diff --git a/.github/scripts/.gitignore b/.github/scripts/.gitignore new file mode 100644 index 000000000..672e477d8 --- /dev/null +++ b/.github/scripts/.gitignore @@ -0,0 +1 @@ +piper_phonemize.html diff --git a/.github/scripts/generate-piper-phonemize-page.py b/.github/scripts/generate-piper-phonemize-page.py new file mode 100755 index 000000000..3784d5fa5 --- /dev/null +++ b/.github/scripts/generate-piper-phonemize-page.py @@ -0,0 +1,29 @@ +#!/usr/bin/env python3 + + +def main(): + prefix = ( + "https://github.com/csukuangfj/piper-phonemize/releases/download/2023.12.5/" + ) + files = [ + "piper_phonemize-1.2.0-cp310-cp310-macosx_10_14_x86_64.whl", + "piper_phonemize-1.2.0-cp310-cp310-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", + "piper_phonemize-1.2.0-cp311-cp311-macosx_10_14_x86_64.whl", + "piper_phonemize-1.2.0-cp311-cp311-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", + "piper_phonemize-1.2.0-cp312-cp312-macosx_10_14_x86_64.whl", + "piper_phonemize-1.2.0-cp312-cp312-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", + "piper_phonemize-1.2.0-cp37-cp37m-macosx_10_14_x86_64.whl", + "piper_phonemize-1.2.0-cp37-cp37m-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", + "piper_phonemize-1.2.0-cp38-cp38-macosx_10_14_x86_64.whl", + "piper_phonemize-1.2.0-cp38-cp38-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", + "piper_phonemize-1.2.0-cp39-cp39-macosx_10_14_x86_64.whl", + "piper_phonemize-1.2.0-cp39-cp39-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", + ] + with open("piper_phonemize.html", "w") as f: + for file in files: + url = prefix + file + f.write(f'{file}
\n') + + +if __name__ == "__main__": + main() diff --git a/.github/scripts/librispeech/ASR/run.sh b/.github/scripts/librispeech/ASR/run.sh index 7e9bd8a47..293ed66e5 100755 --- a/.github/scripts/librispeech/ASR/run.sh +++ b/.github/scripts/librispeech/ASR/run.sh @@ -15,9 +15,9 @@ function prepare_data() { # cause OOM error for CI later. mkdir -p download/lm pushd download/lm - wget -q http://www.openslr.org/resources/11/librispeech-vocab.txt - wget -q http://www.openslr.org/resources/11/librispeech-lexicon.txt - wget -q http://www.openslr.org/resources/11/librispeech-lm-norm.txt.gz + wget -q https://huggingface.co/csukuangfj/librispeech-for-ci/resolve/main/librispeech-lm-norm.txt.gz + wget -q https://huggingface.co/csukuangfj/librispeech-for-ci/resolve/main/librispeech-lexicon.txt + wget -q https://huggingface.co/csukuangfj/librispeech-for-ci/resolve/main/librispeech-vocab.txt ls -lh gunzip librispeech-lm-norm.txt.gz diff --git a/.github/scripts/ljspeech/TTS/run.sh b/.github/scripts/ljspeech/TTS/run.sh new file mode 100755 index 000000000..707361782 --- /dev/null +++ b/.github/scripts/ljspeech/TTS/run.sh @@ -0,0 +1,157 @@ +#!/usr/bin/env bash + +set -ex + +python3 -m pip install piper_phonemize -f https://k2-fsa.github.io/icefall/piper_phonemize.html +python3 -m pip install espnet_tts_frontend +python3 -m pip install numba + +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]}) $*" +} + +cd egs/ljspeech/TTS + +sed -i.bak s/600/8/g ./prepare.sh +sed -i.bak s/"first 100"/"first 3"/g ./prepare.sh +sed -i.bak s/500/5/g ./prepare.sh +git diff + +function prepare_data() { + # We have created a subset of the data for testing + # + mkdir download + pushd download + wget -q https://huggingface.co/csukuangfj/ljspeech-subset-for-ci-test/resolve/main/LJSpeech-1.1.tar.bz2 + tar xvf LJSpeech-1.1.tar.bz2 + popd + + ./prepare.sh + tree . +} + +function train() { + pushd ./vits + sed -i.bak s/200/3/g ./train.py + git diff . + popd + + for t in low medium high; do + ./vits/train.py \ + --exp-dir vits/exp-$t \ + --model-type $t \ + --num-epochs 1 \ + --save-every-n 1 \ + --num-buckets 2 \ + --tokens data/tokens.txt \ + --max-duration 20 + + ls -lh vits/exp-$t + done +} + +function infer() { + for t in low medium high; do + ./vits/infer.py \ + --num-buckets 2 \ + --model-type $t \ + --epoch 1 \ + --exp-dir ./vits/exp-$t \ + --tokens data/tokens.txt \ + --max-duration 20 + done +} + +function export_onnx() { + for t in low medium high; do + ./vits/export-onnx.py \ + --model-type $t \ + --epoch 1 \ + --exp-dir ./vits/exp-$t \ + --tokens data/tokens.txt + + ls -lh vits/exp-$t/ + done +} + +function test_medium() { + git clone https://huggingface.co/csukuangfj/icefall-tts-ljspeech-vits-medium-2024-03-12 + + ./vits/export-onnx.py \ + --model-type medium \ + --epoch 820 \ + --exp-dir ./icefall-tts-ljspeech-vits-medium-2024-03-12/exp \ + --tokens ./icefall-tts-ljspeech-vits-medium-2024-03-12/data/tokens.txt + + ls -lh ./icefall-tts-ljspeech-vits-medium-2024-03-12/exp + + ./vits/test_onnx.py \ + --model-filename ./icefall-tts-ljspeech-vits-medium-2024-03-12/exp/vits-epoch-820.onnx \ + --tokens ./icefall-tts-ljspeech-vits-medium-2024-03-12/data/tokens.txt \ + --output-filename /icefall/test-medium.wav + + ls -lh /icefall/test-medium.wav + + d=/icefall/vits-icefall-en_US-ljspeech-medium + mkdir $d + cp -v ./icefall-tts-ljspeech-vits-medium-2024-03-12/data/tokens.txt $d/ + cp -v ./icefall-tts-ljspeech-vits-medium-2024-03-12/exp/vits-epoch-820.onnx $d/model.onnx + + rm -rf icefall-tts-ljspeech-vits-medium-2024-03-12 + + pushd $d + wget -q https://github.com/k2-fsa/sherpa-onnx/releases/download/tts-models/espeak-ng-data.tar.bz2 + tar xf espeak-ng-data.tar.bz2 + rm espeak-ng-data.tar.bz2 + cd .. + tar cjf vits-icefall-en_US-ljspeech-medium.tar.bz2 vits-icefall-en_US-ljspeech-medium + rm -rf vits-icefall-en_US-ljspeech-medium + ls -lh *.tar.bz2 + popd +} + +function test_low() { + git clone https://huggingface.co/csukuangfj/icefall-tts-ljspeech-vits-low-2024-03-12 + + ./vits/export-onnx.py \ + --model-type low \ + --epoch 1600 \ + --exp-dir ./icefall-tts-ljspeech-vits-low-2024-03-12/exp \ + --tokens ./icefall-tts-ljspeech-vits-low-2024-03-12/data/tokens.txt + + ls -lh ./icefall-tts-ljspeech-vits-low-2024-03-12/exp + + ./vits/test_onnx.py \ + --model-filename ./icefall-tts-ljspeech-vits-low-2024-03-12/exp/vits-epoch-1600.onnx \ + --tokens ./icefall-tts-ljspeech-vits-low-2024-03-12/data/tokens.txt \ + --output-filename /icefall/test-low.wav + + ls -lh /icefall/test-low.wav + + d=/icefall/vits-icefall-en_US-ljspeech-low + mkdir $d + cp -v ./icefall-tts-ljspeech-vits-low-2024-03-12/data/tokens.txt $d/ + cp -v ./icefall-tts-ljspeech-vits-low-2024-03-12/exp/vits-epoch-1600.onnx $d/model.onnx + + rm -rf icefall-tts-ljspeech-vits-low-2024-03-12 + + pushd $d + wget -q https://github.com/k2-fsa/sherpa-onnx/releases/download/tts-models/espeak-ng-data.tar.bz2 + tar xf espeak-ng-data.tar.bz2 + rm espeak-ng-data.tar.bz2 + cd .. + tar cjf vits-icefall-en_US-ljspeech-low.tar.bz2 vits-icefall-en_US-ljspeech-low + rm -rf vits-icefall-en_US-ljspeech-low + ls -lh *.tar.bz2 + popd +} + +prepare_data +train +infer +export_onnx +rm -rf vits/exp-{low,medium,high} +test_medium +test_low diff --git a/.github/workflows/build-doc.yml b/.github/workflows/build-doc.yml index d7fe2c964..c622476f2 100644 --- a/.github/workflows/build-doc.yml +++ b/.github/workflows/build-doc.yml @@ -56,11 +56,14 @@ jobs: - name: Build doc shell: bash run: | + .github/scripts/generate-piper-phonemize-page.py cd docs python3 -m pip install -r ./requirements.txt make html touch build/html/.nojekyll + cp -v ../piper_phonemize.html ./build/html/ + - name: Deploy uses: peaceiris/actions-gh-pages@v3 with: diff --git a/.github/workflows/build-docker-image.yml b/.github/workflows/build-docker-image.yml index d5081f7d8..f5796d114 100644 --- a/.github/workflows/build-docker-image.yml +++ b/.github/workflows/build-docker-image.yml @@ -16,7 +16,7 @@ jobs: fail-fast: false matrix: os: [ubuntu-latest] - image: ["torch2.2.0-cuda12.1", "torch2.2.0-cuda11.8", "torch2.1.0-cuda12.1", "torch2.1.0-cuda11.8", "torch2.0.0-cuda11.7", "torch1.13.0-cuda11.6", "torch1.12.1-cuda11.3", "torch1.9.0-cuda10.2"] + image: ["torch2.2.1-cuda12.1", "torch2.2.1-cuda11.8", "torch2.2.0-cuda12.1", "torch2.2.0-cuda11.8", "torch2.1.0-cuda12.1", "torch2.1.0-cuda11.8", "torch2.0.0-cuda11.7", "torch1.13.0-cuda11.6", "torch1.12.1-cuda11.3", "torch1.9.0-cuda10.2"] steps: # refer to https://github.com/actions/checkout diff --git a/.github/workflows/ljspeech.yml b/.github/workflows/ljspeech.yml new file mode 100644 index 000000000..25402275b --- /dev/null +++ b/.github/workflows/ljspeech.yml @@ -0,0 +1,102 @@ +name: ljspeech + +on: + push: + branches: + - master + + pull_request: + branches: + - master + + workflow_dispatch: + +concurrency: + group: ljspeech-${{ github.ref }} + cancel-in-progress: true + +jobs: + generate_build_matrix: + if: github.repository_owner == 'csukuangfj' || github.repository_owner == 'k2-fsa' + # see https://github.com/pytorch/pytorch/pull/50633 + runs-on: ubuntu-latest + outputs: + matrix: ${{ steps.set-matrix.outputs.matrix }} + steps: + - uses: actions/checkout@v4 + with: + fetch-depth: 0 + - name: Generating build matrix + id: set-matrix + run: | + # outputting for debugging purposes + python ./.github/scripts/docker/generate_build_matrix.py + MATRIX=$(python ./.github/scripts/docker/generate_build_matrix.py) + echo "::set-output name=matrix::${MATRIX}" + + ljspeech: + needs: generate_build_matrix + name: py${{ matrix.python-version }} torch${{ matrix.torch-version }} v${{ matrix.version }} + runs-on: ubuntu-latest + strategy: + fail-fast: false + matrix: + ${{ fromJson(needs.generate_build_matrix.outputs.matrix) }} + + steps: + - uses: actions/checkout@v4 + with: + fetch-depth: 0 + + - name: Free space + shell: bash + run: | + ls -lh + df -h + rm -rf /opt/hostedtoolcache + df -h + echo "pwd: $PWD" + echo "github.workspace ${{ github.workspace }}" + + - name: Run tests + uses: addnab/docker-run-action@v3 + with: + image: ghcr.io/${{ github.repository_owner }}/icefall:cpu-py${{ matrix.python-version }}-torch${{ matrix.torch-version }}-v${{ matrix.version }} + options: | + --volume ${{ github.workspace }}/:/icefall + shell: bash + run: | + export PYTHONPATH=/icefall:$PYTHONPATH + cd /icefall + git config --global --add safe.directory /icefall + + .github/scripts/ljspeech/TTS/run.sh + + - name: display files + shell: bash + run: | + ls -lh + + - uses: actions/upload-artifact@v4 + if: matrix.python-version == '3.9' && matrix.torch-version == '2.2.0' + with: + name: generated-test-files-${{ matrix.python-version }}-${{ matrix.torch-version }} + path: ./*.wav + + - uses: actions/upload-artifact@v4 + if: matrix.python-version == '3.9' && matrix.torch-version == '2.2.0' + with: + name: generated-models-py${{ matrix.python-version }}-torch${{ matrix.torch-version }} + path: ./*.wav + + - name: Release exported onnx models + if: matrix.python-version == '3.9' && matrix.torch-version == '2.2.0' + uses: svenstaro/upload-release-action@v2 + with: + file_glob: true + overwrite: true + file: vits-icefall-*.tar.bz2 + repo_name: k2-fsa/sherpa-onnx + repo_token: ${{ secrets.UPLOAD_GH_SHERPA_ONNX_TOKEN }} + tag: tts-models + diff --git a/.github/workflows/run-docker-image.yml b/.github/workflows/run-docker-image.yml index 65ba2cd64..eab31cccc 100644 --- a/.github/workflows/run-docker-image.yml +++ b/.github/workflows/run-docker-image.yml @@ -14,7 +14,7 @@ jobs: fail-fast: false matrix: os: [ubuntu-latest] - image: ["torch2.2.0-cuda12.1", "torch2.2.0-cuda11.8", "torch2.1.0-cuda12.1", "torch2.1.0-cuda11.8", "torch2.0.0-cuda11.7", "torch1.13.0-cuda11.6", "torch1.12.1-cuda11.3", "torch1.9.0-cuda10.2"] + image: ["torch2.2.1-cuda12.1", "torch2.2.1-cuda11.8", "torch2.2.0-cuda12.1", "torch2.2.0-cuda11.8", "torch2.1.0-cuda12.1", "torch2.1.0-cuda11.8", "torch2.0.0-cuda11.7", "torch1.13.0-cuda11.6", "torch1.12.1-cuda11.3", "torch1.9.0-cuda10.2"] steps: # refer to https://github.com/actions/checkout - uses: actions/checkout@v2 diff --git a/.github/workflows/style_check.yml b/.github/workflows/style_check.yml index fc1dcbfd4..1c37f13ed 100644 --- a/.github/workflows/style_check.yml +++ b/.github/workflows/style_check.yml @@ -49,7 +49,7 @@ jobs: - name: Install Python dependencies run: | - python3 -m pip install --upgrade pip black==22.3.0 flake8==5.0.4 click==8.1.0 + python3 -m pip install --upgrade pip black==22.3.0 flake8==5.0.4 click==8.1.0 isort==5.10.1 # Click issue fixed in https://github.com/psf/black/pull/2966 - name: Run flake8 @@ -67,3 +67,9 @@ jobs: working-directory: ${{github.workspace}} run: | black --check --diff . + + - name: Run isort + shell: bash + working-directory: ${{github.workspace}} + run: | + isort --check --diff . diff --git a/.pre-commit-config.yaml b/.pre-commit-config.yaml index 1bb38f6ba..5cb213327 100644 --- a/.pre-commit-config.yaml +++ b/.pre-commit-config.yaml @@ -26,7 +26,7 @@ repos: # E121,E123,E126,E226,E24,E704,W503,W504 - repo: https://github.com/pycqa/isort - rev: 5.11.5 + rev: 5.10.1 hooks: - id: isort args: ["--profile=black"] diff --git a/docker/torch1.12.1-cuda11.3.dockerfile b/docker/torch1.12.1-cuda11.3.dockerfile index cb885e59e..33ecbf4d1 100644 --- a/docker/torch1.12.1-cuda11.3.dockerfile +++ b/docker/torch1.12.1-cuda11.3.dockerfile @@ -5,8 +5,8 @@ ENV LC_ALL C.UTF-8 ARG DEBIAN_FRONTEND=noninteractive # python 3.7 -ARG K2_VERSION="1.24.4.dev20240211+cuda11.3.torch1.12.1" -ARG KALDIFEAT_VERSION="1.25.4.dev20240210+cuda11.3.torch1.12.1" +ARG K2_VERSION="1.24.4.dev20240223+cuda11.3.torch1.12.1" +ARG KALDIFEAT_VERSION="1.25.4.dev20240223+cuda11.3.torch1.12.1" ARG TORCHAUDIO_VERSION="0.12.1+cu113" LABEL authors="Fangjun Kuang " diff --git a/docker/torch1.13.0-cuda11.6.dockerfile b/docker/torch1.13.0-cuda11.6.dockerfile index e238d87aa..b4d62b0bc 100644 --- a/docker/torch1.13.0-cuda11.6.dockerfile +++ b/docker/torch1.13.0-cuda11.6.dockerfile @@ -5,8 +5,8 @@ ENV LC_ALL C.UTF-8 ARG DEBIAN_FRONTEND=noninteractive # python 3.9 -ARG K2_VERSION="1.24.4.dev20240211+cuda11.6.torch1.13.0" -ARG KALDIFEAT_VERSION="1.25.4.dev20240210+cuda11.6.torch1.13.0" +ARG K2_VERSION="1.24.4.dev20240223+cuda11.6.torch1.13.0" +ARG KALDIFEAT_VERSION="1.25.4.dev20240223+cuda11.6.torch1.13.0" ARG TORCHAUDIO_VERSION="0.13.0+cu116" LABEL authors="Fangjun Kuang " diff --git a/docker/torch1.9.0-cuda10.2.dockerfile b/docker/torch1.9.0-cuda10.2.dockerfile index 26d45cafc..4d2d3058a 100644 --- a/docker/torch1.9.0-cuda10.2.dockerfile +++ b/docker/torch1.9.0-cuda10.2.dockerfile @@ -5,8 +5,8 @@ ENV LC_ALL C.UTF-8 ARG DEBIAN_FRONTEND=noninteractive # python 3.7 -ARG K2_VERSION="1.24.4.dev20240211+cuda10.2.torch1.9.0" -ARG KALDIFEAT_VERSION="1.25.4.dev20240210+cuda10.2.torch1.9.0" +ARG K2_VERSION="1.24.4.dev20240223+cuda10.2.torch1.9.0" +ARG KALDIFEAT_VERSION="1.25.4.dev20240223+cuda10.2.torch1.9.0" ARG TORCHAUDIO_VERSION="0.9.0" LABEL authors="Fangjun Kuang " diff --git a/docker/torch2.0.0-cuda11.7.dockerfile b/docker/torch2.0.0-cuda11.7.dockerfile index 02906e53b..ad23f8be7 100644 --- a/docker/torch2.0.0-cuda11.7.dockerfile +++ b/docker/torch2.0.0-cuda11.7.dockerfile @@ -5,8 +5,8 @@ ENV LC_ALL C.UTF-8 ARG DEBIAN_FRONTEND=noninteractive # python 3.10 -ARG K2_VERSION="1.24.4.dev20240211+cuda11.7.torch2.0.0" -ARG KALDIFEAT_VERSION="1.25.4.dev20240210+cuda11.7.torch2.0.0" +ARG K2_VERSION="1.24.4.dev20240223+cuda11.7.torch2.0.0" +ARG KALDIFEAT_VERSION="1.25.4.dev20240223+cuda11.7.torch2.0.0" ARG TORCHAUDIO_VERSION="2.0.0+cu117" LABEL authors="Fangjun Kuang " diff --git a/docker/torch2.1.0-cuda11.8.dockerfile b/docker/torch2.1.0-cuda11.8.dockerfile index c87305922..4e6812b83 100644 --- a/docker/torch2.1.0-cuda11.8.dockerfile +++ b/docker/torch2.1.0-cuda11.8.dockerfile @@ -5,8 +5,8 @@ ENV LC_ALL C.UTF-8 ARG DEBIAN_FRONTEND=noninteractive # python 3.10 -ARG K2_VERSION="1.24.4.dev20240211+cuda11.8.torch2.1.0" -ARG KALDIFEAT_VERSION="1.25.4.dev20240210+cuda11.8.torch2.1.0" +ARG K2_VERSION="1.24.4.dev20240223+cuda11.8.torch2.1.0" +ARG KALDIFEAT_VERSION="1.25.4.dev20240223+cuda11.8.torch2.1.0" ARG TORCHAUDIO_VERSION="2.1.0+cu118" LABEL authors="Fangjun Kuang " diff --git a/docker/torch2.1.0-cuda12.1.dockerfile b/docker/torch2.1.0-cuda12.1.dockerfile index f4c297678..c7de4cf28 100644 --- a/docker/torch2.1.0-cuda12.1.dockerfile +++ b/docker/torch2.1.0-cuda12.1.dockerfile @@ -5,8 +5,8 @@ ENV LC_ALL C.UTF-8 ARG DEBIAN_FRONTEND=noninteractive # python 3.10 -ARG K2_VERSION="1.24.4.dev20240211+cuda12.1.torch2.1.0" -ARG KALDIFEAT_VERSION="1.25.4.dev20240210+cuda12.1.torch2.1.0" +ARG K2_VERSION="1.24.4.dev20240223+cuda12.1.torch2.1.0" +ARG KALDIFEAT_VERSION="1.25.4.dev20240223+cuda12.1.torch2.1.0" ARG TORCHAUDIO_VERSION="2.1.0+cu121" LABEL authors="Fangjun Kuang " diff --git a/docker/torch2.2.0-cuda11.8.dockerfile b/docker/torch2.2.0-cuda11.8.dockerfile index c59661c27..0104ae870 100644 --- a/docker/torch2.2.0-cuda11.8.dockerfile +++ b/docker/torch2.2.0-cuda11.8.dockerfile @@ -5,8 +5,8 @@ ENV LC_ALL C.UTF-8 ARG DEBIAN_FRONTEND=noninteractive # python 3.10 -ARG K2_VERSION="1.24.4.dev20240211+cuda11.8.torch2.2.0" -ARG KALDIFEAT_VERSION="1.25.4.dev20240210+cuda11.8.torch2.2.0" +ARG K2_VERSION="1.24.4.dev20240223+cuda11.8.torch2.2.0" +ARG KALDIFEAT_VERSION="1.25.4.dev20240223+cuda11.8.torch2.2.0" ARG TORCHAUDIO_VERSION="2.2.0+cu118" LABEL authors="Fangjun Kuang " diff --git a/docker/torch2.2.0-cuda12.1.dockerfile b/docker/torch2.2.0-cuda12.1.dockerfile index 2c484efd5..ccd5265b2 100644 --- a/docker/torch2.2.0-cuda12.1.dockerfile +++ b/docker/torch2.2.0-cuda12.1.dockerfile @@ -5,8 +5,8 @@ ENV LC_ALL C.UTF-8 ARG DEBIAN_FRONTEND=noninteractive # python 3.10 -ARG K2_VERSION="1.24.4.dev20240211+cuda12.1.torch2.2.0" -ARG KALDIFEAT_VERSION="1.25.4.dev20240210+cuda12.1.torch2.2.0" +ARG K2_VERSION="1.24.4.dev20240223+cuda12.1.torch2.2.0" +ARG KALDIFEAT_VERSION="1.25.4.dev20240223+cuda12.1.torch2.2.0" ARG TORCHAUDIO_VERSION="2.2.0+cu121" LABEL authors="Fangjun Kuang " diff --git a/docker/torch2.2.1-cuda11.8.dockerfile b/docker/torch2.2.1-cuda11.8.dockerfile new file mode 100644 index 000000000..0528ba72f --- /dev/null +++ b/docker/torch2.2.1-cuda11.8.dockerfile @@ -0,0 +1,70 @@ +FROM pytorch/pytorch:2.2.1-cuda11.8-cudnn8-devel + +ENV LC_ALL C.UTF-8 + +ARG DEBIAN_FRONTEND=noninteractive + +# python 3.10 +ARG K2_VERSION="1.24.4.dev20240223+cuda11.8.torch2.2.1" +ARG KALDIFEAT_VERSION="1.25.4.dev20240223+cuda11.8.torch2.2.1" +ARG TORCHAUDIO_VERSION="2.2.1+cu118" + +LABEL authors="Fangjun Kuang " +LABEL k2_version=${K2_VERSION} +LABEL kaldifeat_version=${KALDIFEAT_VERSION} +LABEL