mirrors/icefall
1
0
Fork 0
mirror of https://github.com/k2-fsa/icefall.git synced 2025-08-09 10:02:22 +00:00
Code Issues Packages Projects Releases Wiki Activity

Combining rnnt loss and k2-ctc loss for Dan's Zipformer (#683)

Browse Source 
* init files

* add ctc as auxiliary loss and ctc_decode.py

* tuning the scalar of HLG score for 1best, nbest and nbest-oracle

* rename to pruned_transducer_stateless7_ctc

* fix doc

* fix bug, recover the hlg scores

* modify ctc_decode.py, move out the hlg scale

* fix hlg_scale

* add export.py and pretrained.py, and so on

* upload files, update README.md and RESULTS.md

* add CI test
This commit is contained in:
Zengwei Yao 2022-12-03 19:01:10 +08:00 committed by GitHub
parent 7700ddcb38
commit 8eb4b9d96d
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
26 changed files with 5396 additions and 14 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

147
.github/scripts/run-librispeech-pruned-transducer-stateless7-ctc-2022-12-01.sh vendored Executable file
View File

@ -0,0 +1,147 @@
#!/usr/bin/env bash
set -e
log() {
# This function is from espnet
local fname=${BASH_SOURCE[1]##*/}
echo -e "$(date '+%Y-%m-%d %H:%M:%S') (${fname}:${BASH_LINENO[0]}:${FUNCNAME[1]}) $*"
}
cd egs/librispeech/ASR
repo_url=https://huggingface.co/Zengwei/icefall-asr-librispeech-pruned-transducer-stateless7-ctc-2022-12-01
log "Downloading pre-trained model from $repo_url"
git lfs install
GIT_LFS_SKIP_SMUDGE=1 git clone $repo_url
repo=$(basename $repo_url)
log "Display test files"
tree $repo/
soxi $repo/test_wavs/*.wav
ls -lh $repo/test_wavs/*.wav
pushd $repo/exp
git lfs pull --include "data/*"
git lfs pull --include "exp/cpu_jit.pt"
git lfs pull --include "exp/pretrained.pt"
ln -s pretrained.pt epoch-99.pt
ls -lh *.pt
popd
log "Export to torchscript model"
./pruned_transducer_stateless7_ctc/export.py \
--exp-dir $repo/exp \
--use-averaged-model false \
--bpe-model $repo/data/lang_bpe_500/bpe.model \
--epoch 99 \
--avg 1 \
--jit 1
ls -lh $repo/exp/*.pt
log "Decode with models exported by torch.jit.script()"
./pruned_transducer_stateless7_ctc/jit_pretrained.py \
--bpe-model $repo/data/lang_bpe_500/bpe.model \
--nn-model-filename $repo/exp/cpu_jit.pt \
$repo/test_wavs/1089-134686-0001.wav \
$repo/test_wavs/1221-135766-0001.wav \
$repo/test_wavs/1221-135766-0002.wav
for m in ctc-decoding 1best; do
./pruned_transducer_stateless7_ctc/jit_pretrained_ctc.py \
--model-filename $repo/exp/cpu_jit.pt \
--words-file $repo/data/lang_bpe_500/words.txt \
--HLG $repo/data/lang_bpe_500/HLG.pt \
--bpe-model $repo/data/lang_bpe_500/bpe.model \
--G $repo/data/lm/G_4_gram.pt \
--method $m \
--sample-rate 16000 \
$repo/test_wavs/1089-134686-0001.wav \
$repo/test_wavs/1221-135766-0001.wav \
$repo/test_wavs/1221-135766-0002.wav
done
for sym in 1 2 3; do
log "Greedy search with --max-sym-per-frame $sym"
./pruned_transducer_stateless7_ctc/pretrained.py \
--method greedy_search \
--max-sym-per-frame $sym \
--checkpoint $repo/exp/pretrained.pt \
--bpe-model $repo/data/lang_bpe_500/bpe.model \
$repo/test_wavs/1089-134686-0001.wav \
$repo/test_wavs/1221-135766-0001.wav \
$repo/test_wavs/1221-135766-0002.wav
done
for method in modified_beam_search beam_search fast_beam_search; do
log "$method"
./pruned_transducer_stateless7_ctc/pretrained.py \
--method $method \
--beam-size 4 \
--checkpoint $repo/exp/pretrained.pt \
--bpe-model $repo/data/lang_bpe_500/bpe.model \
$repo/test_wavs/1089-134686-0001.wav \
$repo/test_wavs/1221-135766-0001.wav \
$repo/test_wavs/1221-135766-0002.wav
done
for m in ctc-decoding 1best; do
./pruned_transducer_stateless7_ctc/pretrained_ctc.py \
--checkpoint $repo/exp/pretrained.pt \
--words-file $repo/data/lang_bpe_500/words.txt \
--HLG $repo/data/lang_bpe_500/HLG.pt \
--bpe-model $repo/data/lang_bpe_500/bpe.model \
--G $repo/data/lm/G_4_gram.pt \
--method $m \
--sample-rate 16000 \
$repo/test_wavs/1089-134686-0001.wav \
$repo/test_wavs/1221-135766-0001.wav \
$repo/test_wavs/1221-135766-0002.wav
done
echo "GITHUB_EVENT_NAME: ${GITHUB_EVENT_NAME}"
echo "GITHUB_EVENT_LABEL_NAME: ${GITHUB_EVENT_LABEL_NAME}"
if [[ x"${GITHUB_EVENT_NAME}" == x"schedule" || x"${GITHUB_EVENT_LABEL_NAME}" == x"run-decode" ]]; then
mkdir -p pruned_transducer_stateless7_ctc/exp
ln -s $PWD/$repo/exp/pretrained.pt pruned_transducer_stateless7_ctc/exp/epoch-999.pt
ln -s $PWD/$repo/data/lang_bpe_500 data/
ls -lh data
ls -lh pruned_transducer_stateless7_ctc/exp
log "Decoding test-clean and test-other"
# use a small value for decoding with CPU
max_duration=100
for method in greedy_search fast_beam_search modified_beam_search; do
log "Decoding with $method"
./pruned_transducer_stateless7_ctc/decode.py \
--decoding-method $method \
--epoch 999 \
--avg 1 \
--use-averaged-model 0 \
--max-duration $max_duration \
--exp-dir pruned_transducer_stateless7_ctc/exp
done
for m in ctc-decoding 1best; do
./pruned_transducer_stateless7_ctc/ctc_decode.py \
--epoch 999 \
--avg 1 \
--exp-dir ./pruned_transducer_stateless7_ctc/exp \
--max-duration $max_duration \
--use-averaged-model 0 \
--decoding-method $m \
--hlg-scale 0.6 \
--lm-dir data/lm
done
rm pruned_transducer_stateless7_ctc/exp/*.pt
fi

163
.github/workflows/run-librispeech-2022-12-01-stateless7-ctc.yml vendored Normal file
View File

@ -0,0 +1,163 @@
# Copyright 2022 Fangjun Kuang (csukuangfj@gmail.com)
# 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.
name: run-librispeech-2022-12-01-stateless7-ctc
# zipformer
on:
push:
branches:
- master
pull_request:
types: [labeled]
schedule:
# minute (0-59)
# hour (0-23)
# day of the month (1-31)
# month (1-12)
# day of the week (0-6)
# nightly build at 15:50 UTC time every day
- cron: "50 15 * * *"
jobs:
run_librispeech_2022_11_11_zipformer:
if: github.event.label.name == 'ready' || github.event.label.name == 'run-decode' || github.event_name == 'push' || github.event_name == 'schedule'
runs-on: ${{ matrix.os }}
strategy:
matrix:
os: [ubuntu-latest]
python-version: [3.8]
fail-fast: false
steps:
- uses: actions/checkout@v2
with:
fetch-depth: 0
- name: Setup Python ${{ matrix.python-version }}
uses: actions/setup-python@v2
with:
python-version: ${{ matrix.python-version }}
cache: 'pip'
cache-dependency-path: '**/requirements-ci.txt'
- name: Install Python dependencies
run: |
grep -v '^#' ./requirements-ci.txt | xargs -n 1 -L 1 pip install
pip uninstall -y protobuf
pip install --no-binary protobuf protobuf
- name: Cache kaldifeat
id: my-cache
uses: actions/cache@v2
with:
path: |
~/tmp/kaldifeat
key: cache-tmp-${{ matrix.python-version }}-2022-09-25
- name: Install kaldifeat
if: steps.my-cache.outputs.cache-hit != 'true'
shell: bash
run: |
.github/scripts/install-kaldifeat.sh
- name: Cache LibriSpeech test-clean and test-other datasets
id: libri-test-clean-and-test-other-data
uses: actions/cache@v2
with:
path: |
~/tmp/download
key: cache-libri-test-clean-and-test-other
- name: Download LibriSpeech test-clean and test-other
if: steps.libri-test-clean-and-test-other-data.outputs.cache-hit != 'true'
shell: bash
run: |
.github/scripts/download-librispeech-test-clean-and-test-other-dataset.sh
- name: Prepare manifests for LibriSpeech test-clean and test-other
shell: bash
run: |
.github/scripts/prepare-librispeech-test-clean-and-test-other-manifests.sh
- name: Cache LibriSpeech test-clean and test-other fbank features
id: libri-test-clean-and-test-other-fbank
uses: actions/cache@v2
with:
path: |
~/tmp/fbank-libri
key: cache-libri-fbank-test-clean-and-test-other-v2
- name: Compute fbank for LibriSpeech test-clean and test-other
if: steps.libri-test-clean-and-test-other-fbank.outputs.cache-hit != 'true'
shell: bash
run: |
.github/scripts/compute-fbank-librispeech-test-clean-and-test-other.sh
- name: Inference with pre-trained model
shell: bash
env:
GITHUB_EVENT_NAME: ${{ github.event_name }}
GITHUB_EVENT_LABEL_NAME: ${{ github.event.label.name }}
run: |
mkdir -p egs/librispeech/ASR/data
ln -sfv ~/tmp/fbank-libri egs/librispeech/ASR/data/fbank
ls -lh egs/librispeech/ASR/data/*
sudo apt-get -qq install git-lfs tree sox
export PYTHONPATH=$PWD:$PYTHONPATH
export PYTHONPATH=~/tmp/kaldifeat/kaldifeat/python:$PYTHONPATH
export PYTHONPATH=~/tmp/kaldifeat/build/lib:$PYTHONPATH
.github/scripts/run-librispeech-pruned-transducer-stateless7-ctc-2022-12-01.sh
- name: Display decoding results for librispeech pruned_transducer_stateless7_ctc
if: github.event_name == 'schedule' || github.event.label.name == 'run-decode'
shell: bash
run: |
cd egs/librispeech/ASR/
tree ./pruned_transducer_stateless7_ctc/exp
cd pruned_transducer_stateless7_ctc
echo "results for pruned_transducer_stateless7_ctc"
echo "===greedy search==="
find exp/greedy_search -name "log-*" -exec grep -n --color "best for test-clean" {} + | sort -n -k2
find exp/greedy_search -name "log-*" -exec grep -n --color "best for test-other" {} + | sort -n -k2
echo "===fast_beam_search==="
find exp/fast_beam_search -name "log-*" -exec grep -n --color "best for test-clean" {} + | sort -n -k2
find exp/fast_beam_search -name "log-*" -exec grep -n --color "best for test-other" {} + | sort -n -k2
echo "===modified beam search==="
find exp/modified_beam_search -name "log-*" -exec grep -n --color "best for test-clean" {} + | sort -n -k2
find exp/modified_beam_search -name "log-*" -exec grep -n --color "best for test-other" {} + | sort -n -k2
echo "===ctc decoding==="
find exp/ctc-decoding -name "log-*" -exec grep -n --color "best for test-clean" {} + | sort -n -k2
find exp/ctc-decoding -name "log-*" -exec grep -n --color "best for test-other" {} + | sort -n -k2
echo "===1best==="
find exp/1best -name "log-*" -exec grep -n --color "best for test-clean" {} + | sort -n -k2
find exp/1best -name "log-*" -exec grep -n --color "best for test-other" {} + | sort -n -k2
- name: Upload decoding results for librispeech pruned_transducer_stateless7_ctc
uses: actions/upload-artifact@v2
if: github.event_name == 'schedule' || github.event.label.name == 'run-decode'
with:
name: torch-${{ matrix.torch }}-python-${{ matrix.python-version }}-ubuntu-18.04-cpu-pruned_transducer_stateless7-ctc-2022-12-01
path: egs/librispeech/ASR/pruned_transducer_stateless7_ctc/exp/

1
egs/librispeech/ASR/README.md
View File

@ -23,6 +23,7 @@ The following table lists the differences among them.
| `pruned_transducer_stateless5` | Conformer(modified) | Embedding + Conv1d | same as pruned_transducer_stateless4 + more layers + random combiner|
| `pruned_transducer_stateless6` | Conformer(modified) | Embedding + Conv1d | same as pruned_transducer_stateless4 + distillation with hubert|
| `pruned_transducer_stateless7` | Zipformer | Embedding + Conv1d | First experiment with Zipformer from Dan|
| `pruned_transducer_stateless7_ctc` | Zipformer | Embedding + Conv1d | Same as pruned_transducer_stateless7, but with extra CTC head|
| `pruned_transducer_stateless8` | Zipformer | Embedding + Conv1d | Same as pruned_transducer_stateless7, but using extra data from GigaSpeech|
| `pruned_stateless_emformer_rnnt2` | Emformer(from torchaudio) | Embedding + Conv1d | Using Emformer from torchaudio for streaming ASR|
| `conv_emformer_transducer_stateless` | ConvEmformer | Embedding + Conv1d | Using ConvEmformer for streaming ASR + mechanisms in reworked model |

79
egs/librispeech/ASR/RESULTS.md
View File

@ -1,5 +1,84 @@
## Results
### pruned_transducer_stateless7_ctc (zipformer with transducer loss and ctc loss)
See <https://github.com/k2-fsa/icefall/pull/683> for more details.
[pruned_transducer_stateless7_ctc](./pruned_transducer_stateless7_ctc)
The tensorboard log can be found at
<https://tensorboard.dev/experiment/hxlGAhOPToGmRLZFnAzPWw/>
You can find a pretrained model, training logs, decoding logs, and decoding
results at:
<https://huggingface.co/Zengwei/icefall-asr-librispeech-pruned-transducer-stateless7-ctc-2022-12-01>
Number of model parameters: 70561891, i.e., 70.56 M
| | test-clean | test-other | comment |
|--------------------------|------------|-------------|--------------------|
| greedy search | 2.23 | 5.19 | --epoch 30 --avg 8 |
| modified beam search | 2.21 | 5.12 | --epoch 30 --avg 8 |
| fast beam search | 2.23 | 5.18 | --epoch 30 --avg 8 |
| ctc decoding | 2.48 | 5.82 | --epoch 30 --avg 9 |
| 1best | 2.43 | 5.22 | --epoch 30 --avg 9 |
| nbest | 2.43 | 5.22 | --epoch 30 --avg 9 |
| nbest rescoring | 2.34 | 5.05 | --epoch 30 --avg 9 |
| whole lattice rescoring | 2.34 | 5.04 | --epoch 30 --avg 9 |
The training commands are:
```bash
export CUDA_VISIBLE_DEVICES="0,1,2,3"
./pruned_transducer_stateless7_ctc/train.py \
--world-size 4 \
--num-epochs 30 \
--full-libri 1 \
--use-fp16 1 \
--max-duration 750 \
--exp-dir pruned_transducer_stateless7_ctc/exp \
--feedforward-dims "1024,1024,2048,2048,1024" \
--ctc-loss-scale 0.2 \
--master-port 12535
```
The decoding commands for the transducer branch are:
```bash
for m in greedy_search fast_beam_search modified_beam_search ; do
for epoch in 30; do
for avg in 8; do
./pruned_transducer_stateless7_ctc/decode.py \
--epoch $epoch \
--avg $avg \
--use-averaged-model 1 \
--exp-dir ./pruned_transducer_stateless7_ctc/exp \
--feedforward-dims "1024,1024,2048,2048,1024" \
--max-duration 600 \
--decoding-method $m
done
done
done
```
The decoding commands for the ctc branch are:
```bash
for m in ctc-decoding nbest nbest-rescoring whole-lattice-rescoring; do
for epoch in 30; do
for avg in 9; do
./pruned_transducer_stateless7_ctc/ctc_decode.py \
--epoch $epoch \
--avg $avg \
--exp-dir ./pruned_transducer_stateless7_ctc/exp \
--max-duration 100 \
--decoding-method $m \
--hlg-scale 0.6 \
--lm-dir data/lm
done
done
done
```
### LibriSpeech BPE training results (Conformer CTC, supporting delay penalty)
#### [conformer_ctc3](./conformer_ctc3)

