mirrors/icefall
1
0
Fork 0
mirror of https://github.com/k2-fsa/icefall.git synced 2025-08-12 03:22:19 +00:00
Code Issues Packages Projects Releases Wiki Activity

add ctc_decode.py

Browse Source
This commit is contained in:
yaozengwei 2023-02-13 21:21:30 +08:00
parent 251ff62880
commit dace53b125
3 changed files with 1091 additions and 30 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

1065
egs/librispeech/ASR/pruned_transducer_stateless4_ctc/ctc_decode.py Executable file
View File

File diff suppressed because it is too large Load Diff

42
egs/librispeech/ASR/pruned_transducer_stateless4_ctc/decode.py
View File

@ -19,36 +19,36 @@
""" """
Usage: Usage:
(1) greedy search (1) greedy search
./pruned_transducer_stateless4/decode.py \ ./pruned_transducer_stateless4_ctc/decode.py \
--epoch 30 \ --epoch 30 \
--avg 15 \ --avg 15 \
--exp-dir ./pruned_transducer_stateless4/exp \ --exp-dir ./pruned_transducer_stateless4_ctc/exp \
--max-duration 600 \ --max-duration 600 \
--decoding-method greedy_search --decoding-method greedy_search
(2) beam search (not recommended) (2) beam search (not recommended)
./pruned_transducer_stateless4/decode.py \ ./pruned_transducer_stateless4_ctc/decode.py \
--epoch 30 \ --epoch 30 \
--avg 15 \ --avg 15 \
--exp-dir ./pruned_transducer_stateless4/exp \ --exp-dir ./pruned_transducer_stateless4_ctc/exp \
--max-duration 600 \ --max-duration 600 \
--decoding-method beam_search \ --decoding-method beam_search \
--beam-size 4 --beam-size 4
(3) modified beam search (3) modified beam search
./pruned_transducer_stateless4/decode.py \ ./pruned_transducer_stateless4_ctc/decode.py \
--epoch 30 \ --epoch 30 \
--avg 15 \ --avg 15 \
--exp-dir ./pruned_transducer_stateless4/exp \ --exp-dir ./pruned_transducer_stateless4_ctc/exp \
--max-duration 600 \ --max-duration 600 \
--decoding-method modified_beam_search \ --decoding-method modified_beam_search \
--beam-size 4 --beam-size 4
(4) fast beam search (one best) (4) fast beam search (one best)
./pruned_transducer_stateless4/decode.py \ ./pruned_transducer_stateless4_ctc/decode.py \
--epoch 30 \ --epoch 30 \
--avg 15 \ --avg 15 \
--exp-dir ./pruned_transducer_stateless4/exp \ --exp-dir ./pruned_transducer_stateless4_ctc/exp \
--max-duration 600 \ --max-duration 600 \
--decoding-method fast_beam_search \ --decoding-method fast_beam_search \
--beam 20.0 \ --beam 20.0 \
@ -56,7 +56,7 @@ Usage:
--max-states 64 --max-states 64
(5) fast beam search (nbest) (5) fast beam search (nbest)
./pruned_transducer_stateless4/decode.py \ ./pruned_transducer_stateless4_ctc/decode.py \
--epoch 30 \ --epoch 30 \
--avg 15 \ --avg 15 \
--exp-dir ./pruned_transducer_stateless3/exp \ --exp-dir ./pruned_transducer_stateless3/exp \
@ -69,10 +69,10 @@ Usage:
--nbest-scale 0.5 --nbest-scale 0.5
(6) fast beam search (nbest oracle WER) (6) fast beam search (nbest oracle WER)
./pruned_transducer_stateless4/decode.py \ ./pruned_transducer_stateless4_ctc/decode.py \
--epoch 30 \ --epoch 30 \
--avg 15 \ --avg 15 \
--exp-dir ./pruned_transducer_stateless4/exp \ --exp-dir ./pruned_transducer_stateless4_ctc/exp \
--max-duration 600 \ --max-duration 600 \
--decoding-method fast_beam_search_nbest_oracle \ --decoding-method fast_beam_search_nbest_oracle \
--beam 20.0 \ --beam 20.0 \
@ -82,10 +82,10 @@ Usage:
--nbest-scale 0.5 --nbest-scale 0.5
(7) fast beam search (with LG) (7) fast beam search (with LG)
./pruned_transducer_stateless4/decode.py \ ./pruned_transducer_stateless4_ctc/decode.py \
--epoch 28 \ --epoch 28 \
--avg 15 \ --avg 15 \
--exp-dir ./pruned_transducer_stateless4/exp \ --exp-dir ./pruned_transducer_stateless4_ctc/exp \
--max-duration 600 \ --max-duration 600 \
--decoding-method fast_beam_search_nbest_LG \ --decoding-method fast_beam_search_nbest_LG \
--beam 20.0 \ --beam 20.0 \
@ -93,14 +93,14 @@ Usage:
--max-states 64 --max-states 64
(8) decode in streaming mode (take greedy search as an example) (8) decode in streaming mode (take greedy search as an example)
./pruned_transducer_stateless4/decode.py \ ./pruned_transducer_stateless4_ctc/decode.py \
--epoch 30 \ --epoch 30 \
--avg 15 \ --avg 15 \
--simulate-streaming 1 \ --simulate-streaming 1 \
--causal-convolution 1 \ --causal-convolution 1 \
--decode-chunk-size 16 \ --decode-chunk-size 16 \
--left-context 64 \ --left-context 64 \
--exp-dir ./pruned_transducer_stateless4/exp \ --exp-dir ./pruned_transducer_stateless4_ctc/exp \
--max-duration 600 \ --max-duration 600 \
--decoding-method greedy_search --decoding-method greedy_search
--beam 20.0 \ --beam 20.0 \
@ -109,16 +109,13 @@ Usage:
To evaluate symbol delay, you should: To evaluate symbol delay, you should:
(1) Generate cuts with word-time alignments: (1) Generate cuts with word-time alignments:
./local/add_alignment_librispeech.py \ ./add_alignments.sh
--alignments-dir data/alignment \
--cuts-in-dir data/fbank \
--cuts-out-dir data/fbank_ali
(2) Set the argument "--manifest-dir data/fbank_ali" while decoding. (2) Set the argument "--manifest-dir data/fbank_ali" while decoding.
For example: For example:
./pruned_transducer_stateless4/decode.py \ ./pruned_transducer_stateless4_ctc/decode.py \
--epoch 40 \ --epoch 40 \
--avg 20 \ --avg 20 \
--exp-dir ./pruned_transducer_stateless4/exp \ --exp-dir ./pruned_transducer_stateless4_ctc/exp \
--max-duration 600 \ --max-duration 600 \
--decoding-method greedy_search \ --decoding-method greedy_search \
--manifest-dir data/fbank_ali --manifest-dir data/fbank_ali
@ -216,7 +213,7 @@ def get_parser():
parser.add_argument( parser.add_argument(
"--exp-dir", "--exp-dir",
type=str, type=str,
default="pruned_transducer_stateless4/exp", default="pruned_transducer_stateless4_ctc/exp",
help="The experiment dir", help="The experiment dir",
) )
@ -528,7 +525,6 @@ def decode_one_batch(
res = DecodingResults(hyps=tokens, timestamps=timestamps) res = DecodingResults(hyps=tokens, timestamps=timestamps)
hyps, timestamps = parse_hyp_and_timestamp( hyps, timestamps = parse_hyp_and_timestamp(
decoding_method=params.decoding_method,
res=res, res=res,
sp=sp, sp=sp,
subsampling_factor=params.subsampling_factor, subsampling_factor=params.subsampling_factor,