github_repo="https://github.com/k2-fsa/icefall" + +RUN apt-get update && \ + apt-get install -y --no-install-recommends \ + curl \ + vim \ + libssl-dev \ + autoconf \ + automake \ + bzip2 \ + ca-certificates \ + ffmpeg \ + g++ \ + gfortran \ + git \ + libtool \ + make \ + patch \ + sox \ + subversion \ + unzip \ + valgrind \ + wget \ + zlib1g-dev \ + && rm -rf /var/lib/apt/lists/* + +# Install dependencies +RUN pip install --no-cache-dir \ + torchaudio==${TORCHAUDIO_VERSION} -f https://download.pytorch.org/whl/torch_stable.html \ + k2==${K2_VERSION} -f https://k2-fsa.github.io/k2/cuda.html \ + git+https://github.com/lhotse-speech/lhotse \ + kaldifeat==${KALDIFEAT_VERSION} -f https://csukuangfj.github.io/kaldifeat/cuda.html \ + kaldi_native_io \ + kaldialign \ + kaldifst \ + kaldilm \ + sentencepiece>=0.1.96 \ + tensorboard \ + typeguard \ + dill \ + onnx \ + onnxruntime \ + onnxmltools \ + multi_quantization \ + typeguard \ + numpy \ + pytest \ + graphviz + +RUN git clone https://github.com/k2-fsa/icefall /workspace/icefall && \ + cd /workspace/icefall && \ + pip install --no-cache-dir -r requirements.txt + +ENV PYTHONPATH /workspace/icefall:$PYTHONPATH + +WORKDIR /workspace/icefall diff --git a/docker/torch2.2.1-cuda12.1.dockerfile b/docker/torch2.2.1-cuda12.1.dockerfile new file mode 100644 index 000000000..3cdbb16ec --- /dev/null +++ b/docker/torch2.2.1-cuda12.1.dockerfile @@ -0,0 +1,70 @@ +FROM pytorch/pytorch:2.2.1-cuda12.1-cudnn8-devel + +ENV LC_ALL C.UTF-8 + +ARG DEBIAN_FRONTEND=noninteractive + +# python 3.10 +ARG K2_VERSION="1.24.4.dev20240223+cuda12.1.torch2.2.1" +ARG KALDIFEAT_VERSION="1.25.4.dev20240223+cuda12.1.torch2.2.1" +ARG TORCHAUDIO_VERSION="2.2.1+cu121" + +LABEL authors="Fangjun Kuang " +LABEL k2_version=${K2_VERSION} +LABEL kaldifeat_version=${KALDIFEAT_VERSION} +LABEL github_repo="https://github.com/k2-fsa/icefall" + +RUN apt-get update && \ + apt-get install -y --no-install-recommends \ + curl \ + vim \ + libssl-dev \ + autoconf \ + automake \ + bzip2 \ + ca-certificates \ + ffmpeg \ + g++ \ + gfortran \ + git \ + libtool \ + make \ + patch \ + sox \ + subversion \ + unzip \ + valgrind \ + wget \ + zlib1g-dev \ + && rm -rf /var/lib/apt/lists/* + +# Install dependencies +RUN pip install --no-cache-dir \ + torchaudio==${TORCHAUDIO_VERSION} -f https://download.pytorch.org/whl/torch_stable.html \ + k2==${K2_VERSION} -f https://k2-fsa.github.io/k2/cuda.html \ + git+https://github.com/lhotse-speech/lhotse \ + kaldifeat==${KALDIFEAT_VERSION} -f https://csukuangfj.github.io/kaldifeat/cuda.html \ + kaldi_native_io \ + kaldialign \ + kaldifst \ + kaldilm \ + sentencepiece>=0.1.96 \ + tensorboard \ + typeguard \ + dill \ + onnx \ + onnxruntime \ + onnxmltools \ + multi_quantization \ + typeguard \ + numpy \ + pytest \ + graphviz + +RUN git clone https://github.com/k2-fsa/icefall /workspace/icefall && \ + cd /workspace/icefall && \ + pip install --no-cache-dir -r requirements.txt + +ENV PYTHONPATH /workspace/icefall:$PYTHONPATH + +WORKDIR /workspace/icefall diff --git a/docs/source/docker/intro.rst b/docs/source/docker/intro.rst index 149970eff..1acaa3d4f 100644 --- a/docs/source/docker/intro.rst +++ b/docs/source/docker/intro.rst @@ -34,6 +34,8 @@ which will give you something like below: .. code-block:: bash + "torch2.2.1-cuda12.1" + "torch2.2.1-cuda11.8" "torch2.2.0-cuda12.1" "torch2.2.0-cuda11.8" "torch2.1.0-cuda12.1" diff --git a/docs/source/recipes/Finetune/adapter/finetune_adapter.rst b/docs/source/recipes/Finetune/adapter/finetune_adapter.rst new file mode 100644 index 000000000..a94b008f6 --- /dev/null +++ b/docs/source/recipes/Finetune/adapter/finetune_adapter.rst @@ -0,0 +1,225 @@ +Finetune from a pre-trained Zipformer model with adapters +========================================================= + +This tutorial shows you how to fine-tune a pre-trained **Zipformer** +transducer model on a new dataset with adapters. +Adapters are compact and efficient module that can be integrated into a pre-trained model +to improve the model's performance on a new domain. Adapters are injected +between different modules in the well-trained neural network. During training, only the parameters +in the adapters will be updated. It achieves competitive performance +while requiring much less GPU memory than full fine-tuning. For more details about adapters, +please refer to the original `paper `_ for more details. + +.. HINT:: + + We assume you have read the page :ref:`install icefall` and have setup + the environment for ``icefall``. + +.. HINT:: + + We recommend you to use a GPU or several GPUs to run this recipe + +For illustration purpose, we fine-tune the Zipformer transducer model +pre-trained on `LibriSpeech`_ on the small subset of `GigaSpeech`_. You could use your +own data for fine-tuning if you create a manifest for your new dataset. + +Data preparation +---------------- + +Please follow the instructions in the `GigaSpeech recipe `_ +to prepare the fine-tune data used in this tutorial. We only require the small subset in GigaSpeech for this tutorial. + + +Model preparation +----------------- + +We are using the Zipformer model trained on full LibriSpeech (960 hours) as the intialization. The +checkpoint of the model can be downloaded via the following command: + +.. code-block:: bash + + $ GIT_LFS_SKIP_SMUDGE=1 git clone https://huggingface.co/Zengwei/icefall-asr-librispeech-zipformer-2023-05-15 + $ cd icefall-asr-librispeech-zipformer-2023-05-15/exp + $ git lfs pull --include "pretrained.pt" + $ ln -s pretrained.pt epoch-99.pt + $ cd ../data/lang_bpe_500 + $ git lfs pull --include bpe.model + $ cd ../../.. + +Before fine-tuning, let's test the model's WER on the new domain. The following command performs +decoding on the GigaSpeech test sets: + +.. code-block:: bash + + ./zipformer/decode_gigaspeech.py \ + --epoch 99 \ + --avg 1 \ + --exp-dir icefall-asr-librispeech-zipformer-2023-05-15/exp \ + --use-averaged-model 0 \ + --max-duration 1000 \ + --decoding-method greedy_search + +You should see the following numbers: + +.. code-block:: + + For dev, WER of different settings are: + greedy_search 20.06 best for dev + + For test, WER of different settings are: + greedy_search 19.27 best for test + + +Fine-tune with adapter +---------------------- + +We insert 4 adapters with residual connection in each ``Zipformer2EncoderLayer``. +The original model parameters remain untouched during training and only the parameters of +the adapters are updated. The following command starts a fine-tuning experiment with adapters: + +.. code-block:: bash + + $ do_finetune=1 + $ use_adapters=1 + $ adapter_dim=8 + + $ ./zipformer_adapter/train.py \ + --world-size 2 \ + --num-epochs 20 \ + --start-epoch 1 \ + --exp-dir zipformer_adapter/exp_giga_finetune_adapters${use_adapters}_adapter_dim${adapter_dim} \ + --use-fp16 1 \ + --base-lr 0.045 \ + --use-adapters $use_adapters --adapter-dim $adapter_dim \ + --bpe-model data/lang_bpe_500/bpe.model \ + --do-finetune $do_finetune \ + --master-port 13022 \ + --finetune-ckpt icefall-asr-librispeech-zipformer-2023-05-15/exp/pretrained.pt \ + --max-duration 1000 + +The following arguments are related to fine-tuning: + +- ``--do-finetune`` + If True, do fine-tuning by initializing the model from a pre-trained checkpoint. + **Note that if you want to resume your fine-tuning experiment from certain epochs, you + need to set this to False.** + +- ``use-adapters`` + If adapters are used during fine-tuning. + +- ``--adapter-dim`` + The bottleneck dimension of the adapter module. Typically a small number. + +You should notice that in the training log, the total number of trainale parameters is shown: + +.. code-block:: + + 2024-02-22 21:22:03,808 INFO [train.py:1277] A total of 761344 trainable parameters (1.148% of the whole model) + +The trainable parameters only makes up 1.15% of the entire model parameters, so the training will be much faster +and requires less memory than full fine-tuning. + + +Decoding +-------- + +After training, let's test the WERs. To test the WERs on the GigaSpeech set, +you can execute the following command: + +.. code-block:: bash + + $ epoch=20 + $ avg=10 + $ use_adapters=1 + $ adapter_dim=8 + + % ./zipformer/decode.py \ + --epoch $epoch \ + --avg $avg \ + --use-averaged-model 1 \ + --exp-dir zipformer_adapter/exp_giga_finetune_adapters${use_adapters}_adapter_dim${adapter_dim} \ + --max-duration 600 \ + --use-adapters $use_adapters \ + --adapter-dim $adapter_dim \ + --decoding-method greedy_search + +You should see the following numbers: + +.. code-block:: + + For dev, WER of different settings are: + greedy_search 15.44 best for dev + + For test, WER of different settings are: + greedy_search 15.42 best for test + + +The WER on test set is improved from 19.27 to 15.42, demonstrating the effectiveness of adapters. + +The same model can be used to perform decoding on LibriSpeech test sets. You can deactivate the adapters +to keep the same performance of the original model: + +.. code-block:: bash + + $ epoch=20 + $ avg=1 + $ use_adapters=0 + $ adapter_dim=8 + + % ./zipformer/decode.py \ + --epoch $epoch \ + --avg $avg \ + --use-averaged-model 1 \ + --exp-dir zipformer_adapter/exp_giga_finetune_adapters${use_adapters}_adapter_dim${adapter_dim} \ + --max-duration 600 \ + --use-adapters $use_adapters \ + --adapter-dim $adapter_dim \ + --decoding-method greedy_search + + +.. code-block:: + + For dev, WER of different settings are: + greedy_search 2.23 best for test-clean + + For test, WER of different settings are: + greedy_search 4.96 best for test-other + +The numbers are the same as reported in `icefall `_. So adapter-based +fine-tuning is also very flexible as the same model can be used for decoding on the original and target domain. + + +Export the model +---------------- + +After training, the model can be exported to ``onnx`` format easily using the following command: + +.. code-block:: bash + + $ use_adapters=1 + $ adapter_dim=16 + + $ ./zipformer_adapter/export-onnx.py \ + --tokens icefall-asr-librispeech-zipformer-2023-05-15/data/lang_bpe_500/tokens.txt \ + --use-averaged-model 1 \ + --epoch 20 \ + --avg 10 \ + --exp-dir zipformer_adapter/exp_giga_finetune_adapters${use_adapters}_adapter_dim${adapter_dim} \ + --use-adapters $use_adapters \ + --adapter-dim $adapter_dim \ + --num-encoder-layers "2,2,3,4,3,2" \ + --downsampling-factor "1,2,4,8,4,2" \ + --feedforward-dim "512,768,1024,1536,1024,768" \ + --num-heads "4,4,4,8,4,4" \ + --encoder-dim "192,256,384,512,384,256" \ + --query-head-dim 32 \ + --value-head-dim 12 \ + --pos-head-dim 4 \ + --pos-dim 48 \ + --encoder-unmasked-dim "192,192,256,256,256,192" \ + --cnn-module-kernel "31,31,15,15,15,31" \ + --decoder-dim 512 \ + --joiner-dim 512 \ + --causal False \ + --chunk-size "16,32,64,-1" \ + --left-context-frames "64,128,256,-1" \ No newline at end of file diff --git a/docs/source/recipes/Finetune/index.rst b/docs/source/recipes/Finetune/index.rst index e62b8980f..7f36d2687 100644 --- a/docs/source/recipes/Finetune/index.rst +++ b/docs/source/recipes/Finetune/index.rst @@ -13,3 +13,4 @@ data to improve the performance on new domains. :caption: Table of Contents from_supervised/finetune_zipformer + adapter/finetune_adapter diff --git a/docs/source/recipes/TTS/ljspeech/vits.rst b/docs/source/recipes/TTS/ljspeech/vits.rst index d08aa0f47..9499a3aea 100644 --- a/docs/source/recipes/TTS/ljspeech/vits.rst +++ b/docs/source/recipes/TTS/ljspeech/vits.rst @@ -1,11 +1,11 @@ -VITS +VITS-LJSpeech =============== This tutorial shows you how to train an VITS model with the `LJSpeech `_ dataset. .. note:: - + TTS related recipes require packages in ``requirements-tts.txt``. .. note:: @@ -13,6 +13,14 @@ with the `LJSpeech `_ dataset. The VITS paper: `Conditional Variational Autoencoder with Adversarial Learning for End-to-End Text-to-Speech `_ +Install extra dependencies +-------------------------- + +.. code-block:: bash + + pip install piper_phonemize -f https://k2-fsa.github.io/icefall/piper_phonemize.html + pip install numba espnet_tts_frontend + Data preparation ---------------- @@ -56,7 +64,8 @@ Training --start-epoch 1 \ --use-fp16 1 \ --exp-dir vits/exp \ - --tokens data/tokens.txt + --tokens data/tokens.txt \ + --model-type high \ --max-duration 500 .. note:: @@ -64,6 +73,11 @@ Training You can adjust the hyper-parameters to control the size of the VITS model and the training configurations. For more details, please run ``./vits/train.py --help``. +.. warning:: + + If you want a model that runs faster on CPU, please use ``--model-type low`` + or ``--model-type medium``. + .. note:: The training can take a long time (usually a couple of days). @@ -95,8 +109,8 @@ training part first. It will save the ground-truth and generated wavs to the dir Export models ------------- -Currently we only support ONNX model exporting. It will generate two files in the given ``exp-dir``: -``vits-epoch-*.onnx`` and ``vits-epoch-*.int8.onnx``. +Currently we only support ONNX model exporting. It will generate one file in the given ``exp-dir``: +``vits-epoch-*.onnx``. .. code-block:: bash @@ -120,4 +134,68 @@ Download pretrained models If you don't want to train from scratch, you can download the pretrained models by visiting the following link: - - ``_ + - ``--model-type=high``: ``_ + - ``--model-type=medium``: ``_ + - ``--model-type=low``: ``_ + +Usage in sherpa-onnx +-------------------- + +The following describes how to test the exported ONNX model in `sherpa-onnx`_. + +.. hint:: + + `sherpa-onnx`_ supports different programming languages, e.g., C++, C, Python, + Kotlin, Java, Swift, Go, C#, etc. It also supports Android and iOS. + + We only describe how to use pre-built binaries from `sherpa-onnx`_ below. + Please refer to ``_ + for more documentation. + +Install sherpa-onnx +^^^^^^^^^^^^^^^^^^^ + +.. code-block:: bash + + pip install sherpa-onnx + +To check that you have installed `sherpa-onnx`_ successfully, please run: + +.. code-block:: bash + + which sherpa-onnx-offline-tts + sherpa-onnx-offline-tts --help + +Download lexicon files +^^^^^^^^^^^^^^^^^^^^^^ + +.. code-block:: bash + + cd /tmp + wget https://github.com/k2-fsa/sherpa-onnx/releases/download/tts-models/espeak-ng-data.tar.bz2 + tar xf espeak-ng-data.tar.bz2 + +Run sherpa-onnx +^^^^^^^^^^^^^^^ + +.. code-block:: bash + + cd egs/ljspeech/TTS + + sherpa-onnx-offline-tts \ + --vits-model=vits/exp/vits-epoch-1000.onnx \ + --vits-tokens=data/tokens.txt \ + --vits-data-dir=/tmp/espeak-ng-data \ + --num-threads=1 \ + --output-filename=./high.wav \ + "Ask not what your country can do for you; ask what you can do for your country." + +.. hint:: + + You can also use ``sherpa-onnx-offline-tts-play`` to play the audio + as it is generating. + +You should get a file ``high.wav`` after running the above command. + +Congratulations! You have successfully trained and exported a text-to-speech +model and run it with `sherpa-onnx`_. diff --git a/docs/source/recipes/TTS/vctk/vits.rst b/docs/source/recipes/TTS/vctk/vits.rst index 34024a5ea..45ae9d9d2 100644 --- a/docs/source/recipes/TTS/vctk/vits.rst +++ b/docs/source/recipes/TTS/vctk/vits.rst @@ -1,11 +1,11 @@ -VITS +VITS-VCTK =============== This tutorial shows you how to train an VITS model with the `VCTK `_ dataset. .. note:: - + TTS related recipes require packages in ``requirements-tts.txt``. .. note:: diff --git a/egs/aishell/ASR/README.md b/egs/aishell/ASR/README.md index b54719162..d088072a7 100644 --- a/egs/aishell/ASR/README.md +++ b/egs/aishell/ASR/README.md @@ -19,7 +19,9 @@ The following table lists the differences among them. | `transducer_stateless_modified` | Conformer | Embedding + Conv1d | with modified transducer from `optimized_transducer` | | `transducer_stateless_modified-2` | Conformer | Embedding + Conv1d | with modified transducer from `optimized_transducer` + extra data | | `pruned_transducer_stateless3` | Conformer (reworked) | Embedding + Conv1d | pruned RNN-T + reworked model with random combiner + using aidatatang_20zh as extra data| -| `pruned_transducer_stateless7` | Zipformer | Embedding | pruned RNN-T + zipformer encoder + stateless decoder with context-size 1 | +| `pruned_transducer_stateless7` | Zipformer | Embedding | pruned RNN-T + zipformer encoder + stateless decoder with context-size set to 1 | +| `zipformer` | Upgraded Zipformer | Embedding + Conv1d | The latest recipe with context-size set to 1 | + The decoder in `transducer_stateless` is modified from the paper [Rnn-Transducer with Stateless Prediction Network](https://ieeexplore.ieee.org/document/9054419/). diff --git a/egs/aishell/ASR/RESULTS.md b/egs/aishell/ASR/RESULTS.md index 46d712fb2..355d1516d 100644 --- a/egs/aishell/ASR/RESULTS.md +++ b/egs/aishell/ASR/RESULTS.md @@ -75,7 +75,7 @@ It's reworked Zipformer with Pruned RNNT loss, trained with Byte-level BPE, `voc | fast beam search | 4.43 | 4.17 | --epoch 40 --avg 10 | ```bash -./prepare.sh +./prepare.sh export CUDA_VISIBLE_DEVICES="0,1" diff --git a/egs/aishell/ASR/prepare.sh b/egs/aishell/ASR/prepare.sh index b7be89bc8..13be69534 100755 --- a/egs/aishell/ASR/prepare.sh +++ b/egs/aishell/ASR/prepare.sh @@ -360,7 +360,7 @@ if [ $stage -le 11 ] && [ $stop_stage -ge 11 ]; then fi if [ $stage -le 12 ] && [ $stop_stage -ge 12 ]; then - log "Stage 11: Train RNN LM model" + log "Stage 12: Train RNN LM model" python ../../../icefall/rnn_lm/train.py \ --start-epoch 0 \ --world-size 1 \ diff --git a/egs/aishell/ASR/pruned_transducer_stateless7/do_not_use_it_directly.py b/egs/aishell/ASR/pruned_transducer_stateless7/do_not_use_it_directly.py index 6027273b2..058d0ff6b 100755 --- a/egs/aishell/ASR/pruned_transducer_stateless7/do_not_use_it_directly.py +++ b/egs/aishell/ASR/pruned_transducer_stateless7/do_not_use_it_directly.py @@ -89,6 +89,7 @@ from icefall.checkpoint import ( ) from icefall.dist import cleanup_dist, setup_dist from icefall.env import get_env_info +from icefall.err import raise_grad_scale_is_too_small_error from icefall.lexicon import Lexicon from icefall.utils import ( AttributeDict, @@ -881,9 +882,7 @@ def train_one_epoch( if cur_grad_scale < 0.01: logging.warning(f"Grad scale is small: {cur_grad_scale}") if cur_grad_scale < 1.0e-05: - raise RuntimeError( - f"grad_scale is too small, exiting: {cur_grad_scale}" - ) + raise_grad_scale_is_too_small_error() if