mirrors/icefall
1
0
Fork 0
mirror of https://github.com/k2-fsa/icefall.git synced 2025-08-27 02:34:21 +00:00
Code Issues Packages Projects Releases Wiki Activity

Update RESULTS.

Browse Source
This commit is contained in:
Fangjun Kuang 2022-02-07 16:31:10 +08:00
parent e1936fa5a8
commit 3b93761199
4 changed files with 109 additions and 40 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

71
.github/workflows/run-pretrained-transducer-stateless.yml vendored
View File

@ -74,24 +74,53 @@ jobs:
mkdir tmp mkdir tmp
cd tmp cd tmp
git lfs install git lfs install
git clone https://huggingface.co/csukuangfj/icefall-asr-librispeech-transducer-stateless-bpe-500-2022-01-10 git clone https://huggingface.co/csukuangfj/icefall-asr-librispeech-transducer-stateless-bpe-500-2022-02-07
cd .. cd ..
tree tmp tree tmp
soxi tmp/icefall-asr-librispeech-transducer-stateless-bpe-500-2022-01-10/test_wavs/*.wav soxi tmp/icefall-asr-librispeech-transducer-stateless-bpe-500-2022-02-07/test_wavs/*.wav
ls -lh tmp/icefall-asr-librispeech-transducer-stateless-bpe-500-2022-01-10/test_wavs/*.wav ls -lh tmp/icefall-asr-librispeech-transducer-stateless-bpe-500-2022-02-07/test_wavs/*.wav
- name: Run greedy search decoding - name: Run greedy search decoding (max-sym-per-frame 1)
shell: bash shell: bash
run: | run: |
export PYTHONPATH=$PWD:PYTHONPATH export PYTHONPATH=$PWD:PYTHONPATH
cd egs/librispeech/ASR cd egs/librispeech/ASR
./transducer_stateless/pretrained.py \ ./transducer_stateless/pretrained.py \
--method greedy_search \ --method greedy_search \
--checkpoint ./tmp/icefall-asr-librispeech-transducer-stateless-bpe-500-2022-01-10/exp/pretrained.pt \ --max-sym-per-frame 1 \
--bpe-model ./tmp/icefall-asr-librispeech-transducer-stateless-bpe-500-2022-01-10/data/lang_bpe_500/bpe.model \ --checkpoint ./tmp/icefall-asr-librispeech-transducer-stateless-bpe-500-2022-02-07/exp/pretrained.pt \
./tmp/icefall-asr-librispeech-transducer-stateless-bpe-500-2022-01-10/test_wavs/1089-134686-0001.wav \ --bpe-model ./tmp/icefall-asr-librispeech-transducer-stateless-bpe-500-2022-02-07/data/lang_bpe_500/bpe.model \
./tmp/icefall-asr-librispeech-transducer-stateless-bpe-500-2022-01-10/test_wavs/1221-135766-0001.wav \ ./tmp/icefall-asr-librispeech-transducer-stateless-bpe-500-2022-02-07/test_wavs/1089-134686-0001.wav \
./tmp/icefall-asr-librispeech-transducer-stateless-bpe-500-2022-01-10/test_wavs/1221-135766-0002.wav ./tmp/icefall-asr-librispeech-transducer-stateless-bpe-500-2022-02-07/test_wavs/1221-135766-0001.wav \
./tmp/icefall-asr-librispeech-transducer-stateless-bpe-500-2022-02-07/test_wavs/1221-135766-0002.wav
- name: Run greedy search decoding (max-sym-per-frame 2)
shell: bash
run: |
export PYTHONPATH=$PWD:PYTHONPATH
cd egs/librispeech/ASR
./transducer_stateless/pretrained.py \
--method greedy_search \
--max-sym-per-frame 2 \
--checkpoint ./tmp/icefall-asr-librispeech-transducer-stateless-bpe-500-2022-02-07/exp/pretrained.pt \
--bpe-model ./tmp/icefall-asr-librispeech-transducer-stateless-bpe-500-2022-02-07/data/lang_bpe_500/bpe.model \
