From 1c067e73646b62db1d215d86a58d41a6f555e4c8 Mon Sep 17 00:00:00 2001 From: yaozengwei Date: Mon, 13 Jun 2022 22:10:50 +0800 Subject: [PATCH 01/12] init files --- .../.emformer.py.swp | Bin 0 -> 90112 bytes .../asr_datamodule.py | 1 + .../beam_search.py | 1 + .../decode.py | 1 + .../decoder.py | 1 + .../emformer.py | 1898 +++++++++++++++++ .../encoder_interface.py | 1 + .../export.py | 1 + .../joiner.py | 1 + .../model.py | 1 + .../optim.py | 1 + .../scaling.py | 1 + .../stream.py | 1 + .../streaming_decode.py | 1 + .../test_emformer.py | 1 + .../train.py | 1 + 16 files changed, 1912 insertions(+) create mode 100644 egs/librispeech/ASR/conv_emformer_transducer_stateless2/.emformer.py.swp create mode 120000 egs/librispeech/ASR/conv_emformer_transducer_stateless2/asr_datamodule.py create mode 120000 egs/librispeech/ASR/conv_emformer_transducer_stateless2/beam_search.py create mode 120000 egs/librispeech/ASR/conv_emformer_transducer_stateless2/decode.py create mode 120000 egs/librispeech/ASR/conv_emformer_transducer_stateless2/decoder.py create mode 100644 egs/librispeech/ASR/conv_emformer_transducer_stateless2/emformer.py create mode 120000 egs/librispeech/ASR/conv_emformer_transducer_stateless2/encoder_interface.py create mode 120000 egs/librispeech/ASR/conv_emformer_transducer_stateless2/export.py create mode 120000 egs/librispeech/ASR/conv_emformer_transducer_stateless2/joiner.py create mode 120000 egs/librispeech/ASR/conv_emformer_transducer_stateless2/model.py create mode 120000 egs/librispeech/ASR/conv_emformer_transducer_stateless2/optim.py create mode 120000 egs/librispeech/ASR/conv_emformer_transducer_stateless2/scaling.py create mode 120000 egs/librispeech/ASR/conv_emformer_transducer_stateless2/stream.py create mode 120000 egs/librispeech/ASR/conv_emformer_transducer_stateless2/streaming_decode.py create mode 120000 egs/librispeech/ASR/conv_emformer_transducer_stateless2/test_emformer.py create mode 120000 egs/librispeech/ASR/conv_emformer_transducer_stateless2/train.py diff --git a/egs/librispeech/ASR/conv_emformer_transducer_stateless2/.emformer.py.swp b/egs/librispeech/ASR/conv_emformer_transducer_stateless2/.emformer.py.swp new file mode 100644 index 0000000000000000000000000000000000000000..8da5e4929e5bfee7cb7c05804c8356b0dd018c37 GIT binary patch literal 90112 zcmeI537lM2mH*2@0Y%gi7)PAZHwj3!^iIOk&`3xEL_z{t*xFdBu72HJq`Io9s?LJ@ zzTpBmI^)8Oira{z<2EWTpreeasDDQP?%UrT_Z=Pa|2=oVuU=Jm!lKTkKKZ4(-n;9$ z>pAzFb0>$-Uw2M&;^1h3&jSjDTUNhr@(BlA{J`Hl;qbzqa^teJKC>&W%FFJbsdV6k z6{9DdG_YfIpxrE2>jNhY3?Dx{I&$K1qr=AyR93DyX=P<)r7%=U=Vu2l+cmVS(cICR zFIUo`bZ)xQoJ*TS)k->DuGLERL8;QH?<}>dbMv+8^q!$~rZrTnPByEp`82J}4z1d} zX-I`|;mNq(np&u&%~GpfZl|@h)f#o_56v0UPch^L)m9f7E>2_XnHb)8>2Fys+=@ZJyW6_lq;{?`58!Z@&M+ z6xhbUk9oeqd{1WHA7Y-LYre0~yuYt`ZtL~11M~x%{yok6i_G`4GVc#H&n+FlnR$PN zc|U2sug|={pLt$3-{+bPZT|N+&r3Fb<~^bt^|ACh$`H)P|4Z|J!hAnD^ZtS6`6lyy zYv%o7=D98ZYSSTX`iGnMwtTP4yf@V?q$b>fro-6y=JEco{S??wf&CQNPl5du*iV7| z6xdIJ{S??wf&CQNPl3CL0_CYfLC*evB#anLq~-tL3&Z&y&;pl&A@Bp3)DME|!3)6i zz?t9x@GcnOCO8jl1S`Q;VX7|!d%zr616G6MK>~gXQ~pKpA@FwaYH$U3BA5d^z$M^x zunL?8mVv(m3Aiu#IE?@0U^Vy+LWD1YFM>~l4}ojIHgGt24?>7{gKNQyz!va8a5KE< zJeUPDU=?^U_%%H2ufWH^8^Ihn4HUrv;1lr3*ML`p-QYs-aBv&E`G0{|fv19}fGMyB z91IQuzkwnCHTV?}S;0%ec5pcOI6{OMf){|NgH7N(a3)v@q^*9qPmFV7K5NUB*#s4d zxwPI+s`W%^Oq6QXRy(OQn$5H|->6Sj>oZBak+f%1g0$0Sy!dZm(3XT%ah5Gxwpj|9D56mII4P+b33*}(rCRBHxrypYP}o9WGH`0LWnsRS zPOOK#`oDIgNqM)V^;V;~ZCih$-oy_zNdpq=K>Lu=w!&08Ee+==USz0CBQ?|ZLbFcY zTej5_`7u%5mW(G8Qt?u`-L97^^p;c}_bX~>)g1S5n@S71B~|H4G}263R4sz~&Shzz zt_kT%q8l3ZbX)QW-D2v!&svH!Nw$)EMRmrB$*N9WUP%s<)&68sa!Kn8XbnN@qAI3R z1=1xl`AH0&2&vV|d(!6MLVdDYpDIpO=f;Qo3KBq)tdXR8A19liVm32FDqAZGRZ?lS zT=KnWl1*(X1d?JaA`O=3=ToZFa+K0_k^u7QMXsP#19C|$of-vch2$FXNlUUVA(g79 zZrdJY%+D#OYf0TFwy9E=d^hq0?OR$m%ck6mq&{&%)zW$~ENnrj;d?P5f*vt3ylqhW zcG2{sAYy+Qj#z30oAJe*qLNLL?b{_k{f<6FdVSlsGxnxB-i)X>lZ`g>^`+&7n#?m4 zd@^mygsMB+V3I6P>WxdwE_I*s$Tku-GqKd5QCnzN8+GPC8Pk$Rb%3ZOOooxmL=Uk%H?K`eoH+g;mSGbPbxjxyN=P z6(*T(Hs)La)#>QkTlK20p2v%2uuPggMP5BydaYb9tK_~cbHR$)8rzp8SFMxr-b7Z) z^&8JDtvz@1y7e13FrMei?b*RvW2QKg4vaAT3k5|CqrN)G;Hoh_&doQPZNlzIOY`Na z(pZfVF(wIV1R_*JiQN*?H*XkGRWWOiR z3nrf4oYdJYw32!~dTJ=p-ZLNMBdyqT4#SQQvuUpFBx zPe%4N6O**<%5kcCdrOe&bd??CWLfr+je21@)v8tJs%Rin^H#D@hs+dHYNCt0)&W78q-LlE!44jsibWY0w3j(S)d$$mKSL z++#x48*SP86M)K1SA|p9jdHr;h3JD6TgFhb3!;UI2(3(dA)!{OGTJx&$@%a*G*m+Q zlA@GhnTfQl?_^bA=>UU+L++ockm?APTDd8mKy@IN$0g(rM6#OfA5tsodfg?cy{1o2*gWz@Gd0;yj1&4v#;Qzl5t^@xJUJMq%(cmHA z+W)Z>kWMWh!^Z9|s9*fITlU>)ZF;b=&@N4y^{}|xGci$sI*Vxd za(S|$v5PdA8Dvv5Tb@so;)edDSllu&(l<0Tnhdb0$RGU)^Z#6avc=tl$|7YR;YD$z zpS6y~jg9lnLSvy->>J!!O?MS1`x92l3jFYNpHGIT74>48dhFJWhJV|wcv6&XAypZR z#VW6YqVv=z=SU&UK}ZI9u^T3^b?HKy!s@V&Cd(Bew)7{GqLOR~VW^#ZR2@tQDFshg z_a|rdCl~mfmM0s9Z)Y>vuc<4cY(&>~o=1|_)Vuuz&OymadNoK+dvx*&lIFU0%d{#j z9G1cEq$SX9y_=LjBe7PJh+t@8vQ=i0NFh)`f{1_xCYUu0ew>a+1;<$AypTuHDtzGN zWJ%&FBuQ0f%k_F%D?#A#0Vq3Ht(WLStdu*`@!`Q0D|uAj9X?vQvLF%V)6wZ^!_J*y zv7iF;-*RDAS^8zwoJ*%z^n}8U8j*ytNFyMt)I|Em-=4v4obINqP+N8p zQqpPJO~8uKkgVuIO0Be=&6>(o%ChE50adAq#b&4?rI6jqc3y1n*G#zl>5}t>mZB<_ z`#4uc#%_@jhLB7*TuaRk$z(oJVTw%&;g}%;891pIB>`n zd4foJ{#?k=t!7e(4l<1VX^{c0Rd$`@BXOfh(f7JLbgNpH4w=02JAfK8vQ{x>>h;0R zup1(eD%Xl$klb2;TVjnZ!QWQfrBcxkqjIa2!kGu;Pfi_APAEuON%NJVKCVON5Pd%sLitQ1p5j z%sQS3N&ajkrC1Gj0qZixDrs3c5vD+-6Ik_wS&jzz=yj+FIp^L^zMDr|lv0Wkez^L^SphZ(CMy!j8AHgOh}?-BrXmNRTDyc(P?^XoZTY zSZ2Rz-g1oDBpb1!3bUHwE2GdIx<&hJyCsxOhiphxPj@7exJHJ}apdTlB&Zq8b|V=X zTEWs^pJ~t1tzf2+Nl2X0DZ6Wf1Vv_!a~e|%H94K~^M+w0bdGAtW{j0|V(ew~b8~9x zW9i4xOaCz_k{To_vgsd|eZSf^Qz_XKyFD|%SNHie!m-Kbll`M)OP+}OOAwo5PXtV* z>1KH@<*Am`Dq7N5CSzB?`~p)zHFp}(-(a;q%_OP#|4%`KbA`8r|F`<>ABDdczF*G$ zzYaftBlrk-6?hRy!B%i8_!YeTN5GX}4ft<(_s@X;0&fM^fCZ2OIsbnU_!K<+KZ6=L z8vFu&{rlimK=}ANSPM=72Ls{d{||Ttcm_BPJOF$NUj2*U17HDc18czX;C|qf@ai{! zX8}3eKNTzozkz4}G`JCb1pEVd2@rmM5*!DXgWtike+0Y;l)wsbAMg?Q_1A#MgR{Ug z@M-w2VZrM=BHw0*?5AummQXILPoeKU|f_ z&1Orm=y%E0ZenbcPuD%|5q{H{4{w(PO9o-a5Zx}otgr*AQ(t(+h5C+TS}<{QOjJtJ z7p|o6AO5sXsx0S=&}c#QIg-i(dLAr}If=+zE%W@Y#zJk%wF3niu-%IZJBoj@3w2yy zvUPq+lpxdzBgz%-z);Ux2LHM2G3@PGwKnnF#* z%*$yaXAZgv)k>j{!9uw~4_y%$ z_~+*`o1{$f$O{Yhw|u*7RK;Y$y3-elh?qU5-#a*6cOT$N!p0n9XW%Lmq#O{b;ch4T z92uddQ`mL)W?0*yMX&BQD({v-ZDpf=oSI=BwT$oPp%P8O5}`|2+EaH>PRxoEQJ{XW zi?PiQ4gZ&X3O$ zP_#byRb16rt@?LoXA9`uZSjaYh3{pB-1hMA_qt~}qV~sqDW^;LEpR5R@10Sp{|LQ= zjmaIIx9*C`C2a4m^9t*uQCoG5>>t4vK+bx)OBjX;^KJF{*yE~&DB>41M;)$cL24s5 zHMZC)HNN*fwPBl@UKdq@`p;~d`l%Ajhb1%hi92sl+)=c6g?o_7Gb_sZ|5fmw2f?EX z|8MQbKMUV~HMkh81jm3c!R!A$Xo4w_fZxLJ{{Y+y-UH;!zY1o-3E=1O`riXLfsXECfn&iB;rVX@a<>0G@EEWS90f$b|4J|o`oSUKKHz%z{j0!c z@Nn=2_BB;GOVF+KHfBt<{HIPyzgz7(3$`SFrJq9-PQeu~?)OAX+xinqB!ZWSwzA&e z`6XkyQn^C)NuLf#uoT5W^wJ8pfyKRF%IqljWs^pB_Hh}x&- zQexuq{DC=caW4n8WEg>Q@pMj3Rbutov^>?Hki7aSzq%)_w;q#obH4gQjY))}2uDO$ zV@f8rh3KSN%M#F>vJ^7}3d|mSMIy44@O_GQ6{)+pins2MIPHSOk&h3g$^YcJJ5M7ygctcuSUw3-o~+h5_D{o7Dp|5n&J*qI^W(BRGH@`4AP^#J4bXr}`$^Ia6syC*zG6a^EgS_guNah0^+uPi zB&Ahngvrcyp)UPLB5D1RNCl@L&*c+SET&WNOLZ%#fd>^8jsKQYW1a*HzI$~Z&s8n$ zvhChTC z;+1L{pHp1gkzfMQ{QK&H&(CxX?Y`oD!n1nRGIt^t2mL@Xa}a%Sy4HYCk4I|qh({y?!-HOr=^O&ao8E%ZTt1t4 z6>&x#<7=y0r%Hr#OS}%_D*}hRRTQ&pdqgOs+Iaxg`fvso?Pz5lsm1Dh#S^2v7wCS) zzqWaPIolnVEZK=7(m9qK@h+4j>E`OT`-)nuTOik5-2zCj(K|79d~P;hC26Xn?`viU zmRXAns`v1p?*;NC6N+_xT=mVVW@DaR6m2KUar6@X*#0U=T=`_;)9E-`7AsfTb(L#_ zCY_*o2v8BpmddSQ2V<7nZbdV%OKR-s*=WEEQ5P$MXxF1M7!*MY9_J-bva2Z@vLF`; z;d6>gM5VjD(Mxqf;@VGVzKY&;^xQ?oQA9X?vuz`V{Welb4St|ti)304yvatRmQiu* zQo8QPqhh3-e z{||-1RQx9V|8~P{-UPod`u}Ud?eO_mgI9nTg7d-s!B64uzX$#k+z8$ZUI5zQ{@@Gn z^&bb9gNwi;z#ri4KMmdhn&6S34-|pu_WuO_{xe`F*Z@ughk`G_-@g$&8HhdqgTP;b zpTOt;82kv_1l|JP3{C~(;5hJL@H=?^&x3b?SAmPbX+U)O4+ftQz90Mtcq_ODh+JSA zYy>O8k>C*US!4m%0nzb49f+R)UxTkA3wS?xA9x|S3~UE$!4UXsa1S7M|8E6X0nzjS z1w8+2z(wFbK-%Kni}L@rOD<2+>1ko}cBZvGn1U8%uRYhmrcY@xX(RkH3O8ac1dKdfJ;&KV9!+Ep*M3}1R?GLSdNti+fH37-#O1= zxRg+(guToRD#DX#dsm9PfR)2)uQxJ0tdhY%ivvY{(5*D9TYh6CJ6R6Z#b7-d5p|;# zBqLgNjk&a4Rhl+Z68Q;}P;NEq{VGjG*JN17p3}D8(i+Q?Gjd2Ty2k7^n{zF5_%4U% zR#}+8cQbHPUXnjY{MC?hqNij@%VIK9TL@I;mM5#Gk>e!u4RqXO_aeK%og>MvG!fOn z-4w6cSeQXwOBxil0qg*284SFtUR64t%OO=j7uSvBd^QqifRU-MOHn+Wrr{nPbA6(HJ$7`-k6=RJ-isvV+_5dQRK zFtUzfKme@)el)7!NIc+VgI6wt<@#w;D_En^DM}a86GZoIYRVdWOGi9r^Qv>!9%Drh zE`=WIu54SO6%DmeQR;R=ICB_uj#fWtOp1tTFxub7DwWb7#u8alq1Be=LZCSi)F-1G zU>hLwIFHnsAR%!xucQ%mzv-bqnee$_Hf3ZZacC$jKdFJ7bIJIylVXu1FsE;iOO4#9 zQXz|K`{4KKXN&%Ov&L+6nX)b8xRS$*)(cj$#a2{mt#P_nCT&#-^wchxdTBd?7UQ4tB(gf}fbixu8 zb26uw6!=Y1&$RE5kui!q67<m5& zh`lYF1Vz%A;Nrez8{R=`tYd0a&tTsZq;Jxdv;sF6n&#pfNFS1&YRRn#C~3;15YSd6D{p- z#(T3_QpXIEGTn}ZvVAi*(4EjdhfVhk(hs_4FtuohHokA`mbR-rD&yKXUeAQ=4UK%f zAU!5=4hr;Y#r_&+OQ?zxDc^&a+py22Wf7JXR4U0TIb_!ZjM18LE0*g5L@zs=di!{} z)BWWnk++-0oxS9H%!yN3%+t#f(%2`jIj(7~-wLHD11{ilKz;QR3P{|+tz+d&E3AAAem{@vi!;CY}8_%`?) z_yBl5SODjNbHO9QZ;=7K3A_^Q0T+Ni@N0Pg4}lkfM}tk^{@}Cl`C{WQzW?3;#3tZE zFa~}N!F~wd46X!cf^qOOc>C+X<=`pc$zTlJ29JLu_y~AC*a=EtGgu46M&Mv@qwxOV zRp6CCd;nI!3Qz>M(MHlv@3pc4JOhRiTh3?gX3_l2+XR{+%ITLJ%Gd)0CX?v* zpg9}!0x&;?NVX!47Za5h_cOxWMR5(kWOtGh90x{OJ8nd0cqiF}g{*QrRFGiu>6`=P z>FKhXtBrV=?CJ3?3WFwZY!;MGXO&k*xBr(s$U8izf}%4r`vr1?4AFe8z3*|s7aOj3jztBi+^ra1< zlG5SRd!}RK!kIRu&HUD(BQLjeQ3p##cwkQdY)3PDWXhpxCbjb^%CBshbiWo~Fsfjb zpdi8IV{#x{^H;XA{mF5Cojo~<=C%GN@DZ5qPT3r=n+!-3l0}^`1tf@zZd0%aKlp?c zmxxI-%?!^X{>RLhhjqGmx}$j85g+f17rDP}fD$Kuwe*=k2=7gSXK zi_2@K430cq^L9MqP&(5BE-{slE2XKddmCp(@9Rt&3;Lk7#b`b!luOXnwLfV^2C>{HQua*BR0t$%eZMUP2Y_bXSADASHYbWF`bMaP+$?TJ8?ifLh&g%Mh9 zN{-eIlNVSCud*?^QZe~EL|rcQ5{_$2y4KW>5$VTa&Y7)CHS5GfYHpSwJ<0nKM#s&o zQ#9jm%AN1+Y@)qD#rs#-L(yUIi%|Y@6}uYQWTp)ttLs9N@QEWwHPT1iYoba-#Pci6 zUN1=zwK9v5Yk)X2!>DdiYr7ZhN*4*lbuGMVwc}MOSh(WRB6j=O^p`|$K-`{;uJp+Y zZA4CY7YeG~{>fEXGVOe-20<*OM$~J3d8|e#)G4Gc%R!B{{uqxvS#(A_XLg*Pz7VzoPNfF7_EUbHdTo5;Le%hT+ zO&#akv+fL}qY0!`@=;-fktC{H&bNsO>kBFv|1DR}8?~e8g%-y*IEXRzvc`U$f_`nD zjT=CHFoptGH?n5i6pucf%#Du>53)ku!4@9r(!#!Khy}2|SI2a6il&acX88-J47hI+ zsX2l+_SsBw+V75<{5U~S8yP=+El+k<%Za$#Gfsq*h(O#E`N~w=TB@gxwg}5Hoj$Y@ zmqtRVMnH3smE(k_EQ(p$QTMbQXSww4iJLvWqGN-%@s`3Tt&6aT|Bno59AX2U2 z!Q-5$f>TF)9L?gyd#1sWBijmHS99hV*3zFLDoGQ4hn|1pu{fZ3b)X_-&5vj!U$A~yBlk+=BP-c{=!XLL%mwtsxIjEpMyClcK z!IPxxNL?j$d)zx;+ZNQD- z@!%YAIQR*?|G$IhfQ!I|;CS$0@PFa;KLJD@@HCKuDKHLhgU1&i0XKtx0#o1!@JV?5 ztHI;IYH(k0EByVZfcOS@Iw*tFf%yGD2K)>j|EJ)3a1{_ge;0#6a1;=~0EdFx;rV|8 zJ_-H>JP&LJ_W%VT@_?JcE5HH}TY0` zv*!d$`RtxaN#&X4p_c1`p!M(=@jLz8$;OFIqG!{=8BOG5=$?{pnX9Z zt54}ky_dB);Ot&bTOyXOu zQA%ec^^mwaidD;g`=zaeoV;l zSD+cHh~91Q9p%<7kSk!f0J0VGn{#7zFLXFfm+6vfdk|;AEE9el`e{M}M&)oUvCJ1D zUy3(#t~foy8PSy&HM&TQ=-m@yQI4Q+%vc^MvEf!RDVLt#;%EJ7=IzI`VKn)B>ttJv^sr{SoCRqa`w8dmV2-!1Ba$&47wil9 zj$T4fRdgJq%*#v~Ej6xRd8~Kjj*b+_QwApX`eczD?UCQDo z^KP<%g2WC#5xd*jh2x@iyf0*v%B@K$mp=}p;-_0`>sLsbj-s?&PS$5m^w!WlIW-v$ ztR2EpX8LDD(x~oXkBUY`PRU%te#nK<^}-SjE~)Hl`&fuZfAutB<2w z5V-5KCzDQxl3)=BVq&s_e;OE$qON~tB6G7jNE()mAVhqf2Hu3(=Pjm>U5CD$I^0>}{YTPd_slpko{&CJFo%B)}pdpv(L zn4?hc(260f=P#o-ET$QLl8fSWq>*9fz68Dq#75wqKy(Jg zPT(nE2iO2k2KNA8gZG#F0bdB#0?`q;8J_->O1NZ}SfzJW49e5f@!4wcbf(L^CMlSGHa5-2Hth_*c2_6A{fIL9_ z2wnj$1IL0dL0GX9cn^>}0-p&cK^cgg;27`~#(3OE@Qz^`aGX}eEZxq$6hYyns)#m}0uSdMKSZ6Mzhi|T{|n7@ z3N{fRZK-BKxu!LHk<6gBBJ`z?`T9IH1Ed|^ItQC+Jz+={^0{Z>>$r4F>DOyfrMY+z z6-5|9W%{^X`83C8xN~}1PPWT(OZw(kLrbG`Feb8$tG{2#G!T-HY!)U1QBK|-vQkl# zMx1R|uG|%lYYfg)W%G5H;`y2`;CWs=jpBeym|%X0j4-F(9mPtL&I&4x0=HT1o|qwG z!ICQ54`bm&Mk)|iuEQ=quIDP-yAdhNGHCrKYPMdarqcPj`HHH;ex-F~G@;kX+``#p zqGkxnmk}_Wef}6ScX3QB!k);c*BA7>FpRRY_3j4I1W_7(O8@0#zrMJEBZ6J;HozyU zTG)tekXI|3b#y(^R`wak`Y$QV->npAXh^ z?Fn698RmHp6)H66&V~%~4gPdR@a6S?WlAnG;=HJFNlh7?34692zNe}p85PQbjP8S&X7(w_ zDQ7{|p8r(YGKM~`?3w4%6guW220)4?GNH+ZAEGC2tHQ9GmXU{Z2oCGAE5cw>T%*l` z##A+JliJu~os4~@xG>0FZr4qtNvsw}xkyCGj5)}pyGc>Bg}l4?GR$kc*-31-83t5M zlsQsvc5AeedNA!{|LJQ0nQeM@%=IYy24|z>$s*%GTLXw;(AqgDi3CvV+uP;Ebn)2e zX{}6o>e@2o5U{3fBW!rrwq4ALnk+AIPC+|q2Zq+x#|B$v+cC%1N4lA%g6?wSAWaLI z3i^TmKVqVs?BY%oXGa+@j;u&~Yb4SRgH0 z;6Xs{`V(IOR|2^M;0W+V_;@+%7e4@(0`UQGPw*{x_IH8H!CLTO@D+IUFN044@dMBV zj{*mRo8iwzw_on>D}q13kG}&v2ULKZ=|3C@pMDd3`TN0R!4y~r#J2yz;AimT-vz?2 zzZ(e8E`I-?0a9=pSO$Izt^N;?bN#Jg5DWmJq1g7n63ChUCLlcg7&r#pABe60dqE4F z2ObK33-A6;@Dj_z`|jRt&!6Z|x7W;}XRzvJL%9d`5Bctuh7H#%jFxvHY!E*H4I86J zvNp)l6BR=}WyN2Kyrb`HLW{-YeN;9ZK7MYWMnOHd7AN7KXMz`OGP=n`bZAMbYi-}} zZ9IG)SZdiM8NhDECmiSq;)rnuZ zO^tutyX5YSMy;bIBVIR7pKPt@j|Ka8l5Er(bVylLK}4XVE_@Iv) z+D$Zgt`@5r0f!dlXfaVB2_#J|2=u8r%~&mv?It#odbllkoMN8%?%ZZNU>rz>mu)EN z3tUI53j2s8OW*6H}*FoV3KKOdu8C(6hkE2SzLd8MlG zT-q(Anx1R&*uFzzX96$@`e|Ydf>HE^mBqcdMeJ~t7aTg0z?iIFZOiQLM1;&;)$Xbl zeUFOLqSd;$WX^UNfAp=?FOd{W#Gt`dElN)}H!PJJ%QRE>z@#vy{aY}?*i&HWl~jGc zN^{ool4^JlCYu-L#Y<{IdWL);hZcP1%1Hgd74eO}kDP8~?m`!DgMMr0xK=#0HZbOq z7;Ye|9#Ed-`uAk}WOZh{ZZsvUr*-^DP~nVJS@|N>?m^*Dd>A4m4>!Jt@MT~o?D(WH zQn)Cnn(`m>hxJV>w2_2l4rMW2pi^XzoM*cY97eiCn~pm>K3U4Ri;^Q2T}AG-tL*U0 z0<%mYi5G=%aX2M6zp8-vq%o6#>0~};NEGpYrY3h!59<%Jhn_(<3h1c1=3f7@vvO3J z+AXzSVV!yq(RUq4Ay+uyANVr*O<686@@qiH%pV-YxOK%-Ly=keAND17!59Qw3%gqs zjk2v*pYr87Cp@)?Tg7LzH4GG|=7%6R2KX})Y4Ps#@R8PxX=!IHaPsf>W*2KcJlc*P zUDe{c#qwm!#x)yp)+@em_MrM-G3LR4oC6I_@@^a4XgSIT68`@Nc+kIvcNPBsT*Lo= z4PIaL|DOrAfe|naz5$Q_M({+i9UKNe3x6+u|EEC;P6J1Po8j?40X_g;2F?b{!L9K4 z9{^W_2`~))A3XjI;Kkrfa5T6b-u{E&1|aA5&jKfb@51B%3wR6=-+qUJAH&;!0=ynP z56C%v37i8C1h>N9Uk9EJg!h*-`Vnv>_yaustw8knUk}8u-=l!=`$OOW@IC1B3GgBC zL2v_jDX4;3@EEWT35D!`pjw}Q@i@oM!bph+T^Fu4zHo_sG}BaSW9CpV=`hg-Cp6eOu- z9in}`R}=fOY8$tdEqBO}mGQ5wSLlDu_2P_ZNn(DrNpU{R{pLW*!A7Rbn?-ohg2a|j zhqAigFw5&mra(!?vk&6(L~kiv9GnMn@IMQ#T}tgXoq4=_K?o(Em~>hTSBqf85ua1k zbxe;@*e$5Y{O5Knnsbd+bq&iLURoFJlh=#hyh}=)F$Tqs_sUj|u(;=Xx9JQLj(43t zEmy6;$5;>l=j++eX1j~Fa&m`#$SnvIt@<4ywi3eS-0_xi>PilQ`KTicw(#mGRjwZO zSB}PPyrQS!iS4AIlL#;7)Eoo?8mWqF(Zfk&|+gqinrIYNdUTCo4Lct$eMuyFoM+$|lz5HuP$3c;<$snFX>`ws)&m zQnhandsIN&sISv)dWuu_p=^2WzYisoWV9@%U(1aHH45pn=V;NFWmUUKz8GDhNG%*c zhO6o(?~y#T+^l2l8yQ!aSk_c8PwsTb1Y4`a+l70|wXD2PW+cU_=9Je>3SC=2Bk$5H zu$@?taf#&0*rfWWF{i3zW3*$S8k!d^7NHe{@bcNh#yXGDMyeC&+=P;e&Sl6qk#wiU zMKR zUrLi#omjpboe$8?I76W8ggv>hi_F6;k1=(8?be4<4HGsbl^0b+LK4^QwE=&oEIv{s zG>NL@!*r982>-tV_VO3PbHe{u4afOqc>U|avq1xh?SB#cH~ju*fY|*%4O|2U!9Bod z;Q3z%o(9eVqVN9|`2BZ*8BhY^6L38^6%>Kk{QnC6UvvT92VMe1A7DK=8XN?^i7eoH z@NRH5cm=o=Yyu~NzX88SF7PSv5pXrw0z^Mx9guqhzJZM3t>7hKCl~}rf#u*g$Orxd zh!4O$;6m^KAaa9yf&Yt);CVeeQ5xV!a>za;Pxt(uAeU)d=_fV|cd*RG9cd1k=wL1j4tj=_TqMz{; z6URY~Z+u-e>s>kPcgRSL(DT)QIr6K7!npRfCzVnhEZh=gb_j96a51!&t{GNIOO`Xy zM4(Wz`jJEke`o6MLt%1hlrM|sVV@5h^RuC=AMw$d@$R!?rD)3UE?MU+IJ;VD9*Gm3 z|!KF(1r6*-UcCg_oZe%Bo_wIxxphyIc?UH(Y>$sRyS#9OW!v z`5;~13=BDp;%X0`HEiT^-Pchyq1bn&?9BE5TLtEK5QoiYjt zh$VEdW^L}!sOXU;`HCO+o!TPRdp9}j=Ty)RHU9}oVH1DR=V z+`yXFL4Po1*mntZHx280oxmz!W-Z(T~tA!nMKa8Cd3 zp6=OdiB~p@B50MhICidN+2oG#pwbNFXzJ|88jUT%=p`CEz07vj+?jSXzJ=6nI{FF) z<3rq4HCPC;wLGgjHHCJEDtvID3+aO02lh6To;_<3no1ZOXtt!8cO^PEJ1gS@T z`dO*LWayH{I#P$nA`*AGfqYgcrnV-YxzV>3TgIzM816lCiJ2%WUlq|T#yVEGvoXBu%OfFIcxox4KN}vkS zIUz5x3``|H9sif4taGsYQ!boeh=mN`yC~$pCsihDzY(cyG?`TQO6|~xWY8aftPIYZpDVpwms`C}(w|qJ; zEsAMs!{Ef$yD3&N`Hvp0Pt0)$J5?<9^$k|r>0GfdIc1!|FgR@O3A*%ffvX0qsy+g^G9D&G9dM5_ z)@{pWp_s0SvWH#-aUDekcByL-CA3Q@EG<0)!AcR@F#Pokyby{3tP3t(N~QtnM9oV0 zm?xTgN&p&do5B+aPp^(Mwnb0S-3hUt4C6C%tNQL~l<;v;RDt}dRd~p059zwZbdzVC zU!F?2^(Mu1U{9Tw>}j;~)R{}#ly-2O?;S9R2rI2sIz+O|#jc4vOrqg8i+dAVGe-$p zD7wl{L)vtM*&?)x$W9QROX^TD>2(UrF%v!bE;5b?+u5)dy~?sFa`)Bils8N`8US5S zT!o-pAt*CWg(>f@vRt@}REzIQwPLYgKO?q^WIULA+pCwtqWwsJ(Q)13I z3Mv);^E13WsdK}JoR679Fpis|9Zfbe^fHfRg;D#koYR#2tilG0HblM+C1S*QQ=V8) zPzqSE>30ODB5`o3SrlA1sl+SlVLW@s+;#JDyuh`^o8h&0;2-54dg1@ehW{7)|F?lFz@x$a zz<1#3zYT5%ZvxK(GhiJ!3ygsyz}MjI<&MB7fKA|Fa1;FeOTmR;7#su+1iyfnzYaVN ztOLuyZSe0u1Rn#!+lx)WC7=%!!99W41&BSs6M&rcZwAMMZ^6&M2fP~yuRjUOUuov0Dc0mFZTg{2>dg60oV$*fK}i=;OFrB zA`93F=D}G&?gcyp90k4)zyFV57uW(e1Ca-;01p8N1GyLA-@p`*_Gp6(!RbKm2>cr$ za)Ddn`#%gm13m`c2*jVjBf;MSkqO)mmhjQtLF_Eep1Y)2gqD^d^?vKO{N3XF5Rp_% zCA;)ErWE>1*@R}xy3}$?xZHg66UaWz{%$H_>8D*1j;;cah=9nBEwWsJf@n{P3LUGe zngBh9yFlP4iVZoL+3nw*r|Lg3Hr;5nE3lzs+gr2M(c#e%TfgnwbatEDT$}>eHj`|g zPw~!+C&o%+Mh;UO^+yzhtdFoBe#Vlu)6>;Tl_N>3O1XuNXN$dnI3(ii8IF5ZrLlk= zK}D>xuvC%l{rbkPWWD(EtL#w;vNKyZH*f5Z)p57Yfy-7Zvy`^dmV4e@l{n4U)CN@UGf3we;?0W0whRk~N;L^_Pzf3m;^tN(RkMa#9X; zE#Fd7^YvsYo6syber}d|S-7tT?##qKSY0rrtUuJDfUcO=I=59-r84eyP~Qm1!CpaM za$M+QbEm6o_msxTw@S|vDLBNk`59FlF+A9&w*bE6<{f>5SL&Q=CasHp|IOntOW-GZ*F z<)xX;^`D@$RFlxpsNHS0Jqh<-I_G6=VRI@)UAr<$yYO*BIY&!v58|~0<6seY5lh3Y z?5KcXXR0V}+Vq7}8#L4RVuI~5P@yWl{zi|ga0yF4$VM#!}G6cZ)NBE9{~4$|)GFk+=(aR4L3dgi1>$H9$|ffI{~utxashm!@c$*lG~Nuq|90?vFb7rvIrIN6{Jq@!cQt5% zGr*DH2k`hG1y_PPSPsP3|Fz&6@M>@ccq*6$Vgs-O+#h@op8sRuHQ-6$Vj$=IM}VKf z>x-}d8^N{Ujo``PNkH`f*MNrtx&QBT;MJfFHiCx&kq5jURKY3W{@`oy^fv?H{pJ2Y z;r}lKYk~L){4xCeJHhk7G&l~7f*-@jzY&N}z&2O`9szy{?|uuo96SX~fd_#@!6)I@ zUk{!O9uI`Ke-OAg_yaWmC=ee2*MP?XX@?Eq98d&u@88$q<=>af&)cq`B&5cCsyek> zZ)_9B8eD|=^a#=}*g#5d$yP_{r$hx3r4o|_^IozK+HG%TcXw)VgNA3s{O&>0@XzI!eDv5imz!^~$Ek5%@FzHT zXPu32boU!_os_%0J$^A!$h@Y|e&6F>vf~u3Zt=0Ca+eLLa&D&BOc(5rbw7L_MO>l& zWNwPLltu-k+JutX`{#R-Hts6Lo}HmD8PZ1qp$?{_hw!25MT@qEBc2@;^}l?~wk2V*lJA@l@7s$VH8%^i}#3=8kZk z%i*nE@nviec53Zm#+lV@jicxtnYJ3aGb}%%YIxI+yvK1tqT>`jdZBpE(>R(@fhDN< zD}J3LVAP)MVyUH{)#482+c_~G5fW*|4Q6DS(bcWLs?Wwr#tBL6&L{FTR0Z#$7cN2N z6X=9@*ALf0tvg87xZ^=X$8((xoYzG0;Zd%shmD@oQt=U&nUM=NLipX$Nyj^$9 zp&5^UM%&pfka|DS9a>haw(~2s#tanY_EFd-oU-_(J-dGotyuSMSI-BFIb9tOvak z^DVS*Mg+d^*a=wgAn>BuUpuow1H`_60jvcN13!hQeNSAvtk!@#Z3{=?u($KTtv)_$8!-vqY9E4C?VsXXcV zSz46*P|^OflFcTFlX(@q35pq~)~L1Y9H?XK*(H-w=I1^otkT%`n63!>ltesdIyt@^ z>$cP>A=6o9;scP~11 zi#pklkOck73BC(1scnsT!YfwY)l|?cVAeB@t1(2?ZpAezPI51-Lc8CB26iEi=C1iX zo#NJ(Ltts%z)D3?ULCC2fZ4-}n4 zEy){oFO_juA(^1HPxEe5@bv*Y`4iFC3F*_j9fMlFwEo@S2lM z>r?%*ewL>(MbOU*DoF633+ZC90d8VkMm+VdOAH@sGDQey9K3Y|mfOcr=!i=aEyf*D zM5Tnx7YkJCC_(oao_NvjVks;9ern1IN=GA|<6aB3pgc_}$926lei(Q_N5G;1nym9{&%aC!Pc zrcLpJERRiaxP@)5Ixh%J-#G~M8SD6pY#n9Ji3^cy45FaqdI%DiKXfw3@~11GdurWG zN=rQ=3{i6`Z8@@uARV$$Rf(2$BdB)4JcU$k?AgvDn?B!6>MwU>lki~zUtO+M^ld)$ zi-oy5ib3d&vatGH0Qbg*7n`c3R}K)^V5qWrw`8KnIHvgj$H0ev8{Sm-f3!E%=V$Qx zqW>?x0YnCHG8hBDfzSUEcmsF>cr@4m?gtJ9pM>9k8+a1f0oH&+!EfRDKL^APK<)q( zU4Wfn02~g)2jDNkH^ECm9jpK%5BM8U0AGXm|0?)2cn6pVCx9cs$B+StEx?7~0w8jM zhlB6I|GyeM7Bs*(5Iul<1GzKcYETB-!O`Hq;Qem_?*elEKMBO<|5$Jc_!>O_4d89y zt>AL-T(AJ9!G++FAOXKX4)A91PvF_01&#+p;CAT!dmwZdz5maGw}96Hxijz-Aom0u z2}EY_ZSb$)*m_OX3}!n+;l+m+1L2y zuD{M??2|Bk`;5%O$88kKTilNlT$IiTIwNbB7Zy9ZVvu`Q^(U-E{4D;}nO{x;@ms%}>h^r^C=+b>m70`Nh3eFY!UtsD!-6erc+Tuk0Xm-shLJ8V z6Fj#Cj3ry{NC9Myo^LiPX$$j#k~U$DH?6){`Y4ecJIg2a$sC*ec(PLHU$Scs6mCz` z30;C|e7bwS(QNxxGTULTuI=m=O$mH3$=9!+@=HG25Omzwz!%iB;J;95vxlXj#BdCj z5_%-uk*Dje!&KpZ(F4(Dhs5NvU|k`cj7IJnObYD7~7m9`W~oLBniiMcmjoW zRjcrhw8S%2ouY8lQ8BCyKhJfdW=Ri3ma^r_k292`=fM_b#1|%}CJCu}<*`w&Ccbgw z0ufnv9t>0F=8L9dkX_epiE=wR9&DCe9;C%Z)LT7^71b{)X7r}GyU*86WK}4tFxz6h z*sC0_Gce(}zolbj^tR}G$~p(0$}__f%IaU+wux1$Pw&U-L|g?xFC{ZI{`2kTM^ccF zO~rSbB^$6?KM4ppMwjZ9KO$V$s`)0-Gm#ec4{$@`q%xK<*Cxu4fE&@5BZzdBI;zXj z3aV=6zYBt!6`_g}rc@;@n@Tf9eYqu;E7)H#5l+HAQzs-z)u|vQ+qWXM>x|-Z=$k=q zo{p+(6*%-QiE(3YwK8u*eI0W^hAkS_Dp{456Mq{o13~$AFI6o$lJ_`L`|_YstX*|E zyF4!|+M8ALPO|e4Ec?|8(y_pxrX<>i*#*f=|6Vr$KLCncvidNsvMfv${lGOVO*}U_ z@+v!IJ=2Iml%I|gKH_4@%tGaM(WpwJYgQDNqfd2SE%&P{ZyT*ryCDKeLIf_MyKqOU*JTm)GUOEy9L1g3OWdO|dUe!-MZZY*bN=TCg!-|aKlLozQ z%Zwqv&`XcoruhFo@R47JpA`OolV<-5--Ory5AY%I4)Bj)3wQ{)4L<+PU=o}Hz7BtX z9T0y181N-UdfBM^UnABTs(2BhF3@G$UDa6A0_zkz3h3&2WnB=~Rm_s@XW zfX9OE;B@d1a4Y=#4dAulSwL+3w}aEb&*9@g3SJJL48*qo1aKI5Aov1&{A+==;XjPO zZ-HNz^Zsjq`0p3p{vX1p%RPT@1Fr;61sg#Dd>p=9?)1ABJO%6kVzd8nV0HV2rhgAE z2hRm!pPztRp!a`z-rZ3A4t_~<%4fc;zcaT@yDcL#bisy*r%7WbygPS>%dzfbr(hG3?~V(~qPI(&!OU)XBv^DhXCX$krn9!~ zcKP%TPS-FEABmeU{v{>#ndm90msnH6<19|+XY3qQU2-Bh=O(tR^a%?qp63dZjQ!-X z*MmE&>8@f>u%HrAPHOKU`vIHUzE(OAbi1@9&D_ccftOgxzFsx5WzVBTf6f*mX*Rf4 znjN`z0lsXy)@U?iMoBl&RCTAPj45;n#q>gA9?z9S<~Qai)jrR)rK6IVnjmj3*SJ02 zj3Bj@qm$@w;j)bLdv!;EHid2Z6)Migb>};cZA}Kc4mZEfMdOMx?4IZzE6I8ng`Z^8 zYE1T96V+|n9K+Gbwy-KkhXOfUi~n3wXE4t%46B$*Jl{uZ;~1=u zCb2@ziP4?}-CH6_9$hD`g1ogKQ}-^Vi76^^I!iwCCslG=nx&L{_^v5=ICmB-o#i{9 z*0vOJiF0~y&_-Q)Zx~|H-WvvX#S9+$-iyhV4}gfc*;ePCS&6Zp5mYt&*!7EXR?;<) zXs>WnW{@EJXw3Dx!pO@ggqV9PD6?{cyIoM0?~)+zKp^LoAef~#rFi72hnMJSGFU-Y zrO`x?HIGh>Qf^lI(n_2h(+zzQKMTrrY{9*witFm|Kp-+UaWE<=7IdEWfB(YC)AEX8 zh1OcU9i^czL$wp(m>0E2og#2BAgAV;o2{bsl!pb^rKqZL;_3zE%#7;HBPuG3i-e*fV~4S;qSx1kB|$z1-u4`jleQ+I5-H1|NmEkmwSuJDT(i(bj@=d} zTbzQ`gln2ESKyE0J$1xYYKd*0k?2yA>e$~eW-=NX5A`w??P(&~xep)jd=ytYW4_nh zJhZD>Xisy{Py`+stS&}LwLv$01)Ht4#zJ(hajWO|Tw%bHx=D)Qt69ZIr#AaWS9Z|T zX<7XC;r7my+4tb(^xz`*N2&ybw;ofwI2$7wIMv1}NN1Oiih{~o(6!JskdRx3hJuHt zzZFy@6>rNdD_@mjRNKi^BV|XjtHD7gxwf>5ErTS*OJQw~UzoXsPr_Ku>EZ@FNaSVX z>*jYh%Qd@Wxin2A+32uYRfmG+P>Yj2t(UD*$DuK-ouxpZr?^zaQ_favS>AZmZv42$ zb2x$FEc3Hn#bu71A$C$|aG4ObbI(vl<2~8XXgi;b>D1b0v(fb2Y_JiO;)r^6ZfTCr z$u;464}T}6M9+;Hg+l$&XD(54$dNOn{PC!fSD2*{BQNG&(;}P&91#}n4F#dB{oojV zv0gx{L>YChpdVF|diRuSuUbJHUlw-!d^sf~@M>SpL z_OOi7h*{AJmH;YZPV$IRI`Z^m^?#QUykFj;HPxSU|2f&fcQ&gvl@E;Q!w#P@ln-a` z%q!_jnpVn4qa@xf!M{rH*xQPh9BF6qOVSnUr-)JR%|!HX;)wAtDTB{^-2d5{(ES5s@}o<*y?oCR6$Zt^!>PWb;p2LDHdzl8s{ z{{Q|RKL5F39XJb&fWyGg;qQM2z6t&lJQb9|Y2a7z_-_Kw1PefP0fgTdAAq6{ApZZK z09FC9|NmDYcl%uiE(BH|;79QB?+0?f-}6Bgh_Ao%!2Q6X;5+d0-xhuz+yLGNUJ4!u z&IN~n-@(^^3w#n>37!is1{ZU z!gHrW4X0}B`0wmV1iAGmYeaEZ#O;oRaP8Lds{RgUXm=)H>8-!_9|?r?7~s&p)m9@| zi9C#kuRIK;@9MxrNzmll1FdTuG=g)6Y)%i@y8KT#TJY77@gCLpkEJ@z?RZR9Yx?5x zv-NJ&k=`QbVxm1X-2)s>^hT*(E}**acR_eyb} z5-z(%pVvNaUKZV`xB}gw?M7H=Ey(V*SICGdyV{GmmDNxuyROQLy0|IKpV4CP8YAH& zGIR6-%w@hU1bOrg9D#^mGwzc0la>l>E?+!Z_84^$&a<05RbN26i;X9z{LAzew9*~1 zJz1DvHkc&qR5%<0g%&YfBbuO>S5}5I^9`)1#b^fiE9x?vnPzzkjX8vh+(5wfMlH?& zl!C`F@6qm?5!0lgBD(wgNl!`}v2k6S-_6hAVy;Yl^2~+qhV{;D_?K!a^uKAji;RNm zjru^>^)FbDqQ%N6=IRA)apR!e6(!d;NcjJ+!*4zv{xm%Q7yJL4fXDzI4_e?%AaVd} z1MoR`ez601ArKkBC^!V%1Ka}d|IgsX;L%_LJPMow9s+)a4B%7XWk6&Co4|v>Ux7a$ z3-~G!`M|Tm6sUl$U|f6PMHcWPa20q8xELG-z7K!@74TjlwgQ)e zc_1=?+u-qq@BaXJ1GpT>JpldSq2K}FH}L(x2I3>|IpEnK1p`3t1o##Z+ky9hSAa=y z0yrN0HTWtrfv%DJx==Kcf7f%E3t}VLZ}`o?=U!Mk?YxA2{?1p;W=i*iZL*J(ut{E)t}+zV|Evt zbKw}GYh|EG1-Wv}i<9p1YsP{b?E0;)zY-uvM+mNID4*PWm#GO)Y&!aSFb?4y z!YKcArmr`Lr50uwv4}#C3(|5q%K^2xSF$?|zoV4CBUfIwQ#PfOdz8Nxd``HSxK69h z&z5ORDcD@ft(uA63$yWM)5=cSD(k4?B55VuuEs)bN`-a_g{37mv2KMn*n{gA2wcUh zvI{PKNXEN#qQFfPfAd6BPppY!lg7}Mp{Lj0U-;&DCpEvx(VTbyi#e!G{dqR)E~cBS z<5>23Hcwz-7~0~dL}g<<5-izwz_}R1qdsH|Zl<}$3|~&_HnKQhw$o>yAIXlwlfbT4 z@(G{Ne*Zj`3`r~1=ksU^+mvpYt+!NYW6(p#9e0nm+%IZ+krYEv?o1FX#5@!|CJlX?z3QHEz=`!VsHEX4S z1qo(9QE+_L#lgjiUBL#GZ(X->%pgjpn$hd1@nZepI33r3_?M}n(2i!ottm( P!8OldyWum{JDmRyt>Pnj literal 0 HcmV?d00001 diff --git a/egs/librispeech/ASR/conv_emformer_transducer_stateless2/asr_datamodule.py b/egs/librispeech/ASR/conv_emformer_transducer_stateless2/asr_datamodule.py new file mode 120000 index 000000000..104eeea5d --- /dev/null +++ b/egs/librispeech/ASR/conv_emformer_transducer_stateless2/asr_datamodule.py @@ -0,0 +1 @@ +../conv_emformer_transducer_stateless/asr_datamodule.py \ No newline at end of file diff --git a/egs/librispeech/ASR/conv_emformer_transducer_stateless2/beam_search.py b/egs/librispeech/ASR/conv_emformer_transducer_stateless2/beam_search.py new file mode 120000 index 000000000..57fd35665 --- /dev/null +++ b/egs/librispeech/ASR/conv_emformer_transducer_stateless2/beam_search.py @@ -0,0 +1 @@ +../conv_emformer_transducer_stateless/beam_search.py \ No newline at end of file diff --git a/egs/librispeech/ASR/conv_emformer_transducer_stateless2/decode.py b/egs/librispeech/ASR/conv_emformer_transducer_stateless2/decode.py new file mode 120000 index 000000000..a9e9e1576 --- /dev/null +++ b/egs/librispeech/ASR/conv_emformer_transducer_stateless2/decode.py @@ -0,0 +1 @@ +../conv_emformer_transducer_stateless/decode.py \ No newline at end of file diff --git a/egs/librispeech/ASR/conv_emformer_transducer_stateless2/decoder.py b/egs/librispeech/ASR/conv_emformer_transducer_stateless2/decoder.py new file mode 120000 index 000000000..1db262df7 --- /dev/null +++ b/egs/librispeech/ASR/conv_emformer_transducer_stateless2/decoder.py @@ -0,0 +1 @@ +../conv_emformer_transducer_stateless/decoder.py \ No newline at end of file diff --git a/egs/librispeech/ASR/conv_emformer_transducer_stateless2/emformer.py b/egs/librispeech/ASR/conv_emformer_transducer_stateless2/emformer.py new file mode 100644 index 000000000..46993da48 --- /dev/null +++ b/egs/librispeech/ASR/conv_emformer_transducer_stateless2/emformer.py @@ -0,0 +1,1898 @@ +# Copyright 2022 Xiaomi Corporation (Author: 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. +# +# It is modified based on +# 1) https://github.com/pytorch/audio/blob/main/torchaudio/models/emformer.py # noqa +# 2) https://github.com/pytorch/audio/blob/main/torchaudio/prototype/models/conv_emformer.py # noqa + +import math +from typing import List, Optional, Tuple + +import torch +import torch.nn as nn +from encoder_interface import EncoderInterface +from scaling import ( + ActivationBalancer, + BasicNorm, + DoubleSwish, + ScaledConv1d, + ScaledConv2d, + ScaledLinear, +) + +from icefall.utils import make_pad_mask + + +LOG_EPSILON = math.log(1e-10) + + +def unstack_states( + states: Tuple[List[List[torch.Tensor]], List[torch.Tensor]] +) -> List[Tuple[List[List[torch.Tensor]], List[torch.Tensor]]]: + """Unstack the emformer state corresponding to a batch of utterances + into a list of states, where the i-th entry is the state from the i-th + utterance in the batch. + + Args: + states: + A tuple of 2 elements. + ``states[0]`` is the attention caches of a batch of utterance. + ``states[1]`` is the convolution caches of a batch of utterance. + ``len(states[0])`` and ``len(states[1])`` both eqaul to number of layers. # noqa + + Returns: + A list of states. + ``states[i]`` is a tuple of 2 elements of i-th utterance. + ``states[i][0]`` is the attention caches of i-th utterance. + ``states[i][1]`` is the convolution caches of i-th utterance. + ``len(states[i][0])`` and ``len(states[i][1])`` both eqaul to number of layers. # noqa + """ + + attn_caches, conv_caches = states + batch_size = conv_caches[0].size(0) + num_layers = len(attn_caches) + + list_attn_caches = [None] * batch_size + for i in range(batch_size): + list_attn_caches[i] = [[] for _ in range(num_layers)] + for li, layer in enumerate(attn_caches): + for s in layer: + s_list = s.unbind(dim=1) + for bi, b in enumerate(list_attn_caches): + b[li].append(s_list[bi]) + + list_conv_caches = [None] * batch_size + for i in range(batch_size): + list_conv_caches[i] = [None] * num_layers + for li, layer in enumerate(conv_caches): + c_list = layer.unbind(dim=0) + for bi, b in enumerate(list_conv_caches): + b[li] = c_list[bi] + + ans = [None] * batch_size + for i in range(batch_size): + ans[i] = [list_attn_caches[i], list_conv_caches[i]] + + return ans + + +def stack_states( + state_list: List[Tuple[List[List[torch.Tensor]], List[torch.Tensor]]] +) -> Tuple[List[List[torch.Tensor]], List[torch.Tensor]]: + """Stack list of emformer states that correspond to separate utterances + into a single emformer state so that it can be used as an input for + emformer when those utterances are formed into a batch. + + Note: + It is the inverse of :func:`unstack_states`. + + Args: + state_list: + Each element in state_list corresponding to the internal state + of the emformer model for a single utterance. + ``states[i]`` is a tuple of 2 elements of i-th utterance. + ``states[i][0]`` is the attention caches of i-th utterance. + ``states[i][1]`` is the convolution caches of i-th utterance. + ``len(states[i][0])`` and ``len(states[i][1])`` both eqaul to number of layers. # noqa + + Returns: + A new state corresponding to a batch of utterances. + See the input argument of :func:`unstack_states` for the meaning + of the returned tensor. + """ + batch_size = len(state_list) + + attn_caches = [] + for layer in state_list[0][0]: + if batch_size > 1: + # Note: We will stack attn_caches[layer][s][] later to get attn_caches[layer][s] # noqa + attn_caches.append([[s] for s in layer]) + else: + attn_caches.append([s.unsqueeze(1) for s in layer]) + for b, states in enumerate(state_list[1:], 1): + for li, layer in enumerate(states[0]): + for si, s in enumerate(layer): + attn_caches[li][si].append(s) + if b == batch_size - 1: + attn_caches[li][si] = torch.stack( + attn_caches[li][si], dim=1 + ) + + conv_caches = [] + for layer in state_list[0][1]: + if batch_size > 1: + # Note: We will stack conv_caches[layer][] later to get conv_caches[layer] # noqa + conv_caches.append([layer]) + else: + conv_caches.append(layer.unsqueeze(0)) + for b, states in enumerate(state_list[1:], 1): + for li, layer in enumerate(states[1]): + conv_caches[li].append(layer) + if b == batch_size - 1: + conv_caches[li] = torch.stack(conv_caches[li], dim=0) + + return [attn_caches, conv_caches] + + +class ConvolutionModule(nn.Module): + """ConvolutionModule. + + Modified from https://github.com/pytorch/audio/blob/main/torchaudio/prototype/models/conv_emformer.py # noqa + + Args: + chunk_length (int): + Length of each chunk. + right_context_length (int): + Length of right context. + channels (int): + The number of input channels and output 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, + chunk_length: int, + right_context_length: int, + channels: int, + kernel_size: int, + bias: bool = True, + ) -> None: + """Construct an ConvolutionModule object.""" + super().__init__() + # kernerl_size should be an odd number for 'SAME' padding + assert (kernel_size - 1) % 2 == 0, kernel_size + + self.chunk_length = chunk_length + self.right_context_length = right_context_length + self.channels = channels + + self.pointwise_conv1 = ScaledConv1d( + 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( + channel_dim=1, max_abs=10.0, min_positive=0.05, max_positive=1.0 + ) + + # make it causal by padding cached (kernel_size - 1) frames on the left + self.cache_size = kernel_size - 1 + self.depthwise_conv = ScaledConv1d( + channels, + channels, + kernel_size, + stride=1, + padding=0, + groups=channels, + bias=bias, + ) + + self.deriv_balancer2 = ActivationBalancer( + channel_dim=1, min_positive=0.05, max_positive=1.0 + ) + + self.activation = DoubleSwish() + + self.pointwise_conv2 = ScaledConv1d( + channels, + channels, + kernel_size=1, + stride=1, + padding=0, + bias=bias, + initial_scale=0.25, + ) + + def _split_right_context( + self, + pad_utterance: torch.Tensor, + right_context: torch.Tensor, + ) -> torch.Tensor: + """ + Args: + pad_utterance: + Its shape is (cache_size + U, B, D). + right_context: + Its shape is (R, B, D). + + Returns: + Right context segments padding with corresponding context. + Its shape is (num_segs * B, D, cache_size + right_context_length). + """ + U_, B, D = pad_utterance.size() + R = right_context.size(0) + assert self.right_context_length != 0 + assert R % self.right_context_length == 0 + num_chunks = R // self.right_context_length + right_context = right_context.reshape( + num_chunks, self.right_context_length, B, D + ) + right_context = right_context.permute(0, 2, 1, 3).reshape( + num_chunks * B, self.right_context_length, D + ) + + intervals = torch.arange( + 0, self.chunk_length * (num_chunks - 1), self.chunk_length + ) + first = torch.arange( + self.chunk_length, self.chunk_length + self.cache_size + ) + indexes = intervals.unsqueeze(1) + first.unsqueeze(0) + indexes = torch.cat( + [indexes, torch.arange(U_ - self.cache_size, U_).unsqueeze(0)] + ) + padding = pad_utterance[indexes] # (num_chunks, cache_size, B, D) + padding = padding.permute(0, 2, 1, 3).reshape( + num_chunks * B, self.cache_size, D + ) + + pad_right_context = torch.cat([padding, right_context], dim=1) + # (num_chunks * B, cache_size + right_context_length, D) + return pad_right_context.permute(0, 2, 1) + + def _merge_right_context( + self, right_context: torch.Tensor, B: int + ) -> torch.Tensor: + """ + Args: + right_context: + Right context segments. + It shape is (num_segs * B, D, right_context_length). + B: + Batch size. + + Returns: + A tensor of shape (B, D, R), where + R = num_segs * right_context_length. + """ + right_context = right_context.reshape( + -1, B, self.channels, self.right_context_length + ) + right_context = right_context.permute(1, 2, 0, 3) + right_context = right_context.reshape(B, self.channels, -1) + return right_context + + def forward( + self, + utterance: torch.Tensor, + right_context: torch.Tensor, + ) -> Tuple[torch.Tensor, torch.Tensor]: + """Causal convolution module. + + Args: + utterance (torch.Tensor): + Utterance tensor of shape (U, B, D). + right_context (torch.Tensor): + Right context tensor of shape (R, B, D). + + Returns: + A tuple of 2 tensors: + - output utterance of shape (U, B, D). + - output right_context of shape (R, B, D). + """ + U, B, D = utterance.size() + R, _, _ = right_context.size() + + # point-wise conv and GLU mechanism + x = torch.cat([right_context, utterance], dim=0) # (R + U, B, D) + x = x.permute(1, 2, 0) # (B, D, R + U) + x = self.pointwise_conv1(x) # (B, 2 * D, R + U) + x = self.deriv_balancer1(x) + x = nn.functional.glu(x, dim=1) # (B, D, R + U) + utterance = x[:, :, R:] # (B, D, U) + right_context = x[:, :, :R] # (B, D, R) + + # make causal convolution + cache = torch.zeros( + B, D, self.cache_size, device=x.device, dtype=x.dtype + ) + pad_utterance = torch.cat( + [cache, utterance], dim=2 + ) # (B, D, cache + U) + + # depth-wise conv on utterance + utterance = self.depthwise_conv(pad_utterance) # (B, D, U) + + if self.right_context_length > 0: + # depth-wise conv on right_context + pad_right_context = self._split_right_context( + pad_utterance.permute(2, 0, 1), right_context.permute(2, 0, 1) + ) # (num_segs * B, D, cache_size + right_context_length) + right_context = self.depthwise_conv( + pad_right_context + ) # (num_segs * B, D, right_context_length) + right_context = self._merge_right_context( + right_context, B + ) # (B, D, R) + + x = torch.cat([right_context, utterance], dim=2) # (B, D, R + U) + x = self.deriv_balancer2(x) + x = self.activation(x) + + # point-wise conv + x = self.pointwise_conv2(x) # (B, D, R + U) + + right_context = x[:, :, :R] # (B, D, R) + utterance = x[:, :, R:] # (B, D, U) + return ( + utterance.permute(2, 0, 1), + right_context.permute(2, 0, 1), + ) + + @torch.jit.export + def infer( + self, + utterance: torch.Tensor, + right_context: torch.Tensor, + cache: torch.Tensor, + ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]: + """Causal convolution module applied on both utterance and right_context. + + Args: + utterance (torch.Tensor): + Utterance tensor of shape (U, B, D). + right_context (torch.Tensor): + Right context tensor of shape (R, B, D). + cache (torch.Tensor, optional): + Cached tensor for left padding of shape (B, D, cache_size). + + Returns: + A tuple of 3 tensors: + - output utterance of shape (U, B, D). + - output right_context of shape (R, B, D). + - updated cache tensor of shape (B, D, cache_size). + """ + U, B, D = utterance.size() + R, _, _ = right_context.size() + + # point-wise conv + x = torch.cat([utterance, right_context], dim=0) # (U + R, B, D) + x = x.permute(1, 2, 0) # (B, D, U + R) + x = self.pointwise_conv1(x) # (B, 2 * D, U + R) + x = self.deriv_balancer1(x) + x = nn.functional.glu(x, dim=1) # (B, D, U + R) + + # make causal convolution + assert cache.shape == (B, D, self.cache_size), cache.shape + x = torch.cat([cache, x], dim=2) # (B, D, cache_size + U + R) + # update cache + new_cache = x[:, :, -R - self.cache_size : -R] + + # 1-D depth-wise conv + x = self.depthwise_conv(x) # (B, D, U + R) + + x = self.deriv_balancer2(x) + x = self.activation(x) + + # point-wise conv + x = self.pointwise_conv2(x) # (B, D, U + R) + + utterance = x[:, :, :U] # (B, D, U) + right_context = x[:, :, U:] # (B, D, R) + return ( + utterance.permute(2, 0, 1), + right_context.permute(2, 0, 1), + new_cache, + ) + + +class EmformerAttention(nn.Module): + r"""Emformer layer attention module. + + Args: + embed_dim (int): + Embedding dimension. + nhead (int): + Number of attention heads in each Emformer layer. + dropout (float, optional): + Dropout probability. (Default: 0.0) + tanh_on_mem (bool, optional): + If ``True``, applies tanh to memory elements. (Default: ``False``) + negative_inf (float, optional): + Value to use for negative infinity in attention weights. (Default: -1e8) + """ + + def __init__( + self, + embed_dim: int, + nhead: int, + dropout: float = 0.0, + tanh_on_mem: bool = False, + negative_inf: float = -1e8, + ): + super().