batch_idx % params.log_interval == 0: cur_lr = scheduler.get_last_lr()[0] cur_grad_scale = scaler._scale.item() if params.use_fp16 else 1.0 diff --git a/egs/aishell/ASR/pruned_transducer_stateless7/train.py b/egs/aishell/ASR/pruned_transducer_stateless7/train.py index 9d9dd4288..2dc835f3b 100755 --- a/egs/aishell/ASR/pruned_transducer_stateless7/train.py +++ b/egs/aishell/ASR/pruned_transducer_stateless7/train.py @@ -85,6 +85,7 @@ from icefall.checkpoint import ( ) from icefall.dist import cleanup_dist, setup_dist from icefall.env import get_env_info +from icefall.err import raise_grad_scale_is_too_small_error from icefall.hooks import register_inf_check_hooks from icefall.lexicon import Lexicon from icefall.utils import ( @@ -878,9 +879,7 @@ def train_one_epoch( if cur_grad_scale < 0.01: logging.warning(f"Grad scale is small: {cur_grad_scale}") if cur_grad_scale < 1.0e-05: - raise RuntimeError( - f"grad_scale is too small, exiting: {cur_grad_scale}" - ) + raise_grad_scale_is_too_small_error(cur_grad_scale) if batch_idx % params.log_interval == 0: cur_lr = scheduler.get_last_lr()[0] cur_grad_scale = scaler._scale.item() if params.use_fp16 else 1.0 diff --git a/egs/aishell/ASR/pruned_transducer_stateless7_bbpe/train.py b/egs/aishell/ASR/pruned_transducer_stateless7_bbpe/train.py index 3858bafd7..811269989 100755 --- a/egs/aishell/ASR/pruned_transducer_stateless7_bbpe/train.py +++ b/egs/aishell/ASR/pruned_transducer_stateless7_bbpe/train.py @@ -78,6 +78,7 @@ from icefall.checkpoint import ( ) from icefall.dist import cleanup_dist, setup_dist from icefall.env import get_env_info +from icefall.err import raise_grad_scale_is_too_small_error from icefall.hooks import register_inf_check_hooks from icefall.utils import ( AttributeDict, @@ -871,9 +872,7 @@ def train_one_epoch( if cur_grad_scale < 0.01: logging.warning(f"Grad scale is small: {cur_grad_scale}") if cur_grad_scale < 1.0e-05: - raise RuntimeError( - f"grad_scale is too small, exiting: {cur_grad_scale}" - ) + raise_grad_scale_is_too_small_error(cur_grad_scale) if batch_idx % params.log_interval == 0: cur_lr = scheduler.get_last_lr()[0] diff --git a/egs/aishell/ASR/pruned_transducer_stateless7_streaming/decode.py b/egs/aishell/ASR/pruned_transducer_stateless7_streaming/decode.py index 99110d6b6..0e783e92b 100755 --- a/egs/aishell/ASR/pruned_transducer_stateless7_streaming/decode.py +++ b/egs/aishell/ASR/pruned_transducer_stateless7_streaming/decode.py @@ -250,7 +250,7 @@ def get_parser(): parser.add_argument( "--context-size", type=int, - default=1, + default=2, help="The context size in the decoder. 1 means bigram; 2 means tri-gram", ) parser.add_argument( diff --git a/egs/aishell/ASR/pruned_transducer_stateless7_streaming/do_not_use_it_directly.py b/egs/aishell/ASR/pruned_transducer_stateless7_streaming/do_not_use_it_directly.py index 0fba3b58f..6653d9d9c 100755 --- a/egs/aishell/ASR/pruned_transducer_stateless7_streaming/do_not_use_it_directly.py +++ b/egs/aishell/ASR/pruned_transducer_stateless7_streaming/do_not_use_it_directly.py @@ -78,6 +78,7 @@ from icefall.checkpoint import ( ) from icefall.dist import cleanup_dist, setup_dist from icefall.env import get_env_info +from icefall.err import raise_grad_scale_is_too_small_error from icefall.hooks import register_inf_check_hooks from icefall.lexicon import Lexicon from icefall.utils import AttributeDict, MetricsTracker, setup_logger, str2bool @@ -882,9 +883,7 @@ def train_one_epoch( if cur_grad_scale < 0.01: logging.warning(f"Grad scale is small: {cur_grad_scale}") if cur_grad_scale < 1.0e-05: - raise RuntimeError( - f"grad_scale is too small, exiting: {cur_grad_scale}" - ) + raise_grad_scale_is_too_small_error(cur_grad_scale) if batch_idx % params.log_interval == 0: cur_lr = scheduler.get_last_lr()[0] diff --git a/egs/aishell/ASR/pruned_transducer_stateless7_streaming/train.py b/egs/aishell/ASR/pruned_transducer_stateless7_streaming/train.py index 2e1044658..f3b0f1e11 100755 --- a/egs/aishell/ASR/pruned_transducer_stateless7_streaming/train.py +++ b/egs/aishell/ASR/pruned_transducer_stateless7_streaming/train.py @@ -78,6 +78,7 @@ from icefall.checkpoint import ( ) from icefall.dist import cleanup_dist, setup_dist from icefall.env import get_env_info +from icefall.err import raise_grad_scale_is_too_small_error from icefall.hooks import register_inf_check_hooks from icefall.lexicon import Lexicon from icefall.utils import AttributeDict, MetricsTracker, setup_logger, str2bool @@ -881,9 +882,7 @@ def train_one_epoch( if cur_grad_scale < 0.01: logging.warning(f"Grad scale is small: {cur_grad_scale}") if cur_grad_scale < 1.0e-05: - raise RuntimeError( - f"grad_scale is too small, exiting: {cur_grad_scale}" - ) + raise_grad_scale_is_too_small_error(cur_grad_scale) if batch_idx % params.log_interval == 0: cur_lr = scheduler.get_last_lr()[0] diff --git a/egs/aishell/ASR/zipformer/train.py b/egs/aishell/ASR/zipformer/train.py index d381649e4..a25979226 100755 --- a/egs/aishell/ASR/zipformer/train.py +++ b/egs/aishell/ASR/zipformer/train.py @@ -86,6 +86,7 @@ from icefall.checkpoint import ( ) from icefall.dist import cleanup_dist, setup_dist from icefall.env import get_env_info +from icefall.err import raise_grad_scale_is_too_small_error from icefall.hooks import register_inf_check_hooks from icefall.lexicon import Lexicon from icefall.utils import ( @@ -985,9 +986,7 @@ def train_one_epoch( logging.warning(f"Grad scale is small: {cur_grad_scale}") if cur_grad_scale < 1.0e-05: save_bad_model() - raise RuntimeError( - f"grad_scale is too small, exiting: {cur_grad_scale}" - ) + raise_grad_scale_is_too_small_error(cur_grad_scale) if batch_idx % params.log_interval == 0: cur_lr = max(scheduler.get_last_lr()) diff --git a/egs/aishell/ASR/zipformer/train_bbpe.py b/egs/aishell/ASR/zipformer/train_bbpe.py index a2bf96b29..0713c5787 100755 --- a/egs/aishell/ASR/zipformer/train_bbpe.py +++ b/egs/aishell/ASR/zipformer/train_bbpe.py @@ -83,6 +83,7 @@ from icefall.checkpoint import ( update_averaged_model, ) from icefall.dist import cleanup_dist, setup_dist +from icefall.err import raise_grad_scale_is_too_small_error from icefall.hooks import register_inf_check_hooks from icefall.utils import ( AttributeDict, @@ -570,9 +571,7 @@ def train_one_epoch( logging.warning(f"Grad scale is small: {cur_grad_scale}") if cur_grad_scale < 1.0e-05: save_bad_model() - raise RuntimeError( - f"grad_scale is too small, exiting: {cur_grad_scale}" - ) + raise_grad_scale_is_too_small_error(cur_grad_scale) if batch_idx % params.log_interval == 0: cur_lr = max(scheduler.get_last_lr()) diff --git a/egs/aishell2/ASR/RESULTS.md b/egs/aishell2/ASR/RESULTS.md index 32ad74b50..0b7ae9299 100644 --- a/egs/aishell2/ASR/RESULTS.md +++ b/egs/aishell2/ASR/RESULTS.md @@ -1,6 +1,6 @@ ## Results -### Aishell2 char-based training results +### Aishell2 char-based training results #### Pruned transducer stateless 5 diff --git a/egs/aishell2/ASR/local/compute_fbank_aishell2.py b/egs/aishell2/ASR/local/compute_fbank_aishell2.py index 1fb1621ff..557f22b0c 100755 --- a/egs/aishell2/ASR/local/compute_fbank_aishell2.py +++ b/egs/aishell2/ASR/local/compute_fbank_aishell2.py @@ -29,7 +29,14 @@ import os from pathlib import Path import torch -from lhotse import CutSet, Fbank, FbankConfig, LilcomChunkyWriter +from lhotse import ( + CutSet, + Fbank, + FbankConfig, + LilcomChunkyWriter, + WhisperFbank, + WhisperFbankConfig, +) from lhotse.recipes.utils import read_manifests_if_cached from icefall.utils import get_executor, str2bool @@ -42,10 +49,12 @@ torch.set_num_threads(1) torch.set_num_interop_threads(1) -def compute_fbank_aishell2(num_mel_bins: int = 80, perturb_speed: bool = False): +def compute_fbank_aishell2( + num_mel_bins: int = 80, perturb_speed: bool = False, whisper_fbank: bool = False +): src_dir = Path("data/manifests") output_dir = Path("data/fbank") - num_jobs = min(15, os.cpu_count()) + num_jobs = min(8, os.cpu_count()) dataset_parts = ( "train", @@ -68,8 +77,12 @@ def compute_fbank_aishell2(num_mel_bins: int = 80, perturb_speed: bool = False): list(manifests.keys()), dataset_parts, ) - - extractor = Fbank(FbankConfig(num_mel_bins=num_mel_bins)) + if whisper_fbank: + extractor = WhisperFbank( + WhisperFbankConfig(num_filters=num_mel_bins, device="cuda") + ) + else: + extractor = Fbank(FbankConfig(num_mel_bins=num_mel_bins)) with get_executor() as ex: # Initialize the executor only once. for partition, m in manifests.items(): @@ -82,7 +95,7 @@ def compute_fbank_aishell2(num_mel_bins: int = 80, perturb_speed: bool = False): supervisions=m["supervisions"], ) if "train" in partition and perturb_speed: - logging.info(f"Doing speed perturb") + logging.info("Doing speed perturb") cut_set = ( cut_set + cut_set.perturb_speed(0.9) + cut_set.perturb_speed(1.1) ) @@ -111,7 +124,12 @@ def get_args(): default=False, help="Enable 0.9 and 1.1 speed perturbation for data augmentation. Default: False.", ) - + parser.add_argument( + "--whisper-fbank", + type=str2bool, + default=False, + help="Use WhisperFbank instead of Fbank. Default: False.", + ) return parser.parse_args() @@ -122,5 +140,7 @@ if __name__ == "__main__": args = get_args() compute_fbank_aishell2( - num_mel_bins=args.num_mel_bins, perturb_speed=args.perturb_speed + num_mel_bins=args.num_mel_bins, + perturb_speed=args.perturb_speed, + whisper_fbank=args.whisper_fbank, ) diff --git a/egs/aishell2/ASR/prepare.sh b/egs/aishell2/ASR/prepare.sh index a5eb9bd13..c959bd4d1 100755 --- a/egs/aishell2/ASR/prepare.sh +++ b/egs/aishell2/ASR/prepare.sh @@ -108,6 +108,16 @@ if [ $stage -le 3 ] && [ $stop_stage -ge 3 ]; then fi fi +whisper_mel_bins=80 +if [ $stage -le 30 ] && [ $stop_stage -ge 30 ]; then + log "Stage 30: Compute whisper fbank for aishell2" + if [ ! -f data/fbank/.aishell2.whisper.done ]; then + mkdir -p data/fbank + ./local/compute_fbank_aishell2.py --perturb-speed ${perturb_speed} --num-mel-bins ${whisper_mel_bins} --whisper-fbank true + touch data/fbank/.aishell2.whisper.done + fi +fi + if [ $stage -le 4 ] && [ $stop_stage -ge 4 ]; then log "Stage 4: Compute fbank for musan" if [ ! -f data/fbank/.msuan.done ]; then diff --git a/egs/aishell4/ASR/README.md b/egs/aishell4/ASR/README.md index 67fa17790..b96161762 100644 --- a/egs/aishell4/ASR/README.md +++ b/egs/aishell4/ASR/README.md @@ -3,7 +3,7 @@ This recipe contains some various ASR models trained with Aishell4 (including S, M and L three subsets). -The AISHELL-4 is a sizable real-recorded Mandarin speech dataset collected by 8-channel circular microphone array for speech processing in conference scenarios. The dataset consists of 211 recorded meeting sessions, each containing 4 to 8 speakers, with a total length of 120 hours. This dataset aims to bridge the advanced research on multi-speaker processing and the practical application scenario in three aspects. With real recorded meetings, AISHELL-4 provides realistic acoustics and rich natural speech characteristics in conversation such as short pause, speech overlap, quick speaker turn, noise, etc. Meanwhile, the accurate transcription and speaker voice activity are provided for each meeting in AISHELL-4. This allows the researchers to explore different aspects in meeting processing, ranging from individual tasks such as speech front-end processing, speech recognition and speaker diarization, to multi-modality modeling and joint optimization of relevant tasks. +The AISHELL-4 is a sizable real-recorded Mandarin speech dataset collected by 8-channel circular microphone array for speech processing in conference scenarios. The dataset consists of 211 recorded meeting sessions, each containing 4 to 8 speakers, with a total length of 120 hours. This dataset aims to bridge the advanced research on multi-speaker processing and the practical application scenario in three aspects. With real recorded meetings, AISHELL-4 provides realistic acoustics and rich natural speech characteristics in conversation such as short pause, speech overlap, quick speaker turn, noise, etc. Meanwhile, the accurate transcription and speaker voice activity are provided for each meeting in AISHELL-4. This allows the researchers to explore different aspects in meeting processing, ranging from individual tasks such as speech front-end processing, speech recognition and speaker diarization, to multi-modality modeling and joint optimization of relevant tasks. (From [Open Speech and Language Resources](https://www.openslr.org/111/)) diff --git a/egs/aishell4/ASR/local/compute_fbank_aishell4.py b/egs/aishell4/ASR/local/compute_fbank_aishell4.py index f19163988..b5f8468ac 100755 --- a/egs/aishell4/ASR/local/compute_fbank_aishell4.py +++ b/egs/aishell4/ASR/local/compute_fbank_aishell4.py @@ -29,7 +29,14 @@ import os from pathlib import Path import torch -from lhotse import ChunkedLilcomHdf5Writer, CutSet, Fbank, FbankConfig +from lhotse import ( + CutSet, + Fbank, + FbankConfig, + LilcomChunkyWriter, + WhisperFbank, + WhisperFbankConfig, +) from lhotse.recipes.utils import read_manifests_if_cached from icefall.utils import get_executor, str2bool @@ -42,10 +49,12 @@ torch.set_num_threads(1) torch.set_num_interop_threads(1) -def compute_fbank_aishell4(num_mel_bins: int = 80, perturb_speed: bool = False): +def compute_fbank_aishell4( + num_mel_bins: int = 80, perturb_speed: bool = False, whisper_fbank: bool = False +): src_dir = Path("data/manifests/aishell4") output_dir = Path("data/fbank") - num_jobs = min(15, os.cpu_count()) + num_jobs = min(8, os.cpu_count()) dataset_parts = ( "train_S", @@ -70,7 +79,12 @@ def compute_fbank_aishell4(num_mel_bins: int = 80, perturb_speed: bool = False): dataset_parts, ) - extractor = Fbank(FbankConfig(num_mel_bins=num_mel_bins)) + if whisper_fbank: + extractor = WhisperFbank( + WhisperFbankConfig(num_filters=num_mel_bins, device="cuda") + ) + else: + extractor = Fbank(FbankConfig(num_mel_bins=num_mel_bins)) with get_executor() as ex: # Initialize the executor only once. for partition, m in manifests.items(): @@ -84,7 +98,7 @@ def compute_fbank_aishell4(num_mel_bins: int = 80, perturb_speed: bool = False): supervisions=m["supervisions"], ) if "train" in partition and perturb_speed: - logging.info(f"Doing speed perturb") + logging.info("Doing speed perturb") cut_set = ( cut_set + cut_set.perturb_speed(0.9) + cut_set.perturb_speed(1.1) ) @@ -95,7 +109,7 @@ def compute_fbank_aishell4(num_mel_bins: int = 80, perturb_speed: bool = False): # when an executor is specified, make more partitions num_jobs=num_jobs if ex is None else 80, executor=ex, - storage_type=ChunkedLilcomHdf5Writer, + storage_type=LilcomChunkyWriter, ) logging.info("About splitting cuts into smaller chunks") @@ -121,7 +135,12 @@ def get_args(): default=False, help="Enable 0.9 and 1.1 speed perturbation for data augmentation. Default: False.", ) - + parser.add_argument( + "--whisper-fbank", + type=str2bool, + default=False, + help="Use WhisperFbank instead of Fbank. Default: False.", + ) return parser.parse_args() @@ -132,5 +151,7 @@ if __name__ == "__main__": args = get_args() compute_fbank_aishell4( - num_mel_bins=args.num_mel_bins, perturb_speed=args.perturb_speed + num_mel_bins=args.num_mel_bins, + perturb_speed=args.perturb_speed, + whisper_fbank=args.whisper_fbank, ) diff --git a/egs/aishell4/ASR/prepare.sh b/egs/aishell4/ASR/prepare.sh index e8d9eb7b9..38a36d97a 100755 --- a/egs/aishell4/ASR/prepare.sh +++ b/egs/aishell4/ASR/prepare.sh @@ -6,7 +6,7 @@ export PROTOCOL_BUFFERS_PYTHON_IMPLEMENTATION=python set -eou pipefail stage=-1 -stop_stage=100 +stop_stage=7 perturb_speed=true @@ -76,11 +76,21 @@ if [ $stage -le 1 ] && [ $stop_stage -ge 1 ]; then fi if [ $stage -le 2 ] && [ $stop_stage -ge 2 ]; then - log "Stage 2: Process aishell4" + log "Stage 2: Compute fbank for aishell4" if [ ! -f data/fbank/aishell4/.fbank.done ]; then - mkdir -p data/fbank/aishell4 + mkdir -p data/fbank ./local/compute_fbank_aishell4.py --perturb-speed ${perturb_speed} - touch data/fbank/aishell4/.fbank.done + touch data/fbank/.fbank.done + fi +fi + +whisper_mel_bins=80 +if [ $stage -le 20 ] && [ $stop_stage -ge 20 ]; then + log "Stage 20: Compute whisper fbank for aishell4" + if [ ! -f data/fbank/aishell4/.fbank.done ]; then + mkdir -p data/fbank + ./local/compute_fbank_aishell4.py --perturb-speed ${perturb_speed} --num-mel-bins ${whisper_mel_bins} --whisper-fbank true + touch data/fbank/.fbank.done fi fi @@ -106,16 +116,7 @@ if [ $stage -le 4 ] && [ $stop_stage -ge 4 ]; then fi if [ $stage -le 5 ] && [ $stop_stage -ge 5 ]; then - log "Stage 5: Compute fbank for aishell4" - if [ ! -f data/fbank/.aishell4.done ]; then - mkdir -p data/fbank - ./local/compute_fbank_aishell4.py --perturb-speed ${perturb_speed} - touch data/fbank/.aishell4.done - fi -fi - -if [ $stage -le 6 ] && [ $stop_stage -ge 6 ]; then - log "Stage 6: Prepare char based lang" + log "Stage 5: Prepare char based lang" lang_char_dir=data/lang_char mkdir -p $lang_char_dir diff --git a/egs/alimeeting/ASR/local/compute_fbank_alimeeting.py b/egs/alimeeting/ASR/local/compute_fbank_alimeeting.py index f8c10648a..09c873a34 100755 --- a/egs/alimeeting/ASR/local/compute_fbank_alimeeting.py +++ b/egs/alimeeting/ASR/local/compute_fbank_alimeeting.py @@ -29,7 +29,14 @@ import os from pathlib import Path import torch -from lhotse import CutSet, Fbank, FbankConfig, LilcomChunkyWriter +from lhotse import ( + CutSet, + Fbank, + FbankConfig, + LilcomChunkyWriter, + WhisperFbank, + WhisperFbankConfig, +) from lhotse.recipes.utils import read_manifests_if_cached from icefall.utils import get_executor, str2bool @@ -42,10 +49,12 @@ torch.set_num_threads(1) torch.set_num_interop_threads(1) -def compute_fbank_alimeeting(num_mel_bins: int = 80, perturb_speed: bool = False): +def compute_fbank_alimeeting( + num_mel_bins: int = 80, perturb_speed: bool = False, whisper_fbank: bool = False +): src_dir = Path("data/manifests/alimeeting") output_dir = Path("data/fbank") - num_jobs = min(15, os.cpu_count()) + num_jobs = min(8, os.cpu_count()) dataset_parts = ( "train", @@ -53,7 +62,7 @@ def compute_fbank_alimeeting(num_mel_bins: int = 80, perturb_speed: bool = False "test", ) - prefix = "alimeeting" + prefix = "alimeeting-far" suffix = "jsonl.gz" manifests = read_manifests_if_cached( dataset_parts=dataset_parts, @@ -70,7 +79,12 @@ def