AttributeDict, + model: Union[nn.Module, DDP], + optimizer: torch.optim.Optimizer or [torch.optim.Optimizer, torch.optim.Optimizer], + scheduler: LRSchedulerType or [LRSchedulerType, LRSchedulerType], + sp: spm.SentencePieceProcessor, + train_dl: torch.utils.data.DataLoader, + valid_dl: torch.utils.data.DataLoader, + scaler: GradScaler, + model_avg: Optional[nn.Module] = None, + tb_writer: Optional[SummaryWriter] = None, + world_size: int = 1, + rank: int = 0, + wb = None, + lora_modules = None, +) -> None: + """Train the model for one epoch. + + The training loss from the mean of all frames is saved in + `params.train_loss`. It runs the validation process every + `params.valid_interval` batches. + + Args: + params: + It is returned by :func:`get_params`. + model: + The model for training. + optimizer: + The optimizer we are using. + scheduler: + The learning rate scheduler, we call step() every step. + train_dl: + Dataloader for the training dataset. + valid_dl: + Dataloader for the validation dataset. + scaler: + The scaler used for mix precision training. + model_avg: + The stored model averaged from the start of training. + tb_writer: + Writer to write log messages to tensorboard. + world_size: + Number of nodes in DDP training. If it is 1, DDP is disabled. + rank: + The rank of the node in DDP training. If no DDP is used, it should + be set to 0. + """ + model.train() + + tot_loss = MetricsTracker() + + cur_batch_idx = params.get("cur_batch_idx", 0) + + if params.multi_optim: + optimizer_enc, optimizer_dec = optimizer[0], optimizer[1] + scheduler_enc, scheduler_dec = scheduler[0], scheduler[1] + + for batch_idx, batch in enumerate(train_dl): + if params.batch_idx_train > params.num_updates: + break + if batch_idx < cur_batch_idx: + continue + cur_batch_idx = batch_idx + + if batch_idx % params.accum_grads == 0: params.batch_idx_train += 1 + batch_size = len(batch["supervisions"]["text"]) + + try: + with torch.cuda.amp.autocast(enabled=params.use_fp16): + loss, loss_info = compute_loss( + params=params, + model=model, + sp=sp, + batch=batch, + is_training=True, + decode = True if batch_idx % params.decode_interval == 0 else False, + ) + + try: loss_info.reduce(loss.device) + except: pass + + numel = params.world_size / (params.accum_grads * loss_info["utterances"]) + loss *= numel ## normalize loss over utts(batch size) + + # summary stats + tot_loss = (tot_loss * (1 - 1 / params.reset_interval)) + loss_info + + # NOTE: We use reduction==sum and loss is computed over utterances + # in the batch and there is no normalization to it so far. + scaler.scale(loss).backward() + + if params.multi_optim and (batch_idx+1) % params.accum_grads == 0: + set_batch_count(model, params.batch_idx_train) + scheduler_enc.step_batch(params.batch_idx_train) + scheduler_dec.step_batch(params.batch_idx_train) + scaler.step(optimizer_enc) + scaler.step(optimizer_dec) + scaler.update() + optimizer_enc.zero_grad() + optimizer_dec.zero_grad() + elif not params.multi_optim and (batch_idx+1) % params.accum_grads == 0: + set_batch_count(model, params.batch_idx_train) + scheduler.step_batch(params.batch_idx_train) + scaler.step(optimizer) + scaler.update() + optimizer.zero_grad() + + except: # noqa + display_and_save_batch(batch, params=params, sp=sp) + raise + + if params.print_diagnostics and batch_idx == 5: + return + + ''' + if ( + rank == 0 + and params.batch_idx_train > 0 + and params.batch_idx_train % params.average_period == 0 + ): + update_averaged_model( + params=params, + model_cur=model, + model_avg=model_avg, + ) + ''' + + if ( + params.batch_idx_train > 0 + and params.batch_idx_train % params.save_every_n == 0 + ): + params.cur_batch_idx = batch_idx + ''' + save_checkpoint_with_global_batch_idx( + out_dir=params.exp_dir, + global_batch_idx=params.batch_idx_train, + model=model, + model_avg=model_avg, + params=params, + optimizer=optimizer, + scheduler=scheduler, + sampler=train_dl.sampler, + scaler=scaler, + rank=rank, + ) + ''' + del params.cur_batch_idx + + if rank == 0: + for i, lora in enumerate(lora_modules): + lora.save_checkpoint(i, params.batch_idx_train, params.exp_dir) + + ''' + remove_checkpoints( + out_dir=params.exp_dir, + topk=params.keep_last_k, + rank=rank, + ) + ''' + + if batch_idx % 100 == 0 and params.use_fp16: + # If the grad scale was less than 1, try increasing it. The _growth_interval + # of the grad scaler is configurable, but we can't configure it to have different + # behavior depending on the current grad scale. + cur_grad_scale = scaler._scale.item() + ''' + if cur_grad_scale < 1.0 or (cur_grad_scale < 8.0 and batch_idx % 400 == 0): + scaler.update(cur_grad_scale * 2.0) + ''' + if cur_grad_scale < 0.01: + logging.warning(f"Grad scale is small: {cur_grad_scale}") + if cur_grad_scale < 1.0e-05: + wb.log({"valid/loss": 10000}) + raise RuntimeError( + f"grad_scale is too small, exiting: {cur_grad_scale}" + ) + + #if params.batch_idx_train > 4000 and loss > 300 and params.wandb: + # wb.log({"valid/loss": 10000}) + # raise RuntimeError( + # f"divergence... exiting: loss={loss}" + # ) + + if batch_idx % (params.log_interval*params.accum_grads) == 0: + #for n, p in model.named_parameters(): + # if 'adapter' in n: + # print(p) + if params.multi_optim: + cur_enc_lr = scheduler_enc.get_last_lr()[0] + cur_dec_lr = scheduler_dec.get_last_lr()[0] + cur_grad_scale = scaler._scale.item() if params.use_fp16 else 1.0 + + logging.info( + f"Epoch {params.cur_epoch}, " + f"batch {batch_idx}, loss[{loss_info}], " + f"tot_loss[{tot_loss}], batch size: {batch_size}, " + f"enc_lr: {cur_enc_lr:.2e}, " + f"dec_lr: {cur_dec_lr:.2e}, " + + (f"grad_scale: {scaler._scale.item()}" if params.use_fp16 else "") + ) + + else: + cur_lr = scheduler.get_last_lr()[0] + cur_grad_scale = scaler._scale.item() if params.use_fp16 else 1.0 + + logging.info( + f"Epoch {params.cur_epoch}, " + f"batch {batch_idx}, loss[{loss_info}], " + f"tot_loss[{tot_loss}], batch size: {batch_size}, " + f"lr: {cur_lr:.2e}, " + + (f"grad_scale: {scaler._scale.item()}" if params.use_fp16 else "") + ) + + if tb_writer is not None: + if params.multi_optim: + tb_writer.add_scalar( + "train/enc_learning_rate", cur_enc_lr, params.batch_idx_train + ) + tb_writer.add_scalar( + "train/dec_learning_rate", cur_dec_lr, params.batch_idx_train + ) + + else: + tb_writer.add_scalar( + "train/learning_rate", cur_lr, params.batch_idx_train + ) + + loss_info.write_summary( + tb_writer, "train/current_", params.batch_idx_train + ) + tot_loss.write_summary(tb_writer, "train/tot_", params.batch_idx_train) + if params.use_fp16: + tb_writer.add_scalar( + "train/grad_scale", + cur_grad_scale, + params.batch_idx_train, + ) + + if wb is not None and rank == 0: + wb.log({"train/loss": loss_info["loss"]*numel}) + wb.log({"train/simple_loss": loss_info["simple_loss"]*numel}) + wb.log({"train/pruned_loss": loss_info["pruned_loss"]*numel}) + wb.log({"train/ctc_loss": loss_info["ctc_loss"]*numel}) + + ''' + logging.info("Computing validation loss") + valid_info = compute_validation_loss( + params=params, + model=model, + sp=sp, + valid_dl=valid_dl, + world_size=world_size, + ) + model.train() + logging.info(f"Epoch {params.cur_epoch}, validation: {valid_info}") + logging.info( + f"Maximum memory allocated so far is {torch.cuda.max_memory_allocated()//1000000}MB" + ) + if tb_writer is not None: + valid_info.write_summary( + tb_writer, "train/valid_", params.batch_idx_train + ) + + if wb is not None and rank == 0: + numel = 1 / (params.accum_grads * valid_info["utterances"]) + #wb.log({"valid/loss": valid_info["loss"]*numel}) + wb.log({"valid/loss": numel*(valid_info["simple_loss"] + +valid_info["pruned_loss"] + +valid_info["ctc_loss"] + )}) + wb.log({"valid/simple_loss": valid_info["simple_loss"]*numel}) + wb.log({"valid/pruned_loss": valid_info["pruned_loss"]*numel}) + wb.log({"valid/ctc_loss": valid_info["ctc_loss"]*numel}) + ''' + loss_value = tot_loss["loss"] / tot_loss["utterances"] + params.train_loss = loss_value + if params.train_loss < params.best_train_loss: + params.best_train_epoch = params.cur_epoch + params.best_train_loss = params.train_loss + + +def run(rank, world_size, args, wb=None): + """ + Args: + rank: + It is a value between 0 and `world_size-1`, which is + passed automatically by `mp.spawn()` in :func:`main`. + The node with rank 0 is responsible for saving checkpoint. + world_size: + Number of GPUs for DDP training. + args: + The return value of get_parser().parse_args() + """ + params = get_params() + params.update(vars(args)) + #params.warm_step *= params.accum_grads + + fix_random_seed(params.seed) + if world_size > 1: + setup_dist(rank, world_size, params.master_port) + + setup_logger(f"{params.exp_dir}/log/log-train") + logging.info("Training started") + + if args.tensorboard and rank == 0: + tb_writer = SummaryWriter(log_dir=f"{params.exp_dir}/tensorboard") + else: + tb_writer = None + + device = torch.device("cpu") + if torch.cuda.is_available(): + device = torch.device("cuda", rank) + logging.info(f"Device: {device}") + + sp = spm.SentencePieceProcessor() + sp.load(params.bpe_model) + + # is defined in local/train_bpe_model.py + params.blank_id = sp.piece_to_id("") + params.vocab_size = sp.get_piece_size() + + logging.info(params) + + logging.info("About to create model") + model = get_transducer_model(params) + logging.info(model) + + num_param = sum([p.numel() for p in model.parameters()]) + logging.info(f"Number of model parameters: {num_param}") + + assert params.save_every_n >= params.average_period + model_avg: Optional[nn.Module] = None + if rank == 0: + # model_avg is only used with rank 0 + model_avg = copy.deepcopy(model).to(torch.float64) + + assert params.start_epoch > 0, params.start_epoch + checkpoints = load_checkpoint_if_available( + params=params, model=model, model_avg=model_avg + ) + + model.to(device) + if world_size > 1: + logging.info("Using DDP") + model = DDP(model, device_ids=[rank], find_unused_parameters=True) + + if params.multi_optim: + logging.info("Using seperate optimizers over encoder, decoder ...") + + enc_param = [] + enc_names = [] + + dec_names = [] + dec_param = [] + + for n, p in model.named_parameters(): + name = n.split('.')[1] + if name == 'encoder' and 'feature_extractor' not in n: + enc_names.append(n) + enc_param.append(p) + elif 'ctc_output' in n: + enc_names.append(n) + enc_param.append(p) + elif 'feature_extractor' not in n: + dec_names.append(n) + dec_param.append(p) + + optimizer_enc = ScaledAdam( + enc_param, + lr=params.peak_enc_lr, + clipping_scale=None, + parameters_names=[enc_names], + ) + optimizer_dec = ScaledAdam( + dec_param, + lr=params.peak_dec_lr, + clipping_scale=5.0, + parameters_names=[dec_names], + ) + + scheduler_enc = Eden(optimizer_enc, params.lr_batches, params.lr_epochs) + scheduler_dec = Eden(optimizer_dec, params.lr_batches, params.lr_epochs) + optimizer = [optimizer_enc, optimizer_dec] + scheduler = [scheduler_enc, scheduler_dec] + + else: + parameters_names = [] + parameters_names.append( + [name_param_pair[0] for name_param_pair in model.named_parameters()] + ) + + logging.info(f"len name = {len(parameters_names)}") + logging.info(f"len param = {len(list(model.parameters()))}") + + optimizer = ScaledAdam( + model.parameters(), + lr=params.base_lr, + clipping_scale=2.0, + parameters_names=parameters_names, + ) + + scheduler = Eden(optimizer, params.lr_batches, params.lr_epochs) + + if checkpoints and ("optimizer" in checkpoints or "optimizer_enc" in checkpoints): + if params.multi_optim: + logging.info("Loading optimizer state dict") + optimizer_enc.load_state_dict(checkpoints["optimizer_enc"]) + optimizer_dec.load_state_dict(checkpoints["optimizer_dec"]) + + else: + logging.info("Loading optimizer state dict") + optimizer.load_state_dict(checkpoints["optimizer"]) + + if checkpoints: + if ( + params.multi_optim + and "scheduler_enc" in checkpoints + and checkpoints["scheduler_enc"] is not None + ): + logging.info("Loading enc/dec scheduler state dict") + scheduler_enc.load_state_dict(checkpoints["scheduler_enc"]) + scheduler_dec.load_state_dict(checkpoints["scheduler_dec"]) + else: + logging.info("Loading scheduler state dict") + scheduler.load_state_dict(checkpoints["scheduler"]) + + if params.print_diagnostics: + opts = diagnostics.TensorDiagnosticOptions( + 2**22 + ) # allow 4 megabytes per sub-module + diagnostic = diagnostics.attach_diagnostics(model, opts) + + if params.inf_check: + register_inf_check_hooks(model) + + librispeech = LibriSpeechAsrDataModule(args) + + train_cuts = librispeech.train_clean_100_cuts() + if params.full_libri: + train_cuts += librispeech.train_clean_360_cuts() + train_cuts += librispeech.train_other_500_cuts() + + def remove_short_and_long_utt(c: Cut): + # Keep only utterances with duration between 1 second and 20 seconds + # + # Caution: There is a reason to select 20.0 here. Please see + # ../local/display_manifest_statistics.py + # + # You should use ../local/display_manifest_statistics.py to get + # an utterance duration distribution for your dataset to select + # the threshold + return 1.0 <= c.duration <= 20.0 + + train_cuts = train_cuts.filter(remove_short_and_long_utt) + + if params.start_batch > 0 and checkpoints and "sampler" in checkpoints: + # We only load the sampler's state dict when it loads a checkpoint + # saved in the middle of an epoch + sampler_state_dict = checkpoints["sampler"] + else: + sampler_state_dict = None + + train_dl = librispeech.train_dataloaders( + train_cuts, sampler_state_dict=sampler_state_dict + ) + + valid_cuts = librispeech.dev_clean_cuts() + valid_cuts += librispeech.dev_other_cuts() + valid_dl = librispeech.valid_dataloaders(valid_cuts) + + ''' + if not params.print_diagnostics: + scan_pessimistic_batches_for_oom( + model=model, + train_dl=train_dl, + optimizer=optimizer, + sp=sp, + params=params, + ) + ''' + + scaler = GradScaler(enabled=params.use_fp16, init_scale=1.0) + if checkpoints and "grad_scaler" in checkpoints: + logging.info("Loading grad scaler state dict") + scaler.load_state_dict(checkpoints["grad_scaler"]) + + for epoch in range(params.start_epoch, params.num_epochs + 1): + if params.multi_optim: + scheduler_enc.step_epoch(epoch - 1) + scheduler_dec.step_epoch(epoch - 1) + else: + scheduler.step_epoch(epoch - 1) + fix_random_seed(params.seed + epoch - 1) + train_dl.sampler.set_epoch(epoch - 1) + + if tb_writer is not None: + tb_writer.add_scalar("train/epoch", epoch, params.batch_idx_train) + + params.cur_epoch = epoch + + train_one_epoch( + params=params, + model=model, + model_avg=model_avg, + optimizer=optimizer, + scheduler=scheduler, + sp=sp, + train_dl=train_dl, + valid_dl=valid_dl, + scaler=scaler, + tb_writer=tb_writer, + world_size=world_size, + rank=rank, + wb=wb, + ) + + if params.print_diagnostics: + diagnostic.print_diagnostics() + break + + ''' + if epoch % 50 == 0: + save_checkpoint( + params=params, + model=model, + model_avg=model_avg, + optimizer=optimizer, + scheduler=scheduler, + sampler=train_dl.sampler, + scaler=scaler, + rank=rank, + ) + ''' + + logging.info("Done!") + + if world_size > 1: + torch.distributed.barrier() + cleanup_dist() + + +def run_adapter(rank, world_size, args, wb=None): + """ + Args: + rank: + It is a value between 0 and `world_size-1`, which is + passed automatically by `mp.spawn()` in :func:`main`. + The node with rank 0 is responsible for saving checkpoint. + world_size: + Number of GPUs for DDP training. + args: + The return value of get_parser().parse_args() + """ + params = get_params() + params.update(vars(args)) + + fix_random_seed(params.seed) + if world_size > 1: + setup_dist(rank, world_size, params.master_port) + + setup_logger(f"{params.exp_dir}/log/log-train") + logging.info("Training started") + + if args.tensorboard and rank == 0: + tb_writer = SummaryWriter(log_dir=f"{params.exp_dir}/tensorboard") + else: + tb_writer = None + + device = torch.device("cpu") + if torch.cuda.is_available(): + device = torch.device("cuda", rank) + logging.info(f"Device: {device}") + + sp = spm.SentencePieceProcessor() + sp.load(params.bpe_model) + + # is defined in local/train_bpe_model.py + params.blank_id = sp.piece_to_id("") + params.vocab_size = sp.get_piece_size() + + logging.info(params) + + logging.info("About to create model") + model = get_transducer_model(params) + + num_param = sum([p.numel() if p.requires_grad else 0 for p in model.parameters()]) + logging.info(f"Number of model parameters: {num_param}") + + assert params.save_every_n >= params.average_period + model_avg: Optional[nn.Module] = None + if rank == 0: + # model_avg is only used with rank 0 + model_avg = copy.deepcopy(model).to(torch.float64) + + assert params.start_epoch > 0, params.start_epoch + checkpoints = load_checkpoint_if_available( + params=params, model=model, model_avg=model_avg + ) + + model.to(device) + if world_size > 1: + logging.info("Using DDP") + model = DDP(model, device_ids=[rank], find_unused_parameters=True) + + lora_modules = [] + for modules