__init__() + + if embed_dim % nhead != 0: + raise ValueError( + f"embed_dim ({embed_dim}) is not a multiple of" + f"nhead ({nhead})." + ) + + self.embed_dim = embed_dim + self.nhead = nhead + self.tanh_on_mem = tanh_on_mem + self.negative_inf = negative_inf + self.head_dim = embed_dim // nhead + self.dropout = dropout + + self.emb_to_key_value = ScaledLinear( + embed_dim, 2 * embed_dim, bias=True + ) + self.emb_to_query = ScaledLinear(embed_dim, embed_dim, bias=True) + self.out_proj = ScaledLinear( + embed_dim, embed_dim, bias=True, initial_scale=0.25 + ) + + def _gen_attention_probs( + self, + attention_weights: torch.Tensor, + attention_mask: torch.Tensor, + padding_mask: Optional[torch.Tensor] = None, + ) -> torch.Tensor: + """Given the entire attention weights, mask out unecessary connections + and optionally with padding positions, to obtain underlying chunk-wise + attention probabilities. + + B: batch size; + Q: length of query; + KV: length of key and value. + + Args: + attention_weights (torch.Tensor): + Attention weights computed on the entire concatenated tensor + with shape (B * nhead, Q, KV). + attention_mask (torch.Tensor): + Mask tensor where chunk-wise connections are filled with `False`, + and other unnecessary connections are filled with `True`, + with shape (Q, KV). + padding_mask (torch.Tensor, optional): + Mask tensor where the padding positions are fill with `True`, + and other positions are filled with `False`, with shapa `(B, KV)`. + + Returns: + A tensor of shape (B * nhead, Q, KV). + """ + attention_weights_float = attention_weights.float() + attention_weights_float = attention_weights_float.masked_fill( + attention_mask.unsqueeze(0), self.negative_inf + ) + if padding_mask is not None: + Q = attention_weights.size(1) + B = attention_weights.size(0) // self.nhead + attention_weights_float = attention_weights_float.view( + B, self.nhead, Q, -1 + ) + attention_weights_float = attention_weights_float.masked_fill( + padding_mask.unsqueeze(1).unsqueeze(2).to(torch.bool), + self.negative_inf, + ) + attention_weights_float = attention_weights_float.view( + B * self.nhead, Q, -1 + ) + + attention_probs = nn.functional.softmax( + attention_weights_float, dim=-1 + ).type_as(attention_weights) + + attention_probs = nn.functional.dropout( + attention_probs, p=self.dropout, training=self.training + ) + return attention_probs + + def _forward_impl( + self, + utterance: torch.Tensor, + right_context: torch.Tensor, + summary: torch.Tensor, + memory: torch.Tensor, + attention_mask: torch.Tensor, + padding_mask: Optional[torch.Tensor] = None, + left_context_key: Optional[torch.Tensor] = None, + left_context_val: Optional[torch.Tensor] = None, + ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor]: + """Underlying chunk-wise attention implementation.""" + U, B, _ = utterance.size() + R = right_context.size(0) + M = memory.size(0) + scaling = float(self.head_dim) ** -0.5 + + # compute query with [right_context, utterance, summary]. + query = self.emb_to_query( + torch.cat([right_context, utterance, summary]) + ) + # compute key and value with [memory, right_context, utterance]. + key, value = self.emb_to_key_value( + torch.cat([memory, right_context, utterance]) + ).chunk(chunks=2, dim=2) + + if left_context_key is not None and left_context_val is not None: + # now compute key and value with + # [memory, right context, left context, uttrance] + # this is used in inference mode + key = torch.cat([key[: M + R], left_context_key, key[M + R :]]) + value = torch.cat( + [value[: M + R], left_context_val, value[M + R :]] + ) + Q = query.size(0) + # KV = key.size(0) + + reshaped_query, reshaped_key, reshaped_value = [ + tensor.contiguous() + .view(-1, B * self.nhead, self.head_dim) + .transpose(0, 1) + for tensor in [query, key, value] + ] # (B * nhead, Q or KV, head_dim) + attention_weights = torch.bmm( + reshaped_query * scaling, reshaped_key.transpose(1, 2) + ) # (B * nhead, Q, KV) + + # compute attention probabilities + attention_probs = self._gen_attention_probs( + attention_weights, attention_mask, padding_mask + ) + + # compute attention outputs + attention = torch.bmm(attention_probs, reshaped_value) + assert attention.shape == (B * self.nhead, Q, self.head_dim) + attention = ( + attention.transpose(0, 1).contiguous().view(Q, B, self.embed_dim) + ) + + # apply output projection + outputs = self.out_proj(attention) + + output_right_context_utterance = outputs[: R + U] + output_memory = outputs[R + U :] + if self.tanh_on_mem: + output_memory = torch.tanh(output_memory) + else: + output_memory = torch.clamp(output_memory, min=-10, max=10) + + return output_right_context_utterance, output_memory, key, value + + def forward( + self, + utterance: torch.Tensor, + right_context: torch.Tensor, + summary: torch.Tensor, + memory: torch.Tensor, + attention_mask: torch.Tensor, + padding_mask: Optional[torch.Tensor] = None, + ) -> Tuple[torch.Tensor, torch.Tensor]: + # TODO: Modify docs. + """Forward pass for training and validation mode. + + B: batch size; + D: embedding dimension; + R: length of the hard-copied right contexts; + U: length of full utterance; + S: length of summary vectors; + M: length of memory vectors. + + It computes a `big` attention matrix on full utterance and + then utilizes a pre-computed mask to simulate chunk-wise attention. + + It concatenates three blocks: hard-copied right contexts, + full utterance, and summary vectors, as a `big` block, + to compute the query tensor: + query = [right_context, utterance, summary], + with length Q = R + U + S. + It concatenates the three blocks: memory vectors, + hard-copied right contexts, and full utterance as another `big` block, + to compute the key and value tensors: + key & value = [memory, right_context, utterance], + with length KV = M + R + U. + Attention scores is computed with above `big` query and key. + + Then the underlying chunk-wise attention is obtained by applying + the attention mask. Suppose + c_i: chunk at index i; + r_i: right context that c_i can use; + l_i: left context that c_i can use; + m_i: past memory vectors from previous layer that c_i can use; + s_i: summary vector of c_i; + The target chunk-wise attention is: + c_i, r_i (in query) -> l_i, c_i, r_i, m_i (in key); + s_i (in query) -> l_i, c_i, r_i (in key). + + Args: + utterance (torch.Tensor): + Full utterance frames, with shape (U, B, D). + right_context (torch.Tensor): + Hard-copied right context frames, with shape (R, B, D), + where R = num_chunks * right_context_length + summary (torch.Tensor): + Summary elements with shape (S, B, D), where S = num_chunks. + It is an empty tensor without using memory. + memory (torch.Tensor): + Memory elements, with shape (M, B, D), where M = num_chunks - 1. + It is an empty tensor without using memory. + attention_mask (torch.Tensor): + Pre-computed attention mask to simulate underlying chunk-wise + attention, with shape (Q, KV). + padding_mask (torch.Tensor): + Padding mask of key tensor, with shape (B, KV). + + Returns: + A tuple containing 2 tensors: + - output of right context and utterance, with shape (R + U, B, D). + - memory output, with shape (M, B, D), where M = S - 1 or M = 0. + """ + ( + output_right_context_utterance, + output_memory, + _, + _, + ) = self._forward_impl( + utterance, + right_context, + summary, + memory, + attention_mask, + padding_mask=padding_mask, + ) + return output_right_context_utterance, output_memory[:-1] + + @torch.jit.export + def infer( + self, + utterance: torch.Tensor, + right_context: torch.Tensor, + summary: torch.Tensor, + memory: torch.Tensor, + left_context_key: torch.Tensor, + left_context_val: torch.Tensor, + padding_mask: Optional[torch.Tensor] = None, + ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor]: + """Forward pass for inference. + + B: batch size; + D: embedding dimension; + R: length of right context; + U: length of utterance, i.e., current chunk; + L: length of cached left context; + S: length of summary vectors, S = 1; + M: length of cached memory vectors. + + It concatenates the right context, utterance (i.e., current chunk) + and summary vector of current chunk, to compute the query tensor: + query = [right_context, utterance, summary], + with length Q = R + U + S. + It concatenates the memory vectors, right context, left context, and + current chunk, to compute the key and value tensors: + key & value = [memory, right_context, left_context, utterance], + with length KV = M + R + L + U. + + The chunk-wise attention is: + chunk, right context (in query) -> + left context, chunk, right context, memory vectors (in key); + summary (in query) -> left context, chunk, right context (in key). + + Args: + utterance (torch.Tensor): + Current chunk frames, with shape (U, B, D), where U = chunk_length. + right_context (torch.Tensor): + Right context frames, with shape (R, B, D), + where R = right_context_length. + summary (torch.Tensor): + Summary vector with shape (1, B, D), or empty tensor. + memory (torch.Tensor): + Memory vectors, with shape (M, B, D), or empty tensor. + left_context_key (torch,Tensor): + Cached attention key of left context from preceding computation, + with shape (L, B, D). + left_context_val (torch.Tensor): + Cached attention value of left context from preceding computation, + with shape (L, B, D). + padding_mask (torch.Tensor): + Padding mask of key tensor, with shape (B, KV). + + Returns: + A tuple containing 4 tensors: + - output of right context and utterance, with shape (R + U, B, D). + - memory output, with shape (1, B, D) or (0, B, D). + - attention key of left context and utterance, which would be cached + for next computation, with shape (L + U, B, D). + - attention value of left context and utterance, which would be + cached for next computation, with shape (L + U, B, D). + """ + U = utterance.size(0) + R = right_context.size(0) + L = left_context_key.size(0) + S = summary.size(0) + M = memory.size(0) + + # TODO: move it outside + # query = [right context, utterance, summary] + Q = R + U + S + # key, value = [memory, right context, left context, uttrance] + KV = M + R + L + U + attention_mask = torch.zeros(Q, KV).to( + dtype=torch.bool, device=utterance.device + ) + # disallow attention bettween the summary vector with the memory bank + attention_mask[-1, :M] = True + ( + output_right_context_utterance, + output_memory, + key, + value, + ) = self._forward_impl( + utterance, + right_context, + summary, + memory, + attention_mask, + padding_mask=padding_mask, + left_context_key=left_context_key, + left_context_val=left_context_val, + ) + return ( + output_right_context_utterance, + output_memory, + key[M + R :], + value[M + R :], + ) + + +class EmformerEncoderLayer(nn.Module): + """Emformer layer that constitutes Emformer. + + Args: + d_model (int): + Input dimension. + nhead (int): + Number of attention heads. + dim_feedforward (int): + Hidden layer dimension of feedforward network. + chunk_length (int): + Length of each input segment. + dropout (float, optional): + Dropout probability. (Default: 0.0) + layer_dropout (float, optional): + Layer dropout probability. (Default: 0.0) + cnn_module_kernel (int): + Kernel size of convolution module. + left_context_length (int, optional): + Length of left context. (Default: 0) + right_context_length (int, optional): + Length of right context. (Default: 0) + memory_size (int, optional): + Number of memory elements to use. (Default: 0) + tanh_on_mem (bool, optional): + If ``True``, applies tanh to memory elements. (Default: ``False``) + negative_inf (float, optional): + Value to use for negative infinity in attention weights. (Default: -1e8) + """ + + def __init__( + self, + d_model: int, + nhead: int, + dim_feedforward: int, + chunk_length: int, + dropout: float = 0.1, + layer_dropout: float = 0.075, + cnn_module_kernel: int = 31, + left_context_length: int = 0, + right_context_length: int = 0, + memory_size: int = 0, + tanh_on_mem: bool = False, + negative_inf: float = -1e8, + ): + super().__init__() + + self.attention = EmformerAttention( + embed_dim=d_model, + nhead=nhead, + dropout=dropout, + tanh_on_mem=tanh_on_mem, + negative_inf=negative_inf, + ) + self.summary_op = nn.AvgPool1d( + kernel_size=chunk_length, stride=chunk_length, ceil_mode=True + ) + + self.feed_forward_macaron = nn.Sequential( + ScaledLinear(d_model, dim_feedforward), + ActivationBalancer(channel_dim=-1), + DoubleSwish(), + nn.Dropout(dropout), + ScaledLinear(dim_feedforward, d_model, initial_scale=0.25), + ) + + self.feed_forward = nn.Sequential( + ScaledLinear(d_model, dim_feedforward), + ActivationBalancer(channel_dim=-1), + DoubleSwish(), + nn.Dropout(dropout), + ScaledLinear(dim_feedforward, d_model, initial_scale=0.25), + ) + + self.conv_module = ConvolutionModule( + chunk_length, + right_context_length, + d_model, + cnn_module_kernel, + ) + + self.norm_final = BasicNorm(d_model) + + # try to ensure the output is close to zero-mean + # (or at least, zero-median). + self.balancer = ActivationBalancer( + channel_dim=-1, min_positive=0.45, max_positive=0.55, max_abs=6.0 + ) + + self.dropout = nn.Dropout(dropout) + + self.layer_dropout = layer_dropout + self.left_context_length = left_context_length + self.chunk_length = chunk_length + self.memory_size = memory_size + self.d_model = d_model + self.use_memory = memory_size > 0 + + def _update_attn_cache( + self, + next_key: torch.Tensor, + next_val: torch.Tensor, + memory: torch.Tensor, + attn_cache: List[torch.Tensor], + ) -> List[torch.Tensor]: + """Update cached attention state: + 1) output memory of current chunk in the lower layer; + 2) attention key and value in current chunk's computation, which would + be resued in next chunk's computation. + """ + new_memory = torch.cat([attn_cache[0], memory]) + new_key = torch.cat([attn_cache[1], next_key]) + new_val = torch.cat([attn_cache[2], next_val]) + attn_cache[0] = new_memory[new_memory.size(0) - self.memory_size :] + attn_cache[1] = new_key[new_key.size(0) - self.left_context_length :] + attn_cache[2] = new_val[new_val.size(0) - self.left_context_length :] + return attn_cache + + def _apply_conv_module_forward( + self, + right_context_utterance: torch.Tensor, + R: int, + ) -> torch.Tensor: + """Apply convolution module in training and validation mode.""" + utterance = right_context_utterance[R:] + right_context = right_context_utterance[:R] + utterance, right_context = self.conv_module(utterance, right_context) + right_context_utterance = torch.cat([right_context, utterance]) + return right_context_utterance + + def _apply_conv_module_infer( + self, + right_context_utterance: torch.Tensor, + R: int, + conv_cache: torch.Tensor, + ) -> Tuple[torch.Tensor, torch.Tensor]: + """Apply convolution module on utterance in inference mode.""" + utterance = right_context_utterance[R:] + right_context = right_context_utterance[:R] + utterance, right_context, conv_cache = self.conv_module.infer( + utterance, right_context, conv_cache + ) + right_context_utterance = torch.cat([right_context, utterance]) + return right_context_utterance, conv_cache + + def _apply_attention_module_forward( + self, + right_context_utterance: torch.Tensor, + R: int, + memory: torch.Tensor, + attention_mask: torch.Tensor, + padding_mask: Optional[torch.Tensor] = None, + ) -> Tuple[torch.Tensor, torch.Tensor]: + """Apply attention module in training and validation mode.""" + utterance = right_context_utterance[R:] + right_context = right_context_utterance[:R] + + if self.use_memory: + summary = self.summary_op(utterance.permute(1, 2, 0)).permute( + 2, 0, 1 + ) + else: + summary = torch.empty(0).to( + dtype=utterance.dtype, device=utterance.device + ) + output_right_context_utterance, output_memory = self.attention( + utterance=utterance, + right_context=right_context, + summary=summary, + memory=memory, + attention_mask=attention_mask, + padding_mask=padding_mask, + ) + + return output_right_context_utterance, output_memory + + def _apply_attention_module_infer( + self, + right_context_utterance: torch.Tensor, + R: int, + memory: torch.Tensor, + attn_cache: List[torch.Tensor], + padding_mask: Optional[torch.Tensor] = None, + ) -> Tuple[torch.Tensor, torch.Tensor, List[torch.Tensor]]: + """Apply attention module in inference mode. + 1) Unpack cached states including: + - memory from previous chunks in the lower layer; + - attention key and value of left context from preceding + chunk's compuation; + 2) Apply attention computation; + 3) Update cached attention states including: + - output memory of current chunk in the lower layer; + - attention key and value in current chunk's computation, which would + be resued in next chunk's computation. + """ + utterance = right_context_utterance[R:] + right_context = right_context_utterance[:R] + + pre_memory = attn_cache[0] + left_context_key = attn_cache[1] + left_context_val = attn_cache[2] + + if self.use_memory: + summary = self.summary_op(utterance.permute(1, 2, 0)).permute( + 2, 0, 1 + ) + summary = summary[:1] + else: + summary = torch.empty(0).to( + dtype=utterance.dtype, device=utterance.device + ) + ( + output_right_context_utterance, + output_memory, + next_key, + next_val, + ) = self.attention.infer( + utterance=utterance, + right_context=right_context, + summary=summary, + memory=pre_memory, + left_context_key=left_context_key, + left_context_val=left_context_val, + padding_mask=padding_mask, + ) + attn_cache = self._update_attn_cache( + next_key, next_val, memory, attn_cache + ) + return output_right_context_utterance, output_memory, attn_cache + + def forward( + self, + utterance: torch.Tensor, + right_context: torch.Tensor, + memory: torch.Tensor, + attention_mask: torch.Tensor, + padding_mask: Optional[torch.Tensor] = None, + warmup: float = 1.0, + ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]: + r"""Forward pass for training and validation mode. + + B: batch size; + D: embedding dimension; + R: length of hard-copied right contexts; + U: length of full utterance; + M: length of memory vectors. + + Args: + utterance (torch.Tensor): + Utterance frames, with shape (U, B, D). + right_context (torch.Tensor): + Right context frames, with shape (R, B, D). + memory (torch.Tensor): + Memory elements, with shape (M, B, D). + It is an empty tensor without using memory. + attention_mask (torch.Tensor): + Attention mask for underlying attention module, + with shape (Q, KV), where Q = R + U + S, KV = M + R + U. + padding_mask (torch.Tensor): + Padding mask of ker tensor, with shape (B, KV). + + Returns: + A tuple containing 3 tensors: + - output utterance, with shape (U, B, D). + - output right context, with shape (R, B, D). + - output memory, with shape (M, B, D). + """ + R = right_context.size(0) + src = torch.cat([right_context, utterance]) + src_orig = src + + warmup_scale = min(0.1 + warmup, 1.0) + # alpha = 1.0 means fully use this encoder layer, 0.0 would mean + # completely bypass it. + if self.training: + alpha = ( + warmup_scale + if torch.rand(()).item() <= (1.0 - self.layer_dropout) + else 0.1 + ) + else: + alpha = 1.0 + + # macaron style feed forward module + src = src + self.dropout(self.feed_forward_macaron(src)) + + # emformer attention module + src_att, output_memory = self._apply_attention_module_forward( + src, R, memory, attention_mask, padding_mask=padding_mask + ) + src = src + self.dropout(src_att) + + # convolution module + src_conv = self._apply_conv_module_forward(src, R) + src = src + self.dropout(src_conv) + + # feed forward module + src = src + self.dropout(self.feed_forward(src)) + + src = self.norm_final(self.balancer(src)) + + if alpha != 1.0: + src = alpha * src + (1 - alpha) * src_orig + + output_utterance = src[R:] + output_right_context = src[:R] + return output_utterance, output_right_context, output_memory + + @torch.jit.export + def infer( + self, + utterance: torch.Tensor, + right_context: torch.Tensor, + memory: torch.Tensor, + attn_cache: List[torch.Tensor], + conv_cache: torch.Tensor, + padding_mask: Optional[torch.Tensor] = None, + ) -> Tuple[ + torch.Tensor, + torch.Tensor, + torch.Tensor, + List[torch.Tensor], + torch.Tensor, + ]: + """Forward pass for inference. + + B: batch size; + D: embedding dimension; + R: length of right_context; + U: length of utterance; + M: length of memory. + + Args: + utterance (torch.Tensor): + Utterance frames, with shape (U, B, D). + right_context (torch.Tensor): + Right context frames, with shape (R, B, D). + memory (torch.Tensor): + Memory elements, with shape (M, B, D). + attn_cache (List[torch.Tensor]): + Cached attention tensors generated in preceding computation, + including memory, key and value of left context. + conv_cache (torch.Tensor, optional): + Cache tensor of left context for causal convolution. + padding_mask (torch.Tensor): + Padding mask of ker tensor. + + Returns: + (Tensor, Tensor, List[torch.Tensor], Tensor): + - output utterance, with shape (U, B, D); + - output right_context, with shape (R, B, D); + - output memory, with shape (1, B, D) or (0, B, D). + - output state. + - updated conv_cache. + """ + R = right_context.size(0) + src = torch.cat([right_context, utterance]) + + # macaron style feed forward module + src = src + self.dropout(self.feed_forward_macaron(src)) + + # emformer attention module + ( + src_att, + output_memory, + attn_cache, + ) = self._apply_attention_module_infer( + src, R, memory, attn_cache, padding_mask=padding_mask + ) + src = src + self.dropout(src_att) + + # convolution module + src_conv, conv_cache = self._apply_conv_module_infer(src, R, conv_cache) + src = src + self.dropout(src_conv) + + # feed forward module + src = src + self.dropout(self.feed_forward(src)) + + src = self.norm_final(self.balancer(src)) + + output_utterance = src[R:] + output_right_context = src[:R] + return ( + output_utterance, + output_right_context, + output_memory, + attn_cache, + conv_cache, + ) + + +def _gen_attention_mask_block( + col_widths: List[int], + col_mask: List[bool], + num_rows: int, + device: torch.device, +) -> torch.Tensor: + assert len(col_widths) == len( + col_mask + ), "Length of col_widths must match that of col_mask" + + mask_block = [ + torch.ones(num_rows, col_width, device=device) + if is_ones_col + else torch.zeros(num_rows, col_width, device=device) + for col_width, is_ones_col in zip(col_widths, col_mask) + ] + return torch.cat(mask_block, dim=1) + + +class EmformerEncoder(nn.Module): + """Implements the Emformer architecture introduced in + *Emformer: Efficient Memory Transformer Based Acoustic Model for Low Latency + Streaming Speech Recognition* + [:footcite:`shi2021emformer`]. + + Args: + d_model (int): + Input dimension. + nhead (int): + Number of attention heads in each emformer layer. + dim_feedforward (int): + Hidden layer dimension of each emformer layer's feedforward network. + num_encoder_layers (int): + Number of emformer layers to instantiate. + chunk_length (int): + Length of each input segment. + dropout (float, optional): + Dropout probability. (default: 0.0) + layer_dropout (float, optional): + Layer dropout probability. (default: 0.0) + cnn_module_kernel (int): + Kernel size of convolution module. + left_context_length (int, optional): + Length of left context. (default: 0) + right_context_length (int, optional): + Length of right context. (default: 0) + memory_size (int, optional): + Number of memory elements to use. (default: 0) + tanh_on_mem (bool, optional): + If ``true``, applies tanh to memory elements. (default: ``false``) + negative_inf (float, optional): + Value to use for negative infinity in attention weights. (default: -1e8) + """ + + def __init__( + self, + chunk_length: int, + d_model: int = 256, + nhead: int = 4, + dim_feedforward: int = 2048, + num_encoder_layers: int = 12, + dropout: float = 0.1, + layer_dropout: float = 0.075, + cnn_module_kernel: int = 31, + left_context_length: int = 0, + right_context_length: int = 0, + memory_size: int = 0, + tanh_on_mem: bool = False, + negative_inf: float = -1e8, + ): + super().__init__() + + self.use_memory = memory_size > 0 + self.init_memory_op = nn.AvgPool1d( + kernel_size=chunk_length, + stride=chunk_length, + ceil_mode=True, + ) + + self.emformer_layers = nn.ModuleList( + [ + EmformerEncoderLayer( + d_model=d_model, + nhead=nhead, + dim_feedforward=dim_feedforward, + chunk_length=chunk_length, + dropout=dropout, + layer_dropout=layer_dropout, + cnn_module_kernel=cnn_module_kernel, + left_context_length=left_context_length, + right_context_length=right_context_length, + memory_size=memory_size, + tanh_on_mem=tanh_on_mem, + negative_inf=negative_inf, + ) + for layer_idx in range(num_encoder_layers) + ] + ) + + self.num_encoder_layers = num_encoder_layers + self.d_model = d_model + self.left_context_length = left_context_length + self.right_context_length = right_context_length + self.chunk_length = chunk_length + self.memory_size = memory_size + self.cnn_module_kernel = cnn_module_kernel + + def _gen_right_context(self, x: torch.Tensor) -> torch.Tensor: + """Hard copy each chunk's right context and concat them.""" + T = x.shape[0] + num_chunks = math.ceil( + (T - self.right_context_length) / self.chunk_length + ) + # first (num_chunks - 1) right context block + intervals = torch.arange( + 0, self.chunk_length * (num_chunks - 1), self.chunk_length + ) + first = torch.arange( + self.chunk_length, self.chunk_length + self.right_context_length + ) + indexes = intervals.unsqueeze(1) + first.unsqueeze(0) + # cat last right context block + indexes = torch.cat( + [ + indexes, + torch.arange(T - self.right_context_length, T).unsqueeze(0), + ] + ) + right_context_blocks = x[indexes.reshape(-1)] + return right_context_blocks + + def _gen_attention_mask_col_widths( + self, chunk_idx: int, U: int + ) -> List[int]: + """Calculate column widths (key, value) in attention mask for the + chunk_idx chunk.""" + num_chunks = math.ceil(U / self.chunk_length) + rc = self.right_context_length + lc = self.left_context_length + rc_start = chunk_idx * rc + rc_end = rc_start + rc + chunk_start = max(chunk_idx * self.chunk_length - lc, 0) + chunk_end = min((chunk_idx + 1) * self.chunk_length, U) + R = rc * num_chunks + + if self.use_memory: + m_start = max(chunk_idx - self.memory_size, 0) + M = num_chunks - 1 + col_widths = [ + m_start, # before memory + chunk_idx - m_start, # memory + M - chunk_idx, # after memory + rc_start, # before right context + rc, # right context + R - rc_end, # after right context + chunk_start, # before chunk + chunk_end - chunk_start, # chunk + U - chunk_end, # after chunk + ] + else: + col_widths = [ + rc_start, # before right context + rc, # right context + R - rc_end, # after right context + chunk_start, # before chunk + chunk_end - chunk_start, # chunk + U - chunk_end, # after chunk + ] + + return col_widths + + def _gen_attention_mask(self, utterance: torch.Tensor) -> torch.Tensor: + """Generate attention mask to simulate underlying chunk-wise attention + computation, where chunk-wise connections are filled with `False`, + and other unnecessary connections beyond chunk are filled with `True`. + + R: length of hard-copied right contexts; + U: length of full utterance; + S: length of summary vectors; + M: length of memory vectors; + Q: length of attention query; + KV: length of attention key and value. + + The shape of attention mask is (Q, KV). + If self.use_memory is `True`: + query = [right_context, utterance, summary]; + key, value = [memory, right_context, utterance]; + Q = R + U + S, KV = M + R + U. + Otherwise: + query = [right_context, utterance] + key, value = [right_context, utterance] + Q = R + U, KV = R + U. + + Suppose: + c_i: chunk at index i; + r_i: right context that c_i can use; + l_i: left context that c_i can use; + m_i: past memory vectors from previous layer that c_i can use; + s_i: summary vector of c_i. + The target chunk-wise attention is: + c_i, r_i (in query) -> l_i, c_i, r_i, m_i (in key); + s_i (in query) -> l_i, c_i, r_i (in key). + """ + U = utterance.size(0) + num_chunks = math.ceil(U / self.chunk_length) + + right_context_mask = [] + utterance_mask = [] + summary_mask = [] + + if self.use_memory: + num_cols = 9 + # right context and utterance both attend to memory, right context, + # utterance + right_context_utterance_cols_mask = [ + idx in [1, 4, 7] for idx in range(num_cols) + ] + # summary attends to right context, utterance + summary_cols_mask = [idx in [4, 7] for idx in range(num_cols)] + masks_to_concat = [right_context_mask, utterance_mask, summary_mask] + else: + num_cols = 6 + # right context and utterance both attend to right context and + # utterance + right_context_utterance_cols_mask = [ + idx in [1, 4] for idx in range(num_cols) + ] + summary_cols_mask = None + masks_to_concat = [right_context_mask, utterance_mask] + + for chunk_idx in range(num_chunks): + col_widths = self._gen_attention_mask_col_widths(chunk_idx, U) + + right_context_mask_block = _gen_attention_mask_block( + col_widths, + right_context_utterance_cols_mask, + self.right_context_length, + utterance.device, + ) + right_context_mask.append(right_context_mask_block) + + utterance_mask_block = _gen_attention_mask_block( + col_widths, + right_context_utterance_cols_mask, + min( + self.chunk_length, + U - chunk_idx * self.chunk_length, + ), + utterance.device, + ) + utterance_mask.append(utterance_mask_block) + + if summary_cols_mask is not None: + summary_mask_block = _gen_attention_mask_block( + col_widths, summary_cols_mask, 1, utterance.device + ) + summary_mask.append(summary_mask_block) + + attention_mask = ( + 1 - torch.cat([torch.cat(mask) for mask in masks_to_concat]) + ).to(torch.bool) + return attention_mask + + def forward( + self, x: torch.Tensor, lengths: torch.Tensor, warmup: float = 1.0 + ) -> Tuple[torch.Tensor, torch.Tensor]: + """Forward pass for training and validation mode. + + B: batch size; + D: input dimension; + U: length of utterance. + + Args: + x (torch.Tensor): + Utterance frames right-padded with right context frames, + with shape (U + right_context_length, B, D). + lengths (torch.Tensor): + With shape (B,) and i-th element representing number of valid + utterance frames for i-th batch element in x, which contains the + right_context at the end. + + Returns: + A tuple of 2 tensors: + - output utterance frames, with shape (U, B, D). + - output_lengths, with shape (B,), without containing the + right_context at the end. + """ + U = x.size(0) - self.right_context_length + + right_context = self._gen_right_context(x) + utterance = x[:U] + output_lengths = torch.clamp(lengths - self.right_context_length, min=0) + attention_mask = self._gen_attention_mask(utterance) + memory = ( + self.init_memory_op(utterance.permute(1, 2, 0)).permute(2, 0, 1)[ + :-1 + ] + if self.use_memory + else torch.empty(0).to(dtype=x.dtype, device=x.device) + ) + padding_mask = make_pad_mask( + memory.size(0) + right_context.size(0) + output_lengths + ) + + output = utterance + for layer in self.emformer_layers: + output, right_context, memory = layer( + output, + right_context, + memory, + attention_mask, + padding_mask=padding_mask, + warmup=warmup, + ) + + return output, output_lengths + + @torch.jit.export + def infer( + self, + x: torch.Tensor, + lengths: torch.Tensor, + num_processed_frames: torch.Tensor, + states: Tuple[List[List[torch.Tensor]], List[torch.Tensor]], + ) -> Tuple[ + torch.Tensor, + torch.Tensor, + Tuple[List[List[torch.Tensor]], List[torch.Tensor]], + ]: + """Forward pass for streaming inference. + + B: batch size; + D: input dimension; + U: length of utterance. + + Args: + x (torch.Tensor): + Utterance frames right-padded with right context frames, + with shape (U + right_context_length, B, D). + lengths (torch.Tensor): + With shape (B,) and i-th element representing number of valid + utterance frames for i-th batch element in x, which contains the + right_context at the end. + states (List[torch.Tensor, List[List[torch.Tensor]], List[torch.Tensor]]: # noqa + Cached states containing: + - past_lens: number of past frames for each sample in batch + - attn_caches: attention states from preceding chunk's computation, + where each element corresponds to each emformer layer + - conv_caches: left context for causal convolution, where each + element corresponds to each layer. + + Returns: + (Tensor, Tensor, List[List[torch.Tensor]], List[torch.Tensor]): + - output utterance frames, with shape (U, B, D). + - output lengths, with shape (B,), without containing the + right_context at the end. + - updated states from current chunk's computation. + """ + assert num_processed_frames.shape == (x.size(1),) + + attn_caches = states[0] + assert len(attn_caches) == self.num_encoder_layers, len(attn_caches) + for i in range(len(attn_caches)): + assert attn_caches[i][0].shape == ( + self.memory_size, + x.size(1), + self.d_model, + ), attn_caches[i][0].shape + assert attn_caches[i][1].shape == ( + self.left_context_length, + x.size(1), + self.d_model, + ), attn_caches[i][1].shape + assert attn_caches[i][2].shape == ( + self.left_context_length, + x.size(1), + self.d_model, + ), attn_caches[i][2].shape + + conv_caches = states[1] + assert len(conv_caches) == self.num_encoder_layers, len(conv_caches) + for i in range(len(conv_caches)): + assert conv_caches[i].shape == ( + x.size(1), + self.d_model, + self.cnn_module_kernel - 1, + ), conv_caches[i].shape + + right_context = x[-self.right_context_length :] + utterance = x[: -self.right_context_length] + output_lengths = torch.clamp(lengths - self.right_context_length, min=0) + memory = ( + self.init_memory_op(utterance.permute(1, 2, 0)).permute(2, 0, 1) + if self.use_memory + else torch.empty(0).to(dtype=x.dtype, device=x.device) + ) + + # calcualte padding mask to mask out initial zero caches + chunk_mask = make_pad_mask(output_lengths).to(x.device) + memory_mask = ( + torch.div( + num_processed_frames, self.chunk_length, rounding_mode="floor" + ).view(x.size(1), 1) + <= torch.arange(self.memory_size, device=x.device).expand( + x.size(1), self.memory_size + ) + ).flip(1) + left_context_mask = ( + num_processed_frames.view(x.size(1), 1) + <= torch.arange(self.left_context_length, device=x.device).expand( + x.size(1), self.left_context_length + ) + ).flip(1) + right_context_mask = torch.zeros( + x.size(1), + self.right_context_length, + dtype=torch.bool, + device=x.device, + ) + padding_mask = torch.cat( + [memory_mask, right_context_mask, left_context_mask, chunk_mask], + dim=1, + ) + + output = utterance + output_attn_caches: List[List[torch.Tensor]] = [] + output_conv_caches: List[torch.Tensor] = [] + for layer_idx, layer in enumerate(self.emformer_layers): + ( + output, + right_context, + memory, + output_attn_cache, + output_conv_cache, + ) = layer.infer( + output, + right_context, + memory, + padding_mask=padding_mask, + attn_cache=attn_caches[layer_idx], + conv_cache=conv_caches[layer_idx], + ) + output_attn_caches.append(output_attn_cache) + output_conv_caches.append(output_conv_cache) + + output_states: Tuple[List[List[torch.Tensor]], List[torch.Tensor]] = ( + output_attn_caches, + output_conv_caches, + ) + return output, output_lengths, output_states + + +class Emformer(EncoderInterface): + def __init__( + self, + num_features: int, + chunk_length: int, + subsampling_factor: int = 4, + d_model: int = 256, + nhead: int = 4, + dim_feedforward: int = 2048, + num_encoder_layers: int = 12, + dropout: float = 0.1, + layer_dropout: float = 0.075, + cnn_module_kernel: int = 3, + left_context_length: int = 0, + right_context_length: int = 0, + memory_size: int = 0, + tanh_on_mem: bool = False, + negative_inf: float = -1e8, + ): + super().__init__() + + self.subsampling_factor = subsampling_factor + self.right_context_length = right_context_length + if subsampling_factor != 4: + raise NotImplementedError("Support only 'subsampling_factor=4'.") + if chunk_length % subsampling_factor != 0: + raise NotImplementedError( + "chunk_length must be a mutiple of subsampling_factor." + ) + if ( + left_context_length != 0 + and left_context_length % subsampling_factor != 0 + ): + raise NotImplementedError( + "left_context_length must be 0 or a mutiple of subsampling_factor." # noqa + ) + if ( + right_context_length != 0 + and right_context_length % subsampling_factor != 0 + ): + raise NotImplementedError( + "right_context_length must be 0 or a mutiple of subsampling_factor." # noqa + ) + + # self.encoder_embed converts the input of shape (N, T, num_features) + # to the shape (N, T//subsampling_factor, d_model). + # That is, it does two things simultaneously: + # (1) subsampling: T -> T//subsampling_factor + # (2) embedding: num_features -> d_model + self.encoder_embed = Conv2dSubsampling(num_features, d_model) + + self.encoder = EmformerEncoder( + chunk_length=chunk_length // subsampling_factor, + d_model=d_model, + nhead=nhead, + dim_feedforward=dim_feedforward, + num_encoder_layers=num_encoder_layers, + dropout=dropout, + layer_dropout=layer_dropout, + cnn_module_kernel=cnn_module_kernel, + left_context_length=left_context_length // subsampling_factor, + right_context_length=right_context_length // subsampling_factor, + memory_size=memory_size, + tanh_on_mem=tanh_on_mem, + negative_inf=negative_inf, + ) + + def forward( + self, x: torch.Tensor, x_lens: torch.Tensor, warmup: float = 1.0 + ) -> Tuple[torch.Tensor, torch.Tensor]: + """Forward pass for training and non-streaming inference. + + B: batch size; + D: feature dimension; + T: length of utterance. + + Args: + x (torch.Tensor): + Utterance frames right-padded with right context frames, + with shape (B, T, D). + x_lens (torch.Tensor): + With shape (B,) and i-th element representing number of valid + utterance frames for i-th batch element in x, containing the + right_context at the end. + warmup: + A floating point value that gradually increases from 0 throughout + training; when it is >= 1.0 we are "fully warmed up". It is used + to turn modules on sequentially. + + Returns: + (Tensor, Tensor): + - output embedding, with shape (B, T', D), where + T' = ((T - 1) // 2 - 1) // 2 - self.right_context_length // 4. + - output lengths, with shape (B,), without containing the + right_context at the end. + """ + x = self.encoder_embed(x) + x = x.permute(1, 0, 2) # (N, T, C) -> (T, N, C) + + x_lens = (((x_lens - 1) >> 1) - 1) >> 1 + assert x.size(0) == x_lens.max().item() + + output, output_lengths = self.encoder( + x, x_lens, warmup=warmup + ) # (T, N, C) + + output = output.permute(1, 0, 2) # (T, N, C) -> (N, T, C) + + return output, output_lengths + + @torch.jit.export + def infer( + self, + x: torch.Tensor, + x_lens: torch.Tensor, + num_processed_frames: torch.Tensor, + states: Tuple[List[List[torch.Tensor]], List[torch.Tensor]], + ) -> Tuple[ + torch.Tensor, + torch.Tensor, + Tuple[List[List[torch.Tensor]], List[torch.Tensor]], + ]: + """Forward pass for streaming inference. + + B: batch size; + D: feature dimension; + T: length of utterance. + + Args: + x (torch.Tensor): + Utterance frames right-padded with right context frames, + with shape (B, T, D). + lengths (torch.Tensor): + With shape (B,) and i-th element representing number of valid + utterance frames for i-th batch element in x, containing the + right_context at the end. + states (List[torch.Tensor, List[List[torch.Tensor]], List[torch.Tensor]]: # noqa + Cached states containing: + - past_lens: number of past frames for each sample in batch + - attn_caches: attention states from preceding chunk's computation, + where each element corresponds to each emformer layer + - conv_caches: left context for causal convolution, where each + element corresponds to each layer. + Returns: + (Tensor, Tensor): + - output embedding, with shape (B, T', D), where + T' = ((T - 1) // 2 - 1) // 2 - self.right_context_length // 4. + - output lengths, with shape (B,), without containing the + right_context at the end. + - updated states from current chunk's computation. + """ + x = self.encoder_embed(x) + # drop the first and last frames + x = x[:, 1:-1, :] + x = x.permute(1, 0, 2) # (N, T, C) -> (T, N, C) + + # Caution: We assume the subsampling factor is 4! + x_lens = (((x_lens - 1) >> 1) - 1) >> 1 + x_lens -= 2 + assert x.size(0) == x_lens.max().item() + + num_processed_frames = num_processed_frames >> 2 + + output, output_lengths, output_states = self.encoder.infer( + x, x_lens, num_processed_frames, states + ) + + output = output.permute(1, 0, 2) # (T, N, C) -> (N, T, C) + + return output, output_lengths, output_states + + +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-1)//2 - 1)//2, which approximates T' == T//4 + + 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, + ) -> 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-1)//2 - 1)//2, out_channels) + layer1_channels: + Number of channels in layer1 + layer1_channels: + Number of channels in layer2 + """ + assert in_channels >= 7 + super().