20
egs/librispeech/ASR/conformer_ctc3/jit_pretrained.py
View File

@ -23,40 +23,44 @@ Usage (for non-streaming mode):
(1) ctc-decoding
./conformer_ctc3/pretrained.py \
--checkpoint conformer_ctc3/exp/pretrained.pt \
--nn-model-filename ./conformer_ctc3/exp/cpu_jit.pt \
--bpe-model data/lang_bpe_500/bpe.model \
--method ctc-decoding \
--sample-rate 16000 \
test_wavs/1089-134686-0001.wav
/path/to/foo.wav \
/path/to/bar.wav
(2) 1best
./conformer_ctc3/pretrained.py \
--checkpoint conformer_ctc3/exp/pretrained.pt \
--nn-model-filename ./conformer_ctc3/exp/cpu_jit.pt \
--HLG data/lang_bpe_500/HLG.pt \
--words-file data/lang_bpe_500/words.txt \
--method 1best \
--sample-rate 16000 \
test_wavs/1089-134686-0001.wav
/path/to/foo.wav \
/path/to/bar.wav
(3) nbest-rescoring
./conformer_ctc3/pretrained.py \
--checkpoint conformer_ctc3/exp/pretrained.pt \
--nn-model-filename ./conformer_ctc3/exp/cpu_jit.pt \
--HLG data/lang_bpe_500/HLG.pt \
--words-file data/lang_bpe_500/words.txt \
--G data/lm/G_4_gram.pt \
--method nbest-rescoring \
--sample-rate 16000 \
test_wavs/1089-134686-0001.wav
/path/to/foo.wav \
/path/to/bar.wav
(4) whole-lattice-rescoring
./conformer_ctc3/pretrained.py \
--checkpoint conformer_ctc3/exp/pretrained.pt \
--nn-model-filename ./conformer_ctc3/exp/cpu_jit.pt \
--HLG data/lang_bpe_500/HLG.pt \
--words-file data/lang_bpe_500/words.txt \
--G data/lm/G_4_gram.pt \
--method whole-lattice-rescoring \
--sample-rate 16000 \
test_wavs/1089-134686-0001.wav
/path/to/foo.wav \
/path/to/bar.wav
"""

0
egs/librispeech/ASR/pruned_transducer_stateless7_ctc/__init__.py Normal file
View File

1
egs/librispeech/ASR/pruned_transducer_stateless7_ctc/asr_datamodule.py Symbolic link
View File

@ -0,0 +1 @@
../pruned_transducer_stateless2/asr_datamodule.py

1
egs/librispeech/ASR/pruned_transducer_stateless7_ctc/beam_search.py Symbolic link
View File

@ -0,0 +1 @@
../pruned_transducer_stateless2/beam_search.py

818
egs/librispeech/ASR/pruned_transducer_stateless7_ctc/ctc_decode.py Executable file
View File

