From df013c37fadb5381b715a1ec39e1496a99adf7be Mon Sep 17 00:00:00 2001
From: dohe0342 <kimdohe1070@gmail.com>
Date: Wed, 15 Feb 2023 13:05:09 +0900
Subject: [PATCH] from local

---
 egs/tedlium2/ASR/conformer_ctc3/.train.py.swp | Bin 61440 -> 65536 bytes
 egs/tedlium2/ASR/conformer_ctc3/train.py      |  54 ++++++++++++++++++
 2 files changed, 54 insertions(+)

diff --git a/egs/tedlium2/ASR/conformer_ctc3/.train.py.swp b/egs/tedlium2/ASR/conformer_ctc3/.train.py.swp
index ab8cb7b4817f3d96fc038b48744c7552b3f31ee4..194e6637a81c89c2d56cd1df0acffdaf5915ca50 100644
GIT binary patch
delta 574
zcmYk&Ur1AN6u|LwyH^)%bN9wtC~RG<=p|&7R0>H&#)e9QVwpr+S>l89vf67~&_h8`
zc;FT?DKen~k#(PvK<%mMC8~!We8`tD7)a1-f4(9?2R^^w;dl7`euq=}YE~*%;!=7j
z+vj$)rG>CXj&|2C=R0eOddz52)1EhSTC;aKvCT5kvbD+M;FY~U_R_Vy=wXgN8Z6w-
zsdN8R`N(#WIMzcVZ}9}97(y2NU}3dUq=o{{qXQ1UHHggO9ww1N3aggLJ3K=fgK+T6
zq=o>uO*tjqBnPn<d+=wQ$Op`#ge;=?P1`@P+5Q_WV+jj}=t?H8y}r}vy5oiGSAx{E
zXXxUY(|Ww#uWmIywUN~Rs->^h50<CSn!3;z($vFb^Y~TQ%NO$#*V@L4lO^wBp)}!k
zx#>>P6Vp5O!h^6@ZnXq6y-8DZPZOHk=u&UySg^agHc-xwiaZL_b7X6r|EB$!@z0@%
z9yH^7lgM-UxPw;2u^tt9k2#EC80XNA27KY|7r23wXvZP&XGmH}c0>Zhx*OJ|n-M+W
R+d+^2)6kJj<iFnMjlam1dUgN+

delta 303
zcmXBMu`feW9L4e9^=`2=y$wo8npc0o%d?RV8hML^&_P#;Mq?y`(UaI1yxgrU1{*`&
z(V&rtO(gyRbdV;x8eY<fuQ<tPJLkl)j^Dk)Mp#>1_UD37h!I(ddiAN<R-u<uS?ie_
z71@><cU?1)OWOXeueY<K8Ij81e?7AKv)9A<VUZbR&>a$az%ll*hbpG=&FTr~SVtAB
z_<IZC<DT~;?4p4Ml#ptXXPn>wH3aaql%?&7T;mj*C?g#Z`NShGu!8`TNR`M7u9W?_
j9C7hiL))Z2X1A7%E%uGQOY*ML?<iOLc~h>~?5zC)@Yz1s

diff --git a/egs/tedlium2/ASR/conformer_ctc3/train.py b/egs/tedlium2/ASR/conformer_ctc3/train.py
index b8e9c53d1..8630d5dc5 100755
--- a/egs/tedlium2/ASR/conformer_ctc3/train.py
+++ b/egs/tedlium2/ASR/conformer_ctc3/train.py
@@ -636,7 +636,61 @@ def compute_loss(
                 ) 
             
             ctc_loss = (1-params.interctc_weight) * ctc_loss + params.interctc_weight * inter_ctc_loss
+        
+        elif params.group_num > 0:
+            dense_fsa_vec = k2.DenseFsaVec(
+                nnet_output[0],
+                supervision_segments,
+                allow_truncate=params.subsampling_factor - 1,
+            )
+            
+            dense_fsa_vec_inter = [
+                #k2.DenseFsaVec(
+                #    nnet_output[1][2],
+                #    supervision_segments,
+                #    allow_truncate=params.subsampling_factor - 1,
+                #),
+                #k2.DenseFsaVec(
+                #    nnet_output[1][5],
+                #    supervision_segments,
+                #    allow_truncate=params.subsampling_factor - 1,
+                #),
+                k2.DenseFsaVec(
+                    nnet_output[1][8],
+                    supervision_segments,
+                    allow_truncate=params.subsampling_factor - 1,
+                ),
+                #k2.DenseFsaVec(
+                #    nnet_output[1][11],
+                #    supervision_segments,
+                #    allow_truncate=params.subsampling_factor - 1,
+                #),
+                #k2.DenseFsaVec(
+                #    nnet_output[1][14],
+                #    supervision_segments,
+                #    allow_truncate=params.subsampling_factor - 1,
+                #)
+            ]
 
+            ctc_loss = k2.ctc_loss(
+                decoding_graph=decoding_graph,
+                dense_fsa_vec=dense_fsa_vec,
+                output_beam=params.beam_size,
+                reduction=params.reduction,
+                use_double_scores=params.use_double_scores,
+            ) 
+            
+            inter_ctc_loss = 0
+            for fsa in dense_fsa_vec_inter:
+                inter_ctc_loss += k2.ctc_loss(
+                    decoding_graph=decoding_graph,
+                    dense_fsa_vec=fsa,
+                    output_beam=params.beam_size,
+                    reduction=params.reduction,
+                    use_double_scores=params.use_double_scores,
+                ) 
+            
+            ctc_loss = (1-params.interctc_weight) * ctc_loss + params.interctc_weight * inter_ctc_loss
         else:
             dense_fsa_vec = k2.DenseFsaVec(
                 nnet_output,