__init__() + + self.conv = nn.Sequential( + ScaledConv2d( + in_channels=1, + out_channels=layer1_channels, + kernel_size=3, + padding=1, + ), + ActivationBalancer(channel_dim=1), + DoubleSwish(), + ScaledConv2d( + in_channels=layer1_channels, + out_channels=layer2_channels, + kernel_size=3, + stride=2, + ), + ActivationBalancer(channel_dim=1), + DoubleSwish(), + ScaledConv2d( + in_channels=layer2_channels, + out_channels=layer3_channels, + kernel_size=3, + stride=2, + ), + ActivationBalancer(channel_dim=1), + DoubleSwish(), + ) + self.out = ScaledLinear( + layer3_channels * (((in_channels - 1) // 2 - 1) // 2), out_channels + ) + # set learn_eps=False because out_norm is preceded by `out`, and `out` + # itself has learned scale, so the extra degree of freedom is not + # needed. + self.out_norm = BasicNorm(out_channels, learn_eps=False) + # constrain median of output to be close to zero. + self.out_balancer = ActivationBalancer( + channel_dim=-1, min_positive=0.45, max_positive=0.55 + ) + + 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-1)//2 - 1)//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-1)//2 - 1)//2, ((idim-1)//2 - 1)//2) + b, c, t, f = x.size() + x = self.out(x.transpose(1, 2).contiguous().view(b, t, c * f)) + # Now x is of shape (N, ((T-1)//2 - 1))//2, odim) + x = self.out_norm(x) + x = self.out_balancer(x) + return x diff --git a/egs/librispeech/ASR/conv_emformer_transducer_stateless2/encoder_interface.py b/egs/librispeech/ASR/conv_emformer_transducer_stateless2/encoder_interface.py new file mode 120000 index 000000000..ee2f09151 --- /dev/null +++ b/egs/librispeech/ASR/conv_emformer_transducer_stateless2/encoder_interface.py @@ -0,0 +1 @@ +../conv_emformer_transducer_stateless/encoder_interface.py \ No newline at end of file diff --git a/egs/librispeech/ASR/conv_emformer_transducer_stateless2/export.py b/egs/librispeech/ASR/conv_emformer_transducer_stateless2/export.py new file mode 120000 index 000000000..f986b6973 --- /dev/null +++ b/egs/librispeech/ASR/conv_emformer_transducer_stateless2/export.py @@ -0,0 +1 @@ +../conv_emformer_transducer_stateless/export.py \ No newline at end of file diff --git a/egs/librispeech/ASR/conv_emformer_transducer_stateless2/joiner.py b/egs/librispeech/ASR/conv_emformer_transducer_stateless2/joiner.py new file mode 120000 index 000000000..1eb4dcc83 --- /dev/null +++ b/egs/librispeech/ASR/conv_emformer_transducer_stateless2/joiner.py @@ -0,0 +1 @@ +../conv_emformer_transducer_stateless/joiner.py \ No newline at end of file diff --git a/egs/librispeech/ASR/conv_emformer_transducer_stateless2/model.py b/egs/librispeech/ASR/conv_emformer_transducer_stateless2/model.py new file mode 120000 index 000000000..322b694e0 --- /dev/null +++ b/egs/librispeech/ASR/conv_emformer_transducer_stateless2/model.py @@ -0,0 +1 @@ +../conv_emformer_transducer_stateless/model.py \ No newline at end of file diff --git a/egs/librispeech/ASR/conv_emformer_transducer_stateless2/optim.py b/egs/librispeech/ASR/conv_emformer_transducer_stateless2/optim.py new file mode 120000 index 000000000..8f19a99da --- /dev/null +++ b/egs/librispeech/ASR/conv_emformer_transducer_stateless2/optim.py @@ -0,0 +1 @@ +../conv_emformer_transducer_stateless/optim.py \ No newline at end of file diff --git a/egs/librispeech/ASR/conv_emformer_transducer_stateless2/scaling.py b/egs/librispeech/ASR/conv_emformer_transducer_stateless2/scaling.py new file mode 120000 index 000000000..12f22cf9c --- /dev/null +++ b/egs/librispeech/ASR/conv_emformer_transducer_stateless2/scaling.py @@ -0,0 +1 @@ +../conv_emformer_transducer_stateless/scaling.py \ No newline at end of file diff --git a/egs/librispeech/ASR/conv_emformer_transducer_stateless2/stream.py b/egs/librispeech/ASR/conv_emformer_transducer_stateless2/stream.py new file mode 120000 index 000000000..bf9cbbe2e --- /dev/null +++ b/egs/librispeech/ASR/conv_emformer_transducer_stateless2/stream.py @@ -0,0 +1 @@ +../conv_emformer_transducer_stateless/stream.py \ No newline at end of file diff --git a/egs/librispeech/ASR/conv_emformer_transducer_stateless2/streaming_decode.py b/egs/librispeech/ASR/conv_emformer_transducer_stateless2/streaming_decode.py new file mode 120000 index 000000000..f6272202f --- /dev/null +++ b/egs/librispeech/ASR/conv_emformer_transducer_stateless2/streaming_decode.py @@ -0,0 +1 @@ +../conv_emformer_transducer_stateless/streaming_decode.py \ No newline at end of file diff --git a/egs/librispeech/ASR/conv_emformer_transducer_stateless2/test_emformer.py b/egs/librispeech/ASR/conv_emformer_transducer_stateless2/test_emformer.py new file mode 120000 index 000000000..d59fea9ee --- /dev/null +++ b/egs/librispeech/ASR/conv_emformer_transducer_stateless2/test_emformer.py @@ -0,0 +1 @@ +../conv_emformer_transducer_stateless/test_emformer.py \ No newline at end of file diff --git a/egs/librispeech/ASR/conv_emformer_transducer_stateless2/train.py b/egs/librispeech/ASR/conv_emformer_transducer_stateless2/train.py new file mode 120000 index 000000000..597332fdf --- /dev/null +++ b/egs/librispeech/ASR/conv_emformer_transducer_stateless2/train.py @@ -0,0 +1 @@ +../conv_emformer_transducer_stateless/train.py \ No newline at end of file From 193b44ed7aa7649546c49806a4fc4d59b3b0ddfe Mon Sep 17 00:00:00 2001 From: yaozengwei Date: Mon, 13 Jun 2022 22:14:24 +0800 Subject: [PATCH 02/12] use average value as memory vector for each chunk --- .../emformer.py | 208 +++++------------- 1 file changed, 53 insertions(+), 155 deletions(-) diff --git a/egs/librispeech/ASR/conv_emformer_transducer_stateless2/emformer.py b/egs/librispeech/ASR/conv_emformer_transducer_stateless2/emformer.py index 46993da48..c5d862ad8 100644 --- a/egs/librispeech/ASR/conv_emformer_transducer_stateless2/emformer.py +++ b/egs/librispeech/ASR/conv_emformer_transducer_stateless2/emformer.py @@ -537,7 +537,6 @@ class EmformerAttention(nn.Module): self, utterance: torch.Tensor, right_context: torch.Tensor, - summary: torch.Tensor, memory: torch.Tensor, attention_mask: torch.Tensor, padding_mask: Optional[torch.Tensor] = None, @@ -550,10 +549,8 @@ class EmformerAttention(nn.Module): M = memory.size(0) scaling = float(self.head_dim) ** -0.5 - # compute query with [right_context, utterance, summary]. - query = self.emb_to_query( - torch.cat([right_context, utterance, summary]) - ) + # compute query with [right_context, utterance]. + query = self.emb_to_query(torch.cat([right_context, utterance])) # compute key and value with [memory, right_context, utterance]. key, value = self.emb_to_key_value( torch.cat([memory, right_context, utterance]) @@ -593,26 +590,18 @@ class EmformerAttention(nn.Module): ) # apply output projection - outputs = self.out_proj(attention) + output_right_context_utterance = self.out_proj(attention) - output_right_context_utterance = outputs[: R + U] - output_memory = outputs[R + U :] - if self.tanh_on_mem: - output_memory = torch.tanh(output_memory) - else: - output_memory = torch.clamp(output_memory, min=-10, max=10) - - return output_right_context_utterance, output_memory, key, value + return output_right_context_utterance, key, value def forward( self, utterance: torch.Tensor, right_context: torch.Tensor, - summary: torch.Tensor, memory: torch.Tensor, attention_mask: torch.Tensor, padding_mask: Optional[torch.Tensor] = None, - ) -> Tuple[torch.Tensor, torch.Tensor]: + ) -> torch.Tensor: # TODO: Modify docs. """Forward pass for training and validation mode. @@ -620,17 +609,16 @@ class EmformerAttention(nn.Module): D: embedding dimension; R: length of the hard-copied right contexts; U: length of full utterance; - S: length of summary vectors; M: length of memory vectors. It computes a `big` attention matrix on full utterance and then utilizes a pre-computed mask to simulate chunk-wise attention. It concatenates three blocks: hard-copied right contexts, - full utterance, and summary vectors, as a `big` block, + and full utterance, as a `big` block, to compute the query tensor: - query = [right_context, utterance, summary], - with length Q = R + U + S. + query = [right_context, utterance], + with length Q = R + U. It concatenates the three blocks: memory vectors, hard-copied right contexts, and full utterance as another `big` block, to compute the key and value tensors: @@ -644,10 +632,8 @@ class EmformerAttention(nn.Module): r_i: right context that c_i can use; l_i: left context that c_i can use; m_i: past memory vectors from previous layer that c_i can use; - s_i: summary vector of c_i; The target chunk-wise attention is: - c_i, r_i (in query) -> l_i, c_i, r_i, m_i (in key); - s_i (in query) -> l_i, c_i, r_i (in key). + c_i, r_i (in query) -> l_i, c_i, r_i, m_i (in key) Args: utterance (torch.Tensor): @@ -655,9 +641,6 @@ class EmformerAttention(nn.Module): right_context (torch.Tensor): Hard-copied right context frames, with shape (R, B, D), where R = num_chunks * right_context_length - summary (torch.Tensor): - Summary elements with shape (S, B, D), where S = num_chunks. - It is an empty tensor without using memory. memory (torch.Tensor): Memory elements, with shape (M, B, D), where M = num_chunks - 1. It is an empty tensor without using memory. @@ -668,31 +651,22 @@ class EmformerAttention(nn.Module): Padding mask of key tensor, with shape (B, KV). Returns: - A tuple containing 2 tensors: - - output of right context and utterance, with shape (R + U, B, D). - - memory output, with shape (M, B, D), where M = S - 1 or M = 0. + Output of right context and utterance, with shape (R + U, B, D). """ - ( - output_right_context_utterance, - output_memory, - _, - _, - ) = self._forward_impl( + output_right_context_utterance, _, _ = self._forward_impl( utterance, right_context, - summary, memory, attention_mask, padding_mask=padding_mask, ) - return output_right_context_utterance, output_memory[:-1] + return output_right_context_utterance @torch.jit.export def infer( self, utterance: torch.Tensor, right_context: torch.Tensor, - summary: torch.Tensor, memory: torch.Tensor, left_context_key: torch.Tensor, left_context_val: torch.Tensor, @@ -705,13 +679,12 @@ class EmformerAttention(nn.Module): R: length of right context; U: length of utterance, i.e., current chunk; L: length of cached left context; - S: length of summary vectors, S = 1; M: length of cached memory vectors. - It concatenates the right context, utterance (i.e., current chunk) - and summary vector of current chunk, to compute the query tensor: - query = [right_context, utterance, summary], - with length Q = R + U + S. + It concatenates the right context and utterance (i.e., current chunk) + of current chunk, to compute the query tensor: + query = [right_context, utterance], + with length Q = R + U. It concatenates the memory vectors, right context, left context, and current chunk, to compute the key and value tensors: key & value = [memory, right_context, left_context, utterance], @@ -719,8 +692,7 @@ class EmformerAttention(nn.Module): The chunk-wise attention is: chunk, right context (in query) -> - left context, chunk, right context, memory vectors (in key); - summary (in query) -> left context, chunk, right context (in key). + left context, chunk, right context, memory vectors (in key). Args: utterance (torch.Tensor): @@ -728,8 +700,6 @@ class EmformerAttention(nn.Module): right_context (torch.Tensor): Right context frames, with shape (R, B, D), where R = right_context_length. - summary (torch.Tensor): - Summary vector with shape (1, B, D), or empty tensor. memory (torch.Tensor): Memory vectors, with shape (M, B, D), or empty tensor. left_context_key (torch,Tensor): @@ -744,7 +714,6 @@ class EmformerAttention(nn.Module): Returns: A tuple containing 4 tensors: - output of right context and utterance, with shape (R + U, B, D). - - memory output, with shape (1, B, D) or (0, B, D). - attention key of left context and utterance, which would be cached for next computation, with shape (L + U, B, D). - attention value of left context and utterance, which would be @@ -753,28 +722,19 @@ class EmformerAttention(nn.Module): U = utterance.size(0) R = right_context.size(0) L = left_context_key.size(0) - S = summary.size(0) M = memory.size(0) - # TODO: move it outside - # query = [right context, utterance, summary] - Q = R + U + S + # query = [right context, utterance] + Q = R + U # key, value = [memory, right context, left context, uttrance] KV = M + R + L + U attention_mask = torch.zeros(Q, KV).to( dtype=torch.bool, device=utterance.device ) - # disallow attention bettween the summary vector with the memory bank - attention_mask[-1, :M] = True - ( - output_right_context_utterance, - output_memory, - key, - value, - ) = self._forward_impl( + + output_right_context_utterance, key, value = self._forward_impl( utterance, right_context, - summary, memory, attention_mask, padding_mask=padding_mask, @@ -783,7 +743,6 @@ class EmformerAttention(nn.Module): ) return ( output_right_context_utterance, - output_memory, key[M + R :], value[M + R :], ) @@ -938,49 +897,46 @@ class EmformerEncoderLayer(nn.Module): self, right_context_utterance: torch.Tensor, R: int, - memory: torch.Tensor, attention_mask: torch.Tensor, padding_mask: Optional[torch.Tensor] = None, - ) -> Tuple[torch.Tensor, torch.Tensor]: + ) -> torch.Tensor: """Apply attention module in training and validation mode.""" utterance = right_context_utterance[R:] right_context = right_context_utterance[:R] if self.use_memory: - summary = self.summary_op(utterance.permute(1, 2, 0)).permute( + memory = self.summary_op(utterance.permute(1, 2, 0)).permute( 2, 0, 1 - ) + )[:-1, :, :] else: - summary = torch.empty(0).to( + memory = torch.empty(0).to( dtype=utterance.dtype, device=utterance.device ) - output_right_context_utterance, output_memory = self.attention( + output_right_context_utterance = self.attention( utterance=utterance, right_context=right_context, - summary=summary, memory=memory, attention_mask=attention_mask, padding_mask=padding_mask, ) - return output_right_context_utterance, output_memory + return output_right_context_utterance def _apply_attention_module_infer( self, right_context_utterance: torch.Tensor, R: int, - memory: torch.Tensor, attn_cache: List[torch.Tensor], padding_mask: Optional[torch.Tensor] = None, - ) -> Tuple[torch.Tensor, torch.Tensor, List[torch.Tensor]]: + ) -> Tuple[torch.Tensor, List[torch.Tensor]]: """Apply attention module in inference mode. 1) Unpack cached states including: - - memory from previous chunks in the lower layer; + - memory from previous chunks; - attention key and value of left context from preceding chunk's compuation; 2) Apply attention computation; 3) Update cached attention states including: - - output memory of current chunk in the lower layer; + - memory of current chunk; - attention key and value in current chunk's computation, which would be resued in next chunk's computation. """ @@ -992,23 +948,20 @@ class EmformerEncoderLayer(nn.Module): left_context_val = attn_cache[2] if self.use_memory: - summary = self.summary_op(utterance.permute(1, 2, 0)).permute( + memory = self.summary_op(utterance.permute(1, 2, 0)).permute( 2, 0, 1 - ) - summary = summary[:1] + )[:1, :, :] else: - summary = torch.empty(0).to( + memory = torch.empty(0).to( dtype=utterance.dtype, device=utterance.device ) ( output_right_context_utterance, - output_memory, next_key, next_val, ) = self.attention.infer( utterance=utterance, right_context=right_context, - summary=summary, memory=pre_memory, left_context_key=left_context_key, left_context_val=left_context_val, @@ -1017,17 +970,16 @@ class EmformerEncoderLayer(nn.Module): attn_cache = self._update_attn_cache( next_key, next_val, memory, attn_cache ) - return output_right_context_utterance, output_memory, attn_cache + return output_right_context_utterance, attn_cache def forward( self, utterance: torch.Tensor, right_context: torch.Tensor, - memory: torch.Tensor, attention_mask: torch.Tensor, padding_mask: Optional[torch.Tensor] = None, warmup: float = 1.0, - ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]: + ) -> Tuple[torch.Tensor, torch.Tensor]: r"""Forward pass for training and validation mode. B: batch size; @@ -1041,20 +993,16 @@ class EmformerEncoderLayer(nn.Module): Utterance frames, with shape (U, B, D). right_context (torch.Tensor): Right context frames, with shape (R, B, D). - memory (torch.Tensor): - Memory elements, with shape (M, B, D). - It is an empty tensor without using memory. attention_mask (torch.Tensor): Attention mask for underlying attention module, - with shape (Q, KV), where Q = R + U + S, KV = M + R + U. + with shape (Q, KV), where Q = R + U, KV = M + R + U. padding_mask (torch.Tensor): Padding mask of ker tensor, with shape (B, KV). Returns: - A tuple containing 3 tensors: + A tuple containing 2 tensors: - output utterance, with shape (U, B, D). - output right context, with shape (R, B, D). - - output memory, with shape (M, B, D). """ R = right_context.size(0) src = torch.cat([right_context, utterance]) @@ -1076,8 +1024,8 @@ class EmformerEncoderLayer(nn.Module): src = src + self.dropout(self.feed_forward_macaron(src)) # emformer attention module - src_att, output_memory = self._apply_attention_module_forward( - src, R, memory, attention_mask, padding_mask=padding_mask + src_att = self._apply_attention_module_forward( + src, R, attention_mask, padding_mask=padding_mask ) src = src + self.dropout(src_att) @@ -1095,24 +1043,17 @@ class EmformerEncoderLayer(nn.Module): output_utterance = src[R:] output_right_context = src[:R] - return output_utterance, output_right_context, output_memory + return output_utterance, output_right_context @torch.jit.export def infer( self, utterance: torch.Tensor, right_context: torch.Tensor, - memory: torch.Tensor, attn_cache: List[torch.Tensor], conv_cache: torch.Tensor, padding_mask: Optional[torch.Tensor] = None, - ) -> Tuple[ - torch.Tensor, - torch.Tensor, - torch.Tensor, - List[torch.Tensor], - torch.Tensor, - ]: + ) -> Tuple[torch.Tensor, torch.Tensor, List[torch.Tensor], torch.Tensor]: """Forward pass for inference. B: batch size; @@ -1126,8 +1067,6 @@ class EmformerEncoderLayer(nn.Module): Utterance frames, with shape (U, B, D). right_context (torch.Tensor): Right context frames, with shape (R, B, D). - memory (torch.Tensor): - Memory elements, with shape (M, B, D). attn_cache (List[torch.Tensor]): Cached attention tensors generated in preceding computation, including memory, key and value of left context. @@ -1140,9 +1079,8 @@ class EmformerEncoderLayer(nn.Module): (Tensor, Tensor, List[torch.Tensor], Tensor): - output utterance, with shape (U, B, D); - output right_context, with shape (R, B, D); - - output memory, with shape (1, B, D) or (0, B, D). - - output state. - - updated conv_cache. + - output attention cache; + - output convolution cache. """ R = right_context.size(0) src = torch.cat([right_context, utterance]) @@ -1151,12 +1089,8 @@ class EmformerEncoderLayer(nn.Module): src = src + self.dropout(self.feed_forward_macaron(src)) # emformer attention module - ( - src_att, - output_memory, - attn_cache, - ) = self._apply_attention_module_infer( - src, R, memory, attn_cache, padding_mask=padding_mask + src_att, attn_cache = self._apply_attention_module_infer( + src, R, attn_cache, padding_mask=padding_mask ) src = src + self.dropout(src_att) @@ -1174,7 +1108,6 @@ class EmformerEncoderLayer(nn.Module): return ( output_utterance, output_right_context, - output_memory, attn_cache, conv_cache, ) @@ -1253,11 +1186,6 @@ class EmformerEncoder(nn.Module): super().__init__() self.use_memory = memory_size > 0 - self.init_memory_op = nn.AvgPool1d( - kernel_size=chunk_length, - stride=chunk_length, - ceil_mode=True, - ) self.emformer_layers = nn.ModuleList( [ @@ -1358,16 +1286,15 @@ class EmformerEncoder(nn.Module): R: length of hard-copied right contexts; U: length of full utterance; - S: length of summary vectors; M: length of memory vectors; Q: length of attention query; KV: length of attention key and value. The shape of attention mask is (Q, KV). If self.use_memory is `True`: - query = [right_context, utterance, summary]; + query = [right_context, utterance]; key, value = [memory, right_context, utterance]; - Q = R + U + S, KV = M + R + U. + Q = R + U, KV = M + R + U. Otherwise: query = [right_context, utterance] key, value = [right_context, utterance] @@ -1378,17 +1305,14 @@ class EmformerEncoder(nn.Module): r_i: right context that c_i can use; l_i: left context that c_i can use; m_i: past memory vectors from previous layer that c_i can use; - s_i: summary vector of c_i. The target chunk-wise attention is: - c_i, r_i (in query) -> l_i, c_i, r_i, m_i (in key); - s_i (in query) -> l_i, c_i, r_i (in key). + c_i, r_i (in query) -> l_i, c_i, r_i, m_i (in key). """ U = utterance.size(0) num_chunks = math.ceil(U / self.chunk_length) right_context_mask = [] utterance_mask = [] - summary_mask = [] if self.use_memory: num_cols = 9 @@ -1397,9 +1321,6 @@ class EmformerEncoder(nn.Module): right_context_utterance_cols_mask = [ idx in [1, 4, 7] for idx in range(num_cols) ] - # summary attends to right context, utterance - summary_cols_mask = [idx in [4, 7] for idx in range(num_cols)] - masks_to_concat = [right_context_mask, utterance_mask, summary_mask] else: num_cols = 6 # right context and utterance both attend to right context and @@ -1407,8 +1328,7 @@ class EmformerEncoder(nn.Module): right_context_utterance_cols_mask = [ idx in [1, 4] for idx in range(num_cols) ] - summary_cols_mask = None - masks_to_concat = [right_context_mask, utterance_mask] + masks_to_concat = [right_context_mask, utterance_mask] for chunk_idx in range(num_chunks): col_widths = self._gen_attention_mask_col_widths(chunk_idx, U) @@ -1432,12 +1352,6 @@ class EmformerEncoder(nn.Module): ) utterance_mask.append(utterance_mask_block) - if summary_cols_mask is not None: - summary_mask_block = _gen_attention_mask_block( - col_widths, summary_cols_mask, 1, utterance.device - ) - summary_mask.append(summary_mask_block) - attention_mask = ( 1 - torch.cat([torch.cat(mask) for mask in masks_to_concat]) ).to(torch.bool) @@ -1473,23 +1387,15 @@ class EmformerEncoder(nn.Module): utterance = x[:U] output_lengths = torch.clamp(lengths - self.right_context_length, min=0) attention_mask = self._gen_attention_mask(utterance) - memory = ( - self.init_memory_op(utterance.permute(1, 2, 0)).permute(2, 0, 1)[ - :-1 - ] - if self.use_memory - else torch.empty(0).to(dtype=x.dtype, device=x.device) - ) - padding_mask = make_pad_mask( - memory.size(0) + right_context.size(0) + output_lengths - ) + + M = right_context.size(0) // self.chunk_length - 1 + padding_mask = make_pad_mask(M + right_context.size(0) + output_lengths) output = utterance for layer in self.emformer_layers: - output, right_context, memory = layer( + output, right_context = layer( output, right_context, - memory, attention_mask, padding_mask=padding_mask, warmup=warmup, @@ -1525,7 +1431,6 @@ class EmformerEncoder(nn.Module): right_context at the end. states (List[torch.Tensor, List[List[torch.Tensor]], List[torch.Tensor]]: # noqa Cached states containing: - - past_lens: number of past frames for each sample in batch - attn_caches: attention states from preceding chunk's computation, where each element corresponds to each emformer layer - conv_caches: left context for causal convolution, where each @@ -1571,11 +1476,6 @@ class EmformerEncoder(nn.Module): right_context = x[-self.right_context_length :] utterance = x[: -self.right_context_length] output_lengths = torch.clamp(lengths - self.right_context_length, min=0) - memory = ( - self.init_memory_op(utterance.permute(1, 2, 0)).permute(2, 0, 1) - if self.use_memory - else torch.empty(0).to(dtype=x.dtype, device=x.device) - ) # calcualte padding mask to mask out initial zero caches chunk_mask = make_pad_mask(output_lengths).to(x.device) @@ -1611,13 +1511,11 @@ class EmformerEncoder(nn.Module): ( output, right_context, - memory, output_attn_cache, output_conv_cache, ) = layer.infer( output, right_context, - memory, padding_mask=padding_mask, attn_cache=attn_caches[layer_idx], conv_cache=conv_caches[layer_idx], From 5d877ef25e73ad62977fa82be73a29f05afaf59a Mon Sep 17 00:00:00 2001 From: yaozengwei Date: Fri, 17 Jun 2022 18:54:55 +0800 Subject: [PATCH 03/12] change tail padding length from right_context_length to chunk_length --- .../ASR/conv_emformer_transducer_stateless/decode.py | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/egs/librispeech/ASR/conv_emformer_transducer_stateless/decode.py b/egs/librispeech/ASR/conv_emformer_transducer_stateless/decode.py index aadac2ae4..287fb94df 100755 --- a/egs/librispeech/ASR/conv_emformer_transducer_stateless/decode.py +++ b/egs/librispeech/ASR/conv_emformer_transducer_stateless/decode.py @@ -277,10 +277,10 @@ def decode_one_batch( supervisions = batch["supervisions"] feature_lens = supervisions["num_frames"].to(device) - feature_lens += params.right_context_length + feature_lens += params.chunk_length feature = torch.nn.functional.pad( feature, - pad=(0, 0, 0, params.right_context_length), + pad=(0, 0, 0, params.chunk_length), value=LOG_EPS, ) From c27bb1c554cf174e37ebffb069b329d63a9da98f Mon Sep 17 00:00:00 2001 From: yaozengwei Date: Fri, 17 Jun 2022 19:16:22 +0800 Subject: [PATCH 04/12] correct the files, ln -> cp --- .../.emformer.py.swp | Bin 90112 -> 0 bytes .../beam_search.py | 2 +- .../decode.py | 658 +++++++++- .../export.py | 288 ++++- .../streaming_decode.py | 979 +++++++++++++- .../test_emformer.py | 195 ++- .../train.py | 1137 ++++++++++++++++- 7 files changed, 3253 insertions(+), 6 deletions(-) delete mode 100644 egs/librispeech/ASR/conv_emformer_transducer_stateless2/.emformer.py.swp mode change 120000 => 100755 egs/librispeech/ASR/conv_emformer_transducer_stateless2/decode.py mode change 120000 => 100755 egs/librispeech/ASR/conv_emformer_transducer_stateless2/export.py mode change 120000 => 100755 egs/librispeech/ASR/conv_emformer_transducer_stateless2/streaming_decode.py mode change 120000 => 100644 egs/librispeech/ASR/conv_emformer_transducer_stateless2/test_emformer.py mode change 120000 => 100755 egs/librispeech/ASR/conv_emformer_transducer_stateless2/train.py diff --git a/egs/librispeech/ASR/conv_emformer_transducer_stateless2/.emformer.py.swp b/egs/librispeech/ASR/conv_emformer_transducer_stateless2/.emformer.py.swp deleted file mode 100644 index 8da5e4929e5bfee7cb7c05804c8356b0dd018c37..0000000000000000000000000000000000000000 GIT binary patch literal 0 HcmV?d00001 literal 90112 zcmeI537lM2mH*2@0Y%gi7)PAZHwj3!^iIOk&`3xEL_z{t*xFdBu72HJq`Io9s?LJ@ zzTpBmI^)8Oira{z<2EWTpreeasDDQP?%UrT_Z=Pa|2=oVuU=Jm!lKTkKKZ4(-n;9$ z>pAzFb0>$-Uw2M&;^1h3&jSjDTUNhr@(BlA{J`Hl;qbzqa^teJKC>&W%FFJbsdV6k z6{9DdG_YfIpxrE2>jNhY3?Dx{I&$K1qr=AyR93DyX=P<)r7%=U=Vu2l+cmVS(cICR zFIUo`bZ)xQoJ*TS)k->DuGLERL8;QH?<}>dbMv+8^q!$~rZrTnPByEp`82J}4z1d} zX-I`|;mNq(np&u&%~GpfZl|@h)f#o_56v0UPch^L)m9f7E>2_XnHb)8>2Fys+=@ZJyW6_lq;{?`58!Z@&M+ z6xhbUk9oeqd{1WHA7Y-LYre0~yuYt`ZtL~11M~x%{yok6i_G`4GVc#H&n+FlnR$PN zc|U2sug|={pLt$3-{+bPZT|N+&r3Fb<~^bt^|ACh$`H)P|4Z|J!hAnD^ZtS6`6lyy zYv%o7=D98ZYSSTX`iGnMwtTP4yf@V?q$b>fro-6y=JEco{S??wf&CQNPl5du*iV7| z6xdIJ{S??wf&CQNPl3CL0_CYfLC*evB#anLq~-tL3&Z&y&;pl&A@Bp3)DME|!3)6i zz?t9x@GcnOCO8jl1S`Q;VX7|!d%zr616G6MK>~gXQ~pKpA@FwaYH$U3BA5d^z$M^x zunL?8mVv(m3Aiu#IE?@0U^Vy+LWD1YFM>~l4}ojIHgGt24?>7{gKNQyz!va8a5KE< zJeUPDU=?^U_%%H2ufWH^8^Ihn4HUrv;1lr3*ML`p-QYs-aBv&E`G0{|fv19}fGMyB z91IQuzkwnCHTV?}S;0%ec5pcOI6{OMf){|NgH7N(a3)v@q^*9qPmFV7K5NUB*#s4d zxwPI+s`W%^Oq6QXRy(OQn$5H|->6Sj>oZBak+f%1g0$0Sy!dZm(3XT%ah5Gxwpj|9D56mII4P+b33*}(rCRBHxrypYP}o9WGH`0LWnsRS zPOOK#`oDIgNqM)V^;V;~ZCih$-oy_zNdpq=K>Lu=w!&08Ee+==USz0CBQ?|ZLbFcY zTej5_`7u%5mW(G8Qt?u`-L97^^p;c}_bX~>)g1S5n@S71B~|H4G}263R4sz~&Shzz zt_kT%q8l3ZbX)QW-D2v!&svH!Nw$)EMRmrB$*N9WUP%s<)&68sa!Kn8XbnN@qAI3R z1=1xl`AH0&2&vV|d(!6MLVdDYpDIpO=f;Qo3KBq)tdXR8A19liVm32FDqAZGRZ?lS zT=KnWl1*(X1d?JaA`O=3=ToZFa+K0_k^u7QMXsP#19C|$of-vch2$FXNlUUVA(g79 zZrdJY%+D#OYf0TFwy9E=d^hq0?OR$m%ck6mq&{&%)zW$~ENnrj;d?P5f*vt3ylqhW zcG2{sAYy+Qj#z30oAJe*qLNLL?b{_k{f<6FdVSlsGxnxB-i)X>lZ`g>^`+&7n#?m4 zd@^mygsMB+V3I6P>WxdwE_I*s$Tku-GqKd5QCnzN8+GPC8Pk$Rb%3ZOOooxmL=Uk%H?K`eoH+g;mSGbPbxjxyN=P z6(*T(Hs)La)#>QkTlK20p2v%2uuPggMP5BydaYb9tK_~cbHR$)8rzp8SFMxr-b7Z) z^&8JDtvz@1y7e13FrMei?b*RvW2QKg4vaAT3k5|CqrN)G;Hoh_&doQPZNlzIOY`Na z(pZfVF(wIV1R_*JiQN*?H*XkGRWWOiR z3nrf4oYdJYw32!~dTJ=p-ZLNMBdyqT4#SQQvuUpFBx zPe%4N6O**<%5kcCdrOe&bd??CWLfr+je21@)v8tJs%Rin^H#D@hs+dHYNCt0)&W78q-LlE!44jsibWY0w3j(S)d$$mKSL z++#x48*SP86M)K1SA|p9jdHr;h3JD6TgFhb3!;UI2(3(dA)!{OGTJx&$@%a*G*m+Q zlA@GhnTfQl?_^bA=>UU+L++ockm?APTDd8mKy@IN$0g(rM6#OfA5tsodfg?cy{1o2*gWz@Gd0;yj1&4v#;Qzl5t^@xJUJMq%(cmHA z+W)Z>kWMWh!^Z9|s9*fITlU>)ZF;b=&@N4y^{}|xGci$sI*Vxd za(S|$v5PdA8Dvv5Tb@so;)edDSllu&(l<0Tnhdb0$RGU)^Z#6avc=tl$|7YR;YD$z zpS6y~jg9lnLSvy->>J!!O?MS1`x92l3jFYNpHGIT74>48dhFJWhJV|wcv6&XAypZR z#VW6YqVv=z=SU&UK}ZI9u^T3^b?HKy!s@V&Cd(Bew)7{GqLOR~VW^#ZR2@tQDFshg z_a|rdCl~mfmM0s9Z)Y>vuc<4cY(&>~o=1|_)Vuuz&OymadNoK+dvx*&lIFU0%d{#j z9G1cEq$SX9y_=LjBe7PJh+t@8vQ=i0NFh)`f{1_xCYUu0ew>a+1;<$AypTuHDtzGN zWJ%&FBuQ0f%k_F%D?#A#0Vq3Ht(WLStdu*`@!`Q0D|uAj9X?vQvLF%V)6wZ^!_J*y zv7iF;-*RDAS^8zwoJ*%z^n}8U8j*ytNFyMt)I|Em-=4v4obINqP+N8p zQqpPJO~8uKkgVuIO0Be=&6>(o%ChE50adAq#b&4?rI6jqc3y1n*G#zl>5}t>mZB<_ z`#4uc#%_@jhLB7*TuaRk$z(oJVTw%&;g}%;891pIB>`n zd4foJ{#?k=t!7e(4l<1VX^{c0Rd$`@BXOfh(f7JLbgNpH4w=02JAfK8vQ{x>>h;0R zup1(eD%Xl$klb2;TVjnZ!QWQfrBcxkqjIa2!kGu;Pfi_APAEuON%NJVKCVON5Pd%sLitQ1p5j z%sQS3N&ajkrC1Gj0qZixDrs3c5vD+-6Ik_wS&jzz=yj+FIp^L^zMDr|lv0Wkez^L^SphZ(CMy!j8AHgOh}?-BrXmNRTDyc(P?^XoZTY zSZ2Rz-g1oDBpb1!3bUHwE2GdIx<&hJyCsxOhiphxPj@7exJHJ}apdTlB&Zq8b|V=X zTEWs^pJ~t1tzf2+Nl2X0DZ6Wf1Vv_!a~e|%H94K~^M+w0bdGAtW{j0|V(ew~b8~9x zW9i4xOaCz_k{To_vgsd|eZSf^Qz_XKyFD|%SNHie!m-Kbll`M)OP+}OOAwo5PXtV* z>1KH@<*Am`Dq7N5CSzB?`~p)zHFp}(-(a;q%_OP#|4%`KbA`8r|F`<>ABDdczF*G$ zzYaftBlrk-6?hRy!B%i8_!YeTN5GX}4ft<(_s@X;0&fM^fCZ2OIsbnU_!K<+KZ6=L z8vFu&{rlimK=}ANSPM=72Ls{d{||Ttcm_BPJOF$NUj2*U17HDc18czX;C|qf@ai{! zX8}3eKNTzozkz4}G`JCb1pEVd2@rmM5*!DXgWtike+0Y;l)wsbAMg?Q_1A#MgR{Ug z@M-w2VZrM=BHw0*?5AummQXILPoeKU|f_ z&1Orm=y%E0ZenbcPuD%|5q{H{4{w(PO9o-a5Zx}otgr*AQ(t(+h5C+TS}<{QOjJtJ z7p|o6AO5sXsx0S=&}c#QIg-i(dLAr}If=+zE%W@Y#zJk%wF3niu-%IZJBoj@3w2yy zvUPq+lpxdzBgz%-z);Ux2LHM2G3@PGwKnnF#* z%*$yaXAZgv)k>j{!9uw~4_y%$ z_~+*`o1{$f$O{Yhw|u*7RK;Y$y3-elh?qU5-#a*6cOT$N!p0n9XW%Lmq#O{b;ch4T z92uddQ`mL)W?0*yMX&BQD({v-ZDpf=oSI=BwT$oPp%P8O5}`|2+EaH>PRxoEQJ{XW zi?PiQ4gZ&X3O$ zP_#byRb16rt@?LoXA9`uZSjaYh3{pB-1hMA_qt~}qV~sqDW^;LEpR5R@10Sp{|LQ= zjmaIIx9*C`C2a4m^9t*uQCoG5>>t4vK+bx)OBjX;^KJF{*yE~&DB>41M;)$cL24s5 zHMZC)HNN*fwPBl@UKdq@`p;~d`l%Ajhb1%hi92sl+)=c6g?o_7Gb_sZ|5fmw2f?EX z|8MQbKMUV~HMkh81jm3c!R!A$Xo4w_fZxLJ{{Y+y-UH;!zY1o-3E=1O`riXLfsXECfn&iB;rVX@a<>0G@EEWS90f$b|4J|o`oSUKKHz%z{j0!c z@Nn=2_BB;GOVF+KHfBt<{HIPyzgz7(3$`SFrJq9-PQeu~?)OAX+xinqB!ZWSwzA&e z`6XkyQn^C)NuLf#uoT5W^wJ8pfyKRF%IqljWs^pB_Hh}x&- zQexuq{DC=caW4n8WEg>Q@pMj3Rbutov^>?Hki7aSzq%)_w;q#obH4gQjY))}2uDO$ zV@f8rh3KSN%M#F>vJ^7}3d|mSMIy44@O_GQ6{)+pins2MIPHSOk&h3g$^YcJJ5M7ygctcuSUw3-o~+h5_D{o7Dp|5n&J*qI^W(BRGH@`4AP^#J4bXr}`$^Ia6syC*zG6a^EgS_guNah0^+uPi zB&Ahngvrcyp)UPLB5D1RNCl@L&*c+SET&WNOLZ%#fd>^8jsKQYW1a*HzI$~Z&s8n$ zvhChTC z;+1L{pHp1gkzfMQ{QK&H&(CxX?Y`oD!n1nRGIt^t2mL@Xa}a%Sy4HYCk4I|qh({y?!-HOr=^O&ao8E%ZTt1t4 z6>&x#<7=y0r%Hr#OS}%_D*}hRRTQ&pdqgOs+Iaxg`fvso?Pz5lsm1Dh#S^2v7wCS) zzqWaPIolnVEZK=7(m9qK@h+4j>E`OT`-)nuTOik5-2zCj(K|79d~P;hC26Xn?`viU zmRXAns`v1p?*;NC6N+_xT=mVVW@DaR6m2KUar6@X*#0U=T=`_;)9E-`7AsfTb(L#_ zCY_*o2v8BpmddSQ2V<7nZbdV%OKR-s*=WEEQ5P$MXxF1M7!*MY9_J-bva2Z@vLF`; z;d6>gM5VjD(Mxqf;@VGVzKY&;^xQ?oQA9X?vuz`V{Welb4St|ti)304yvatRmQiu* zQo8QPqhh3-e z{||-1RQx9V|8~P{-UPod`u}Ud?eO_mgI9nTg7d-s!B64uzX$#k+z8$ZUI5zQ{@@Gn z^&bb9gNwi;z#ri4KMmdhn&6S34-|pu_WuO_{xe`F*Z@ughk`G_-@g$&8HhdqgTP;b zpTOt;82kv_1l|JP3{C~(;5hJL@H=?^&x3b?SAmPbX+U)O4+ftQz90Mtcq_ODh+JSA zYy>O8k>C*US!4m%0nzb49f+R)UxTkA3wS?xA9x|S3~UE$!4UXsa1S7M|8E6X0nzjS z1w8+2z(wFbK-%Kni}L@rOD<2+>1ko}cBZvGn1U8%uRYhmrcY@xX(RkH3O8ac1dKdfJ;&KV9!+Ep*M3}1R?GLSdNti+fH37-#O1= zxRg+(guToRD#DX#dsm9PfR)2)uQxJ0tdhY%ivvY{(5*D9TYh6CJ6R6Z#b7-d5p|;# zBqLgNjk&a4Rhl+Z68Q;}P;NEq{VGjG*JN17p3}D8(i+Q?Gjd2Ty2k7^n{zF5_%4U% zR#}+8cQbHPUXnjY{MC?hqNij@%VIK9TL@I;mM5#Gk>e!u4RqXO_aeK%og>MvG!fOn z-4w6cSeQXwOBxil0qg*284SFtUR64t%OO=j7uSvBd^QqifRU-MOHn+Wrr{nPbA6(HJ$7`-k6=RJ-isvV+_5dQRK zFtUzfKme@)el)7!NIc+VgI6wt<@#w;D_En^DM}a86GZoIYRVdWOGi9r^Qv>!9%Drh zE`=WIu54SO6%DmeQR;R=ICB_uj#fWtOp1tTFxub7DwWb7#u8alq1Be=LZCSi)F-1G zU>hLwIFHnsAR%!xucQ%mzv-bqnee$_Hf3ZZacC$jKdFJ7bIJIylVXu1FsE;iOO4#9 zQXz|K`{4KKXN&%Ov&L+6nX)b8xRS$*)(cj$#a2{mt#P_nCT&#-^wchxdTBd?7UQ4tB(gf}fbixu8 zb26uw6!=Y1&$RE5kui!q67<m5& zh`lYF1Vz%A;Nrez8{R=`tYd0a&tTsZq;Jxdv;sF6n&#pfNFS1&YRRn#C~3;15YSd6D{p- z#(T3_QpXIEGTn}ZvVAi*(4EjdhfVhk(hs_4FtuohHokA`mbR-rD&yKXUeAQ=4UK%f zAU!5=4hr;Y#r_&+OQ?zxDc^&a+py22Wf7JXR4U0TIb_!ZjM18LE0*g5L@zs=di!{} z)BWWnk++-0oxS9H%!yN3%+t#f(%2`jIj(7~-wLHD11{ilKz;QR3P{|+tz+d&E3AAAem{@vi!;CY}8_%`?) z_yBl5SODjNbHO9QZ;=7K3A_^Q0T+Ni@N0Pg4}lkfM}tk^{@}Cl`C{WQzW?3;#3tZE zFa~}N!F~wd46X!cf^qOOc>C+X<=`pc$zTlJ29JLu_y~AC*a=EtGgu46M&Mv@qwxOV zRp6CCd;nI!3Qz>M(MHlv@3pc4JOhRiTh3?gX3_l2+XR{+%ITLJ%Gd)0CX?v* zpg9}!0x&;?NVX!47Za5h_cOxWMR5(kWOtGh90x{OJ8nd0cqiF}g{*QrRFGiu>6`=P z>FKhXtBrV=?CJ3?3WFwZY!;MGXO&k*xBr(s$U8izf}%4r`vr1?4AFe8z3*|s7aOj3jztBi+^ra1< zlG5SRd!}RK!kIRu&HUD(BQLjeQ3p##cwkQdY)3PDWXhpxCbjb^%CBshbiWo~Fsfjb zpdi8IV{#x{^H;XA{mF5Cojo~<=C%GN@DZ5qPT3r=n+!-3l0}^`1tf@zZd0%aKlp?c zmxxI-%?!^X{>RLhhjqGmx}$j85g+f17rDP}fD$Kuwe*=k2=7gSXK zi_2@K430cq^L9MqP&(5BE-{slE2XKddmCp(@9Rt&3;Lk7#b`b!luOXnwLfV^2C>{HQua*BR0t$%eZMUP2Y_bXSADASHYbWF`bMaP+$?TJ8?ifLh&g%Mh9 zN{-eIlNVSCud*?^QZe~EL|rcQ5{_$2y4KW>5$VTa&Y7)CHS5GfYHpSwJ<0nKM#s&o zQ#9jm%AN1+Y@)qD#rs#-L(yUIi%|Y@6}uYQWTp)ttLs9N@QEWwHPT1iYoba-#Pci6 zUN1=zwK9v5Yk)X2!>DdiYr7ZhN*4*lbuGMVwc}MOSh(WRB6j=O^p`|$K-`{;uJp+Y zZA4CY7YeG~{>fEXGVOe-20<*OM$~J3d8|e#)G4Gc%R!B{{uqxvS#(A_XLg*Pz7VzoPNfF7_EUbHdTo5;Le%hT+ zO&#akv+fL}qY0!`@=;-fktC{H&bNsO>kBFv|1DR}8?~e8g%-y*IEXRzvc`U$f_`nD zjT=CHFoptGH?n5i6pucf%#Du>53)ku!4@9r(!#!Khy}2|SI2a6il&acX88-J47hI+ zsX2l+_SsBw+V75<{5U~S8yP=+El+k<%Za$#Gfsq*h(O#E`N~w=TB@gxwg}5Hoj$Y@ zmqtRVMnH3smE(k_EQ(p$QTMbQXSww4iJLvWqGN-%@s`3Tt&6aT|Bno59AX2U2 z!Q-5$f>TF)9L?gyd#1sWBijmHS99hV*3zFLDoGQ4hn|1pu{fZ3b)X_-&5vj!U$A~yBlk+=BP-c{=!XLL%mwtsxIjEpMyClcK z!IPxxNL?j$d)zx;+ZNQD- z@!%YAIQR*?|G$IhfQ!I|;CS$0@PFa;KLJD@@HCKuDKHLhgU1&i0XKtx0#o1!@JV?5 ztHI;IYH(k0EByVZfcOS@Iw*tFf%yGD2K)>j|EJ)3a1{_ge;0#6a1;=~0EdFx;rV|8 zJ_-H>JP&LJ_W%VT@_?JcE5HH}TY0` zv*!d$`RtxaN#&X4p_c1`p!M(=@jLz8$;OFIqG!{=8BOG5=$?{pnX9Z zt54}ky_dB);Ot&bTOyXOu zQA%ec^^mwaidD;g`=zaeoV;l zSD+cHh~91Q9p%<7kSk!f0J0VGn{#7zFLXFfm+6vfdk|;AEE9el`e{M}M&)oUvCJ1D zUy3(#t~foy8PSy&HM&TQ=-m@yQI4Q+%vc^MvEf!RDVLt#;%EJ7=IzI`VKn)B>ttJv^sr{SoCRqa`w8dmV2-!1Ba$&47wil9 zj$T4fRdgJq%*#v~Ej6xRd8~Kjj*b+_QwApX`eczD?UCQDo z^KP<%g2WC#5xd*jh2x@iyf0*v%B@K$mp=}p;-_0`>sLsbj-s?&PS$5m^w!WlIW-v$ ztR2EpX8LDD(x~oXkBUY`PRU%te#nK<^}-SjE~)Hl`&fuZfAutB<2w z5V-5KCzDQxl3)=BVq&s_e;OE$qON~tB6G7jNE()mAVhqf2Hu3(=Pjm>U5CD$I^0>}{YTPd_slpko{&CJFo%B)}pdpv(L zn4?hc(260f=P#o-ET$QLl8fSWq>*9fz68Dq#75wqKy(Jg zPT(nE2iO2k2KNA8gZG#F0bdB#0?`q;8J_->O1NZ}SfzJW49e5f@!4wcbf(L^CMlSGHa5-2Hth_*c2_6A{fIL9_ z2wnj$1IL0dL0GX9cn^>}0-p&cK^cgg;27`~#(3OE@Qz^`aGX}eEZxq$6hYyns)#m}0uSdMKSZ6Mzhi|T{|n7@ z3N{fRZK-BKxu!LHk<6gBBJ`z?`T9IH1Ed|^ItQC+Jz+={^0{Z>>$r4F>DOyfrMY+z z6-5|9W%{^X`83C8xN~}1PPWT(OZw(kLrbG`Feb8$tG{2#G!T-HY!)U1QBK|-vQkl# zMx1R|uG|%lYYfg)W%G5H;`y2`;CWs=jpBeym|%X0j4-F(9mPtL&I&4x0=HT1o|qwG z!ICQ54`bm&Mk)|iuEQ=quIDP-yAdhNGHCrKYPMdarqcPj`HHH;ex-F~G@;kX+``#p zqGkxnmk}_Wef}6ScX3QB!k);c*BA7>FpRRY_3j4I1W_7(O8@0#zrMJEBZ6J;HozyU zTG)tekXI|3b#y(^R`wak`Y$QV->npAXh^ z?Fn698RmHp6)H66&V~%~4gPdR@a6S?WlAnG;=HJFNlh7?34692zNe}p85PQbjP8S&X7(w_ zDQ7{|p8r(YGKM~`?3w4%6guW220)4?GNH+ZAEGC2tHQ9GmXU{Z2oCGAE5cw>T%*l` z##A+JliJu~os4~@xG>0FZr4qtNvsw}xkyCGj5)}pyGc>Bg}l4?GR$kc*-31-83t5M zlsQsvc5AeedNA!{|LJQ0nQeM@%=IYy24|z>$s*%GTLXw;(AqgDi3CvV+uP;Ebn)2e zX{}6o>e@2o5U{3fBW!rrwq4ALnk+AIPC+|q2Zq+x#|B$v+cC%1N4lA%g6?wSAWaLI z3i^TmKVqVs?BY%oXGa+@j;u&~Yb4SRgH0 z;6Xs{`V(IOR|2^M;0W+V_;@+%7e4@(0`UQGPw*{x_IH8H!CLTO@D+IUFN044@dMBV zj{*mRo8iwzw_on>D}q13kG}&v2ULKZ=|3C@pMDd3`TN0R!4y~r#J2yz;AimT-vz?2 zzZ(e8E`I-?0a9=pSO$Izt^N;?bN#Jg5DWmJq1g7n63ChUCLlcg7&r#pABe60dqE4F z2ObK33-A6;@Dj_z`|jRt&!6Z|x7W;}XRzvJL%9d`5Bctuh7H#%jFxvHY!E*H4I86J zvNp)l6BR=}WyN2Kyrb`HLW{-YeN;9ZK7MYWMnOHd7AN7KXMz`OGP=n`bZAMbYi-}} zZ9IG)SZdiM8NhDECmiSq;)rnuZ zO^tutyX5YSMy;bIBVIR7pKPt@j|Ka8l5Er(bVylLK}4XVE_@Iv) z+D$Zgt`@5r0f!dlXfaVB2_#J|2=u8r%~&mv?It#odbllkoMN8%?%ZZNU>rz>mu)EN z3tUI53j2s8OW*6H}*FoV3KKOdu8C(6hkE2SzLd8MlG zT-q(Anx1R&*uFzzX96$@`e|Ydf>HE^mBqcdMeJ~t7aTg0z?iIFZOiQLM1;&;)$Xbl zeUFOLqSd;$WX^UNfAp=?FOd{W#Gt`dElN)}H!PJJ%QRE>z@#vy{aY}?*i&HWl~jGc zN^{ool4^JlCYu-L#Y<{IdWL);hZcP1%1Hgd74eO}kDP8~?m`!DgMMr0xK=#0HZbOq z7;Ye|9#Ed-`uAk}WOZh{ZZsvUr*-^DP~nVJS@|N>?m^*Dd>A4m4>!Jt@MT~o?D(WH zQn)Cnn(`m>hxJV>w2_2l4rMW2pi^XzoM*cY97eiCn~pm>K3U4Ri;^Q2T}AG-tL*U0 z0<%mYi5G=%aX2M6zp8-vq%o6#>0~};NEGpYrY3h!59<%Jhn_(<3h1c1=3f7@vvO3J z+AXzSVV!yq(RUq4Ay+uyANVr*O<686@@qiH%pV-YxOK%-Ly=keAND17!59Qw3%gqs zjk2v*pYr87Cp@)?Tg7LzH4GG|=7%6R2KX})Y4Ps#@R8PxX=!IHaPsf>W*2KcJlc*P zUDe{c#qwm!#x)yp)+@em_MrM-G3LR4oC6I_@@^a4XgSIT68`@Nc+kIvcNPBsT*Lo= z4PIaL|DOrAfe|naz5$Q_M({+i9UKNe3x6+u|EEC;P6J1Po8j?40X_g;2F?b{!L9K4 z9{^W_2`~))A3XjI;Kkrfa5T6b-u{E&1|aA5&jKfb@51B%3wR6=-+qUJAH&;!0=ynP z56C%v37i8C1h>N9Uk9EJg!h*-`Vnv>_yaustw8knUk}8u-=l!=`$OOW@IC1B3GgBC zL2v_jDX4;3@EEWT35D!`pjw}Q@i@oM!bph+T^Fu4zHo_sG}BaSW9CpV=`hg-Cp6eOu- z9in}`R}=fOY8$tdEqBO}mGQ5wSLlDu_2P_ZNn(DrNpU{R{pLW*!A7Rbn?-ohg2a|j zhqAigFw5&mra(!?vk&6(L~kiv9GnMn@IMQ#T}tgXoq4=_K?o(Em~>hTSBqf85ua1k zbxe;@*e$5Y{O5Knnsbd+bq&iLURoFJlh=#hyh}=)F$Tqs_sUj|u(;=Xx9JQLj(43t zEmy6;$5;>l=j++eX1j~Fa&m`#$SnvIt@<4ywi3eS-0_xi>PilQ`KTicw(#mGRjwZO zSB}PPyrQS!iS4AIlL#;7)Eoo?8mWqF(Zfk&|+gqinrIYNdUTCo4Lct$eMuyFoM+$|lz5HuP$3c;<$snFX>`ws)&m zQnhandsIN&sISv)dWuu_p=^2WzYisoWV9@%U(1aHH45pn=V;NFWmUUKz8GDhNG%*c zhO6o(?~y#T+^l2l8yQ!aSk_c8PwsTb1Y4`a+l70|wXD2PW+cU_=9Je>3SC=2Bk$5H zu$@?taf#&0*rfWWF{i3zW3*$S8k!d^7NHe{@bcNh#yXGDMyeC&+=P;e&Sl6qk#wiU zMKR zUrLi#omjpboe$8?I76W8ggv>hi_F6;k1=(8?be4<4HGsbl^0b+LK4^QwE=&oEIv{s zG>NL@!*r982>-tV_VO3PbHe{u4afOqc>U|avq1xh?SB#cH~ju*fY|*%4O|2U!9Bod z;Q3z%o(9eVqVN9|`2BZ*8BhY^6L38^6%>Kk{QnC6UvvT92VMe1A7DK=8XN?^i7eoH z@NRH5cm=o=Yyu~NzX88SF7PSv5pXrw0z^Mx9guqhzJZM3t>7hKCl~}rf#u*g$Orxd zh!4O$;6m^KAaa9yf&Yt);CVeeQ5xV!a>za;Pxt(uAeU)d=_fV|cd*RG9cd1k=wL1j4tj=_TqMz{; z6URY~Z+u-e>s>kPcgRSL(DT)QIr6K7!npRfCzVnhEZh=gb_j96a51!