14
egs/librispeech/ASR/pruned_transducer_stateless4_ctc/train.py
View File

@ -22,31 +22,31 @@ Usage:
export CUDA_VISIBLE_DEVICES="0,1,2,3" export CUDA_VISIBLE_DEVICES="0,1,2,3"
./pruned_transducer_stateless4/train.py \ ./pruned_transducer_stateless4_ctc/train.py \
--world-size 4 \ --world-size 4 \
--num-epochs 30 \ --num-epochs 30 \
--start-epoch 1 \ --start-epoch 1 \
--exp-dir pruned_transducer_stateless2/exp \ --exp-dir pruned_transducer_stateless4_ctc/exp \
--full-libri 1 \ --full-libri 1 \
--max-duration 300 --max-duration 300
# For mix precision training: # For mix precision training:
./pruned_transducer_stateless4/train.py \ ./pruned_transducer_stateless4_ctc/train.py \
--world-size 4 \ --world-size 4 \
--num-epochs 30 \ --num-epochs 30 \
--start-epoch 1 \ --start-epoch 1 \
--use-fp16 1 \ --use-fp16 1 \
--exp-dir pruned_transducer_stateless2/exp \ --exp-dir pruned_transducer_stateless4_ctc/exp \
--full-libri 1 \ --full-libri 1 \
--max-duration 550 --max-duration 550
# train a streaming model # train a streaming model
./pruned_transducer_stateless4/train.py \ ./pruned_transducer_stateless4_ctc/train.py \
--world-size 4 \ --world-size 4 \
--num-epochs 30 \ --num-epochs 30 \
--start-epoch 1 \ --start-epoch 1 \
--exp-dir pruned_transducer_stateless4/exp \ --exp-dir pruned_transducer_stateless4_ctc/exp \
--full-libri 1 \ --full-libri 1 \
--dynamic-chunk-training 1 \ --dynamic-chunk-training 1 \
--causal-convolution 1 \ --causal-convolution 1 \
@ -196,7 +196,7 @@ def get_parser():
parser.add_argument( parser.add_argument(
"--exp-dir", "--exp-dir",
type=str, type=str,
default="pruned_transducer_stateless2/exp", default="pruned_transducer_stateless4_ctc/exp",
help="""The experiment dir. help="""The experiment dir.
It specifies the directory where all training related It specifies the directory where all training related
files, e.g., checkpoints, log, etc, are saved files, e.g., checkpoints, log, etc, are saved