mirrors/icefall
1
0
Fork 0
mirror of https://github.com/k2-fsa/icefall.git synced 2025-09-04 14:44:18 +00:00
Code Issues Packages Projects Releases Wiki Activity

Merge remote-tracking branch 'upstream/master' into gigaspeech_rnnt

Browse Source
This commit is contained in:
Guanbo Wang 2022-05-12 00:59:45 -04:00
commit cfebfd21f6
40 changed files with 2458 additions and 156 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

1
.flake8
View File

@ -10,6 +10,7 @@ per-file-ignores =
egs/gigaspeech/ASR/*/conformer.py: E501,
egs/librispeech/ASR/pruned_transducer_stateless2/*.py: E501,
egs/gigaspeech/ASR/pruned_transducer_stateless2/*.py: E501,
egs/librispeech/ASR/pruned_transducer_stateless4/*.py: E501,
egs/librispeech/ASR/*/optim.py: E501,
egs/librispeech/ASR/*/scaling.py: E501,

17
.github/scripts/compute-fbank-librispeech-test-clean-and-test-other.sh vendored Executable file
View File

@ -0,0 +1,17 @@
#!/usr/bin/env bash
# This script computes fbank features for the test-clean and test-other datasets.
# The computed features are saved to ~/tmp/fbank-libri and are
# cached for later runs
export PYTHONPATH=$PWD:$PYTHONPATH
echo $PYTHONPATH
mkdir ~/tmp/fbank-libri
cd egs/librispeech/ASR
mkdir -p data
cd data
[ ! -e fbank ] && ln -s ~/tmp/fbank-libri fbank
cd ..
./local/compute_fbank_librispeech.py
ls -lh data/fbank/

23
.github/scripts/download-librispeech-test-clean-and-test-other-dataset.sh vendored Executable file
View File

@ -0,0 +1,23 @@
#!/usr/bin/env bash
# This script downloads the test-clean and test-other datasets
# of LibriSpeech and unzip them to the folder ~/tmp/download,
# which is cached by GitHub actions for later runs.
#
# You will find directories ~/tmp/download/LibriSpeech after running
# this script.
mkdir ~/tmp/download
cd egs/librispeech/ASR
ln -s ~/tmp/download .
cd download
wget -q --no-check-certificate https://www.openslr.org/resources/12/test-clean.tar.gz
tar xf test-clean.tar.gz
rm test-clean.tar.gz
wget -q --no-check-certificate https://www.openslr.org/resources/12/test-other.tar.gz
tar xf test-other.tar.gz
rm test-other.tar.gz
pwd
ls -lh
ls -lh LibriSpeech

13
.github/scripts/install-kaldifeat.sh vendored Executable file
View File

@ -0,0 +1,13 @@
#!/usr/bin/env bash
# This script installs kaldifeat into the directory ~/tmp/kaldifeat
# which is cached by GitHub actions for later runs.
mkdir -p ~/tmp
cd ~/tmp
git clone https://github.com/csukuangfj/kaldifeat
cd kaldifeat
mkdir build
cd build
cmake -DCMAKE_BUILD_TYPE=Release ..
make -j2 _kaldifeat

11
.github/scripts/prepare-librispeech-test-clean-and-test-other-manifests.sh vendored Executable file
View File

@ -0,0 +1,11 @@
#!/usr/bin/env bash
# This script assumes that test-clean and test-other are downloaded
# to egs/librispeech/ASR/download/LibriSpeech and generates manifest
# files in egs/librispeech/ASR/data/manifests
cd egs/librispeech/ASR
[ ! -e download ] && ln -s ~/tmp/download .
mkdir -p data/manifests
lhotse prepare librispeech -j 2 -p test-clean -p test-other ./download/LibriSpeech data/manifests
ls -lh data/manifests

28
.github/scripts/run-librispeech-pruned-transducer-stateless-2022-03-12.sh vendored
View File

@ -45,3 +45,31 @@ for method in modified_beam_search beam_search; do
$repo/test_wavs/1221-135766-0001.wav \
$repo/test_wavs/1221-135766-0002.wav
done
echo "GITHUB_EVENT_NAME: ${GITHUB_EVENT_NAME}"
if [[ x"${GITHUB_EVENT_NAME}" == x"schedule" ]]; then
mkdir -p pruned_transducer_stateless/exp
ln -s $PWD/$repo/exp/pretrained.pt pruned_transducer_stateless/exp/epoch-999.pt
ln -s $PWD/$repo/data/lang_bpe_500 data/
ls -lh data
ls -lh pruned_transducer_stateless/exp
log "Decoding test-clean and test-other"
# use a small value for decoding with CPU
max_duration=50
for method in greedy_search fast_beam_search; do
log "Decoding with $method"
./pruned_transducer_stateless/decode.py \
--decoding-method $method \
--epoch 999 \
--avg 1 \
--max-duration $max_duration \
--exp-dir pruned_transducer_stateless/exp
done
rm pruned_transducer_stateless/exp/*.pt
fi

28
.github/scripts/run-librispeech-pruned-transducer-stateless2-2022-04-29.sh vendored
View File

@ -49,3 +49,31 @@ for method in modified_beam_search beam_search fast_beam_search; do
$repo/test_wavs/1221-135766-0001.wav \
$repo/test_wavs/1221-135766-0002.wav
done
echo "GITHUB_EVENT_NAME: ${GITHUB_EVENT_NAME}"
if [[ x"${GITHUB_EVENT_NAME}" == x"schedule" ]]; then
mkdir -p pruned_transducer_stateless2/exp
ln -s $PWD/$repo/exp/pretrained.pt pruned_transducer_stateless2/exp/epoch-999.pt
ln -s $PWD/$repo/data/lang_bpe_500 data/
ls -lh data
ls -lh pruned_transducer_stateless2/exp
log "Decoding test-clean and test-other"
# use a small value for decoding with CPU
max_duration=50
for method in greedy_search fast_beam_search; do
log "Decoding with $method"
./pruned_transducer_stateless2/decode.py \
--decoding-method $method \
--epoch 999 \
--avg 1 \
--max-duration $max_duration \
--exp-dir pruned_transducer_stateless2/exp
done
rm pruned_transducer_stateless2/exp/*.pt
fi

