From 1c067e73646b62db1d215d86a58d41a6f555e4c8 Mon Sep 17 00:00:00 2001 From: yaozengwei Date: Mon, 13 Jun 2022 22:10:50 +0800 Subject: [PATCH] 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