./tmp/icefall-asr-librispeech-transducer-stateless-bpe-500-2022-02-07/test_wavs/1089-134686-0001.wav \
./tmp/icefall-asr-librispeech-transducer-stateless-bpe-500-2022-02-07/test_wavs/1221-135766-0001.wav \
./tmp/icefall-asr-librispeech-transducer-stateless-bpe-500-2022-02-07/test_wavs/1221-135766-0002.wav
- name: Run greedy search decoding (max-sym-per-frame 3)
shell: bash
run: |
export PYTHONPATH=$PWD:PYTHONPATH
cd egs/librispeech/ASR
./transducer_stateless/pretrained.py \
--method greedy_search \
--max-sym-per-frame 3 \
--checkpoint ./tmp/icefall-asr-librispeech-transducer-stateless-bpe-500-2022-02-07/exp/pretrained.pt \
--bpe-model ./tmp/icefall-asr-librispeech-transducer-stateless-bpe-500-2022-02-07/data/lang_bpe_500/bpe.model \
./tmp/icefall-asr-librispeech-transducer-stateless-bpe-500-2022-02-07/test_wavs/1089-134686-0001.wav \
./tmp/icefall-asr-librispeech-transducer-stateless-bpe-500-2022-02-07/test_wavs/1221-135766-0001.wav \
./tmp/icefall-asr-librispeech-transducer-stateless-bpe-500-2022-02-07/test_wavs/1221-135766-0002.wav
- name: Run beam search decoding - name: Run beam search decoding
shell: bash shell: bash
@ -101,8 +130,22 @@ jobs:
./transducer_stateless/pretrained.py \ ./transducer_stateless/pretrained.py \
--method beam_search \ --method beam_search \
--beam-size 4 \ --beam-size 4 \
--checkpoint ./tmp/icefall-asr-librispeech-transducer-stateless-bpe-500-2022-01-10/exp/pretrained.pt \ --checkpoint ./tmp/icefall-asr-librispeech-transducer-stateless-bpe-500-2022-02-07/exp/pretrained.pt \
--bpe-model ./tmp/icefall-asr-librispeech-transducer-stateless-bpe-500-2022-01-10/data/lang_bpe_500/bpe.model \ --bpe-model ./tmp/icefall-asr-librispeech-transducer-stateless-bpe-500-2022-02-07/data/lang_bpe_500/bpe.model \
./tmp/icefall-asr-librispeech-transducer-stateless-bpe-500-2022-01-10/test_wavs/1089-134686-0001.wav \ ./tmp/icefall-asr-librispeech-transducer-stateless-bpe-500-2022-02-07/test_wavs/1089-134686-0001.wav \
./tmp/icefall-asr-librispeech-transducer-stateless-bpe-500-2022-01-10/test_wavs/1221-135766-0001.wav \ ./tmp/icefall-asr-librispeech-transducer-stateless-bpe-500-2022-02-07/test_wavs/1221-135766-0001.wav \
./tmp/icefall-asr-librispeech-transducer-stateless-bpe-500-2022-01-10/test_wavs/1221-135766-0002.wav ./tmp/icefall-asr-librispeech-transducer-stateless-bpe-500-2022-02-07/test_wavs/1221-135766-0002.wav
- name: Run modified beam search decoding
shell: bash
run: |
export PYTHONPATH=$PWD:$PYTHONPATH
cd egs/librispeech/ASR
./transducer_stateless/pretrained.py \
--method modified_beam_search \
--beam-size 4 \
--checkpoint ./tmp/icefall-asr-librispeech-transducer-stateless-bpe-500-2022-02-07/exp/pretrained.pt \
--bpe-model ./tmp/icefall-asr-librispeech-transducer-stateless-bpe-500-2022-02-07/data/lang_bpe_500/bpe.model \
./tmp/icefall-asr-librispeech-transducer-stateless-bpe-500-2022-02-07/test_wavs/1089-134686-0001.wav \
./tmp/icefall-asr-librispeech-transducer-stateless-bpe-500-2022-02-07/test_wavs/1221-135766-0001.wav \
./tmp/icefall-asr-librispeech-transducer-stateless-bpe-500-2022-02-07/test_wavs/1221-135766-0002.wav