28
.github/scripts/run-librispeech-pruned-transducer-stateless3-2022-04-29.sh vendored
View File

@ -49,3 +49,31 @@ for method in modified_beam_search beam_search fast_beam_search; do
$repo/test_wavs/1221-135766-0001.wav \
$repo/test_wavs/1221-135766-0002.wav
done
echo "GITHUB_EVENT_NAME: ${GITHUB_EVENT_NAME}"
if [[ x"${GITHUB_EVENT_NAME}" == x"schedule" ]]; then
mkdir -p pruned_transducer_stateless3/exp
ln -s $PWD/$repo/exp/pretrained.pt pruned_transducer_stateless3/exp/epoch-999.pt
ln -s $PWD/$repo/data/lang_bpe_500 data/
ls -lh data
ls -lh pruned_transducer_stateless3/exp
log "Decoding test-clean and test-other"
# use a small value for decoding with CPU
max_duration=50
for method in greedy_search fast_beam_search; do
log "Decoding with $method"
./pruned_transducer_stateless3/decode.py \
--decoding-method $method \
--epoch 999 \
--avg 1 \
--max-duration $max_duration \
--exp-dir pruned_transducer_stateless3/exp
done
rm pruned_transducer_stateless3/exp/*.pt
fi

28
.github/scripts/run-librispeech-transducer-stateless2-2022-04-19.sh vendored
View File

@ -45,3 +45,31 @@ for method in modified_beam_search beam_search; do
$repo/test_wavs/1221-135766-0001.wav \
$repo/test_wavs/1221-135766-0002.wav
done
echo "GITHUB_EVENT_NAME: ${GITHUB_EVENT_NAME}"
if [[ x"${GITHUB_EVENT_NAME}" == x"schedule" ]]; then
mkdir -p transducer_stateless2/exp
ln -s $PWD/$repo/exp/pretrained.pt transducer_stateless2/exp/epoch-999.pt
ln -s $PWD/$repo/data/lang_bpe_500 data/
ls -lh data
ls -lh transducer_stateless2/exp
log "Decoding test-clean and test-other"
# use a small value for decoding with CPU
max_duration=50
for method in greedy_search modified_beam_search; do
log "Decoding with $method"
./transducer_stateless2/decode.py \
--decoding-method $method \
--epoch 999 \
--avg 1 \
--max-duration $max_duration \
--exp-dir transducer_stateless2/exp
done
rm transducer_stateless2/exp/*.pt
fi

85
.github/workflows/run-librispeech-2022-03-12.yml vendored
View File

@ -24,9 +24,18 @@ on:
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_03_12:
if: github.event.label.name == 'ready' || github.event_name == 'push'
if: github.event.label.name == 'ready' || github.event_name == 'push' || github.event_name == 'schedule'
runs-on: ${{ matrix.os }}
strategy:
matrix:
@ -63,20 +72,78 @@ jobs:
if: steps.my-cache.outputs.cache-hit != 'true'
shell: bash
run: |
mkdir -p ~/tmp
cd ~/tmp
git clone https://github.com/csukuangfj/kaldifeat
cd kaldifeat
mkdir build
cd build
cmake -DCMAKE_BUILD_TYPE=Release ..
make -j2 _kaldifeat
.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
- 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: ${{ github.event_name }}
run: |
mkdir -p egs/librispeech/ASR/data
ln -sfv ~/tmp/fbank-libri egs/librispeech/ASR/data/fbank
ls -lh egs/librispeech/ASR/data/*
sudo apt-get -qq install git-lfs tree sox
export PYTHONPATH=$PWD:$PYTHONPATH
export PYTHONPATH=~/tmp/kaldifeat/kaldifeat/python:$PYTHONPATH
export PYTHONPATH=~/tmp/kaldifeat/build/lib:$PYTHONPATH
.github/scripts/run-librispeech-pruned-transducer-stateless-2022-03-12.sh
- name: Display decoding results
if: github.event_name == 'schedule'
shell: bash
run: |
cd egs/librispeech/ASR/
tree ./pruned_transducer_stateless/exp
cd pruned_transducer_stateless
echo "results for pruned_transducer_stateless"
echo "===greedy search==="
find exp/greedy_search -name "log-*" -exec grep -n --color "best for test-clean" {} + | sort -n -k2
find exp/greedy_search -name "log-*" -exec grep -n --color "best for test-other" {} + | sort -n -k2
echo "===fast_beam_search==="
find exp/fast_beam_search -name "log-*" -exec grep -n --color "best for test-clean" {} + | sort -n -k2
find exp/fast_beam_search -name "log-*" -exec grep -n --color "best for test-other" {} + | sort -n -k2
- name: Upload decoding results for pruned_transducer_stateless
uses: actions/upload-artifact@v2
if: github.event_name == 'schedule'
with:
name: torch-${{ matrix.torch }}-python-${{ matrix.python-version }}-ubuntu-18.04-cpu-pruned_transducer_stateless-2022-03-12
path: egs/librispeech/ASR/pruned_transducer_stateless/exp/