in model.modules(): + if isinstance(modules, fairseq.modules.multihead_attention.MultiheadAttention): + for module in modules.modules(): + if isinstance(module, torch.nn.Linear): + lora_modules.append(LoRAHook(module)) + + if world_size > 1: + logging.info("Using DDP for LoRA") + for lora in lora_modules: + lora.lora = lora.lora.to(device) + lora.lora = DDP(lora.lora, device_ids=[rank], find_unused_parameters=False) + + adapter_names = [] + adapter_param = [] + for i, lora in enumerate(lora_modules): + for n, p in lora.lora.named_parameters(): + new_n = str(i) + n + adapter_names.append(new_n) + adapter_param.append(p) + + ''' + for n, p in model.named_parameters(): + if 'joiner' in n or 'simple' in n or 'ctc' in n: + adapter_names.append(n) + adapter_param.append(p) + p.requires_grad = True + else: + p.requires_grad = False + ''' + #for lora in lora_modules: + # print(lora.lora.state_dict()) + #print(adapter_names) + #exit() + ''' + for n, p in model.named_parameters(): + print(n) + if 'adapters' in n:# or 'joiner' in n or 'simple' in n or 'ctc' in n: + adapter_names.append(n) + adapter_param.append(p) + elif 'joiner' in n or 'simple' in n or 'ctc' in n: + p.requires_grad = True + else: + p.requires_grad = False + ''' + optimizer_adapter = ScaledAdam( + adapter_param, + lr=params.adapter_lr, + clipping_scale=5.0, + parameters_names=[adapter_names], + ) + scheduler_adapter = Eden(optimizer_adapter, 10000, 7) #params.lr_batche, params.lr_epochs) + + optimizer, scheduler = optimizer_adapter, scheduler_adapter + + librispeech = LibriSpeechAsrDataModule(args) + + ''' + if params.hpo: + train_cuts = librispeech.train_clean_10_cuts(option=params.gender) + else: + train_cuts = librispeech.train_clean_100_cuts(option=params.gender) + if params.full_libri: + train_cuts += librispeech.train_clean_360_cuts(option=params.gender) + train_cuts += librispeech.train_other_500_cuts(option=params.gender) + ''' + + #train_cuts = librispeech.train_clean_10_cuts(option='male') + #train_cuts = librispeech.test_clean_user(option='big') + train_cuts = librispeech.vox_cuts(option=params.spk_id) + + def remove_short_and_long_utt(c: Cut): + return 1.0 <= c.duration <= 20.0 + + train_cuts = train_cuts.filter(remove_short_and_long_utt) + + sampler_state_dict = None + + train_dl = librispeech.train_dataloaders( + train_cuts, sampler_state_dict=sampler_state_dict + ) + #train_dl = librispeech.test_dataloaders( + # train_cuts + #) + + ''' + print('\n'*5) + print('-'*30) + for batch in train_dl: + print(batch) + print('-'*30) + print('\n'*5) + exit() + ''' + + valid_cuts = librispeech.dev_clean_cuts(option=params.gender) + valid_cuts += librispeech.dev_other_cuts(option=params.gender) + valid_dl = librispeech.valid_dataloaders(valid_cuts) + + scaler = GradScaler(enabled=params.use_fp16, init_scale=1.0) + + for epoch in range(params.start_epoch, params.num_epochs + 1): + logging.info(f"update num : {params.batch_idx_train}") + scheduler.step_epoch(epoch - 1) + fix_random_seed(params.seed + epoch - 1) + train_dl.sampler.set_epoch(epoch - 1) + + if tb_writer is not None: + tb_writer.add_scalar("train/epoch", epoch, params.batch_idx_train) + + params.cur_epoch = epoch + + train_one_epoch( + params=params, + model=model, + model_avg=model_avg, + optimizer=optimizer, + scheduler=scheduler, + sp=sp, + train_dl=train_dl, + valid_dl=valid_dl, + scaler=scaler, + tb_writer=tb_writer, + world_size=world_size, + rank=rank, + wb=wb, + lora_modules=lora_modules, + ) + + if params.print_diagnostics: + diagnostic.print_diagnostics() + break + + ''' + if epoch % 10 == 0: + save_checkpoint( + params=params, + model=model, + model_avg=model_avg, + optimizer=optimizer, + scheduler=scheduler, + sampler=train_dl.sampler, + scaler=scaler, + rank=rank, + ) + ''' + + logging.info("Done!") + + if world_size > 1: + torch.distributed.barrier() + cleanup_dist() + + +def display_and_save_batch( + batch: dict, + params: AttributeDict, + sp: spm.SentencePieceProcessor, +) -> None: + """Display the batch statistics and save the batch into disk. + + Args: + batch: + A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()` + for the content in it. + params: + Parameters for training. See :func:`get_params`. + sp: + The BPE model. + """ + from lhotse.utils import uuid4 + + filename = f"{params.exp_dir}/batch-{uuid4()}.pt" + logging.info(f"Saving batch to {filename}") + torch.save(batch, filename) + + supervisions = batch["supervisions"] + features = batch["inputs"] + + logging.info(f"features shape: {features.shape}") + + y = sp.encode(supervisions["text"], out_type=int) + num_tokens = sum(len(i) for i in y) + logging.info(f"num tokens: {num_tokens}") + + +def scan_pessimistic_batches_for_oom( + model: Union[nn.Module, DDP], + train_dl: torch.utils.data.DataLoader, + optimizer: torch.optim.Optimizer, + sp: spm.SentencePieceProcessor, + params: AttributeDict, +): + from lhotse.dataset import find_pessimistic_batches + + logging.info( + "Sanity check -- see if any of the batches in epoch 1 would cause OOM." + ) + batches, crit_values = find_pessimistic_batches(train_dl.sampler) + for criterion, cuts in batches.items(): + batch = train_dl.dataset[cuts] + try: + with torch.cuda.amp.autocast(enabled=params.use_fp16): + loss, _ = compute_loss( + params=params, + model=model, + sp=sp, + batch=batch, + is_training=True, + ) + loss.backward() + optimizer.zero_grad() + except Exception as e: + if "CUDA out of memory" in str(e): + logging.error( + "Your GPU ran out of memory with the current " + "max_duration setting. We recommend decreasing " + "max_duration and trying again.\n" + f"Failing criterion: {criterion} " + f"(={crit_values[criterion]}) ..." + ) + display_and_save_batch(batch, params=params, sp=sp) + raise + logging.info( + f"Maximum memory allocated so far is {torch.cuda.max_memory_allocated()//1000000}MB" + ) + + +def main(): + parser = get_parser() + LibriSpeechAsrDataModule.add_arguments(parser) + args = parser.parse_args() + if args.wandb: args.exp_dir = args.exp_dir + str(random.randint(0,400)) + args.exp_dir = Path(args.exp_dir) + + logging.info("save arguments to config.yaml...") + save_args(args) + + if args.wandb: wb = wandb.init(project="d2v-adapter", entity="dohe0342", config=vars(args)) + else: wb = None + + world_size = args.world_size + assert world_size >= 1 + if world_size > 1: + mp.spawn(run if not args.add_adapter else run_adapter, + args=(world_size, args, wb), + nprocs=world_size, + join=True + ) + else: + if args.add_adapter: run_adapter(rank=0, world_size=1, args=args, wb=wb) + else: run(rank=0, world_size=1, args=args, wb=wb) + +torch.set_num_threads(1) +torch.set_num_interop_threads(1) + +if __name__ == "__main__": + main() diff --git a/egs/tedlium3/ASR/pruned_transducer_stateless_d2v_v2/train_uda.py b/egs/tedlium3/ASR/pruned_transducer_stateless_d2v_v2/train_uda.py new file mode 100755 index 000000000..c25ab54aa --- /dev/null +++ b/egs/tedlium3/ASR/pruned_transducer_stateless_d2v_v2/train_uda.py @@ -0,0 +1,1960 @@ +#!/usr/bin/env python3 +# Copyright 2021-2022 Xiaomi Corp. (authors: Fangjun Kuang, +# Wei Kang, +# Mingshuang Luo,) +# Zengwei Yao) +# +# See ../../../../LICENSE for clarification regarding multiple authors +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +""" +Usage: + +export CUDA_VISIBLE_DEVICES="0,1,2,3" + +./pruned_transducer_stateless7_ctc/train.py \ + --world-size 4 \ + --num-epochs 30 \ + --start-epoch 1 \ + --exp-dir pruned_transducer_stateless7_ctc/exp \ + --full-libri 1 \ + --max-duration 300 + +# For mix precision training: + +./pruned_transducer_stateless7_ctc/train.py \ + --world-size 4 \ + --num-epochs 30 \ + --start-epoch 1 \ + --use-fp16 1 \ + --exp-dir pruned_transducer_stateless7_ctc/exp \ + --full-libri 1 \ + --max-duration 550 + +# For d2v-T training: +export CUDA_VISIBLE_DEVICES="0,1,2,3,4,5,6,7" + +./pruned_transducer_stateless_d2v_v2/train.py \ + --wandb true \ + --input-strategy AudioSamples \ + --enable-spec-aug False \ + --multi-optim True \ + --world-size 8 \ + --num-epochs 30 \ + --start-epoch 1 \ + --full-libri 0 \ + --exp-dir ./pruned_transducer_stateless_d2v_v2/$1 \ + --max-duration 250 \ + --freeze-finetune-updates 2000 \ + --use-fp16 1 \ + --peak-enc-lr 0.001 \ + --peak-dec-lr 0.05 \ + --accum-grads 1 \ + --encoder-type d2v \ + --additional-block True \ + --encoder-dim 768 \ + --decoder-dim 768 \ + --joiner-dim 768 \ + --prune-range 20 \ + --context-size 2 \ + --ctc-loss-scale 0.2 + +""" + + +import random +import argparse +import copy +import logging +import warnings +from pathlib import Path +from shutil import copyfile +from typing import Any, Dict, Optional, Tuple, Union + +import k2 +import optim +import sentencepiece as spm +import torch +import torch.multiprocessing as mp +import torch.nn as nn +from asr_datamodule import LibriSpeechAsrDataModule +from decoder import Decoder +from joiner import Joiner +from lhotse.cut import Cut +from lhotse.dataset.sampling.base import CutSampler +from lhotse.utils import fix_random_seed +from model import Transducer +from optim import Eden, ScaledAdam +from torch import Tensor +from torch.cuda.amp import GradScaler +from torch.nn.parallel import DistributedDataParallel as DDP +from torch.utils.tensorboard import SummaryWriter +from zipformer import Zipformer +from data2vec_encoder import FairSeqData2VecEncoder + +from icefall import diagnostics +from icefall.checkpoint import remove_checkpoints +from icefall.checkpoint import update_averaged_model +from checkpoint import ( + save_checkpoint as save_checkpoint_impl, + save_checkpoint_with_global_batch_idx, + load_checkpoint +) +from icefall.dist import cleanup_dist, setup_dist +from icefall.env import get_env_info +from icefall.hooks import register_inf_check_hooks +from icefall.utils import ( + AttributeDict, + MetricsTracker, + encode_supervisions, + setup_logger, + str2bool, + save_args, +) + +import wandb + +#from icefall.checkpoint import save_checkpoint as save_checkpoint_impl +LRSchedulerType = Union[torch.optim.lr_scheduler._LRScheduler, optim.LRScheduler] + + +def set_batch_count(model: Union[nn.Module, DDP], batch_count: float) -> None: + if isinstance(model, DDP): + # get underlying nn.Module + model = model.module + for module in model.modules(): + if hasattr(module, "batch_count"): + module.batch_count = batch_count + model.encoder.num_updates = int(batch_count) + + +def add_adapter_arguments(parser: argparse.ArgumentParser): + parser.add_argument( + "--add-adapter", + type=str2bool, + default=False, + help="add adapter to rep model's encoder" + ) + + parser.add_argument( + "--adapter-lr", + type=float, + default=0.0001, + help="adapter learning rate" + ) + + parser.add_argument( + "--gender", + type=str, + default='male', + help="select gender" + ) + + +def add_rep_arguments(parser: argparse.ArgumentParser): + parser.add_argument( + "--wandb", + type=str2bool, + default=True, + help="Use wandb for MLOps", + ) + parser.add_argument( + "--hpo", + type=str2bool, + default=False, + help="Use small db for HPO", + ) + + parser.add_argument( + "--accum-grads", + type=int, + default=1, + help="accum-grad num.", + ) + + parser.add_argument( + "--multi-optim", + type=str2bool, + default=True, + help="use sperate optimizer (enc / dec)", + ) + + parser.add_argument( + "--peak-enc-lr", + type=float, + default=0.0001, + help="The initial learning rate. This value should not need to be changed.", + ) + + parser.add_argument( + "--peak-dec-lr", + type=float, + default=0.001, + help="The initial learning rate. This value should not need to be changed.", + ) + + parser.add_argument( + "--encoder-type", + type=str, + default='d2v', + help="Type of encoder (e.g. conformer, w2v, d2v...", + ) + + parser.add_argument( + "--encoder-dim", + type=int, + default=768, + help="encoder embedding dimension", + ) + + parser.add_argument( + "--freeze-finetune-updates", + type=int, + default=0 + ) + + parser.add_argument( + "--additional-block", + type=str2bool, + default=True, + ) + + parser.add_argument( + "--decode-interval", + type=int, + default=200, + help="decode interval", + ) + + +def add_model_arguments(parser: argparse.ArgumentParser): + parser.add_argument( + "--num-encoder-layers", + type=str, + default="2,4,3,2,4", + help="Number of zipformer encoder layers, comma separated.", + ) + + parser.add_argument( + "--feedforward-dims", + type=str, + default="1024,1024,2048,2048,1024", + help="Feedforward dimension of the zipformer encoder layers, comma separated.", + ) + + parser.add_argument( + "--nhead", + type=str, + default="8,8,8,8,8", + help="Number of attention heads in the zipformer encoder layers.", + ) + + parser.add_argument( + "--encoder-dims", + type=str, + default="384,384,384,384,384", + help="Embedding dimension in the 2 blocks of zipformer encoder layers, comma separated", + ) + + parser.add_argument( + "--attention-dims", + type=str, + default="192,192,192,192,192", + help="""Attention dimension in the 2 blocks of zipformer encoder layers, comma separated; + not the same as embedding dimension.""", + ) + + parser.add_argument( + "--encoder-unmasked-dims", + type=str, + default="256,256,256,256,256", + help="Unmasked dimensions in the encoders, relates to augmentation during training. " + "Must be <= each of encoder_dims. Empirically, less than 256 seems to make performance " + " worse.", + ) + + parser.add_argument( + "--zipformer-downsampling-factors", + type=str, + default="1,2,4,8,2", + help="Downsampling factor for each stack of encoder layers.", + ) + + parser.add_argument( + "--cnn-module-kernels", + type=str, + default="31,31,31,31,31", + help="Sizes of kernels in convolution modules", + ) + + parser.add_argument( + "--decoder-dim", + type=int, + default=768, + help="Embedding dimension in the decoder model.", + ) + + parser.add_argument( + "--joiner-dim", + type=int, + default=768, + help="""Dimension used in the joiner model. + Outputs from the encoder and decoder model are projected + to this dimension before adding. + """, + ) + + +def get_parser(): + parser = argparse.ArgumentParser( + formatter_class=argparse.ArgumentDefaultsHelpFormatter + ) + + parser.add_argument( + "--world-size", + type=int, + default=1, + help="Number of GPUs for DDP training.", + ) + + parser.add_argument( + "--master-port", + type=int, + default=12354, + help="Master port to use for DDP training.", + ) + + parser.add_argument( + "--tensorboard", + type=str2bool, + default=True, + help="Should various information be logged in tensorboard.", + ) + + parser.add_argument( + "--num-epochs", + type=int, + default=30, + help="Number of epochs to train.", + ) + + parser.add_argument( + "--start-epoch", + type=int, + default=1, + help="""Resume training from this epoch. It should be positive. + If larger than 1, it will load checkpoint from + exp-dir/epoch-{start_epoch-1}.pt + """, + ) + + parser.add_argument( + "--start-batch", + type=int, + default=0, + help="""If positive, --start-epoch is ignored and + it loads the checkpoint from exp-dir/checkpoint-{start_batch}.pt + """, + ) + + parser.add_argument( + "--exp-dir", + type=str, + default="pruned_transducer_stateless7_ctc/exp", + help="""The experiment dir. + It specifies the directory where all training related + files, e.g., checkpoints, log, etc, are saved + """, + ) + + parser.add_argument( + "--bpe-model", + type=str, + default="data/lang_bpe_500/bpe.model", + help="Path to the BPE model", + ) + + parser.add_argument( + "--base-lr", type=float, default=0.05, help="The base learning rate." + ) + + parser.add_argument( + "--lr-batches", + type=float, + default=5000, + help="""Number of steps that affects how rapidly the learning rate + decreases. We suggest not to change this.""", + ) + + parser.add_argument( + "--lr-epochs", + type=float, + default=3.5, + help="""Number of epochs that affects how rapidly the learning rate decreases. + """, + ) + + parser.add_argument( + "--context-size", + type=int, + default=2, + help="The context size in the decoder. 