@ -0,0 +1,818 @@
#!/usr/bin/env python3
#
# Copyright 2021-2022 Xiaomi Corporation (Author: Fangjun Kuang,
# Liyong Guo,
# Quandong Wang,
# 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.
"""
Usage:
(1) ctc-decoding
./pruned_transducer_stateless7_ctc/ctc_decode.py \
--epoch 30 \
--avg 15 \
--exp-dir ./pruned_transducer_stateless7_ctc/exp \
--max-duration 600 \
--decoding-method ctc-decoding
(2) 1best
./pruned_transducer_stateless7_ctc/ctc_decode.py \
--epoch 30 \
--avg 15 \
--exp-dir ./pruned_transducer_stateless7_ctc/exp \
--max-duration 600 \
--hlg-scale 0.8 \
--decoding-method 1best
(3) nbest
./pruned_transducer_stateless7_ctc/ctc_decode.py \
--epoch 30 \
--avg 15 \
--exp-dir ./pruned_transducer_stateless7_ctc/exp \
--max-duration 600 \
--hlg-scale 0.8 \
--decoding-method 1best
(4) nbest-rescoring
./pruned_transducer_stateless7_ctc/ctc_decode.py \
--epoch 30 \
--avg 15 \
--exp-dir ./pruned_transducer_stateless7_ctc/exp \
--max-duration 600 \
--hlg-scale 0.8 \
--lm-dir data/lm \
--decoding-method nbest-rescoring
(5) whole-lattice-rescoring
./pruned_transducer_stateless7_ctc/ctc_decode.py \
--epoch 30 \
--avg 15 \
--exp-dir ./pruned_transducer_stateless7_ctc/exp \
--max-duration 600 \
--hlg-scale 0.8 \
--lm-dir data/lm \
--decoding-method whole-lattice-rescoring
"""
import argparse
import logging
import math
from collections import defaultdict
from pathlib import Path
from typing import Dict, List, Optional, Tuple
import k2
import sentencepiece as spm
import torch
import torch.nn as nn
from asr_datamodule import LibriSpeechAsrDataModule
from train import add_model_arguments, get_params, get_transducer_model
from icefall.checkpoint import (
average_checkpoints,
average_checkpoints_with_averaged_model,
find_checkpoints,
load_checkpoint,
)
from icefall.decode import (
get_lattice,
nbest_decoding,
nbest_oracle,
one_best_decoding,
rescore_with_n_best_list,
rescore_with_whole_lattice,
)
from icefall.lexicon import Lexicon
from icefall.utils import (
AttributeDict,
get_texts,
setup_logger,
store_transcripts,
str2bool,
write_error_stats,
)
LOG_EPS = math.log(1e-10)
def get_parser():
parser = argparse.ArgumentParser(
formatter_class=argparse.ArgumentDefaultsHelpFormatter
)
parser.add_argument(
"--epoch",
type=int,
default=30,
help="""It specifies the checkpoint to use for decoding.
Note: Epoch counts from 1.
You can specify --avg to use more checkpoints for model averaging.""",
)
parser.add_argument(
"--iter",
type=int,
default=0,
help="""If positive, --epoch is ignored and it
will use the checkpoint exp_dir/checkpoint-iter.pt.
You can specify --avg to use more checkpoints for model averaging.
""",
)
parser.add_argument(
"--avg",
type=int,
default=15,
help="Number of checkpoints to average. Automatically select "
"consecutive checkpoints before the checkpoint specified by "
"'--epoch' and '--iter'",
)
parser.add_argument(
"--use-averaged-model",
type=str2bool,
default=True,
help="Whether to load averaged model. Currently it only supports "
"using --epoch. If True, it would decode with the averaged model "
"over the epoch range from `epoch-avg` (excluded) to `epoch`."
"Actually only the models with epoch number of `epoch-avg` and "
"`epoch` are loaded for averaging. ",
)
parser.add_argument(
"--exp-dir",
type=str,
default="pruned_transducer_stateless7_ctc/exp",
help="The experiment dir",
)
parser.add_argument(
"--bpe-model",
type=str,
default="data/lang_bpe_500/bpe.model",
help="Path to the BPE model",
)
parser.add_argument(
"--lang-dir",
type=Path,
default="data/lang_bpe_500",
help="The lang dir containing word table and LG graph",
)
parser.add_argument(
"--context-size",
type=int,
default=2,
help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
)
parser.add_argument(
"--decoding-method",
type=str,
default="ctc-decoding",
help="""Decoding method.
Supported values are:
- (1) ctc-decoding. Use CTC decoding. It uses a sentence piece
model, i.e., lang_dir/bpe.model, to convert word pieces to words.
It needs neither a lexicon nor an n-gram LM.
- (2) 1best. Extract the best path from the decoding lattice as the
decoding result.
- (3) nbest. Extract n paths from the decoding lattice; the path
with the highest score is the decoding result.
- (4) nbest-rescoring. Extract n paths from the decoding lattice,
rescore them with an n-gram LM (e.g., a 4-gram LM), the path with
the highest score is the decoding result.
- (5) whole-lattice-rescoring. Rescore the decoding lattice with an
n-gram LM (e.g., a 4-gram LM), the best path of rescored lattice
is the decoding result.
you have trained an RNN LM using ./rnn_lm/train.py
- (6) nbest-oracle. Its WER is the lower bound of any n-best
rescoring method can achieve. Useful for debugging n-best
rescoring method.
""",
)
parser.add_argument(
"--num-paths",
type=int,
default=100,
help="""Number of paths for n-best based decoding method.
Used only when "method" is one of the following values:
nbest, nbest-rescoring, and nbest-oracle
""",
)
parser.add_argument(
"--nbest-scale",
type=float,
default=0.5,
help="""The scale to be applied to `lattice.scores`.
It's needed if you use any kinds of n-best based rescoring.
Used only when "method" is one of the following values:
nbest, nbest-rescoring, and nbest-oracle
A smaller value results in more unique paths.
""",
)
parser.add_argument(
"--hlg-scale",
type=float,
default=0.8,
help="""The scale to be applied to `hlg.scores`.
""",
)
parser.add_argument(
"--lm-dir",
type=str,
default="data/lm",
help="""The n-gram LM dir.
It should contain either G_4_gram.pt or G_4_gram.fst.txt
""",
)
add_model_arguments(parser)
return parser
def get_decoding_params() -> AttributeDict:
"""Parameters for decoding."""
params = AttributeDict(
{
"frame_shift_ms": 10,
"search_beam": 20,
"output_beam": 8,
"min_active_states": 30,
"max_active_states": 10000,
"use_double_scores": True,
}
)
return params
def decode_one_batch(
params: AttributeDict,
model: nn.Module,
HLG: Optional[k2.Fsa],
H: Optional[k2.Fsa],
bpe_model: Optional[spm.SentencePieceProcessor],
batch: dict,
word_table: k2.SymbolTable,
G: Optional[k2.Fsa] = None,
) -> Dict[str, List[List[str]]]:
"""Decode one batch and return the result in a dict. The dict has the
following format:
- key: It indicates the setting used for decoding. For example,
if no rescoring is used, the key is the string `no_rescore`.
If LM rescoring is used, the key is the string `lm_scale_xxx`,
where `xxx` is the value of `lm_scale`. An example key is
`lm_scale_0.7`
- value: It contains the decoding result. `len(value)` equals to
batch size. `value[i]` is the decoding result for the i-th
utterance in the given batch.
Args:
params:
It's the return value of :func:`get_params`.
- params.decoding_method is "1best", it uses 1best decoding without LM rescoring.
- params.decoding_method is "nbest", it uses nbest decoding without LM rescoring.
- params.decoding_method is "nbest-rescoring", it uses nbest LM rescoring.
- params.decoding_method is "whole-lattice-rescoring", it uses whole lattice LM
rescoring.
model:
The neural model.
HLG:
The decoding graph. Used only when params.decoding_method is NOT ctc-decoding.
H:
The ctc topo. Used only when params.decoding_method is ctc-decoding.
bpe_model:
The BPE model. Used only when params.decoding_method is ctc-decoding.
batch:
It is the return value from iterating
`lhotse.dataset.K2SpeechRecognitionDataset`. See its documentation
for the format of the `batch`.
word_table:
The word symbol table.
G:
An LM. It is not None when params.decoding_method is "nbest-rescoring"
or "whole-lattice-rescoring". In general, the G in HLG
is a 3-gram LM, while this G is a 4-gram LM.
Returns:
Return the decoding result. See above description for the format of
the returned dict. Note: If it decodes to nothing, then return None.
"""
if HLG is not None:
device = HLG.device
else:
device = H.device
feature = batch["inputs"]
assert feature.ndim == 3
feature = feature.to(device)
# at entry, feature is (N, T, C)
supervisions = batch["supervisions"]
feature_lens = supervisions["num_frames"].to(device)
encoder_out, encoder_out_lens = model.encoder(feature, feature_lens)
nnet_output = model.ctc_output(encoder_out)
# nnet_output is (N, T, C)
supervision_segments = torch.stack(
(
supervisions["sequence_idx"],
supervisions["start_frame"] // params.subsampling_factor,
supervisions["num_frames"] // params.subsampling_factor,
),
1,
).to(torch.int32)
if H is None:
assert HLG is not None
decoding_graph = HLG
else:
assert HLG is None
assert bpe_model is not None
decoding_graph = H
lattice = get_lattice(
nnet_output=nnet_output,
decoding_graph=decoding_graph,
supervision_segments=supervision_segments,
search_beam=params.search_beam,
output_beam=params.output_beam,
min_active_states=params.min_active_states,
max_active_states=params.max_active_states,
subsampling_factor=params.subsampling_factor,
)
if params.decoding_method == "ctc-decoding":
best_path = one_best_decoding(
lattice=lattice, use_double_scores=params.use_double_scores
)
# Note: `best_path.aux_labels` contains token IDs, not word IDs
# since we are using H, not HLG here.
#
# token_ids is a lit-of-list of IDs
token_ids = get_texts(best_path)
# hyps is a list of str, e.g., ['xxx yyy zzz', ...]
hyps = bpe_model.decode(token_ids)
# hyps is a list of list of str, e.g., [['xxx', 'yyy', 'zzz'], ... ]
hyps = [s.split() for s in hyps]
key = "ctc-decoding"
return {key: hyps}
if params.decoding_method == "nbest-oracle":
# Note: You can also pass rescored lattices to it.
# We choose the HLG decoded lattice for speed reasons
# as HLG decoding is faster and the oracle WER
# is only slightly worse than that of rescored lattices.
best_path = nbest_oracle(
lattice=lattice,
num_paths=params.num_paths,
ref_texts=supervisions["text"],
word_table=word_table,
nbest_scale=params.nbest_scale,
oov="<UNK>",
)
hyps = get_texts(best_path)
hyps = [[word_table[i] for i in ids] for ids in hyps]
key = f"oracle_{params.num_paths}_nbest_scale_{params.nbest_scale}" # noqa
return {key: hyps}
if params.decoding_method in ["1best", "nbest"]:
if params.decoding_method == "1best":
best_path = one_best_decoding(
lattice=lattice, use_double_scores=params.use_double_scores
)
key = "no_rescore"
else:
best_path = nbest_decoding(
lattice=lattice,
num_paths=params.num_paths,
use_double_scores=params.use_double_scores,
nbest_scale=params.nbest_scale,
)
key = f"no_rescore-nbest-scale-{params.nbest_scale}-{params.num_paths}" # noqa
hyps = get_texts(best_path)
hyps = [[word_table[i] for i in ids] for ids in hyps]
return {key: hyps}
assert params.decoding_method in [
"nbest-rescoring",
"whole-lattice-rescoring",
]
lm_scale_list = [0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7]
lm_scale_list += [0.8, 0.9, 1.0, 1.1, 1.2, 1.3]
lm_scale_list += [1.4, 1.5, 1.6, 1.7, 1.8, 1.9, 2.0]
if params.decoding_method == "nbest-rescoring":
best_path_dict = rescore_with_n_best_list(
lattice=lattice,
G=G,
num_paths=params.num_paths,
lm_scale_list=lm_scale_list,
nbest_scale=params.nbest_scale,
)
elif params.decoding_method == "whole-lattice-rescoring":
best_path_dict = rescore_with_whole_lattice(
lattice=lattice,
G_with_epsilon_loops=G,
lm_scale_list=lm_scale_list,
)
else:
assert False, f"Unsupported decoding method: {params.decoding_method}"
ans = dict()
if best_path_dict is not None:
for lm_scale_str, best_path in best_path_dict.items():
hyps = get_texts(best_path)
hyps = [[word_table[i] for i in ids] for ids in hyps]
ans[lm_scale_str] = hyps
else:
ans = None
return ans
def decode_dataset(
dl: torch.utils.data.DataLoader,
params: AttributeDict,
model: nn.Module,
HLG: Optional[k2.Fsa],
H: Optional[k2.Fsa],
bpe_model: Optional[spm.SentencePieceProcessor],
word_table: k2.SymbolTable,
G: Optional[k2.Fsa] = None,
) -> Dict[str, List[Tuple[str, List[str], List[str]]]]:
"""Decode dataset.
Args:
dl:
PyTorch's dataloader containing the dataset to decode.
params:
It is returned by :func:`get_params`.
model:
The neural model.
HLG:
The decoding graph. Used only when params.decoding_method is NOT ctc-decoding.
H:
The ctc topo. Used only when params.decoding_method is ctc-decoding.
bpe_model:
The BPE model. Used only when params.decoding_method is ctc-decoding.
word_table:
It is the word symbol table.
G:
An LM. It is not None when params.decoding_method is "nbest-rescoring"
or "whole-lattice-rescoring". In general, the G in HLG
is a 3-gram LM, while this G is a 4-gram LM.
Returns:
Return a dict, whose key may be "no-rescore" if no LM rescoring
is used, or it may be "lm_scale_0.7" if LM rescoring is used.
Its value is a list of tuples. Each tuple contains two elements:
The first is the reference transcript, and the second is the
predicted result.
"""
num_cuts = 0
try:
num_batches = len(dl)
except TypeError:
num_batches = "?"
results = defaultdict(list)
for batch_idx, batch in enumerate(dl):
texts = batch["supervisions"]["text"]
cut_ids = [cut.id for cut in batch["supervisions"]["cut"]]
hyps_dict = decode_one_batch(
params=params,
model=model,
HLG=HLG,
H=H,
bpe_model=bpe_model,
batch=batch,
word_table=word_table,
G=G,
)
for name, hyps in hyps_dict.items():
this_batch = []
assert len(hyps) == len(texts)
for cut_id, hyp_words, ref_text in zip(cut_ids, hyps, texts):
ref_words = ref_text.split()
this_batch.append((cut_id, ref_words, hyp_words))
results[name].extend(this_batch)
num_cuts += len(texts)
if batch_idx % 100 == 0:
batch_str = f"{batch_idx}/{num_batches}"
logging.info(f"batch {batch_str}, cuts processed until now is {num_cuts}")
return results
def save_results(
params: AttributeDict,
test_set_name: str,
results_dict: Dict[str, List[Tuple[str, List[str], List[str]]]],
):
test_set_wers = dict()
for key, results in results_dict.items():
recog_path = (
params.res_dir / f"recogs-{test_set_name}-{key}-{params.suffix}.txt"
)
results = sorted(results)
store_transcripts(filename=recog_path, texts=results)
logging.info(f"The transcripts are stored in {recog_path}")
# The following prints out WERs, per-word error statistics and aligned
# ref/hyp pairs.
errs_filename = (
params.res_dir / f"errs-{test_set_name}-{key}-{params.suffix}.txt"
)
with open(errs_filename, "w") as f:
wer = write_error_stats(f, f"{test_set_name}-{key}", results)
test_set_wers[key] = wer
logging.info("Wrote detailed error stats to {}".format(errs_filename))
test_set_wers = sorted(test_set_wers.items(), key=lambda x: x[1])
errs_info = (
params.res_dir / f"wer-summary-{test_set_name}-{key}-{params.suffix}.txt"
)
with open(errs_info, "w") as f:
print("settings\tWER", file=f)
for key, val in test_set_wers:
print("{}\t{}".format(key, val), file=f)
s = "\nFor {}, WER of different settings are:\n".format(test_set_name)
note = "\tbest for {}".format(test_set_name)
for key, val in test_set_wers:
s += "{}\t{}{}\n".format(key, val, note)
note = ""
logging.info(s)
@torch.no_grad()
def main():
parser = get_parser()
LibriSpeechAsrDataModule.add_arguments(parser)
args = parser.parse_args()
args.exp_dir = Path(args.exp_dir)
args.lang_dir = Path(args.lang_dir)
args.lm_dir = Path(args.lm_dir)
params = get_params()
# add decoding params
params.update(get_decoding_params())
params.update(vars(args))
assert params.decoding_method in (
"ctc-decoding",
"1best",
"nbest",
"nbest-rescoring",
"whole-lattice-rescoring",
"nbest-oracle",
)
params.res_dir = params.exp_dir / params.decoding_method
if params.iter > 0:
params.suffix = f"iter-{params.iter}-avg-{params.avg}"
else:
params.suffix = f"epoch-{params.epoch}-avg-{params.avg}"
if params.use_averaged_model:
params.suffix += "-use-averaged-model"
setup_logger(f"{params.res_dir}/log-decode-{params.suffix}")
logging.info("Decoding started")
device = torch.device("cpu")
if torch.cuda.is_available():
device = torch.device("cuda", 0)
logging.info(f"Device: {device}")
logging.info(params)
lexicon = Lexicon(params.lang_dir)
max_token_id = max(lexicon.tokens)
num_classes = max_token_id + 1 # +1 for the blank
params.vocab_size = num_classes
# <blk> and <unk> are defined in local/train_bpe_model.py
params.blank_id = 0
if params.decoding_method == "ctc-decoding":
HLG = None
H = k2.ctc_topo(
max_token=max_token_id,
modified=False,
device=device,
)
bpe_model = spm.SentencePieceProcessor()
bpe_model.load(str(params.lang_dir / "bpe.model"))
else:
H = None
bpe_model = None
HLG = k2.Fsa.from_dict(
torch.load(f"{params.lang_dir}/HLG.pt", map_location=device)
)
assert HLG.requires_grad is False
HLG.scores *= params.hlg_scale
if not hasattr(HLG, "lm_scores"):
HLG.lm_scores = HLG.scores.clone()
if params.decoding_method in (
"nbest-rescoring",
"whole-lattice-rescoring",
):
if not (params.lm_dir / "G_4_gram.pt").is_file():
logging.info("Loading G_4_gram.fst.txt")
logging.warning("It may take 8 minutes.")
with open(params.lm_dir / "G_4_gram.fst.txt") as f:
first_word_disambig_id = lexicon.word_table["#0"]
G = k2.Fsa.from_openfst(f.read(), acceptor=False)
# G.aux_labels is not needed in later computations, so
# remove it here.
del G.aux_labels
# CAUTION: The following line is crucial.
# Arcs entering the back-off state have label equal to #0.
# We have to change it to 0 here.
G.labels[G.labels >= first_word_disambig_id] = 0
# See https://github.com/k2-fsa/k2/issues/874
# for why we need to set G.properties to None
G.__dict__["_properties"] = None
G = k2.Fsa.from_fsas([G]).to(device)
G = k2.arc_sort(G)
# Save a dummy value so that it can be loaded in C++.
# See https://github.com/pytorch/pytorch/issues/67902
# for why we need to do this.
G.dummy = 1
torch.save(G.as_dict(), params.lm_dir / "G_4_gram.pt")
else:
logging.info("Loading pre-compiled G_4_gram.pt")
d = torch.load(params.lm_dir / "G_4_gram.pt", map_location=device)
G = k2.Fsa.from_dict(d)
if params.decoding_method == "whole-lattice-rescoring":
# Add epsilon self-loops to G as we will compose
# it with the whole lattice later
G = k2.add_epsilon_self_loops(G)
G = k2.arc_sort(G)
G = G.to(device)
# G.lm_scores is used to replace HLG.lm_scores during
# LM rescoring.
G.lm_scores = G.scores.clone()
else:
G = None
logging.info("About to create model")
model = get_transducer_model(params)
if not params.use_averaged_model:
if params.iter > 0:
filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
: params.avg
]
if len(filenames) == 0:
raise ValueError(
f"No checkpoints found for"
f" --iter {params.iter}, --avg {params.avg}"
)
elif len(filenames) < params.avg:
raise ValueError(
f"Not enough checkpoints ({len(filenames)}) found for"
f" --iter {params.iter}, --avg {params.avg}"
)
logging.info(f"averaging {filenames}")
model.to(device)
model.load_state_dict(average_checkpoints(filenames, device=device))
elif params.avg == 1:
load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
else:
start = params.epoch - params.avg + 1
filenames = []
for i in range(start, params.epoch + 1):
if i >= 1:
filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
logging.info(f"averaging {filenames}")
model.to(device)
model.load_state_dict(average_checkpoints(filenames, device=device))
else:
if params.iter > 0:
filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
: params.avg + 1
]
if len(filenames) == 0:
raise ValueError(
f"No checkpoints found for"
f" --iter {params.iter}, --avg {params.avg}"
)
elif len(filenames) < params.avg + 1:
raise ValueError(
f"Not enough checkpoints ({len(filenames)}) found for"
f" --iter {params.iter}, --avg {params.avg}"
)
filename_start = filenames[-1]
filename_end = filenames[0]
logging.info(
"Calculating the averaged model over iteration checkpoints"
f" from {filename_start} (excluded) to {filename_end}"
)
model.to(device)
model.load_state_dict(
average_checkpoints_with_averaged_model(
filename_start=filename_start,
filename_end=filename_end,
device=device,
)
)
else:
assert params.avg > 0, params.avg
start = params.epoch - params.avg
assert start >= 1, start
filename_start = f"{params.exp_dir}/epoch-{start}.pt"
filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt"
logging.info(
f"Calculating the averaged model over epoch range from "
f"{start} (excluded) to {params.epoch}"
)
model.to(device)
model.load_state_dict(
average_checkpoints_with_averaged_model(
filename_start=filename_start,
filename_end=filename_end,
device=device,
)
)
model.to(device)
model.eval()
num_param = sum([p.numel() for p in model.parameters()])
logging.info(f"Number of model parameters: {num_param}")
# we need cut ids to display recognition results.
args.return_cuts = True
librispeech = LibriSpeechAsrDataModule(args)
test_clean_cuts = librispeech.test_clean_cuts()
test_other_cuts = librispeech.test_other_cuts()
test_clean_dl = librispeech.test_dataloaders(test_clean_cuts)
test_other_dl = librispeech.test_dataloaders(test_other_cuts)
test_sets = ["test-clean", "test-other"]
test_dl = [test_clean_dl, test_other_dl]
for test_set, test_dl in zip(test_sets, test_dl):
results_dict = decode_dataset(
dl=test_dl,
params=params,
model=model,
HLG=HLG,
H=H,
bpe_model=bpe_model,
word_table=lexicon.word_table,
G=G,
)
save_results(
params=params,
test_set_name=test_set,
results_dict=results_dict,
)
logging.info("Done!")
if __name__ == "__main__":
main()

841
egs/librispeech/ASR/pruned_transducer_stateless7_ctc/decode.py Executable file
View File