6
README.md
View File

@ -80,16 +80,16 @@ We provide a Colab notebook to run a pre-trained RNN-T conformer model: [![Open
Using Conformer as encoder. The decoder consists of 1 embedding layer Using Conformer as encoder. The decoder consists of 1 embedding layer
and 1 convolutional layer. and 1 convolutional layer.
The best WER using beam search with beam size 4 is: The best WER using modified beam search with beam size 4 is:
| | test-clean | test-other | | | test-clean | test-other |
|-----|------------|------------| |-----|------------|------------|
| WER | 2.68 | 6.72 | | WER | 2.67 | 6.64 |
Note: No auxiliary losses are used in the training and no LMs are used Note: No auxiliary losses are used in the training and no LMs are used
in the decoding. in the decoding.
We provide a Colab notebook to run a pre-trained transducer conformer + stateless decoder model: [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/drive/1Rc4Is-3Yp9LbcEz_Iy8hfyenyHsyjvqE?usp=sharing) We provide a Colab notebook to run a pre-trained transducer conformer + stateless decoder model: [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/drive/1CO1bXJ-2khDckZIW8zjOPHGSKLHpTDlp?usp=sharing)
### Aishell ### Aishell

51
egs/librispeech/ASR/RESULTS.md
View File

@ -4,7 +4,7 @@
#### Conformer encoder + embedding decoder #### Conformer encoder + embedding decoder
Using commit `4c1b3665ee6efb935f4dd93a80ff0e154b13efb6`. Using commit `TODO`.
Conformer encoder + non-recurrent decoder. The decoder Conformer encoder + non-recurrent decoder. The decoder
contains only an embedding layer and a Conv1d (with kernel size 2). contains only an embedding layer and a Conv1d (with kernel size 2).
@ -12,54 +12,65 @@ contains only an embedding layer and a Conv1d (with kernel size 2).
The WERs are The WERs are
| | test-clean | test-other | comment | | | test-clean | test-other | comment |
|---------------------------|------------|------------|------------------------------------------| |-------------------------------------|------------|------------|------------------------------------------|
| greedy search | 2.69 | 6.81 | --epoch 71, --avg 15, --max-duration 100 | | greedy search (max sym per frame 1) | 2.68 | 6.71 | --epoch 61, --avg 18, --max-duration 100 |
| beam search (beam size 4) | 2.68 | 6.72 | --epoch 71, --avg 15, --max-duration 100 | | greedy search (max sym per frame 2) | 2.69 | 6.71 | --epoch 61, --avg 18, --max-duration 100 |
| greedy search (max sym per frame 3) | 2.69 | 6.71 | --epoch 61, --avg 18, --max-duration 100 |
| modified beam search (beam size 4) | 2.67 | 6.64 | --epoch 61, --avg 18, --max-duration 100 |
The training command for reproducing is given below: The training command for reproducing is given below:
``` ```
cd egs/librispeech/ASR/
./prepare.sh
export CUDA_VISIBLE_DEVICES="0,1,2,3" export CUDA_VISIBLE_DEVICES="0,1,2,3"
./transducer_stateless/train.py \ ./transducer_stateless/train.py \
--world-size 4 \ --world-size 4 \
--num-epochs 76 \ --num-epochs 76 \
--start-epoch 0 \ --start-epoch 0 \
--exp-dir transducer_stateless/exp-full \ --exp-dir transducer_stateless/exp-full \
--full-libri 1 \ --full-libri 1 \
--max-duration 250 \ --max-duration 300 \
--lr-factor 3 --lr-factor 5 \
--bpe-model data/lang_bpe_500/bpe.model \
--modified-transducer-prob 0.25
``` ```
The tensorboard training log can be found at The tensorboard training log can be found at
<https://tensorboard.dev/experiment/qGdqzHnxS0WJ695OXfZDzA/#scalars&_smoothingWeight=0> <https://tensorboard.dev/experiment/qgvWkbF2R46FYA6ZMNmOjA/#scalars>
The decoding command is: The decoding command is:
``` ```
epoch=71 epoch=61
avg=15 avg=18
## greedy search ## greedy search
./transducer_stateless/decode.py \ for sym in 1 2 3; do
--epoch $epoch \
--avg $avg \
--exp-dir transducer_stateless/exp-full \
--bpe-model ./data/lang_bpe_500/bpe.model \
--max-duration 100
## beam search
./transducer_stateless/decode.py \ ./transducer_stateless/decode.py \
--epoch $epoch \ --epoch $epoch \
--avg $avg \ --avg $avg \
--exp-dir transducer_stateless/exp-full \ --exp-dir transducer_stateless/exp-full \
--bpe-model ./data/lang_bpe_500/bpe.model \ --bpe-model ./data/lang_bpe_500/bpe.model \
--max-duration 100 \ --max-duration 100 \
--decoding-method beam_search \ --max-sym-per-frame $sym
done
## modified beam search
./transducer_stateless/decode.py \
--epoch $epoch \
--avg $avg \
--exp-dir transducer_stateless/exp-full \
--bpe-model ./data/lang_bpe_500/bpe.model \
--max-duration 100 \
--context-size 2 \
--decoding-method modified_beam_search \
--beam-size 4 --beam-size 4
``` ```
You can find a pretrained model by visiting You can find a pretrained model by visiting
<https://huggingface.co/csukuangfj/icefall-asr-librispeech-transducer-stateless-bpe-500-2022-01-10> <https://huggingface.co/csukuangfj/icefall-asr-librispeech-transducer-stateless-bpe-500-2022-02-07>
#### Conformer encoder + LSTM decoder #### Conformer encoder + LSTM decoder