101
.github/workflows/run-librispeech-2022-04-29.yml vendored
View File

@ -24,9 +24,18 @@ on:
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_04_29:
if: github.event.label.name == 'ready' || github.event_name == 'push'
if: github.event.label.name == 'ready' || github.event_name == 'push' || github.event_name == 'schedule'
runs-on: ${{ matrix.os }}
strategy:
matrix:
@ -63,18 +72,50 @@ jobs:
if: steps.my-cache.outputs.cache-hit != 'true'
shell: bash
run: |
mkdir -p ~/tmp
cd ~/tmp
git clone https://github.com/csukuangfj/kaldifeat
cd kaldifeat
mkdir build
cd build
cmake -DCMAKE_BUILD_TYPE=Release ..
make -j2 _kaldifeat
.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
- 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: ${{ github.event_name }}
run: |
mkdir -p egs/librispeech/ASR/data
ln -sfv ~/tmp/fbank-libri egs/librispeech/ASR/data/fbank
ls -lh egs/librispeech/ASR/data/*
sudo apt-get -qq install git-lfs tree sox
export PYTHONPATH=$PWD:$PYTHONPATH
export PYTHONPATH=~/tmp/kaldifeat/kaldifeat/python:$PYTHONPATH
@ -83,3 +124,45 @@ jobs:
.github/scripts/run-librispeech-pruned-transducer-stateless2-2022-04-29.sh
.github/scripts/run-librispeech-pruned-transducer-stateless3-2022-04-29.sh
- name: Display decoding results
if: github.event_name == 'schedule'
shell: bash
run: |
cd egs/librispeech/ASR
tree pruned_transducer_stateless2/exp
cd pruned_transducer_stateless2
echo "results for pruned_transducer_stateless2"
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
cd ../
tree pruned_transducer_stateless3/exp
cd pruned_transducer_stateless3
echo "results for pruned_transducer_stateless3"
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
- name: Upload decoding results for pruned_transducer_stateless2
uses: actions/upload-artifact@v2
if: github.event_name == 'schedule'
with:
name: torch-${{ matrix.torch }}-python-${{ matrix.python-version }}-ubuntu-18.04-cpu-pruned_transducer_stateless2-2022-04-29
path: egs/librispeech/ASR/pruned_transducer_stateless2/exp/
- name: Upload decoding results for pruned_transducer_stateless3
uses: actions/upload-artifact@v2
if: github.event_name == 'schedule'
with:
name: torch-${{ matrix.torch }}-python-${{ matrix.python-version }}-ubuntu-18.04-cpu-pruned_transducer_stateless3-2022-04-29
path: egs/librispeech/ASR/pruned_transducer_stateless3/exp/

84
.github/workflows/run-librispeech-transducer-stateless2-2022-04-19.yml vendored
View File

@ -24,9 +24,18 @@ on:
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_04_19:
if: github.event.label.name == 'ready' || github.event_name == 'push'
if: github.event.label.name == 'ready' || github.event_name == 'push' || github.event_name == 'schedule'
runs-on: ${{ matrix.os }}
strategy:
matrix:
@ -63,20 +72,77 @@ jobs:
if: steps.my-cache.outputs.cache-hit != 'true'
shell: bash
run: |
mkdir -p ~/tmp
cd ~/tmp
git clone https://github.com/csukuangfj/kaldifeat
cd kaldifeat
mkdir build
cd build
cmake -DCMAKE_BUILD_TYPE=Release ..
make -j2 _kaldifeat
.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
- 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: ${{ github.event_name }}
run: |
mkdir -p egs/librispeech/ASR/data
ln -sfv ~/tmp/fbank-libri egs/librispeech/ASR/data/fbank
ls -lh egs/librispeech/ASR/data/*
sudo apt-get -qq install git-lfs tree sox
export PYTHONPATH=$PWD:$PYTHONPATH
export PYTHONPATH=~/tmp/kaldifeat/kaldifeat/python:$PYTHONPATH
export PYTHONPATH=~/tmp/kaldifeat/build/lib:$PYTHONPATH
.github/scripts/run-librispeech-transducer-stateless2-2022-04-19.sh
- name: Display decoding results
if: github.event_name == 'schedule'
shell: bash
run: |
cd egs/librispeech/ASR/
tree ./transducer_stateless2/exp
cd transducer_stateless2
echo "results for transducer_stateless2"
echo "===greedy search==="
find exp/greedy_search -name "log-*" -exec grep -n --color "best for test-clean" {} + | sort -n -k2
find exp/greedy_search -name "log-*" -exec grep -n --color "best for test-other" {} + | sort -n -k2
echo "===modified_beam_search==="
find exp/modified_beam_search -name "log-*" -exec grep -n --color "best for test-clean" {} + | sort -n -k2
find exp/modified_beam_search -name "log-*" -exec grep -n --color "best for test-other" {} + | sort -n -k2
- name: Upload decoding results for transducer_stateless2
uses: actions/upload-artifact@v2
if: github.event_name == 'schedule'
with:
name: torch-${{ matrix.torch }}-python-${{ matrix.python-version }}-ubuntu-18.04-cpu-transducer_stateless2-2022-04-19
path: egs/librispeech/ASR/transducer_stateless2/exp/

9
.github/workflows/run-pretrained-conformer-ctc.yml vendored
View File

@ -62,14 +62,7 @@ jobs:
if: steps.my-cache.outputs.cache-hit != 'true'
shell: bash
run: |
mkdir -p ~/tmp
cd ~/tmp
git clone https://github.com/csukuangfj/kaldifeat
cd kaldifeat
mkdir build
cd build
cmake -DCMAKE_BUILD_TYPE=Release ..
make -j2 _kaldifeat
.github/scripts/install-kaldifeat.sh
- name: Inference with pre-trained model
shell: bash

9
.github/workflows/run-pretrained-transducer-stateless-librispeech-100h.yml vendored
View File

@ -62,14 +62,7 @@ jobs:
if: steps.my-cache.outputs.cache-hit != 'true'
shell: bash
run: |
mkdir -p ~/tmp
cd ~/tmp
git clone https://github.com/csukuangfj/kaldifeat
cd kaldifeat
mkdir build
cd build
cmake -DCMAKE_BUILD_TYPE=Release ..
make -j2 _kaldifeat
.github/scripts/install-kaldifeat.sh
- name: Inference with pre-trained model
shell: bash

9
.github/workflows/run-pretrained-transducer-stateless-librispeech-multi-datasets.yml vendored
View File

@ -62,14 +62,7 @@ jobs:
if: steps.my-cache.outputs.cache-hit != 'true'
shell: bash
run: |
mkdir -p ~/tmp
cd ~/tmp
git clone https://github.com/csukuangfj/kaldifeat
cd kaldifeat
mkdir build
cd build
cmake -DCMAKE_BUILD_TYPE=Release ..
make -j2 _kaldifeat
.github/scripts/install-kaldifeat.sh
- name: Inference with pre-trained model
shell: bash