@ -0,0 +1,841 @@
#!/usr/bin/env python3
#
# Copyright 2021-2022 Xiaomi Corporation (Author: Fangjun Kuang,
# 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.
"""
Usage:
(1) greedy search
./pruned_transducer_stateless7_ctc/decode.py \
--epoch 28 \
--avg 15 \
--exp-dir ./pruned_transducer_stateless7_ctc/exp \
--max-duration 600 \
--decoding-method greedy_search
(2) beam search (not recommended)
./pruned_transducer_stateless7_ctc/decode.py \
--epoch 28 \
--avg 15 \
--exp-dir ./pruned_transducer_stateless7_ctc/exp \
--max-duration 600 \
--decoding-method beam_search \
--beam-size 4
(3) modified beam search
./pruned_transducer_stateless7_ctc/decode.py \
--epoch 28 \
--avg 15 \
--exp-dir ./pruned_transducer_stateless7_ctc/exp \
--max-duration 600 \
--decoding-method modified_beam_search \
--beam-size 4
(4) fast beam search (one best)
./pruned_transducer_stateless7_ctc/decode.py \
--epoch 28 \
--avg 15 \
--exp-dir ./pruned_transducer_stateless7_ctc/exp \
--max-duration 600 \
--decoding-method fast_beam_search \
--beam 20.0 \
--max-contexts 8 \
--max-states 64
(5) fast beam search (nbest)
./pruned_transducer_stateless7_ctc/decode.py \
--epoch 28 \
--avg 15 \
--exp-dir ./pruned_transducer_stateless7_ctc/exp \
--max-duration 600 \
--decoding-method fast_beam_search_nbest \
--beam 20.0 \
--max-contexts 8 \
--max-states 64 \
--num-paths 200 \
--nbest-scale 0.5
(6) fast beam search (nbest oracle WER)
./pruned_transducer_stateless7_ctc/decode.py \
--epoch 28 \
--avg 15 \
--exp-dir ./pruned_transducer_stateless7_ctc/exp \
--max-duration 600 \
--decoding-method fast_beam_search_nbest_oracle \
--beam 20.0 \
--max-contexts 8 \
--max-states 64 \
--num-paths 200 \
--nbest-scale 0.5
(7) fast beam search (with LG)
./pruned_transducer_stateless7_ctc/decode.py \
--epoch 28 \
--avg 15 \
--exp-dir ./pruned_transducer_stateless7_ctc/exp \
--max-duration 600 \
--decoding-method fast_beam_search_nbest_LG \
--beam 20.0 \
--max-contexts 8 \
--max-states 64
"""
import argparse
import logging
import math
from collections import defaultdict
from pathlib import Path
from typing import Dict, List, Optional, Tuple
import k2
import sentencepiece as spm
import torch
import torch.nn as nn
from asr_datamodule import LibriSpeechAsrDataModule
from beam_search import (
beam_search,
fast_beam_search_nbest,
fast_beam_search_nbest_LG,
fast_beam_search_nbest_oracle,
fast_beam_search_one_best,
greedy_search,
greedy_search_batch,
modified_beam_search,
)
from train import add_model_arguments, get_params, get_transducer_model
from icefall.checkpoint import (
average_checkpoints,
average_checkpoints_with_averaged_model,
find_checkpoints,
load_checkpoint,
)
from icefall.lexicon import Lexicon
from icefall.utils import (
AttributeDict,
setup_logger,
store_transcripts,
str2bool,
write_error_stats,
)
LOG_EPS = math.log(1e-10)
def get_parser():
parser = argparse.ArgumentParser(
formatter_class=argparse.ArgumentDefaultsHelpFormatter
)
parser.add_argument(
"--epoch",
type=int,
default=30,
help="""It specifies the checkpoint to use for decoding.
Note: Epoch counts from 1.
You can specify --avg to use more checkpoints for model averaging.""",
)
parser.add_argument(
"--iter",
type=int,
default=0,
help="""If positive, --epoch is ignored and it
will use the checkpoint exp_dir/checkpoint-iter.pt.
You can specify --avg to use more checkpoints for model averaging.
""",
)
parser.add_argument(
"--avg",
type=int,
default=9,
help="Number of checkpoints to average. Automatically select "
"consecutive checkpoints before the checkpoint specified by "
"'--epoch' and '--iter'",
)
parser.add_argument(
"--use-averaged-model",
type=str2bool,
default=True,
help="Whether to load averaged model. Currently it only supports "
"using --epoch. If True, it would decode with the averaged model "
"over the epoch range from `epoch-avg` (excluded) to `epoch`."
"Actually only the models with epoch number of `epoch-avg` and "
"`epoch` are loaded for averaging. ",
)
parser.add_argument(
"--exp-dir",
type=str,
default="pruned_transducer_stateless7_ctc/exp",
help="The experiment dir",
)
parser.add_argument(
"--bpe-model",
type=str,
default="data/lang_bpe_500/bpe.model",
help="Path to the BPE model",
)
parser.add_argument(
"--lang-dir",
type=Path,
default="data/lang_bpe_500",
help="The lang dir containing word table and LG graph",
)
parser.add_argument(
"--decoding-method",
type=str,
default="greedy_search",
help="""Possible values are:
- greedy_search
- beam_search
- modified_beam_search
- fast_beam_search
- fast_beam_search_nbest
- fast_beam_search_nbest_oracle
- fast_beam_search_nbest_LG
If you use fast_beam_search_nbest_LG, you have to specify
`--lang-dir`, which should contain `LG.pt`.
""",
)
parser.add_argument(
"--beam-size",
type=int,
default=4,
help="""An integer indicating how many candidates we will keep for each
frame. Used only when --decoding-method is beam_search or
modified_beam_search.""",
)
parser.add_argument(
"--beam",
type=float,
default=20.0,
help="""A floating point value to calculate the cutoff score during beam
search (i.e., `cutoff = max-score - beam`), which is the same as the
`beam` in Kaldi.
Used only when --decoding-method is fast_beam_search,
fast_beam_search_nbest, fast_beam_search_nbest_LG,
and fast_beam_search_nbest_oracle
""",
)
parser.add_argument(
"--ngram-lm-scale",
type=float,
default=0.01,
help="""
Used only when --decoding_method is fast_beam_search_nbest_LG.
It specifies the scale for n-gram LM scores.
""",
)
parser.add_argument(
"--max-contexts",
type=int,
default=8,
help="""Used only when --decoding-method is
fast_beam_search, fast_beam_search_nbest, fast_beam_search_nbest_LG,
and fast_beam_search_nbest_oracle""",
)
parser.add_argument(
"--max-states",
type=int,
default=64,
help="""Used only when --decoding-method is
fast_beam_search, fast_beam_search_nbest, fast_beam_search_nbest_LG,
and fast_beam_search_nbest_oracle""",
)
parser.add_argument(
"--context-size",
type=int,
default=2,
help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
)
parser.add_argument(
"--max-sym-per-frame",
type=int,
default=1,
help="""Maximum number of symbols per frame.
Used only when --decoding_method is greedy_search""",
)
parser.add_argument(
"--num-paths",
type=int,
default=200,
help="""Number of paths for nbest decoding.
Used only when the decoding method is fast_beam_search_nbest,
fast_beam_search_nbest_LG, and fast_beam_search_nbest_oracle""",
)
parser.add_argument(
"--nbest-scale",
type=float,
default=0.5,
help="""Scale applied to lattice scores when computing nbest paths.
Used only when the decoding method is fast_beam_search_nbest,
fast_beam_search_nbest_LG, and fast_beam_search_nbest_oracle""",
)
parser.add_argument(
"--simulate-streaming",
type=str2bool,
default=False,
help="""Whether to simulate streaming in decoding, this is a good way to
test a streaming model.
""",
)
parser.add_argument(
"--decode-chunk-size",
type=int,
default=16,
help="The chunk size for decoding (in frames after subsampling)",
)
parser.add_argument(
"--left-context",
type=int,
default=64,
help="left context can be seen during decoding (in frames after subsampling)",
)
add_model_arguments(parser)
return parser
def decode_one_batch(
params: AttributeDict,
model: nn.Module,
sp: spm.SentencePieceProcessor,
batch: dict,
word_table: Optional[k2.SymbolTable] = None,
decoding_graph: Optional[k2.Fsa] = None,
) -> Dict[str, List[List[str]]]:
"""Decode one batch and return the result in a dict. The dict has the
following format:
- key: It indicates the setting used for decoding. For example,
if greedy_search is used, it would be "greedy_search"
If beam search with a beam size of 7 is used, it would be
"beam_7"
- value: It contains the decoding result. `len(value)` equals to
batch size. `value[i]` is the decoding result for the i-th
utterance in the given batch.
Args:
params:
It's the return value of :func:`get_params`.
model:
The neural model.
sp:
The BPE model.
batch:
It is the return value from iterating
`lhotse.dataset.K2SpeechRecognitionDataset`. See its documentation
for the format of the `batch`.
word_table:
The word symbol table.
decoding_graph:
The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used
only when --decoding_method is fast_beam_search, fast_beam_search_nbest,
fast_beam_search_nbest_oracle, and fast_beam_search_nbest_LG.
Returns:
Return the decoding result. See above description for the format of
the returned dict.
"""
device = next(model.parameters()).device
feature = batch["inputs"]
assert feature.ndim == 3
feature = feature.to(device)
# at entry, feature is (N, T, C)
supervisions = batch["supervisions"]
feature_lens = supervisions["num_frames"].to(device)
if params.simulate_streaming:
feature_lens += params.left_context
feature = torch.nn.functional.pad(
feature,
pad=(0, 0, 0, params.left_context),
value=LOG_EPS,
)
encoder_out, encoder_out_lens, _ = model.encoder.streaming_forward(
x=feature,
x_lens=feature_lens,
chunk_size=params.decode_chunk_size,
left_context=params.left_context,
simulate_streaming=True,
)
else:
encoder_out, encoder_out_lens = model.encoder(x=feature, x_lens=feature_lens)
hyps = []
if params.decoding_method == "fast_beam_search":
hyp_tokens = fast_beam_search_one_best(
model=model,
decoding_graph=decoding_graph,
encoder_out=encoder_out,
encoder_out_lens=encoder_out_lens,
beam=params.beam,
max_contexts=params.max_contexts,
max_states=params.max_states,
)
for hyp in sp.decode(hyp_tokens):
hyps.append(hyp.split())
elif params.decoding_method == "fast_beam_search_nbest_LG":
hyp_tokens = fast_beam_search_nbest_LG(
model=model,
decoding_graph=decoding_graph,
encoder_out=encoder_out,
encoder_out_lens=encoder_out_lens,
beam=params.beam,
max_contexts=params.max_contexts,
max_states=params.max_states,
num_paths=params.num_paths,
nbest_scale=params.nbest_scale,
)
for hyp in hyp_tokens:
hyps.append([word_table[i] for i in hyp])
elif params.decoding_method == "fast_beam_search_nbest":
hyp_tokens = fast_beam_search_nbest(
model=model,
decoding_graph=decoding_graph,
encoder_out=encoder_out,
encoder_out_lens=encoder_out_lens,
beam=params.beam,
max_contexts=params.max_contexts,
max_states=params.max_states,
num_paths=params.num_paths,
nbest_scale=params.nbest_scale,
)
for hyp in sp.decode(hyp_tokens):
hyps.append(hyp.split())
elif params.decoding_method == "fast_beam_search_nbest_oracle":
hyp_tokens = fast_beam_search_nbest_oracle(
model=model,
decoding_graph=decoding_graph,
encoder_out=encoder_out,
encoder_out_lens=encoder_out_lens,
beam=params.beam,
max_contexts=params.max_contexts,
max_states=params.max_states,
num_paths=params.num_paths,
ref_texts=sp.encode(supervisions["text"]),
nbest_scale=params.nbest_scale,
)
for hyp in sp.decode(hyp_tokens):
hyps.append(hyp.split())
elif params.decoding_method == "greedy_search" and params.max_sym_per_frame == 1:
hyp_tokens = greedy_search_batch(
model=model,
encoder_out=encoder_out,
encoder_out_lens=encoder_out_lens,
)
for hyp in sp.decode(hyp_tokens):
hyps.append(hyp.split())
elif params.decoding_method == "modified_beam_search":
hyp_tokens = modified_beam_search(
model=model,
encoder_out=encoder_out,
encoder_out_lens=encoder_out_lens,
beam=params.beam_size,
)
for hyp in sp.decode(hyp_tokens):
hyps.append(hyp.split())
else:
batch_size = encoder_out.size(0)
for i in range(batch_size):
# fmt: off
encoder_out_i = encoder_out[i:i+1, :encoder_out_lens[i]]
# fmt: on
if params.decoding_method == "greedy_search":
hyp = greedy_search(
model=model,
encoder_out=encoder_out_i,
max_sym_per_frame=params.max_sym_per_frame,
)
elif params.decoding_method == "beam_search":
hyp = beam_search(
model=model,
encoder_out=encoder_out_i,
beam=params.beam_size,
)
else:
raise ValueError(
f"Unsupported decoding method: {params.decoding_method}"
)
hyps.append(sp.decode(hyp).split())
if params.decoding_method == "greedy_search":
return {"greedy_search": hyps}
elif "fast_beam_search" in params.decoding_method:
key = f"beam_{params.beam}_"
key += f"max_contexts_{params.max_contexts}_"
key += f"max_states_{params.max_states}"
if "nbest" in params.decoding_method:
key += f"_num_paths_{params.num_paths}_"
key += f"nbest_scale_{params.nbest_scale}"
if "LG" in params.decoding_method:
key += f"_ngram_lm_scale_{params.ngram_lm_scale}"
return {key: hyps}
else:
return {f"beam_size_{params.beam_size}": hyps}
def decode_dataset(
dl: torch.utils.data.DataLoader,
params: AttributeDict,
model: nn.Module,
sp: spm.SentencePieceProcessor,
word_table: Optional[k2.SymbolTable] = None,
decoding_graph: Optional[k2.Fsa] = None,
) -> Dict[str, List[Tuple[str, List[str], List[str]]]]:
"""Decode dataset.
Args:
dl:
PyTorch's dataloader containing the dataset to decode.
params:
It is returned by :func:`get_params`.
model:
The neural model.
sp:
The BPE model.
word_table:
The word symbol table.
decoding_graph:
The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used
only when --decoding_method is fast_beam_search, fast_beam_search_nbest,
fast_beam_search_nbest_oracle, and fast_beam_search_nbest_LG.
Returns:
Return a dict, whose key may be "greedy_search" if greedy search
is used, or it may be "beam_7" if beam size of 7 is used.
Its value is a list of tuples. Each tuple contains two elements:
The first is the reference transcript, and the second is the
predicted result.
"""
num_cuts = 0
try:
num_batches = len(dl)
except TypeError:
num_batches = "?"
if params.decoding_method == "greedy_search":
log_interval = 50
else:
log_interval = 20
results = defaultdict(list)
for batch_idx, batch in enumerate(dl):
texts = batch["supervisions"]["text"]
cut_ids = [cut.id for cut in batch["supervisions"]["cut"]]
hyps_dict = decode_one_batch(
params=params,
model=model,
sp=sp,
decoding_graph=decoding_graph,
word_table=word_table,
batch=batch,
)
for name, hyps in hyps_dict.items():
this_batch = []
assert len(hyps) == len(texts)
for cut_id, hyp_words, ref_text in zip(cut_ids, hyps, texts):
ref_words = ref_text.split()
this_batch.append((cut_id, ref_words, hyp_words))
results[name].extend(this_batch)
num_cuts += len(texts)
if batch_idx % log_interval == 0:
batch_str = f"{batch_idx}/{num_batches}"
logging.info(f"batch {batch_str}, cuts processed until now is {num_cuts}")
return results
def save_results(
params: AttributeDict,
test_set_name: str,
results_dict: Dict[str, List[Tuple[str, List[str], List[str]]]],
):
test_set_wers = dict()
for key, results in results_dict.items():
recog_path = (
params.res_dir / f"recogs-{test_set_name}-{key}-{params.suffix}.txt"
)
results = sorted(results)
store_transcripts(filename=recog_path, texts=results)
logging.info(f"The transcripts are stored in {recog_path}")
# The following prints out WERs, per-word error statistics and aligned
# ref/hyp pairs.
errs_filename = (
params.res_dir / f"errs-{test_set_name}-{key}-{params.suffix}.txt"
)
with open(errs_filename, "w") as f:
wer = write_error_stats(
f, f"{test_set_name}-{key}", results, enable_log=True
)
test_set_wers[key] = wer
logging.info("Wrote detailed error stats to {}".format(errs_filename))
test_set_wers = sorted(test_set_wers.items(), key=lambda x: x[1])
errs_info = (
params.res_dir / f"wer-summary-{test_set_name}-{key}-{params.suffix}.txt"
)
with open(errs_info, "w") as f:
print("settings\tWER", file=f)
for key, val in test_set_wers:
print("{}\t{}".format(key, val), file=f)
s = "\nFor {}, WER of different settings are:\n".format(test_set_name)
note = "\tbest for {}".format(test_set_name)
for key, val in test_set_wers:
s += "{}\t{}{}\n".format(key, val, note)
note = ""
logging.info(s)
@torch.no_grad()
def main():
parser = get_parser()
LibriSpeechAsrDataModule.add_arguments(parser)
args = parser.parse_args()
args.exp_dir = Path(args.exp_dir)
params = get_params()
params.update(vars(args))
assert params.decoding_method in (
"greedy_search",
"beam_search",
"fast_beam_search",
"fast_beam_search_nbest",
"fast_beam_search_nbest_LG",
"fast_beam_search_nbest_oracle",
"modified_beam_search",
)
params.res_dir = params.exp_dir / params.decoding_method
if params.iter > 0:
params.suffix = f"iter-{params.iter}-avg-{params.avg}"
else:
params.suffix = f"epoch-{params.epoch}-avg-{params.avg}"
if params.simulate_streaming:
params.suffix += f"-streaming-chunk-size-{params.decode_chunk_size}"
params.suffix += f"-left-context-{params.left_context}"
if "fast_beam_search" in params.decoding_method:
params.suffix += f"-beam-{params.beam}"
params.suffix += f"-max-contexts-{params.max_contexts}"
params.suffix += f"-max-states-{params.max_states}"
if "nbest" in params.decoding_method:
params.suffix += f"-nbest-scale-{params.nbest_scale}"
params.suffix += f"-num-paths-{params.num_paths}"
if "LG" in params.decoding_method:
params.suffix += f"-ngram-lm-scale-{params.ngram_lm_scale}"
elif "beam_search" in params.decoding_method:
params.suffix += f"-{params.decoding_method}-beam-size-{params.beam_size}"
else:
params.suffix += f"-context-{params.context_size}"
params.suffix += f"-max-sym-per-frame-{params.max_sym_per_frame}"
if params.use_averaged_model:
params.suffix += "-use-averaged-model"
setup_logger(f"{params.res_dir}/log-decode-{params.suffix}")
logging.info("Decoding started")
device = torch.device("cpu")
if torch.cuda.is_available():
device = torch.device("cuda", 0)
logging.info(f"Device: {device}")
sp = spm.SentencePieceProcessor()
sp.load(params.bpe_model)
# <blk> and <unk> are defined in local/train_bpe_model.py
params.blank_id = sp.piece_to_id("<blk>")
params.unk_id = sp.piece_to_id("<unk>")
params.vocab_size = sp.get_piece_size()
if params.simulate_streaming:
assert (
params.causal_convolution
), "Decoding in streaming requires causal convolution"
logging.info(params)
logging.info("About to create model")
model = get_transducer_model(params)
if not params.use_averaged_model:
if params.iter > 0:
filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
: params.avg
]
if len(filenames) == 0:
raise ValueError(
f"No checkpoints found for"
f" --iter {params.iter}, --avg {params.avg}"
)
elif len(filenames) < params.avg:
raise ValueError(
f"Not enough checkpoints ({len(filenames)}) found for"
f" --iter {params.iter}, --avg {params.avg}"
)
logging.info(f"averaging {filenames}")
model.to(device)
model.load_state_dict(average_checkpoints(filenames, device=device))
elif params.avg == 1:
load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
else:
start = params.epoch - params.avg + 1
filenames = []
for i in range(start, params.epoch + 1):
if i >= 1:
filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
logging.info(f"averaging {filenames}")
model.to(device)
model.load_state_dict(average_checkpoints(filenames, device=device))
else:
if params.iter > 0:
filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
: params.avg + 1
]
if len(filenames) == 0:
raise ValueError(
f"No checkpoints found for"
f" --iter {params.iter}, --avg {params.avg}"
)
elif len(filenames) < params.avg + 1:
raise ValueError(
f"Not enough checkpoints ({len(filenames)}) found for"
f" --iter {params.iter}, --avg {params.avg}"
)
filename_start = filenames[-1]
filename_end = filenames[0]
logging.info(
"Calculating the averaged model over iteration checkpoints"
f" from {filename_start} (excluded) to {filename_end}"
)
model.to(device)
model.load_state_dict(
average_checkpoints_with_averaged_model(
filename_start=filename_start,
filename_end=filename_end,
device=device,
)
)
else:
assert params.avg > 0, params.avg
start = params.epoch - params.avg
assert start >= 1, start
filename_start = f"{params.exp_dir}/epoch-{start}.pt"
filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt"
logging.info(
f"Calculating the averaged model over epoch range from "
f"{start} (excluded) to {params.epoch}"
)
model.to(device)
model.load_state_dict(
average_checkpoints_with_averaged_model(
filename_start=filename_start,
filename_end=filename_end,
device=device,
)
)
model.to(device)
model.eval()
if "fast_beam_search" in params.decoding_method:
if params.decoding_method == "fast_beam_search_nbest_LG":
lexicon = Lexicon(params.lang_dir)
word_table = lexicon.word_table
lg_filename = params.lang_dir / "LG.pt"
logging.info(f"Loading {lg_filename}")
decoding_graph = k2.Fsa.from_dict(
torch.load(lg_filename, map_location=device)
)
decoding_graph.scores *= params.ngram_lm_scale
else:
word_table = None
decoding_graph = k2.trivial_graph(params.vocab_size - 1, device=device)
else:
decoding_graph = None
word_table = None
num_param = sum([p.numel() for p in model.parameters()])
logging.info(f"Number of model parameters: {num_param}")
# we need cut ids to display recognition results.
args.return_cuts = True
librispeech = LibriSpeechAsrDataModule(args)
test_clean_cuts = librispeech.test_clean_cuts()
test_other_cuts = librispeech.test_other_cuts()
test_clean_dl = librispeech.test_dataloaders(test_clean_cuts)
test_other_dl = librispeech.test_dataloaders(test_other_cuts)
test_sets = ["test-clean", "test-other"]
test_dl = [test_clean_dl, test_other_dl]
for test_set, test_dl in zip(test_sets, test_dl):
results_dict = decode_dataset(
dl=test_dl,
params=params,
model=model,
sp=sp,
word_table=word_table,
decoding_graph=decoding_graph,
)
save_results(
params=params,
test_set_name=test_set,
results_dict=results_dict,
)
logging.info("Done!")
if __name__ == "__main__":
main()