1 means bigram; 2 means tri-gram", + ) + + parser.add_argument( + "--prune-range", + type=int, + default=5, + help="The prune range for rnnt loss, it means how many symbols(context)" + "we are using to compute the loss", + ) + + parser.add_argument( + "--lm-scale", + type=float, + default=0.25, + help="The scale to smooth the loss with lm " + "(output of prediction network) part.", + ) + + parser.add_argument( + "--am-scale", + type=float, + default=0.0, + help="The scale to smooth the loss with am (output of encoder network) part.", + ) + + parser.add_argument( + "--simple-loss-scale", + type=float, + default=0.5, + help="To get pruning ranges, we will calculate a simple version" + "loss(joiner is just addition), this simple loss also uses for" + "training (as a regularization item). We will scale the simple loss" + "with this parameter before adding to the final loss.", + ) + + parser.add_argument( + "--ctc-loss-scale", + type=float, + default=0.2, + help="Scale for CTC loss.", + ) + + parser.add_argument( + "--seed", + type=int, + default=42, + help="The seed for random generators intended for reproducibility", + ) + + parser.add_argument( + "--print-diagnostics", + type=str2bool, + default=False, + help="Accumulate stats on activations, print them and exit.", + ) + + parser.add_argument( + "--inf-check", + type=str2bool, + default=False, + help="Add hooks to check for infinite module outputs and gradients.", + ) + + parser.add_argument( + "--save-every-n", + type=int, + default=2000, + help="""Save checkpoint after processing this number of batches" + periodically. We save checkpoint to exp-dir/ whenever + params.batch_idx_train % save_every_n == 0. The checkpoint filename + has the form: f'exp-dir/checkpoint-{params.batch_idx_train}.pt' + Note: It also saves checkpoint to `exp-dir/epoch-xxx.pt` at the + end of each epoch where `xxx` is the epoch number counting from 0. + """, + ) + + parser.add_argument( + "--keep-last-k", + type=int, + default=30, + help="""Only keep this number of checkpoints on disk. + For instance, if it is 3, there are only 3 checkpoints + in the exp-dir with filenames `checkpoint-xxx.pt`. + It does not affect checkpoints with name `epoch-xxx.pt`. + """, + ) + + parser.add_argument( + "--average-period", + type=int, + default=200, + help="""Update the averaged model, namely `model_avg`, after processing + this number of batches. `model_avg` is a separate version of model, + in which each floating-point parameter is the average of all the + parameters from the start of training. Each time we take the average, + we do: `model_avg = model * (average_period / batch_idx_train) + + model_avg * ((batch_idx_train - average_period) / batch_idx_train)`. + """, + ) + + parser.add_argument( + "--use-fp16", + type=str2bool, + default=True, + help="Whether to use half precision training.", + ) + + add_model_arguments(parser) + add_rep_arguments(parser) + add_adapter_arguments(parser) + + return parser + + +def get_params() -> AttributeDict: + """Return a dict containing training parameters. + + All training related parameters that are not passed from the commandline + are saved in the variable `params`. + + Commandline options are merged into `params` after they are parsed, so + you can also access them via `params`. + + Explanation of options saved in `params`: + + - best_train_loss: Best training loss so far. It is used to select + the model that has the lowest training loss. It is + updated during the training. + + - best_valid_loss: Best validation loss so far. It is used to select + the model that has the lowest validation loss. It is + updated during the training. + + - best_train_epoch: It is the epoch that has the best training loss. + + - best_valid_epoch: It is the epoch that has the best validation loss. + + - batch_idx_train: Used to writing statistics to tensorboard. It + contains number of batches trained so far across + epochs. + + - log_interval: Print training loss if batch_idx % log_interval` is 0 + + - reset_interval: Reset statistics if batch_idx % reset_interval is 0 + + - valid_interval: Run validation if batch_idx % valid_interval is 0 + + - feature_dim: The model input dim. It has to match the one used + in computing features. + + - subsampling_factor: The subsampling factor for the model. + + - encoder_dim: Hidden dim for multi-head attention model. + + - num_decoder_layers: Number of decoder layer of transformer decoder. + + - warm_step: The warmup period that dictates the decay of the + scale on "simple" (un-pruned) loss. + """ + params = AttributeDict( + { + "best_train_loss": float("inf"), + "best_valid_loss": float("inf"), + "best_train_epoch": -1, + "best_valid_epoch": -1, + "batch_idx_train": 0, + "log_interval": 50, + "reset_interval": 200, + "valid_interval": 3000, # For the 100h subset, use 800 + # parameters for zipformer + "feature_dim": 80, + "subsampling_factor": 320, # not passed in, this is fixed. + # parameters for ctc loss + "beam_size": 10, + "use_double_scores": True, + "warm_step": 0, + #"warm_step": 4000, + #"warm_step": 3000, + "env_info": get_env_info(), + } + ) + + return params + + +def get_encoder_model(params: AttributeDict) -> nn.Module: + # TODO: We can add an option to switch between Zipformer and Transformer + def to_int_tuple(s: str): + return tuple(map(int, s.split(","))) + + if params.encoder_type == 'd2v': + encoder = FairSeqData2VecEncoder( + input_size=params.encoder_dim, + w2v_url='None', + output_size=params.encoder_dim, + freeze_finetune_updates=params.freeze_finetune_updates, + additional_block=params.additional_block, + ) + else: + encoder = Zipformer( + num_features=params.feature_dim, + output_downsampling_factor=2, + zipformer_downsampling_factors=to_int_tuple( + params.zipformer_downsampling_factors + ), + encoder_dims=to_int_tuple(params.encoder_dims), + attention_dim=to_int_tuple(params.attention_dims), + encoder_unmasked_dims=to_int_tuple(params.encoder_unmasked_dims), + nhead=to_int_tuple(params.nhead), + feedforward_dim=to_int_tuple(params.feedforward_dims), + cnn_module_kernels=to_int_tuple(params.cnn_module_kernels), + num_encoder_layers=to_int_tuple(params.num_encoder_layers), + ) + + return encoder + + +def get_decoder_model(params: AttributeDict) -> nn.Module: + decoder = Decoder( + vocab_size=params.vocab_size, + decoder_dim=params.decoder_dim, + blank_id=params.blank_id, + context_size=params.context_size, + ) + return decoder + + +def get_joiner_model(params: AttributeDict) -> nn.Module: + joiner = Joiner( + encoder_dim=params.encoder_dim if params.encoder_type == 'd2v' else int(params.encoder_dims.split(",")[-1]), + decoder_dim=params.decoder_dim, + joiner_dim=params.joiner_dim, + vocab_size=params.vocab_size, + ) + return joiner + + +def get_transducer_model(params: AttributeDict) -> nn.Module: + encoder = get_encoder_model(params) + decoder = get_decoder_model(params) + joiner = get_joiner_model(params) + + model = Transducer( + encoder=encoder, + decoder=decoder, + joiner=joiner, + encoder_dim=params.encoder_dim if params.encoder_type == 'd2v' else int(params.encoder_dims.split(",")[-1]), + decoder_dim=params.decoder_dim, + joiner_dim=params.joiner_dim, + vocab_size=params.vocab_size, + ) + return model + + +def load_checkpoint_if_available( + params: AttributeDict, + model: nn.Module, + model_avg: nn.Module = None, + optimizer: Optional[torch.optim.Optimizer] = None, + scheduler: Optional[LRSchedulerType] = None, +) -> Optional[Dict[str, Any]]: + """Load checkpoint from file. + + If params.start_batch is positive, it will load the checkpoint from + `params.exp_dir/checkpoint-{params.start_batch}.pt`. Otherwise, if + params.start_epoch is larger than 1, it will load the checkpoint from + `params.start_epoch - 1`. + + Apart from loading state dict for `model` and `optimizer` it also updates + `best_train_epoch`, `best_train_loss`, `best_valid_epoch`, + and `best_valid_loss` in `params`. + + Args: + params: + The return value of :func:`get_params`. + model: + The training model. + model_avg: + The stored model averaged from the start of training. + optimizer: + The optimizer that we are using. + scheduler: + The scheduler that we are using. + Returns: + Return a dict containing previously saved training info. + """ + if params.start_batch > 0: + filename = params.exp_dir / f"checkpoint-{params.start_batch}.pt" + elif params.start_epoch > 1: + filename = params.exp_dir / f"epoch-{params.start_epoch-1}.pt" + elif params.add_adapter: + filename = params.exp_dir / f"../d2v-base-T.pt" + else: + return None + + assert filename.is_file(), f"{filename} does not exist!" + + saved_params = load_checkpoint( + filename, + model=model, + model_avg=model_avg, + optimizer=optimizer, + scheduler=scheduler, + strict=True if not params.add_adapter else False, + ) + + keys = [ + "best_train_epoch", + "best_valid_epoch", + "batch_idx_train", + "best_train_loss", + "best_valid_loss", + ] + for k in keys: + params[k] = saved_params[k] + + if params.start_batch > 0: + if "cur_epoch" in saved_params: + params["start_epoch"] = saved_params["cur_epoch"] + + if "cur_batch_idx" in saved_params: + params["cur_batch_idx"] = saved_params["cur_batch_idx"] + + return saved_params + + +def save_checkpoint( + params: AttributeDict, + model: Union[nn.Module, DDP], + model_avg: Optional[nn.Module] = None, + optimizer: Optional[torch.optim.Optimizer] = None, + scheduler: Optional[LRSchedulerType] = None, + sampler: Optional[CutSampler] = None, + scaler: Optional[GradScaler] = None, + rank: int = 0, +) -> None: + """Save model, optimizer, scheduler and training stats to file. + + Args: + params: + It is returned by :func:`get_params`. + model: + The training model. + model_avg: + The stored model averaged from the start of training. + optimizer: + The optimizer used in the training. + sampler: + The sampler for the training dataset. + scaler: + The scaler used for mix precision training. + """ + if rank != 0: + return + filename = params.exp_dir / f"epoch-{params.cur_epoch}.pt" + save_checkpoint_impl( + filename=filename, + model=model, + model_avg=model_avg, + params=params, + optimizer=optimizer, + scheduler=scheduler, + sampler=sampler, + scaler=scaler, + rank=rank, + ) + + if params.best_train_epoch == params.cur_epoch: + best_train_filename = params.exp_dir / "best-train-loss.pt" + copyfile(src=filename, dst=best_train_filename) + + if params.best_valid_epoch == params.cur_epoch: + best_valid_filename = params.exp_dir / "best-valid-loss.pt" + copyfile(src=filename, dst=best_valid_filename) + + +def compute_loss( + params: AttributeDict, + model: Union[nn.Module, DDP], + sp: spm.SentencePieceProcessor, + batch: dict, + is_training: bool, + decode: bool = False, +) -> Tuple[Tensor, MetricsTracker]: + """ + Compute transducer loss given the model and its inputs. + + Args: + params: + Parameters for training. See :func:`get_params`. + model: + The model for training. It is an instance of Zipformer in our case. + batch: + A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()` + for the content in it. + is_training: + True for training. False for validation. When it is True, this + function enables autograd during computation; when it is False, it + disables autograd. + warmup: a floating point value which increases throughout training; + values >= 1.0 are fully warmed up and have all modules present. + """ + device = model.device if isinstance(model, DDP) else next(model.parameters()).device + feature = batch["inputs"] + # at entry, feature is (N, T, C) + assert feature.ndim == 2 or feature.ndim == 3 + feature = feature.to(device) + + supervisions = batch["supervisions"] + + if feature.ndim == 2: + feature_lens = [] + for supervision in supervisions['cut']: + try: feature_lens.append(supervision.tracks[0].cut.recording.num_samples) + except: feature_lens.append(supervision.recording.num_samples) + feature_lens = torch.tensor(feature_lens) + + elif feature.ndim == 3: + feature_lens = supervisions["num_frames"].to(device) + + batch_idx_train = params.batch_idx_train + warm_step = params.warm_step + + texts = batch["supervisions"]["text"] + + token_ids = sp.encode(texts, out_type=int) + y = k2.RaggedTensor(token_ids).to(device) + + with torch.set_grad_enabled(is_training): + simple_loss, pruned_loss, ctc_output = model( + x=feature, + x_lens=feature_lens, + y=y, + prune_range=params.prune_range, + am_scale=params.am_scale, + lm_scale=params.lm_scale, + ) + + s = params.simple_loss_scale + # take down the scale on the simple loss from 1.0 at the start + # to params.simple_loss scale by warm_step. + simple_loss_scale = ( + s + if batch_idx_train >= warm_step + else 1.0 - (batch_idx_train / warm_step) * (1.0 - s) + ) + pruned_loss_scale = ( + 1.0 + if batch_idx_train >= warm_step + else 0.1 + 0.9 * (batch_idx_train / warm_step) + ) + + loss = simple_loss_scale * simple_loss + pruned_loss_scale * pruned_loss + + info = MetricsTracker() + + if params.ctc_loss_scale > 0: + # Compute ctc loss + + # NOTE: We need `encode_supervisions` to sort sequences with + # different duration in decreasing order, required by + # `k2.intersect_dense` called in `k2.ctc_loss` + with warnings.catch_warnings(): + warnings.simplefilter("ignore") + supervision_segments, token_ids = encode_supervisions( + supervisions, + subsampling_factor=params.subsampling_factor, + token_ids=token_ids, + ) + + # Works with a BPE model + decoding_graph = k2.ctc_graph(token_ids, modified=False, device=device) + dense_fsa_vec = k2.DenseFsaVec( + ctc_output, + supervision_segments, + allow_truncate=params.subsampling_factor - 1, + ) + + ctc_loss = k2.ctc_loss( + decoding_graph=decoding_graph, + dense_fsa_vec=dense_fsa_vec, + output_beam=params.beam_size, + reduction="sum", + use_double_scores=params.use_double_scores, + ) + assert ctc_loss.requires_grad == is_training + loss += params.ctc_loss_scale * ctc_loss + + info["ctc_loss"] = ctc_loss.detach().cpu().item() + + assert loss.requires_grad == is_training + + if decode: + model.eval() + with torch.no_grad(): + hypos = model.module.decode( + x=feature, + x_lens=feature_lens, + y=y, + sp=sp + ) + logging.info(f'ref: {batch["supervisions"]["text"][0]}') + logging.info(f'hyp: {" ".join(hypos[0])}') + model.train() + + with warnings.catch_warnings(): + warnings.simplefilter("ignore") + info["frames"] = (feature_lens // params.subsampling_factor).sum().item() + + # Note: We use reduction=sum while computing the loss. + info["utterances"] = feature.size(0) + info["loss"] = loss.detach().cpu().item() + info["simple_loss"] = simple_loss.detach().cpu().item() + info["pruned_loss"] = pruned_loss.detach().cpu().item() + + return loss, info + + +def compute_validation_loss( + params: AttributeDict, + model: Union[nn.Module, DDP], + sp: spm.SentencePieceProcessor, + valid_dl: torch.utils.data.DataLoader, + world_size: int = 1, +) -> MetricsTracker: + """Run the validation process.""" + model.eval() + + tot_loss = MetricsTracker() + + for batch_idx, batch in enumerate(valid_dl): + loss, loss_info = compute_loss( + params=params, + model=model, + sp=sp, + batch=batch, + is_training=False, + ) + assert loss.requires_grad is False + tot_loss = tot_loss + loss_info + + if world_size > 1: + tot_loss.reduce(loss.device) + + loss_value = tot_loss["loss"] / tot_loss["utterances"] + if loss_value < params.best_valid_loss: + params.best_valid_epoch = params.cur_epoch + params.best_valid_loss = loss_value + + return tot_loss + + +def train_one_epoch( + params: AttributeDict, + model: Union[nn.Module, DDP], + optimizer: torch.optim.Optimizer or [torch.optim.Optimizer, torch.optim.Optimizer], + scheduler: LRSchedulerType or [LRSchedulerType, LRSchedulerType], + sp: spm.SentencePieceProcessor, + train_dl: torch.utils.data.DataLoader or [torch.utils.data.DataLoader, torch.utils.data.DataLoader], + valid_dl: torch.utils.data.DataLoader, + scaler: GradScaler, + model_avg: Optional[nn.Module] = None, + tb_writer: Optional[SummaryWriter] = None, + world_size: int = 1, + rank: int = 0, + wb = None, +) -> None: + """Train the model for one epoch. + + The training loss from the mean of all frames is saved in + `params.train_loss`. It runs the validation process every + `params.valid_interval` batches. + + Args: + params: + It is returned by :func:`get_params`. + model: + The model for training. + optimizer: + The optimizer we are using. + scheduler: + The learning rate scheduler, we call step() every step. + train_dl: + Dataloader for the training dataset. + valid_dl: + Dataloader for the validation dataset. + scaler: + The scaler used for mix precision training. + model_avg: + The stored model averaged from the start of training. + tb_writer: + Writer to write log messages to tensorboard. + world_size: + Number of nodes in DDP training. If it is 1, DDP is disabled. + rank: + The rank of the node in DDP training. If no DDP is used, it should + be set to 0. + """ + model.train() + + tot_loss = MetricsTracker() + + cur_batch_idx = params.get("cur_batch_idx", 0) + + if params.multi_optim: + optimizer_enc, optimizer_dec = optimizer[0], optimizer[1] + scheduler_enc, scheduler_dec = scheduler[0], scheduler[1] + + if type(train_dl) == list: + train_dl_uda = train_dl[1] + train_dl = train_dl[0] + #for batch_idx, batch in enumerate(train_dl): + for batch_idx, batch in enumerate(zip(train_dl, train_dl_uda)): + if batch_idx < cur_batch_idx: + continue + cur_batch_idx = batch_idx + + if batch_idx % params.accum_grads == 0: params.batch_idx_train += 1 + batch_size = len(batch["supervisions"]["text"]) + + try: + with torch.cuda.amp.autocast(enabled=params.use_fp16): + loss, loss_info = compute_loss( + params=params, + model=model, + sp=sp, + batch=batch, + is_training=True, + decode = True if batch_idx % params.decode_interval == 0 else False, + ) + loss_info.reduce(loss.device) + + numel = params.world_size / (params.accum_grads * loss_info["utterances"]) + loss *= numel ## normalize loss over utts(batch size) + + # summary stats + tot_loss = (tot_loss * (1 - 1 / params.reset_interval)) + loss_info + + # NOTE: We use reduction==sum and loss is computed over utterances + # in the batch and there is no normalization to it so far. + scaler.scale(loss).backward() + + if params.multi_optim and (batch_idx+1) % params.accum_grads == 0: + set_batch_count(model, params.batch_idx_train) + scheduler_enc.step_batch(params.batch_idx_train) + scheduler_dec.step_batch(params.batch_idx_train) + scaler.step(optimizer_enc) + scaler.step(optimizer_dec) + scaler.update() + optimizer_enc.zero_grad() + optimizer_dec.zero_grad() + elif not params.multi_optim and (batch_idx+1) % params.accum_grads == 0: + set_batch_count(model, params.batch_idx_train) + scheduler.step_batch(params.batch_idx_train) + scaler.step(optimizer) + scaler.update() + optimizer.zero_grad() + + except: # noqa + display_and_save_batch(batch, params=params, sp=sp) + raise + + if params.print_diagnostics and batch_idx == 5: + return + + if ( + rank == 0 + and params.batch_idx_train > 0 + and params.batch_idx_train % params.average_period == 0 + ): + update_averaged_model( + params=params, + model_cur=model, + model_avg=model_avg, + ) + + if ( + params.batch_idx_train > 0 + and params.batch_idx_train % params.save_every_n == 0 + ): + params.cur_batch_idx = batch_idx + save_checkpoint_with_global_batch_idx( + out_dir=params.exp_dir, + global_batch_idx=params.batch_idx_train, + model=model, + model_avg=model_avg, + params=params, + optimizer=optimizer, + scheduler=scheduler, + sampler=train_dl.sampler, + scaler=scaler, + rank=rank, + ) + del params.cur_batch_idx + remove_checkpoints( + out_dir=params.exp_dir, + topk=params.keep_last_k, + rank=rank, + ) + + if batch_idx % 100 == 0 and params.use_fp16: + # If the grad scale was less than 1, try increasing it. The _growth_interval + # of the grad scaler is configurable, but we can't configure it to have different + # behavior depending on the current grad scale. + cur_grad_scale = scaler._scale.item() + ''' + if cur_grad_scale < 1.0 or (cur_grad_scale < 8.0 and batch_idx % 400 == 0): + scaler.update(cur_grad_scale * 2.0) + ''' + if cur_grad_scale < 0.01: + logging.warning(f"Grad scale is small: {cur_grad_scale}") + if cur_grad_scale < 1.0e-05: + wb.log({"valid/loss": 10000}) + raise RuntimeError( + f"grad_scale is too small, exiting: {cur_grad_scale}" + ) + + #if params.batch_idx_train > 4000 and loss > 300 and params.wandb: + # wb.log({"valid/loss": 10000}) + # raise RuntimeError( + # f"divergence... exiting: loss={loss}" + # ) + + if batch_idx % (params.log_interval*params.accum_grads) == 0: + #for n, p in model.named_parameters(): + # if 'adapter' in n: + # print(p) + if params.multi_optim: + cur_enc_lr = scheduler_enc.get_last_lr()[0] + cur_dec_lr = scheduler_dec.get_last_lr()[0] + cur_grad_scale = scaler._scale.item() if params.use_fp16 else 1.0 + + logging.info( + f"Epoch {params.cur_epoch}, " + f"batch {batch_idx}, loss[{loss_info}], " + f"tot_loss[{tot_loss}], batch size: {batch_size}, " + f"enc_lr: {cur_enc_lr:.2e}, " + f"dec_lr: {cur_dec_lr:.2e}, " + + (f"grad_scale: {scaler._scale.item()}" if params.use_fp16 else "") + ) + + else: + cur_lr = scheduler.get_last_lr()[0] + cur_grad_scale = scaler._scale.item() if params.use_fp16 else 1.0 + + logging.info( + f"Epoch {params.cur_epoch}, " + f"batch {batch_idx}, loss[{loss_info}], " + f"tot_loss[{tot_loss}], batch size: {batch_size}, " + f"lr: {cur_lr:.2e}, " + + (f"grad_scale: {scaler._scale.item()}" if params.use_fp16 else "") + ) + + if tb_writer is not None: + if params.multi_optim: + tb_writer.add_scalar( + "train/enc_learning_rate", cur_enc_lr, params.batch_idx_train + ) + tb_writer.add_scalar( + "train/dec_learning_rate", cur_dec_lr, params.batch_idx_train + ) + + else: + tb_writer.add_scalar( + "train/learning_rate", cur_lr, params.batch_idx_train + ) + + loss_info.write_summary( + tb_writer, "train/current_", params.batch_idx_train + ) + tot_loss.write_summary(tb_writer, "train/tot_", params.batch_idx_train) + if params.use_fp16: + tb_writer.add_scalar( + "train/grad_scale", + cur_grad_scale, + params.batch_idx_train, + ) + + if wb is not None and rank == 0: + wb.log({"train/loss": loss_info["loss"]*numel}) + wb.log({"train/simple_loss": loss_info["simple_loss"]*numel}) + wb.log({"train/pruned_loss": loss_info["pruned_loss"]*numel}) + wb.log({"train/ctc_loss": loss_info["ctc_loss"]*numel}) + + ''' + logging.info("Computing validation loss") + valid_info = compute_validation_loss( + params=params, + model=model, + sp=sp, + valid_dl=valid_dl, + world_size=world_size, + ) + model.train() + logging.info(f"Epoch {params.cur_epoch}, validation: {valid_info}") + logging.info( + f"Maximum memory allocated so far is {torch.cuda.max_memory_allocated()//1000000}MB" + ) + if tb_writer is not None: + valid_info.write_summary( + tb_writer, "train/valid_", params.batch_idx_train + ) + + if wb is not None and rank == 0: + numel = 1 / (params.accum_grads * valid_info["utterances"]) + #wb.log({"valid/loss": valid_info["loss"]*numel}) + wb.log({"valid/loss": numel*(valid_info["simple_loss"] + +valid_info["pruned_loss"] + +valid_info["ctc_loss"] + )}) + wb.log({"valid/simple_loss": valid_info["simple_loss"]*numel}) + wb.log({"valid/pruned_loss": valid_info["pruned_loss"]*numel}) + wb.log({"valid/ctc_loss": valid_info["ctc_loss"]*numel}) + ''' + loss_value = tot_loss["loss"] / tot_loss["utterances"] + params.train_loss = loss_value + if params.train_loss < params.best_train_loss: + params.best_train_epoch = params.cur_epoch + params.best_train_loss = params.train_loss + + +def run(rank, world_size, args, wb=None): + """ + Args: + rank: + It is a value between 0 and `world_size-1`, which is + passed automatically by `mp.spawn()` in :func:`main`. + The node with rank 0 is responsible for saving checkpoint. + world_size: + Number of GPUs for DDP training. + args: + The return value of get_parser().parse_args() + """ + params = get_params() + params.update(vars(args)) + #params.warm_step *= params.accum_grads + + fix_random_seed(params.seed) + if world_size > 1: + setup_dist(rank, world_size, params.master_port) + + setup_logger(f"{params.exp_dir}/log/log-train") + logging.info("Training started") + + if args.tensorboard and rank == 0: + tb_writer = SummaryWriter(log_dir=f"{params.exp_dir}/tensorboard") + else: + tb_writer = None + + device = torch.device("cpu") + if torch.cuda.is_available(): + device = torch.device("cuda", rank) + logging.info(f"Device: {device}") + + sp = spm.SentencePieceProcessor() + sp.load(params.bpe_model) + + # is defined in local/train_bpe_model.py + params.blank_id = sp.piece_to_id("") + params.vocab_size = sp.get_piece_size() + + logging.info(params) + + logging.info("About to create model") + model = get_transducer_model(params) + logging.info(model) + + num_param = sum([p.numel() for p in model.parameters()]) + logging.info(f"Number of model parameters: {num_param}") + + assert params.save_every_n >= params.average_period + model_avg: Optional[nn.Module] = None + if rank == 0: + # model_avg is only used with rank 0 + model_avg = copy.deepcopy(model).to(torch.float64) + + assert params.start_epoch > 0, params.start_epoch + checkpoints = load_checkpoint_if_available( + params=params, model=model, model_avg=model_avg + ) + + model.to(device) + if world_size > 1: + logging.info("Using DDP") + model = DDP(model, device_ids=[rank], find_unused_parameters=True) + + if params.multi_optim: + logging.info("Using seperate optimizers over encoder, decoder ...") + + enc_param = [] + enc_names = [] + + dec_names = [] + dec_param = [] + + for n, p in model.named_parameters(): + name = n.split('.')[1] + if name == 'encoder' and 'feature_extractor' not in n: + enc_names.append(n) + enc_param.append(p) + elif 'ctc_output' in n: + enc_names.append(n) + enc_param.append(p) + elif 'feature_extractor' not in n: + dec_names.append(n) + dec_param.append(p) + + optimizer_enc = ScaledAdam( + enc_param, + lr=params.peak_enc_lr, + clipping_scale=None, + parameters_names=[enc_names], + ) + optimizer_dec = ScaledAdam( + dec_param, + lr=params.peak_dec_lr, + clipping_scale=5.0, + parameters_names=[dec_names], + ) + + scheduler_enc = Eden(optimizer_enc, params.lr_batches, params.lr_epochs) + scheduler_dec = Eden(optimizer_dec, params.lr_batches, params.lr_epochs) + optimizer = [optimizer_enc, optimizer_dec] + scheduler = [scheduler_enc, scheduler_dec] + + else: + parameters_names = [] + parameters_names.append( + [name_param_pair[0] for name_param_pair in model.named_parameters()] + ) + + logging.info(f"len name = {len(parameters_names)}") + logging.info(f"len param = {len(list(model.parameters()))}") + + optimizer = ScaledAdam( + model.parameters(), + lr=params.base_lr, + clipping_scale=2.0, + parameters_names=parameters_names, + ) + + scheduler = Eden(optimizer, params.lr_batches, params.lr_epochs) + + if checkpoints and ("optimizer" in checkpoints or "optimizer_enc" in checkpoints): + if params.multi_optim: + logging.info("Loading optimizer state dict") + optimizer_enc.load_state_dict(checkpoints["optimizer_enc"]) + optimizer_dec.load_state_dict(checkpoints["optimizer_dec"]) + + else: + logging.info("Loading optimizer state dict") + optimizer.load_state_dict(checkpoints["optimizer"]) + + if checkpoints: + if ( + params.multi_optim + and "scheduler_enc" in checkpoints + and checkpoints["scheduler_enc"] is not None + ): + logging.info("Loading enc/dec scheduler state dict") + scheduler_enc.load_state_dict(checkpoints["scheduler_enc"]) + scheduler_dec.load_state_dict(checkpoints["scheduler_dec"]) + else: + logging.info("Loading scheduler state dict") + scheduler.load_state_dict(checkpoints["scheduler"]) + + if params.print_diagnostics: + opts = diagnostics.TensorDiagnosticOptions( + 2**22 + ) # allow 4 megabytes per sub-module + diagnostic = diagnostics.attach_diagnostics(model, opts) + + if params.inf_check: + register_inf_check_hooks(model) + + librispeech = LibriSpeechAsrDataModule(args) + + train_cuts = librispeech.train_clean_100_cuts() + if params.full_libri: + train_cuts += librispeech.train_clean_360_cuts() + train_cuts += librispeech.train_other_500_cuts() + + def remove_short_and_long_utt(c: Cut): + # Keep only utterances with duration between 1 second and 20 seconds + # + # Caution: There is a reason to select 20.0 here. Please see + # ../local/display_manifest_statistics.py + # + # You should use ../local/display_manifest_statistics.py to get + # an utterance duration distribution for your dataset to select + # the threshold + return 1.0 <= c.duration <= 20.0 + + train_cuts = train_cuts.filter(remove_short_and_long_utt) + + if params.start_batch > 0 and checkpoints and "sampler" in checkpoints: + # We only load the sampler's state dict when it loads a checkpoint + # saved in the middle of an epoch + sampler_state_dict = checkpoints["sampler"] + else: + sampler_state_dict = None + + train_dl = librispeech.train_dataloaders( + train_cuts, sampler_state_dict=sampler_state_dict + ) + + valid_cuts = librispeech.dev_clean_cuts() + valid_cuts += librispeech.dev_other_cuts() + valid_dl = librispeech.valid_dataloaders(valid_cuts) + + ''' + if not params.print_diagnostics: + scan_pessimistic_batches_for_oom( + model=model, + train_dl=train_dl, + optimizer=optimizer, + sp=sp, + params=params, + ) + ''' + + scaler = GradScaler(enabled=params.use_fp16, init_scale=1.0) + if checkpoints and "grad_scaler" in checkpoints: + logging.info("Loading grad scaler state dict") + scaler.load_state_dict(checkpoints["grad_scaler"]) + + for epoch in range(params.start_epoch, params.num_epochs + 1): + if params.multi_optim: + scheduler_enc.step_epoch(epoch - 1) + scheduler_dec.step_epoch(epoch - 1) + else: + scheduler.step_epoch(epoch - 1) + fix_random_seed(params.seed + epoch - 1) + train_dl.sampler.set_epoch(epoch - 1) + + if tb_writer is not None: + tb_writer.add_scalar("train/epoch", epoch, params.batch_idx_train) + + params.cur_epoch = epoch + + train_one_epoch( + params=params, + model=model, + model_avg=model_avg, + optimizer=optimizer, + scheduler=scheduler, + sp=sp, + train_dl=train_dl, + valid_dl=valid_dl, + scaler=scaler, + tb_writer=tb_writer, + world_size=world_size, + rank=rank, + wb=wb, + ) + + if params.print_diagnostics: + diagnostic.print_diagnostics() + break + + if epoch % 10 == 0: + save_checkpoint( + params=params, + model=model, + model_avg=model_avg, + optimizer=optimizer, + scheduler=scheduler, + sampler=train_dl.sampler, + scaler=scaler, + rank=rank, + ) + + logging.info("Done!") + + if world_size > 1: + torch.distributed.barrier() + cleanup_dist() + + +def run_adapter(rank, world_size, args, wb=None): + """ + Args: + rank: + It is a value between 0 and `world_size-1`, which is + passed automatically by `mp.spawn()` in :func:`main`. + The node with rank 0 is responsible for saving checkpoint. + world_size: + Number of GPUs for DDP training. + args: + The return value of get_parser().parse_args() + """ + params = get_params() + params.update(vars(args)) + + fix_random_seed(params.seed) + if world_size > 1: + setup_dist(rank, world_size, params.master_port) + + setup_logger(f"{params.exp_dir}/log/log-train") + logging.info("Training started") + + if args.tensorboard and rank == 0: + tb_writer = SummaryWriter(log_dir=f"{params.exp_dir}/tensorboard") + else: + tb_writer = None + + device = torch.device("cpu") + if torch.cuda.is_available(): + device = torch.device("cuda", rank) + logging.info(f"Device: {device}") + + sp = spm.SentencePieceProcessor() + sp.load(params.bpe_model) + + # is defined in local/train_bpe_model.py + params.blank_id = sp.piece_to_id("") + params.vocab_size = sp.get_piece_size() + + logging.info(params) + + logging.info("About to create model") + model = get_transducer_model(params) + + num_param = sum([p.numel() if p.requires_grad else 0 for p in model.parameters()]) + logging.info(f"Number of model parameters: {num_param}") + + assert params.save_every_n >= params.average_period + model_avg: Optional[nn.Module] = None + if rank == 0: + # model_avg is only used with rank 0 + model_avg = copy.deepcopy(model).to(torch.float64) + + assert params.start_epoch > 0, params.start_epoch + checkpoints = load_checkpoint_if_available( + params=params, model=model, model_avg=model_avg + ) + + model.to(device) + if world_size > 1: + logging.info("Using DDP") + model = DDP(model, device_ids=[rank], find_unused_parameters=True) + + adapter_names = [] + adapter_param = [] + for n, p in model.named_parameters(): + if 'adapters' in n:# or 'joiner' in n or 'simple' in n or 'ctc' in n: + adapter_names.append(n) + adapter_param.append(p) + elif 'joiner' in n or 'simple' in n or 'ctc' in n: + p.requires_grad = True + else: + p.requires_grad = False + + optimizer_adapter = ScaledAdam( + adapter_param, + lr=params.adapter_lr, + clipping_scale=5.0, + parameters_names=[adapter_names], + ) + scheduler_adapter = Eden(optimizer_adapter, 10000, 7) #params.lr_batche, params.lr_epochs) + + optimizer, scheduler = optimizer_adapter, scheduler_adapter + + librispeech = LibriSpeechAsrDataModule(args) + + ''' + if params.hpo: + train_cuts = librispeech.train_clean_10_cuts(option=params.gender) + else: + train_cuts = librispeech.train_clean_100_cuts(option=params.gender) + if params.full_libri: + train_cuts += librispeech.train_clean_360_cuts(option=params.gender) + train_cuts += librispeech.train_other_500_cuts(option=params.gender) + ''' + + #train_cuts = librispeech.train_clean_10_cuts(option='male') + #train_cuts = librispeech.test_clean_user(option='big') + train_cuts = librispeech.vox_cuts(option=params.spk_id) + + def remove_short_and_long_utt(c: Cut): + return 1.0 <= c.duration <= 20.0 + + train_cuts = train_cuts.filter(remove_short_and_long_utt) + + sampler_state_dict = None + + train_dl = librispeech.train_dataloaders( + train_cuts, sampler_state_dict=sampler_state_dict + ) + #train_dl = librispeech.test_dataloaders( + # train_cuts + #) + + ''' + print('\n'*5) + print('-'*30) + for batch in train_dl: + print(batch) + print('-'*30) + print('\n'*5) + exit() + ''' + + valid_cuts = librispeech.dev_clean_cuts(option=params.gender) + valid_cuts += librispeech.dev_other_cuts(option=params.gender) + valid_dl = librispeech.valid_dataloaders(valid_cuts) + + scaler = GradScaler(enabled=params.use_fp16, init_scale=1.0) + + for epoch in range(params.start_epoch, params.num_epochs + 1): + scheduler.step_epoch(epoch - 1) + fix_random_seed(params.seed + epoch - 1) + train_dl.sampler.set_epoch(epoch - 1) + + if tb_writer is not None: + tb_writer.add_scalar("train/epoch", epoch, params.batch_idx_train) + + params.cur_epoch = epoch + + train_one_epoch( + params=params, + model=model, + model_avg=model_avg, + optimizer=optimizer, + scheduler=scheduler, + sp=sp, + train_dl=train_dl, + valid_dl=valid_dl, + scaler=scaler, + tb_writer=tb_writer, + world_size=world_size, + rank=rank, + wb=wb, + ) + + if params.print_diagnostics: + diagnostic.print_diagnostics() + break + + if epoch % 10 == 0: + save_checkpoint( + params=params, + model=model, + model_avg=model_avg, + optimizer=optimizer, + scheduler=scheduler, + sampler=train_dl.sampler, + scaler=scaler, + rank=rank, + ) + + logging.info("Done!") + + if world_size > 1: + torch.distributed.barrier() + cleanup_dist() + + +def run_adapter_uda(rank, world_size, args, wb=None): + """ + Args: + rank: + It is a value between 0 and `world_size-1`, which is + passed automatically by `mp.spawn()` in :func:`main`. + The node with rank 0 is responsible for saving checkpoint. + world_size: + Number of GPUs for DDP training. + args: + The return value of get_parser().parse_args() + """ + params = get_params() + params.update(vars(args)) + + fix_random_seed(params.seed) + if world_size > 1: + setup_dist(rank, world_size, params.master_port) + + setup_logger(f"{params.exp_dir}/log/log-train") + logging.info("Training started") + + if args.tensorboard and rank == 0: + tb_writer = SummaryWriter(log_dir=f"{params.exp_dir}/tensorboard") + else: + tb_writer = None + + device = torch.device("cpu") + if torch.cuda.is_available(): + device = torch.device("cuda", rank) + logging.info(f"Device: {device}") + + sp = spm.SentencePieceProcessor() + sp.load(params.bpe_model) + + # is defined in local/train_bpe_model.py + params.blank_id = sp.piece_to_id("") + params.vocab_size = sp.get_piece_size() + + logging.info(params) + + logging.info("About to create model") + model = get_transducer_model(params) + + num_param = sum([p.numel() if p.requires_grad else 0 for p in model.parameters()]) + logging.info(f"Number of model parameters: {num_param}") + + assert params.save_every_n >= params.average_period + model_avg: Optional[nn.Module] = None + if rank == 0: + # model_avg is only used with rank 0 + model_avg = copy.deepcopy(model).to(torch.float64) + + assert params.start_epoch > 0, params.start_epoch + checkpoints = load_checkpoint_if_available( + params=params, model=model, model_avg=model_avg + ) + + model.to(device) + if world_size > 1: + logging.info("Using DDP") + model = DDP(model, device_ids=[rank], find_unused_parameters=True) + + adapter_names = [] + adapter_param = [] + for n, p in model.named_parameters(): + if 'adapters' in n:# or 'joiner' in n or 'simple' in n or 'ctc' in n: + adapter_names.append(n) + adapter_param.append(p) + elif 'joiner' in n or 'simple' in n or 'ctc' in n: + p.requires_grad = True + else: + p.requires_grad = False + + optimizer_adapter = ScaledAdam( + adapter_param, + lr=params.adapter_lr, + clipping_scale=5.0, + parameters_names=[adapter_names], + ) + scheduler_adapter = Eden(optimizer_adapter, 10000, 7) #params.lr_batche, params.lr_epochs) + + optimizer, scheduler = optimizer_adapter, scheduler_adapter + + librispeech = LibriSpeechAsrDataModule(args) + librispeech_uda = LibriSpeechAsrDataModule(args) + + ''' + if params.hpo: + train_cuts = librispeech.train_clean_10_cuts(option=params.gender) + else: + train_cuts = librispeech.train_clean_100_cuts(option=params.gender) + if params.full_libri: + train_cuts += librispeech.train_clean_360_cuts(option=params.gender) + train_cuts += librispeech.train_other_500_cuts(option=params.gender) + ''' + + #train_cuts = librispeech.train_clean_10_cuts(option='male') + #train_cuts = librispeech.test_clean_user(option='big') + train_cuts = librispeech.vox_cuts(option=params.spk_id) + train_cuts_uda = librispeech_uda.vox_cuts(option=params.spk_id) + + def remove_short_and_long_utt(c: Cut): + return 1.0 <= c.duration <= 20.0 + + train_cuts = train_cuts.filter(remove_short_and_long_utt) + train_cuts_uda = train_cuts_uda.filter(remove_short_and_long_utt) + + sampler_state_dict = None + + train_dl = librispeech.train_dataloaders( + train_cuts, sampler_state_dict=sampler_state_dict + ) + train_dl_uda = librispeech.train_dataloaders( + train_cuts_uda, sampler_state_dict=sampler_state_dict + ) + + #train_dl = librispeech.test_dataloaders( + # train_cuts + #) + + ''' + print('\n'*5) + print('-'*30) + for batch in train_dl: + print(batch) + print('-'*30) + print('\n'*5) + exit() + ''' + + valid_cuts = librispeech.dev_clean_cuts(option=params.gender) + valid_cuts += librispeech.dev_other_cuts(option=params.gender) + valid_dl = librispeech.valid_dataloaders(valid_cuts) + + scaler = GradScaler(enabled=params.use_fp16, init_scale=1.0) + + for epoch in range(params.start_epoch, params.num_epochs + 1): + scheduler.step_epoch(epoch - 1) + fix_random_seed(params.seed + epoch - 1) + train_dl.sampler.set_epoch(epoch - 1) + + if tb_writer is not None: + tb_writer.add_scalar("train/epoch", epoch, params.batch_idx_train) + + params.cur_epoch = epoch + + train_one_epoch( + params=params, + model=model, + model_avg=model_avg, + optimizer=optimizer, + scheduler=scheduler, + sp=sp, + train_dl=[train_dl, train_dl_uda], + valid_dl=valid_dl, + scaler=scaler, + tb_writer=tb_writer, + world_size=world_size, + rank=rank, + wb=wb, + ) + + if params.print_diagnostics: + diagnostic.print_diagnostics() + break + + if epoch % 10 == 0: + save_checkpoint( + params=params, + model=model, + model_avg=model_avg, + optimizer=optimizer, + scheduler=scheduler, + sampler=train_dl.sampler, + scaler=scaler, + rank=rank, + ) + + logging.info("Done!") + + if world_size > 1: + torch.distributed.barrier() + cleanup_dist() + + + +def display_and_save_batch( + batch: dict, + params: AttributeDict, + sp: spm.SentencePieceProcessor, +) -> None: + """Display the batch statistics and save the batch into disk. + + Args: + batch: + A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()` + for the content in it. + params: + Parameters for training. See :func:`get_params`. + sp: + The BPE model. + """ + from lhotse.utils import uuid4 + + filename = f"{params.exp_dir}/batch-{uuid4()}.pt" + logging.info(f"Saving batch to {filename}") + torch.save(batch, filename) + + supervisions = batch["supervisions"] + features = batch["inputs"] + + logging.info(f"features shape: {features.shape}") + + y = sp.encode(supervisions["text"], out_type=int) + num_tokens = sum(len(i) for i in y) + logging.info(f"num tokens: {num_tokens}") + + +def scan_pessimistic_batches_for_oom( + model: Union[nn.Module, DDP], + train_dl: torch.utils.data.DataLoader, + optimizer: torch.optim.Optimizer, + sp: spm.SentencePieceProcessor, + params: AttributeDict, +): + from lhotse.dataset import find_pessimistic_batches + + logging.info( + "Sanity check -- see if any of the batches in epoch 1 would cause OOM." + ) + batches, crit_values = find_pessimistic_batches(train_dl.sampler) + for criterion, cuts in batches.items(): + batch = train_dl.dataset[cuts] + try: + with torch.cuda.amp.autocast(enabled=params.use_fp16): + loss, _ = compute_loss( + params=params, + model=model, + sp=sp, + batch=batch, + is_training=True, + ) + loss.backward() + optimizer.zero_grad() + except Exception as e: + if "CUDA out of memory" in str(e): + logging.error( + "Your GPU ran out of memory with the current " + "max_duration setting. We recommend decreasing " + "max_duration and trying again.\n" + f"Failing criterion: {criterion} " + f"(={crit_values[criterion]}) ..." + ) + display_and_save_batch(batch, params=params, sp=sp) + raise + logging.info( + f"Maximum memory allocated so far is {torch.cuda.max_memory_allocated()//1000000}MB" + ) + + +def main(): + parser = get_parser() + LibriSpeechAsrDataModule.add_arguments(parser) + args = parser.parse_args() + if args.wandb: args.exp_dir = args.exp_dir + str(random.randint(0,400)) + args.exp_dir = Path(args.exp_dir) + + logging.info("save arguments to config.yaml...") + save_args(args) + + if args.wandb: wb = wandb.init(project="d2v-adapter", entity="dohe0342", config=vars(args)) + else: wb = None + + world_size = args.world_size + assert world_size >= 1 + if world_size > 1: + mp.spawn(run if not args.add_adapter else run_adapter, + args=(world_size, args, wb), + nprocs=world_size, + join=True + ) + else: + if not args.add_adapter: run(rank=0, world_size=1, args=args, wb=wb) + else: run(rank=0, world_size=1, args=args, wb=wb) + +torch.set_num_threads(1) +torch.set_num_interop_threads(1) + +if __name__ == "__main__": + main() diff --git a/egs/tedlium3/ASR/pruned_transducer_stateless_d2v_v2/utils.py b/egs/tedlium3/ASR/pruned_transducer_stateless_d2v_v2/utils.py new file mode 100644 index 000000000..dd52d8624 --- /dev/null +++ b/egs/tedlium3/ASR/pruned_transducer_stateless_d2v_v2/utils.py @@ -0,0 +1,21 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# This source code is licensed under the MIT license found in the +# LICENSE file in the root directory of this source tree. + +import math +import torch.nn.functional as F + + +def pad_to_multiple(x, multiple, dim=-1, value=0): + # Inspired from https://github.com/lucidrains/local-attention/blob/master/local_attention/local_attention.py#L41 + if x is None: + return None, 0 + tsz = x.size(dim) + m = tsz / multiple + remainder = math.ceil(m) * multiple - tsz + if m.is_integer(): + return x, 0 + pad_offset = (0,) * (-1 - dim) * 2 + + return F.pad(x, (*pad_offset, 0, remainder), value=value), remainder diff --git a/egs/tedlium3/ASR/pruned_transducer_stateless_d2v_v2/zipformer.py b/egs/tedlium3/ASR/pruned_transducer_stateless_d2v_v2/zipformer.py new file mode 100644 index 000000000..b007a7308 --- /dev/null +++ b/egs/tedlium3/ASR/pruned_transducer_stateless_d2v_v2/zipformer.py @@ -0,0 +1,1866 @@ +#!/usr/bin/env python3 +# Copyright (c) 2021 University of Chinese Academy of Sciences (author: Han Zhu) +# +# See ../../../../LICENSE for clarification regarding multiple authors +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import copy +import itertools +import logging +import math +import random +import warnings +from typing import List, Optional, Tuple, Union + +import torch +from encoder_interface import EncoderInterface +from scaling import ( + ScaledLinear, # not as in other dirs.. just scales down initial parameter values. +) +from scaling import ( + ActivationBalancer, + BasicNorm, + DoubleSwish, + Identity, + MaxEig, + ScaledConv1d, + Whiten, + _diag, + penalize_abs_values_gt, + random_clamp, + softmax, +) +from torch import Tensor, nn + +from icefall.dist import get_rank +from icefall.utils import make_pad_mask + + +class Zipformer(EncoderInterface): + """ + Args: + num_features (int): Number of input features + d_model: (int,int): embedding dimension of 2 encoder stacks + attention_dim: (int,int): attention dimension of 2 encoder stacks + nhead (int, int): number of heads + dim_feedforward (int, int): feedforward dimension in 2 encoder stacks + num_encoder_layers (int): number of encoder layers + dropout (float): dropout rate + cnn_module_kernel (int): Kernel size of convolution module + vgg_frontend (bool): whether to use vgg frontend. + warmup_batches (float): number of batches to warm up over + """ + + def __init__( + self, + num_features: int, + output_downsampling_factor: int = 2, + encoder_dims: Tuple[int] = (384, 384), + attention_dim: Tuple[int] = (256, 256), + encoder_unmasked_dims: Tuple[int] = (256, 256), + zipformer_downsampling_factors: Tuple[int] = (2, 4), + nhead: Tuple[int] = (8, 8), + feedforward_dim: Tuple[int] = (1536, 2048), + num_encoder_layers: Tuple[int] = (12, 12), + dropout: float = 0.1, + cnn_module_kernels: Tuple[int] = (31, 31), + pos_dim: int = 4, + warmup_batches: float = 4000.0, + ) -> None: + super(Zipformer, self).__init__() + + self.num_features = num_features + self.encoder_unmasked_dims = encoder_unmasked_dims + assert 0 < encoder_dims[0] <= encoder_dims[1] + self.encoder_dims = encoder_dims + self.encoder_unmasked_dims = encoder_unmasked_dims + self.zipformer_downsampling_factors = zipformer_downsampling_factors + self.output_downsampling_factor = output_downsampling_factor + + # will be written to, see set_batch_count() + self.batch_count = 0 + self.warmup_end = warmup_batches + + for u, d in zip(encoder_unmasked_dims, encoder_dims): + assert u <= d, (u, d) + + # self.encoder_embed converts the input of shape (N, T, num_features) + # to the shape (N, (T - 7)//2, encoder_dims). + # That is, it does two things simultaneously: + # (1) subsampling: T -> (T - 7)//2 + # (2) embedding: num_features -> encoder_dims + self.encoder_embed = Conv2dSubsampling( + num_features, encoder_dims[0], dropout=dropout + ) + + # each one will be ZipformerEncoder or DownsampledZipformerEncoder + encoders = [] + + num_encoders = len(encoder_dims) + for i in range(num_encoders): + encoder_layer = ZipformerEncoderLayer( + encoder_dims[i], + attention_dim[i], + nhead[i], + feedforward_dim[i], + dropout, + cnn_module_kernels[i], + pos_dim, + ) + + # For the segment of the warmup period, we let the Conv2dSubsampling + # layer learn something. Then we start to warm up the other encoders. + encoder = ZipformerEncoder( + encoder_layer, + num_encoder_layers[i], + dropout, + warmup_begin=warmup_batches * (i + 1) / (num_encoders + 1), + warmup_end=warmup_batches * (i + 2) / (num_encoders + 1), + ) + + if zipformer_downsampling_factors[i] != 1: + encoder = DownsampledZipformerEncoder( + encoder, + input_dim=encoder_dims[i - 1] if i > 0 else encoder_dims[0], + output_dim=encoder_dims[i], + downsample=zipformer_downsampling_factors[i], + ) + encoders.append(encoder) + self.encoders = nn.ModuleList(encoders) + + # initializes self.skip_layers and self.skip_modules + self._init_skip_modules() + + self.downsample_output = AttentionDownsample( + encoder_dims[-1], encoder_dims[-1], downsample=output_downsampling_factor + ) + + def _get_layer_skip_dropout_prob(self): + if not self.training: + return 0.0 + batch_count = self.batch_count + min_dropout_prob = 0.025 + + if batch_count > self.warmup_end: + return min_dropout_prob + else: + return 0.5 - (batch_count / self.warmup_end) * (0.5 - min_dropout_prob) + + def _init_skip_modules(self): + """ + If self.zipformer_downampling_factors = (1, 2, 4, 8, 4, 2), then at the input of layer + indexed 4 (in zero indexing), with has subsapling_factor=4, we combine the output of + layers 2 and 3; and at the input of layer indexed 5, which which has subsampling_factor=2, + we combine the outputs of layers 1 and 5. + """ + skip_layers = [] + skip_modules = [] + z = self.zipformer_downsampling_factors + for i in range(len(z)): + if i <= 1 or z[i - 1] <= z[i]: + skip_layers.append(None) + skip_modules.append(SimpleCombinerIdentity()) + else: + # TEMP + for j in range(i - 2, -1, -1): + if z[j] <= z[i] or j == 0: + # TEMP logging statement. + logging.info( + f"At encoder stack {i}, which has downsampling_factor={z[i]}, we will " + f"combine the outputs of layers {j} and {i-1}, with downsampling_factors={z[j]} and {z[i-1]}." + ) + skip_layers.append(j) + skip_modules.append( + SimpleCombiner( + self.encoder_dims[j], + self.encoder_dims[i - 1], + min_weight=(0.0, 0.25), + ) + ) + break + self.skip_layers = skip_layers + self.skip_modules = nn.ModuleList(skip_modules) + + def get_feature_masks(self, x: torch.Tensor) -> List[float]: + # Note: The actual return type is Union[List[float], List[Tensor]], + # but to make torch.jit.script() work, we use List[float] + """ + In eval mode, returns [1.0] * num_encoders; in training mode, returns a number of + randomized feature masks, one per encoder. + On e.g. 15% of frames, these masks will zero out all enocder dims larger than + some supplied number, e.g. >256, so in effect on those frames we are using + a smaller encoer dim. + + We generate the random masks at this level because we want the 2 masks to 'agree' + all the way up the encoder stack. This will mean that the 1st mask will have + mask values repeated self.zipformer_subsampling_factor times. + + Args: + x: the embeddings (needed for the shape and dtype and device), of shape + (num_frames, batch_size, encoder_dims0) + """ + num_encoders = len(self.encoder_dims) + if torch.jit.is_scripting() or not self.training: + return [1.0] * num_encoders + + (num_frames0, batch_size, _encoder_dims0) = x.shape + + assert self.encoder_dims[0] == _encoder_dims0, ( + self.encoder_dims, + _encoder_dims0, + ) + + max_downsampling_factor = max(self.zipformer_downsampling_factors) + + num_frames_max = num_frames0 + max_downsampling_factor - 1 + + feature_mask_dropout_prob = 0.15 + + # frame_mask_max shape: (num_frames_max, batch_size, 1) + frame_mask_max = ( + torch.rand(num_frames_max, batch_size, 1, device=x.device) + > feature_mask_dropout_prob + ).to(x.dtype) + + feature_masks = [] + for i in range(num_encoders): + ds = self.zipformer_downsampling_factors[i] + upsample_factor = max_downsampling_factor // ds + + frame_mask = ( + frame_mask_max.unsqueeze(1) + .expand(num_frames_max, upsample_factor, batch_size, 1) + .reshape(num_frames_max * upsample_factor, batch_size, 1) + ) + num_frames = (num_frames0 + ds - 1) // ds + frame_mask = frame_mask[:num_frames] + feature_mask = torch.ones( + num_frames, + batch_size, + self.encoder_dims[i], + dtype=x.dtype, + device=x.device, + ) + u = self.encoder_unmasked_dims[i] + feature_mask[:, :, u:] *= frame_mask + feature_masks.append(feature_mask) + + return feature_masks + + def forward( + self, + x: torch.Tensor, + x_lens: torch.Tensor, + ) -> Tuple[torch.Tensor, torch.Tensor]: + """ + Args: + x: + The input tensor. Its shape is (batch_size, seq_len, feature_dim). + x_lens: + A tensor of shape (batch_size,) containing the number of frames in + `x` before padding. + Returns: + Return a tuple containing 2 tensors: + - embeddings: its shape is (batch_size, output_seq_len, encoder_dims[-1]) + - lengths, a tensor of shape (batch_size,) containing the number + of frames in `embeddings` before padding. + """ + x = self.encoder_embed(x) + + x = x.permute(1, 0, 2) # (N, T, C) -> (T, N, C) + + lengths = (x_lens - 7) >> 1 + assert x.size(0) == lengths.max().item(), (x.shape, lengths, lengths.max()) + mask = make_pad_mask(lengths) + + outputs = [] + feature_masks = self.get_feature_masks(x) + + for i, (module, skip_module) in enumerate( + zip(self.encoders, self.skip_modules) + ): + ds = self.zipformer_downsampling_factors[i] + k = self.skip_layers[i] + if isinstance(k, int): + layer_skip_dropout_prob = self._get_layer_skip_dropout_prob() + if torch.jit.is_scripting(): + x = skip_module(outputs[k], x) + elif (not self.training) or random.random() > layer_skip_dropout_prob: + x = skip_module(outputs[k], x) + x = module( + x, + feature_mask=feature_masks[i], + src_key_padding_mask=None if mask is None else mask[..., ::ds], + ) + outputs.append(x) + + x = self.downsample_output(x) + # class Downsample has this rounding behavior.. + assert self.output_downsampling_factor == 2, self.output_downsampling_factor + lengths = (lengths + 1) >> 1 + + x = x.permute(1, 0, 2) # (T, N, C) ->(N, T, C) + + return x, lengths + + +class ZipformerEncoderLayer(nn.Module): + """ + ZipformerEncoderLayer is made up of self-attn, feedforward and convolution networks. + + Args: + d_model: the number of expected features in the input (required). + nhead: the number of heads in the multiheadattention models (required). + feedforward_dim: the dimension of the feedforward network model (default=2048). + dropout: the dropout value (default=0.1). + cnn_module_kernel (int): Kernel size of convolution module. + + Examples:: + >>> encoder_layer = ZipformerEncoderLayer(d_model=512, nhead=8) + >>> src = torch.rand(10, 32, 512) + >>> pos_emb = torch.rand(32, 19, 512) + >>> out = encoder_layer(src, pos_emb) + """ + + def __init__( + self, + d_model: int, + attention_dim: int, + nhead: int, + feedforward_dim: int = 2048, + dropout: float = 0.1, + cnn_module_kernel: int = 31, + pos_dim: int = 4, + ) -> None: + super(ZipformerEncoderLayer, self).