&t{GNIOO`Xy zM4(Wz`jJEke`o6MLt%1hlrM|sVV@5h^RuC=AMw$d@$R!?rD)3UE?MU+IJ;VD9*Gm3 z|!KF(1r6*-UcCg_oZe%Bo_wIxxphyIc?UH(Y>$sRyS#9OW!v z`5;~13=BDp;%X0`HEiT^-Pchyq1bn&?9BE5TLtEK5QoiYjt zh$VEdW^L}!sOXU;`HCO+o!TPRdp9}j=Ty)RHU9}oVH1DR=V z+`yXFL4Po1*mntZHx280oxmz!W-Z(T~tA!nMKa8Cd3 zp6=OdiB~p@B50MhICidN+2oG#pwbNFXzJ|88jUT%=p`CEz07vj+?jSXzJ=6nI{FF) z<3rq4HCPC;wLGgjHHCJEDtvID3+aO02lh6To;_<3no1ZOXtt!8cO^PEJ1gS@T z`dO*LWayH{I#P$nA`*AGfqYgcrnV-YxzV>3TgIzM816lCiJ2%WUlq|T#yVEGvoXBu%OfFIcxox4KN}vkS zIUz5x3``|H9sif4taGsYQ!boeh=mN`yC~$pCsihDzY(cyG?`TQO6|~xWY8aftPIYZpDVpwms`C}(w|qJ; zEsAMs!{Ef$yD3&N`Hvp0Pt0)$J5?<9^$k|r>0GfdIc1!|FgR@O3A*%ffvX0qsy+g^G9D&G9dM5_ z)@{pWp_s0SvWH#-aUDekcByL-CA3Q@EG<0)!AcR@F#Pokyby{3tP3t(N~QtnM9oV0 zm?xTgN&p&do5B+aPp^(Mwnb0S-3hUt4C6C%tNQL~l<;v;RDt}dRd~p059zwZbdzVC zU!F?2^(Mu1U{9Tw>}j;~)R{}#ly-2O?;S9R2rI2sIz+O|#jc4vOrqg8i+dAVGe-$p zD7wl{L)vtM*&?)x$W9QROX^TD>2(UrF%v!bE;5b?+u5)dy~?sFa`)Bils8N`8US5S zT!o-pAt*CWg(>f@vRt@}REzIQwPLYgKO?q^WIULA+pCwtqWwsJ(Q)13I z3Mv);^E13WsdK}JoR679Fpis|9Zfbe^fHfRg;D#koYR#2tilG0HblM+C1S*QQ=V8) zPzqSE>30ODB5`o3SrlA1sl+SlVLW@s+;#JDyuh`^o8h&0;2-54dg1@ehW{7)|F?lFz@x$a zz<1#3zYT5%ZvxK(GhiJ!3ygsyz}MjI<&MB7fKA|Fa1;FeOTmR;7#su+1iyfnzYaVN ztOLuyZSe0u1Rn#!+lx)WC7=%!!99W41&BSs6M&rcZwAMMZ^6&M2fP~yuRjUOUuov0Dc0mFZTg{2>dg60oV$*fK}i=;OFrB zA`93F=D}G&?gcyp90k4)zyFV57uW(e1Ca-;01p8N1GyLA-@p`*_Gp6(!RbKm2>cr$ za)Ddn`#%gm13m`c2*jVjBf;MSkqO)mmhjQtLF_Eep1Y)2gqD^d^?vKO{N3XF5Rp_% zCA;)ErWE>1*@R}xy3}$?xZHg66UaWz{%$H_>8D*1j;;cah=9nBEwWsJf@n{P3LUGe zngBh9yFlP4iVZoL+3nw*r|Lg3Hr;5nE3lzs+gr2M(c#e%TfgnwbatEDT$}>eHj`|g zPw~!+C&o%+Mh;UO^+yzhtdFoBe#Vlu)6>;Tl_N>3O1XuNXN$dnI3(ii8IF5ZrLlk= zK}D>xuvC%l{rbkPWWD(EtL#w;vNKyZH*f5Z)p57Yfy-7Zvy`^dmV4e@l{n4U)CN@UGf3we;?0W0whRk~N;L^_Pzf3m;^tN(RkMa#9X; zE#Fd7^YvsYo6syber}d|S-7tT?##qKSY0rrtUuJDfUcO=I=59-r84eyP~Qm1!CpaM za$M+QbEm6o_msxTw@S|vDLBNk`59FlF+A9&w*bE6<{f>5SL&Q=CasHp|IOntOW-GZ*F z<)xX;^`D@$RFlxpsNHS0Jqh<-I_G6=VRI@)UAr<$yYO*BIY&!v58|~0<6seY5lh3Y z?5KcXXR0V}+Vq7}8#L4RVuI~5P@yWl{zi|ga0yF4$VM#!}G6cZ)NBE9{~4$|)GFk+=(aR4L3dgi1>$H9$|ffI{~utxashm!@c$*lG~Nuq|90?vFb7rvIrIN6{Jq@!cQt5% zGr*DH2k`hG1y_PPSPsP3|Fz&6@M>@ccq*6$Vgs-O+#h@op8sRuHQ-6$Vj$=IM}VKf z>x-}d8^N{Ujo``PNkH`f*MNrtx&QBT;MJfFHiCx&kq5jURKY3W{@`oy^fv?H{pJ2Y z;r}lKYk~L){4xCeJHhk7G&l~7f*-@jzY&N}z&2O`9szy{?|uuo96SX~fd_#@!6)I@ zUk{!O9uI`Ke-OAg_yaWmC=ee2*MP?XX@?Eq98d&u@88$q<=>af&)cq`B&5cCsyek> zZ)_9B8eD|=^a#=}*g#5d$yP_{r$hx3r4o|_^IozK+HG%TcXw)VgNA3s{O&>0@XzI!eDv5imz!^~$Ek5%@FzHT zXPu32boU!_os_%0J$^A!$h@Y|e&6F>vf~u3Zt=0Ca+eLLa&D&BOc(5rbw7L_MO>l& zWNwPLltu-k+JutX`{#R-Hts6Lo}HmD8PZ1qp$?{_hw!25MT@qEBc2@;^}l?~wk2V*lJA@l@7s$VH8%^i}#3=8kZk z%i*nE@nviec53Zm#+lV@jicxtnYJ3aGb}%%YIxI+yvK1tqT>`jdZBpE(>R(@fhDN< zD}J3LVAP)MVyUH{)#482+c_~G5fW*|4Q6DS(bcWLs?Wwr#tBL6&L{FTR0Z#$7cN2N z6X=9@*ALf0tvg87xZ^=X$8((xoYzG0;Zd%shmD@oQt=U&nUM=NLipX$Nyj^$9 zp&5^UM%&pfka|DS9a>haw(~2s#tanY_EFd-oU-_(J-dGotyuSMSI-BFIb9tOvak z^DVS*Mg+d^*a=wgAn>BuUpuow1H`_60jvcN13!hQeNSAvtk!@#Z3{=?u($KTtv)_$8!-vqY9E4C?VsXXcV zSz46*P|^OflFcTFlX(@q35pq~)~L1Y9H?XK*(H-w=I1^otkT%`n63!>ltesdIyt@^ z>$cP>A=6o9;scP~11 zi#pklkOck73BC(1scnsT!YfwY)l|?cVAeB@t1(2?ZpAezPI51-Lc8CB26iEi=C1iX zo#NJ(Ltts%z)D3?ULCC2fZ4-}n4 zEy){oFO_juA(^1HPxEe5@bv*Y`4iFC3F*_j9fMlFwEo@S2lM z>r?%*ewL>(MbOU*DoF633+ZC90d8VkMm+VdOAH@sGDQey9K3Y|mfOcr=!i=aEyf*D zM5Tnx7YkJCC_(oao_NvjVks;9ern1IN=GA|<6aB3pgc_}$926lei(Q_N5G;1nym9{&%aC!Pc zrcLpJERRiaxP@)5Ixh%J-#G~M8SD6pY#n9Ji3^cy45FaqdI%DiKXfw3@~11GdurWG zN=rQ=3{i6`Z8@@uARV$$Rf(2$BdB)4JcU$k?AgvDn?B!6>MwU>lki~zUtO+M^ld)$ zi-oy5ib3d&vatGH0Qbg*7n`c3R}K)^V5qWrw`8KnIHvgj$H0ev8{Sm-f3!E%=V$Qx zqW>?x0YnCHG8hBDfzSUEcmsF>cr@4m?gtJ9pM>9k8+a1f0oH&+!EfRDKL^APK<)q( zU4Wfn02~g)2jDNkH^ECm9jpK%5BM8U0AGXm|0?)2cn6pVCx9cs$B+StEx?7~0w8jM zhlB6I|GyeM7Bs*(5Iul<1GzKcYETB-!O`Hq;Qem_?*elEKMBO<|5$Jc_!>O_4d89y zt>AL-T(AJ9!G++FAOXKX4)A91PvF_01&#+p;CAT!dmwZdz5maGw}96Hxijz-Aom0u z2}EY_ZSb$)*m_OX3}!n+;l+m+1L2y zuD{M??2|Bk`;5%O$88kKTilNlT$IiTIwNbB7Zy9ZVvu`Q^(U-E{4D;}nO{x;@ms%}>h^r^C=+b>m70`Nh3eFY!UtsD!-6erc+Tuk0Xm-shLJ8V z6Fj#Cj3ry{NC9Myo^LiPX$$j#k~U$DH?6){`Y4ecJIg2a$sC*ec(PLHU$Scs6mCz` z30;C|e7bwS(QNxxGTULTuI=m=O$mH3$=9!+@=HG25Omzwz!%iB;J;95vxlXj#BdCj z5_%-uk*Dje!&KpZ(F4(Dhs5NvU|k`cj7IJnObYD7~7m9`W~oLBniiMcmjoW zRjcrhw8S%2ouY8lQ8BCyKhJfdW=Ri3ma^r_k292`=fM_b#1|%}CJCu}<*`w&Ccbgw z0ufnv9t>0F=8L9dkX_epiE=wR9&DCe9;C%Z)LT7^71b{)X7r}GyU*86WK}4tFxz6h z*sC0_Gce(}zolbj^tR}G$~p(0$}__f%IaU+wux1$Pw&U-L|g?xFC{ZI{`2kTM^ccF zO~rSbB^$6?KM4ppMwjZ9KO$V$s`)0-Gm#ec4{$@`q%xK<*Cxu4fE&@5BZzdBI;zXj z3aV=6zYBt!6`_g}rc@;@n@Tf9eYqu;E7)H#5l+HAQzs-z)u|vQ+qWXM>x|-Z=$k=q zo{p+(6*%-QiE(3YwK8u*eI0W^hAkS_Dp{456Mq{o13~$AFI6o$lJ_`L`|_YstX*|E zyF4!|+M8ALPO|e4Ec?|8(y_pxrX<>i*#*f=|6Vr$KLCncvidNsvMfv${lGOVO*}U_ z@+v!IJ=2Iml%I|gKH_4@%tGaM(WpwJYgQDNqfd2SE%&P{ZyT*ryCDKeLIf_MyKqOU*JTm)GUOEy9L1g3OWdO|dUe!-MZZY*bN=TCg!-|aKlLozQ z%Zwqv&`XcoruhFo@R47JpA`OolV<-5--Ory5AY%I4)Bj)3wQ{)4L<+PU=o}Hz7BtX z9T0y181N-UdfBM^UnABTs(2BhF3@G$UDa6A0_zkz3h3&2WnB=~Rm_s@XW zfX9OE;B@d1a4Y=#4dAulSwL+3w}aEb&*9@g3SJJL48*qo1aKI5Aov1&{A+==;XjPO zZ-HNz^Zsjq`0p3p{vX1p%RPT@1Fr;61sg#Dd>p=9?)1ABJO%6kVzd8nV0HV2rhgAE z2hRm!pPztRp!a`z-rZ3A4t_~<%4fc;zcaT@yDcL#bisy*r%7WbygPS>%dzfbr(hG3?~V(~qPI(&!OU)XBv^DhXCX$krn9!~ zcKP%TPS-FEABmeU{v{>#ndm90msnH6<19|+XY3qQU2-Bh=O(tR^a%?qp63dZjQ!-X z*MmE&>8@f>u%HrAPHOKU`vIHUzE(OAbi1@9&D_ccftOgxzFsx5WzVBTf6f*mX*Rf4 znjN`z0lsXy)@U?iMoBl&RCTAPj45;n#q>gA9?z9S<~Qai)jrR)rK6IVnjmj3*SJ02 zj3Bj@qm$@w;j)bLdv!;EHid2Z6)Migb>};cZA}Kc4mZEfMdOMx?4IZzE6I8ng`Z^8 zYE1T96V+|n9K+Gbwy-KkhXOfUi~n3wXE4t%46B$*Jl{uZ;~1=u zCb2@ziP4?}-CH6_9$hD`g1ogKQ}-^Vi76^^I!iwCCslG=nx&L{_^v5=ICmB-o#i{9 z*0vOJiF0~y&_-Q)Zx~|H-WvvX#S9+$-iyhV4}gfc*;ePCS&6Zp5mYt&*!7EXR?;<) zXs>WnW{@EJXw3Dx!pO@ggqV9PD6?{cyIoM0?~)+zKp^LoAef~#rFi72hnMJSGFU-Y zrO`x?HIGh>Qf^lI(n_2h(+zzQKMTrrY{9*witFm|Kp-+UaWE<=7IdEWfB(YC)AEX8 zh1OcU9i^czL$wp(m>0E2og#2BAgAV;o2{bsl!pb^rKqZL;_3zE%#7;HBPuG3i-e*fV~4S;qSx1kB|$z1-u4`jleQ+I5-H1|NmEkmwSuJDT(i(bj@=d} zTbzQ`gln2ESKyE0J$1xYYKd*0k?2yA>e$~eW-=NX5A`w??P(&~xep)jd=ytYW4_nh zJhZD>Xisy{Py`+stS&}LwLv$01)Ht4#zJ(hajWO|Tw%bHx=D)Qt69ZIr#AaWS9Z|T zX<7XC;r7my+4tb(^xz`*N2&ybw;ofwI2$7wIMv1}NN1Oiih{~o(6!JskdRx3hJuHt zzZFy@6>rNdD_@mjRNKi^BV|XjtHD7gxwf>5ErTS*OJQw~UzoXsPr_Ku>EZ@FNaSVX z>*jYh%Qd@Wxin2A+32uYRfmG+P>Yj2t(UD*$DuK-ouxpZr?^zaQ_favS>AZmZv42$ zb2x$FEc3Hn#bu71A$C$|aG4ObbI(vl<2~8XXgi;b>D1b0v(fb2Y_JiO;)r^6ZfTCr z$u;464}T}6M9+;Hg+l$&XD(54$dNOn{PC!fSD2*{BQNG&(;}P&91#}n4F#dB{oojV zv0gx{L>YChpdVF|diRuSuUbJHUlw-!d^sf~@M>SpL z_OOi7h*{AJmH;YZPV$IRI`Z^m^?#QUykFj;HPxSU|2f&fcQ&gvl@E;Q!w#P@ln-a` z%q!_jnpVn4qa@xf!M{rH*xQPh9BF6qOVSnUr-)JR%|!HX;)wAtDTB{^-2d5{(ES5s@}o<*y?oCR6$Zt^!>PWb;p2LDHdzl8s{ z{{Q|RKL5F39XJb&fWyGg;qQM2z6t&lJQb9|Y2a7z_-_Kw1PefP0fgTdAAq6{ApZZK z09FC9|NmDYcl%uiE(BH|;79QB?+0?f-}6Bgh_Ao%!2Q6X;5+d0-xhuz+yLGNUJ4!u z&IN~n-@(^^3w#n>37!is1{ZU z!gHrW4X0}B`0wmV1iAGmYeaEZ#O;oRaP8Lds{RgUXm=)H>8-!_9|?r?7~s&p)m9@| zi9C#kuRIK;@9MxrNzmll1FdTuG=g)6Y)%i@y8KT#TJY77@gCLpkEJ@z?RZR9Yx?5x zv-NJ&k=`QbVxm1X-2)s>^hT*(E}**acR_eyb} z5-z(%pVvNaUKZV`xB}gw?M7H=Ey(V*SICGdyV{GmmDNxuyROQLy0|IKpV4CP8YAH& zGIR6-%w@hU1bOrg9D#^mGwzc0la>l>E?+!Z_84^$&a<05RbN26i;X9z{LAzew9*~1 zJz1DvHkc&qR5%<0g%&YfBbuO>S5}5I^9`)1#b^fiE9x?vnPzzkjX8vh+(5wfMlH?& zl!C`F@6qm?5!0lgBD(wgNl!`}v2k6S-_6hAVy;Yl^2~+qhV{;D_?K!a^uKAji;RNm zjru^>^)FbDqQ%N6=IRA)apR!e6(!d;NcjJ+!*4zv{xm%Q7yJL4fXDzI4_e?%AaVd} z1MoR`ez601ArKkBC^!V%1Ka}d|IgsX;L%_LJPMow9s+)a4B%7XWk6&Co4|v>Ux7a$ z3-~G!`M|Tm6sUl$U|f6PMHcWPa20q8xELG-z7K!@74TjlwgQ)e zc_1=?+u-qq@BaXJ1GpT>JpldSq2K}FH}L(x2I3>|IpEnK1p`3t1o##Z+ky9hSAa=y z0yrN0HTWtrfv%DJx==Kcf7f%E3t}VLZ}`o?=U!Mk?YxA2{?1p;W=i*iZL*J(ut{E)t}+zV|Evt zbKw}GYh|EG1-Wv}i<9p1YsP{b?E0;)zY-uvM+mNID4*PWm#GO)Y&!aSFb?4y z!YKcArmr`Lr50uwv4}#C3(|5q%K^2xSF$?|zoV4CBUfIwQ#PfOdz8Nxd``HSxK69h z&z5ORDcD@ft(uA63$yWM)5=cSD(k4?B55VuuEs)bN`-a_g{37mv2KMn*n{gA2wcUh zvI{PKNXEN#qQFfPfAd6BPppY!lg7}Mp{Lj0U-;&DCpEvx(VTbyi#e!G{dqR)E~cBS z<5>23Hcwz-7~0~dL}g<<5-izwz_}R1qdsH|Zl<}$3|~&_HnKQhw$o>yAIXlwlfbT4 z@(G{Ne*Zj`3`r~1=ksU^+mvpYt+!NYW6(p#9e0nm+%IZ+krYEv?o1FX#5@!|CJlX?z3QHEz=`!VsHEX4S z1qo(9QE+_L#lgjiUBL#GZ(X->%pgjpn$hd1@nZepI33r3_?M}n(2i!ottm( P!8OldyWum{JDmRyt>Pnj diff --git a/egs/librispeech/ASR/conv_emformer_transducer_stateless2/beam_search.py b/egs/librispeech/ASR/conv_emformer_transducer_stateless2/beam_search.py index 57fd35665..8554e44cc 120000 --- a/egs/librispeech/ASR/conv_emformer_transducer_stateless2/beam_search.py +++ b/egs/librispeech/ASR/conv_emformer_transducer_stateless2/beam_search.py @@ -1 +1 @@ -../conv_emformer_transducer_stateless/beam_search.py \ No newline at end of file +../pruned_transducer_stateless2/beam_search.py \ No newline at end of file diff --git a/egs/librispeech/ASR/conv_emformer_transducer_stateless2/decode.py b/egs/librispeech/ASR/conv_emformer_transducer_stateless2/decode.py deleted file mode 120000 index a9e9e1576..000000000 --- a/egs/librispeech/ASR/conv_emformer_transducer_stateless2/decode.py +++ /dev/null @@ -1 +0,0 @@ -../conv_emformer_transducer_stateless/decode.py \ No newline at end of file diff --git a/egs/librispeech/ASR/conv_emformer_transducer_stateless2/decode.py b/egs/librispeech/ASR/conv_emformer_transducer_stateless2/decode.py new file mode 100755 index 000000000..287fb94df --- /dev/null +++ b/egs/librispeech/ASR/conv_emformer_transducer_stateless2/decode.py @@ -0,0 +1,657 @@ +#!/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: +(1) greedy search +./conv_emformer_transducer_stateless/decode.py \ + --epoch 30 \ + --avg 10 \ + --exp-dir conv_emformer_transducer_stateless/exp \ + --max-duration 300 \ + --num-encoder-layers 12 \ + --chunk-length 32 \ + --cnn-module-kernel 31 \ + --left-context-length 32 \ + --right-context-length 8 \ + --memory-size 32 \ + --decoding-method greedy_search \ + --use-averaged-model True + +(2) modified beam search +./conv_emformer_transducer_stateless/decode.py \ + --epoch 30 \ + --avg 10 \ + --exp-dir conv_emformer_transducer_stateless/exp \ + --max-duration 300 \ + --num-encoder-layers 12 \ + --chunk-length 32 \ + --cnn-module-kernel 31 \ + --left-context-length 32 \ + --right-context-length 8 \ + --memory-size 32 \ + --decoding-method modified_beam_search \ + --use-averaged-model True \ + --beam-size 4 + +(3) fast beam search +./conv_emformer_transducer_stateless/decode.py \ + --epoch 30 \ + --avg 10 \ + --exp-dir conv_emformer_transducer_stateless/exp \ + --max-duration 300 \ + --num-encoder-layers 12 \ + --chunk-length 32 \ + --cnn-module-kernel 31 \ + --left-context-length 32 \ + --right-context-length 8 \ + --memory-size 32 \ + --decoding-method fast_beam_search \ + --use-averaged-model True \ + --beam 4 \ + --max-contexts 4 \ + --max-states 8 +""" + + +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_one_best, + greedy_search, + greedy_search_batch, + modified_beam_search, +) +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 ( + 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( + "--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=10, + 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_stateless4/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( + "--decoding-method", + type=str, + default="greedy_search", + help="""Possible values are: + - greedy_search + - modified_beam_search + - fast_beam_search + """, + ) + + 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=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 --decoding-method is fast_beam_search""", + ) + + parser.add_argument( + "--max-contexts", + type=int, + default=4, + help="""Used only when --decoding-method is + fast_beam_search""", + ) + + parser.add_argument( + "--max-states", + type=int, + default=8, + help="""Used only when --decoding-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 --decoding_method is greedy_search""", + ) + + add_model_arguments(parser) + + return parser + + +def decode_one_batch( + params: AttributeDict, + model: nn.Module, + sp: spm.SentencePieceProcessor, + batch: dict, + 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`. + decoding_graph: + The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used + only when --decoding_method is fast_beam_search. + 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 == 3 + + feature = feature.to(device) + # at entry, feature is (N, T, C) + + supervisions = batch["supervisions"] + feature_lens = supervisions["num_frames"].to(device) + + feature_lens += params.chunk_length + feature = torch.nn.functional.pad( + feature, + pad=(0, 0, 0, params.chunk_length), + value=LOG_EPS, + ) + + 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 == "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 params.decoding_method == "fast_beam_search": + return { + ( + f"beam_{params.beam}_" + f"max_contexts_{params.max_contexts}_" + f"max_states_{params.max_states}" + ): 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, + decoding_graph: Optional[k2.Fsa] = None, +) -> Dict[str, List[Tuple[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. + decoding_graph: + The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used + only when --decoding_method is fast_beam_search. + 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 = 100 + else: + log_interval = 2 + + results = defaultdict(list) + for batch_idx, batch in enumerate(dl): + texts = batch["supervisions"]["text"] + + hyps_dict = decode_one_batch( + params=params, + model=model, + sp=sp, + decoding_graph=decoding_graph, + batch=batch, + ) + + for name, hyps in hyps_dict.items(): + this_batch = [] + assert len(hyps) == len(texts) + for hyp_words, ref_text in zip(hyps, texts): + ref_words = ref_text.split() + this_batch.append((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[List[int], List[int]]]], +): + 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" + ) + 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", + "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 "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}" + 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 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(params) + + 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 + start = params.epoch - params.avg + assert start >= 1 + 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 params.decoding_method == "fast_beam_search": + decoding_graph = k2.trivial_graph(params.vocab_size - 1, device=device) + else: + decoding_graph = None + + num_param = sum([p.numel() for p in model.parameters()]) + logging.info(f"Number of model parameters: {num_param}") + + 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, + sp=sp, + 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/librispeech/ASR/conv_emformer_transducer_stateless2/export.py b/egs/librispeech/ASR/conv_emformer_transducer_stateless2/export.py deleted file mode 120000 index f986b6973..000000000 --- a/egs/librispeech/ASR/conv_emformer_transducer_stateless2/export.py +++ /dev/null @@ -1 +0,0 @@ -../conv_emformer_transducer_stateless/export.py \ No newline at end of file diff --git a/egs/librispeech/ASR/conv_emformer_transducer_stateless2/export.py b/egs/librispeech/ASR/conv_emformer_transducer_stateless2/export.py new file mode 100755 index 000000000..4930881ea --- /dev/null +++ b/egs/librispeech/ASR/conv_emformer_transducer_stateless2/export.py @@ -0,0 +1,287 @@ +#!/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: +./conv_emformer_transducer_stateless/export.py \ + --exp-dir ./conv_emformer_transducer_stateless/exp \ + --bpe-model data/lang_bpe_500/bpe.model \ + --epoch 30 \ + --avg 10 \ + --use-averaged-model=True \ + --num-encoder-layers 12 \ + --chunk-length 32 \ + --cnn-module-kernel 31 \ + --left-context-length 32 \ + --right-context-length 8 \ + --memory-size 32 \ + --jit False + +It will generate a file exp_dir/pretrained.pt + +To use the generated file with `conv_emformer_transducer_stateless/decode.py`, +you can do: + + cd /path/to/exp_dir + ln -s pretrained.pt epoch-9999.pt + + cd /path/to/egs/librispeech/ASR + ./conv_emformer_transducer_stateless/decode.py \ + --exp-dir ./conv_emformer_transducer_stateless/exp \ + --epoch 9999 \ + --avg 1 \ + --max-duration 100 \ + --bpe-model data/lang_bpe_500/bpe.model \ + --use-averaged-model=False \ + --num-encoder-layers 12 \ + --chunk-length 32 \ + --cnn-module-kernel 31 \ + --left-context-length 32 \ + --right-context-length 8 \ + --memory-size 32 +""" + +import argparse +import logging +from pathlib import Path + +import sentencepiece as spm +import torch +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=28, + help="""It specifies the checkpoint to use for averaging. + Note: Epoch counts from 0. + 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( + "--exp-dir", + type=str, + default="pruned_transducer_stateless2/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. + """, + ) + + 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( + "--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. ", + ) + + add_model_arguments(parser) + + return parser + + +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") + + 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) + + 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.eval() + + if params.jit: + # 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 torch.jit.script") + # 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/librispeech/ASR/conv_emformer_transducer_stateless2/streaming_decode.py b/egs/librispeech/ASR/conv_emformer_transducer_stateless2/streaming_decode.py deleted file mode 120000 index f6272202f..000000000 --- a/egs/librispeech/ASR/conv_emformer_transducer_stateless2/streaming_decode.py +++ /dev/null @@ -1 +0,0 @@ -../conv_emformer_transducer_stateless/streaming_decode.py \ No newline at end of file diff --git a/egs/librispeech/ASR/conv_emformer_transducer_stateless2/streaming_decode.py b/egs/librispeech/ASR/conv_emformer_transducer_stateless2/streaming_decode.py new file mode 100755 index 000000000..4fac405b0 --- /dev/null +++ b/egs/librispeech/ASR/conv_emformer_transducer_stateless2/streaming_decode.py @@ -0,0 +1,978 @@ +#!/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: +(1) greedy search +./conv_emformer_transducer_stateless/streaming_decode.py \ + --epoch 30 \ + --avg 10 \ + --exp-dir conv_emformer_transducer_stateless/exp \ + --num-decode-streams 2000 \ + --num-encoder-layers 12 \ + --chunk-length 32 \ + --cnn-module-kernel 31 \ + --left-context-length 32 \ + --right-context-length 8 \ + --memory-size 32 \ + --decoding-method greedy_search \ + --use-averaged-model True + +(2) modified beam search +./conv_emformer_transducer_stateless/streaming_decode.py \ + --epoch 30 \ + --avg 10 \ + --exp-dir conv_emformer_transducer_stateless/exp \ + --num-decode-streams 2000 \ + --num-encoder-layers 12 \ + --chunk-length 32 \ + --cnn-module-kernel 31 \ + --left-context-length 32 \ + --right-context-length 8 \ + --memory-size 32 \ + --decoding-method modified_beam_search \ + --use-averaged-model True \ + --beam-size 4 + +(3) fast beam search +./conv_emformer_transducer_stateless/streaming_decode.py \ + --epoch 30 \ + --avg 10 \ + --exp-dir conv_emformer_transducer_stateless/exp \ + --num-decode-streams 2000 \ + --num-encoder-layers 12 \ + --chunk-length 32 \ + --cnn-module-kernel 31 \ + --left-context-length 32 \ + --right-context-length 8 \ + --memory-size 32 \ + --decoding-method fast_beam_search \ + --use-averaged-model True \ + --beam 4 \ + --max-contexts 4 \ + --max-states 8 +""" +import argparse +import logging +import warnings +from pathlib import Path +from typing import Dict, List, Optional, Tuple + +import k2 +from lhotse import CutSet +import numpy as np +import sentencepiece as spm +import torch +import torch.nn as nn +from asr_datamodule import LibriSpeechAsrDataModule +from beam_search import Hypothesis, HypothesisList, get_hyps_shape +from emformer import LOG_EPSILON, stack_states, unstack_states +from kaldifeat import Fbank, FbankOptions +from stream import Stream +from torch.nn.utils.rnn import pad_sequence +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 one_best_decoding +from icefall.utils import ( + AttributeDict, + get_texts, + setup_logger, + store_transcripts, + str2bool, + write_error_stats, +) + + +def get_parser(): + parser = argparse.ArgumentParser( + formatter_class=argparse.ArgumentDefaultsHelpFormatter + ) + + parser.add_argument( + "--epoch", + type=int, + default=28, + help="It specifies the checkpoint to use for decoding." + "Note: Epoch counts from 0.", + ) + + 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'. ", + ) + + parser.add_argument( + "--use-averaged-model", + type=str2bool, + default=False, + 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="transducer_emformer/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( + "--decoding-method", + type=str, + default="greedy_search", + help="""Possible values are: + - greedy_search + - modified_beam_search + - fast_beam_search + """, + ) + + parser.add_argument( + "--beam-size", + type=int, + default=4, + help="""An interger 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=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 --decoding-method is fast_beam_search""", + ) + + parser.add_argument( + "--max-contexts", + type=int, + default=4, + help="""Used only when --decoding-method is + fast_beam_search""", + ) + + parser.add_argument( + "--max-states", + type=int, + default=8, + help="""Used only when --decoding-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 --decoding_method is greedy_search""", + ) + + parser.add_argument( + "--sampling-rate", + type=float, + default=16000, + help="Sample rate of the audio", + ) + + parser.add_argument( + "--num-decode-streams", + type=int, + default=2000, + help="The number of streams that can be decoded parallel", + ) + + add_model_arguments(parser) + + return parser + + +def greedy_search( + model: nn.Module, + encoder_out: torch.Tensor, + streams: List[Stream], +) -> None: + """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. + streams: + A list of Stream objects. + """ + assert len(streams) == encoder_out.size(0) + assert encoder_out.ndim == 3 + + blank_id = model.decoder.blank_id + context_size = model.decoder.context_size + device = next(model.parameters()).device + T = encoder_out.size(1) + + encoder_out = model.joiner.encoder_proj(encoder_out) + + decoder_input = torch.tensor( + [stream.hyp[-context_size:] for stream in streams], + device=device, + dtype=torch.int64, + ) + # decoder_out is of shape (batch_size, 1, decoder_out_dim) + decoder_out = model.decoder(decoder_input, need_pad=False) + decoder_out = model.joiner.decoder_proj(decoder_out) + + for t in range(T): + # current_encoder_out's shape: (batch_size, 1, encoder_out_dim) + current_encoder_out = encoder_out[:, t : t + 1, :] # noqa + + logits = model.joiner( + current_encoder_out.unsqueeze(2), + decoder_out.unsqueeze(1), + project_input=False, + ) + # logits'shape (batch_size, vocab_size) + logits = logits.squeeze(1).squeeze(1) + + assert logits.ndim == 2, logits.shape + y = logits.argmax(dim=1).tolist() + emitted = False + for i, v in enumerate(y): + if v != blank_id: + streams[i].hyp.append(v) + emitted = True + if emitted: + # update decoder output + decoder_input = torch.tensor( + [stream.hyp[-context_size:] for stream in streams], + device=device, + dtype=torch.int64, + ) + decoder_out = model.decoder( + decoder_input, + need_pad=False, + ) + decoder_out = model.joiner.decoder_proj(decoder_out) + + +def modified_beam_search( + model: nn.Module, + encoder_out: torch.Tensor, + streams: List[Stream], + beam: int = 4, +): + """Beam search in batch mode with --max-sym-per-frame=1 being hardcoded. + + Args: + model: + The RNN-T model. + encoder_out: + A 3-D tensor of shape (N, T, encoder_out_dim) containing the output of + the encoder model. + streams: + A list of stream objects. + beam: + Number of active paths during the beam search. + """ + assert encoder_out.ndim == 3, encoder_out.shape + assert len(streams) == encoder_out.size(0) + + blank_id = model.decoder.blank_id + context_size = model.decoder.context_size + device = next(model.parameters()).device + batch_size = len(streams) + T = encoder_out.size(1) + + B = [stream.hyps for stream in streams] + + encoder_out = model.joiner.encoder_proj(encoder_out) + + for t in range(T): + current_encoder_out = encoder_out[:, t].unsqueeze(1).unsqueeze(1) + # current_encoder_out's shape: (batch_size, 1, 1, encoder_out_dim) + + 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.stack( + [hyp.log_prob.reshape(1) for hyps in A for hyp in hyps], dim=0 + ) # (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, decoder_output_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, 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) + + 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 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 != blank_id: + new_ys.append(new_token) + + new_log_prob = topk_log_probs[k] + new_hyp = Hypothesis(ys=new_ys, log_prob=new_log_prob) + B[i].add(new_hyp) + + for i in range(batch_size): + streams[i].hyps = B[i] + + +def fast_beam_search_one_best( + model: nn.Module, + streams: List[Stream], + encoder_out: torch.Tensor, + processed_lens: torch.Tensor, + beam: float, + max_states: int, + max_contexts: int, +) -> None: + """It limits the maximum number of symbols per frame to 1. + + A lattice is first obtained using modified beam search, and then + the shortest path within the lattice is used as the final output. + + Args: + model: + An instance of `Transducer`. + streams: + A list of stream objects. + encoder_out: + A tensor of shape (N, T, C) from the encoder. + processed_lens: + A tensor of shape (N,) containing the number of processed 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. + """ + assert encoder_out.ndim == 3 + + context_size = model.decoder.context_size + vocab_size = model.decoder.vocab_size + + B, T, C = encoder_out.shape + assert B == len(streams) + + 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(streams[i].rnnt_decoding_stream) + 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.log_softmax(dim=-1) + decoding_streams.advance(log_probs) + + decoding_streams.terminate_and_flush_to_streams() + + lattice = decoding_streams.format_output(processed_lens.tolist()) + + best_path = one_best_decoding(lattice) + hyps = get_texts(best_path) + + for i in range(B): + streams[i].hyp = hyps[i] + + +def decode_one_chunk( + model: nn.Module, + streams: List[Stream], + params: AttributeDict, + decoding_graph: Optional[k2.Fsa] = None, +) -> List[int]: + """ + Args: + model: + The Transducer model. + streams: + A list of Stream objects. + params: + It is returned by :func:`get_params`. + decoding_graph: + The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used + only when --decoding_method is fast_beam_search. + + Returns: + A list of indexes indicating the finished streams. + """ + device = next(model.parameters()).device + + feature_list = [] + feature_len_list = [] + state_list = [] + num_processed_frames_list = [] + + for stream in streams: + # We should first get `stream.num_processed_frames` + # before calling `stream.get_feature_chunk()` + # since `stream.num_processed_frames` would be updated + num_processed_frames_list.append(stream.num_processed_frames) + feature = stream.get_feature_chunk() + feature_len = feature.size(0) + feature_list.append(feature) + feature_len_list.append(feature_len) + state_list.append(stream.states) + + features = pad_sequence( + feature_list, batch_first=True, padding_value=LOG_EPSILON + ).to(device) + feature_lens = torch.tensor(feature_len_list, device=device) + num_processed_frames = torch.tensor( + num_processed_frames_list, device=device + ) + + # Make sure it has at least 1 frame after subsampling, first-and-last-frame cutting, and right context cutting # noqa + tail_length = ( + 3 * params.subsampling_factor + params.right_context_length + 3 + ) + if features.size(1) < tail_length: + pad_length = tail_length - features.size(1) + feature_lens += pad_length + features = torch.nn.functional.pad( + features, + (0, 0, 0, pad_length), + mode="constant", + value=LOG_EPSILON, + ) + + # Stack states of all streams + states = stack_states(state_list) + + encoder_out, encoder_out_lens, states = model.encoder.infer( + x=features, + x_lens=feature_lens, + states=states, + num_processed_frames=num_processed_frames, + ) + + if params.decoding_method == "greedy_search": + greedy_search( + model=model, + streams=streams, + encoder_out=encoder_out, + ) + elif params.decoding_method == "modified_beam_search": + modified_beam_search( + model=model, + streams=streams, + encoder_out=encoder_out, + beam=params.beam_size, + ) + elif params.decoding_method == "fast_beam_search": + # feature_len is needed to get partial results. + # The rnnt_decoding_stream for fast_beam_search. + fast_beam_search_one_best( + model=model, + streams=streams, + encoder_out=encoder_out, + processed_lens=(num_processed_frames >> 2) + encoder_out_lens, + beam=params.beam, + max_contexts=params.max_contexts, + max_states=params.max_states, + ) + else: + raise ValueError( + f"Unsupported decoding method: {params.decoding_method}" + ) + + # Update cached states of each stream + state_list = unstack_states(states) + for i, s in enumerate(state_list): + streams[i].states = s + + finished_streams = [i for i, stream in enumerate(streams) if stream.done] + return finished_streams + + +def create_streaming_feature_extractor() -> Fbank: + """Create a CPU streaming feature extractor. + + At present, we assume it returns a fbank feature extractor with + fixed options. In the future, we will support passing in the options + from outside. + + Returns: + Return a CPU streaming feature extractor. + """ + opts = FbankOptions() + opts.device = "cpu" + opts.frame_opts.dither = 0 + opts.frame_opts.snip_edges = False + opts.frame_opts.samp_freq = 16000 + opts.mel_opts.num_bins = 80 + return Fbank(opts) + + +def decode_dataset( + cuts: CutSet, + model: nn.Module, + params: AttributeDict, + sp: spm.SentencePieceProcessor, + decoding_graph: Optional[k2.Fsa] = None, +): + """Decode dataset. + + Args: + cuts: + Lhotse Cutset containing the dataset to decode. + params: + It is returned by :func:`get_params`. + model: + The Transducer model. + sp: + The BPE model. + decoding_graph: + The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used + only when --decoding_method is fast_beam_search. + + 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. + """ + device = next(model.parameters()).device + + log_interval = 300 + + fbank = create_streaming_feature_extractor() + + decode_results = [] + streams = [] + for num, cut in enumerate(cuts): + # Each utterance has a Stream. + stream = Stream( + params=params, + decoding_graph=decoding_graph, + device=device, + LOG_EPS=LOG_EPSILON, + ) + + audio: np.ndarray = cut.load_audio() + # audio.shape: (1, num_samples) + assert len(audio.shape) == 2 + assert audio.shape[0] == 1, "Should be single channel" + assert audio.dtype == np.float32, audio.dtype + # The trained model is using normalized samples + assert audio.max() <= 1, "Should be normalized to [-1, 1])" + + samples = torch.from_numpy(audio).squeeze(0) + feature = fbank(samples) + stream.set_feature(feature) + stream.set_ground_truth(cut.supervisions[0].text) + + streams.append(stream) + + while len(streams) >= params.num_decode_streams: + finished_streams = decode_one_chunk( + model=model, + streams=streams, + params=params, + decoding_graph=decoding_graph, + ) + + for i in sorted(finished_streams, reverse=True): + decode_results.append( + ( + streams[i].ground_truth.split(), + sp.decode(streams[i].decoding_result()).split(), + ) + ) + del streams[i] + + if num % log_interval == 0: + logging.info(f"Cuts processed until now is {num}.") + + while len(streams) > 0: + finished_streams = decode_one_chunk( + model=model, + streams=streams, + params=params, + decoding_graph=decoding_graph, + ) + + for i in sorted(finished_streams, reverse=True): + decode_results.append( + ( + streams[i].ground_truth.split(), + sp.decode(streams[i].decoding_result()).split(), + ) + ) + del streams[i] + + if params.decoding_method == "greedy_search": + key = "greedy_search" + elif params.decoding_method == "fast_beam_search": + key = ( + f"beam_{params.beam}_" + f"max_contexts_{params.max_contexts}_" + f"max_states_{params.max_states}" + ) + else: + key = f"beam_size_{params.beam_size}" + + return {key: decode_results} + + +def save_results( + params: AttributeDict, + test_set_name: str, + results_dict: Dict[str, List[Tuple[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" + ) + store_transcripts(filename=recog_path, texts=sorted(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", + "fast_beam_search", + "modified_beam_search", + ) + params.res_dir = params.exp_dir / "streaming" / 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}" + + # for streaming + params.suffix += f"-streaming-chunk-length-{params.chunk_length}" + params.suffix += f"-left-context-length-{params.left_context_length}" + params.suffix += f"-right-context-length-{params.right_context_length}" + params.suffix += f"-memory-size-{params.memory_size}" + + 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}" + 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-streaming-decode") + 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() + + params.device = device + + logging.info(params) + + 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.eval() + + if params.decoding_method == "fast_beam_search": + decoding_graph = k2.trivial_graph(params.vocab_size - 1, device=device) + else: + decoding_graph = None + + num_param = sum([p.numel() for p in model.parameters()]) + logging.info(f"Number of model parameters: {num_param}") + + librispeech = LibriSpeechAsrDataModule(args) + + test_clean_cuts = librispeech.test_clean_cuts() + test_other_cuts = librispeech.test_other_cuts() + + test_sets = ["test-clean", "test-other"] + test_cuts = [test_clean_cuts, test_other_cuts] + + for test_set, test_cut in zip(test_sets, test_cuts): + results_dict = decode_dataset( + cuts=test_cut, + model=model, + params=params, + sp=sp, + decoding_graph=decoding_graph, + ) + + save_results( + params=params, + test_set_name=test_set, + results_dict=results_dict, + ) + + logging.info("Done!") + + +if __name__ == "__main__": + torch.manual_seed(20220410) + main() diff --git a/egs/librispeech/ASR/conv_emformer_transducer_stateless2/test_emformer.py b/egs/librispeech/ASR/conv_emformer_transducer_stateless2/test_emformer.py deleted file mode 120000 index d59fea9ee..000000000 --- a/egs/librispeech/ASR/conv_emformer_transducer_stateless2/test_emformer.py +++ /dev/null @@ -1 +0,0 @@ -../conv_emformer_transducer_stateless/test_emformer.py \ No newline at end of file diff --git a/egs/librispeech/ASR/conv_emformer_transducer_stateless2/test_emformer.py b/egs/librispeech/ASR/conv_emformer_transducer_stateless2/test_emformer.py new file mode 100644 index 000000000..8cde6205b --- /dev/null +++ b/egs/librispeech/ASR/conv_emformer_transducer_stateless2/test_emformer.py @@ -0,0 +1,194 @@ +#!/usr/bin/env python3 +# +# Copyright 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. + + +import torch +from emformer import ConvolutionModule, Emformer, stack_states, unstack_states + + +def test_convolution_module_forward(): + B, D = 2, 256 + chunk_length = 4 + right_context_length = 2 + num_chunks = 3 + U = num_chunks * chunk_length + R = num_chunks * right_context_length + kernel_size = 31 + conv_module = ConvolutionModule( + chunk_length, + right_context_length, + D, + kernel_size, + ) + + utterance = torch.randn(U, B, D) + right_context = torch.randn(R, B, D) + + utterance, right_context = conv_module(utterance, right_context) + assert utterance.shape == (U, B, D), utterance.shape + assert right_context.shape == (R, B, D), right_context.shape + + +def test_convolution_module_infer(): + from emformer import ConvolutionModule + + B, D = 2, 256 + chunk_length = 4 + right_context_length = 2 + num_chunks = 1 + U = num_chunks * chunk_length + R = num_chunks * right_context_length + kernel_size = 31 + conv_module = ConvolutionModule( + chunk_length, + right_context_length, + D, + kernel_size, + ) + + utterance = torch.randn(U, B, D) + right_context = torch.randn(R, B, D) + cache = torch.randn(B, D, kernel_size - 1) + + utterance, right_context, new_cache = conv_module.infer( + utterance, right_context, cache + ) + assert utterance.shape == (U, B, D), utterance.shape + assert right_context.shape == (R, B, D), right_context.shape + assert new_cache.shape == (B, D, kernel_size - 1), new_cache.shape + + +def test_state_stack_unstack(): + num_features = 80 + chunk_length = 32 + encoder_dim = 512 + num_encoder_layers = 2 + kernel_size = 31 + left_context_length = 32 + right_context_length = 8 + memory_size = 32 + + model = Emformer( + num_features=num_features, + chunk_length=chunk_length, + subsampling_factor=4, + d_model=encoder_dim, + num_encoder_layers=num_encoder_layers, + cnn_module_kernel=kernel_size, + left_context_length=left_context_length, + right_context_length=right_context_length, + memory_size=memory_size, + ) + + for batch_size in [1, 2]: + attn_caches = [ + [ + torch.zeros(memory_size, batch_size, encoder_dim), + torch.zeros(left_context_length // 4, batch_size, encoder_dim), + torch.zeros( + left_context_length // 4, + batch_size, + encoder_dim, + ), + ] + for _ in range(num_encoder_layers) + ] + conv_caches = [ + torch.zeros(batch_size, encoder_dim, kernel_size - 1) + for _ in range(num_encoder_layers) + ] + states = [attn_caches, conv_caches] + x = torch.randn(batch_size, 23, num_features) + x_lens = torch.full((batch_size,), 23) + num_processed_frames = torch.full((batch_size,), 0) + y, y_lens, states = model.infer( + x, x_lens, num_processed_frames=num_processed_frames, states=states + ) + + state_list = unstack_states(states) + states2 = stack_states(state_list) + + for ss, ss2 in zip(states[0], states2[0]): + for s, s2 in zip(ss, ss2): + assert torch.allclose(s, s2), f"{s.sum()}, {s2.sum()}" + + for s, s2 in zip(states[1], states2[1]): + assert torch.allclose(s, s2), f"{s.sum()}, {s2.sum()}" + + +def test_torchscript_consistency_infer(): + r"""Verify that scripting Emformer does not change the behavior of method `infer`.""" # noqa + num_features = 80 + chunk_length = 32 + encoder_dim = 512 + num_encoder_layers = 2 + kernel_size = 31 + left_context_length = 32 + right_context_length = 8 + memory_size = 32 + batch_size = 2 + + model = Emformer( + num_features=num_features, + chunk_length=chunk_length, + subsampling_factor=4, + d_model=encoder_dim, + num_encoder_layers=num_encoder_layers, + cnn_module_kernel=kernel_size, + left_context_length=left_context_length, + right_context_length=right_context_length, + memory_size=memory_size, + ).eval() + attn_caches = [ + [ + torch.zeros(memory_size, batch_size, encoder_dim), + torch.zeros(left_context_length // 4, batch_size, encoder_dim), + torch.zeros( + left_context_length // 4, + batch_size, + encoder_dim, + ), + ] + for _ in range(num_encoder_layers) + ] + conv_caches = [ + torch.zeros(batch_size, encoder_dim, kernel_size - 1) + for _ in range(num_encoder_layers) + ] + states = [attn_caches, conv_caches] + x = torch.randn(batch_size, 23, num_features) + x_lens = torch.full((batch_size,), 23) + num_processed_frames = torch.full((batch_size,), 0) + y, y_lens, out_states = model.infer( + x, x_lens, num_processed_frames=num_processed_frames, states=states + ) + + sc_model = torch.jit.script(model).eval() + sc_y, sc_y_lens, sc_out_states = sc_model.infer( + x, x_lens, num_processed_frames=num_processed_frames, states=states + ) + + assert torch.allclose(y, sc_y) + + +if __name__ == "__main__": + test_convolution_module_forward() + test_convolution_module_infer() + test_state_stack_unstack() + test_torchscript_consistency_infer() diff --git a/egs/librispeech/ASR/conv_emformer_transducer_stateless2/train.py b/egs/librispeech/ASR/conv_emformer_transducer_stateless2/train.py deleted file mode 120000 index 597332fdf..000000000 --- a/egs/librispeech/ASR/conv_emformer_transducer_stateless2/train.py +++ /dev/null @@ -1 +0,0 @@ -../conv_emformer_transducer_stateless/train.py \ No newline at end of file diff --git a/egs/librispeech/ASR/conv_emformer_transducer_stateless2/train.py b/egs/librispeech/ASR/conv_emformer_transducer_stateless2/train.py new file mode 100755 index 000000000..106f3e511 --- /dev/null +++ b/egs/librispeech/ASR/conv_emformer_transducer_stateless2/train.py @@ -0,0 +1,1136 @@ +#!