1
egs/librispeech/ASR/pruned_transducer_stateless7_ctc/decoder.py Symbolic link
View File

@ -0,0 +1 @@
../pruned_transducer_stateless7/decoder.py

1
egs/librispeech/ASR/pruned_transducer_stateless7_ctc/encoder_interface.py Symbolic link
View File

@ -0,0 +1 @@
../pruned_transducer_stateless2/encoder_interface.py

320
egs/librispeech/ASR/pruned_transducer_stateless7_ctc/export.py Executable file
View File

@ -0,0 +1,320 @@
#!/usr/bin/env python3
#
# Copyright 2021 Xiaomi Corporation (Author: Fangjun Kuang)
#
# 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.
# This script converts several saved checkpoints
# to a single one using model averaging.
"""
Usage:
(1) Export to torchscript model using torch.jit.script()
./pruned_transducer_stateless7_ctc/export.py \
--exp-dir ./pruned_transducer_stateless7_ctc/exp \
--bpe-model data/lang_bpe_500/bpe.model \
--epoch 30 \
--avg 9 \
--jit 1
It will generate a file `cpu_jit.pt` in the given `exp_dir`. You can later
load it by `torch.jit.load("cpu_jit.pt")`.
Note `cpu` in the name `cpu_jit.pt` means the parameters when loaded into Python
are on CPU. You can use `to("cuda")` to move them to a CUDA device.
Check
https://github.com/k2-fsa/sherpa
for how to use the exported models outside of icefall.
(2) Export `model.state_dict()`
./pruned_transducer_stateless7_ctc/export.py \
--exp-dir ./pruned_transducer_stateless7_ctc/exp \
--bpe-model data/lang_bpe_500/bpe.model \
--epoch 20 \
--avg 10
It will generate a file `pretrained.pt` in the given `exp_dir`. You can later
load it by `icefall.checkpoint.load_checkpoint()`.
To use the generated file with `pruned_transducer_stateless7_ctc/decode.py`,
you can do:
cd /path/to/exp_dir
ln -s pretrained.pt epoch-9999.pt
cd /path/to/egs/librispeech/ASR
./pruned_transducer_stateless7_ctc/decode.py \
--exp-dir ./pruned_transducer_stateless7_ctc/exp \
--epoch 9999 \
--avg 1 \
--max-duration 600 \
--decoding-method greedy_search \
--bpe-model data/lang_bpe_500/bpe.model
Check ./pretrained.py for its usage.
Note: If you don't want to train a model from scratch, we have
provided one for you. You can get it at
https://huggingface.co/csukuangfj/icefall-asr-librispeech-pruned-transducer-stateless7-2022-11-11
with the following commands:
sudo apt-get install git-lfs
git lfs install
git clone https://huggingface.co/csukuangfj/icefall-asr-librispeech-pruned-transducer-stateless7-2022-11-11
# You will find the pre-trained model in icefall-asr-librispeech-pruned-transducer-stateless7-2022-11-11/exp
"""
import argparse
import logging
from pathlib import Path
import sentencepiece as spm
import torch
from scaling_converter import convert_scaled_to_non_scaled
from train import add_model_arguments, get_params, get_transducer_model
from icefall.checkpoint import (
average_checkpoints,
average_checkpoints_with_averaged_model,
find_checkpoints,
load_checkpoint,
)
from icefall.utils import str2bool
def get_parser():
parser = argparse.ArgumentParser(
formatter_class=argparse.ArgumentDefaultsHelpFormatter
)
parser.add_argument(
"--epoch",
type=int,
default=30,
help="""It specifies the checkpoint to use for decoding.
Note: Epoch counts from 1.
You can specify --avg to use more checkpoints for model averaging.""",
)
parser.add_argument(
"--iter",
type=int,
default=0,
help="""If positive, --epoch is ignored and it
will use the checkpoint exp_dir/checkpoint-iter.pt.
You can specify --avg to use more checkpoints for model averaging.
""",
)
parser.add_argument(
"--avg",
type=int,
default=9,
help="Number of checkpoints to average. Automatically select "
"consecutive checkpoints before the checkpoint specified by "
"'--epoch' and '--iter'",
)
parser.add_argument(
"--use-averaged-model",
type=str2bool,
default=True,
help="Whether to load averaged model. Currently it only supports "
"using --epoch. If True, it would decode with the averaged model "
"over the epoch range from `epoch-avg` (excluded) to `epoch`."
"Actually only the models with epoch number of `epoch-avg` and "
"`epoch` are loaded for averaging. ",
)
parser.add_argument(
"--exp-dir",
type=str,
default="pruned_transducer_stateless7/exp",
help="""It specifies the directory where all training related
files, e.g., checkpoints, log, etc, are saved
""",
)
parser.add_argument(
"--bpe-model",
type=str,
default="data/lang_bpe_500/bpe.model",
help="Path to the BPE model",
)
parser.add_argument(
"--jit",
type=str2bool,
default=False,
help="""True to save a model after applying torch.jit.script.
It will generate a file named cpu_jit.pt
Check ./jit_pretrained.py for how to use it.
""",
)
parser.add_argument(
"--context-size",
type=int,
default=2,
help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
)
add_model_arguments(parser)
return parser
@torch.no_grad()
def main():
args = get_parser().parse_args()
args.exp_dir = Path(args.exp_dir)
params = get_params()
params.update(vars(args))
device = torch.device("cpu")
if torch.cuda.is_available():
device = torch.device("cuda", 0)
logging.info(f"device: {device}")
sp = spm.SentencePieceProcessor()
sp.load(params.bpe_model)
# <blk> is defined in local/train_bpe_model.py
params.blank_id = sp.piece_to_id("<blk>")
params.vocab_size = sp.get_piece_size()
logging.info(params)
logging.info("About to create model")
model = get_transducer_model(params)
model.to(device)
if not params.use_averaged_model:
if params.iter > 0:
filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
: params.avg
]
if len(filenames) == 0:
raise ValueError(
f"No checkpoints found for"
f" --iter {params.iter}, --avg {params.avg}"
)
elif len(filenames) < params.avg:
raise ValueError(
f"Not enough checkpoints ({len(filenames)}) found for"
f" --iter {params.iter}, --avg {params.avg}"
)
logging.info(f"averaging {filenames}")
model.to(device)
model.load_state_dict(average_checkpoints(filenames, device=device))
elif params.avg == 1:
load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model)
else:
start = params.epoch - params.avg + 1
filenames = []
for i in range(start, params.epoch + 1):
if i >= 1:
filenames.append(f"{params.exp_dir}/epoch-{i}.pt")
logging.info(f"averaging {filenames}")
model.to(device)
model.load_state_dict(average_checkpoints(filenames, device=device))
else:
if params.iter > 0:
filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[
: params.avg + 1
]
if len(filenames) == 0:
raise ValueError(
f"No checkpoints found for"
f" --iter {params.iter}, --avg {params.avg}"
)
elif len(filenames) < params.avg + 1:
raise ValueError(
f"Not enough checkpoints ({len(filenames)}) found for"
f" --iter {params.iter}, --avg {params.avg}"
)
filename_start = filenames[-1]
filename_end = filenames[0]
logging.info(
"Calculating the averaged model over iteration checkpoints"
f" from {filename_start} (excluded) to {filename_end}"
)
model.to(device)
model.load_state_dict(
average_checkpoints_with_averaged_model(
filename_start=filename_start,
filename_end=filename_end,
device=device,
)
)
else:
assert params.avg > 0, params.avg
start = params.epoch - params.avg
assert start >= 1, start
filename_start = f"{params.exp_dir}/epoch-{start}.pt"
filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt"
logging.info(
f"Calculating the averaged model over epoch range from "
f"{start} (excluded) to {params.epoch}"
)
model.to(device)
model.load_state_dict(
average_checkpoints_with_averaged_model(
filename_start=filename_start,
filename_end=filename_end,
device=device,
)
)
model.to("cpu")
model.eval()
if params.jit is True:
convert_scaled_to_non_scaled(model, inplace=True)
logging.info("Using torch.jit.script()")
# We won't use the forward() method of the model in C++, so just ignore
# it here.
# Otherwise, one of its arguments is a ragged tensor and is not
# torch scriptabe.
model.__class__.forward = torch.jit.ignore(model.__class__.forward)
logging.info("Using torch.jit.script")
model = torch.jit.script(model)
filename = params.exp_dir / "cpu_jit.pt"
model.save(str(filename))
logging.info(f"Saved to {filename}")
else:
logging.info("Not using torchscript. Export model.state_dict()")
# Save it using a format so that it can be loaded
# by :func:`load_checkpoint`
filename = params.exp_dir / "pretrained.pt"
torch.save({"model": model.state_dict()}, str(filename))
logging.info(f"Saved to {filename}")
if __name__ == "__main__":
formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
logging.basicConfig(format=formatter, level=logging.INFO)
main()

271
egs/librispeech/ASR/pruned_transducer_stateless7_ctc/jit_pretrained.py Executable file
View File