__init__() + + self.d_model = d_model + + # will be written to, see set_batch_count() + self.batch_count = 0 + + self.self_attn = RelPositionMultiheadAttention( + d_model, + attention_dim, + nhead, + pos_dim, + dropout=0.0, + ) + + self.pooling = PoolingModule(d_model) + + self.feed_forward1 = FeedforwardModule(d_model, feedforward_dim, dropout) + + self.feed_forward2 = FeedforwardModule(d_model, feedforward_dim, dropout) + + self.feed_forward3 = FeedforwardModule(d_model, feedforward_dim, dropout) + + self.conv_module1 = ConvolutionModule(d_model, cnn_module_kernel) + + self.conv_module2 = ConvolutionModule(d_model, cnn_module_kernel) + + self.norm_final = BasicNorm(d_model) + + self.bypass_scale = nn.Parameter(torch.tensor(0.5)) + + # try to ensure the output is close to zero-mean (or at least, zero-median). + self.balancer = ActivationBalancer( + d_model, + channel_dim=-1, + min_positive=0.45, + max_positive=0.55, + max_abs=6.0, + ) + self.whiten = Whiten( + num_groups=1, whitening_limit=5.0, prob=(0.025, 0.25), grad_scale=0.01 + ) + + def get_bypass_scale(self): + if torch.jit.is_scripting() or not self.training: + return self.bypass_scale + if random.random() < 0.1: + # ensure we get grads if self.bypass_scale becomes out of range + return self.bypass_scale + # hardcode warmup period for bypass scale + warmup_period = 20000.0 + initial_clamp_min = 0.75 + final_clamp_min = 0.25 + if self.batch_count > warmup_period: + clamp_min = final_clamp_min + else: + clamp_min = initial_clamp_min - (self.batch_count / warmup_period) * ( + initial_clamp_min - final_clamp_min + ) + return self.bypass_scale.clamp(min=clamp_min, max=1.0) + + def get_dynamic_dropout_rate(self): + # return dropout rate for the dynamic modules (self_attn, pooling, convolution); this + # starts at 0.2 and rapidly decreases to 0. Its purpose is to keep the training stable + # at the beginning, by making the network focus on the feedforward modules. + if torch.jit.is_scripting() or not self.training: + return 0.0 + warmup_period = 2000.0 + initial_dropout_rate = 0.2 + final_dropout_rate = 0.0 + if self.batch_count > warmup_period: + return final_dropout_rate + else: + return initial_dropout_rate - ( + initial_dropout_rate * final_dropout_rate + ) * (self.batch_count / warmup_period) + + def forward( + self, + src: Tensor, + pos_emb: Tensor, + src_mask: Optional[Tensor] = None, + src_key_padding_mask: Optional[Tensor] = None, + ) -> Tensor: + """ + Pass the input through the encoder layer. + + Args: + src: the sequence to the encoder layer (required). + pos_emb: Positional embedding tensor (required). + src_mask: the mask for the src sequence (optional). + src_key_padding_mask: the mask for the src keys per batch (optional). + batch_split: if not None, this layer will only be applied to + + Shape: + src: (S, N, E). + pos_emb: (N, 2*S-1, E) + src_mask: (S, S). + src_key_padding_mask: (N, S). + S is the source sequence length, N is the batch size, E is the feature number + """ + src_orig = src + + # macaron style feed forward module + src = src + self.feed_forward1(src) + + # dropout rate for submodules that interact with time. + dynamic_dropout = self.get_dynamic_dropout_rate() + + # pooling module + if torch.jit.is_scripting(): + src = src + self.pooling(src, key_padding_mask=src_key_padding_mask) + elif random.random() > dynamic_dropout: + src = src + self.pooling(src, key_padding_mask=src_key_padding_mask) + + if torch.jit.is_scripting(): + src_att, attn_weights = self.self_attn( + src, + pos_emb=pos_emb, + attn_mask=src_mask, + key_padding_mask=src_key_padding_mask, + ) + src = src + src_att + + src = src + self.conv_module1( + src, src_key_padding_mask=src_key_padding_mask + ) + + src = src + self.feed_forward2(src) + + src = src + self.self_attn.forward2(src, attn_weights) + + src = src + self.conv_module2( + src, src_key_padding_mask=src_key_padding_mask + ) + else: + use_self_attn = random.random() > dynamic_dropout + if use_self_attn: + src_att, attn_weights = self.self_attn( + src, + pos_emb=pos_emb, + attn_mask=src_mask, + key_padding_mask=src_key_padding_mask, + ) + src = src + src_att + + if random.random() > dynamic_dropout: + src = src + self.conv_module1( + src, src_key_padding_mask=src_key_padding_mask + ) + + src = src + self.feed_forward2(src) + if use_self_attn: + src = src + self.self_attn.forward2(src, attn_weights) + + if random.random() > dynamic_dropout: + src = src + self.conv_module2( + src, src_key_padding_mask=src_key_padding_mask + ) + + src = src + self.feed_forward3(src) + + src = self.norm_final(self.balancer(src)) + + delta = src - src_orig + + src = src_orig + delta * self.get_bypass_scale() + + return self.whiten(src) + + +class ZipformerEncoder(nn.Module): + r"""ZipformerEncoder is a stack of N encoder layers + + Args: + encoder_layer: an instance of the ZipformerEncoderLayer() class (required). + num_layers: the number of sub-encoder-layers in the encoder (required). + + Examples:: + >>> encoder_layer = ZipformerEncoderLayer(d_model=512, nhead=8) + >>> zipformer_encoder = ZipformerEncoder(encoder_layer, num_layers=6) + >>> src = torch.rand(10, 32, 512) + >>> out = zipformer_encoder(src) + """ + + def __init__( + self, + encoder_layer: nn.Module, + num_layers: int, + dropout: float, + warmup_begin: float, + warmup_end: float, + ) -> None: + super().__init__() + # will be written to, see set_batch_count() Note: in inference time this + # may be zero but should be treated as large, we can check if + # self.training is true. + self.batch_count = 0 + self.warmup_begin = warmup_begin + self.warmup_end = warmup_end + # module_seed is for when we need a random number that is unique to the module but + # shared across jobs. It's used to randomly select how many layers to drop, + # so that we can keep this consistent across worker tasks (for efficiency). + self.module_seed = torch.randint(0, 1000, ()).item() + + self.encoder_pos = RelPositionalEncoding(encoder_layer.d_model, dropout) + + self.layers = nn.ModuleList( + [copy.deepcopy(encoder_layer) for i in range(num_layers)] + ) + self.num_layers = num_layers + + assert 0 <= warmup_begin <= warmup_end, (warmup_begin, warmup_end) + + delta = (1.0 / num_layers) * (warmup_end - warmup_begin) + cur_begin = warmup_begin + for i in range(num_layers): + self.layers[i].warmup_begin = cur_begin + cur_begin += delta + self.layers[i].warmup_end = cur_begin + + def get_layers_to_drop(self, rnd_seed: int): + ans = set() + if not self.training: + return ans + + batch_count = self.batch_count + num_layers = len(self.layers) + + def get_layerdrop_prob(layer: int) -> float: + layer_warmup_begin = self.layers[layer].warmup_begin + layer_warmup_end = self.layers[layer].warmup_end + + initial_layerdrop_prob = 0.5 + final_layerdrop_prob = 0.05 + + if batch_count == 0: + # As a special case, if batch_count == 0, return 0 (drop no + # layers). This is rather ugly, I'm afraid; it is intended to + # enable our scan_pessimistic_batches_for_oom() code to work correctly + # so if we are going to get OOM it will happen early. + # also search for 'batch_count' with quotes in this file to see + # how we initialize the warmup count to a random number between + # 0 and 10. + return 0.0 + elif batch_count < layer_warmup_begin: + return initial_layerdrop_prob + elif batch_count > layer_warmup_end: + return final_layerdrop_prob + else: + # linearly interpolate + t = (batch_count - layer_warmup_begin) / layer_warmup_end + assert 0.0 <= t < 1.001, t + return initial_layerdrop_prob + t * ( + final_layerdrop_prob - initial_layerdrop_prob + ) + + shared_rng = random.Random(batch_count + self.module_seed) + independent_rng = random.Random(rnd_seed) + + layerdrop_probs = [get_layerdrop_prob(i) for i in range(num_layers)] + tot = sum(layerdrop_probs) + # Instead of drawing the samples independently, we first randomly decide + # how many layers to drop out, using the same random number generator between + # jobs so that all jobs drop out the same number (this is for speed). + # Then we use an approximate approach to drop out the individual layers + # with their specified probs while reaching this exact target. + num_to_drop = int(tot) + int(shared_rng.random() < (tot - int(tot))) + + layers = list(range(num_layers)) + independent_rng.shuffle(layers) + + # go through the shuffled layers until we get the required number of samples. + if num_to_drop > 0: + for layer in itertools.cycle(layers): + if independent_rng.random() < layerdrop_probs[layer]: + ans.add(layer) + if len(ans) == num_to_drop: + break + if shared_rng.random() < 0.005 or __name__ == "__main__": + logging.info( + f"warmup_begin={self.warmup_begin:.1f}, warmup_end={self.warmup_end:.1f}, " + f"batch_count={batch_count:.1f}, num_to_drop={num_to_drop}, layers_to_drop={ans}" + ) + return ans + + def forward( + self, + src: Tensor, + # Note: The type of feature_mask should be Union[float, Tensor], + # but to make torch.jit.script() work, we use `float` here + feature_mask: float = 1.0, + mask: Optional[Tensor] = None, + src_key_padding_mask: Optional[Tensor] = None, + ) -> Tensor: + r"""Pass the input through the encoder layers in turn. + + Args: + src: the sequence to the encoder (required). + feature_mask: something that broadcasts with src, that we'll multiply `src` + by at every layer. + mask: the mask for the src sequence (optional). + src_key_padding_mask: the mask for the src keys per batch (optional). + + Shape: + src: (S, N, E). + pos_emb: (N, 2*S-1, E) + mask: (S, S). + src_key_padding_mask: (N, S). + S is the source sequence length, T is the target sequence length, N is the batch size, E is the feature number + + Returns: (x, x_no_combine), both of shape (S, N, E) + """ + pos_emb = self.encoder_pos(src) + output = src + + if torch.jit.is_scripting(): + layers_to_drop = [] + else: + rnd_seed = src.numel() + random.randint(0, 1000) + layers_to_drop = self.get_layers_to_drop(rnd_seed) + + output = output * feature_mask + + for i, mod in enumerate(self.layers): + if not torch.jit.is_scripting(): + if i in layers_to_drop: + continue + output = mod( + output, + pos_emb, + src_mask=mask, + src_key_padding_mask=src_key_padding_mask, + ) + + output = output * feature_mask + + return output + + +class DownsampledZipformerEncoder(nn.Module): + r""" + DownsampledZipformerEncoder is a zipformer encoder evaluated at a reduced frame rate, + after convolutional downsampling, and then upsampled again at the output, and combined + with the origin input, so that the output has the same shape as the input. + """ + + def __init__( + self, encoder: nn.Module, input_dim: int, output_dim: int, downsample: int + ): + super(DownsampledZipformerEncoder, self).__init__() + self.downsample_factor = downsample + self.downsample = AttentionDownsample(input_dim, output_dim, downsample) + self.encoder = encoder + self.upsample = SimpleUpsample(output_dim, downsample) + self.out_combiner = SimpleCombiner( + input_dim, output_dim, min_weight=(0.0, 0.25) + ) + + def forward( + self, + src: Tensor, + # Note: the type of feature_mask should be Unino[float, Tensor], + # but to make torch.jit.script() happ, we use float here + feature_mask: float = 1.0, + mask: Optional[Tensor] = None, + src_key_padding_mask: Optional[Tensor] = None, + ) -> Tensor: + r"""Downsample, go through encoder, upsample. + + Args: + src: the sequence to the encoder (required). + feature_mask: something that broadcasts with src, that we'll multiply `src` + by at every layer. feature_mask is expected to be already downsampled by + self.downsample_factor. + mask: the mask for the src sequence (optional). CAUTION: we need to downsample + this, if we are to support it. Won't work correctly yet. + src_key_padding_mask: the mask for the src keys per batch (optional). Should + be downsampled already. + + Shape: + src: (S, N, E). + mask: (S, S). + src_key_padding_mask: (N, S). + S is the source sequence length, T is the target sequence length, N is the batch size, E is the feature number + + Returns: output of shape (S, N, F) where F is the number of output features + (output_dim to constructor) + """ + src_orig = src + src = self.downsample(src) + ds = self.downsample_factor + if mask is not None: + mask = mask[::ds, ::ds] + + src = self.encoder( + src, + feature_mask=feature_mask, + mask=mask, + src_key_padding_mask=mask, + ) + src = self.upsample(src) + # remove any extra frames that are not a multiple of downsample_factor + src = src[: src_orig.shape[0]] + + return self.out_combiner(src_orig, src) + + +class AttentionDownsample(torch.nn.Module): + """ + Does downsampling with attention, by weighted sum, and a projection.. + """ + + def __init__(self, in_channels: int, out_channels: int, downsample: int): + """ + Require out_channels > in_channels. + """ + super(AttentionDownsample, self).__init__() + self.query = nn.Parameter(torch.randn(in_channels) * (in_channels**-0.5)) + + # fill in the extra dimensions with a projection of the input + if out_channels > in_channels: + self.extra_proj = nn.Linear( + in_channels * downsample, out_channels - in_channels, bias=False + ) + else: + self.extra_proj = None + self.downsample = downsample + + def forward(self, src: Tensor) -> Tensor: + """ + x: (seq_len, batch_size, in_channels) + Returns a tensor of shape + ( (seq_len+downsample-1)//downsample, batch_size, out_channels) + """ + (seq_len, batch_size, in_channels) = src.shape + ds = self.downsample + d_seq_len = (seq_len + ds - 1) // ds + + # Pad to an exact multiple of self.downsample + if seq_len != d_seq_len * ds: + # right-pad src, repeating the last element. + pad = d_seq_len * ds - seq_len + src_extra = src[src.shape[0] - 1 :].expand(pad, src.shape[1], src.shape[2]) + src = torch.cat((src, src_extra), dim=0) + assert src.shape[0] == d_seq_len * ds, (src.shape[0], d_seq_len, ds) + + src = src.reshape(d_seq_len, ds, batch_size, in_channels) + scores = (src * self.query).sum(dim=-1, keepdim=True) + + scores = penalize_abs_values_gt(scores, limit=10.0, penalty=1.0e-04) + + weights = scores.softmax(dim=1) + + # ans1 is the first `in_channels` channels of the output + ans = (src * weights).sum(dim=1) + src = src.permute(0, 2, 1, 3).reshape(d_seq_len, batch_size, ds * in_channels) + + if self.extra_proj is not None: + ans2 = self.extra_proj(src) + ans = torch.cat((ans, ans2), dim=2) + return ans + + +class SimpleUpsample(torch.nn.Module): + """ + A very simple form of upsampling that mostly just repeats the input, but + also adds a position-specific bias. + """ + + def __init__(self, num_channels: int, upsample: int): + super(SimpleUpsample, self).