9
.github/workflows/run-pretrained-transducer-stateless-modified-2-aishell.yml vendored
View File

@ -62,14 +62,7 @@ jobs:
if: steps.my-cache.outputs.cache-hit != 'true'
shell: bash
run: |
mkdir -p ~/tmp
cd ~/tmp
git clone https://github.com/csukuangfj/kaldifeat
cd kaldifeat
mkdir build
cd build
cmake -DCMAKE_BUILD_TYPE=Release ..
make -j2 _kaldifeat
.github/scripts/install-kaldifeat.sh
- name: Inference with pre-trained model
shell: bash

9
.github/workflows/run-pretrained-transducer-stateless-modified-aishell.yml vendored
View File

@ -62,14 +62,7 @@ jobs:
if: steps.my-cache.outputs.cache-hit != 'true'
shell: bash
run: |
mkdir -p ~/tmp
cd ~/tmp
git clone https://github.com/csukuangfj/kaldifeat
cd kaldifeat
mkdir build
cd build
cmake -DCMAKE_BUILD_TYPE=Release ..
make -j2 _kaldifeat
.github/scripts/install-kaldifeat.sh
- name: Inference with pre-trained model
shell: bash

9
.github/workflows/run-pretrained-transducer-stateless.yml vendored
View File

@ -62,14 +62,7 @@ jobs:
if: steps.my-cache.outputs.cache-hit != 'true'
shell: bash
run: |
mkdir -p ~/tmp
cd ~/tmp
git clone https://github.com/csukuangfj/kaldifeat
cd kaldifeat
mkdir build
cd build
cmake -DCMAKE_BUILD_TYPE=Release ..
make -j2 _kaldifeat
.github/scripts/install-kaldifeat.sh
- name: Inference with pre-trained model
shell: bash

7
.github/workflows/run-pretrained-transducer.yml vendored
View File

@ -62,13 +62,6 @@ jobs:
if: steps.my-cache.outputs.cache-hit != 'true'
shell: bash
run: |
mkdir -p ~/tmp
cd ~/tmp
git clone https://github.com/csukuangfj/kaldifeat
cd kaldifeat
mkdir build
cd build
cmake -DCMAKE_BUILD_TYPE=Release ..
make -j2 _kaldifeat
- name: Inference with pre-trained model

23
egs/aishell/ASR/transducer_stateless/export.py
View File

@ -1,6 +1,7 @@
#!/usr/bin/env python3
#
# Copyright 2021 Xiaomi Corporation (Author: Fangjun Kuang)
# 2022 Xiaomi Corporation (Author: Mingshuang Luo)
#
# See ../../../../LICENSE for clarification regarding multiple authors
#
@ -22,7 +23,7 @@
Usage:
./transducer_stateless/export.py \
--exp-dir ./transducer_stateless/exp \
--bpe-model data/lang_bpe_500/bpe.model \
--lang-dir data/lang_char \
--epoch 20 \
--avg 10
@ -33,20 +34,19 @@ To use the generated file with `transducer_stateless/decode.py`, you can do:
cd /path/to/exp_dir
ln -s pretrained.pt epoch-9999.pt
cd /path/to/egs/librispeech/ASR
cd /path/to/egs/aishell/ASR
./transducer_stateless/decode.py \
--exp-dir ./transducer_stateless/exp \
--epoch 9999 \
--avg 1 \
--max-duration 1 \
--bpe-model data/lang_bpe_500/bpe.model
--lang-dir data/lang_char
"""
import argparse
import logging
from pathlib import Path
import sentencepiece as spm
import torch
import torch.nn as nn
from conformer import Conformer
@ -56,6 +56,7 @@ from model import Transducer
from icefall.checkpoint import average_checkpoints, load_checkpoint
from icefall.env import get_env_info
from icefall.lexicon import Lexicon
from icefall.utils import AttributeDict, str2bool
@ -91,10 +92,10 @@ def get_parser():
)
parser.add_argument(
"--bpe-model",
"--lang-dir",
type=str,
default="data/lang_bpe_500/bpe.model",
help="Path to the BPE model",
default="data/lang_char",
help="The lang dir",
)
parser.add_argument(
@ -194,12 +195,10 @@ def main():
logging.info(f"device: {device}")
sp = spm.SentencePieceProcessor()
sp.load(params.bpe_model)
lexicon = Lexicon(params.lang_dir)
# <blk> is defined in local/train_bpe_model.py
params.blank_id = sp.piece_to_id("<blk>")
params.vocab_size = sp.get_piece_size()
params.blank_id = 0
params.vocab_size = max(lexicon.tokens) + 1
logging.info(params)

4
egs/gigaspeech/ASR/conformer_ctc/decode.py
View File

@ -177,8 +177,8 @@ def post_processing(
) -> List[Tuple[List[str], List[str]]]:
new_results = []
for ref, hyp in results:
new_ref = asr_text_post_processing(" ".join(ref))
new_hyp = asr_text_post_processing(" ".join(hyp))
new_ref = asr_text_post_processing(" ".join(ref)).split()
new_hyp = asr_text_post_processing(" ".join(hyp)).split()
new_results.append((new_ref, new_hyp))
return new_results

13
egs/librispeech/ASR/pruned_transducer_stateless/beam_search.py
View File

@ -15,6 +15,7 @@
# See the License for the specific language governing permissions and
# limitations under the License.
import warnings
from dataclasses import dataclass
from typing import Dict, List, Optional
@ -565,8 +566,10 @@ def modified_beam_search(
for i in range(batch_size):
topk_log_probs, topk_indexes = ragged_log_probs[i].topk(beam)
topk_hyp_indexes = (topk_indexes // vocab_size).tolist()
topk_token_indexes = (topk_indexes % vocab_size).tolist()
with warnings.catch_warnings():
warnings.simplefilter("ignore")
topk_hyp_indexes = (topk_indexes // vocab_size).tolist()
topk_token_indexes = (topk_indexes % vocab_size).tolist()
for k in range(len(topk_hyp_indexes)):
hyp_idx = topk_hyp_indexes[k]
@ -679,8 +682,10 @@ def _deprecated_modified_beam_search(
topk_hyp_indexes = topk_indexes // logits.size(-1)
topk_token_indexes = topk_indexes % logits.size(-1)
topk_hyp_indexes = topk_hyp_indexes.tolist()
topk_token_indexes = topk_token_indexes.tolist()
with warnings.catch_warnings():
warnings.simplefilter("ignore")
topk_hyp_indexes = topk_hyp_indexes.tolist()
topk_token_indexes = topk_token_indexes.tolist()
for i in range(len(topk_hyp_indexes)):
hyp = A[topk_hyp_indexes[i]]