compute_fbank_alimeeting(num_mel_bins: int = 80, perturb_speed: bool = False dataset_parts, ) - extractor = Fbank(FbankConfig(num_mel_bins=num_mel_bins)) + if whisper_fbank: + extractor = WhisperFbank( + WhisperFbankConfig(num_filters=num_mel_bins, device="cuda") + ) + else: + extractor = Fbank(FbankConfig(num_mel_bins=num_mel_bins)) with get_executor() as ex: # Initialize the executor only once. for partition, m in manifests.items(): @@ -83,7 +97,7 @@ def compute_fbank_alimeeting(num_mel_bins: int = 80, perturb_speed: bool = False supervisions=m["supervisions"], ) if "train" in partition and perturb_speed: - logging.info(f"Doing speed perturb") + logging.info("Doing speed perturb") cut_set = ( cut_set + cut_set.perturb_speed(0.9) + cut_set.perturb_speed(1.1) ) @@ -121,7 +135,12 @@ def get_args(): default=False, help="Enable 0.9 and 1.1 speed perturbation for data augmentation. Default: False.", ) - + parser.add_argument( + "--whisper-fbank", + type=str2bool, + default=False, + help="Use the Whisper Fbank feature extractor. Default: False.", + ) return parser.parse_args() @@ -132,5 +151,7 @@ if __name__ == "__main__": args = get_args() compute_fbank_alimeeting( - num_mel_bins=args.num_mel_bins, perturb_speed=args.perturb_speed + num_mel_bins=args.num_mel_bins, + perturb_speed=args.perturb_speed, + whisper_fbank=args.whisper_fbank, ) diff --git a/egs/alimeeting/ASR/prepare.sh b/egs/alimeeting/ASR/prepare.sh index c8fed658d..996a1da2d 100755 --- a/egs/alimeeting/ASR/prepare.sh +++ b/egs/alimeeting/ASR/prepare.sh @@ -6,7 +6,7 @@ export PROTOCOL_BUFFERS_PYTHON_IMPLEMENTATION=python set -eou pipefail stage=-1 -stop_stage=100 +stop_stage=7 perturb_speed=true # We assume dl_dir (download dir) contains the following @@ -15,7 +15,7 @@ perturb_speed=true # # - $dl_dir/alimeeting # This directory contains the following files downloaded from -# https://openslr.org/62/ +# https://openslr.org/119/ # # - Train_Ali_far.tar.gz # - Train_Ali_near.tar.gz @@ -66,10 +66,21 @@ if [ $stage -le 1 ] && [ $stop_stage -ge 1 ]; then fi if [ $stage -le 2 ] && [ $stop_stage -ge 2 ]; then - log "Stage 2: Process alimeeting" - if [ ! -f data/fbank/alimeeting/.fbank.done ]; then - mkdir -p data/fbank/alimeeting + log "Stage 2: compute fbank for alimeeting" + if [ ! -f data/fbank/.fbank.done ]; then + mkdir -p data/fbank ./local/compute_fbank_alimeeting.py --perturb-speed ${perturb_speed} + touch data/fbank/.fbank.done + fi +fi + +whisper_mel_bins=80 +if [ $stage -le 20 ] && [ $stop_stage -ge 20 ]; then + log "Stage 20: compute whisper fbank for alimeeting" + if [ ! -f data/fbank/.fbank.done ]; then + mkdir -p data/fbank + ./local/compute_fbank_alimeeting.py --perturb-speed ${perturb_speed} --num-mel-bins ${whisper_mel_bins} --whisper-fbank true + touch data/fbank/.fbank.done fi fi @@ -95,16 +106,7 @@ if [ $stage -le 4 ] && [ $stop_stage -ge 4 ]; then fi if [ $stage -le 5 ] && [ $stop_stage -ge 5 ]; then - log "Stage 5: Compute fbank for alimeeting" - if [ ! -f data/fbank/.alimeeting.done ]; then - mkdir -p data/fbank - ./local/compute_fbank_alimeeting.py --perturb-speed True - touch data/fbank/.alimeeting.done - fi -fi - -if [ $stage -le 6 ] && [ $stop_stage -ge 6 ]; then - log "Stage 6: Prepare char based lang" + log "Stage 5: Prepare char based lang" lang_char_dir=data/lang_char mkdir -p $lang_char_dir diff --git a/egs/alimeeting/ASR_v2/prepare.sh b/egs/alimeeting/ASR_v2/prepare.sh index 1098840f8..15c20692d 100755 --- a/egs/alimeeting/ASR_v2/prepare.sh +++ b/egs/alimeeting/ASR_v2/prepare.sh @@ -12,7 +12,7 @@ use_gss=true # Use GSS-based enhancement with MDM setting # # - $dl_dir/alimeeting # This directory contains the following files downloaded from -# https://openslr.org/62/ +# https://openslr.org/119/ # # - Train_Ali_far.tar.gz # - Train_Ali_near.tar.gz diff --git a/egs/alimeeting/ASR_v2/pruned_transducer_stateless7/train.py b/egs/alimeeting/ASR_v2/pruned_transducer_stateless7/train.py index 8f09f1aa5..30879d8d2 100755 --- a/egs/alimeeting/ASR_v2/pruned_transducer_stateless7/train.py +++ b/egs/alimeeting/ASR_v2/pruned_transducer_stateless7/train.py @@ -70,6 +70,7 @@ from icefall.checkpoint import ( ) from icefall.dist import cleanup_dist, setup_dist from icefall.env import get_env_info +from icefall.err import raise_grad_scale_is_too_small_error from icefall.hooks import register_inf_check_hooks from icefall.lexicon import Lexicon from icefall.utils import AttributeDict, MetricsTracker, setup_logger, str2bool @@ -851,9 +852,7 @@ def train_one_epoch( if cur_grad_scale < 0.01: logging.warning(f"Grad scale is small: {cur_grad_scale}") if cur_grad_scale < 1.0e-05: - raise RuntimeError( - f"grad_scale is too small, exiting: {cur_grad_scale}" - ) + raise_grad_scale_is_too_small_error(cur_grad_scale) if batch_idx % params.log_interval == 0: cur_lr = scheduler.get_last_lr()[0] diff --git a/egs/ami/ASR/pruned_transducer_stateless7/train.py b/egs/ami/ASR/pruned_transducer_stateless7/train.py index 9b67141c0..d62cdadb7 100755 --- a/egs/ami/ASR/pruned_transducer_stateless7/train.py +++ b/egs/ami/ASR/pruned_transducer_stateless7/train.py @@ -69,6 +69,7 @@ from icefall.checkpoint import ( ) from icefall.dist import cleanup_dist, setup_dist from icefall.env import get_env_info +from icefall.err import raise_grad_scale_is_too_small_error from icefall.hooks import register_inf_check_hooks from icefall.utils import AttributeDict, MetricsTracker, setup_logger, str2bool @@ -842,9 +843,7 @@ def train_one_epoch( if cur_grad_scale < 0.01: logging.warning(f"Grad scale is small: {cur_grad_scale}") if cur_grad_scale < 1.0e-05: - raise RuntimeError( - f"grad_scale is too small, exiting: {cur_grad_scale}" - ) + raise_grad_scale_is_too_small_error(cur_grad_scale) if batch_idx % params.log_interval == 0: cur_lr = scheduler.get_last_lr()[0] diff --git a/egs/ami/SURT/dprnn_zipformer/train.py b/egs/ami/SURT/dprnn_zipformer/train.py index cd5fafc34..adc6a8495 100755 --- a/egs/ami/SURT/dprnn_zipformer/train.py +++ b/egs/ami/SURT/dprnn_zipformer/train.py @@ -75,6 +75,7 @@ from icefall.checkpoint import ( ) from icefall.dist import cleanup_dist, setup_dist from icefall.env import get_env_info +from icefall.err import raise_grad_scale_is_too_small_error from icefall.utils import AttributeDict, MetricsTracker, setup_logger, str2bool LRSchedulerType = Union[torch.optim.lr_scheduler._LRScheduler, optim.LRScheduler] @@ -1138,9 +1139,7 @@ def train_one_epoch( if cur_grad_scale < 0.01: logging.warning(f"Grad scale is small: {cur_grad_scale}") if cur_grad_scale < 1.0e-05: - raise RuntimeError( - f"grad_scale is too small, exiting: {cur_grad_scale}" - ) + raise_grad_scale_is_too_small_error(cur_grad_scale) if batch_idx % params.log_interval == 0: cur_lr = scheduler.get_last_lr()[0] diff --git a/egs/ami/SURT/dprnn_zipformer/train_adapt.py b/egs/ami/SURT/dprnn_zipformer/train_adapt.py index 9f3b4425f..ac5b0dadc 100755 --- a/egs/ami/SURT/dprnn_zipformer/train_adapt.py +++ b/egs/ami/SURT/dprnn_zipformer/train_adapt.py @@ -75,6 +75,7 @@ from icefall.checkpoint import ( ) from icefall.dist import cleanup_dist, setup_dist from icefall.env import get_env_info +from icefall.err import raise_grad_scale_is_too_small_error from icefall.utils import AttributeDict, MetricsTracker, setup_logger, str2bool LRSchedulerType = Union[torch.optim.lr_scheduler._LRScheduler, optim.LRScheduler] @@ -1129,9 +1130,7 @@ def train_one_epoch( if cur_grad_scale < 0.01: logging.warning(f"Grad scale is small: {cur_grad_scale}") if cur_grad_scale < 1.0e-05: - raise RuntimeError( - f"grad_scale is too small, exiting: {cur_grad_scale}" - ) + raise_grad_scale_is_too_small_error(cur_grad_scale) if batch_idx % params.log_interval == 0: cur_lr = scheduler.get_last_lr()[0] diff --git a/egs/commonvoice/ASR/local/compile_hlg.py b/egs/commonvoice/ASR/local/compile_hlg.py deleted file mode 120000 index 471aa7fb4..000000000 --- a/egs/commonvoice/ASR/local/compile_hlg.py +++ /dev/null @@ -1 +0,0 @@ -../../../librispeech/ASR/local/compile_hlg.py \ No newline at end of file diff --git a/egs/commonvoice/ASR/local/compile_hlg.py b/egs/commonvoice/ASR/local/compile_hlg.py new file mode 100755 index 000000000..6512aa68b --- /dev/null +++ b/egs/commonvoice/ASR/local/compile_hlg.py @@ -0,0 +1,168 @@ +#!