__init__() + self.bias = nn.Parameter(torch.randn(upsample, num_channels) * 0.01) + + def forward(self, src: Tensor) -> Tensor: + """ + x: (seq_len, batch_size, num_channels) + Returns a tensor of shape + ( (seq_len*upsample), batch_size, num_channels) + """ + upsample = self.bias.shape[0] + (seq_len, batch_size, num_channels) = src.shape + src = src.unsqueeze(1).expand(seq_len, upsample, batch_size, num_channels) + src = src + self.bias.unsqueeze(1) + src = src.reshape(seq_len * upsample, batch_size, num_channels) + return src + + +class SimpleCombinerIdentity(nn.Module): + def __init__(self, *args, **kwargs): + super().__init__() + + def forward(self, src1: Tensor, src2: Tensor) -> Tensor: + return src1 + + +class SimpleCombiner(torch.nn.Module): + """ + A very simple way of combining 2 vectors of 2 different dims, via a + learned weighted combination in the shared part of the dim. + Args: + dim1: the dimension of the first input, e.g. 256 + dim2: the dimension of the second input, e.g. 384. + The output will have the same dimension as dim2. + """ + + def __init__(self, dim1: int, dim2: int, min_weight: Tuple[float] = (0.0, 0.0)): + super(SimpleCombiner, self).__init__() + assert dim2 >= dim1, (dim2, dim1) + self.weight1 = nn.Parameter(torch.zeros(())) + self.min_weight = min_weight + + def forward(self, src1: Tensor, src2: Tensor) -> Tensor: + """ + src1: (*, dim1) + src2: (*, dim2) + + Returns: a tensor of shape (*, dim2) + """ + assert src1.shape[:-1] == src2.shape[:-1], (src1.shape, src2.shape) + + weight1 = self.weight1 + if not torch.jit.is_scripting(): + if ( + self.training + and random.random() < 0.25 + and self.min_weight != (0.0, 0.0) + ): + weight1 = weight1.clamp( + min=self.min_weight[0], max=1.0 - self.min_weight[1] + ) + + src1 = src1 * weight1 + src2 = src2 * (1.0 - weight1) + + src1_dim = src1.shape[-1] + src2_dim = src2.shape[-1] + if src1_dim != src2_dim: + if src1_dim < src2_dim: + src1 = torch.nn.functional.pad(src1, (0, src2_dim - src1_dim)) + else: + src1 = src1[:src2_dim] + + return src1 + src2 + + +class RelPositionalEncoding(torch.nn.Module): + """Relative positional encoding module. + + See : Appendix B in "Transformer-XL: Attentive Language Models Beyond a Fixed-Length Context" + Modified from https://github.com/espnet/espnet/blob/master/espnet/nets/pytorch_backend/transformer/embedding.py + + Args: + d_model: Embedding dimension. + dropout_rate: Dropout rate. + max_len: Maximum input length. + + """ + + def __init__(self, d_model: int, dropout_rate: float, max_len: int = 5000) -> None: + """Construct a PositionalEncoding object.""" + super(RelPositionalEncoding, self).__init__() + self.d_model = d_model + self.dropout = torch.nn.Dropout(dropout_rate) + self.pe = None + self.extend_pe(torch.tensor(0.0).expand(1, max_len)) + + def extend_pe(self, x: Tensor) -> None: + """Reset the positional encodings.""" + if self.pe is not None: + # self.pe contains both positive and negative parts + # the length of self.pe is 2 * input_len - 1 + if self.pe.size(1) >= x.size(0) * 2 - 1: + # Note: TorchScript doesn't implement operator== for torch.Device + if self.pe.dtype != x.dtype or str(self.pe.device) != str(x.device): + self.pe = self.pe.to(dtype=x.dtype, device=x.device) + return + # Suppose `i` means to the position of query vecotr and `j` means the + # position of key vector. We use position relative positions when keys + # are to the left (i>j) and negative relative positions otherwise (i Tensor: + """Add positional encoding. + + Args: + x (torch.Tensor): Input tensor (time, batch, `*`). + + Returns: + torch.Tensor: Encoded tensor (batch, time, `*`). + torch.Tensor: Encoded tensor (batch, 2*time-1, `*`). + + """ + self.extend_pe(x) + pos_emb = self.pe[ + :, + self.pe.size(1) // 2 + - x.size(0) + + 1 : self.pe.size(1) // 2 # noqa E203 + + x.size(0), + ] + return self.dropout(pos_emb) + + +class RelPositionMultiheadAttention(nn.Module): + r"""Multi-Head Attention layer with relative position encoding + + This is a quite heavily modified from: "Transformer-XL: Attentive Language Models Beyond a Fixed-Length Context", + we have to write up the differences. + + + Args: + embed_dim: total dimension of the model. + attention_dim: dimension in the attention module, may be less or more than embed_dim + but must be a multiple of num_heads. + num_heads: parallel attention heads. + dropout: a Dropout layer on attn_output_weights. Default: 0.0. + + Examples:: + + >>> rel_pos_multihead_attn = RelPositionMultiheadAttention(embed_dim, num_heads) + >>> attn_output, attn_output_weights = multihead_attn(query, key, value, pos_emb) + """ + + def __init__( + self, + embed_dim: int, + attention_dim: int, + num_heads: int, + pos_dim: int, + dropout: float = 0.0, + ) -> None: + super(RelPositionMultiheadAttention, self).__init__() + self.embed_dim = embed_dim + self.attention_dim = attention_dim + self.num_heads = num_heads + self.dropout = dropout + self.head_dim = attention_dim // num_heads + self.pos_dim = pos_dim + assert self.head_dim % 2 == 0, self.head_dim + assert self.head_dim * num_heads == attention_dim, ( + self.head_dim, + num_heads, + attention_dim, + ) + + # the initial_scale is supposed to take over the "scaling" factor of + # head_dim ** -0.5, dividing it between the query and key. + in_proj_dim = ( + 2 * attention_dim + + attention_dim // 2 # query, key + + pos_dim * num_heads # value + ) # positional encoding query + + self.in_proj = ScaledLinear( + embed_dim, in_proj_dim, bias=True, initial_scale=self.head_dim**-0.25 + ) + + # self.whiten_values is applied on the values in forward(); + # it just copies the keys but prevents low-rank distribution by modifying grads. + self.whiten_values = Whiten( + num_groups=num_heads, + whitening_limit=2.0, + prob=(0.025, 0.25), + grad_scale=0.025, + ) + self.whiten_keys = Whiten( + num_groups=num_heads, + whitening_limit=2.0, + prob=(0.025, 0.25), + grad_scale=0.025, + ) + + # linear transformation for positional encoding. + self.linear_pos = ScaledLinear( + embed_dim, num_heads * pos_dim, bias=False, initial_scale=0.05 + ) + + # the following are for diagnosics only, see --print-diagnostics option. + # they only copy their inputs. + self.copy_pos_query = Identity() + self.copy_query = Identity() + + self.out_proj = ScaledLinear( + attention_dim // 2, embed_dim, bias=True, initial_scale=0.05 + ) + + self.in_proj2 = nn.Linear(embed_dim, attention_dim // 2, bias=False) + self.out_proj2 = ScaledLinear( + attention_dim // 2, embed_dim, bias=True, initial_scale=0.05 + ) + # self.whiten_values2 is applied on the values in forward2() + self.whiten_values2 = Whiten( + num_groups=num_heads, + whitening_limit=2.0, + prob=(0.025, 0.25), + grad_scale=0.025, + ) + + def forward( + self, + x: Tensor, + pos_emb: Tensor, + key_padding_mask: Optional[Tensor] = None, + attn_mask: Optional[Tensor] = None, + ) -> Tuple[Tensor, Tensor]: + r""" + Args: + x: input to be projected to query, key, value + pos_emb: Positional embedding tensor + key_padding_mask: if provided, specified padding elements in the key will + be ignored by the attention. When given a binary mask and a value is True, + the corresponding value on the attention layer will be ignored. When given + a byte mask and a value is non-zero, the corresponding value on the attention + layer will be ignored + attn_mask: 2D or 3D mask that prevents attention to certain positions. A 2D mask will be broadcasted for all + the batches while a 3D mask allows to specify a different mask for the entries of each batch. + + Shape: + - Inputs: + - x: :math:`(L, N, E)` where L is the target sequence length, N is the batch size, E is + the embedding dimension. + - pos_emb: :math:`(N, 2*L-1, E)` where L is the target sequence length, N is the batch size, E is + the embedding dimension. + - key_padding_mask: :math:`(N, S)` where N is the batch size, S is the source sequence length. + If a ByteTensor is provided, the non-zero positions will be ignored while the position + with the zero positions will be unchanged. If a BoolTensor is provided, the positions with the + value of ``True`` will be ignored while the position with the value of ``False`` will be unchanged. + - attn_mask: 2D mask :math:`(L, S)` where L is the target sequence length, S is the source sequence length. + 3D mask :math:`(N*num_heads, L, S)` where N is the batch size, L is the target sequence length, + S is the source sequence length. attn_mask ensure that position i is allowed to attend the unmasked + positions. If a ByteTensor is provided, the non-zero positions are not allowed to attend + while the zero positions will be unchanged. If a BoolTensor is provided, positions with ``True`` + is not allowed to attend while ``False`` values will be unchanged. If a FloatTensor + is provided, it will be added to the attention weight. + + - Returns: (attn_output, attn_weights) + + - attn_output: :math:`(S, N, E)` where S is the sequence length, N is the batch size, + E is the embedding dimension. + - attn_weights: :math:`(N * N, S, S)` where N is the batch size, H is the num-heads + and S is the sequence length. + """ + x, weights = self.multi_head_attention_forward( + self.in_proj(x), + self.linear_pos(pos_emb), + self.attention_dim, + self.num_heads, + self.dropout, + self.out_proj.weight, + self.out_proj.bias, + training=self.training, + key_padding_mask=key_padding_mask, + attn_mask=attn_mask, + ) + return x, weights + + def multi_head_attention_forward( + self, + x_proj: Tensor, + pos: Tensor, + attention_dim: int, + num_heads: int, + dropout_p: float, + out_proj_weight: Tensor, + out_proj_bias: Tensor, + training: bool = True, + key_padding_mask: Optional[Tensor] = None, + attn_mask: Optional[Tensor] = None, + ) -> Tuple[Tensor, Tensor]: + r""" + Args: + x_proj: the projected input, to be split into query, key, value. + pos: head-specific biases arising from the positional embeddings. + attention_dim: dimension inside attention mechanism + num_heads: parallel attention heads. + dropout_p: probability of an element to be zeroed. + out_proj_weight, out_proj_bias: the output projection weight and bias. + training: apply dropout if is ``True``. + key_padding_mask: if provided, specified padding elements in the key will + be ignored by the attention. This is an binary mask. When the value is True, + the corresponding value on the attention layer will be filled with -inf. + attn_mask: 2D or 3D mask that prevents attention to certain positions. A 2D mask will be broadcasted for all + the batches while a 3D mask allows to specify a different mask for the entries of each batch. + + Shape: + Inputs: + - x: :math:`(L, N, 7 * A // 2)` where L is the target sequence length, N is the batch size, A is + the attention dimension. Will be split into (query, key, value, pos). + - pos: :math:`(N, 2*L-1, A//2)` or :math:`(1, 2*L-1, A//2)` where L is the sequence + length, N is the batch size, and A is the attention dim. + - key_padding_mask: :math:`(N, S)` where N is the batch size, S is the source sequence length. + If a ByteTensor is provided, the non-zero positions will be ignored while the zero positions + will be unchanged. If a BoolTensor is provided, the positions with the + value of ``True`` will be ignored while the position with the value of ``False`` will be unchanged. + - attn_mask: 2D mask :math:`(L, S)` where L is the target sequence length, S is the source sequence length. + 3D mask :math:`(N*num_heads, L, S)` where N is the batch size, L is the target sequence length, + S is the source sequence length. attn_mask ensures that position i is allowed to attend the unmasked + positions. If a ByteTensor is provided, the non-zero positions are not allowed to attend + while the zero positions will be unchanged. If a BoolTensor is provided, positions with ``True`` + are not allowed to attend while ``False`` values will be unchanged. If a FloatTensor + is provided, it will be added to the attention weight. + + Outputs: + - attn_output: :math:`(L, N, E)` where L is the target sequence length, N is the batch size, + E is the embedding dimension. + - attn_weights: :math:`(N * H, S, S)` where N is the batch size, + H is the num-heads, S is the sequence length. + """ + + seq_len, bsz, _ = x_proj.size() + + head_dim = attention_dim // num_heads + pos_dim = self.pos_dim # positional-encoding dim per head + assert ( + head_dim * num_heads == attention_dim + ), f"attention_dim must be divisible by num_heads: {head_dim}, {num_heads}, {attention_dim}" + + # self-attention + q = x_proj[..., 0:attention_dim] + k = x_proj[..., attention_dim : 2 * attention_dim] + value_dim = attention_dim // 2 + v = x_proj[..., 2 * attention_dim : 2 * attention_dim + value_dim] + # p is the position-encoding query, its dimension is num_heads*pos_dim.. + p = x_proj[..., 2 * attention_dim + value_dim :] + + k = self.whiten_keys(k) # does nothing in the forward pass. + v = self.whiten_values(v) # does nothing in the forward pass. + q = self.copy_query(q) # for diagnostics only, does nothing. + p = self.copy_pos_query(p) # for diagnostics only, does nothing. + + if attn_mask is not None: + assert ( + attn_mask.dtype == torch.float32 + or attn_mask.dtype == torch.float64 + or attn_mask.dtype == torch.float16 + or attn_mask.dtype == torch.uint8 + or attn_mask.dtype == torch.bool + ), "Only float, byte, and bool types are supported for attn_mask, not {}".format( + attn_mask.dtype + ) + if attn_mask.dtype == torch.uint8: + warnings.warn( + "Byte tensor for attn_mask is deprecated. Use bool tensor instead." + ) + attn_mask = attn_mask.to(torch.bool) + + if attn_mask.dim() == 2: + attn_mask = attn_mask.unsqueeze(0) + if list(attn_mask.size()) != [1, seq_len, seq_len]: + raise RuntimeError("The size of the 2D attn_mask is not correct.") + elif attn_mask.dim() == 3: + if list(attn_mask.size()) != [ + bsz * num_heads, + seq_len, + seq_len, + ]: + raise RuntimeError("The size of the 3D attn_mask is not correct.") + else: + raise RuntimeError( + "attn_mask's dimension {} is not supported".format(attn_mask.dim()) + ) + # attn_mask's dim is 3 now. + + # convert ByteTensor key_padding_mask to bool + if key_padding_mask is not None and key_padding_mask.dtype == torch.uint8: + warnings.warn( + "Byte tensor for key_padding_mask is deprecated. Use bool tensor instead." + ) + key_padding_mask = key_padding_mask.to(torch.bool) + + q = q.reshape(seq_len, bsz, num_heads, head_dim) + p = p.reshape(seq_len, bsz, num_heads, pos_dim) + k = k.reshape(seq_len, bsz, num_heads, head_dim) + v = v.reshape(seq_len, bsz * num_heads, head_dim // 2).transpose(0, 1) + + if key_padding_mask is not None: + assert key_padding_mask.size(0) == bsz, "{} == {}".format( + key_padding_mask.size(0), bsz + ) + assert key_padding_mask.size(1) == seq_len, "{} == {}".format( + key_padding_mask.size(1), seq_len + ) + + q = q.permute(1, 2, 0, 3) # (batch, head, time1, head_dim) + p = p.permute(1, 2, 0, 3) # (batch, head, time1, pos_dim) + k = k.permute(1, 2, 3, 0) # (batch, head, d_k, time2) + + seq_len2 = 2 * seq_len - 1 + pos = pos.reshape(1, seq_len2, num_heads, pos_dim).permute(0, 2, 3, 1) + # pos shape now: (batch, head, pos_dim, seq_len2) + + # (batch, head, time1, pos_dim) x (1, head, pos_dim, seq_len2) -> (batch, head, time1, seq_len2) + # [where seq_len2 represents relative position.] + pos_weights = torch.matmul(p, pos) + # the following .as_strided() expression converts the last axis of pos_weights from relative + # to absolute position. I don't know whether I might have got the time-offsets backwards or + # not, but let this code define which way round it is supposed to be. + pos_weights = pos_weights.as_strided( + (bsz, num_heads, seq_len, seq_len), + ( + pos_weights.stride(0), + pos_weights.stride(1), + pos_weights.stride(2) - pos_weights.stride(3), + pos_weights.stride(3), + ), + storage_offset=pos_weights.stride(3) * (seq_len - 1), + ) + + # caution: they are really scores at this point. + attn_output_weights = torch.matmul(q, k) + pos_weights + + if not torch.jit.is_scripting(): + if training and random.random() < 0.1: + # This is a harder way of limiting the attention scores to not be too large. + # It incurs a penalty if any of them has an absolute value greater than 50.0. + # this should be outside the normal range of the attention scores. We use + # this mechanism instead of, say, a limit on entropy, because once the entropy + # gets very small gradients through the softmax can become very small, and + # some mechanisms like that become ineffective. + attn_output_weights = penalize_abs_values_gt( + attn_output_weights, limit=25.0, penalty=1.0e-04 + ) + + # attn_output_weights: (batch, head, time1, time2) + attn_output_weights = attn_output_weights.view( + bsz * num_heads, seq_len, seq_len + ) + + assert list(attn_output_weights.size()) == [ + bsz * num_heads, + seq_len, + seq_len, + ] + + if attn_mask is