10
egs/librispeech/ASR/pruned_transducer_stateless2/beam_search.py
View File

@ -276,7 +276,7 @@ def greedy_search(
context_size = model.decoder.context_size
unk_id = getattr(model, "unk_id", blank_id)
device = model.device
device = next(model.parameters()).device
decoder_input = torch.tensor(
[blank_id] * context_size, device=device, dtype=torch.int64
@ -350,7 +350,7 @@ def greedy_search_batch(
assert encoder_out.ndim == 3
assert encoder_out.size(0) >= 1, encoder_out.size(0)
device = model.device
device = next(model.parameters()).device
batch_size = encoder_out.size(0)
T = encoder_out.size(1)
@ -580,7 +580,7 @@ def modified_beam_search(
blank_id = model.decoder.blank_id
unk_id = getattr(model, "unk_id", blank_id)
context_size = model.decoder.context_size
device = model.device
device = next(model.parameters()).device
B = [HypothesisList() for _ in range(batch_size)]
for i in range(batch_size):
B[i].add(
@ -705,7 +705,7 @@ def _deprecated_modified_beam_search(
unk_id = getattr(model, "unk_id", blank_id)
context_size = model.decoder.context_size
device = model.device
device = next(model.parameters()).device
T = encoder_out.size(1)
@ -813,7 +813,7 @@ def beam_search(
unk_id = getattr(model, "unk_id", blank_id)
context_size = model.decoder.context_size
device = model.device
device = next(model.parameters()).device
decoder_input = torch.tensor(
[blank_id] * context_size,

82
egs/librispeech/ASR/pruned_transducer_stateless2/train.py
View File

@ -156,15 +156,16 @@ def get_parser():
"--initial-lr",
type=float,
default=0.003,
help="The initial learning rate. This value should not need to be changed.",
help="The initial learning rate. This value should not need to "
"be changed.",
)
parser.add_argument(
"--lr-batches",
type=float,
default=5000,
help="""Number of steps that affects how rapidly the learning rate decreases.
We suggest not to change this.""",
help="""Number of steps that affects how rapidly the learning rate
decreases. We suggest not to change this.""",
)
parser.add_argument(
@ -670,25 +671,29 @@ def train_one_epoch(
params.batch_idx_train += 1
batch_size = len(batch["supervisions"]["text"])
with torch.cuda.amp.autocast(enabled=params.use_fp16):
loss, loss_info = compute_loss(
params=params,
model=model,
sp=sp,
batch=batch,
is_training=True,
warmup=(params.batch_idx_train / params.model_warm_step),
)
# summary stats
tot_loss = (tot_loss * (1 - 1 / params.reset_interval)) + loss_info
try:
with torch.cuda.amp.autocast(enabled=params.use_fp16):
loss, loss_info = compute_loss(
params=params,
model=model,
sp=sp,
batch=batch,
is_training=True,
warmup=(params.batch_idx_train / params.model_warm_step),
)
# summary stats
tot_loss = (tot_loss * (1 - 1 / params.reset_interval)) + loss_info
# NOTE: We use reduction==sum and loss is computed over utterances
# in the batch and there is no normalization to it so far.
scaler.scale(loss).backward()
scheduler.step_batch(params.batch_idx_train)
scaler.step(optimizer)
scaler.update()
optimizer.zero_grad()
# NOTE: We use reduction==sum and loss is computed over utterances
# in the batch and there is no normalization to it so far.
scaler.scale(loss).backward()
scheduler.step_batch(params.batch_idx_train)
scaler.step(optimizer)
scaler.update()
optimizer.zero_grad()
except: # noqa
display_and_save_batch(batch, params=params, sp=sp)
raise
if params.print_diagnostics and batch_idx == 5:
return
@ -933,6 +938,38 @@ def run(rank, world_size, args):
cleanup_dist()
def display_and_save_batch(
batch: dict,
params: AttributeDict,
sp: spm.SentencePieceProcessor,
) -> None:
"""Display the batch statistics and save the batch into disk.
Args:
batch:
A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()`
for the content in it.
params:
Parameters for training. See :func:`get_params`.
sp:
The BPE model.
"""
from lhotse.utils import uuid4
filename = f"{params.exp_dir}/batch-{uuid4()}.pt"
logging.info(f"Saving batch to {filename}")
torch.save(batch, filename)
supervisions = batch["supervisions"]
features = batch["inputs"]
logging.info(f"features shape: {features.shape}")
y = sp.encode(supervisions["text"], out_type=int)
num_tokens = sum(len(i) for i in y)
logging.info(f"num tokens: {num_tokens}")
def scan_pessimistic_batches_for_oom(
model: nn.Module,
train_dl: torch.utils.data.DataLoader,
@ -964,7 +1001,7 @@ def scan_pessimistic_batches_for_oom(
loss.backward()
optimizer.step()
optimizer.zero_grad()
except RuntimeError as e:
except Exception as e:
if "CUDA out of memory" in str(e):
logging.error(
"Your GPU ran out of memory with the current "
@ -973,6 +1010,7 @@ def scan_pessimistic_batches_for_oom(
f"Failing criterion: {criterion} "
f"(={crit_values[criterion]}) ..."
)
display_and_save_batch(batch, params=params, sp=sp)
raise