/usr/bin/env python3 +# Copyright 2021 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" + +./conv_emformer_transducer_stateless/train.py \ + --world-size 4 \ + --num-epochs 30 \ + --start-epoch 1 \ + --exp-dir conv_emformer_transducer_stateless/exp \ + --full-libri 1 \ + --max-duration 300 \ + --master-port 12321 \ + --num-encoder-layers 12 \ + --chunk-length 32 \ + --cnn-module-kernel 31 \ + --left-context-length 32 \ + --right-context-length 8 \ + --memory-size 32 + +# For mix precision training: +./conv_emformer_transducer_stateless/train.py \ + --world-size 4 \ + --num-epochs 30 \ + --start-epoch 1 \ + --use-fp16 1 \ + --exp-dir conv_emformer_transducer_stateless/exp \ + --full-libri 1 \ + --max-duration 300 \ + --master-port 12321 \ + --num-encoder-layers 12 \ + --chunk-length 32 \ + --cnn-module-kernel 31 \ + --left-context-length 32 \ + --right-context-length 8 \ + --memory-size 32 +""" + + +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 emformer import Emformer +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, Eve +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 icefall import diagnostics +from icefall.checkpoint import load_checkpoint, remove_checkpoints +from icefall.checkpoint import save_checkpoint as save_checkpoint_impl +from icefall.checkpoint import ( + save_checkpoint_with_global_batch_idx, + update_averaged_model, +) +from icefall.dist import cleanup_dist, setup_dist +from icefall.env import get_env_info +from icefall.utils import AttributeDict, MetricsTracker, setup_logger, str2bool + +LRSchedulerType = Union[ + torch.optim.lr_scheduler._LRScheduler, optim.LRScheduler +] + + +def add_model_arguments(parser: argparse.ArgumentParser): + parser.add_argument( + "--encoder-dim", + type=int, + default=512, + help="Attention dim for the Emformer", + ) + + parser.add_argument( + "--nhead", + type=int, + default=8, + help="Number of attention heads for the Emformer", + ) + + parser.add_argument( + "--dim-feedforward", + type=int, + default=2048, + help="Feed-forward dimension for the Emformer", + ) + + parser.add_argument( + "--num-encoder-layers", + type=int, + default=12, + help="Number of encoder layers for the Emformer", + ) + + parser.add_argument( + "--cnn-module-kernel", + type=int, + default=31, + help="Kernel size for the convolution module.", + ) + + parser.add_argument( + "--left-context-length", + type=int, + default=32, + help="""Number of frames before subsampling for left context + in the Emformer.""", + ) + + parser.add_argument( + "--chunk-length", + type=int, + default=32, + help="""Number of frames before subsampling for each chunk + in the Emformer.""", + ) + + parser.add_argument( + "--right-context-length", + type=int, + default=8, + help="""Number of frames before subsampling for right context + in the Emformer.""", + ) + + parser.add_argument( + "--memory-size", + type=int, + default=0, + help="Number of entries in the memory for the Emformer", + ) + + +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_stateless2/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( + "--initial-lr", + type=float, + default=0.003, + help="""The initial learning rate. This value should not need to be + changed.""", + ) + + 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=6, + 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( + "--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( + "--save-every-n", + type=int, + default=8000, + 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=20, + 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=100, + 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=False, + help="Whether to use half precision training.", + ) + + add_model_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 warm_step for Noam optimizer. + """ + 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 Emformer + "feature_dim": 80, + "subsampling_factor": 4, + # parameters for decoder + "decoder_dim": 512, + # parameters for joiner + "joiner_dim": 512, + # parameters for Noam + "model_warm_step": 3000, # arg given to model, not for lrate + "env_info": get_env_info(), + } + ) + + return params + + +def get_encoder_model(params: AttributeDict) -> nn.Module: + # TODO: We can add an option to switch between Conformer and Transformer + encoder = Emformer( + num_features=params.feature_dim, + chunk_length=params.chunk_length, + subsampling_factor=params.subsampling_factor, + d_model=params.encoder_dim, + nhead=params.nhead, + dim_feedforward=params.dim_feedforward, + num_encoder_layers=params.num_encoder_layers, + cnn_module_kernel=params.cnn_module_kernel, + left_context_length=params.left_context_length, + right_context_length=params.right_context_length, + memory_size=params.memory_size, + ) + 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, + 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, + 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" + 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, + ) + + 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, + warmup: float = 1.0, +) -> Tuple[Tensor, MetricsTracker]: + """ + Compute RNN-T 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 Conformer 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 == 3 + feature = feature.to(device) + + supervisions = batch["supervisions"] + feature_lens = supervisions["num_frames"].to(device) + + texts = batch["supervisions"]["text"] + y = sp.encode(texts, out_type=int) + y = k2.RaggedTensor(y).to(device) + + with torch.set_grad_enabled(is_training): + simple_loss, pruned_loss = model( + x=feature, + x_lens=feature_lens, + y=y, + prune_range=params.prune_range, + am_scale=params.am_scale, + lm_scale=params.lm_scale, + warmup=warmup, + ) + # after the main warmup step, we keep pruned_loss_scale small + # for the same amount of time (model_warm_step), to avoid + # overwhelming the simple_loss and causing it to diverge, + # in case it had not fully learned the alignment yet. + pruned_loss_scale = ( + 0.0 + if warmup < 1.0 + else (0.1 if warmup > 1.0 and warmup < 2.0 else 1.0) + ) + loss = ( + params.simple_loss_scale * simple_loss + + pruned_loss_scale * pruned_loss + ) + + assert loss.requires_grad == is_training + + info = MetricsTracker() + 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["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["frames"] + 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, + scheduler: 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, +) -> 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) + + for batch_idx, batch in enumerate(train_dl): + if batch_idx < cur_batch_idx: + continue + cur_batch_idx = batch_idx + + params.batch_idx_train += 1 + batch_size = len(batch["supervisions"]["text"]) + + 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, + warmup=(params.batch_idx_train / params.model_warm_step), + ) + # 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() + scheduler.step_batch(params.batch_idx_train) + scaler.step(optimizer) + scaler.update() + optimizer.zero_grad() + + 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 % params.log_interval == 0: + cur_lr = scheduler.get_last_lr()[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}" + ) + + if tb_writer is not None: + 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 batch_idx > 0 and batch_idx % params.valid_interval == 0: + 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}") + if tb_writer is not None: + valid_info.write_summary( + tb_writer, "train/valid_", params.batch_idx_train + ) + + loss_value = tot_loss["loss"] / tot_loss["frames"] + 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): + """ + 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)) + if params.full_libri is False: + params.valid_interval = 1600 + + 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() 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) + + 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]) + + optimizer = Eve(model.parameters(), lr=params.initial_lr) + + scheduler = Eden(optimizer, params.lr_batches, params.lr_epochs) + + if checkpoints and "optimizer" in checkpoints: + logging.info("Loading optimizer state dict") + optimizer.load_state_dict(checkpoints["optimizer"]) + + if ( + checkpoints + and "scheduler" in checkpoints + and checkpoints["scheduler"] is not None + ): + 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) + + 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) + 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): + 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, + ) + + 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 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: + # warmup = 0.0 is so that the derivs for the pruned loss stay zero + # (i.e. are not remembered by the decaying-average in adam), because + # we want to avoid these params being subject to shrinkage in adam. + with torch.cuda.amp.autocast(enabled=params.use_fp16): + loss, _ = compute_loss( + params=params, + model=model, + sp=sp, + batch=batch, + is_training=True, + warmup=0.0, + ) + loss.backward() + optimizer.step() + optimizer.zero_grad() + except RuntimeError 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]}) ..." + ) + raise + + +def main(): + parser = get_parser() + LibriSpeechAsrDataModule.add_arguments(parser) + args = parser.parse_args() + args.exp_dir = Path(args.exp_dir) + + world_size = args.world_size + assert world_size >= 1 + if world_size > 1: + mp.spawn(run, args=(world_size, args), nprocs=world_size, join=True) + else: + run(rank=0, world_size=1, args=args) + + +torch.set_num_threads(1) +torch.set_num_interop_threads(1) + +if __name__ == "__main__": + main() From 208bbb6325da2e31054770a85b26d1749d9fdfa9 Mon Sep 17 00:00:00 2001 From: yaozengwei Date: Fri, 17 Jun 2022 22:52:05 +0800 Subject: [PATCH 05/12] fix bug in conv_emformer_transducer_stateless2/emformer.py --- .../ASR/conv_emformer_transducer_stateless2/emformer.py | 6 +++--- 1 file changed, 3 insertions(+), 3 deletions(-) diff --git a/egs/librispeech/ASR/conv_emformer_transducer_stateless2/emformer.py b/egs/librispeech/ASR/conv_emformer_transducer_stateless2/emformer.py index c5d862ad8..45ca03dd2 100644 --- a/egs/librispeech/ASR/conv_emformer_transducer_stateless2/emformer.py +++ b/egs/librispeech/ASR/conv_emformer_transducer_stateless2/emformer.py @@ -542,7 +542,7 @@ class EmformerAttention(nn.Module): padding_mask: Optional[torch.Tensor] = None, left_context_key: Optional[torch.Tensor] = None, left_context_val: Optional[torch.Tensor] = None, - ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor]: + ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]: """Underlying chunk-wise attention implementation.""" U, B, _ = utterance.size() R = right_context.size(0) @@ -671,7 +671,7 @@ class EmformerAttention(nn.Module): left_context_key: torch.Tensor, left_context_val: torch.Tensor, padding_mask: Optional[torch.Tensor] = None, - ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor]: + ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]: """Forward pass for inference. B: batch size; @@ -1388,7 +1388,7 @@ class EmformerEncoder(nn.Module): output_lengths = torch.clamp(lengths - self.right_context_length, min=0) attention_mask = self._gen_attention_mask(utterance) - M = right_context.size(0) // self.chunk_length - 1 + M = right_context.size(0) // self.right_context_length - 1 padding_mask = make_pad_mask(M + right_context.size(0) + output_lengths) output = utterance From 5b19011edb675688d401b27ac07c7b9c0dd4905f Mon Sep 17 00:00:00 2001 From: yaozengwei Date: Tue, 21 Jun 2022 20:52:42 +0800 Subject: [PATCH 06/12] fix doc in conv_emformer_transducer_stateless/emformer.py --- .../ASR/conv_emformer_transducer_stateless/emformer.py | 5 ++--- 1 file changed, 2 insertions(+), 3 deletions(-) diff --git a/egs/librispeech/ASR/conv_emformer_transducer_stateless/emformer.py b/egs/librispeech/ASR/conv_emformer_transducer_stateless/emformer.py index 46993da48..327cba2d3 100644 --- a/egs/librispeech/ASR/conv_emformer_transducer_stateless/emformer.py +++ b/egs/librispeech/ASR/conv_emformer_transducer_stateless/emformer.py @@ -1141,8 +1141,8 @@ class EmformerEncoderLayer(nn.Module): - output utterance, with shape (U, B, D); - output right_context, with shape (R, B, D); - output memory, with shape (1, B, D) or (0, B, D). - - output state. - - updated conv_cache. + - updated attention cache. + - updated convolution cache. """ R = right_context.size(0) src = torch.cat([right_context, utterance]) @@ -1525,7 +1525,6 @@ class EmformerEncoder(nn.Module): right_context at the end. states (List[torch.Tensor, List[List[torch.Tensor]], List[torch.Tensor]]: # noqa Cached states containing: - - past_lens: number of past frames for each sample in batch - attn_caches: attention states from preceding chunk's computation, where each element corresponds to each emformer layer - conv_caches: left context for causal convolution, where each From 42e3e883fdc0ba2e7606f4c4df4725321fa173a0 Mon Sep 17 00:00:00 2001 From: yaozengwei Date: Tue, 21 Jun 2022 22:07:36 +0800 Subject: [PATCH 07/12] refactor init states for stream --- .../emformer.py | 28 +++++++++++++++ .../stream.py | 36 +++---------------- .../streaming_decode.py | 2 ++ .../decode.py | 12 +++---- .../emformer.py | 28 +++++++++++++++ .../streaming_decode.py | 14 ++++---- .../train.py | 8 ++--- 7 files changed, 81 insertions(+), 47 deletions(-) diff --git a/egs/librispeech/ASR/conv_emformer_transducer_stateless/emformer.py b/egs/librispeech/ASR/conv_emformer_transducer_stateless/emformer.py index 327cba2d3..753e5c473 100644 --- a/egs/librispeech/ASR/conv_emformer_transducer_stateless/emformer.py +++ b/egs/librispeech/ASR/conv_emformer_transducer_stateless/emformer.py @@ -1630,6 +1630,30 @@ class EmformerEncoder(nn.Module): ) return output, output_lengths, output_states + def init_states(self, device: torch.device = torch.device("cpu")): + """Create initial states.""" + attn_caches = [ + [ + torch.zeros(self.memory_size, self.d_model, device=device), + torch.zeros( + self.left_context_length, self.d_model, device=device + ), + torch.zeros( + self.left_context_length, self.d_model, device=device + ), + ] + for _ in range(self.num_encoder_layers) + ] + conv_caches = [ + torch.zeros(self.d_model, self.cnn_module_kernel - 1, device=device) + for _ in range(self.num_encoder_layers) + ] + states: Tuple[List[List[torch.Tensor]], List[torch.Tensor]] = ( + attn_caches, + conv_caches, + ) + return states + class Emformer(EncoderInterface): def __init__( @@ -1802,6 +1826,10 @@ class Emformer(EncoderInterface): return output, output_lengths, output_states + def init_states(self, device: torch.device = torch.device("cpu")): + """Create initial states.""" + return self.encoder.init_states(device) + class Conv2dSubsampling(nn.Module): """Convolutional 2D subsampling (to 1/4 length). diff --git a/egs/librispeech/ASR/conv_emformer_transducer_stateless/stream.py b/egs/librispeech/ASR/conv_emformer_transducer_stateless/stream.py index 31ad3f50a..69ee7ee9a 100644 --- a/egs/librispeech/ASR/conv_emformer_transducer_stateless/stream.py +++ b/egs/librispeech/ASR/conv_emformer_transducer_stateless/stream.py @@ -43,15 +43,12 @@ class Stream(object): device: The device to run this stream. """ - self.device = device self.LOG_EPS = LOG_EPS # Containing attention caches and convolution caches self.states: Optional[ Tuple[List[List[torch.Tensor]], List[torch.Tensor]] ] = None - # Initailize zero states. - self.init_states(params) # It uses different attributes for different decoding methods. self.context_size = params.context_size @@ -107,34 +104,11 @@ class Stream(object): def set_ground_truth(self, ground_truth: str) -> None: self.ground_truth = ground_truth - def init_states(self, params: AttributeDict) -> None: - attn_caches = [ - [ - torch.zeros( - params.memory_size, params.encoder_dim, device=self.device - ), - torch.zeros( - params.left_context_length // params.subsampling_factor, - params.encoder_dim, - device=self.device, - ), - torch.zeros( - params.left_context_length // params.subsampling_factor, - params.encoder_dim, - device=self.device, - ), - ] - for _ in range(params.num_encoder_layers) - ] - conv_caches = [ - torch.zeros( - params.encoder_dim, - params.cnn_module_kernel - 1, - device=self.device, - ) - for _ in range(params.num_encoder_layers) - ] - self.states = (attn_caches, conv_caches) + def set_states( + self, states: Tuple[List[List[torch.Tensor]], List[torch.Tensor]] + ) -> None: + """Set states.""" + self.states = states def get_feature_chunk(self) -> torch.Tensor: """Get a chunk of feature frames. diff --git a/egs/librispeech/ASR/conv_emformer_transducer_stateless/streaming_decode.py b/egs/librispeech/ASR/conv_emformer_transducer_stateless/streaming_decode.py index 4fac405b0..0a6bbfa8b 100755 --- a/egs/librispeech/ASR/conv_emformer_transducer_stateless/streaming_decode.py +++ b/egs/librispeech/ASR/conv_emformer_transducer_stateless/streaming_decode.py @@ -683,6 +683,8 @@ def decode_dataset( LOG_EPS=LOG_EPSILON, ) + stream.set_states(model.encoder.init_states(device)) + audio: np.ndarray = cut.load_audio() # audio.shape: (1, num_samples) assert len(audio.shape) == 2 diff --git a/egs/librispeech/ASR/conv_emformer_transducer_stateless2/decode.py b/egs/librispeech/ASR/conv_emformer_transducer_stateless2/decode.py index 287fb94df..402ec4293 100755 --- a/egs/librispeech/ASR/conv_emformer_transducer_stateless2/decode.py +++ b/egs/librispeech/ASR/conv_emformer_transducer_stateless2/decode.py @@ -19,10 +19,10 @@ """ Usage: (1) greedy search -./conv_emformer_transducer_stateless/decode.py \ +./conv_emformer_transducer_stateless2/decode.py \ --epoch 30 \ --avg 10 \ - --exp-dir conv_emformer_transducer_stateless/exp \ + --exp-dir conv_emformer_transducer_stateless2/exp \ --max-duration 300 \ --num-encoder-layers 12 \ --chunk-length 32 \ @@ -34,10 +34,10 @@ Usage: --use-averaged-model True (2) modified beam search -./conv_emformer_transducer_stateless/decode.py \ +./conv_emformer_transducer_stateless2/decode.py \ --epoch 30 \ --avg 10 \ - --exp-dir conv_emformer_transducer_stateless/exp \ + --exp-dir conv_emformer_transducer_stateless2/exp \ --max-duration 300 \ --num-encoder-layers 12 \ --chunk-length 32 \ @@ -50,10 +50,10 @@ Usage: --beam-size 4 (3) fast beam search -./conv_emformer_transducer_stateless/decode.py \ +./conv_emformer_transducer_stateless2/decode.py \ --epoch 30 \ --avg 10 \ - --exp-dir conv_emformer_transducer_stateless/exp \ + --exp-dir conv_emformer_transducer_stateless2/exp \ --max-duration 300 \ --num-encoder-layers 12 \ --chunk-length 32 \ diff --git a/egs/librispeech/ASR/conv_emformer_transducer_stateless2/emformer.py b/egs/librispeech/ASR/conv_emformer_transducer_stateless2/emformer.py index 45ca03dd2..e3a598b0e 100644 --- a/egs/librispeech/ASR/conv_emformer_transducer_stateless2/emformer.py +++ b/egs/librispeech/ASR/conv_emformer_transducer_stateless2/emformer.py @@ -1529,6 +1529,30 @@ class EmformerEncoder(nn.Module): ) return output, output_lengths, output_states + def init_states(self, device: torch.device = torch.device("cpu")): + """Create initial states.""" + attn_caches = [ + [ + torch.zeros(self.memory_size, self.d_model, device=device), + torch.zeros( + self.left_context_length, self.d_model, device=device + ), + torch.zeros( + self.left_context_length, self.d_model, device=device + ), + ] + for _ in range(self.num_encoder_layers) + ] + conv_caches = [ + torch.zeros(self.d_model, self.cnn_module_kernel - 1, device=device) + for _ in range(self.num_encoder_layers) + ] + states: Tuple[List[List[torch.Tensor]], List[torch.Tensor]] = ( + attn_caches, + conv_caches, + ) + return states + class Emformer(EncoderInterface): def __init__( @@ -1701,6 +1725,10 @@ class Emformer(EncoderInterface): return output, output_lengths, output_states + def init_states(self, device: torch.device = torch.device("cpu")): + """Create initial states.""" + return self.encoder.init_states(device) + class Conv2dSubsampling(nn.Module): """Convolutional 2D subsampling (to 1/4 length). diff --git a/egs/librispeech/ASR/conv_emformer_transducer_stateless2/streaming_decode.py b/egs/librispeech/ASR/conv_emformer_transducer_stateless2/streaming_decode.py index 4fac405b0..0f687898f 100755 --- a/egs/librispeech/ASR/conv_emformer_transducer_stateless2/streaming_decode.py +++ b/egs/librispeech/ASR/conv_emformer_transducer_stateless2/streaming_decode.py @@ -19,10 +19,10 @@ """ Usage: (1) greedy search -./conv_emformer_transducer_stateless/streaming_decode.py \ +./conv_emformer_transducer_stateless2/streaming_decode.py \ --epoch 30 \ --avg 10 \ - --exp-dir conv_emformer_transducer_stateless/exp \ + --exp-dir conv_emformer_transducer_stateless2/exp \ --num-decode-streams 2000 \ --num-encoder-layers 12 \ --chunk-length 32 \ @@ -34,10 +34,10 @@ Usage: --use-averaged-model True (2) modified beam search -./conv_emformer_transducer_stateless/streaming_decode.py \ +./conv_emformer_transducer_stateless2/streaming_decode.py \ --epoch 30 \ --avg 10 \ - --exp-dir conv_emformer_transducer_stateless/exp \ + --exp-dir conv_emformer_transducer_stateless2/exp \ --num-decode-streams 2000 \ --num-encoder-layers 12 \ --chunk-length 32 \ @@ -50,10 +50,10 @@ Usage: --beam-size 4 (3) fast beam search -./conv_emformer_transducer_stateless/streaming_decode.py \ +./conv_emformer_transducer_stateless2/streaming_decode.py \ --epoch 30 \ --avg 10 \ - --exp-dir conv_emformer_transducer_stateless/exp \ + --exp-dir conv_emformer_transducer_stateless2/exp \ --num-decode-streams 2000 \ --num-encoder-layers 12 \ --chunk-length 32 \ @@ -683,6 +683,8 @@ def decode_dataset( LOG_EPS=LOG_EPSILON, ) + stream.set_states(model.encoder.init_states(device)) + audio: np.ndarray = cut.load_audio() # audio.shape: (1, num_samples) assert len(audio.shape) == 2 diff --git a/egs/librispeech/ASR/conv_emformer_transducer_stateless2/train.py b/egs/librispeech/ASR/conv_emformer_transducer_stateless2/train.py index 106f3e511..716ecc8b1 100755 --- a/egs/librispeech/ASR/conv_emformer_transducer_stateless2/train.py +++ b/egs/librispeech/ASR/conv_emformer_transducer_stateless2/train.py @@ -22,11 +22,11 @@ Usage: export CUDA_VISIBLE_DEVICES="0,1,2,3" -./conv_emformer_transducer_stateless/train.py \ +./conv_emformer_transducer_stateless2/train.py \ --world-size 4 \ --num-epochs 30 \ --start-epoch 1 \ - --exp-dir conv_emformer_transducer_stateless/exp \ + --exp-dir conv_emformer_transducer_stateless2/exp \ --full-libri 1 \ --max-duration 300 \ --master-port 12321 \ @@ -38,12 +38,12 @@ export CUDA_VISIBLE_DEVICES="0,1,2,3" --memory-size 32 # For mix precision training: -./conv_emformer_transducer_stateless/train.py \ +./conv_emformer_transducer_stateless2/train.py \ --world-size 4 \ --num-epochs 30 \ --start-epoch 1 \ --use-fp16 1 \ - --exp-dir conv_emformer_transducer_stateless/exp \ + --exp-dir conv_emformer_transducer_stateless2/exp \ --full-libri 1 \ --max-duration 300 \ --master-port 12321 \ From 9c37c163268666063569665f070bc1b50786e75e Mon Sep 17 00:00:00 2001 From: yaozengwei Date: Wed, 22 Jun 2022 12:28:26 +0800 Subject: [PATCH 08/12] modify .flake8 --- .flake8 | 1 + .../ASR/conv_emformer_transducer_stateless/train.py | 2 +- .../ASR/conv_emformer_transducer_stateless2/export.py | 10 +++++----- .../ASR/conv_emformer_transducer_stateless2/train.py | 2 +- 4 files changed, 8 insertions(+), 7 deletions(-) diff --git a/.flake8 b/.flake8 index c7c4f1855..9dd8d6207 100644 --- a/.flake8 +++ b/.flake8 @@ -10,6 +10,7 @@ per-file-ignores = egs/*/ASR/*/optim.py: E501, egs/*/ASR/*/scaling.py: E501, egs/librispeech/ASR/conv_emformer_transducer_stateless/*.py: E501, E203 + egs/librispeech/ASR/conv_emformer_transducer_stateless2/*.py: E501, E203 # invalid escape sequence (cause by tex formular), W605 icefall/utils.py: E501, W605 diff --git a/egs/librispeech/ASR/conv_emformer_transducer_stateless/train.py b/egs/librispeech/ASR/conv_emformer_transducer_stateless/train.py index 106f3e511..71c8eec27 100755 --- a/egs/librispeech/ASR/conv_emformer_transducer_stateless/train.py +++ b/egs/librispeech/ASR/conv_emformer_transducer_stateless/train.py @@ -28,7 +28,7 @@ export CUDA_VISIBLE_DEVICES="0,1,2,3" --start-epoch 1 \ --exp-dir conv_emformer_transducer_stateless/exp \ --full-libri 1 \ - --max-duration 300 \ + --max-duration 280 \ --master-port 12321 \ --num-encoder-layers 12 \ --chunk-length 32 \ diff --git a/egs/librispeech/ASR/conv_emformer_transducer_stateless2/export.py b/egs/librispeech/ASR/conv_emformer_transducer_stateless2/export.py index 4930881ea..ab15e0241 100755 --- a/egs/librispeech/ASR/conv_emformer_transducer_stateless2/export.py +++ b/egs/librispeech/ASR/conv_emformer_transducer_stateless2/export.py @@ -20,8 +20,8 @@ # to a single one using model averaging. """ Usage: -./conv_emformer_transducer_stateless/export.py \ - --exp-dir ./conv_emformer_transducer_stateless/exp \ +./conv_emformer_transducer_stateless2/export.py \ + --exp-dir ./conv_emformer_transducer_stateless2/exp \ --bpe-model data/lang_bpe_500/bpe.model \ --epoch 30 \ --avg 10 \ @@ -36,15 +36,15 @@ Usage: It will generate a file exp_dir/pretrained.pt -To use the generated file with `conv_emformer_transducer_stateless/decode.py`, +To use the generated file with `conv_emformer_transducer_stateless2/decode.py`, you can do: cd /path/to/exp_dir ln -s pretrained.pt epoch-9999.pt cd /path/to/egs/librispeech/ASR - ./conv_emformer_transducer_stateless/decode.py \ - --exp-dir ./conv_emformer_transducer_stateless/exp \ + ./conv_emformer_transducer_stateless2/decode.py \ + --exp-dir ./conv_emformer_transducer_stateless2/exp \ --epoch 9999 \ --avg 1 \ --max-duration 100 \ diff --git a/egs/librispeech/ASR/conv_emformer_transducer_stateless2/train.py b/egs/librispeech/ASR/conv_emformer_transducer_stateless2/train.py index 716ecc8b1..dfe1b6136 100755 --- a/egs/librispeech/ASR/conv_emformer_transducer_stateless2/train.py +++ b/egs/librispeech/ASR/conv_emformer_transducer_stateless2/train.py @@ -28,7 +28,7 @@ export CUDA_VISIBLE_DEVICES="0,1,2,3" --start-epoch 1 \ --exp-dir conv_emformer_transducer_stateless2/exp \ --full-libri 1 \ - --max-duration 300 \ + --max-duration 280 \ --master-port 12321 \ --num-encoder-layers 12 \ --chunk-length 32 \ From 4929faeed4f11af20f93dde76672786688f30c5d Mon Sep 17 00:00:00 2001 From: yaozengwei Date: Wed, 22 Jun 2022 14:41:04 +0800 Subject: [PATCH 09/12] copy files from conv_emformer_transducer_stateless2 --- .../asr_datamodule.py | 1 + .../beam_search.py | 1 + .../decode.py | 657 ++++++ .../decoder.py | 1 + .../emformer.py | 1824 +++++++++++++++++ .../encoder_interface.py | 1 + .../export.py | 287 +++ .../joiner.py | 1 + .../model.py | 1 + .../optim.py | 1 + .../scaling.py | 1 + .../stream.py | 1 + .../streaming_decode.py | 980 +++++++++ .../test_emformer.py | 194 ++ .../train.py | 1136 ++++++++++ 15 files changed, 5087 insertions(+) create mode 120000 egs/librispeech/ASR/conv_emformer_transducer_stateless3/asr_datamodule.py create mode 120000 egs/librispeech/ASR/conv_emformer_transducer_stateless3/beam_search.py create mode 100755 egs/librispeech/ASR/conv_emformer_transducer_stateless3/decode.py create mode 120000 egs/librispeech/ASR/conv_emformer_transducer_stateless3/decoder.py create mode 100644 egs/librispeech/ASR/conv_emformer_transducer_stateless3/emformer.py create mode 120000 egs/librispeech/ASR/conv_emformer_transducer_stateless3/encoder_interface.py create mode 100755 egs/librispeech/ASR/conv_emformer_transducer_stateless3/export.py create mode 120000 egs/librispeech/ASR/conv_emformer_transducer_stateless3/joiner.py create mode 120000 egs/librispeech/ASR/conv_emformer_transducer_stateless3/model.py create mode 120000 egs/librispeech/ASR/conv_emformer_transducer_stateless3/optim.py create mode 120000 egs/librispeech/ASR/conv_emformer_transducer_stateless3/scaling.py create mode 120000 egs/librispeech/ASR/conv_emformer_transducer_stateless3/stream.py create mode 100755 egs/librispeech/ASR/conv_emformer_transducer_stateless3/streaming_decode.py create mode 100644 egs/librispeech/ASR/conv_emformer_transducer_stateless3/test_emformer.py create mode 100755 egs/librispeech/ASR/conv_emformer_transducer_stateless3/train.py diff --git a/egs/librispeech/ASR/conv_emformer_transducer_stateless3/asr_datamodule.py b/egs/librispeech/ASR/conv_emformer_transducer_stateless3/asr_datamodule.py new file mode 120000 index 000000000..104eeea5d --- /dev/null +++ b/egs/librispeech/ASR/conv_emformer_transducer_stateless3/asr_datamodule.py @@ -0,0 +1 @@ +../conv_emformer_transducer_stateless/asr_datamodule.py \ No newline at end of file diff --git a/egs/librispeech/ASR/conv_emformer_transducer_stateless3/beam_search.py b/egs/librispeech/ASR/conv_emformer_transducer_stateless3/beam_search.py new file mode 120000 index 000000000..8554e44cc --- /dev/null +++ b/egs/librispeech/ASR/conv_emformer_transducer_stateless3/beam_search.py @@ -0,0 +1 @@ +../pruned_transducer_stateless2/beam_search.py \ No newline at end of file diff --git a/egs/librispeech/ASR/conv_emformer_transducer_stateless3/decode.py b/egs/librispeech/ASR/conv_emformer_transducer_stateless3/decode.py new file mode 100755 index 000000000..402ec4293 --- /dev/null +++ b/egs/librispeech/ASR/conv_emformer_transducer_stateless3/decode.py @@ -0,0 +1,657 @@ +#!/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: +(1) greedy search +./conv_emformer_transducer_stateless2/decode.py \ + --epoch 30 \ + --avg 10 \ + --exp-dir conv_emformer_transducer_stateless2/exp \ + --max-duration 300 \ + --num-encoder-layers 12 \ + --chunk-length 32 \ + --cnn-module-kernel 31 \ + --left-context-length 32 \ + --right-context-length 8 \ + --memory-size 32 \ + --decoding-method greedy_search \ + --use-averaged-model True + +(2) modified beam search +./conv_emformer_transducer_stateless2/decode.py \ + --epoch 30 \ + --avg 10 \ + --exp-dir conv_emformer_transducer_stateless2/exp \ + --max-duration 300 \ + --num-encoder-layers 12 \ + --chunk-length 32 \ + --cnn-module-kernel 31 \ + --left-context-length 32 \ + --right-context-length 8 \ + --memory-size 32 \ + --decoding-method modified_beam_search \ + --use-averaged-model True \ + --beam-size 4 + +(3) fast beam search +./conv_emformer_transducer_stateless2/decode.py \ + --epoch 30 \ + --avg 10 \ + --exp-dir conv_emformer_transducer_stateless2/exp \ + --max-duration 300 \ + --num-encoder-layers 12 \ + --chunk-length 32 \ + --cnn-module-kernel 31 \ + --left-context-length 32 \ + --right-context-length 8 \ + --memory-size 32 \ + --decoding-method fast_beam_search \ + --use-averaged-model True \ + --beam 4 \ + --max-contexts 4 \ + --max-states 8 +""" + + +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_one_best, + greedy_search, + greedy_search_batch, + modified_beam_search, +) +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 ( + 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( + "--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=10, + 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_stateless4/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( + "--decoding-method", + type=str, + default="greedy_search", + help="""Possible values are: + - greedy_search + - modified_beam_search + - fast_beam_search + """, + ) + + 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=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 --decoding-method is fast_beam_search""", + ) + + parser.add_argument( + "--max-contexts", + type=int, + default=4, + help="""Used only when --decoding-method is + fast_beam_search""", + ) + + parser.add_argument( + "--max-states", + type=int, + default=8, + help="""Used only when --decoding-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 --decoding_method is greedy_search""", + ) + + add_model_arguments(parser) + + return parser + + +def decode_one_batch( + params: AttributeDict, + model: nn.Module, + sp: spm.SentencePieceProcessor, + batch: dict, + 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`. + decoding_graph: + The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used + only when --decoding_method is fast_beam_search. + 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 == 3 + + feature = feature.to(device) + # at entry, feature is (N, T, C) + + supervisions = batch["supervisions"] + feature_lens = supervisions["num_frames"].to(device) + + feature_lens += params.chunk_length + feature = torch.nn.functional.pad( + feature, + pad=(0, 0, 0, params.chunk_length), + value=LOG_EPS, + ) + + 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 == "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 params.decoding_method == "fast_beam_search": + return { + ( + f"beam_{params.beam}_" + f"max_contexts_{params.max_contexts}_" + f"max_states_{params.max_states}" + ): 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, + decoding_graph: Optional[k2.Fsa] = None, +) -> Dict[str, List[Tuple[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. + decoding_graph: + The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used + only when --decoding_method is fast_beam_search. + 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 = 100 + else: + log_interval = 2 + + results = defaultdict(list) + for batch_idx, batch in enumerate(dl): + texts = batch["supervisions"]["text"] + + hyps_dict = decode_one_batch( + params=params, + model=model, + sp=sp, + decoding_graph=decoding_graph, + batch=batch, + ) + + for name, hyps in hyps_dict.items(): + this_batch = [] + assert len(hyps) == len(texts) + for hyp_words, ref_text in zip(hyps, texts): + ref_words = ref_text.split() + this_batch.append((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[List[int], List[int]]]], +): + 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" + ) + 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", + "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 "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}" + 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 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(params) + + 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 + start = params.epoch - params.avg + assert start >= 1 + 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 params.decoding_method == "fast_beam_search": + decoding_graph = k2.trivial_graph(params.vocab_size - 1, device=device) + else: + decoding_graph = None + + num_param = sum([p.numel() for p in model.parameters()]) + logging.info(f"Number of model parameters: {num_param}") + + 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, + sp=sp, + 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/librispeech/ASR/conv_emformer_transducer_stateless3/decoder.py b/egs/librispeech/ASR/conv_emformer_transducer_stateless3/decoder.py new file mode 120000 index 000000000..1db262df7 --- /dev/null +++ b/egs/librispeech/ASR/conv_emformer_transducer_stateless3/decoder.py @@ -0,0 +1 @@ +../conv_emformer_transducer_stateless/decoder.py \ No newline at end of file diff --git a/egs/librispeech/ASR/conv_emformer_transducer_stateless3/emformer.py b/egs/librispeech/ASR/conv_emformer_transducer_stateless3/emformer.py new file mode 100644 index 000000000..e3a598b0e --- /dev/null +++ b/egs/librispeech/ASR/conv_emformer_transducer_stateless3/emformer.py @@ -0,0 +1,1824 @@ +# Copyright 2022 Xiaomi Corporation (Author: 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. +# +# It is modified based on +# 1) https://github.com/pytorch/audio/blob/main/torchaudio/models/emformer.py # noqa +# 2) https://github.com/pytorch/audio/blob/main/torchaudio/prototype/models/conv_emformer.py # noqa + +import math +from typing import List, Optional, Tuple + +import torch +import torch.nn as nn +from encoder_interface import EncoderInterface +from scaling import ( + ActivationBalancer, + BasicNorm, + DoubleSwish, + ScaledConv1d, + ScaledConv2d, + ScaledLinear, +) + +from icefall.utils import make_pad_mask + + +LOG_EPSILON = math.log(1e-10) + + +def unstack_states( + states: Tuple[List[List[torch.Tensor]], List[torch.Tensor]] +) -> List[Tuple[List[List[torch.Tensor]], List[torch.Tensor]]]: + """Unstack the emformer state corresponding to a batch of utterances + into a list of states, where the i-th entry is the state from the i-th + utterance in the batch. + + Args: + states: + A tuple of 2 elements. + ``states[0]`` is the attention caches of a batch of utterance. + ``states[1]`` is the convolution caches of a batch of utterance. + ``len(states[0])`` and ``len(states[1])`` both eqaul to number of layers. # noqa + + Returns: + A list of states. + ``states[i]`` is a tuple of 2 elements of i-th utterance. + ``states[i][0]`` is the attention caches of i-th utterance. + ``states[i][1]`` is the convolution caches of i-th utterance. + ``len(states[i][0])`` and ``len(states[i][1])`` both eqaul to number of layers. # noqa + """ + + attn_caches, conv_caches = states + batch_size = conv_caches[0].size(0) + num_layers = len(attn_caches) + + list_attn_caches = [None] * batch_size + for i in range(batch_size): + list_attn_caches[i] = [[] for _ in range(num_layers)] + for li, layer in enumerate(attn_caches): + for s in layer: + s_list = s.unbind(dim=1) + for bi, b in enumerate(list_attn_caches): + b[li].append(s_list[bi]) + + list_conv_caches = [None] * batch_size + for i in range(batch_size): + list_conv_caches[i] = [None] * num_layers + for li, layer in enumerate(conv_caches): + c_list = layer.unbind(dim=0) + for bi, b in enumerate(list_conv_caches): + b[li] = c_list[bi] + + ans = [None] * batch_size + for i in range(batch_size): + ans[i] = [list_attn_caches[i], list_conv_caches[i]] + + return ans + + +def stack_states( + state_list: List[Tuple[List[List[torch.Tensor]], List[torch.Tensor]]] +) -> Tuple[List[List[torch.Tensor]], List[torch.Tensor]]: + """Stack list of emformer states that correspond to separate utterances + into a single emformer state so that it can be used as an input for + emformer when those utterances are formed into a batch. + + Note: + It is the inverse of :func:`unstack_states`. + + Args: + state_list: + Each element in state_list corresponding to the internal state + of the emformer model for a single utterance. + ``states[i]`` is a tuple of 2 elements of i-th utterance. + ``states[i][0]`` is the attention caches of i-th utterance. + ``states[i][1]`` is the convolution caches of i-th utterance. + ``len(states[i][0])`` and ``len(states[i][1])`` both eqaul to number of layers. # noqa + + Returns: + A new state corresponding to a batch of utterances. + See the input argument of :func:`unstack_states` for the meaning + of the returned tensor. + """ + batch_size = len(state_list) + + attn_caches = [] + for layer in state_list[0][0]: + if batch_size > 1: + # Note: We will stack attn_caches[layer][s][] later to get attn_caches[layer][s] # noqa + attn_caches.append([[s] for s in layer]) + else: + attn_caches.append([s.unsqueeze(1) for s in layer]) + for b, states in enumerate(state_list[1:], 1): + for li, layer in enumerate(states[0]): + for si, s in enumerate(layer): + attn_caches[li][si].append(s) + if b == batch_size - 1: + attn_caches[li][si] = torch.stack( + attn_caches[li][si], dim=1 + ) + + conv_caches = [] + for layer in state_list[0][1]: + if batch_size > 1: + # Note: We will stack conv_caches[layer][] later to get conv_caches[layer] # noqa + conv_caches.append([layer]) + else: + conv_caches.append(layer.unsqueeze(0)) + for b, states in enumerate(state_list[1:], 1): + for li, layer in enumerate(states[1]): + conv_caches[li].append(layer) + if b == batch_size - 1: + conv_caches[li] = torch.stack(conv_caches[li], dim=0) + + return [attn_caches, conv_caches] + + +class ConvolutionModule(nn.Module): + """ConvolutionModule. + + Modified from https://github.com/pytorch/audio/blob/main/torchaudio/prototype/models/conv_emformer.py # noqa + + Args: + chunk_length (int): + Length of each chunk. + right_context_length (int): + Length of right context. + channels (int): + The number of input channels and output 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, + chunk_length: int, + right_context_length: int, + channels: int, + kernel_size: int, + bias: bool = True, + ) -> None: + """Construct an ConvolutionModule object.""" + super().