@ -0,0 +1,271 @@
#!/usr/bin/env python3
# Copyright 2022 Xiaomi Corp. (authors: Fangjun Kuang)
#
# 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.
"""
This script loads torchscript models, exported by `torch.jit.script()`
and uses them to decode waves.
You can use the following command to get the exported models:
./pruned_transducer_stateless7_ctc/export.py \
--exp-dir ./pruned_transducer_stateless7_ctc/exp \
--bpe-model data/lang_bpe_500/bpe.model \
--epoch 20 \
--avg 10 \
--jit 1
Usage of this script:
./pruned_transducer_stateless7_ctc/jit_pretrained.py \
--nn-model-filename ./pruned_transducer_stateless7_ctc/exp/cpu_jit.pt \
/path/to/foo.wav \
/path/to/bar.wav
"""
import argparse
import logging
import math
from typing import List
import kaldifeat
import sentencepiece as spm
import torch
import torchaudio
from torch.nn.utils.rnn import pad_sequence
def get_parser():
parser = argparse.ArgumentParser(
formatter_class=argparse.ArgumentDefaultsHelpFormatter
)
parser.add_argument(
"--nn-model-filename",
type=str,
required=True,
help="Path to the torchscript model cpu_jit.pt",
)
parser.add_argument(
"--bpe-model",
type=str,
help="""Path to bpe.model.""",
)
parser.add_argument(
"sound_files",
type=str,
nargs="+",
help="The input sound file(s) to transcribe. "
"Supported formats are those supported by torchaudio.load(). "
"For example, wav and flac are supported. "
"The sample rate has to be 16kHz.",
)
return parser
def read_sound_files(
filenames: List[str], expected_sample_rate: float = 16000
) -> List[torch.Tensor]:
"""Read a list of sound files into a list 1-D float32 torch tensors.
Args:
filenames:
A list of sound filenames.
expected_sample_rate:
The expected sample rate of the sound files.
Returns:
Return a list of 1-D float32 torch tensors.
"""
ans = []
for f in filenames:
wave, sample_rate = torchaudio.load(f)
assert (
sample_rate == expected_sample_rate
), f"Expected sample rate: {expected_sample_rate}. Given: {sample_rate}"
# We use only the first channel
ans.append(wave[0])
return ans
def greedy_search(
model: torch.jit.ScriptModule,
encoder_out: torch.Tensor,
encoder_out_lens: torch.Tensor,
) -> List[List[int]]:
"""Greedy search in batch mode. It hardcodes --max-sym-per-frame=1.
Args:
model:
The transducer model.
encoder_out:
A 3-D tensor of shape (N, T, C)
encoder_out_lens:
A 1-D tensor of shape (N,).
Returns:
Return the decoded results for each utterance.
"""
assert encoder_out.ndim == 3
assert encoder_out.size(0) >= 1, encoder_out.size(0)
packed_encoder_out = torch.nn.utils.rnn.pack_padded_sequence(
input=encoder_out,
lengths=encoder_out_lens.cpu(),
batch_first=True,
enforce_sorted=False,
)
device = encoder_out.device
blank_id = 0 # hard-code to 0
batch_size_list = packed_encoder_out.batch_sizes.tolist()
N = encoder_out.size(0)
assert torch.all(encoder_out_lens > 0), encoder_out_lens
assert N == batch_size_list[0], (N, batch_size_list)
context_size = model.decoder.context_size
hyps = [[blank_id] * context_size for _ in range(N)]
decoder_input = torch.tensor(
hyps,
device=device,
dtype=torch.int64,
) # (N, context_size)
decoder_out = model.decoder(
decoder_input,
need_pad=torch.tensor([False]),
).squeeze(1)
offset = 0
for batch_size in batch_size_list:
start = offset
end = offset + batch_size
current_encoder_out = packed_encoder_out.data[start:end]
current_encoder_out = current_encoder_out
# current_encoder_out's shape: (batch_size, encoder_out_dim)
offset = end
decoder_out = decoder_out[:batch_size]
logits = model.joiner(
current_encoder_out,
decoder_out,
)
# logits'shape (batch_size, vocab_size)
assert logits.ndim == 2, logits.shape
y = logits.argmax(dim=1).tolist()
emitted = False
for i, v in enumerate(y):
if v != blank_id:
hyps[i].append(v)
emitted = True
if emitted:
# update decoder output
decoder_input = [h[-context_size:] for h in hyps[:batch_size]]
decoder_input = torch.tensor(
decoder_input,
device=device,
dtype=torch.int64,
)
decoder_out = model.decoder(
decoder_input,
need_pad=torch.tensor([False]),
)
decoder_out = decoder_out.squeeze(1)
sorted_ans = [h[context_size:] for h in hyps]
ans = []
unsorted_indices = packed_encoder_out.unsorted_indices.tolist()
for i in range(N):
ans.append(sorted_ans[unsorted_indices[i]])
return ans
@torch.no_grad()
def main():
parser = get_parser()
args = parser.parse_args()
logging.info(vars(args))
device = torch.device("cpu")
if torch.cuda.is_available():
device = torch.device("cuda", 0)
logging.info(f"device: {device}")
model = torch.jit.load(args.nn_model_filename)
model.eval()
model.to(device)
sp = spm.SentencePieceProcessor()
sp.load(args.bpe_model)
logging.info("Constructing Fbank computer")
opts = kaldifeat.FbankOptions()
opts.device = device
opts.frame_opts.dither = 0
opts.frame_opts.snip_edges = False
opts.frame_opts.samp_freq = 16000
opts.mel_opts.num_bins = 80
fbank = kaldifeat.Fbank(opts)
logging.info(f"Reading sound files: {args.sound_files}")
waves = read_sound_files(
filenames=args.sound_files,
)
waves = [w.to(device) for w in waves]
logging.info("Decoding started")
features = fbank(waves)
feature_lengths = [f.size(0) for f in features]
features = pad_sequence(
features,
batch_first=True,
padding_value=math.log(1e-10),
)
feature_lengths = torch.tensor(feature_lengths, device=device)
encoder_out, encoder_out_lens = model.encoder(
x=features,
x_lens=feature_lengths,
)
hyps = greedy_search(
model=model,
encoder_out=encoder_out,
encoder_out_lens=encoder_out_lens,
)
s = "\n"
for filename, hyp in zip(args.sound_files, hyps):
words = sp.decode(hyp)
s += f"{filename}:\n{words}\n\n"
logging.info(s)
logging.info("Decoding Done")
if __name__ == "__main__":
formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
logging.basicConfig(format=formatter, level=logging.INFO)
main()

423
egs/librispeech/ASR/pruned_transducer_stateless7_ctc/jit_pretrained_ctc.py Executable file
View File

@ -0,0 +1,423 @@
#!/usr/bin/env python3
# Copyright 2022 Xiaomi Corp. (authors: Fangjun Kuang,
# 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.
"""
This script loads torchscript models, exported by `torch.jit.script()`
and uses them to decode waves.
You can use the following command to get the exported models:
./pruned_transducer_stateless7_ctc/export.py \
--exp-dir ./pruned_transducer_stateless7_ctc/exp \
--bpe-model data/lang_bpe_500/bpe.model \
--epoch 20 \
--avg 10 \
--jit 1
Usage of this script:
(1) ctc-decoding
./pruned_transducer_stateless7_ctc/jit_pretrained_ctc.py \
--nn-model-filename ./pruned_transducer_stateless7_ctc/exp/cpu_jit.pt \
--bpe-model data/lang_bpe_500/bpe.model \
--method ctc-decoding \
--sample-rate 16000 \
/path/to/foo.wav \
/path/to/bar.wav
(2) 1best
./pruned_transducer_stateless7_ctc/jit_pretrained_ctc.py \
--nn-model-filename ./pruned_transducer_stateless7_ctc/exp/cpu_jit.pt \
--HLG data/lang_bpe_500/HLG.pt \
--words-file data/lang_bpe_500/words.txt \
--method 1best \
--sample-rate 16000 \
/path/to/foo.wav \
/path/to/bar.wav
(3) nbest-rescoring
./pruned_transducer_stateless7_ctc/jit_pretrained_ctc.py \
--nn-model-filename ./pruned_transducer_stateless7_ctc/exp/cpu_jit.pt \
--HLG data/lang_bpe_500/HLG.pt \
--words-file data/lang_bpe_500/words.txt \
--G data/lm/G_4_gram.pt \
--method nbest-rescoring \
--sample-rate 16000 \
/path/to/foo.wav \
/path/to/bar.wav
(4) whole-lattice-rescoring
./pruned_transducer_stateless7_ctc/jit_pretrained_ctc.py \
--nn-model-filename ./pruned_transducer_stateless7_ctc/exp/cpu_jit.pt \
--HLG data/lang_bpe_500/HLG.pt \
--words-file data/lang_bpe_500/words.txt \
--G data/lm/G_4_gram.pt \
--method whole-lattice-rescoring \
--sample-rate 16000 \
/path/to/foo.wav \
/path/to/bar.wav
"""
import argparse
import logging
import math
from typing import List
import k2
import kaldifeat
import sentencepiece as spm
import torch
import torchaudio
from ctc_decode import get_decoding_params
from torch.nn.utils.rnn import pad_sequence
from train import get_params
from icefall.decode import (
get_lattice,
one_best_decoding,
rescore_with_n_best_list,
rescore_with_whole_lattice,
)
from icefall.utils import get_texts
def get_parser():
parser = argparse.ArgumentParser(
formatter_class=argparse.ArgumentDefaultsHelpFormatter
)
parser.add_argument(
"--model-filename",
type=str,
required=True,
help="Path to the torchscript model.",
)
parser.add_argument(
"--words-file",
type=str,
help="""Path to words.txt.
Used only when method is not ctc-decoding.
""",
)
parser.add_argument(
"--HLG",
type=str,
help="""Path to HLG.pt.
Used only when method is not ctc-decoding.
""",
)
parser.add_argument(
"--bpe-model",
type=str,
help="""Path to bpe.model.
Used only when method is ctc-decoding.
""",
)
parser.add_argument(
"--method",
type=str,
default="1best",
help="""Decoding method.
Possible values are:
(0) ctc-decoding - Use CTC decoding. It uses a sentence
piece model, i.e., lang_dir/bpe.model, to convert
word pieces to words. It needs neither a lexicon
nor an n-gram LM.
(1) 1best - Use the best path as decoding output. Only
the transformer encoder output is used for decoding.
We call it HLG decoding.
(2) nbest-rescoring. Extract n paths from the decoding lattice,
rescore them with an LM, the path with
the highest score is the decoding result.
We call it HLG decoding + n-gram LM rescoring.
(3) whole-lattice-rescoring - Use an LM to rescore the
decoding lattice and then use 1best to decode the
rescored lattice.
We call it HLG decoding + n-gram LM rescoring.
""",
)
parser.add_argument(
"--G",
type=str,
help="""An LM for rescoring.
Used only when method is
whole-lattice-rescoring or nbest-rescoring.
It's usually a 4-gram LM.
""",
)
parser.add_argument(
"--num-paths",
type=int,
default=100,
help="""
Used only when method is attention-decoder.
It specifies the size of n-best list.""",
)
parser.add_argument(
"--ngram-lm-scale",
type=float,
default=1.3,
help="""
Used only when method is whole-lattice-rescoring and nbest-rescoring.
It specifies the scale for n-gram LM scores.
(Note: You need to tune it on a dataset.)
""",
)
parser.add_argument(
"--nbest-scale",
type=float,
default=0.5,
help="""
Used only when method is nbest-rescoring.
It specifies the scale for lattice.scores when
extracting n-best lists. A smaller value results in
more unique number of paths with the risk of missing
the best path.
""",
)
parser.add_argument(
"--num-classes",
type=int,
default=500,
help="""
Vocab size in the BPE model.
""",
)
parser.add_argument(
"--sample-rate",
type=int,
default=16000,
help="The sample rate of the input sound file",
)
parser.add_argument(
"sound_files",
type=str,
nargs="+",
help="The input sound file(s) to transcribe. "
"Supported formats are those supported by torchaudio.load(). "
"For example, wav and flac are supported. "
"The sample rate has to be 16kHz.",
)
return parser
def read_sound_files(
filenames: List[str], expected_sample_rate: float = 16000
) -> List[torch.Tensor]:
"""Read a list of sound files into a list 1-D float32 torch tensors.
Args:
filenames:
A list of sound filenames.
expected_sample_rate:
The expected sample rate of the sound files.
Returns:
Return a list of 1-D float32 torch tensors.
"""
ans = []
for f in filenames:
wave, sample_rate = torchaudio.load(f)
assert (
sample_rate == expected_sample_rate
), f"Expected sample rate: {expected_sample_rate}. Given: {sample_rate}"
# We use only the first channel
ans.append(wave[0])
return ans
@torch.no_grad()
def main():
parser = get_parser()
args = parser.parse_args()
params = get_params()
# add decoding params
params.update(get_decoding_params())
params.update(vars(args))
logging.info(f"{params}")
device = torch.device("cpu")
if torch.cuda.is_available():
device = torch.device("cuda", 0)
logging.info(f"device: {device}")
model = torch.jit.load(args.model_filename)
model.to(device)
model.eval()
logging.info("Constructing Fbank computer")
opts = kaldifeat.FbankOptions()
opts.device = device
opts.frame_opts.dither = 0
opts.frame_opts.snip_edges = False
opts.frame_opts.samp_freq = params.sample_rate
opts.mel_opts.num_bins = params.feature_dim
fbank = kaldifeat.Fbank(opts)
logging.info(f"Reading sound files: {params.sound_files}")
waves = read_sound_files(
filenames=params.sound_files, expected_sample_rate=params.sample_rate
)
waves = [w.to(device) for w in waves]
logging.info("Decoding started")
features = fbank(waves)
feature_lengths = [f.size(0) for f in features]
features = pad_sequence(features, batch_first=True, padding_value=math.log(1e-10))
feature_lengths = torch.tensor(feature_lengths, device=device)
encoder_out, encoder_out_lens = model.encoder(
x=features,
x_lens=feature_lengths,
)
nnet_output = model.ctc_output(encoder_out)
batch_size = nnet_output.shape[0]
supervision_segments = torch.tensor(
[[i, 0, nnet_output.shape[1]] for i in range(batch_size)],
dtype=torch.int32,
)
if params.method == "ctc-decoding":
logging.info("Use CTC decoding")
bpe_model = spm.SentencePieceProcessor()
bpe_model.load(params.bpe_model)
max_token_id = params.num_classes - 1
H = k2.ctc_topo(
max_token=max_token_id,
modified=False,
device=device,
)
lattice = get_lattice(
nnet_output=nnet_output,
decoding_graph=H,
supervision_segments=supervision_segments,
search_beam=params.search_beam,
output_beam=params.output_beam,
min_active_states=params.min_active_states,
max_active_states=params.max_active_states,
subsampling_factor=params.subsampling_factor,
)
best_path = one_best_decoding(
lattice=lattice, use_double_scores=params.use_double_scores
)
token_ids = get_texts(best_path)
hyps = bpe_model.decode(token_ids)
hyps = [s.split() for s in hyps]
elif params.method in [
"1best",
"nbest-rescoring",
"whole-lattice-rescoring",
]:
logging.info(f"Loading HLG from {params.HLG}")
HLG = k2.Fsa.from_dict(torch.load(params.HLG, map_location="cpu"))
HLG = HLG.to(device)
if not hasattr(HLG, "lm_scores"):
# For whole-lattice-rescoring and attention-decoder
HLG.lm_scores = HLG.scores.clone()
if params.method in [
"nbest-rescoring",
"whole-lattice-rescoring",
]:
logging.info(f"Loading G from {params.G}")
G = k2.Fsa.from_dict(torch.load(params.G, map_location="cpu"))
G = G.to(device)
if params.method == "whole-lattice-rescoring":
# Add epsilon self-loops to G as we will compose
# it with the whole lattice later
G = k2.add_epsilon_self_loops(G)
G = k2.arc_sort(G)
# G.lm_scores is used to replace HLG.lm_scores during
# LM rescoring.
G.lm_scores = G.scores.clone()
lattice = get_lattice(
nnet_output=nnet_output,
decoding_graph=HLG,
supervision_segments=supervision_segments,
search_beam=params.search_beam,
output_beam=params.output_beam,
min_active_states=params.min_active_states,
max_active_states=params.max_active_states,
subsampling_factor=params.subsampling_factor,
)
if params.method == "1best":
logging.info("Use HLG decoding")
best_path = one_best_decoding(
lattice=lattice, use_double_scores=params.use_double_scores
)
if params.method == "nbest-rescoring":
logging.info("Use HLG decoding + LM rescoring")
best_path_dict = rescore_with_n_best_list(
lattice=lattice,
G=G,
num_paths=params.num_paths,
lm_scale_list=[params.ngram_lm_scale],
nbest_scale=params.nbest_scale,
)
best_path = next(iter(best_path_dict.values()))
elif params.method == "whole-lattice-rescoring":
logging.info("Use HLG decoding + LM rescoring")
best_path_dict = rescore_with_whole_lattice(
lattice=lattice,
G_with_epsilon_loops=G,
lm_scale_list=[params.ngram_lm_scale],
)
best_path = next(iter(best_path_dict.values()))
hyps = get_texts(best_path)
word_sym_table = k2.SymbolTable.from_file(params.words_file)
hyps = [[word_sym_table[i] for i in ids] for ids in hyps]
else:
raise ValueError(f"Unsupported decoding method: {params.method}")
s = "\n"
for filename, hyp in zip(params.sound_files, hyps):
words = " ".join(hyp)
s += f"{filename}:\n{words}\n\n"
logging.info(s)
logging.info("Decoding Done")
if __name__ == "__main__":
formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
logging.basicConfig(format=formatter, level=logging.INFO)
main()