131
egs/librispeech/ASR/pruned_transducer_stateless3/decode-giga.py
View File

@ -69,7 +69,8 @@ import torch.nn as nn
from asr_datamodule import AsrDataModule
from beam_search import (
beam_search,
fast_beam_search,
fast_beam_search_nbest_oracle,
fast_beam_search_one_best,
greedy_search,
greedy_search_batch,
modified_beam_search,
@ -100,27 +101,28 @@ def get_parser():
"--epoch",
type=int,
default=28,
help="It specifies the checkpoint to use for decoding."
"Note: Epoch counts from 0.",
help="""It specifies the checkpoint to use for decoding.
Note: Epoch counts from 0.
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'. ",
)
parser.add_argument(
"--avg-last-n",
type=int,
default=0,
help="""If positive, --epoch and --avg are ignored and it
will use the last n checkpoints exp_dir/checkpoint-xxx.pt
where xxx is the number of processed batches while
saving that checkpoint.
""",
"'--epoch' and '--iter'",
)
parser.add_argument(
@ -146,6 +148,7 @@ def get_parser():
- beam_search
- modified_beam_search
- fast_beam_search
- fast_beam_search_nbest_oracle
""",
)
@ -165,7 +168,8 @@ def get_parser():
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""",
Used only when --decoding-method is
fast_beam_search or fast_beam_search_nbest_oracle""",
)
parser.add_argument(
@ -173,7 +177,7 @@ def get_parser():
type=int,
default=4,
help="""Used only when --decoding-method is
fast_beam_search""",
fast_beam_search or fast_beam_search_nbest_oracle""",
)
parser.add_argument(
@ -181,7 +185,7 @@ def get_parser():
type=int,
default=8,
help="""Used only when --decoding-method is
fast_beam_search""",
fast_beam_search or fast_beam_search_nbest_oracle""",
)
parser.add_argument(
@ -199,6 +203,23 @@ def get_parser():
Used only when --decoding_method is greedy_search""",
)
parser.add_argument(
"--num-paths",
type=int,
default=100,
help="""Number of paths for computed nbest oracle WER
when the decoding method is 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_oracle.
""",
)
return parser
@ -243,7 +264,8 @@ def decode_one_batch(
for the format of the `batch`.
decoding_graph:
The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used
only when --decoding_method is fast_beam_search.
only when --decoding_method is
fast_beam_search or fast_beam_search_nbest_oracle.
Returns:
Return the decoding result. See above description for the format of
the returned dict.
@ -264,7 +286,7 @@ def decode_one_batch(
hyps = []
if params.decoding_method == "fast_beam_search":
hyp_tokens = fast_beam_search(
hyp_tokens = fast_beam_search_one_best(
model=model,
decoding_graph=decoding_graph,
encoder_out=encoder_out,
@ -275,6 +297,21 @@ def decode_one_batch(
)
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
@ -328,6 +365,16 @@ def decode_one_batch(
f"max_states_{params.max_states}"
): hyps
}
elif params.decoding_method == "fast_beam_search_nbest_oracle":
return {
(
f"beam_{params.beam}_"
f"max_contexts_{params.max_contexts}_"
f"max_states_{params.max_states}_"
f"num_paths_{params.num_paths}_"
f"nbest_scale_{params.nbest_scale}"
): hyps
}
else:
return {f"beam_size_{params.beam_size}": hyps}
@ -463,17 +510,30 @@ def main():
"greedy_search",
"beam_search",
"fast_beam_search",
"fast_beam_search_nbest_oracle",
"modified_beam_search",
)
params.res_dir = params.exp_dir / "giga" / params.decoding_method
params.suffix = f"epoch-{params.epoch}-avg-{params.avg}"
if "fast_beam_search" in 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.decoding_method == "fast_beam_search":
params.suffix += f"-beam-{params.beam}"
params.suffix += f"-max-contexts-{params.max_contexts}"
params.suffix += f"-max-states-{params.max_states}"
elif params.decoding_method == "fast_beam_search_nbest_oracle":
params.suffix += f"-beam-{params.beam}"
params.suffix += f"-max-contexts-{params.max_contexts}"
params.suffix += f"-max-states-{params.max_states}"
params.suffix += f"-num-paths-{params.num_paths}"
params.suffix += f"-nbest-scale-{params.nbest_scale}"
elif "beam_search" in params.decoding_method:
params.suffix += f"-beam-{params.beam_size}"
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}"
@ -490,8 +550,9 @@ def main():
sp = spm.SentencePieceProcessor()
sp.load(params.bpe_model)
# <blk> is defined in local/train_bpe_model.py
# <blk> and <unk> is defined in local/train_bpe_model.py
params.blank_id = sp.piece_to_id("<blk>")
params.unk_id = sp.unk_id()
params.vocab_size = sp.get_piece_size()
logging.info(params)
@ -499,8 +560,20 @@ def main():
logging.info("About to create model")
model = get_transducer_model(params)
if params.avg_last_n > 0:
filenames = find_checkpoints(params.exp_dir)[: params.avg_last_n]
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))
@ -519,13 +592,17 @@ def main():
model.to(device)
model.eval()
model.device = device
model.unk_id = params.unk_id
# In beam_search.py, we are using model.decoder() and model.joiner(),
# so we have to switch to the branch for the GigaSpeech dataset.
model.decoder = model.decoder_giga
model.joiner = model.joiner_giga
if params.decoding_method == "fast_beam_search":
if params.decoding_method in (
"fast_beam_search",
"fast_beam_search_nbest_oracle",
):
decoding_graph = k2.trivial_graph(params.vocab_size - 1, device=device)
else:
decoding_graph = None

1
egs/librispeech/ASR/pruned_transducer_stateless4/__init__.py Symbolic link
View File

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

1
egs/librispeech/ASR/pruned_transducer_stateless4/asr_datamodule.py Symbolic link
View File

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

1
egs/librispeech/ASR/pruned_transducer_stateless4/beam_search.py Symbolic link
View File

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

1
egs/librispeech/ASR/pruned_transducer_stateless4/conformer.py Symbolic link
View File

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

631
egs/librispeech/ASR/pruned_transducer_stateless4/decode.py Executable file
View File