__init__() + # kernerl_size should be an odd number for 'SAME' padding + assert (kernel_size - 1) % 2 == 0, kernel_size + + self.chunk_length = chunk_length + self.right_context_length = right_context_length + self.channels = channels + + self.pointwise_conv1 = ScaledConv1d( + 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( + channel_dim=1, max_abs=10.0, min_positive=0.05, max_positive=1.0 + ) + + # make it causal by padding cached (kernel_size - 1) frames on the left + self.cache_size = kernel_size - 1 + self.depthwise_conv = ScaledConv1d( + channels, + channels, + kernel_size, + stride=1, + padding=0, + groups=channels, + bias=bias, + ) + + self.deriv_balancer2 = ActivationBalancer( + channel_dim=1, min_positive=0.05, max_positive=1.0 + ) + + self.activation = DoubleSwish() + + self.pointwise_conv2 = ScaledConv1d( + channels, + channels, + kernel_size=1, + stride=1, + padding=0, + bias=bias, + initial_scale=0.25, + ) + + def _split_right_context( + self, + pad_utterance: torch.Tensor, + right_context: torch.Tensor, + ) -> torch.Tensor: + """ + Args: + pad_utterance: + Its shape is (cache_size + U, B, D). + right_context: + Its shape is (R, B, D). + + Returns: + Right context segments padding with corresponding context. + Its shape is (num_segs * B, D, cache_size + right_context_length). + """ + U_, B, D = pad_utterance.size() + R = right_context.size(0) + assert self.right_context_length != 0 + assert R % self.right_context_length == 0 + num_chunks = R // self.right_context_length + right_context = right_context.reshape( + num_chunks, self.right_context_length, B, D + ) + right_context = right_context.permute(0, 2, 1, 3).reshape( + num_chunks * B, self.right_context_length, D + ) + + intervals = torch.arange( + 0, self.chunk_length * (num_chunks - 1), self.chunk_length + ) + first = torch.arange( + self.chunk_length, self.chunk_length + self.cache_size + ) + indexes = intervals.unsqueeze(1) + first.unsqueeze(0) + indexes = torch.cat( + [indexes, torch.arange(U_ - self.cache_size, U_).unsqueeze(0)] + ) + padding = pad_utterance[indexes] # (num_chunks, cache_size, B, D) + padding = padding.permute(0, 2, 1, 3).reshape( + num_chunks * B, self.cache_size, D + ) + + pad_right_context = torch.cat([padding, right_context], dim=1) + # (num_chunks * B, cache_size + right_context_length, D) + return pad_right_context.permute(0, 2, 1) + + def _merge_right_context( + self, right_context: torch.Tensor, B: int + ) -> torch.Tensor: + """ + Args: + right_context: + Right context segments. + It shape is (num_segs * B, D, right_context_length). + B: + Batch size. + + Returns: + A tensor of shape (B, D, R), where + R = num_segs * right_context_length. + """ + right_context = right_context.reshape( + -1, B, self.channels, self.right_context_length + ) + right_context = right_context.permute(1, 2, 0, 3) + right_context = right_context.reshape(B, self.channels, -1) + return right_context + + def forward( + self, + utterance: torch.Tensor, + right_context: torch.Tensor, + ) -> Tuple[torch.Tensor, torch.Tensor]: + """Causal convolution module. + + Args: + utterance (torch.Tensor): + Utterance tensor of shape (U, B, D). + right_context (torch.Tensor): + Right context tensor of shape (R, B, D). + + Returns: + A tuple of 2 tensors: + - output utterance of shape (U, B, D). + - output right_context of shape (R, B, D). + """ + U, B, D = utterance.size() + R, _, _ = right_context.size() + + # point-wise conv and GLU mechanism + x = torch.cat([right_context, utterance], dim=0) # (R + U, B, D) + x = x.permute(1, 2, 0) # (B, D, R + U) + x = self.pointwise_conv1(x) # (B, 2 * D, R + U) + x = self.deriv_balancer1(x) + x = nn.functional.glu(x, dim=1) # (B, D, R + U) + utterance = x[:, :, R:] # (B, D, U) + right_context = x[:, :, :R] # (B, D, R) + + # make causal convolution + cache = torch.zeros( + B, D, self.cache_size, device=x.device, dtype=x.dtype + ) + pad_utterance = torch.cat( + [cache, utterance], dim=2 + ) # (B, D, cache + U) + + # depth-wise conv on utterance + utterance = self.depthwise_conv(pad_utterance) # (B, D, U) + + if self.right_context_length > 0: + # depth-wise conv on right_context + pad_right_context = self._split_right_context( + pad_utterance.permute(2, 0, 1), right_context.permute(2, 0, 1) + ) # (num_segs * B, D, cache_size + right_context_length) + right_context = self.depthwise_conv( + pad_right_context + ) # (num_segs * B, D, right_context_length) + right_context = self._merge_right_context( + right_context, B + ) # (B, D, R) + + x = torch.cat([right_context, utterance], dim=2) # (B, D, R + U) + x = self.deriv_balancer2(x) + x = self.activation(x) + + # point-wise conv + x = self.pointwise_conv2(x) # (B, D, R + U) + + right_context = x[:, :, :R] # (B, D, R) + utterance = x[:, :, R:] # (B, D, U) + return ( + utterance.permute(2, 0, 1), + right_context.permute(2, 0, 1), + ) + + @torch.jit.export + def infer( + self, + utterance: torch.Tensor, + right_context: torch.Tensor, + cache: torch.Tensor, + ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]: + """Causal convolution module applied on both utterance and right_context. + + Args: + utterance (torch.Tensor): + Utterance tensor of shape (U, B, D). + right_context (torch.Tensor): + Right context tensor of shape (R, B, D). + cache (torch.Tensor, optional): + Cached tensor for left padding of shape (B, D, cache_size). + + Returns: + A tuple of 3 tensors: + - output utterance of shape (U, B, D). + - output right_context of shape (R, B, D). + - updated cache tensor of shape (B, D, cache_size). + """ + U, B, D = utterance.size() + R, _, _ = right_context.size() + + # point-wise conv + x = torch.cat([utterance, right_context], dim=0) # (U + R, B, D) + x = x.permute(1, 2, 0) # (B, D, U + R) + x = self.pointwise_conv1(x) # (B, 2 * D, U + R) + x = self.deriv_balancer1(x) + x = nn.functional.glu(x, dim=1) # (B, D, U + R) + + # make causal convolution + assert cache.shape == (B, D, self.cache_size), cache.shape + x = torch.cat([cache, x], dim=2) # (B, D, cache_size + U + R) + # update cache + new_cache = x[:, :, -R - self.cache_size : -R] + + # 1-D depth-wise conv + x = self.depthwise_conv(x) # (B, D, U + R) + + x = self.deriv_balancer2(x) + x = self.activation(x) + + # point-wise conv + x = self.pointwise_conv2(x) # (B, D, U + R) + + utterance = x[:, :, :U] # (B, D, U) + right_context = x[:, :, U:] # (B, D, R) + return ( + utterance.permute(2, 0, 1), + right_context.permute(2, 0, 1), + new_cache, + ) + + +class EmformerAttention(nn.Module): + r"""Emformer layer attention module. + + Args: + embed_dim (int): + Embedding dimension. + nhead (int): + Number of attention heads in each Emformer layer. + dropout (float, optional): + Dropout probability. (Default: 0.0) + tanh_on_mem (bool, optional): + If ``True``, applies tanh to memory elements. (Default: ``False``) + negative_inf (float, optional): + Value to use for negative infinity in attention weights. (Default: -1e8) + """ + + def __init__( + self, + embed_dim: int, + nhead: int, + dropout: float = 0.0, + tanh_on_mem: bool = False, + negative_inf: float = -1e8, + ): + super().__init__() + + if embed_dim % nhead != 0: + raise ValueError( + f"embed_dim ({embed_dim}) is not a multiple of" + f"nhead ({nhead})." + ) + + self.embed_dim = embed_dim + self.nhead = nhead + self.tanh_on_mem = tanh_on_mem + self.negative_inf = negative_inf + self.head_dim = embed_dim // nhead + self.dropout = dropout + + self.emb_to_key_value = ScaledLinear( + embed_dim, 2 * embed_dim, bias=True + ) + self.emb_to_query = ScaledLinear(embed_dim, embed_dim, bias=True) + self.out_proj = ScaledLinear( + embed_dim, embed_dim, bias=True, initial_scale=0.25 + ) + + def _gen_attention_probs( + self, + attention_weights: torch.Tensor, + attention_mask: torch.Tensor, + padding_mask: Optional[torch.Tensor] = None, + ) -> torch.Tensor: + """Given the entire attention weights, mask out unecessary connections + and optionally with padding positions, to obtain underlying chunk-wise + attention probabilities. + + B: batch size; + Q: length of query; + KV: length of key and value. + + Args: + attention_weights (torch.Tensor): + Attention weights computed on the entire concatenated tensor + with shape (B * nhead, Q, KV). + attention_mask (torch.Tensor): + Mask tensor where chunk-wise connections are filled with `False`, + and other unnecessary connections are filled with `True`, + with shape (Q, KV). + padding_mask (torch.Tensor, optional): + Mask tensor where the padding positions are fill with `True`, + and other positions are filled with `False`, with shapa `(B, KV)`. + + Returns: + A tensor of shape (B * nhead, Q, KV). + """ + attention_weights_float = attention_weights.float() + attention_weights_float = attention_weights_float.masked_fill( + attention_mask.unsqueeze(0), self.negative_inf + ) + if padding_mask is not None: + Q = attention_weights.size(1) + B = attention_weights.size(0) // self.nhead + attention_weights_float = attention_weights_float.view( + B, self.nhead, Q, -1 + ) + attention_weights_float = attention_weights_float.masked_fill( + padding_mask.unsqueeze(1).unsqueeze(2).to(torch.bool), + self.negative_inf, + ) + attention_weights_float = attention_weights_float.view( + B * self.nhead, Q, -1 + ) + + attention_probs = nn.functional.softmax( + attention_weights_float, dim=-1 + ).type_as(attention_weights) + + attention_probs = nn.functional.dropout( + attention_probs, p=self.dropout, training=self.training + ) + return attention_probs + + def _forward_impl( + self, + utterance: torch.Tensor, + right_context: torch.Tensor, + memory: torch.Tensor, + attention_mask: torch.Tensor, + padding_mask: Optional[torch.Tensor] = None, + left_context_key: Optional[torch.Tensor] = None, + left_context_val: Optional[torch.Tensor] = None, + ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]: + """Underlying chunk-wise attention implementation.""" + U, B, _ = utterance.size() + R = right_context.size(0) + M = memory.size(0) + scaling = float(self.head_dim) ** -0.5 + + # compute query with [right_context, utterance]. + query = self.emb_to_query(torch.cat([right_context, utterance])) + # compute key and value with [memory, right_context, utterance]. + key, value = self.emb_to_key_value( + torch.cat([memory, right_context, utterance]) + ).chunk(chunks=2, dim=2) + + if left_context_key is not None and left_context_val is not None: + # now compute key and value with + # [memory, right context, left context, uttrance] + # this is used in inference mode + key = torch.cat([key[: M + R], left_context_key, key[M + R :]]) + value = torch.cat( + [value[: M + R], left_context_val, value[M + R :]] + ) + Q = query.size(0) + # KV = key.size(0) + + reshaped_query, reshaped_key, reshaped_value = [ + tensor.contiguous() + .view(-1, B * self.nhead, self.head_dim) + .transpose(0, 1) + for tensor in [query, key, value] + ] # (B * nhead, Q or KV, head_dim) + attention_weights = torch.bmm( + reshaped_query * scaling, reshaped_key.transpose(1, 2) + ) # (B * nhead, Q, KV) + + # compute attention probabilities + attention_probs = self._gen_attention_probs( + attention_weights, attention_mask, padding_mask + ) + + # compute attention outputs + attention = torch.bmm(attention_probs, reshaped_value) + assert attention.shape == (B * self.nhead, Q, self.head_dim) + attention = ( + attention.transpose(0, 1).contiguous().view(Q, B, self.embed_dim) + ) + + # apply output projection + output_right_context_utterance = self.out_proj(attention) + + return output_right_context_utterance, key, value + + def forward( + self, + utterance: torch.Tensor, + right_context: torch.Tensor, + memory: torch.Tensor, + attention_mask: torch.Tensor, + padding_mask: Optional[torch.Tensor] = None, + ) -> torch.Tensor: + # TODO: Modify docs. + """Forward pass for training and validation mode. + + B: batch size; + D: embedding dimension; + R: length of the hard-copied right contexts; + U: length of full utterance; + M: length of memory vectors. + + It computes a `big` attention matrix on full utterance and + then utilizes a pre-computed mask to simulate chunk-wise attention. + + It concatenates three blocks: hard-copied right contexts, + and full utterance, as a `big` block, + to compute the query tensor: + query = [right_context, utterance], + with length Q = R + U. + It concatenates the three blocks: memory vectors, + hard-copied right contexts, and full utterance as another `big` block, + to compute the key and value tensors: + key & value = [memory, right_context, utterance], + with length KV = M + R + U. + Attention scores is computed with above `big` query and key. + + Then the underlying chunk-wise attention is obtained by applying + the attention mask. Suppose + c_i: chunk at index i; + r_i: right context that c_i can use; + l_i: left context that c_i can use; + m_i: past memory vectors from previous layer that c_i can use; + The target chunk-wise attention is: + c_i, r_i (in query) -> l_i, c_i, r_i, m_i (in key) + + Args: + utterance (torch.Tensor): + Full utterance frames, with shape (U, B, D). + right_context (torch.Tensor): + Hard-copied right context frames, with shape (R, B, D), + where R = num_chunks * right_context_length + memory (torch.Tensor): + Memory elements, with shape (M, B, D), where M = num_chunks - 1. + It is an empty tensor without using memory. + attention_mask (torch.Tensor): + Pre-computed attention mask to simulate underlying chunk-wise + attention, with shape (Q, KV). + padding_mask (torch.Tensor): + Padding mask of key tensor, with shape (B, KV). + + Returns: + Output of right context and utterance, with shape (R + U, B, D). + """ + output_right_context_utterance, _, _ = self._forward_impl( + utterance, + right_context, + memory, + attention_mask, + padding_mask=padding_mask, + ) + return output_right_context_utterance + + @torch.jit.export + def infer( + self, + utterance: torch.Tensor, + right_context: torch.Tensor, + memory: torch.Tensor, + left_context_key: torch.Tensor, + left_context_val: torch.Tensor, + padding_mask: Optional[torch.Tensor] = None, + ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]: + """Forward pass for inference. + + B: batch size; + D: embedding dimension; + R: length of right context; + U: length of utterance, i.e., current chunk; + L: length of cached left context; + M: length of cached memory vectors. + + It concatenates the right context and utterance (i.e., current chunk) + of current chunk, to compute the query tensor: + query = [right_context, utterance], + with length Q = R + U. + It concatenates the memory vectors, right context, left context, and + current chunk, to compute the key and value tensors: + key & value = [memory, right_context, left_context, utterance], + with length KV = M + R + L + U. + + The chunk-wise attention is: + chunk, right context (in query) -> + left context, chunk, right context, memory vectors (in key). + + Args: + utterance (torch.Tensor): + Current chunk frames, with shape (U, B, D), where U = chunk_length. + right_context (torch.Tensor): + Right context frames, with shape (R, B, D), + where R = right_context_length. + memory (torch.Tensor): + Memory vectors, with shape (M, B, D), or empty tensor. + left_context_key (torch,Tensor): + Cached attention key of left context from preceding computation, + with shape (L, B, D). + left_context_val (torch.Tensor): + Cached attention value of left context from preceding computation, + with shape (L, B, D). + padding_mask (torch.Tensor): + Padding mask of key tensor, with shape (B, KV). + + Returns: + A tuple containing 4 tensors: + - output of right context and utterance, with shape (R + U, B, D). + - attention key of left context and utterance, which would be cached + for next computation, with shape (L + U, B, D). + - attention value of left context and utterance, which would be + cached for next computation, with shape (L + U, B, D). + """ + U = utterance.size(0) + R = right_context.size(0) + L = left_context_key.size(0) + M = memory.size(0) + + # query = [right context, utterance] + Q = R + U + # key, value = [memory, right context, left context, uttrance] + KV = M + R + L + U + attention_mask = torch.zeros(Q, KV).to( + dtype=torch.bool, device=utterance.device + ) + + output_right_context_utterance, key, value = self._forward_impl( + utterance, + right_context, + memory, + attention_mask, + padding_mask=padding_mask, + left_context_key=left_context_key, + left_context_val=left_context_val, + ) + return ( + output_right_context_utterance, + key[M + R :], + value[M + R :], + ) + + +class EmformerEncoderLayer(nn.Module): + """Emformer layer that constitutes Emformer. + + Args: + d_model (int): + Input dimension. + nhead (int): + Number of attention heads. + dim_feedforward (int): + Hidden layer dimension of feedforward network. + chunk_length (int): + Length of each input segment. + dropout (float, optional): + Dropout probability. (Default: 0.0) + layer_dropout (float, optional): + Layer dropout probability. (Default: 0.0) + cnn_module_kernel (int): + Kernel size of convolution module. + left_context_length (int, optional): + Length of left context. (Default: 0) + right_context_length (int, optional): + Length of right context. (Default: 0) + memory_size (int, optional): + Number of memory elements to use. (Default: 0) + tanh_on_mem (bool, optional): + If ``True``, applies tanh to memory elements. (Default: ``False``) + negative_inf (float, optional): + Value to use for negative infinity in attention weights. (Default: -1e8) + """ + + def __init__( + self, + d_model: int, + nhead: int, + dim_feedforward: int, + chunk_length: int, + dropout: float = 0.1, + layer_dropout: float = 0.075, + cnn_module_kernel: int = 31, + left_context_length: int = 0, + right_context_length: int = 0, + memory_size: int = 0, + tanh_on_mem: bool = False, + negative_inf: float = -1e8, + ): + super().__init__() + + self.attention = EmformerAttention( + embed_dim=d_model, + nhead=nhead, + dropout=dropout, + tanh_on_mem=tanh_on_mem, + negative_inf=negative_inf, + ) + self.summary_op = nn.AvgPool1d( + kernel_size=chunk_length, stride=chunk_length, ceil_mode=True + ) + + self.feed_forward_macaron = nn.Sequential( + ScaledLinear(d_model, dim_feedforward), + ActivationBalancer(channel_dim=-1), + DoubleSwish(), + nn.Dropout(dropout), + ScaledLinear(dim_feedforward, d_model, initial_scale=0.25), + ) + + self.feed_forward = nn.Sequential( + ScaledLinear(d_model, dim_feedforward), + ActivationBalancer(channel_dim=-1), + DoubleSwish(), + nn.Dropout(dropout), + ScaledLinear(dim_feedforward, d_model, initial_scale=0.25), + ) + + self.conv_module = ConvolutionModule( + chunk_length, + right_context_length, + d_model, + cnn_module_kernel, + ) + + self.norm_final = BasicNorm(d_model) + + # try to ensure the output is close to zero-mean + # (or at least, zero-median). + self.balancer = ActivationBalancer( + channel_dim=-1, min_positive=0.45, max_positive=0.55, max_abs=6.0 + ) + + self.dropout = nn.Dropout(dropout) + + self.layer_dropout = layer_dropout + self.left_context_length = left_context_length + self.chunk_length = chunk_length + self.memory_size = memory_size + self.d_model = d_model + self.use_memory = memory_size > 0 + + def _update_attn_cache( + self, + next_key: torch.Tensor, + next_val: torch.Tensor, + memory: torch.Tensor, + attn_cache: List[torch.Tensor], + ) -> List[torch.Tensor]: + """Update cached attention state: + 1) output memory of current chunk in the lower layer; + 2) attention key and value in current chunk's computation, which would + be resued in next chunk's computation. + """ + new_memory = torch.cat([attn_cache[0], memory]) + new_key = torch.cat([attn_cache[1], next_key]) + new_val = torch.cat([attn_cache[2], next_val]) + attn_cache[0] = new_memory[new_memory.size(0) - self.memory_size :] + attn_cache[1] = new_key[new_key.size(0) - self.left_context_length :] + attn_cache[2] = new_val[new_val.size(0) - self.left_context_length :] + return attn_cache + + def _apply_conv_module_forward( + self, + right_context_utterance: torch.Tensor, + R: int, + ) -> torch.Tensor: + """Apply convolution module in training and validation mode.""" + utterance = right_context_utterance[R:] + right_context = right_context_utterance[:R] + utterance, right_context = self.conv_module(utterance, right_context) + right_context_utterance = torch.cat([right_context, utterance]) + return right_context_utterance + + def _apply_conv_module_infer( + self, + right_context_utterance: torch.Tensor, + R: int, + conv_cache: torch.Tensor, + ) -> Tuple[torch.Tensor, torch.Tensor]: + """Apply convolution module on utterance in inference mode.""" + utterance = right_context_utterance[R:] + right_context = right_context_utterance[:R] + utterance, right_context, conv_cache = self.conv_module.infer( + utterance, right_context, conv_cache + ) + right_context_utterance = torch.cat([right_context, utterance]) + return right_context_utterance, conv_cache + + def _apply_attention_module_forward( + self, + right_context_utterance: torch.Tensor, + R: int, + attention_mask: torch.Tensor, + padding_mask: Optional[torch.Tensor] = None, + ) -> torch.Tensor: + """Apply attention module in training and validation mode.""" + utterance = right_context_utterance[R:] + right_context = right_context_utterance[:R] + + if self.use_memory: + memory = self.summary_op(utterance.permute(1, 2, 0)).permute( + 2, 0, 1 + )[:-1, :, :] + else: + memory = torch.empty(0).to( + dtype=utterance.dtype, device=utterance.device + ) + output_right_context_utterance = self.attention( + utterance=utterance, + right_context=right_context, + memory=memory, + attention_mask=attention_mask, + padding_mask=padding_mask, + ) + + return output_right_context_utterance + + def _apply_attention_module_infer( + self, + right_context_utterance: torch.Tensor, + R: int, + attn_cache: List[torch.Tensor], + padding_mask: Optional[torch.Tensor] = None, + ) -> Tuple[torch.Tensor, List[torch.Tensor]]: + """Apply attention module in inference mode. + 1) Unpack cached states including: + - memory from previous chunks; + - attention key and value of left context from preceding + chunk's compuation; + 2) Apply attention computation; + 3) Update cached attention states including: + - memory of current chunk; + - attention key and value in current chunk's computation, which would + be resued in next chunk's computation. + """ + utterance = right_context_utterance[R:] + right_context = right_context_utterance[:R] + + pre_memory = attn_cache[0] + left_context_key = attn_cache[1] + left_context_val = attn_cache[2] + + if self.use_memory: + memory = self.summary_op(utterance.permute(1, 2, 0)).permute( + 2, 0, 1 + )[:1, :, :] + else: + memory = torch.empty(0).to( + dtype=utterance.dtype, device=utterance.device + ) + ( + output_right_context_utterance, + next_key, + next_val, + ) = self.attention.infer( + utterance=utterance, + right_context=right_context, + memory=pre_memory, + left_context_key=left_context_key, + left_context_val=left_context_val, + padding_mask=padding_mask, + ) + attn_cache = self._update_attn_cache( + next_key, next_val, memory, attn_cache + ) + return output_right_context_utterance, attn_cache + + def forward( + self, + utterance: torch.Tensor, + right_context: torch.Tensor, + attention_mask: torch.Tensor, + padding_mask: Optional[torch.Tensor] = None, + warmup: float = 1.0, + ) -> Tuple[torch.Tensor, torch.Tensor]: + r"""Forward pass for training and validation mode. + + B: batch size; + D: embedding dimension; + R: length of hard-copied right contexts; + U: length of full utterance; + M: length of memory vectors. + + Args: + utterance (torch.Tensor): + Utterance frames, with shape (U, B, D). + right_context (torch.Tensor): + Right context frames, with shape (R, B, D). + attention_mask (torch.Tensor): + Attention mask for underlying attention module, + with shape (Q, KV), where Q = R + U, KV = M + R + U. + padding_mask (torch.Tensor): + Padding mask of ker tensor, with shape (B, KV). + + Returns: + A tuple containing 2 tensors: + - output utterance, with shape (U, B, D). + - output right context, with shape (R, B, D). + """ + R = right_context.size(0) + src = torch.cat([right_context, utterance]) + src_orig = src + + warmup_scale = min(0.1 + warmup, 1.0) + # alpha = 1.0 means fully use this encoder layer, 0.0 would mean + # completely bypass it. + if self.training: + alpha = ( + warmup_scale + if torch.rand(()).item() <= (1.0 - self.layer_dropout) + else 0.1 + ) + else: + alpha = 1.0 + + # macaron style feed forward module + src = src + self.dropout(self.feed_forward_macaron(src)) + + # emformer attention module + src_att = self._apply_attention_module_forward( + src, R, attention_mask, padding_mask=padding_mask + ) + src = src + self.dropout(src_att) + + # convolution module + src_conv = self._apply_conv_module_forward(src, R) + src = src + self.dropout(src_conv) + + # feed forward module + src = src + self.dropout(self.feed_forward(src)) + + src = self.norm_final(self.balancer(src)) + + if alpha != 1.0: + src = alpha * src + (1 - alpha) * src_orig + + output_utterance = src[R:] + output_right_context = src[:R] + return output_utterance, output_right_context + + @torch.jit.export + def infer( + self, + utterance: torch.Tensor, + right_context: torch.Tensor, + attn_cache: List[torch.Tensor], + conv_cache: torch.Tensor, + padding_mask: Optional[torch.Tensor] = None, + ) -> Tuple[torch.Tensor, torch.Tensor, List[torch.Tensor], torch.Tensor]: + """Forward pass for inference. + + B: batch size; + D: embedding dimension; + R: length of right_context; + U: length of utterance; + M: length of memory. + + Args: + utterance (torch.Tensor): + Utterance frames, with shape (U, B, D). + right_context (torch.Tensor): + Right context frames, with shape (R, B, D). + attn_cache (List[torch.Tensor]): + Cached attention tensors generated in preceding computation, + including memory, key and value of left context. + conv_cache (torch.Tensor, optional): + Cache tensor of left context for causal convolution. + padding_mask (torch.Tensor): + Padding mask of ker tensor. + + Returns: + (Tensor, Tensor, List[torch.Tensor], Tensor): + - output utterance, with shape (U, B, D); + - output right_context, with shape (R, B, D); + - output attention cache; + - output convolution cache. + """ + R = right_context.size(0) + src = torch.cat([right_context, utterance]) + + # macaron style feed forward module + src = src + self.dropout(self.feed_forward_macaron(src)) + + # emformer attention module + src_att, attn_cache = self._apply_attention_module_infer( + src, R, attn_cache, padding_mask=padding_mask + ) + src = src + self.dropout(src_att) + + # convolution module + src_conv, conv_cache = self._apply_conv_module_infer(src, R, conv_cache) + src = src + self.dropout(src_conv) + + # feed forward module + src = src + self.dropout(self.feed_forward(src)) + + src = self.norm_final(self.balancer(src)) + + output_utterance = src[R:] + output_right_context = src[:R] + return ( + output_utterance, + output_right_context, + attn_cache, + conv_cache, + ) + + +def _gen_attention_mask_block( + col_widths: List[int], + col_mask: List[bool], + num_rows: int, + device: torch.device, +) -> torch.Tensor: + assert len(col_widths) == len( + col_mask + ), "Length of col_widths must match that of col_mask" + + mask_block = [ + torch.ones(num_rows, col_width, device=device) + if is_ones_col + else torch.zeros(num_rows, col_width, device=device) + for col_width, is_ones_col in zip(col_widths, col_mask) + ] + return torch.cat(mask_block, dim=1) + + +class EmformerEncoder(nn.Module): + """Implements the Emformer architecture introduced in + *Emformer: Efficient Memory Transformer Based Acoustic Model for Low Latency + Streaming Speech Recognition* + [:footcite:`shi2021emformer`]. + + Args: + d_model (int): + Input dimension. + nhead (int): + Number of attention heads in each emformer layer. + dim_feedforward (int): + Hidden layer dimension of each emformer layer's feedforward network. + num_encoder_layers (int): + Number of emformer layers to instantiate. + chunk_length (int): + Length of each input segment. + dropout (float, optional): + Dropout probability. (default: 0.0) + layer_dropout (float, optional): + Layer dropout probability. (default: 0.0) + cnn_module_kernel (int): + Kernel size of convolution module. + left_context_length (int, optional): + Length of left context. (default: 0) + right_context_length (int, optional): + Length of right context. (default: 0) + memory_size (int, optional): + Number of memory elements to use. (default: 0) + tanh_on_mem (bool, optional): + If ``true``, applies tanh to memory elements. (default: ``false``) + negative_inf (float, optional): + Value to use for negative infinity in attention weights. (default: -1e8) + """ + + def __init__( + self, + chunk_length: int, + d_model: int = 256, + nhead: int = 4, + dim_feedforward: int = 2048, + num_encoder_layers: int = 12, + dropout: float = 0.1, + layer_dropout: float = 0.075, + cnn_module_kernel: int = 31, + left_context_length: int = 0, + right_context_length: int = 0, + memory_size: int = 0, + tanh_on_mem: bool = False, + negative_inf: float = -1e8, + ): + super().