1
egs/librispeech/ASR/pruned_transducer_stateless7_ctc/joiner.py Symbolic link
View File

@ -0,0 +1 @@
../pruned_transducer_stateless7/joiner.py

198
egs/librispeech/ASR/pruned_transducer_stateless7_ctc/model.py Normal file
View File

@ -0,0 +1,198 @@
# Copyright 2021 Xiaomi Corp. (authors: Fangjun Kuang, Wei Kang)
#
# 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.
from typing import Tuple
import k2
import torch
import torch.nn as nn
from encoder_interface import EncoderInterface
from icefall.utils import add_sos
class Transducer(nn.Module):
"""It implements https://arxiv.org/pdf/1211.3711.pdf
"Sequence Transduction with Recurrent Neural Networks"
"""
def __init__(
self,
encoder: EncoderInterface,
decoder: nn.Module,
joiner: nn.Module,
encoder_dim: int,
decoder_dim: int,
joiner_dim: int,
vocab_size: int,
):
"""
Args:
encoder:
It is the transcription network in the paper. Its accepts
two inputs: `x` of (N, T, encoder_dim) and `x_lens` of shape (N,).
It returns two tensors: `logits` of shape (N, T, encoder_dm) and
`logit_lens` of shape (N,).
decoder:
It is the prediction network in the paper. Its input shape
is (N, U) and its output shape is (N, U, decoder_dim).
It should contain one attribute: `blank_id`.
joiner:
It has two inputs with shapes: (N, T, encoder_dim) and (N, U, decoder_dim).
Its output shape is (N, T, U, vocab_size). Note that its output contains
unnormalized probs, i.e., not processed by log-softmax.
"""
super().__init__()
assert isinstance(encoder, EncoderInterface), type(encoder)
assert hasattr(decoder, "blank_id")
self.encoder = encoder
self.decoder = decoder
self.joiner = joiner
self.simple_am_proj = nn.Linear(
encoder_dim,
vocab_size,
)
self.simple_lm_proj = nn.Linear(decoder_dim, vocab_size)
self.ctc_output = nn.Sequential(
nn.Dropout(p=0.1),
nn.Linear(encoder_dim, vocab_size),
nn.LogSoftmax(dim=-1),
)
def forward(
self,
x: torch.Tensor,
x_lens: torch.Tensor,
y: k2.RaggedTensor,
prune_range: int = 5,
am_scale: float = 0.0,
lm_scale: float = 0.0,
) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
"""
Args:
x:
A 3-D tensor of shape (N, T, C).
x_lens:
A 1-D tensor of shape (N,). It contains the number of frames in `x`
before padding.
y:
A ragged tensor with 2 axes [utt][label]. It contains labels of each
utterance.
prune_range:
The prune range for rnnt loss, it means how many symbols(context)
we are considering for each frame to compute the loss.
am_scale:
The scale to smooth the loss with am (output of encoder network)
part
lm_scale:
The scale to smooth the loss with lm (output of predictor network)
part
Returns:
Return a tuple containing simple loss, pruned loss, and ctc-output.
Note:
Regarding am_scale & lm_scale, it will make the loss-function one of
the form:
lm_scale * lm_probs + am_scale * am_probs +
(1-lm_scale-am_scale) * combined_probs
"""
assert x.ndim == 3, x.shape
assert x_lens.ndim == 1, x_lens.shape
assert y.num_axes == 2, y.num_axes
assert x.size(0) == x_lens.size(0) == y.dim0
encoder_out, x_lens = self.encoder(x, x_lens)
assert torch.all(x_lens > 0)
# compute ctc log-probs
ctc_output = self.ctc_output(encoder_out)
# Now for the decoder, i.e., the prediction network
row_splits = y.shape.row_splits(1)
y_lens = row_splits[1:] - row_splits[:-1]
blank_id = self.decoder.blank_id
sos_y = add_sos(y, sos_id=blank_id)
# sos_y_padded: [B, S + 1], start with SOS.
sos_y_padded = sos_y.pad(mode="constant", padding_value=blank_id)
# decoder_out: [B, S + 1, decoder_dim]
decoder_out = self.decoder(sos_y_padded)
# Note: y does not start with SOS
# y_padded : [B, S]
y_padded = y.pad(mode="constant", padding_value=0)
y_padded = y_padded.to(torch.int64)
boundary = torch.zeros((x.size(0), 4), dtype=torch.int64, device=x.device)
boundary[:, 2] = y_lens
boundary[:, 3] = x_lens
lm = self.simple_lm_proj(decoder_out)
am = self.simple_am_proj(encoder_out)
with torch.cuda.amp.autocast(enabled=False):
simple_loss, (px_grad, py_grad) = k2.rnnt_loss_smoothed(
lm=lm.float(),
am=am.float(),
symbols=y_padded,
termination_symbol=blank_id,
lm_only_scale=lm_scale,
am_only_scale=am_scale,
boundary=boundary,
reduction="sum",
return_grad=True,
)
# ranges : [B, T, prune_range]
ranges = k2.get_rnnt_prune_ranges(
px_grad=px_grad,
py_grad=py_grad,
boundary=boundary,
s_range=prune_range,
)
# am_pruned : [B, T, prune_range, encoder_dim]
# lm_pruned : [B, T, prune_range, decoder_dim]
am_pruned, lm_pruned = k2.do_rnnt_pruning(
am=self.joiner.encoder_proj(encoder_out),
lm=self.joiner.decoder_proj(decoder_out),
ranges=ranges,
)
# logits : [B, T, prune_range, vocab_size]
# project_input=False since we applied the decoder's input projections
# prior to do_rnnt_pruning (this is an optimization for speed).
logits = self.joiner(am_pruned, lm_pruned, project_input=False)
with torch.cuda.amp.autocast(enabled=False):
pruned_loss = k2.rnnt_loss_pruned(
logits=logits.float(),
symbols=y_padded,
ranges=ranges,
termination_symbol=blank_id,
boundary=boundary,
reduction="sum",
)
return (simple_loss, pruned_loss, ctc_output)

1
egs/librispeech/ASR/pruned_transducer_stateless7_ctc/optim.py Symbolic link
View File

@ -0,0 +1 @@
../pruned_transducer_stateless7/optim.py

353
egs/librispeech/ASR/pruned_transducer_stateless7_ctc/pretrained.py Executable file
View File

@ -0,0 +1,353 @@
#!/usr/bin/env python3
# Copyright 2021 Xiaomi Corp. (authors: Fangjun Kuang)
#
# 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.
"""
This script loads a checkpoint and uses it to decode waves.
You can generate the checkpoint with the following command:
./pruned_transducer_stateless7_ctc/export.py \
--exp-dir ./pruned_transducer_stateless7_ctc/exp \
--bpe-model data/lang_bpe_500/bpe.model \
--epoch 20 \
--avg 10
Usage of this script:
(1) greedy search
./pruned_transducer_stateless7_ctc/pretrained.py \
--checkpoint ./pruned_transducer_stateless7_ctc/exp/pretrained.pt \
--bpe-model ./data/lang_bpe_500/bpe.model \
--method greedy_search \
/path/to/foo.wav \
/path/to/bar.wav
(2) beam search
./pruned_transducer_stateless7_ctc/pretrained.py \
--checkpoint ./pruned_transducer_stateless7_ctc/exp/pretrained.pt \
--bpe-model ./data/lang_bpe_500/bpe.model \
--method beam_search \
--beam-size 4 \
/path/to/foo.wav \
/path/to/bar.wav
(3) modified beam search
./pruned_transducer_stateless7_ctc/pretrained.py \
--checkpoint ./pruned_transducer_stateless7_ctc/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
(4) fast beam search
./pruned_transducer_stateless7_ctc/pretrained.py \
--checkpoint ./pruned_transducer_stateless7_ctc/exp/pretrained.pt \
--bpe-model ./data/lang_bpe_500/bpe.model \
--method fast_beam_search \
--beam-size 4 \
/path/to/foo.wav \
/path/to/bar.wav
You can also use `./pruned_transducer_stateless7_ctc/exp/epoch-xx.pt`.
Note: ./pruned_transducer_stateless7_ctc/exp/pretrained.pt is generated by
./pruned_transducer_stateless7_ctc/export.py
"""
import argparse
import logging
import math
from typing import List
import k2
import kaldifeat
import sentencepiece as spm
import torch
import torchaudio
from beam_search import (
beam_search,
fast_beam_search_one_best,
greedy_search,
greedy_search_batch,
modified_beam_search,
)
from torch.nn.utils.rnn import pad_sequence
from train import add_model_arguments, get_params, get_transducer_model
def get_parser():
parser = argparse.ArgumentParser(
formatter_class=argparse.ArgumentDefaultsHelpFormatter
)
parser.add_argument(
"--checkpoint",
type=str,
required=True,
help="Path to the checkpoint. "
"The checkpoint is assumed to be saved by "
"icefall.checkpoint.save_checkpoint().",
)
parser.add_argument(
"--bpe-model",
type=str,
help="""Path to bpe.model.""",
)
parser.add_argument(
"--method",
type=str,
default="greedy_search",
help="""Possible values are:
- greedy_search
- beam_search
- modified_beam_search
- fast_beam_search
""",
)
parser.add_argument(
"sound_files",
type=str,
nargs="+",
help="The input sound file(s) to transcribe. "
"Supported formats are those supported by torchaudio.load(). "
"For example, wav and flac are supported. "
"The sample rate has to be 16kHz.",
)
parser.add_argument(
"--sample-rate",
type=int,
default=16000,
help="The sample rate of the input sound file",
)
parser.add_argument(
"--beam-size",
type=int,
default=4,
help="""An integer indicating how many candidates we will keep for each
frame. Used only when --method is beam_search or
modified_beam_search.""",
)
parser.add_argument(
"--beam",
type=float,
default=4,
help="""A floating point value to calculate the cutoff score during beam
search (i.e., `cutoff = max-score - beam`), which is the same as the
`beam` in Kaldi.
Used only when --method is fast_beam_search""",
)
parser.add_argument(
"--max-contexts",
type=int,
default=4,
help="""Used only when --method is fast_beam_search""",
)
parser.add_argument(
"--max-states",
type=int,
default=8,
help="""Used only when --method is fast_beam_search""",
)
parser.add_argument(
"--context-size",
type=int,
default=2,
help="The context size in the decoder. 1 means bigram; 2 means tri-gram",
)
parser.add_argument(
"--max-sym-per-frame",
type=int,
default=1,
help="""Maximum number of symbols per frame. Used only when
--method is greedy_search.
""",
)
add_model_arguments(parser)
return parser
def read_sound_files(
filenames: List[str], expected_sample_rate: float
) -> List[torch.Tensor]:
"""Read a list of sound files into a list 1-D float32 torch tensors.
Args:
filenames:
A list of sound filenames.
expected_sample_rate:
The expected sample rate of the sound files.
Returns:
Return a list of 1-D float32 torch tensors.
"""
ans = []
for f in filenames:
wave, sample_rate = torchaudio.load(f)
assert (
sample_rate == expected_sample_rate
), f"Expected sample rate: {expected_sample_rate}. Given: {sample_rate}"
# We use only the first channel
ans.append(wave[0])
return ans
@torch.no_grad()
def main():
parser = get_parser()
args = parser.parse_args()
params = get_params()
params.update(vars(args))
sp = spm.SentencePieceProcessor()
sp.load(params.bpe_model)
# <blk> is defined in local/train_bpe_model.py
params.blank_id = sp.piece_to_id("<blk>")
params.unk_id = sp.piece_to_id("<unk>")
params.vocab_size = sp.get_piece_size()
logging.info(f"{params}")
device = torch.device("cpu")
if torch.cuda.is_available():
device = torch.device("cuda", 0)
logging.info(f"device: {device}")
logging.info("Creating model")
model = get_transducer_model(params)
num_param = sum([p.numel() for p in model.parameters()])
logging.info(f"Number of model parameters: {num_param}")
checkpoint = torch.load(args.checkpoint, map_location="cpu")
model.load_state_dict(checkpoint["model"], strict=False)
model.to(device)
model.eval()
model.device = device
logging.info("Constructing Fbank computer")
opts = kaldifeat.FbankOptions()
opts.device = device
opts.frame_opts.dither = 0
opts.frame_opts.snip_edges = False
opts.frame_opts.samp_freq = params.sample_rate
opts.mel_opts.num_bins = params.feature_dim
fbank = kaldifeat.Fbank(opts)
logging.info(f"Reading sound files: {params.sound_files}")
waves = read_sound_files(
filenames=params.sound_files, expected_sample_rate=params.sample_rate
)
waves = [w.to(device) for w in waves]
logging.info("Decoding started")
features = fbank(waves)
feature_lengths = [f.size(0) for f in features]
features = pad_sequence(features, batch_first=True, padding_value=math.log(1e-10))
feature_lengths = torch.tensor(feature_lengths, device=device)
encoder_out, encoder_out_lens = model.encoder(x=features, x_lens=feature_lengths)
num_waves = encoder_out.size(0)
hyps = []
msg = f"Using {params.method}"
if params.method == "beam_search":
msg += f" with beam size {params.beam_size}"
logging.info(msg)
if params.method == "fast_beam_search":
decoding_graph = k2.trivial_graph(params.vocab_size - 1, device=device)
hyp_tokens = fast_beam_search_one_best(
model=model,
decoding_graph=decoding_graph,
encoder_out=encoder_out,
encoder_out_lens=encoder_out_lens,
beam=params.beam,
max_contexts=params.max_contexts,
max_states=params.max_states,
)
for hyp in sp.decode(hyp_tokens):
hyps.append(hyp.split())
elif params.method == "modified_beam_search":
hyp_tokens = modified_beam_search(
model=model,
encoder_out=encoder_out,
encoder_out_lens=encoder_out_lens,
beam=params.beam_size,
)
for hyp in sp.decode(hyp_tokens):
hyps.append(hyp.split())
elif params.method == "greedy_search" and params.max_sym_per_frame == 1:
hyp_tokens = greedy_search_batch(
model=model,
encoder_out=encoder_out,
encoder_out_lens=encoder_out_lens,
)
for hyp in sp.decode(hyp_tokens):
hyps.append(hyp.split())
else:
for i in range(num_waves):
# fmt: off
encoder_out_i = encoder_out[i:i+1, :encoder_out_lens[i]]
# fmt: on
if params.method == "greedy_search":
hyp = greedy_search(
model=model,
encoder_out=encoder_out_i,
max_sym_per_frame=params.max_sym_per_frame,
)
elif params.method == "beam_search":
hyp = beam_search(
model=model,
encoder_out=encoder_out_i,
beam=params.beam_size,
)
else:
raise ValueError(f"Unsupported method: {params.method}")
hyps.append(sp.decode(hyp).split())
s = "\n"
for filename, hyp in zip(params.sound_files, hyps):
words = " ".join(hyp)
s += f"{filename}:\n{words}\n\n"
logging.info(s)
logging.info("Decoding Done")
if __name__ == "__main__":
formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
logging.basicConfig(format=formatter, level=logging.INFO)
main()