not None: + if attn_mask.dtype == torch.bool: + attn_output_weights.masked_fill_(attn_mask, float("-inf")) + else: + attn_output_weights += attn_mask + + if key_padding_mask is not None: + attn_output_weights = attn_output_weights.view( + bsz, num_heads, seq_len, seq_len + ) + attn_output_weights = attn_output_weights.masked_fill( + key_padding_mask.unsqueeze(1).unsqueeze(2), + float("-inf"), + ) + attn_output_weights = attn_output_weights.view( + bsz * num_heads, seq_len, seq_len + ) + + # Using this version of softmax, defined in scaling.py, + # should save a little of the memory used in backprop by, if + # we are in automatic mixed precision mode (amp) == autocast, + # only storing the half-precision output for backprop purposes. + attn_output_weights = softmax(attn_output_weights, dim=-1) + + attn_output_weights = nn.functional.dropout( + attn_output_weights, p=dropout_p, training=training + ) + + attn_output = torch.bmm(attn_output_weights, v) + assert list(attn_output.size()) == [bsz * num_heads, seq_len, head_dim // 2] + attn_output = ( + attn_output.transpose(0, 1) + .contiguous() + .view(seq_len, bsz, attention_dim // 2) + ) + attn_output = nn.functional.linear(attn_output, out_proj_weight, out_proj_bias) + + return attn_output, attn_output_weights + + def forward2( + self, + x: Tensor, + attn_weights: Tensor, + ) -> Tensor: + """ + Second forward function, where we re-use the attn_weights returned by the first forward function + but with different input. + Args: + x: input, of shape (seq_len, batch_size, embed_dim) + attn_weights: attention weights returned by forward(), of shape (batch_size * num_heads, seq_len, seq_len) + Returns: + output of the same shape as x, i.e. (seq_len, batch_size, embed_dim) + """ + num_heads = self.num_heads + (seq_len, bsz, embed_dim) = x.shape + head_dim = self.attention_dim // num_heads + # v: (tgt_len, bsz, embed_dim // 2) + v = self.in_proj2(x) + v = self.whiten_values2(v) # does nothing in the forward pass. + v = v.reshape(seq_len, bsz * num_heads, head_dim // 2).transpose(0, 1) + + # now v: (bsz * num_heads, seq_len, head_dim // 2) + attn_output = torch.bmm(attn_weights, v) + + if not torch.jit.is_scripting(): + if random.random() < 0.001 or __name__ == "__main__": + self._print_attn_stats(attn_weights, attn_output) + + # attn_output: (bsz * num_heads, seq_len, head_dim) + attn_output = ( + attn_output.transpose(0, 1) + .contiguous() + .view(seq_len, bsz, self.attention_dim // 2) + ) + # returned value is of shape (seq_len, bsz, embed_dim), like x. + return self.out_proj2(attn_output) + + def _print_attn_stats(self, attn_weights: Tensor, attn_output: Tensor): + # attn_weights: (batch_size * num_heads, seq_len, seq_len) + # attn_output: (bsz * num_heads, seq_len, head_dim) + (n, seq_len, head_dim) = attn_output.shape + num_heads = self.num_heads + bsz = n // num_heads + + with torch.no_grad(): + with torch.cuda.amp.autocast(enabled=False): + attn_weights = attn_weights.to(torch.float32) + attn_output = attn_output.to(torch.float32) + attn_weights_entropy = ( + -((attn_weights + 1.0e-20).log() * attn_weights) + .sum(dim=-1) + .reshape(bsz, num_heads, seq_len) + .mean(dim=(0, 2)) + ) + attn_output = attn_output.reshape(bsz, num_heads, seq_len, head_dim) + attn_output = attn_output.permute(1, 0, 2, 3).reshape( + num_heads, bsz * seq_len, head_dim + ) + attn_output_mean = attn_output.mean(dim=1, keepdim=True) + attn_output = attn_output - attn_output_mean + attn_covar = torch.matmul(attn_output.transpose(1, 2), attn_output) / ( + bsz * seq_len + ) + # attn_covar: (num_heads, head_dim, head_dim) + # eigs, _ = torch.symeig(attn_covar) + # logging.info(f"attn_weights_entropy = {attn_weights_entropy}, output_eigs = {eigs}") + + attn_covar = _diag(attn_covar).mean(dim=1) # (num_heads,) + embed_dim = self.in_proj2.weight.shape[1] + in_proj_covar = ( + self.in_proj2.weight.reshape(num_heads, head_dim, embed_dim) ** 2 + ).mean(dim=(1, 2)) + out_proj_covar = ( + self.out_proj2.weight.reshape(embed_dim, num_heads, head_dim) ** 2 + ).mean(dim=(0, 2)) + logging.info( + f"attn_weights_entropy = {attn_weights_entropy}, covar={attn_covar}, in_proj_covar={in_proj_covar}, out_proj_covar={out_proj_covar}" + ) + + +class PoolingModule(nn.Module): + """ + Averages the input over the time dimension and project with a square matrix. + """ + + def __init__(self, d_model: int): + super().__init__() + self.proj = ScaledLinear(d_model, d_model, initial_scale=0.1, bias=False) + + def forward(self, x: Tensor, key_padding_mask: Optional[Tensor] = None): + """ + Args: + x: a Tensor of shape (T, N, C) + key_padding_mask: a Tensor of bool, of shape (N, T), with True in masked + positions. + Returns: + a Tensor of shape (1, N, C) + """ + if key_padding_mask is not None: + pooling_mask = key_padding_mask.logical_not().to(x.dtype) # (N, T) + pooling_mask = pooling_mask / pooling_mask.sum(dim=1, keepdim=True) + pooling_mask = pooling_mask.transpose(0, 1).contiguous().unsqueeze(-1) + # now pooling_mask: (T, N, 1) + x = (x * pooling_mask).sum(dim=0, keepdim=True) + else: + num_frames = x.shape[0] + pooling_mask = 1.0 / num_frames + x = (x * pooling_mask).sum(dim=0, keepdim=True) + + x = self.proj(x) + return x + + +class FeedforwardModule(nn.Module): + """Feedforward module in Zipformer model.""" + + def __init__(self, d_model: int, feedforward_dim: int, dropout: float): + super(FeedforwardModule, self).__init__() + self.in_proj = nn.Linear(d_model, feedforward_dim) + self.balancer = ActivationBalancer( + feedforward_dim, channel_dim=-1, max_abs=10.0, min_prob=0.25 + ) + self.activation = DoubleSwish() + self.dropout = nn.Dropout(dropout) + self.out_proj = ScaledLinear(feedforward_dim, d_model, initial_scale=0.01) + + def forward(self, x: Tensor): + x = self.in_proj(x) + x = self.balancer(x) + x = self.activation(x) + x = self.dropout(x) + x = self.out_proj(x) + return x + + +class ConvolutionModule(nn.Module): + """ConvolutionModule in Zipformer model. + Modified from https://github.com/espnet/espnet/blob/master/espnet/nets/pytorch_backend/zipformer/convolution.py + + Args: + channels (int): The number of channels of conv layers. + kernel_size (int): Kernerl size of conv layers. + bias (bool): Whether to use bias in conv layers (default=True). + + """ + + def __init__(self, channels: int, kernel_size: int, bias: bool = True) -> None: + """Construct an ConvolutionModule object.""" + super(ConvolutionModule, self).__init__() + # kernerl_size should be a odd number for 'SAME' padding + assert (kernel_size - 1) % 2 == 0, kernel_size + + self.pointwise_conv1 = nn.Conv1d( + channels, + 2 * channels, + kernel_size=1, + stride=1, + padding=0, + bias=bias, + ) + + # after pointwise_conv1 we put x through a gated linear unit (nn.functional.glu). + # For most layers the normal rms value of channels of x seems to be in the range 1 to 4, + # but sometimes, for some reason, for layer 0 the rms ends up being very large, + # between 50 and 100 for different channels. This will cause very peaky and + # sparse derivatives for the sigmoid gating function, which will tend to make + # the loss function not learn effectively. (for most layers the average absolute values + # are in the range 0.5..9.0, and the average p(x>0), i.e. positive proportion, + # at the output of pointwise_conv1.output is around 0.35 to 0.45 for different + # layers, which likely breaks down as 0.5 for the "linear" half and + # 0.2 to 0.3 for the part that goes into the sigmoid. The idea is that if we + # constrain the rms values to a reasonable range via a constraint of max_abs=10.0, + # it will be in a better position to start learning something, i.e. to latch onto + # the correct range. + self.deriv_balancer1 = ActivationBalancer( + 2 * channels, + channel_dim=1, + max_abs=10.0, + min_positive=0.05, + max_positive=1.0, + ) + + self.depthwise_conv = nn.Conv1d( + channels, + channels, + kernel_size, + stride=1, + padding=(kernel_size - 1) // 2, + groups=channels, + bias=bias, + ) + + self.deriv_balancer2 = ActivationBalancer( + channels, + channel_dim=1, + min_positive=0.05, + max_positive=1.0, + max_abs=20.0, + ) + + self.activation = DoubleSwish() + + self.pointwise_conv2 = ScaledConv1d( + channels, + channels, + kernel_size=1, + stride=1, + padding=0, + bias=bias, + initial_scale=0.05, + ) + + def forward( + self, + x: Tensor, + src_key_padding_mask: Optional[Tensor] = None, + ) -> Tensor: + """Compute convolution module. + + Args: + x: Input tensor (#time, batch, channels). + src_key_padding_mask: the mask for the src keys per batch (optional): + (batch, #time), contains bool in masked positions. + + Returns: + Tensor: Output tensor (#time, batch, channels). + + """ + # exchange the temporal dimension and the feature dimension + x = x.permute(1, 2, 0) # (#batch, channels, time). + + # GLU mechanism + x = self.pointwise_conv1(x) # (batch, 2*channels, time) + + x = self.deriv_balancer1(x) + x = nn.functional.glu(x, dim=1) # (batch, channels, time) + + if src_key_padding_mask is not None: + x.masked_fill_(src_key_padding_mask.unsqueeze(1).expand_as(x), 0.0) + + # 1D Depthwise Conv + x = self.depthwise_conv(x) + + x = self.deriv_balancer2(x) + x = self.activation(x) + + x = self.pointwise_conv2(x) # (batch, channel, time) + + return x.permute(2, 0, 1) + + +class Conv2dSubsampling(nn.Module): + """Convolutional 2D subsampling (to 1/4 length). + + Convert an input of shape (N, T, idim) to an output + with shape (N, T', odim), where + T' = (T-3)//2 - 2 == (T-7)//2 + + It is based on + https://github.com/espnet/espnet/blob/master/espnet/nets/pytorch_backend/transformer/subsampling.py # noqa + """ + + def __init__( + self, + in_channels: int, + out_channels: int, + layer1_channels: int = 8, + layer2_channels: int = 32, + layer3_channels: int = 128, + dropout: float = 0.1, + ) -> None: + """ + Args: + in_channels: + Number of channels in. The input shape is (N, T, in_channels). + Caution: It requires: T >=7, in_channels >=7 + out_channels + Output dim. The output shape is (N, (T-7)//2, out_channels) + layer1_channels: + Number of channels in layer1 + layer2_channels: + Number of channels in layer2 + layer3_channels: + Number of channels in layer3 + """ + assert in_channels >= 7, in_channels + super().__init__() + + self.conv = nn.Sequential( + nn.Conv2d( + in_channels=1, + out_channels=layer1_channels, + kernel_size=3, + padding=(0, 1), # (time, freq) + ), + ActivationBalancer(layer1_channels, channel_dim=1), + DoubleSwish(), + nn.Conv2d( + in_channels=layer1_channels, + out_channels=layer2_channels, + kernel_size=3, + stride=2, + padding=0, + ), + ActivationBalancer(layer2_channels, channel_dim=1), + DoubleSwish(), + nn.Conv2d( + in_channels=layer2_channels, + out_channels=layer3_channels, + kernel_size=3, + stride=(1, 2), # (time, freq) + ), + ActivationBalancer(layer3_channels, channel_dim=1), + DoubleSwish(), + ) + out_height = (((in_channels - 1) // 2) - 1) // 2 + self.out = ScaledLinear(out_height * layer3_channels, out_channels) + self.dropout = nn.Dropout(dropout) + + def forward(self, x: torch.Tensor) -> torch.Tensor: + """Subsample x. + + Args: + x: + Its shape is (N, T, idim). + + Returns: + Return a tensor of shape (N, (T-7)//2, odim) + """ + # On entry, x is (N, T, idim) + x = x.unsqueeze(1) # (N, T, idim) -> (N, 1, T, idim) i.e., (N, C, H, W) + x = self.conv(x) + # Now x is of shape (N, odim, (T-7)//2, ((idim-1)//2 - 1)//2) + b, c, t, f = x.size() + x = self.out(x.transpose(1, 2).reshape(b, t, c * f)) + # Now x is of shape (N, (T-7)//2, odim) + x = self.dropout(x) + return x + + +class AttentionCombine(nn.Module): + """ + This module combines a list of Tensors, all with the same shape, to + produce a single output of that same shape which, in training time, + is a random combination of all the inputs; but which in test time + will be just the last input. + + All but the last input will have a linear transform before we + randomly combine them; these linear transforms will be initialized + to the identity transform. + + The idea is that the list of Tensors will be a list of outputs of multiple + zipformer layers. This has a similar effect as iterated loss. (See: + DEJA-VU: DOUBLE FEATURE PRESENTATION AND ITERATED LOSS IN DEEP TRANSFORMER + NETWORKS). + """ + + def __init__( + self, + num_channels: int, + num_inputs: int, + random_prob: float = 0.25, + single_prob: float = 0.333, + ) -> None: + """ + Args: + num_channels: + the number of channels + num_inputs: + The number of tensor inputs, which equals the number of layers' + outputs that are fed into this module. E.g. in an 18-layer neural + net if we output layers 16, 12, 18, num_inputs would be 3. + random_prob: + the probability with which we apply a nontrivial mask, in training + mode. + single_prob: + the probability with which we mask to allow just a single + module's output (in training) + """ + super().__init__() + + self.random_prob = random_prob + self.single_prob = single_prob + self.weight = torch.nn.Parameter(torch.zeros(num_channels, num_inputs)) + self.bias = torch.nn.Parameter(torch.zeros(num_inputs)) + + assert 0 <= random_prob <= 1, random_prob + assert 0 <= single_prob <= 1, single_prob + + def forward(self, inputs: List[Tensor]) -> Tensor: + """Forward function. + Args: + inputs: + A list of Tensor, e.g. from various layers of a transformer. + All must be the same shape, of (*, num_channels) + Returns: + A Tensor of shape (*, num_channels). In test mode + this is just the final input. + """ + num_inputs = self.weight.shape[1] + assert len(inputs) == num_inputs + + # Shape of weights: (*, num_inputs) + num_channels = inputs[0].shape[-1] + num_frames = inputs[0].numel() // num_channels + + ndim = inputs[0].ndim + # stacked_inputs: (num_frames, num_channels, num_inputs) + stacked_inputs = torch.stack(inputs, dim=ndim).reshape( + (num_frames, num_channels, num_inputs) + ) + + scores = (stacked_inputs * self.weight).sum(dim=(1,)) + self.bias + + if random.random() < 0.002: + logging.info(f"Average scores are {scores.softmax(dim=1).mean(dim=0)}") + + if self.training: + # random masking.. + mask_start = torch.randint( + low=1, + high=int(num_inputs / self.random_prob), + size=(num_frames,), + device=scores.device, + ).unsqueeze(1) + # mask will have rows like: [ False, False, False, True, True, .. ] + arange = ( + torch.arange(num_inputs, device=scores.device) + .unsqueeze(0) + .expand(num_frames, num_inputs) + ) + mask = arange >= mask_start + + apply_single_prob = torch.logical_and( + torch.rand(size=(num_frames, 1), device=scores.device) + < self.single_prob, + mask_start < num_inputs, + ) + single_prob_mask = torch.logical_and( + apply_single_prob, arange < mask_start - 1 + ) + + mask = torch.logical_or(mask, single_prob_mask) + + scores = scores.masked_fill(mask, float("-inf")) + + if self.training and random.random() < 0.1: + scores = penalize_abs_values_gt(scores, limit=10.0, penalty=1.0e-04) + + weights = scores.softmax(dim=1) + + # (num_frames, num_channels, num_inputs) * (num_frames, num_inputs, 1) -> (num_frames, num_channels, 1), + ans = torch.matmul(stacked_inputs, weights.unsqueeze(2)) + # ans: (*, num_channels) + ans = ans.reshape(*tuple(inputs[0].shape[:-1]), num_channels) + + if __name__ == "__main__": + # for testing only... + print("Weights = ", weights.reshape(num_frames, num_inputs)) + return ans + + +def _test_random_combine(): + print("_test_random_combine()") + num_inputs = 3 + num_channels = 50 + m = AttentionCombine( + num_channels=num_channels, + num_inputs=num_inputs, + random_prob=0.5, + single_prob=0.0, + ) + + x = [torch.ones(3, 4, num_channels) for _ in range(num_inputs)] + + y = m(x) + assert y.shape == x[0].shape + assert torch.allclose(y, x[0]) # .. since actually all ones. + + +def _test_zipformer_main(): + feature_dim = 50 + batch_size = 5 + seq_len = 20 + feature_dim = 50 + # Just make sure the forward pass runs. + + c = Zipformer( + num_features=feature_dim, + encoder_dims=(64, 96), + encoder_unmasked_dims=(48, 64), + nhead=(4, 4), + ) + batch_size = 5 + seq_len = 20 + # Just make sure the forward pass runs. + f = c( + torch.randn(batch_size, seq_len, feature_dim), + torch.full((batch_size,), seq_len, dtype=torch.int64), + ) + assert ((seq_len - 7) // 2 + 1) // 2 == f[0].shape[1], (seq_len, f.shape[1]) + f[0].sum().backward() + c.eval() + f = c( + torch.randn(batch_size, seq_len, feature_dim), + torch.full((batch_size,), seq_len, dtype=torch.int64), + ) + f # to remove flake8 warnings + + +def _test_conv2d_subsampling(): + num_features = 80 + encoder_dims = 384 + dropout = 0.1 + encoder_embed = Conv2dSubsampling(num_features, encoder_dims, dropout=dropout) + for i in range(20, 40): + x = torch.rand(2, i, num_features) + y = encoder_embed(x) + assert (x.shape[1] - 7) // 2 == y.shape[1], (x.shape[1], y.shape[1]) + + +if __name__ == "__main__": + logging.getLogger().setLevel(logging.INFO) + torch.set_num_threads(1) + torch.set_num_interop_threads(1) + _test_random_combine() + _test_zipformer_main() + _test_conv2d_subsampling()