__init__() + + self.use_memory = memory_size > 0 + + self.emformer_layers = nn.ModuleList( + [ + EmformerEncoderLayer( + d_model=d_model, + nhead=nhead, + dim_feedforward=dim_feedforward, + chunk_length=chunk_length, + dropout=dropout, + layer_dropout=layer_dropout, + cnn_module_kernel=cnn_module_kernel, + left_context_length=left_context_length, + right_context_length=right_context_length, + memory_size=memory_size, + tanh_on_mem=tanh_on_mem, + negative_inf=negative_inf, + ) + for layer_idx in range(num_encoder_layers) + ] + ) + + self.num_encoder_layers = num_encoder_layers + self.d_model = d_model + self.left_context_length = left_context_length + self.right_context_length = right_context_length + self.chunk_length = chunk_length + self.memory_size = memory_size + self.cnn_module_kernel = cnn_module_kernel + + def _gen_right_context(self, x: torch.Tensor) -> torch.Tensor: + """Hard copy each chunk's right context and concat them.""" + T = x.shape[0] + num_chunks = math.ceil( + (T - self.right_context_length) / self.chunk_length + ) + # first (num_chunks - 1) right context block + intervals = torch.arange( + 0, self.chunk_length * (num_chunks - 1), self.chunk_length + ) + first = torch.arange( + self.chunk_length, self.chunk_length + self.right_context_length + ) + indexes = intervals.unsqueeze(1) + first.unsqueeze(0) + # cat last right context block + indexes = torch.cat( + [ + indexes, + torch.arange(T - self.right_context_length, T).unsqueeze(0), + ] + ) + right_context_blocks = x[indexes.reshape(-1)] + return right_context_blocks + + def _gen_attention_mask_col_widths( + self, chunk_idx: int, U: int + ) -> List[int]: + """Calculate column widths (key, value) in attention mask for the + chunk_idx chunk.""" + num_chunks = math.ceil(U / self.chunk_length) + rc = self.right_context_length + lc = self.left_context_length + rc_start = chunk_idx * rc + rc_end = rc_start + rc + chunk_start = max(chunk_idx * self.chunk_length - lc, 0) + chunk_end = min((chunk_idx + 1) * self.chunk_length, U) + R = rc * num_chunks + + if self.use_memory: + m_start = max(chunk_idx - self.memory_size, 0) + M = num_chunks - 1 + col_widths = [ + m_start, # before memory + chunk_idx - m_start, # memory + M - chunk_idx, # after memory + rc_start, # before right context + rc, # right context + R - rc_end, # after right context + chunk_start, # before chunk + chunk_end - chunk_start, # chunk + U - chunk_end, # after chunk + ] + else: + col_widths = [ + rc_start, # before right context + rc, # right context + R - rc_end, # after right context + chunk_start, # before chunk + chunk_end - chunk_start, # chunk + U - chunk_end, # after chunk + ] + + return col_widths + + def _gen_attention_mask(self, utterance: torch.Tensor) -> torch.Tensor: + """Generate attention mask to simulate underlying chunk-wise attention + computation, where chunk-wise connections are filled with `False`, + and other unnecessary connections beyond chunk are filled with `True`. + + R: length of hard-copied right contexts; + U: length of full utterance; + M: length of memory vectors; + Q: length of attention query; + KV: length of attention key and value. + + The shape of attention mask is (Q, KV). + If self.use_memory is `True`: + query = [right_context, utterance]; + key, value = [memory, right_context, utterance]; + Q = R + U, KV = M + R + U. + Otherwise: + query = [right_context, utterance] + key, value = [right_context, utterance] + Q = R + U, KV = R + U. + + Suppose: + c_i: chunk at index i; + r_i: right context that c_i can use; + l_i: left context that c_i can use; + m_i: past memory vectors from previous layer that c_i can use; + The target chunk-wise attention is: + c_i, r_i (in query) -> l_i, c_i, r_i, m_i (in key). + """ + U = utterance.size(0) + num_chunks = math.ceil(U / self.chunk_length) + + right_context_mask = [] + utterance_mask = [] + + if self.use_memory: + num_cols = 9 + # right context and utterance both attend to memory, right context, + # utterance + right_context_utterance_cols_mask = [ + idx in [1, 4, 7] for idx in range(num_cols) + ] + else: + num_cols = 6 + # right context and utterance both attend to right context and + # utterance + right_context_utterance_cols_mask = [ + idx in [1, 4] for idx in range(num_cols) + ] + masks_to_concat = [right_context_mask, utterance_mask] + + for chunk_idx in range(num_chunks): + col_widths = self._gen_attention_mask_col_widths(chunk_idx, U) + + right_context_mask_block = _gen_attention_mask_block( + col_widths, + right_context_utterance_cols_mask, + self.right_context_length, + utterance.device, + ) + right_context_mask.append(right_context_mask_block) + + utterance_mask_block = _gen_attention_mask_block( + col_widths, + right_context_utterance_cols_mask, + min( + self.chunk_length, + U - chunk_idx * self.chunk_length, + ), + utterance.device, + ) + utterance_mask.append(utterance_mask_block) + + attention_mask = ( + 1 - torch.cat([torch.cat(mask) for mask in masks_to_concat]) + ).to(torch.bool) + return attention_mask + + def forward( + self, x: torch.Tensor, lengths: torch.Tensor, warmup: float = 1.0 + ) -> Tuple[torch.Tensor, torch.Tensor]: + """Forward pass for training and validation mode. + + B: batch size; + D: input dimension; + U: length of utterance. + + Args: + x (torch.Tensor): + Utterance frames right-padded with right context frames, + with shape (U + right_context_length, B, D). + lengths (torch.Tensor): + With shape (B,) and i-th element representing number of valid + utterance frames for i-th batch element in x, which contains the + right_context at the end. + + Returns: + A tuple of 2 tensors: + - output utterance frames, with shape (U, B, D). + - output_lengths, with shape (B,), without containing the + right_context at the end. + """ + U = x.size(0) - self.right_context_length + + right_context = self._gen_right_context(x) + utterance = x[:U] + output_lengths = torch.clamp(lengths - self.right_context_length, min=0) + attention_mask = self._gen_attention_mask(utterance) + + M = right_context.size(0) // self.right_context_length - 1 + padding_mask = make_pad_mask(M + right_context.size(0) + output_lengths) + + output = utterance + for layer in self.emformer_layers: + output, right_context = layer( + output, + right_context, + attention_mask, + padding_mask=padding_mask, + warmup=warmup, + ) + + return output, output_lengths + + @torch.jit.export + def infer( + self, + x: torch.Tensor, + lengths: torch.Tensor, + num_processed_frames: torch.Tensor, + states: Tuple[List[List[torch.Tensor]], List[torch.Tensor]], + ) -> Tuple[ + torch.Tensor, + torch.Tensor, + Tuple[List[List[torch.Tensor]], List[torch.Tensor]], + ]: + """Forward pass for streaming inference. + + B: batch size; + D: input dimension; + U: length of utterance. + + Args: + x (torch.Tensor): + Utterance frames right-padded with right context frames, + with shape (U + right_context_length, B, D). + lengths (torch.Tensor): + With shape (B,) and i-th element representing number of valid + utterance frames for i-th batch element in x, which contains the + right_context at the end. + states (List[torch.Tensor, List[List[torch.Tensor]], List[torch.Tensor]]: # noqa + Cached states containing: + - attn_caches: attention states from preceding chunk's computation, + where each element corresponds to each emformer layer + - conv_caches: left context for causal convolution, where each + element corresponds to each layer. + + Returns: + (Tensor, Tensor, List[List[torch.Tensor]], List[torch.Tensor]): + - output utterance frames, with shape (U, B, D). + - output lengths, with shape (B,), without containing the + right_context at the end. + - updated states from current chunk's computation. + """ + assert num_processed_frames.shape == (x.size(1),) + + attn_caches = states[0] + assert len(attn_caches) == self.num_encoder_layers, len(attn_caches) + for i in range(len(attn_caches)): + assert attn_caches[i][0].shape == ( + self.memory_size, + x.size(1), + self.d_model, + ), attn_caches[i][0].shape + assert attn_caches[i][1].shape == ( + self.left_context_length, + x.size(1), + self.d_model, + ), attn_caches[i][1].shape + assert attn_caches[i][2].shape == ( + self.left_context_length, + x.size(1), + self.d_model, + ), attn_caches[i][2].shape + + conv_caches = states[1] + assert len(conv_caches) == self.num_encoder_layers, len(conv_caches) + for i in range(len(conv_caches)): + assert conv_caches[i].shape == ( + x.size(1), + self.d_model, + self.cnn_module_kernel - 1, + ), conv_caches[i].shape + + right_context = x[-self.right_context_length :] + utterance = x[: -self.right_context_length] + output_lengths = torch.clamp(lengths - self.right_context_length, min=0) + + # calcualte padding mask to mask out initial zero caches + chunk_mask = make_pad_mask(output_lengths).to(x.device) + memory_mask = ( + torch.div( + num_processed_frames, self.chunk_length, rounding_mode="floor" + ).view(x.size(1), 1) + <= torch.arange(self.memory_size, device=x.device).expand( + x.size(1), self.memory_size + ) + ).flip(1) + left_context_mask = ( + num_processed_frames.view(x.size(1), 1) + <= torch.arange(self.left_context_length, device=x.device).expand( + x.size(1), self.left_context_length + ) + ).flip(1) + right_context_mask = torch.zeros( + x.size(1), + self.right_context_length, + dtype=torch.bool, + device=x.device, + ) + padding_mask = torch.cat( + [memory_mask, right_context_mask, left_context_mask, chunk_mask], + dim=1, + ) + + output = utterance + output_attn_caches: List[List[torch.Tensor]] = [] + output_conv_caches: List[torch.Tensor] = [] + for layer_idx, layer in enumerate(self.emformer_layers): + ( + output, + right_context, + output_attn_cache, + output_conv_cache, + ) = layer.infer( + output, + right_context, + padding_mask=padding_mask, + attn_cache=attn_caches[layer_idx], + conv_cache=conv_caches[layer_idx], + ) + output_attn_caches.append(output_attn_cache) + output_conv_caches.append(output_conv_cache) + + output_states: Tuple[List[List[torch.Tensor]], List[torch.Tensor]] = ( + output_attn_caches, + output_conv_caches, + ) + return output, output_lengths, output_states + + def init_states(self, device: torch.device = torch.device("cpu")): + """Create initial states.""" + attn_caches = [ + [ + torch.zeros(self.memory_size, self.d_model, device=device), + torch.zeros( + self.left_context_length, self.d_model, device=device + ), + torch.zeros( + self.left_context_length, self.d_model, device=device + ), + ] + for _ in range(self.num_encoder_layers) + ] + conv_caches = [ + torch.zeros(self.d_model, self.cnn_module_kernel - 1, device=device) + for _ in range(self.num_encoder_layers) + ] + states: Tuple[List[List[torch.Tensor]], List[torch.Tensor]] = ( + attn_caches, + conv_caches, + ) + return states + + +class Emformer(EncoderInterface): + def __init__( + self, + num_features: int, + chunk_length: int, + subsampling_factor: int = 4, + d_model: int = 256, + nhead: int = 4, + dim_feedforward: int = 2048, + num_encoder_layers: int = 12, + dropout: float = 0.1, + layer_dropout: float = 0.075, + cnn_module_kernel: int = 3, + left_context_length: int = 0, + right_context_length: int = 0, + memory_size: int = 0, + tanh_on_mem: bool = False, + negative_inf: float = -1e8, + ): + super().__init__() + + self.subsampling_factor = subsampling_factor + self.right_context_length = right_context_length + if subsampling_factor != 4: + raise NotImplementedError("Support only 'subsampling_factor=4'.") + if chunk_length % subsampling_factor != 0: + raise NotImplementedError( + "chunk_length must be a mutiple of subsampling_factor." + ) + if ( + left_context_length != 0 + and left_context_length % subsampling_factor != 0 + ): + raise NotImplementedError( + "left_context_length must be 0 or a mutiple of subsampling_factor." # noqa + ) + if ( + right_context_length != 0 + and right_context_length % subsampling_factor != 0 + ): + raise NotImplementedError( + "right_context_length must be 0 or a mutiple of subsampling_factor." # noqa + ) + + # self.encoder_embed converts the input of shape (N, T, num_features) + # to the shape (N, T//subsampling_factor, d_model). + # That is, it does two things simultaneously: + # (1) subsampling: T -> T//subsampling_factor + # (2) embedding: num_features -> d_model + self.encoder_embed = Conv2dSubsampling(num_features, d_model) + + self.encoder = EmformerEncoder( + chunk_length=chunk_length // subsampling_factor, + d_model=d_model, + nhead=nhead, + dim_feedforward=dim_feedforward, + num_encoder_layers=num_encoder_layers, + dropout=dropout, + layer_dropout=layer_dropout, + cnn_module_kernel=cnn_module_kernel, + left_context_length=left_context_length // subsampling_factor, + right_context_length=right_context_length // subsampling_factor, + memory_size=memory_size, + tanh_on_mem=tanh_on_mem, + negative_inf=negative_inf, + ) + + def forward( + self, x: torch.Tensor, x_lens: torch.Tensor, warmup: float = 1.0 + ) -> Tuple[torch.Tensor, torch.Tensor]: + """Forward pass for training and non-streaming inference. + + B: batch size; + D: feature dimension; + T: length of utterance. + + Args: + x (torch.Tensor): + Utterance frames right-padded with right context frames, + with shape (B, T, D). + x_lens (torch.Tensor): + With shape (B,) and i-th element representing number of valid + utterance frames for i-th batch element in x, containing the + right_context at the end. + warmup: + A floating point value that gradually increases from 0 throughout + training; when it is >= 1.0 we are "fully warmed up". It is used + to turn modules on sequentially. + + Returns: + (Tensor, Tensor): + - output embedding, with shape (B, T', D), where + T' = ((T - 1) // 2 - 1) // 2 - self.right_context_length // 4. + - output lengths, with shape (B,), without containing the + right_context at the end. + """ + x = self.encoder_embed(x) + x = x.permute(1, 0, 2) # (N, T, C) -> (T, N, C) + + x_lens = (((x_lens - 1) >> 1) - 1) >> 1 + assert x.size(0) == x_lens.max().item() + + output, output_lengths = self.encoder( + x, x_lens, warmup=warmup + ) # (T, N, C) + + output = output.permute(1, 0, 2) # (T, N, C) -> (N, T, C) + + return output, output_lengths + + @torch.jit.export + def infer( + self, + x: torch.Tensor, + x_lens: torch.Tensor, + num_processed_frames: torch.Tensor, + states: Tuple[List[List[torch.Tensor]], List[torch.Tensor]], + ) -> Tuple[ + torch.Tensor, + torch.Tensor, + Tuple[List[List[torch.Tensor]], List[torch.Tensor]], + ]: + """Forward pass for streaming inference. + + B: batch size; + D: feature dimension; + T: length of utterance. + + Args: + x (torch.Tensor): + Utterance frames right-padded with right context frames, + with shape (B, T, D). + lengths (torch.Tensor): + With shape (B,) and i-th element representing number of valid + utterance frames for i-th batch element in x, containing the + right_context at the end. + states (List[torch.Tensor, List[List[torch.Tensor]], List[torch.Tensor]]: # noqa + Cached states containing: + - past_lens: number of past frames for each sample in batch + - attn_caches: attention states from preceding chunk's computation, + where each element corresponds to each emformer layer + - conv_caches: left context for causal convolution, where each + element corresponds to each layer. + Returns: + (Tensor, Tensor): + - output embedding, with shape (B, T', D), where + T' = ((T - 1) // 2 - 1) // 2 - self.right_context_length // 4. + - output lengths, with shape (B,), without containing the + right_context at the end. + - updated states from current chunk's computation. + """ + x = self.encoder_embed(x) + # drop the first and last frames + x = x[:, 1:-1, :] + x = x.permute(1, 0, 2) # (N, T, C) -> (T, N, C) + + # Caution: We assume the subsampling factor is 4! + x_lens = (((x_lens - 1) >> 1) - 1) >> 1 + x_lens -= 2 + assert x.size(0) == x_lens.max().item() + + num_processed_frames = num_processed_frames >> 2 + + output, output_lengths, output_states = self.encoder.infer( + x, x_lens, num_processed_frames, states + ) + + output = output.permute(1, 0, 2) # (T, N, C) -> (N, T, C) + + return output, output_lengths, output_states + + def init_states(self, device: torch.device = torch.device("cpu")): + """Create initial states.""" + return self.encoder.init_states(device) + + +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-1)//2 - 1)//2, which approximates T' == T//4 + + 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, + ) -> 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-1)//2 - 1)//2, out_channels) + layer1_channels: + Number of channels in layer1 + layer1_channels: + Number of channels in layer2 + """ + assert in_channels >= 7 + super().__init__() + + self.conv = nn.Sequential( + ScaledConv2d( + in_channels=1, + out_channels=layer1_channels, + kernel_size=3, + padding=1, + ), + ActivationBalancer(channel_dim=1), + DoubleSwish(), + ScaledConv2d( + in_channels=layer1_channels, + out_channels=layer2_channels, + kernel_size=3, + stride=2, + ), + ActivationBalancer(channel_dim=1), + DoubleSwish(), + ScaledConv2d( + in_channels=layer2_channels, + out_channels=layer3_channels, + kernel_size=3, + stride=2, + ), + ActivationBalancer(channel_dim=1), + DoubleSwish(), + ) + self.out = ScaledLinear( + layer3_channels * (((in_channels - 1) // 2 - 1) // 2), out_channels + ) + # set learn_eps=False because out_norm is preceded by `out`, and `out` + # itself has learned scale, so the extra degree of freedom is not + # needed. + self.out_norm = BasicNorm(out_channels, learn_eps=False) + # constrain median of output to be close to zero. + self.out_balancer = ActivationBalancer( + channel_dim=-1, min_positive=0.45, max_positive=0.55 + ) + + 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-1)//2 - 1)//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-1)//2 - 1)//2, ((idim-1)//2 - 1)//2) + b, c, t, f = x.size() + x = self.out(x.transpose(1, 2).contiguous().view(b, t, c * f)) + # Now x is of shape (N, ((T-1)//2 - 1))//2, odim) + x = self.out_norm(x) + x = self.out_balancer(x) + return x diff --git a/egs/librispeech/ASR/conv_emformer_transducer_stateless3/encoder_interface.py b/egs/librispeech/ASR/conv_emformer_transducer_stateless3/encoder_interface.py new file mode 120000 index 000000000..ee2f09151 --- /dev/null +++ b/egs/librispeech/ASR/conv_emformer_transducer_stateless3/encoder_interface.py @@ -0,0 +1 @@ +../conv_emformer_transducer_stateless/encoder_interface.py \ No newline at end of file diff --git a/egs/librispeech/ASR/conv_emformer_transducer_stateless3/export.py b/egs/librispeech/ASR/conv_emformer_transducer_stateless3/export.py new file mode 100755 index 000000000..ab15e0241 --- /dev/null +++ b/egs/librispeech/ASR/conv_emformer_transducer_stateless3/export.py @@ -0,0 +1,287 @@ +#!/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: +./conv_emformer_transducer_stateless2/export.py \ + --exp-dir ./conv_emformer_transducer_stateless2/exp \ + --bpe-model data/lang_bpe_500/bpe.model \ + --epoch 30 \ + --avg 10 \ + --use-averaged-model=True \ + --num-encoder-layers 12 \ + --chunk-length 32 \ + --cnn-module-kernel 31 \ + --left-context-length 32 \ + --right-context-length 8 \ + --memory-size 32 \ + --jit False + +It will generate a file exp_dir/pretrained.pt + +To use the generated file with `conv_emformer_transducer_stateless2/decode.py`, +you can do: + + cd /path/to/exp_dir + ln -s pretrained.pt epoch-9999.pt + + cd /path/to/egs/librispeech/ASR + ./conv_emformer_transducer_stateless2/decode.py \ + --exp-dir ./conv_emformer_transducer_stateless2/exp \ + --epoch 9999 \ + --avg 1 \ + --max-duration 100 \ + --bpe-model data/lang_bpe_500/bpe.model \ + --use-averaged-model=False \ + --num-encoder-layers 12 \ + --chunk-length 32 \ + --cnn-module-kernel 31 \ + --left-context-length 32 \ + --right-context-length 8 \ + --memory-size 32 +""" + +import argparse +import logging +from pathlib import Path + +import sentencepiece as spm +import torch +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=28, + help="""It specifies the checkpoint to use for averaging. + Note: Epoch counts from 0. + 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( + "--exp-dir", + type=str, + default="pruned_transducer_stateless2/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. + """, + ) + + 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( + "--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. ", + ) + + add_model_arguments(parser) + + return parser + + +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") + + 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) + + 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.eval() + + if params.jit: + # 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 torch.jit.script") + # 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/librispeech/ASR/conv_emformer_transducer_stateless3/joiner.py b/egs/librispeech/ASR/conv_emformer_transducer_stateless3/joiner.py new file mode 120000 index 000000000..1eb4dcc83 --- /dev/null +++ b/egs/librispeech/ASR/conv_emformer_transducer_stateless3/joiner.py @@ -0,0 +1 @@ +../conv_emformer_transducer_stateless/joiner.py \ No newline at end of file diff --git a/egs/librispeech/ASR/conv_emformer_transducer_stateless3/model.py b/egs/librispeech/ASR/conv_emformer_transducer_stateless3/model.py new file mode 120000 index 000000000..322b694e0 --- /dev/null +++ b/egs/librispeech/ASR/conv_emformer_transducer_stateless3/model.py @@ -0,0 +1 @@ +../conv_emformer_transducer_stateless/model.py \ No newline at end of file diff --git a/egs/librispeech/ASR/conv_emformer_transducer_stateless3/optim.py b/egs/librispeech/ASR/conv_emformer_transducer_stateless3/optim.py new file mode 120000 index 000000000..8f19a99da --- /dev/null +++ b/egs/librispeech/ASR/conv_emformer_transducer_stateless3/optim.py @@ -0,0 +1 @@ +../conv_emformer_transducer_stateless/optim.py \ No newline at end of file diff --git a/egs/librispeech/ASR/conv_emformer_transducer_stateless3/scaling.py b/egs/librispeech/ASR/conv_emformer_transducer_stateless3/scaling.py new file mode 120000 index 000000000..12f22cf9c --- /dev/null +++ b/egs/librispeech/ASR/conv_emformer_transducer_stateless3/scaling.py @@ -0,0 +1 @@ +../conv_emformer_transducer_stateless/scaling.py \ No newline at end of file diff --git a/egs/librispeech/ASR/conv_emformer_transducer_stateless3/stream.py b/egs/librispeech/ASR/conv_emformer_transducer_stateless3/stream.py new file mode 120000 index 000000000..bf9cbbe2e --- /dev/null +++ b/egs/librispeech/ASR/conv_emformer_transducer_stateless3/stream.py @@ -0,0 +1 @@ +../conv_emformer_transducer_stateless/stream.py \ No newline at end of file diff --git a/egs/librispeech/ASR/conv_emformer_transducer_stateless3/streaming_decode.py b/egs/librispeech/ASR/conv_emformer_transducer_stateless3/streaming_decode.py new file mode 100755 index 000000000..0f687898f --- /dev/null +++ b/egs/librispeech/ASR/conv_emformer_transducer_stateless3/streaming_decode.py @@ -0,0 +1,980 @@ +#!/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: +(1) greedy search +./conv_emformer_transducer_stateless2/streaming_decode.py \ + --epoch 30 \ + --avg 10 \ + --exp-dir conv_emformer_transducer_stateless2/exp \ + --num-decode-streams 2000 \ + --num-encoder-layers 12 \ + --chunk-length 32 \ + --cnn-module-kernel 31 \ + --left-context-length 32 \ + --right-context-length 8 \ + --memory-size 32 \ + --decoding-method greedy_search \ + --use-averaged-model True + +(2) modified beam search +./conv_emformer_transducer_stateless2/streaming_decode.py \ + --epoch 30 \ + --avg 10 \ + --exp-dir conv_emformer_transducer_stateless2/exp \ + --num-decode-streams 2000 \ + --num-encoder-layers 12 \ + --chunk-length 32 \ + --cnn-module-kernel 31 \ + --left-context-length 32 \ + --right-context-length 8 \ + --memory-size 32 \ + --decoding-method modified_beam_search \ + --use-averaged-model True \ + --beam-size 4 + +(3) fast beam search +./conv_emformer_transducer_stateless2/streaming_decode.py \ + --epoch 30 \ + --avg 10 \ + --exp-dir conv_emformer_transducer_stateless2/exp \ + --num-decode-streams 2000 \ + --num-encoder-layers 12 \ + --chunk-length 32 \ + --cnn-module-kernel 31 \ + --left-context-length 32 \ + --right-context-length 8 \ + --memory-size 32 \ + --decoding-method fast_beam_search \ + --use-averaged-model True \ + --beam 4 \ + --max-contexts 4 \ + --max-states 8 +""" +import argparse +import logging +import warnings +from pathlib import Path +from typing import Dict, List, Optional, Tuple + +import k2 +from lhotse import CutSet +import numpy as np +import sentencepiece as spm +import torch +import torch.nn as nn +from asr_datamodule import LibriSpeechAsrDataModule +from beam_search import Hypothesis, HypothesisList, get_hyps_shape +from emformer import LOG_EPSILON, stack_states, unstack_states +from kaldifeat import Fbank, FbankOptions +from stream import Stream +from torch.nn.utils.rnn import pad_sequence +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 one_best_decoding +from icefall.utils import ( + AttributeDict, + get_texts, + setup_logger, + store_transcripts, + str2bool, + write_error_stats, +) + + +def get_parser(): + parser = argparse.ArgumentParser( + formatter_class=argparse.ArgumentDefaultsHelpFormatter + ) + + parser.add_argument( + "--epoch", + type=int, + default=28, + help="It specifies the checkpoint to use for decoding." + "Note: Epoch counts from 0.", + ) + + 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'. ", + ) + + parser.add_argument( + "--use-averaged-model", + type=str2bool, + default=False, + 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="transducer_emformer/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( + "--decoding-method", + type=str, + default="greedy_search", + help="""Possible values are: + - greedy_search + - modified_beam_search + - fast_beam_search + """, + ) + + parser.add_argument( + "--beam-size", + type=int, + default=4, + help="""An interger 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=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 --decoding-method is fast_beam_search""", + ) + + parser.add_argument( + "--max-contexts", + type=int, + default=4, + help="""Used only when --decoding-method is + fast_beam_search""", + ) + + parser.add_argument( + "--max-states", + type=int, + default=8, + help="""Used only when --decoding-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 --decoding_method is greedy_search""", + ) + + parser.add_argument( + "--sampling-rate", + type=float, + default=16000, + help="Sample rate of the audio", + ) + + parser.add_argument( + "--num-decode-streams", + type=int, + default=2000, + help="The number of streams that can be decoded parallel", + ) + + add_model_arguments(parser) + + return parser + + +def greedy_search( + model: nn.Module, + encoder_out: torch.Tensor, + streams: List[Stream], +) -> None: + """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. + streams: + A list of Stream objects. + """ + assert len(streams) == encoder_out.size(0) + assert encoder_out.ndim == 3 + + blank_id = model.decoder.blank_id + context_size = model.decoder.context_size + device = next(model.parameters()).device + T = encoder_out.size(1) + + encoder_out = model.joiner.encoder_proj(encoder_out) + + decoder_input = torch.tensor( + [stream.hyp[-context_size:] for stream in streams], + device=device, + dtype=torch.int64, + ) + # decoder_out is of shape (batch_size, 1, decoder_out_dim) + decoder_out = model.decoder(decoder_input, need_pad=False) + decoder_out = model.joiner.decoder_proj(decoder_out) + + for t in range(T): + # current_encoder_out's shape: (batch_size, 1, encoder_out_dim) + current_encoder_out = encoder_out[:, t : t + 1, :] # noqa + + logits = model.joiner( + current_encoder_out.unsqueeze(2), + decoder_out.unsqueeze(1), + project_input=False, + ) + # logits'shape (batch_size, vocab_size) + logits = logits.squeeze(1).squeeze(1) + + assert logits.ndim == 2, logits.shape + y = logits.argmax(dim=1).tolist() + emitted = False + for i, v in enumerate(y): + if v != blank_id: + streams[i].hyp.append(v) + emitted = True + if emitted: + # update decoder output + decoder_input = torch.tensor( + [stream.hyp[-context_size:] for stream in streams], + device=device, + dtype=torch.int64, + ) + decoder_out = model.decoder( + decoder_input, + need_pad=False, + ) + decoder_out = model.joiner.decoder_proj(decoder_out) + + +def modified_beam_search( + model: nn.Module, + encoder_out: torch.Tensor, + streams: List[Stream], + beam: int = 4, +): + """Beam search in batch mode with --max-sym-per-frame=1 being hardcoded. + + Args: + model: + The RNN-T model. + encoder_out: + A 3-D tensor of shape (N, T, encoder_out_dim) containing the output of + the encoder model. + streams: + A list of stream objects. + beam: + Number of active paths during the beam search. + """ + assert encoder_out.ndim == 3, encoder_out.shape + assert len(streams) == encoder_out.size(0) + + blank_id = model.decoder.blank_id + context_size = model.decoder.context_size + device = next(model.parameters()).device + batch_size = len(streams) + T = encoder_out.size(1) + + B = [stream.hyps for stream in streams] + + encoder_out = model.joiner.encoder_proj(encoder_out) + + for t in range(T): + current_encoder_out = encoder_out[:, t].unsqueeze(1).unsqueeze(1) + # current_encoder_out's shape: (batch_size, 1, 1, encoder_out_dim) + + 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.stack( + [hyp.log_prob.reshape(1) for hyps in A for hyp in hyps], dim=0 + ) # (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, decoder_output_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, 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) + + 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 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 != blank_id: + new_ys.append(new_token) + + new_log_prob = topk_log_probs[k] + new_hyp = Hypothesis(ys=new_ys, log_prob=new_log_prob) + B[i].add(new_hyp) + + for i in range(batch_size): + streams[i].hyps = B[i] + + +def fast_beam_search_one_best( + model: nn.Module, + streams: List[Stream], + encoder_out: torch.Tensor, + processed_lens: torch.Tensor, + beam: float, + max_states: int, + max_contexts: int, +) -> None: + """It limits the maximum number of symbols per frame to 1. + + A lattice is first obtained using modified beam search, and then + the shortest path within the lattice is used as the final output. + + Args: + model: + An instance of `Transducer`. + streams: + A list of stream objects. + encoder_out: + A tensor of shape (N, T, C) from the encoder. + processed_lens: + A tensor of shape (N,) containing the number of processed 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. + """ + assert encoder_out.ndim == 3 + + context_size = model.decoder.context_size + vocab_size = model.decoder.vocab_size + + B, T, C = encoder_out.shape + assert B == len(streams) + + 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(streams[i].rnnt_decoding_stream) + 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.log_softmax(dim=-1) + decoding_streams.advance(log_probs) + + decoding_streams.terminate_and_flush_to_streams() + + lattice = decoding_streams.format_output(processed_lens.tolist()) + + best_path = one_best_decoding(lattice) + hyps = get_texts(best_path) + + for i in range(B): + streams[i].hyp = hyps[i] + + +def decode_one_chunk( + model: nn.Module, + streams: List[Stream], + params: AttributeDict, + decoding_graph: Optional[k2.Fsa] = None, +) -> List[int]: + """ + Args: + model: + The Transducer model. + streams: + A list of Stream objects. + params: + It is returned by :func:`get_params`. + decoding_graph: + The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used + only when --decoding_method is fast_beam_search. + + Returns: + A list of indexes indicating the finished streams. + """ + device = next(model.parameters()).device + + feature_list = [] + feature_len_list = [] + state_list = [] + num_processed_frames_list = [] + + for stream in streams: + # We should first get `stream.num_processed_frames` + # before calling `stream.get_feature_chunk()` + # since `stream.num_processed_frames` would be updated + num_processed_frames_list.append(stream.num_processed_frames) + feature = stream.get_feature_chunk() + feature_len = feature.size(0) + feature_list.append(feature) + feature_len_list.append(feature_len) + state_list.append(stream.states) + + features = pad_sequence( + feature_list, batch_first=True, padding_value=LOG_EPSILON + ).to(device) + feature_lens = torch.tensor(feature_len_list, device=device) + num_processed_frames = torch.tensor( + num_processed_frames_list, device=device + ) + + # Make sure it has at least 1 frame after subsampling, first-and-last-frame cutting, and right context cutting # noqa + tail_length = ( + 3 * params.subsampling_factor + params.right_context_length + 3 + ) + if features.size(1) < tail_length: + pad_length = tail_length - features.size(1) + feature_lens += pad_length + features = torch.nn.functional.pad( + features, + (0, 0, 0, pad_length), + mode="constant", + value=LOG_EPSILON, + ) + + # Stack states of all streams + states = stack_states(state_list) + + encoder_out, encoder_out_lens, states = model.encoder.infer( + x=features, + x_lens=feature_lens, + states=states, + num_processed_frames=num_processed_frames, + ) + + if params.decoding_method == "greedy_search": + greedy_search( + model=model, + streams=streams, + encoder_out=encoder_out, + ) + elif params.decoding_method == "modified_beam_search": + modified_beam_search( + model=model, + streams=streams, + encoder_out=encoder_out, + beam=params.beam_size, + ) + elif params.decoding_method == "fast_beam_search": + # feature_len is needed to get partial results. + # The rnnt_decoding_stream for fast_beam_search. + fast_beam_search_one_best( + model=model, + streams=streams, + encoder_out=encoder_out, + processed_lens=(num_processed_frames >> 2) + encoder_out_lens, + beam=params.beam, + max_contexts=params.max_contexts, + max_states=params.max_states, + ) + else: + raise ValueError( + f"Unsupported decoding method: {params.decoding_method}" + ) + + # Update cached states of each stream + state_list = unstack_states(states) + for i, s in enumerate(state_list): + streams[i].states = s + + finished_streams = [i for i, stream in enumerate(streams) if stream.done] + return finished_streams + + +def create_streaming_feature_extractor() -> Fbank: + """Create a CPU streaming feature extractor. + + At present, we assume it returns a fbank feature extractor with + fixed options. In the future, we will support passing in the options + from outside. + + Returns: + Return a CPU streaming feature extractor. + """ + opts = FbankOptions() + opts.device = "cpu" + opts.frame_opts.dither = 0 + opts.frame_opts.snip_edges = False + opts.frame_opts.samp_freq = 16000 + opts.mel_opts.num_bins = 80 + return Fbank(opts) + + +def decode_dataset( + cuts: CutSet, + model: nn.Module, + params: AttributeDict, + sp: spm.SentencePieceProcessor, + decoding_graph: Optional[k2.Fsa] = None, +): + """Decode dataset. + + Args: + cuts: + Lhotse Cutset containing the dataset to decode. + params: + It is returned by :func:`get_params`. + model: + The Transducer model. + sp: + The BPE model. + decoding_graph: + The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used + only when --decoding_method is fast_beam_search. + + 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. + """ + device = next(model.parameters()).device + + log_interval = 300 + + fbank = create_streaming_feature_extractor() + + decode_results = [] + streams = [] + for num, cut in enumerate(cuts): + # Each utterance has a Stream. + stream = Stream( + params=params, + decoding_graph=decoding_graph, + device=device, + LOG_EPS=LOG_EPSILON, + ) + + stream.set_states(model.encoder.init_states(device)) + + audio: np.ndarray = cut.load_audio() + # audio.shape: (1, num_samples) + assert len(audio.shape) == 2 + assert audio.shape[0] == 1, "Should be single channel" + assert audio.dtype == np.float32, audio.dtype + # The trained model is using normalized samples + assert audio.max() <= 1, "Should be normalized to [-1, 1])" + + samples = torch.from_numpy(audio).squeeze(0) + feature = fbank(samples) + stream.set_feature(feature) + stream.set_ground_truth(cut.supervisions[0].text) + + streams.append(stream) + + while len(streams) >= params.num_decode_streams: + finished_streams = decode_one_chunk( + model=model, + streams=streams, + params=params, + decoding_graph=decoding_graph, + ) + + for i in sorted(finished_streams, reverse=True): + decode_results.append( + ( + streams[i].ground_truth.split(), + sp.decode(streams[i].decoding_result()).split(), + ) + ) + del streams[i] + + if num % log_interval == 0: + logging.info(f"Cuts processed until now is {num}.") + + while len(streams) > 0: + finished_streams = decode_one_chunk( + model=model, + streams=streams, + params=params, + decoding_graph=decoding_graph, + ) + + for i in sorted(finished_streams, reverse=True): + decode_results.append( + ( + streams[i].ground_truth.split(), + sp.decode(streams[i].decoding_result()).split(), + ) + ) + del streams[i] + + if params.decoding_method == "greedy_search": + key = "greedy_search" + elif params.decoding_method == "fast_beam_search": + key = ( + f"beam_{params.beam}_" + f"max_contexts_{params.max_contexts}_" + f"max_states_{params.max_states}" + ) + else: + key = f"beam_size_{params.beam_size}" + + return {key: decode_results} + + +def save_results( + params: AttributeDict, + test_set_name: str, + results_dict: Dict[str, List[Tuple[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" + ) + store_transcripts(filename=recog_path, texts=sorted(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", + "fast_beam_search", + "modified_beam_search", + ) + params.res_dir = params.exp_dir / "streaming" / 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}" + + # for streaming + params.suffix += f"-streaming-chunk-length-{params.chunk_length}" + params.suffix += f"-left-context-length-{params.left_context_length}" + params.suffix += f"-right-context-length-{params.right_context_length}" + params.suffix += f"-memory-size-{params.memory_size}" + + 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}" + 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-streaming-decode") + 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() + + params.device = device + + logging.info(params) + + 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.eval() + + if params.decoding_method == "fast_beam_search": + decoding_graph = k2.trivial_graph(params.vocab_size - 1, device=device) + else: + decoding_graph = None + + num_param = sum([p.numel() for p in model.parameters()]) + logging.info(f"Number of model parameters: {num_param}") + + librispeech = LibriSpeechAsrDataModule(args) + + test_clean_cuts = librispeech.test_clean_cuts() + test_other_cuts = librispeech.test_other_cuts() + + test_sets = ["test-clean", "test-other"] + test_cuts = [test_clean_cuts, test_other_cuts] + + for test_set, test_cut in zip(test_sets, test_cuts): + results_dict = decode_dataset( + cuts=test_cut, + model=model, + params=params, + sp=sp, + decoding_graph=decoding_graph, + ) + + save_results( + params=params, + test_set_name=test_set, + results_dict=results_dict, + ) + + logging.info("Done!") + + +if __name__ == "__main__": + torch.manual_seed(20220410) + main() diff --git a/egs/librispeech/ASR/conv_emformer_transducer_stateless3/test_emformer.py b/egs/librispeech/ASR/conv_emformer_transducer_stateless3/test_emformer.py new file mode 100644 index 000000000..8cde6205b --- /dev/null +++ b/egs/librispeech/ASR/conv_emformer_transducer_stateless3/test_emformer.py @@ -0,0 +1,194 @@ +#!/usr/bin/env python3 +# +# Copyright 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. + + +import torch +from emformer import ConvolutionModule, Emformer, stack_states, unstack_states + + +def test_convolution_module_forward(): + B, D = 2, 256 + chunk_length = 4 + right_context_length = 2 + num_chunks = 3 + U = num_chunks * chunk_length + R = num_chunks * right_context_length + kernel_size = 31 + conv_module = ConvolutionModule( + chunk_length, + right_context_length, + D, + kernel_size, + ) + + utterance = torch.randn(U, B, D) + right_context = torch.randn(R, B, D) + + utterance, right_context = conv_module(utterance, right_context) + assert utterance.shape == (U, B, D), utterance.shape + assert right_context.shape == (R, B, D), right_context.shape + + +def test_convolution_module_infer(): + from emformer import ConvolutionModule + + B, D = 2, 256 + chunk_length = 4 + right_context_length = 2 + num_chunks = 1 + U = num_chunks * chunk_length + R = num_chunks * right_context_length + kernel_size = 31 + conv_module = ConvolutionModule( + chunk_length, + right_context_length, + D, + kernel_size, + ) + + utterance = torch.randn(U, B, D) + right_context = torch.randn(R, B, D) + cache = torch.randn(B, D, kernel_size - 1) + + utterance, right_context, new_cache = conv_module.infer( + utterance, right_context, cache + ) + assert utterance.shape == (U, B, D), utterance.shape + assert right_context.shape == (R, B, D), right_context.shape + assert new_cache.shape == (B, D, kernel_size - 1), new_cache.shape + + +def test_state_stack_unstack(): + num_features = 80 + chunk_length = 32 + encoder_dim = 512 + num_encoder_layers = 2 + kernel_size = 31 + left_context_length = 32 + right_context_length = 8 + memory_size = 32 + + model = Emformer( + num_features=num_features, + chunk_length=chunk_length, + subsampling_factor=4, + d_model=encoder_dim, + num_encoder_layers=num_encoder_layers, + cnn_module_kernel=kernel_size, + left_context_length=left_context_length, + right_context_length=right_context_length, + memory_size=memory_size, + ) + + for batch_size in [1, 2]: + attn_caches = [ + [ + torch.zeros(memory_size, batch_size, encoder_dim), + torch.zeros(left_context_length // 4, batch_size, encoder_dim), + torch.zeros( + left_context_length // 4, + batch_size, + encoder_dim, + ), + ] + for _ in range(num_encoder_layers) + ] + conv_caches = [ + torch.zeros(batch_size, encoder_dim, kernel_size - 1) + for _ in range(num_encoder_layers) + ] + states = [attn_caches, conv_caches] + x = torch.randn(batch_size, 23, num_features) + x_lens = torch.full((batch_size,), 23) + num_processed_frames = torch.full((batch_size,), 0) + y, y_lens, states = model.infer( + x, x_lens, num_processed_frames=num_processed_frames, states=states + ) + + state_list = unstack_states(states) + states2 = stack_states(state_list) + + for ss, ss2 in zip(states[0], states2[0]): + for s, s2 in zip(ss, ss2): + assert torch.allclose(s, s2), f"{s.sum()}, {s2.sum()}" + + for s, s2 in zip(states[1], states2[1]): + assert torch.allclose(s, s2), f"{s.sum()}, {s2.sum()}" + + +def test_torchscript_consistency_infer(): + r"""Verify that scripting Emformer does not change the behavior of method `infer`.""" # noqa + num_features = 80 + chunk_length = 32 + encoder_dim = 512 + num_encoder_layers = 2 + kernel_size = 31 + left_context_length = 32 + right_context_length = 8 + memory_size = 32 + batch_size = 2 + + model = Emformer( + num_features=num_features, + chunk_length=chunk_length, + subsampling_factor=4, + d_model=encoder_dim, + num_encoder_layers=num_encoder_layers, + cnn_module_kernel=kernel_size, + left_context_length=left_context_length, + right_context_length=right_context_length, + memory_size=memory_size, + ).eval() + attn_caches = [ + [ + torch.zeros(memory_size, batch_size, encoder_dim), + torch.zeros(left_context_length // 4, batch_size, encoder_dim), + torch.zeros( + left_context_length // 4, + batch_size, + encoder_dim, + ), + ] + for _ in range(num_encoder_layers) + ] + conv_caches = [ + torch.zeros(batch_size, encoder_dim, kernel_size - 1) + for _ in range(num_encoder_layers) + ] + states = [attn_caches, conv_caches] + x = torch.randn(batch_size, 23, num_features) + x_lens = torch.full((batch_size,), 23) + num_processed_frames = torch.full((batch_size,), 0) + y, y_lens, out_states = model.infer( + x, x_lens, num_processed_frames=num_processed_frames, states=states + ) + + sc_model = torch.jit.script(model).eval() + sc_y, sc_y_lens, sc_out_states = sc_model.infer( + x, x_lens, num_processed_frames=num_processed_frames, states=states + ) + + assert torch.allclose(y, sc_y) + + +if __name__ == "__main__": + test_convolution_module_forward() + test_convolution_module_infer() + test_state_stack_unstack() + test_torchscript_consistency_infer() diff --git a/egs/librispeech/ASR/conv_emformer_transducer_stateless3/train.py b/egs/librispeech/ASR/conv_emformer_transducer_stateless3/train.py new file mode 100755 index 000000000..dfe1b6136 --- /dev/null +++ b/egs/librispeech/ASR/conv_emformer_transducer_stateless3/train.py @@ -0,0 +1,1136 @@ +#!