mirrors/icefall
1
0
Fork 0
mirror of https://github.com/k2-fsa/icefall.git synced 2025-12-11 06:55:27 +00:00
Code Issues Packages Projects Releases Wiki Activity

from local

Browse Source
This commit is contained in:
dohe0342 2023-02-14 22:57:25 +09:00
parent 8a28a3304a
commit a6c53bbdeb
3 changed files with 55 additions and 0 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

BIN
egs/tedlium2/ASR/conformer_ctc3/.train.py.swp
View File

Binary file not shown.

BIN
egs/tedlium2/ASR/conformer_ctc3/.transformer.py.swp
View File

Binary file not shown.

55
egs/tedlium2/ASR/conformer_ctc3/train.py
View File

@ -638,6 +638,7 @@ def compute_loss(
ctc_loss = (1-params.interctc_weight) * ctc_loss + params.interctc_weight * inter_ctc_loss
else:
'''
dense_fsa_vec = k2.DenseFsaVec(
nnet_output,
supervision_segments,
@ -651,6 +652,60 @@ def compute_loss(
reduction=params.reduction,
use_double_scores=params.use_double_scores,
)
'''
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
if params.att_rate > 0.0:
with torch.set_grad_enabled(is_training):