441
egs/librispeech/ASR/pruned_transducer_stateless7_ctc/pretrained_ctc.py Executable file
View File

@ -0,0 +1,441 @@
#!/usr/bin/env python3
# Copyright 2022 Xiaomi Corp. (authors: Fangjun Kuang,
# 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.
"""
This script loads torchscript models, exported by `torch.jit.script()`
and uses them to decode waves.
You can use the following command to get the exported models:
./pruned_transducer_stateless7_ctc/export.py \
--exp-dir ./pruned_transducer_stateless7_ctc/exp \
--bpe-model data/lang_bpe_500/bpe.model \
--epoch 20 \
--avg 10
Usage of this script:
(1) ctc-decoding
./pruned_transducer_stateless7_ctc/jit_pretrained_ctc.py \
--checkpoint ./pruned_transducer_stateless7_ctc/exp/pretrained.pt \
--bpe-model data/lang_bpe_500/bpe.model \
--method ctc-decoding \
--sample-rate 16000 \
/path/to/foo.wav \
/path/to/bar.wav
(2) 1best
./pruned_transducer_stateless7_ctc/jit_pretrained_ctc.py \
--checkpoint ./pruned_transducer_stateless7_ctc/exp/pretrained.pt \
--HLG data/lang_bpe_500/HLG.pt \
--words-file data/lang_bpe_500/words.txt \
--method 1best \
--sample-rate 16000 \
/path/to/foo.wav \
/path/to/bar.wav
(3) nbest-rescoring
./bruned_transducer_stateless7_ctc/jit_pretrained_ctc.py \
--checkpoint ./pruned_transducer_stateless7_ctc/exp/pretrained.pt \
--HLG data/lang_bpe_500/HLG.pt \
--words-file data/lang_bpe_500/words.txt \
--G data/lm/G_4_gram.pt \
--method nbest-rescoring \
--sample-rate 16000 \
/path/to/foo.wav \
/path/to/bar.wav
(4) whole-lattice-rescoring
./pruned_transducer_stateless7_ctc/jit_pretrained_ctc.py \
--checkpoint ./pruned_transducer_stateless7_ctc/exp/pretrained.pt \
--HLG data/lang_bpe_500/HLG.pt \
--words-file data/lang_bpe_500/words.txt \
--G data/lm/G_4_gram.pt \
--method whole-lattice-rescoring \
--sample-rate 16000 \
/path/to/foo.wav \
/path/to/bar.wav
"""
import argparse
import logging
import math
from typing import List
import k2
import kaldifeat
import sentencepiece as spm
import torch
import torchaudio
from ctc_decode import get_decoding_params
from torch.nn.utils.rnn import pad_sequence
from train import add_model_arguments, get_params, get_transducer_model
from icefall.decode import (
get_lattice,
one_best_decoding,
rescore_with_n_best_list,
rescore_with_whole_lattice,
)
from icefall.utils import get_texts
def get_parser():
parser = argparse.ArgumentParser(
formatter_class=argparse.ArgumentDefaultsHelpFormatter
)
parser.add_argument(
"--checkpoint",
type=str,
required=True,
help="Path to the checkpoint. "
"The checkpoint is assumed to be saved by "
"icefall.checkpoint.save_checkpoint().",
)
parser.add_argument(
"--context-size",
type=int,
default=2,
help="The context size in the decoder. 1 means bigram; " "2 means tri-gram",
)
parser.add_argument(
"--words-file",
type=str,
help="""Path to words.txt.
Used only when method is not ctc-decoding.
""",
)
parser.add_argument(
"--HLG",
type=str,
help="""Path to HLG.pt.
Used only when method is not ctc-decoding.
""",
)
parser.add_argument(
"--bpe-model",
type=str,
help="""Path to bpe.model.
Used only when method is ctc-decoding.
""",
)
parser.add_argument(
"--method",
type=str,
default="1best",
help="""Decoding method.
Possible values are:
(0) ctc-decoding - Use CTC decoding. It uses a sentence
piece model, i.e., lang_dir/bpe.model, to convert
word pieces to words. It needs neither a lexicon
nor an n-gram LM.
(1) 1best - Use the best path as decoding output. Only
the transformer encoder output is used for decoding.
We call it HLG decoding.
(2) nbest-rescoring. Extract n paths from the decoding lattice,
rescore them with an LM, the path with
the highest score is the decoding result.
We call it HLG decoding + n-gram LM rescoring.
(3) whole-lattice-rescoring - Use an LM to rescore the
decoding lattice and then use 1best to decode the
rescored lattice.
We call it HLG decoding + n-gram LM rescoring.
""",
)
parser.add_argument(
"--G",
type=str,
help="""An LM for rescoring.
Used only when method is
whole-lattice-rescoring or nbest-rescoring.
It's usually a 4-gram LM.
""",
)
parser.add_argument(
"--num-paths",
type=int,
default=100,
help="""
Used only when method is attention-decoder.
It specifies the size of n-best list.""",
)
parser.add_argument(
"--ngram-lm-scale",
type=float,
default=1.3,
help="""
Used only when method is whole-lattice-rescoring and nbest-rescoring.
It specifies the scale for n-gram LM scores.
(Note: You need to tune it on a dataset.)
""",
)
parser.add_argument(
"--nbest-scale",
type=float,
default=0.5,
help="""
Used only when method is nbest-rescoring.
It specifies the scale for lattice.scores when
extracting n-best lists. A smaller value results in
more unique number of paths with the risk of missing
the best path.
""",
)
parser.add_argument(
"--num-classes",
type=int,
default=500,
help="""
Vocab size in the BPE model.
""",
)
parser.add_argument(
"--sample-rate",
type=int,
default=16000,
help="The sample rate of the input sound file",
)
parser.add_argument(
"sound_files",
type=str,
nargs="+",
help="The input sound file(s) to transcribe. "
"Supported formats are those supported by torchaudio.load(). "
"For example, wav and flac are supported. "
"The sample rate has to be 16kHz.",
)
add_model_arguments(parser)
return parser
def read_sound_files(
filenames: List[str], expected_sample_rate: float = 16000
) -> List[torch.Tensor]:
"""Read a list of sound files into a list 1-D float32 torch tensors.
Args:
filenames:
A list of sound filenames.
expected_sample_rate:
The expected sample rate of the sound files.
Returns:
Return a list of 1-D float32 torch tensors.
"""
ans = []
for f in filenames:
wave, sample_rate = torchaudio.load(f)
assert sample_rate == expected_sample_rate, (
f"expected sample rate: {expected_sample_rate}. " f"Given: {sample_rate}"
)
# We use only the first channel
ans.append(wave[0])
return ans
@torch.no_grad()
def main():
parser = get_parser()
args = parser.parse_args()
params = get_params()
# add decoding params
params.update(get_decoding_params())
params.update(vars(args))
params.vocab_size = params.num_classes
params.blank_id = 0
logging.info(f"{params}")
device = torch.device("cpu")
if torch.cuda.is_available():
device = torch.device("cuda", 0)
logging.info(f"device: {device}")
logging.info("Creating model")
model = get_transducer_model(params)
num_param = sum([p.numel() for p in model.parameters()])
logging.info(f"Number of model parameters: {num_param}")
checkpoint = torch.load(args.checkpoint, map_location="cpu")
model.load_state_dict(checkpoint["model"], strict=False)
model.to(device)
model.eval()
logging.info("Constructing Fbank computer")
opts = kaldifeat.FbankOptions()
opts.device = device
opts.frame_opts.dither = 0
opts.frame_opts.snip_edges = False
opts.frame_opts.samp_freq = params.sample_rate
opts.mel_opts.num_bins = params.feature_dim
fbank = kaldifeat.Fbank(opts)
logging.info(f"Reading sound files: {params.sound_files}")
waves = read_sound_files(
filenames=params.sound_files, expected_sample_rate=params.sample_rate
)
waves = [w.to(device) for w in waves]
logging.info("Decoding started")
features = fbank(waves)
feature_lengths = [f.size(0) for f in features]
features = pad_sequence(features, batch_first=True, padding_value=math.log(1e-10))
feature_lengths = torch.tensor(feature_lengths, device=device)
encoder_out, encoder_out_lens = model.encoder(
x=features,
x_lens=feature_lengths,
)
nnet_output = model.ctc_output(encoder_out)
batch_size = nnet_output.shape[0]
supervision_segments = torch.tensor(
[[i, 0, nnet_output.shape[1]] for i in range(batch_size)],
dtype=torch.int32,
)
if params.method == "ctc-decoding":
logging.info("Use CTC decoding")
bpe_model = spm.SentencePieceProcessor()
bpe_model.load(params.bpe_model)
max_token_id = params.num_classes - 1
H = k2.ctc_topo(
max_token=max_token_id,
modified=False,
device=device,
)
lattice = get_lattice(
nnet_output=nnet_output,
decoding_graph=H,
supervision_segments=supervision_segments,
search_beam=params.search_beam,
output_beam=params.output_beam,
min_active_states=params.min_active_states,
max_active_states=params.max_active_states,
subsampling_factor=params.subsampling_factor,
)
best_path = one_best_decoding(
lattice=lattice, use_double_scores=params.use_double_scores
)
token_ids = get_texts(best_path)
hyps = bpe_model.decode(token_ids)
hyps = [s.split() for s in hyps]
elif params.method in [
"1best",
"nbest-rescoring",
"whole-lattice-rescoring",
]:
logging.info(f"Loading HLG from {params.HLG}")
HLG = k2.Fsa.from_dict(torch.load(params.HLG, map_location="cpu"))
HLG = HLG.to(device)
if not hasattr(HLG, "lm_scores"):
# For whole-lattice-rescoring and attention-decoder
HLG.lm_scores = HLG.scores.clone()
if params.method in [
"nbest-rescoring",
"whole-lattice-rescoring",
]:
logging.info(f"Loading G from {params.G}")
G = k2.Fsa.from_dict(torch.load(params.G, map_location="cpu"))
G = G.to(device)
if params.method == "whole-lattice-rescoring":
# Add epsilon self-loops to G as we will compose
# it with the whole lattice later
G = k2.add_epsilon_self_loops(G)
G = k2.arc_sort(G)
# G.lm_scores is used to replace HLG.lm_scores during
# LM rescoring.
G.lm_scores = G.scores.clone()
lattice = get_lattice(
nnet_output=nnet_output,
decoding_graph=HLG,
supervision_segments=supervision_segments,
search_beam=params.search_beam,
output_beam=params.output_beam,
min_active_states=params.min_active_states,
max_active_states=params.max_active_states,
subsampling_factor=params.subsampling_factor,
)
if params.method == "1best":
logging.info("Use HLG decoding")
best_path = one_best_decoding(
lattice=lattice, use_double_scores=params.use_double_scores
)
if params.method == "nbest-rescoring":
logging.info("Use HLG decoding + LM rescoring")
best_path_dict = rescore_with_n_best_list(
lattice=lattice,
G=G,
num_paths=params.num_paths,
lm_scale_list=[params.ngram_lm_scale],
nbest_scale=params.nbest_scale,
)
best_path = next(iter(best_path_dict.values()))
elif params.method == "whole-lattice-rescoring":
logging.info("Use HLG decoding + LM rescoring")
best_path_dict = rescore_with_whole_lattice(
lattice=lattice,
G_with_epsilon_loops=G,
lm_scale_list=[params.ngram_lm_scale],
)
best_path = next(iter(best_path_dict.values()))
hyps = get_texts(best_path)
word_sym_table = k2.SymbolTable.from_file(params.words_file)
hyps = [[word_sym_table[i] for i in ids] for ids in hyps]
else:
raise ValueError(f"Unsupported decoding method: {params.method}")
s = "\n"
for filename, hyp in zip(params.sound_files, hyps):
words = " ".join(hyp)
s += f"{filename}:\n{words}\n\n"
logging.info(s)
logging.info("Decoding Done")
if __name__ == "__main__":
formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s"
logging.basicConfig(format=formatter, level=logging.INFO)
main()

1
egs/librispeech/ASR/pruned_transducer_stateless7_ctc/scaling.py Symbolic link
View File

@ -0,0 +1 @@
../pruned_transducer_stateless7/scaling.py

1
egs/librispeech/ASR/pruned_transducer_stateless7_ctc/scaling_converter.py Symbolic link
View File

@ -0,0 +1 @@
../pruned_transducer_stateless7/scaling_converter.py

56
egs/librispeech/ASR/pruned_transducer_stateless7_ctc/test_model.py Executable file
View File

@ -0,0 +1,56 @@
#!/usr/bin/env python3
# Copyright 2022 Xiaomi Corp. (authors: Fangjun Kuang)
#
# 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.
"""
To run this file, do:
cd icefall/egs/librispeech/ASR
python ./pruned_transducer_stateless7_ctc/test_model.py
"""
from train import get_params, get_transducer_model
def test_model_1():
params = get_params()
params.vocab_size = 500
params.blank_id = 0
params.context_size = 2
params.num_encoder_layers = "2,4,3,2,4"
# params.feedforward_dims = "1024,1024,1536,1536,1024"
params.feedforward_dims = "1024,1024,2048,2048,1024"
params.nhead = "8,8,8,8,8"
params.encoder_dims = "384,384,384,384,384"
params.attention_dims = "192,192,192,192,192"
params.encoder_unmasked_dims = "256,256,256,256,256"
params.zipformer_downsampling_factors = "1,2,4,8,2"
params.cnn_module_kernels = "31,31,31,31,31"
params.decoder_dim = 512
params.joiner_dim = 512
model = get_transducer_model(params)
num_param = sum([p.numel() for p in model.parameters()])
print(f"Number of model parameters: {num_param}")
def main():
test_model_1()
if __name__ == "__main__":
main()

1252
egs/librispeech/ASR/pruned_transducer_stateless7_ctc/train.py Executable file
View File

File diff suppressed because it is too large Load Diff

1
egs/librispeech/ASR/pruned_transducer_stateless7_ctc/zipformer.py Symbolic link
View File

@ -0,0 +1 @@
../pruned_transducer_stateless7/zipformer.py

18
icefall/utils.py
View File

@ -175,11 +175,13 @@ class AttributeDict(dict):
def encode_supervisions(
supervisions: dict, subsampling_factor: int
) -> Tuple[torch.Tensor, List[str]]:
supervisions: dict,
subsampling_factor: int,
token_ids: Optional[List[List[int]]] = None,
) -> Tuple[torch.Tensor, Union[List[str], List[List[int]]]]:
"""
Encodes Lhotse's ``batch["supervisions"]`` dict into
a pair of torch Tensor, and a list of transcription strings.
a pair of torch Tensor, and a list of transcription strings or token indexes
The supervision tensor has shape ``(batch_size, 3)``.
Its second dimension contains information about sequence index [0],
@ -208,10 +210,14 @@ def encode_supervisions(
indices = torch.argsort(supervision_segments[:, 2], descending=True)
supervision_segments = supervision_segments[indices]
texts = supervisions["text"]
texts = [texts[idx] for idx in indices]
return supervision_segments, texts
if token_ids is None:
texts = supervisions["text"]
res = [texts[idx] for idx in indices]
else:
res = [token_ids[idx] for idx in indices]
return supervision_segments, res
def get_texts(