/usr/bin/env python3 +# Copyright 2021 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" + +./conv_emformer_transducer_stateless2/train.py \ + --world-size 4 \ + --num-epochs 30 \ + --start-epoch 1 \ + --exp-dir conv_emformer_transducer_stateless2/exp \ + --full-libri 1 \ + --max-duration 280 \ + --master-port 12321 \ + --num-encoder-layers 12 \ + --chunk-length 32 \ + --cnn-module-kernel 31 \ + --left-context-length 32 \ + --right-context-length 8 \ + --memory-size 32 + +# For mix precision training: +./conv_emformer_transducer_stateless2/train.py \ + --world-size 4 \ + --num-epochs 30 \ + --start-epoch 1 \ + --use-fp16 1 \ + --exp-dir conv_emformer_transducer_stateless2/exp \ + --full-libri 1 \ + --max-duration 300 \ + --master-port 12321 \ + --num-encoder-layers 12 \ + --chunk-length 32 \ + --cnn-module-kernel 31 \ + --left-context-length 32 \ + --right-context-length 8 \ + --memory-size 32 +""" + + +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 emformer import Emformer +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, Eve +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 icefall import diagnostics +from icefall.checkpoint import load_checkpoint, remove_checkpoints +from icefall.checkpoint import save_checkpoint as save_checkpoint_impl +from icefall.checkpoint import ( + save_checkpoint_with_global_batch_idx, + update_averaged_model, +) +from icefall.dist import cleanup_dist, setup_dist +from icefall.env import get_env_info +from icefall.utils import AttributeDict, MetricsTracker, setup_logger, str2bool + +LRSchedulerType = Union[ + torch.optim.lr_scheduler._LRScheduler, optim.LRScheduler +] + + +def add_model_arguments(parser: argparse.ArgumentParser): + parser.add_argument( + "--encoder-dim", + type=int, + default=512, + help="Attention dim for the Emformer", + ) + + parser.add_argument( + "--nhead", + type=int, + default=8, + help="Number of attention heads for the Emformer", + ) + + parser.add_argument( + "--dim-feedforward", + type=int, + default=2048, + help="Feed-forward dimension for the Emformer", + ) + + parser.add_argument( + "--num-encoder-layers", + type=int, + default=12, + help="Number of encoder layers for the Emformer", + ) + + parser.add_argument( + "--cnn-module-kernel", + type=int, + default=31, + help="Kernel size for the convolution module.", + ) + + parser.add_argument( + "--left-context-length", + type=int, + default=32, + help="""Number of frames before subsampling for left context + in the Emformer.""", + ) + + parser.add_argument( + "--chunk-length", + type=int, + default=32, + help="""Number of frames before subsampling for each chunk + in the Emformer.""", + ) + + parser.add_argument( + "--right-context-length", + type=int, + default=8, + help="""Number of frames before subsampling for right context + in the Emformer.""", + ) + + parser.add_argument( + "--memory-size", + type=int, + default=0, + help="Number of entries in the memory for the Emformer", + ) + + +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_stateless2/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( + "--initial-lr", + type=float, + default=0.003, + help="""The initial learning rate. This value should not need to be + changed.""", + ) + + 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=6, + 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( + "--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( + "--save-every-n", + type=int, + default=8000, + 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=20, + 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=100, + 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=False, + help="Whether to use half precision training.", + ) + + add_model_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 warm_step for Noam optimizer. + """ + 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 Emformer + "feature_dim": 80, + "subsampling_factor": 4, + # parameters for decoder + "decoder_dim": 512, + # parameters for joiner + "joiner_dim": 512, + # parameters for Noam + "model_warm_step": 3000, # arg given to model, not for lrate + "env_info": get_env_info(), + } + ) + + return params + + +def get_encoder_model(params: AttributeDict) -> nn.Module: + # TODO: We can add an option to switch between Conformer and Transformer + encoder = Emformer( + num_features=params.feature_dim, + chunk_length=params.chunk_length, + subsampling_factor=params.subsampling_factor, + d_model=params.encoder_dim, + nhead=params.nhead, + dim_feedforward=params.dim_feedforward, + num_encoder_layers=params.num_encoder_layers, + cnn_module_kernel=params.cnn_module_kernel, + left_context_length=params.left_context_length, + right_context_length=params.right_context_length, + memory_size=params.memory_size, + ) + 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, + 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, + 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" + 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, + ) + + 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, + warmup: float = 1.0, +) -> Tuple[Tensor, MetricsTracker]: + """ + Compute RNN-T 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 Conformer 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 == 3 + feature = feature.to(device) + + supervisions = batch["supervisions"] + feature_lens = supervisions["num_frames"].to(device) + + texts = batch["supervisions"]["text"] + y = sp.encode(texts, out_type=int) + y = k2.RaggedTensor(y).to(device) + + with torch.set_grad_enabled(is_training): + simple_loss, pruned_loss = model( + x=feature, + x_lens=feature_lens, + y=y, + prune_range=params.prune_range, + am_scale=params.am_scale, + lm_scale=params.lm_scale, + warmup=warmup, + ) + # after the main warmup step, we keep pruned_loss_scale small + # for the same amount of time (model_warm_step), to avoid + # overwhelming the simple_loss and causing it to diverge, + # in case it had not fully learned the alignment yet. + pruned_loss_scale = ( + 0.0 + if warmup < 1.0 + else (0.1 if warmup > 1.0 and warmup < 2.0 else 1.0) + ) + loss = ( + params.simple_loss_scale * simple_loss + + pruned_loss_scale * pruned_loss + ) + + assert loss.requires_grad == is_training + + info = MetricsTracker() + 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["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["frames"] + 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, + scheduler: 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, +) -> 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) + + for batch_idx, batch in enumerate(train_dl): + if batch_idx < cur_batch_idx: + continue + cur_batch_idx = batch_idx + + params.batch_idx_train += 1 + batch_size = len(batch["supervisions"]["text"]) + + 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, + warmup=(params.batch_idx_train / params.model_warm_step), + ) + # 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() + scheduler.step_batch(params.batch_idx_train) + scaler.step(optimizer) + scaler.update() + optimizer.zero_grad() + + 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 % params.log_interval == 0: + cur_lr = scheduler.get_last_lr()[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}" + ) + + if tb_writer is not None: + 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 batch_idx > 0 and batch_idx % params.valid_interval == 0: + 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}") + if tb_writer is not None: + valid_info.write_summary( + tb_writer, "train/valid_", params.batch_idx_train + ) + + loss_value = tot_loss["loss"] / tot_loss["frames"] + 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): + """ + 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)) + if params.full_libri is False: + params.valid_interval = 1600 + + 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() 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) + + 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]) + + optimizer = Eve(model.parameters(), lr=params.initial_lr) + + scheduler = Eden(optimizer, params.lr_batches, params.lr_epochs) + + if checkpoints and "optimizer" in checkpoints: + logging.info("Loading optimizer state dict") + optimizer.load_state_dict(checkpoints["optimizer"]) + + if ( + checkpoints + and "scheduler" in checkpoints + and checkpoints["scheduler"] is not None + ): + 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) + + 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) + 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): + 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, + ) + + 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 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: + # warmup = 0.0 is so that the derivs for the pruned loss stay zero + # (i.e. are not remembered by the decaying-average in adam), because + # we want to avoid these params being subject to shrinkage in adam. + with torch.cuda.amp.autocast(enabled=params.use_fp16): + loss, _ = compute_loss( + params=params, + model=model, + sp=sp, + batch=batch, + is_training=True, + warmup=0.0, + ) + loss.backward() + optimizer.step() + optimizer.zero_grad() + except RuntimeError 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]}) ..." + ) + raise + + +def main(): + parser = get_parser() + LibriSpeechAsrDataModule.add_arguments(parser) + args = parser.parse_args() + args.exp_dir = Path(args.exp_dir) + + world_size = args.world_size + assert world_size >= 1 + if world_size > 1: + mp.spawn(run, args=(world_size, args), nprocs=world_size, join=True) + else: + run(rank=0, world_size=1, args=args) + + +torch.set_num_threads(1) +torch.set_num_interop_threads(1) + +if __name__ == "__main__": + main() From 630626a092234f65bd827b4d7689da7bc791c84d Mon Sep 17 00:00:00 2001 From: yaozengwei Date: Sun, 26 Jun 2022 21:39:06 +0800 Subject: [PATCH 10/12] support position encoding --- .flake8 | 3 +- .../emformer.py | 386 ++++++++++++++++-- .../test_emformer.py | 44 +- 3 files changed, 394 insertions(+), 39 deletions(-) diff --git a/.flake8 b/.flake8 index 9dd8d6207..d67fc1542 100644 --- a/.flake8 +++ b/.flake8 @@ -9,8 +9,7 @@ per-file-ignores = egs/*/ASR/pruned_transducer_stateless*/*.py: E501, egs/*/ASR/*/optim.py: E501, egs/*/ASR/*/scaling.py: E501, - egs/librispeech/ASR/conv_emformer_transducer_stateless/*.py: E501, E203 - egs/librispeech/ASR/conv_emformer_transducer_stateless2/*.py: E501, E203 + egs/librispeech/ASR/conv_emformer_transducer_stateless*/emformer.py: E501, E203 # invalid escape sequence (cause by tex formular), W605 icefall/utils.py: E501, W605 diff --git a/egs/librispeech/ASR/conv_emformer_transducer_stateless3/emformer.py b/egs/librispeech/ASR/conv_emformer_transducer_stateless3/emformer.py index e3a598b0e..06fc880df 100644 --- a/egs/librispeech/ASR/conv_emformer_transducer_stateless3/emformer.py +++ b/egs/librispeech/ASR/conv_emformer_transducer_stateless3/emformer.py @@ -35,7 +35,6 @@ from scaling import ( from icefall.utils import make_pad_mask - LOG_EPSILON = math.log(1e-10) @@ -434,6 +433,10 @@ class EmformerAttention(nn.Module): r"""Emformer layer attention module. Args: + chunk_length (int): + Length of chunk. + right_context_length (int): + Length of right context. embed_dim (int): Embedding dimension. nhead (int): @@ -448,6 +451,8 @@ class EmformerAttention(nn.Module): def __init__( self, + chunk_length: int, + right_context_length: int, embed_dim: int, nhead: int, dropout: float = 0.0, @@ -455,6 +460,8 @@ class EmformerAttention(nn.Module): negative_inf: float = -1e8, ): super().__init__() + self.chunk_length = chunk_length + self.right_context_length = right_context_length if embed_dim % nhead != 0: raise ValueError( @@ -477,6 +484,26 @@ class EmformerAttention(nn.Module): embed_dim, embed_dim, bias=True, initial_scale=0.25 ) + # linear transformation for positional encoding. + self.linear_pos = ScaledLinear(embed_dim, embed_dim, bias=False) + # these two learnable bias are used in matrix c and matrix d + # as described in "Transformer-XL: Attentive Language Models Beyond a Fixed-Length Context" Section 3.3 # noqa + self.pos_bias_u = nn.Parameter(torch.Tensor(nhead, self.head_dim)) + self.pos_bias_v = nn.Parameter(torch.Tensor(nhead, self.head_dim)) + self.pos_bias_u_scale = nn.Parameter(torch.zeros(()).detach()) + self.pos_bias_v_scale = nn.Parameter(torch.zeros(()).detach()) + self._reset_parameters() + + def _pos_bias_u(self): + return self.pos_bias_u * self.pos_bias_u_scale.exp() + + def _pos_bias_v(self): + return self.pos_bias_v * self.pos_bias_v_scale.exp() + + def _reset_parameters(self) -> None: + nn.init.normal_(self.pos_bias_u, std=0.01) + nn.init.normal_(self.pos_bias_v, std=0.01) + def _gen_attention_probs( self, attention_weights: torch.Tensor, @@ -539,6 +566,8 @@ class EmformerAttention(nn.Module): right_context: torch.Tensor, memory: torch.Tensor, attention_mask: torch.Tensor, + pos_emb: torch.Tensor, + rel_pos: torch.Tensor, padding_mask: Optional[torch.Tensor] = None, left_context_key: Optional[torch.Tensor] = None, left_context_val: Optional[torch.Tensor] = None, @@ -556,26 +585,79 @@ class EmformerAttention(nn.Module): torch.cat([memory, right_context, utterance]) ).chunk(chunks=2, dim=2) + is_streaming_infer = False if left_context_key is not None and left_context_val is not None: # now compute key and value with # [memory, right context, left context, uttrance] - # this is used in inference mode + # this is used in streaming inference mode + is_streaming_infer = True key = torch.cat([key[: M + R], left_context_key, key[M + R :]]) value = torch.cat( [value[: M + R], left_context_val, value[M + R :]] ) Q = query.size(0) - # KV = key.size(0) + KV = key.size(0) - reshaped_query, reshaped_key, reshaped_value = [ - tensor.contiguous() - .view(-1, B * self.nhead, self.head_dim) + reshaped_key = ( + key.contiguous() + .view(KV, B, self.nhead, self.head_dim) + .permute(1, 2, 0, 3) + ) # (B, nhead, KV, head_dim) + reshaped_value = ( + value.contiguous() + .view(KV, B * self.nhead, self.head_dim) .transpose(0, 1) - for tensor in [query, key, value] - ] # (B * nhead, Q or KV, head_dim) - attention_weights = torch.bmm( - reshaped_query * scaling, reshaped_key.transpose(1, 2) - ) # (B * nhead, Q, KV) + ) # (B * nhead, KV, head_dim) + query = ( + (query * scaling).contiguous().view(Q, B, self.nhead, self.head_dim) + ) + # (B, nhead, Q, head_dim) + query_with_bias_u = (query + self._pos_bias_u()).permute(1, 2, 0, 3) + query_with_bias_v = (query + self._pos_bias_v()).permute(1, 2, 0, 3) + + PE = pos_emb.size(0) + # pos_emb contains flipped positive part and negative part + # for relative position i - j between query (i) and key (j) + if is_streaming_infer: + # i is the first frame in current chunk (query) + # j is the last frame in right context (key) + # Note: R is equal to self.right_context_length here + min_neg_abs = U + R - 1 + # i is the last frame in right context (query) + # j is the first frame in the past context that memory bank can cover (key) # noqa + max_pos_abs = U + R + M * self.chunk_length - 1 + else: + # i is the first frame in utterance (query) + # j is the last frame in the last chunk's right context (key) + min_neg_abs = U + self.right_context_length - 1 + # i is the last frame in the last chunk's right context (query) + # j is the first frame in the utterance (key) + max_pos_abs = U + self.right_context_length - 1 + assert PE == min_neg_abs + max_pos_abs + 1 + pos_emb = ( + self.linear_pos(pos_emb) + .view(1, PE, self.nhead, self.head_dim) + .transpose(1, 2) + ) # (1, nhead, PE, head_dim) + + # content-based matrix-ac + matrix_ac = torch.matmul( + query_with_bias_u, reshaped_key.transpose(-2, -1) + ) # (B, nhead, Q, KV) + + # position-based matrix-bd + # (B, nhead, Q, PE) + matrix_bd = torch.matmul(query_with_bias_v, pos_emb.transpose(-2, -1)) + # gather position-related scores using pre-computed relative position + assert rel_pos.shape == (Q, KV) + rel_pos = rel_pos.unsqueeze(0).unsqueeze(1).expand(B, self.nhead, Q, KV) + matrix_bd = torch.gather( + matrix_bd, + dim=-1, + index=rel_pos, + ) # (B, nhead, Q, KV) + + attention_weights = (matrix_ac + matrix_bd).view(B * self.nhead, Q, KV) # compute attention probabilities attention_probs = self._gen_attention_probs( @@ -600,6 +682,8 @@ class EmformerAttention(nn.Module): right_context: torch.Tensor, memory: torch.Tensor, attention_mask: torch.Tensor, + pos_emb: torch.Tensor, + rel_pos: torch.Tensor, padding_mask: Optional[torch.Tensor] = None, ) -> torch.Tensor: # TODO: Modify docs. @@ -647,6 +731,11 @@ class EmformerAttention(nn.Module): attention_mask (torch.Tensor): Pre-computed attention mask to simulate underlying chunk-wise attention, with shape (Q, KV). + pos_emb (torch.Tensor): + Position embedding, with shape (PE, D), + where PE = 2 * (U + right_context_length) - 1. + rel_pos (torch.Tensor): + Relative positions, with shape (Q, KV). padding_mask (torch.Tensor): Padding mask of key tensor, with shape (B, KV). @@ -654,10 +743,12 @@ class EmformerAttention(nn.Module): Output of right context and utterance, with shape (R + U, B, D). """ output_right_context_utterance, _, _ = self._forward_impl( - utterance, - right_context, - memory, - attention_mask, + utterance=utterance, + right_context=right_context, + memory=memory, + attention_mask=attention_mask, + pos_emb=pos_emb, + rel_pos=rel_pos, padding_mask=padding_mask, ) return output_right_context_utterance @@ -667,6 +758,8 @@ class EmformerAttention(nn.Module): self, utterance: torch.Tensor, right_context: torch.Tensor, + pos_emb: torch.Tensor, + rel_pos: torch.Tensor, memory: torch.Tensor, left_context_key: torch.Tensor, left_context_val: torch.Tensor, @@ -700,6 +793,11 @@ class EmformerAttention(nn.Module): right_context (torch.Tensor): Right context frames, with shape (R, B, D), where R = right_context_length. + pos_emb (torch.Tensor): + Position embedding, with shape (PE, D), + where PE = 2 * (U + R) + M * chunk_length - 1. + rel_pos (torch.Tensor): + Relative positions, with shape (Q, KV). memory (torch.Tensor): Memory vectors, with shape (M, B, D), or empty tensor. left_context_key (torch,Tensor): @@ -733,10 +831,12 @@ class EmformerAttention(nn.Module): ) output_right_context_utterance, key, value = self._forward_impl( - utterance, - right_context, - memory, - attention_mask, + utterance=utterance, + right_context=right_context, + memory=memory, + attention_mask=attention_mask, + pos_emb=pos_emb, + rel_pos=rel_pos, padding_mask=padding_mask, left_context_key=left_context_key, left_context_val=left_context_val, @@ -796,6 +896,8 @@ class EmformerEncoderLayer(nn.Module): super().__init__() self.attention = EmformerAttention( + chunk_length=chunk_length, + right_context_length=right_context_length, embed_dim=d_model, nhead=nhead, dropout=dropout, @@ -898,6 +1000,8 @@ class EmformerEncoderLayer(nn.Module): right_context_utterance: torch.Tensor, R: int, attention_mask: torch.Tensor, + pos_emb: torch.Tensor, + rel_pos: torch.Tensor, padding_mask: Optional[torch.Tensor] = None, ) -> torch.Tensor: """Apply attention module in training and validation mode.""" @@ -917,15 +1021,18 @@ class EmformerEncoderLayer(nn.Module): right_context=right_context, memory=memory, attention_mask=attention_mask, + pos_emb=pos_emb, + rel_pos=rel_pos, padding_mask=padding_mask, ) - return output_right_context_utterance def _apply_attention_module_infer( self, right_context_utterance: torch.Tensor, R: int, + pos_emb: torch.Tensor, + rel_pos: torch.Tensor, attn_cache: List[torch.Tensor], padding_mask: Optional[torch.Tensor] = None, ) -> Tuple[torch.Tensor, List[torch.Tensor]]: @@ -962,6 +1069,8 @@ class EmformerEncoderLayer(nn.Module): ) = self.attention.infer( utterance=utterance, right_context=right_context, + pos_emb=pos_emb, + rel_pos=rel_pos, memory=pre_memory, left_context_key=left_context_key, left_context_val=left_context_val, @@ -977,6 +1086,8 @@ class EmformerEncoderLayer(nn.Module): utterance: torch.Tensor, right_context: torch.Tensor, attention_mask: torch.Tensor, + pos_emb: torch.Tensor, + rel_pos: torch.Tensor, padding_mask: Optional[torch.Tensor] = None, warmup: float = 1.0, ) -> Tuple[torch.Tensor, torch.Tensor]: @@ -996,6 +1107,11 @@ class EmformerEncoderLayer(nn.Module): attention_mask (torch.Tensor): Attention mask for underlying attention module, with shape (Q, KV), where Q = R + U, KV = M + R + U. + pos_emb (torch.Tensor): + Position embedding, with shape (PE, D), + where PE = 2 * (U + right_context_length) - 1. + rel_pos (torch.Tensor): + Relative positions, with shape (Q, KV). padding_mask (torch.Tensor): Padding mask of ker tensor, with shape (B, KV). @@ -1025,7 +1141,12 @@ class EmformerEncoderLayer(nn.Module): # emformer attention module src_att = self._apply_attention_module_forward( - src, R, attention_mask, padding_mask=padding_mask + right_context_utterance=src, + R=R, + attention_mask=attention_mask, + pos_emb=pos_emb, + rel_pos=rel_pos, + padding_mask=padding_mask, ) src = src + self.dropout(src_att) @@ -1050,6 +1171,8 @@ class EmformerEncoderLayer(nn.Module): self, utterance: torch.Tensor, right_context: torch.Tensor, + pos_emb: torch.Tensor, + rel_pos: torch.Tensor, attn_cache: List[torch.Tensor], conv_cache: torch.Tensor, padding_mask: Optional[torch.Tensor] = None, @@ -1067,6 +1190,12 @@ class EmformerEncoderLayer(nn.Module): Utterance frames, with shape (U, B, D). right_context (torch.Tensor): Right context frames, with shape (R, B, D). + pos_emb (torch.Tensor): + Position embedding, with shape (PE, D), + where PE = 2 * (U + R) + M * chunk_length - 1. + rel_pos (torch.Tensor): + Relative positions, with shape (Q, KV), + where Q = R + U, KV = M + R + L + U. attn_cache (List[torch.Tensor]): Cached attention tensors generated in preceding computation, including memory, key and value of left context. @@ -1090,7 +1219,12 @@ class EmformerEncoderLayer(nn.Module): # emformer attention module src_att, attn_cache = self._apply_attention_module_infer( - src, R, attn_cache, padding_mask=padding_mask + right_context_utterance=src, + R=R, + pos_emb=pos_emb, + rel_pos=rel_pos, + attn_cache=attn_cache, + padding_mask=padding_mask, ) src = src + self.dropout(src_att) @@ -1187,6 +1321,7 @@ class EmformerEncoder(nn.Module): self.use_memory = memory_size > 0 + self.encoder_pos = RelPositionalEncoding(d_model, dropout) self.emformer_layers = nn.ModuleList( [ EmformerEncoderLayer( @@ -1215,7 +1350,9 @@ class EmformerEncoder(nn.Module): self.memory_size = memory_size self.cnn_module_kernel = cnn_module_kernel - def _gen_right_context(self, x: torch.Tensor) -> torch.Tensor: + def _gen_right_context( + self, x: torch.Tensor + ) -> Tuple[torch.Tensor, torch.Tensor]: """Hard copy each chunk's right context and concat them.""" T = x.shape[0] num_chunks = math.ceil( @@ -1235,9 +1372,9 @@ class EmformerEncoder(nn.Module): indexes, torch.arange(T - self.right_context_length, T).unsqueeze(0), ] - ) - right_context_blocks = x[indexes.reshape(-1)] - return right_context_blocks + ).reshape(-1) + right_context_blocks = x[indexes] + return right_context_blocks, indexes def _gen_attention_mask_col_widths( self, chunk_idx: int, U: int @@ -1381,10 +1518,33 @@ class EmformerEncoder(nn.Module): - output_lengths, with shape (B,), without containing the right_context at the end. """ - U = x.size(0) - self.right_context_length + x, pos_emb = self.encoder_pos(x, pos_len=x.size(0), neg_len=x.size(0)) - right_context = self._gen_right_context(x) + U = x.size(0) - self.right_context_length + right_context, right_context_indexes = self._gen_right_context(x) + utterance_indexes = torch.arange(0, U) utterance = x[:U] + num_chunks = math.ceil(U / self.chunk_length) + memory_indexes = ( + torch.arange( + self.chunk_length // 2, + (num_chunks - 1) * self.chunk_length, + self.chunk_length, + ) + if num_chunks > 1 + else torch.empty(0).to(dtype=utterance_indexes.dtype) + ) + query_indexes = torch.cat( + [right_context_indexes, utterance_indexes] + ).to(device=x.device) + key_indexes = torch.cat( + [memory_indexes, right_context_indexes, utterance_indexes] + ).to(device=x.device) + # calculate relative position and flip sign + rel_pos = -(query_indexes.unsqueeze(1) - key_indexes.unsqueeze(0)) + # shift to start from zero + rel_pos = rel_pos - rel_pos.min() + output_lengths = torch.clamp(lengths - self.right_context_length, min=0) attention_mask = self._gen_attention_mask(utterance) @@ -1394,9 +1554,11 @@ class EmformerEncoder(nn.Module): output = utterance for layer in self.emformer_layers: output, right_context = layer( - output, - right_context, - attention_mask, + utterance=output, + right_context=right_context, + attention_mask=attention_mask, + pos_emb=pos_emb, + rel_pos=rel_pos, padding_mask=padding_mask, warmup=warmup, ) @@ -1445,6 +1607,7 @@ class EmformerEncoder(nn.Module): """ assert num_processed_frames.shape == (x.size(1),) + # check the shapes of states attn_caches = states[0] assert len(attn_caches) == self.num_encoder_layers, len(attn_caches) for i in range(len(attn_caches)): @@ -1473,6 +1636,11 @@ class EmformerEncoder(nn.Module): self.cnn_module_kernel - 1, ), conv_caches[i].shape + tot_past_length = self.memory_size * self.chunk_length + x, pos_emb = self.encoder_pos( + x, pos_len=x.size(0) + tot_past_length, neg_len=x.size(0) + ) + right_context = x[-self.right_context_length :] utterance = x[: -self.right_context_length] output_lengths = torch.clamp(lengths - self.right_context_length, min=0) @@ -1504,6 +1672,36 @@ class EmformerEncoder(nn.Module): dim=1, ) + # calculate relative position + memory_indexes = torch.arange( + self.chunk_length // 2, tot_past_length, self.chunk_length + ) + left_context_indexes = torch.arange( + tot_past_length - self.left_context_length, tot_past_length + ) + utterance_indexes = torch.arange( + tot_past_length, tot_past_length + utterance.size(0) + ) + right_context_indexes = torch.arange( + tot_past_length + utterance.size(0), + tot_past_length + utterance.size(0) + right_context.size(0), + ) + query_indexes = torch.cat( + [right_context_indexes, utterance_indexes] + ).to(device=x.device) + key_indexes = torch.cat( + [ + memory_indexes, + right_context_indexes, + left_context_indexes, + utterance_indexes, + ] + ).to(device=x.device) + # calculate relative position and flip sign + rel_pos = -(query_indexes.unsqueeze(1) - key_indexes.unsqueeze(0)) + # shift to start from zero + rel_pos = rel_pos - rel_pos.min() + output = utterance output_attn_caches: List[List[torch.Tensor]] = [] output_conv_caches: List[torch.Tensor] = [] @@ -1514,8 +1712,10 @@ class EmformerEncoder(nn.Module): output_attn_cache, output_conv_cache, ) = layer.infer( - output, - right_context, + utterance=output, + right_context=right_context, + pos_emb=pos_emb, + rel_pos=rel_pos, padding_mask=padding_mask, attn_cache=attn_caches[layer_idx], conv_cache=conv_caches[layer_idx], @@ -1597,6 +1797,10 @@ class Emformer(EncoderInterface): raise NotImplementedError( "right_context_length must be 0 or a mutiple of subsampling_factor." # noqa ) + if memory_size > 0 and memory_size * chunk_length < left_context_length: + raise NotImplementedError( + "memory_size * chunk_length must not be smaller than left_context_length." # noqa + ) # self.encoder_embed converts the input of shape (N, T, num_features) # to the shape (N, T//subsampling_factor, d_model). @@ -1822,3 +2026,119 @@ class Conv2dSubsampling(nn.Module): x = self.out_norm(x) x = self.out_balancer(x) return x + + +class RelPositionalEncoding(torch.nn.Module): + """Relative positional encoding module. + + See : Appendix B in "Transformer-XL: Attentive Language Models Beyond a Fixed-Length Context" # noqa + Modified from https://github.com/espnet/espnet/blob/master/espnet/nets/pytorch_backend/transformer/embedding.py # noqa + + Suppose: + i -> position of query, + j -> position of key(value), + we use positive relative position embedding when key(value) is to the + left of query(i.e., i > j) and negative embedding otherwise. + + Args: + d_model: Embedding dimension. + dropout: Dropout rate. + max_len: Maximum input length. + """ + + def __init__( + self, d_model: int, dropout: float, max_len: int = 5000 + ) -> None: + """Construct an PositionalEncoding object.""" + super(RelPositionalEncoding, self).__init__() + self.d_model = d_model + self.dropout = torch.nn.Dropout(p=dropout) + self.pe = None + self.pos_len = max_len + self.neg_len = max_len + self.gen_pe_positive() + self.gen_pe_negative() + + def gen_pe_positive(self) -> None: + """Generate the positive positional encodings.""" + pe_positive = torch.zeros(self.pos_len, self.d_model) + position_positive = torch.arange( + 0, self.pos_len, dtype=torch.float32 + ).unsqueeze(1) + div_term = torch.exp( + torch.arange(0, self.d_model, 2, dtype=torch.float32) + * -(math.log(10000.0) / self.d_model) + ) + pe_positive[:, 0::2] = torch.sin(position_positive * div_term) + pe_positive[:, 1::2] = torch.cos(position_positive * div_term) + # Reserve the order of positive indices and concat both positive and + # negative indices. This is used to support the shifting trick + # as in "Transformer-XL: Attentive Language Models Beyond a Fixed-Length Context" # noqa + self.pe_positive = torch.flip(pe_positive, [0]) + + def gen_pe_negative(self) -> None: + """Generate the negative positional encodings.""" + # Suppose `i` means to the position of query vecotr and `j` means the + # position of key vector. We use positive relative positions when keys + # are to the left (i>j) and negative relative positions otherwise (i torch.Tensor: + """Get positional encoding given positive length and negative length.""" + if self.pe_positive.dtype != dtype or str( + self.pe_positive.device + ) != str(device): + self.pe_positive = self.pe_positive.to(dtype=dtype, device=device) + if self.pe_negative.dtype != dtype or str( + self.pe_negative.device + ) != str(device): + self.pe_negative = self.pe_negative.to(dtype=dtype, device=device) + pe = torch.cat( + [ + self.pe_positive[self.pos_len - pos_len :], + self.pe_negative[1:neg_len], + ], + dim=0, + ) + return pe + + def forward( + self, + x: torch.Tensor, + pos_len: int, + neg_len: int, + ) -> Tuple[torch.Tensor, torch.Tensor]: + """Scale input x and get positional encoding. + Args: + x (torch.Tensor): Input tensor (`*`). + + Returns: + torch.Tensor: + Encoded tensor of shape (`*`). + torch.Tensor: + Position embedding of shape (pos_len + neg_len - 1, `*`). + """ + if pos_len > self.pos_len: + self.pos_len = pos_len + self.gen_pe_positive() + if neg_len > self.neg_len: + self.neg_len = neg_len + self.gen_pe_negative() + pos_emb = self.get_pe(pos_len, neg_len, x.device, x.dtype) + return self.dropout(x), self.dropout(pos_emb) diff --git a/egs/librispeech/ASR/conv_emformer_transducer_stateless3/test_emformer.py b/egs/librispeech/ASR/conv_emformer_transducer_stateless3/test_emformer.py index 8cde6205b..91c50ea3e 100644 --- a/egs/librispeech/ASR/conv_emformer_transducer_stateless3/test_emformer.py +++ b/egs/librispeech/ASR/conv_emformer_transducer_stateless3/test_emformer.py @@ -114,8 +114,12 @@ def test_state_stack_unstack(): for _ in range(num_encoder_layers) ] states = [attn_caches, conv_caches] - x = torch.randn(batch_size, 23, num_features) - x_lens = torch.full((batch_size,), 23) + x = torch.randn( + batch_size, chunk_length + right_context_length + 3, num_features + ) + x_lens = torch.full( + (batch_size,), chunk_length + right_context_length + 3 + ) num_processed_frames = torch.full((batch_size,), 0) y, y_lens, states = model.infer( x, x_lens, num_processed_frames=num_processed_frames, states=states @@ -172,8 +176,10 @@ def test_torchscript_consistency_infer(): for _ in range(num_encoder_layers) ] states = [attn_caches, conv_caches] - x = torch.randn(batch_size, 23, num_features) - x_lens = torch.full((batch_size,), 23) + x = torch.randn( + batch_size, chunk_length + right_context_length + 3, num_features + ) + x_lens = torch.full((batch_size,), chunk_length + right_context_length + 3) num_processed_frames = torch.full((batch_size,), 0) y, y_lens, out_states = model.infer( x, x_lens, num_processed_frames=num_processed_frames, states=states @@ -187,8 +193,38 @@ def test_torchscript_consistency_infer(): assert torch.allclose(y, sc_y) +def test_emformer_forward_shape(): + num_features = 80 + chunk_length = 32 + encoder_dim = 512 + num_encoder_layers = 2 + kernel_size = 31 + left_context_length = 32 + right_context_length = 8 + memory_size = 32 + batch_size = 2 + + model = Emformer( + num_features=num_features, + chunk_length=chunk_length, + subsampling_factor=4, + d_model=encoder_dim, + num_encoder_layers=num_encoder_layers, + cnn_module_kernel=kernel_size, + left_context_length=left_context_length, + right_context_length=right_context_length, + memory_size=memory_size, + ) + U = 2 * chunk_length + x = torch.randn(batch_size, U + right_context_length + 3, num_features) + x_lens = torch.full((batch_size,), U + right_context_length + 3) + output, output_lengths = model(x, x_lens) + assert output.shape == (batch_size, U >> 2, encoder_dim) + + if __name__ == "__main__": test_convolution_module_forward() test_convolution_module_infer() test_state_stack_unstack() test_torchscript_consistency_infer() + test_emformer_forward_shape() From 7b15596495c1c0fde3a02cbabea1c1e49669a6e3 Mon Sep 17 00:00:00 2001 From: yaozengwei Date: Sun, 26 Jun 2022 22:08:40 +0800 Subject: [PATCH 11/12] fix bug of relative position --- .../emformer.py | 20 +++++++++---------- 1 file changed, 9 insertions(+), 11 deletions(-) diff --git a/egs/librispeech/ASR/conv_emformer_transducer_stateless3/emformer.py b/egs/librispeech/ASR/conv_emformer_transducer_stateless3/emformer.py index 06fc880df..c8e202bac 100644 --- a/egs/librispeech/ASR/conv_emformer_transducer_stateless3/emformer.py +++ b/egs/librispeech/ASR/conv_emformer_transducer_stateless3/emformer.py @@ -1540,8 +1540,8 @@ class EmformerEncoder(nn.Module): key_indexes = torch.cat( [memory_indexes, right_context_indexes, utterance_indexes] ).to(device=x.device) - # calculate relative position and flip sign - rel_pos = -(query_indexes.unsqueeze(1) - key_indexes.unsqueeze(0)) + # calculate relative position + rel_pos = query_indexes.unsqueeze(1) - key_indexes.unsqueeze(0) # shift to start from zero rel_pos = rel_pos - rel_pos.min() @@ -1697,8 +1697,8 @@ class EmformerEncoder(nn.Module): utterance_indexes, ] ).to(device=x.device) - # calculate relative position and flip sign - rel_pos = -(query_indexes.unsqueeze(1) - key_indexes.unsqueeze(0)) + # calculate relative position + rel_pos = query_indexes.unsqueeze(1) - key_indexes.unsqueeze(0) # shift to start from zero rel_pos = rel_pos - rel_pos.min() @@ -2071,10 +2071,7 @@ class RelPositionalEncoding(torch.nn.Module): ) pe_positive[:, 0::2] = torch.sin(position_positive * div_term) pe_positive[:, 1::2] = torch.cos(position_positive * div_term) - # Reserve the order of positive indices and concat both positive and - # negative indices. This is used to support the shifting trick - # as in "Transformer-XL: Attentive Language Models Beyond a Fixed-Length Context" # noqa - self.pe_positive = torch.flip(pe_positive, [0]) + self.pe_positive = pe_positive def gen_pe_negative(self) -> None: """Generate the negative positional encodings.""" @@ -2091,7 +2088,8 @@ class RelPositionalEncoding(torch.nn.Module): ) pe_negative[:, 0::2] = torch.sin(-1 * position_negative * div_term) pe_negative[:, 1::2] = torch.cos(-1 * position_negative * div_term) - self.pe_negative = pe_negative + # Reserve the order of negative indices + self.pe_negative = torch.flip(pe_negative, [0]) def get_pe( self, @@ -2111,8 +2109,8 @@ class RelPositionalEncoding(torch.nn.Module): self.pe_negative = self.pe_negative.to(dtype=dtype, device=device) pe = torch.cat( [ - self.pe_positive[self.pos_len - pos_len :], - self.pe_negative[1:neg_len], + self.pe_negative[self.neg_len - neg_len :], + self.pe_positive[1:pos_len], ], dim=0, ) From 5ea58a446514f0cd1786c66a5058f7d735a53806 Mon Sep 17 00:00:00 2001 From: yaozengwei Date: Sun, 26 Jun 2022 22:34:40 +0800 Subject: [PATCH 12/12] add doc --- .../ASR/conv_emformer_transducer_stateless3/emformer.py | 2 ++ 1 file changed, 2 insertions(+) diff --git a/egs/librispeech/ASR/conv_emformer_transducer_stateless3/emformer.py b/egs/librispeech/ASR/conv_emformer_transducer_stateless3/emformer.py index c8e202bac..21ef4b18e 100644 --- a/egs/librispeech/ASR/conv_emformer_transducer_stateless3/emformer.py +++ b/egs/librispeech/ASR/conv_emformer_transducer_stateless3/emformer.py @@ -2109,6 +2109,8 @@ class RelPositionalEncoding(torch.nn.Module): self.pe_negative = self.pe_negative.to(dtype=dtype, device=device) pe = torch.cat( [ + # it starts from the min negative value of relative position + # and it is bound to be gathered self.pe_negative[self.neg_len - neg_len :], self.pe_positive[1:pos_len], ],