21
egs/librispeech/ASR/transducer_stateless/pretrained.py
View File

@ -22,10 +22,11 @@ Usage:
--checkpoint ./transducer_stateless/exp/pretrained.pt \ --checkpoint ./transducer_stateless/exp/pretrained.pt \
--bpe-model ./data/lang_bpe_500/bpe.model \ --bpe-model ./data/lang_bpe_500/bpe.model \
--method greedy_search \ --method greedy_search \
--max-sym-per-frame 1 \
/path/to/foo.wav \ /path/to/foo.wav \
/path/to/bar.wav \ /path/to/bar.wav \
(1) beam search (2) beam search
./transducer_stateless/pretrained.py \ ./transducer_stateless/pretrained.py \
--checkpoint ./transducer_stateless/exp/pretrained.pt \ --checkpoint ./transducer_stateless/exp/pretrained.pt \
--bpe-model ./data/lang_bpe_500/bpe.model \ --bpe-model ./data/lang_bpe_500/bpe.model \
@ -34,6 +35,15 @@ Usage:
/path/to/foo.wav \ /path/to/foo.wav \
/path/to/bar.wav \ /path/to/bar.wav \
(3) modified beam search
./transducer_stateless/pretrained.py \
--checkpoint ./transducer_stateless/exp/pretrained.pt \
--bpe-model ./data/lang_bpe_500/bpe.model \
--method modified_beam_search \
--beam-size 4 \
/path/to/foo.wav \
/path/to/bar.wav \
You can also use `./transducer_stateless/exp/epoch-xx.pt`. You can also use `./transducer_stateless/exp/epoch-xx.pt`.
Note: ./transducer_stateless/exp/pretrained.pt is generated by Note: ./transducer_stateless/exp/pretrained.pt is generated by
@ -50,7 +60,7 @@ import kaldifeat
import sentencepiece as spm import sentencepiece as spm
import torch import torch
import torchaudio import torchaudio
from beam_search import beam_search, greedy_search from beam_search import beam_search, greedy_search, modified_beam_search
from conformer import Conformer from conformer import Conformer
from decoder import Decoder from decoder import Decoder
from joiner import Joiner from joiner import Joiner
@ -90,6 +100,7 @@ def get_parser():
help="""Possible values are: help="""Possible values are:
- greedy_search - greedy_search
- beam_search - beam_search
- modified_beam_search
""", """,
) )
@ -107,7 +118,7 @@ def get_parser():
"--beam-size", "--beam-size",
type=int, type=int,
default=4, default=4,
help="Used only when --method is beam_search", help="Used only when --method is beam_search and modified_beam_search ",
) )
parser.add_argument( parser.add_argument(
@ -300,6 +311,10 @@ def main():
hyp = beam_search( hyp = beam_search(
model=model, encoder_out=encoder_out_i, beam=params.beam_size model=model, encoder_out=encoder_out_i, beam=params.beam_size
) )
elif params.decoding_method == "modified_beam_search":
hyp = modified_beam_search(
model=model, encoder_out=encoder_out_i, beam=params.beam_size
)
else: else:
raise ValueError(f"Unsupported method: {params.method}") raise ValueError(f"Unsupported method: {params.method}")