@ -0,0 +1,631 @@
#!/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_stateless4/decode.py \
--epoch 30 \
--avg 15 \
--exp-dir ./pruned_transducer_stateless2/exp \
--max-duration 100 \
--decoding-method greedy_search
(2) beam search
./pruned_transducer_stateless4/decode.py \
--epoch 30 \
--avg 15 \
--exp-dir ./pruned_transducer_stateless2/exp \
--max-duration 100 \
--decoding-method beam_search \
--beam-size 4
(3) modified beam search
./pruned_transducer_stateless4/decode.py \
--epoch 30 \
--avg 15 \
--exp-dir ./pruned_transducer_stateless2/exp \
--max-duration 100 \
--decoding-method modified_beam_search \
--beam-size 4
(4) fast beam search
./pruned_transducer_stateless4/decode.py \
--epoch 30 \
--avg 15 \
--exp-dir ./pruned_transducer_stateless2/exp \
--max-duration 1500 \
--decoding-method fast_beam_search \
--beam 4 \
--max-contexts 4 \
--max-states 8
"""
import argparse
import logging
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,
greedy_search,
greedy_search_batch,
modified_beam_search,
)
from train import get_params, get_transducer_model
from icefall.checkpoint import (
average_checkpoints,
average_checkpoints_with_averaged_model,
find_checkpoints,
load_checkpoint,
)
from icefall.utils import (
AttributeDict,
setup_logger,
store_transcripts,
str2bool,
write_error_stats,
)
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=False,
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_stateless4/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(
"--decoding-method",
type=str,
default="greedy_search",
help="""Possible values are:
- greedy_search
- beam_search
- modified_beam_search
- fast_beam_search
""",
)
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=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 --decoding-method is fast_beam_search""",
)
parser.add_argument(
"--max-contexts",
type=int,
default=4,
help="""Used only when --decoding-method is
fast_beam_search""",
)
parser.add_argument(
"--max-states",
type=int,
default=8,
help="""Used only when --decoding-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 --decoding_method is greedy_search""",
)
return parser
def decode_one_batch(
params: AttributeDict,
model: nn.Module,
sp: spm.SentencePieceProcessor,
batch: dict,
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`.
decoding_graph:
The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used
only when --decoding_method is fast_beam_search.
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)
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(
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 == "greedy_search"
and params.max_sym_per_frame == 1
):
hyp_tokens = greedy_search_batch(
model=model,
encoder_out=encoder_out,
)
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,
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 params.decoding_method == "fast_beam_search":
return {
(
f"beam_{params.beam}_"
f"max_contexts_{params.max_contexts}_"
f"max_states_{params.max_states}"
): 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,
decoding_graph: Optional[k2.Fsa] = None,
) -> Dict[str, List[Tuple[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.
decoding_graph:
The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used
only when --decoding_method is fast_beam_search.
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 = 100
else:
log_interval = 2
results = defaultdict(list)
for batch_idx, batch in enumerate(dl):
texts = batch["supervisions"]["text"]
hyps_dict = decode_one_batch(
params=params,
model=model,
sp=sp,
decoding_graph=decoding_graph,
batch=batch,
)
for name, hyps in hyps_dict.items():
this_batch = []
assert len(hyps) == len(texts)
for hyp_words, ref_text in zip(hyps, texts):
ref_words = ref_text.split()
this_batch.append((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[List[int], List[int]]]],
):
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"
)
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",
"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 "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}"
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> 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(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
start = params.epoch - params.avg
assert start >= 1
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 params.decoding_method == "fast_beam_search":
decoding_graph = k2.trivial_graph(params.vocab_size - 1, device=device)
else:
decoding_graph = None
num_param = sum([p.numel() for p in model.parameters()])
logging.info(f"Number of model parameters: {num_param}")
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,
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_stateless4/decoder.py Symbolic link
View File

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

1
egs/librispeech/ASR/pruned_transducer_stateless4/encoder_interface.py Symbolic link
View File

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

1
egs/librispeech/ASR/pruned_transducer_stateless4/export.py Symbolic link
View File

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

1
egs/librispeech/ASR/pruned_transducer_stateless4/joiner.py Symbolic link
View File

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

1
egs/librispeech/ASR/pruned_transducer_stateless4/model.py Symbolic link
View File

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

1
egs/librispeech/ASR/pruned_transducer_stateless4/optim.py Symbolic link
View File

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

1
egs/librispeech/ASR/pruned_transducer_stateless4/scaling.py Symbolic link
View File

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

1053
egs/librispeech/ASR/pruned_transducer_stateless4/train.py Executable file
View File

File diff suppressed because it is too large Load Diff

146
icefall/checkpoint.py
View File

@ -1,4 +1,5 @@
# Copyright 2021 Xiaomi Corp. (authors: Fangjun Kuang)
# Copyright 2021-2022 Xiaomi Corporation (authors: Fangjun Kuang,
# Zengwei Yao)
#
# See ../../LICENSE for clarification regarding multiple authors
#
@ -25,6 +26,7 @@ from typing import Any, Dict, List, Optional, Union
import torch
import torch.nn as nn
from lhotse.dataset.sampling.base import CutSampler
from torch import Tensor
from torch.cuda.amp import GradScaler
from torch.nn.parallel import DistributedDataParallel as DDP
from torch.optim import Optimizer
@ -37,6 +39,7 @@ LRSchedulerType = object
def save_checkpoint(
filename: Path,
model: Union[nn.Module, DDP],
model_avg: Optional[nn.Module] = None,
params: Optional[Dict[str, Any]] = None,
optimizer: Optional[Optimizer] = None,
scheduler: Optional[LRSchedulerType] = None,
@ -51,6 +54,8 @@ def save_checkpoint(
The checkpoint filename.
model:
The model to be saved. We only save its `state_dict()`.
model_avg:
The stored model averaged from the start of training.
params:
User defined parameters, e.g., epoch, loss.
optimizer:
@ -80,6 +85,9 @@ def save_checkpoint(
"sampler": sampler.state_dict() if sampler is not None else None,
}
if model_avg is not None:
checkpoint["model_avg"] = model_avg.state_dict()
if params:
for k, v in params.items():
assert k not in checkpoint
@ -91,6 +99,7 @@ def save_checkpoint(
def load_checkpoint(
filename: Path,
model: nn.Module,
model_avg: Optional[nn.Module] = None,
optimizer: Optional[Optimizer] = None,
scheduler: Optional[LRSchedulerType] = None,
scaler: Optional[GradScaler] = None,
@ -118,6 +127,11 @@ def load_checkpoint(
checkpoint.pop("model")
if model_avg is not None and "model_avg" in checkpoint:
logging.info("Loading averaged model")
model_avg.load_state_dict(checkpoint["model_avg"], strict=strict)
checkpoint.pop("model_avg")
def load(name, obj):
s = checkpoint.get(name, None)
if obj and s:
@ -181,6 +195,7 @@ def save_checkpoint_with_global_batch_idx(
out_dir: Path,
global_batch_idx: int,
model: Union[nn.Module, DDP],
model_avg: Optional[nn.Module] = None,
params: Optional[Dict[str, Any]] = None,
optimizer: Optional[Optimizer] = None,
scheduler: Optional[LRSchedulerType] = None,
@ -201,6 +216,8 @@ def save_checkpoint_with_global_batch_idx(
model:
The neural network model whose `state_dict` will be saved in the
checkpoint.
model_avg:
The stored model averaged from the start of training.
params:
A dict of training configurations to be saved.
optimizer:
@ -223,6 +240,7 @@ def save_checkpoint_with_global_batch_idx(
save_checkpoint(
filename=filename,
model=model,
model_avg=model_avg,
params=params,
optimizer=optimizer,
scheduler=scheduler,
@ -327,3 +345,129 @@ def remove_checkpoints(
to_remove = checkpoints[topk:]
for c in to_remove:
os.remove(c)
def update_averaged_model(
params: Dict[str, Tensor],
model_cur: Union[nn.Module, DDP],
model_avg: nn.Module,
) -> None:
"""Update the averaged model:
model_avg = model_cur * (average_period / batch_idx_train)
+ model_avg * ((batch_idx_train - average_period) / batch_idx_train)
Args:
params:
User defined parameters, e.g., epoch, loss.
model_cur:
The current model.
model_avg:
The averaged model to be updated.
"""
weight_cur = params.average_period / params.batch_idx_train
weight_avg = 1 - weight_cur
if isinstance(model_cur, DDP):
model_cur = model_cur.module
cur = model_cur.state_dict()
avg = model_avg.state_dict()
average_state_dict(
state_dict_1=avg,
state_dict_2=cur,
weight_1=weight_avg,
weight_2=weight_cur,
)
def average_checkpoints_with_averaged_model(
filename_start: str,
filename_end: str,
device: torch.device = torch.device("cpu"),
) -> Dict[str, Tensor]:
"""Average model parameters over the range with given
start model (excluded) and end model.
Let start = batch_idx_train of model-start;
end = batch_idx_train of model-end;
interval = end - start.
Then the average model over range from start (excluded) to end is
(1) avg = (model_end * end - model_start * start) / interval.
It can be written as
(2) avg = model_end * weight_end + model_start * weight_start,
where weight_end = end / interval,
weight_start = -start / interval = 1 - weight_end.
Since the terms `weight_end` and `weight_start` would be large
if the model has been trained for lots of batches, which would cause
overflow when multiplying the model parameters.
To avoid this, we rewrite (2) as:
(3) avg = (model_end + model_start * (weight_start / weight_end))
* weight_end
The model index could be epoch number or iteration number.
Args:
filename_start:
Checkpoint filename of the start model. We assume it
is saved by :func:`save_checkpoint`.
filename_end:
Checkpoint filename of the end model. We assume it
is saved by :func:`save_checkpoint`.
device:
Move checkpoints to this device before averaging.
"""
state_dict_start = torch.load(filename_start, map_location=device)
state_dict_end = torch.load(filename_end, map_location=device)
batch_idx_train_start = state_dict_start["batch_idx_train"]
batch_idx_train_end = state_dict_end["batch_idx_train"]
interval = batch_idx_train_end - batch_idx_train_start
assert interval > 0, interval
weight_end = batch_idx_train_end / interval
weight_start = 1 - weight_end
model_end = state_dict_end["model_avg"]
model_start = state_dict_start["model_avg"]
avg = model_end
# scale the weight to avoid overflow
average_state_dict(
state_dict_1=avg,
state_dict_2=model_start,
weight_1=1.0,
weight_2=weight_start / weight_end,
scaling_factor=weight_end,
)
return avg
def average_state_dict(
state_dict_1: Dict[str, Tensor],
state_dict_2: Dict[str, Tensor],
weight_1: float,
weight_2: float,
scaling_factor: float = 1.0,
) -> Dict[str, Tensor]:
"""Average two state_dict with given weights:
state_dict_1 = (state_dict_1 * weight_1 + state_dict_2 * weight_2)
* scaling_factor
It is an in-place operation on state_dict_1 itself.
"""
# Identify shared parameters. Two parameters are said to be shared
# if they have the same data_ptr
uniqued: Dict[int, str] = dict()
for k, v in state_dict_1.items():
v_data_ptr = v.data_ptr()
if v_data_ptr in uniqued:
continue
uniqued[v_data_ptr] = k
uniqued_names = list(uniqued.values())
for k in uniqued_names:
state_dict_1[k] *= weight_1
state_dict_1[k] += (
state_dict_2[k].to(device=state_dict_1[k].device) * weight_2
)
state_dict_1[k] *= scaling_factor

2
icefall/env.py
View File

@ -95,7 +95,7 @@ def get_env_info() -> Dict[str, Any]:
"k2-git-sha1": k2.version.__git_sha1__,
"k2-git-date": k2.version.__git_date__,
"lhotse-version": lhotse.__version__,
"torch-version": torch.__version__,
"torch-version": str(torch.__version__),
"torch-cuda-available": torch.cuda.is_available(),
"torch-cuda-version": torch.version.cuda,
"python-version": sys.version[:3],