/usr/bin/env python3 +# Copyright 2021-2024 Xiaomi Corp. (authors: Fangjun Kuang, +# Zengrui Jin,) +# +# 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 takes as input lang_dir and generates HLG from + + - H, the ctc topology, built from tokens contained in lang_dir/lexicon.txt + - L, the lexicon, built from lang_dir/L_disambig.pt + + Caution: We use a lexicon that contains disambiguation symbols + + - G, the LM, built from data/lm/G_n_gram.fst.txt + +The generated HLG is saved in $lang_dir/HLG.pt +""" +import argparse +import logging +from pathlib import Path + +import k2 +import torch + +from icefall.lexicon import Lexicon + + +def get_args(): + parser = argparse.ArgumentParser() + parser.add_argument( + "--lm", + type=str, + default="G_3_gram", + help="""Stem name for LM used in HLG compiling. + """, + ) + parser.add_argument( + "--lang-dir", + type=str, + help="""Input and output directory. + """, + ) + + return parser.parse_args() + + +def compile_HLG(lang_dir: str, lm: str = "G_3_gram") -> k2.Fsa: + """ + Args: + lang_dir: + The language directory, e.g., data/lang_phone or data/lang_bpe_5000. + lm: + The language stem base name. + + Return: + An FSA representing HLG. + """ + lexicon = Lexicon(lang_dir) + max_token_id = max(lexicon.tokens) + logging.info(f"Building ctc_topo. max_token_id: {max_token_id}") + H = k2.ctc_topo(max_token_id) + L = k2.Fsa.from_dict(torch.load(f"{lang_dir}/L_disambig.pt")) + + if Path(f"{lang_dir}/lm/{lm}.pt").is_file(): + logging.info(f"Loading pre-compiled {lm}") + d = torch.load(f"{lang_dir}/lm/{lm}.pt") + G = k2.Fsa.from_dict(d) + else: + logging.info(f"Loading {lm}.fst.txt") + with open(f"{lang_dir}/lm/{lm}.fst.txt") as f: + G = k2.Fsa.from_openfst(f.read(), acceptor=False) + torch.save(G.as_dict(), f"{lang_dir}/lm/{lm}.pt") + + first_token_disambig_id = lexicon.token_table["#0"] + first_word_disambig_id = lexicon.word_table["#0"] + + L = k2.arc_sort(L) + G = k2.arc_sort(G) + + logging.info("Intersecting L and G") + LG = k2.compose(L, G) + logging.info(f"LG shape: {LG.shape}") + + logging.info("Connecting LG") + LG = k2.connect(LG) + logging.info(f"LG shape after k2.connect: {LG.shape}") + + logging.info(type(LG.aux_labels)) + logging.info("Determinizing LG") + + LG = k2.determinize(LG) + logging.info(type(LG.aux_labels)) + + logging.info("Connecting LG after k2.determinize") + LG = k2.connect(LG) + + logging.info("Removing disambiguation symbols on LG") + + # LG.labels[LG.labels >= first_token_disambig_id] = 0 + # see https://github.com/k2-fsa/k2/pull/1140 + labels = LG.labels + labels[labels >= first_token_disambig_id] = 0 + LG.labels = labels + + assert isinstance(LG.aux_labels, k2.RaggedTensor) + LG.aux_labels.values[LG.aux_labels.values >= first_word_disambig_id] = 0 + + LG = k2.remove_epsilon(LG) + logging.info(f"LG shape after k2.remove_epsilon: {LG.shape}") + + LG = k2.connect(LG) + LG.aux_labels = LG.aux_labels.remove_values_eq(0) + + logging.info("Arc sorting LG") + LG = k2.arc_sort(LG) + + logging.info("Composing H and LG") + # CAUTION: The name of the inner_labels is fixed + # to `tokens`. If you want to change it, please + # also change other places in icefall that are using + # it. + HLG = k2.compose(H, LG, inner_labels="tokens") + + logging.info("Connecting LG") + HLG = k2.connect(HLG) + + logging.info("Arc sorting LG") + HLG = k2.arc_sort(HLG) + logging.info(f"HLG.shape: {HLG.shape}") + + return HLG + + +def main(): + args = get_args() + lang_dir = Path(args.lang_dir) + + if (lang_dir / "HLG.pt").is_file(): + logging.info(f"{lang_dir}/HLG.pt already exists - skipping") + return + + logging.info(f"Processing {lang_dir}") + + HLG = compile_HLG(lang_dir, args.lm) + logging.info(f"Saving HLG.pt to {lang_dir}") + torch.save(HLG.as_dict(), f"{lang_dir}/HLG.pt") + + +if __name__ == "__main__": + formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s" + + logging.basicConfig(format=formatter, level=logging.INFO) + + main() diff --git a/egs/commonvoice/ASR/local/compile_lg.py b/egs/commonvoice/ASR/local/compile_lg.py deleted file mode 120000 index 462d6d3fb..000000000 --- a/egs/commonvoice/ASR/local/compile_lg.py +++ /dev/null @@ -1 +0,0 @@ -../../../librispeech/ASR/local/compile_lg.py \ No newline at end of file diff --git a/egs/commonvoice/ASR/local/compile_lg.py b/egs/commonvoice/ASR/local/compile_lg.py new file mode 100755 index 000000000..76dacb5b2 --- /dev/null +++ b/egs/commonvoice/ASR/local/compile_lg.py @@ -0,0 +1,149 @@ +#!/usr/bin/env python3 +# Copyright 2021-2024 Xiaomi Corp. (authors: Fangjun Kuang, +# Kang Wei, +# Zengrui Jin,) +# +# 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 takes as input lang_dir and generates LG from + + - L, the lexicon, built from lang_dir/L_disambig.pt + + Caution: We use a lexicon that contains disambiguation symbols + + - G, the LM, built from lang_dir/lm/G_3_gram.fst.txt + +The generated LG is saved in $lang_dir/LG.pt +""" +import argparse +import logging +from pathlib import Path + +import k2 +import torch + +from icefall.lexicon import Lexicon + + +def get_args(): + parser = argparse.ArgumentParser() + parser.add_argument( + "--lang-dir", + type=str, + help="""Input and output directory. + """, + ) + parser.add_argument( + "--lm", + type=str, + default="G_3_gram", + help="""Stem name for LM used in HLG compiling. + """, + ) + + return parser.parse_args() + + +def compile_LG(lang_dir: str, lm: str = "G_3_gram") -> k2.Fsa: + """ + Args: + lang_dir: + The language directory, e.g., data/lang_phone or data/lang_bpe_5000. + + Return: + An FSA representing LG. + """ + lexicon = Lexicon(lang_dir) + L = k2.Fsa.from_dict(torch.load(f"{lang_dir}/L_disambig.pt")) + + if Path(f"{lang_dir}/lm/{lm}.pt").is_file(): + logging.info(f"Loading pre-compiled {lm}") + d = torch.load(f"{lang_dir}/lm/{lm}.pt") + G = k2.Fsa.from_dict(d) + else: + logging.info(f"Loading {lm}.fst.txt") + with open(f"{lang_dir}/lm/{lm}.fst.txt") as f: + G = k2.Fsa.from_openfst(f.read(), acceptor=False) + torch.save(G.as_dict(), f"{lang_dir}/lm/{lm}.pt") + + first_token_disambig_id = lexicon.token_table["#0"] + first_word_disambig_id = lexicon.word_table["#0"] + + L = k2.arc_sort(L) + G = k2.arc_sort(G) + + logging.info("Intersecting L and G") + LG = k2.compose(L, G) + logging.info(f"LG shape: {LG.shape}") + + logging.info("Connecting LG") + LG = k2.connect(LG) + logging.info(f"LG shape after k2.connect: {LG.shape}") + + logging.info(type(LG.aux_labels)) + logging.info("Determinizing LG") + + LG = k2.determinize(LG, k2.DeterminizeWeightPushingType.kLogWeightPushing) + logging.info(type(LG.aux_labels)) + + logging.info("Connecting LG after k2.determinize") + LG = k2.connect(LG) + + logging.info("Removing disambiguation symbols on LG") + + # LG.labels[LG.labels >= first_token_disambig_id] = 0 + # see https://github.com/k2-fsa/k2/pull/1140 + labels = LG.labels + labels[labels >= first_token_disambig_id] = 0 + LG.labels = labels + + assert isinstance(LG.aux_labels, k2.RaggedTensor) + LG.aux_labels.values[LG.aux_labels.values >= first_word_disambig_id] = 0 + + LG = k2.remove_epsilon(LG) + logging.info(f"LG shape after k2.remove_epsilon: {LG.shape}") + + LG = k2.connect(LG) + LG.aux_labels = LG.aux_labels.remove_values_eq(0) + + logging.info("Arc sorting LG") + LG = k2.arc_sort(LG) + + return LG + + +def main(): + args = get_args() + lang_dir = Path(args.lang_dir) + + if (lang_dir / "LG.pt").is_file(): + logging.info(f"{lang_dir}/LG.pt already exists - skipping") + return + + logging.info(f"Processing {lang_dir}") + + LG = compile_LG(lang_dir, args.lm) + logging.info(f"Saving LG.pt to {lang_dir}") + torch.save(LG.as_dict(), f"{lang_dir}/LG.pt") + + +if __name__ == "__main__": + formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s" + + logging.basicConfig(format=formatter, level=logging.INFO) + + main() diff --git a/egs/commonvoice/ASR/local/preprocess_commonvoice.py b/egs/commonvoice/ASR/local/preprocess_commonvoice.py index 5f6aa3ec0..dbacdd821 100755 --- a/egs/commonvoice/ASR/local/preprocess_commonvoice.py +++ b/egs/commonvoice/ASR/local/preprocess_commonvoice.py @@ -48,8 +48,27 @@ def normalize_text(utt: str, language: str) -> str: utt = re.sub("’", "'", utt) if language == "en": return re.sub(r"[^a-zA-Z\s]", "", utt).upper() - if language == "fr": + elif language == "fr": return re.sub(r"[^A-ZÀÂÆÇÉÈÊËÎÏÔŒÙÛÜ' ]", "", utt).upper() + elif language == "pl": + return re.sub(r"[^a-ząćęłńóśźżA-ZĄĆĘŁŃÓŚŹŻ' ]", "", utt).upper() + elif language == "yue": + return ( + utt.replace(" ", "") + .replace("，", "") + .replace("。", " ") + .replace("？", "") + .replace("！", "") + .replace("?", "") + ) + else: + raise NotImplementedError( + f""" + Text normalization not implemented for language: {language}, + please consider implementing it in the local/preprocess_commonvoice.py + or raise an issue on GitHub to request it. + """ + ) def preprocess_commonvoice( diff --git a/egs/commonvoice/ASR/pruned_transducer_stateless7/asr_datamodule.py b/egs/commonvoice/ASR/pruned_transducer_stateless7/asr_datamodule.py index c40d9419b..41009831c 100644 --- a/egs/commonvoice/ASR/pruned_transducer_stateless7/asr_datamodule.py +++ b/egs/commonvoice/ASR/pruned_transducer_stateless7/asr_datamodule.py @@ -381,9 +381,11 @@ class CommonVoiceAsrDataModule: def test_dataloaders(self, cuts: CutSet) -> DataLoader: logging.debug("About to create test dataset") test = K2SpeechRecognitionDataset( - input_strategy=OnTheFlyFeatures(Fbank(FbankConfig(num_mel_bins=80))) - if self.args.on_the_fly_feats - else eval(self.args.input_strategy)(), + input_strategy=( + OnTheFlyFeatures(Fbank(FbankConfig(num_mel_bins=80))) + if self.args.on_the_fly_feats + else eval(self.args.input_strategy)() + ), return_cuts=self.args.return_cuts, ) sampler = DynamicBucketingSampler( diff --git a/egs/commonvoice/ASR/pruned_transducer_stateless7/onnx_check.py b/egs/commonvoice/ASR/pruned_transducer_stateless7/onnx_check.py index 19c518eaf..f04537660 100755 --- a/egs/commonvoice/ASR/pruned_transducer_stateless7/onnx_check.py +++ b/egs/commonvoice/ASR/pruned_transducer_stateless7/onnx_check.py @@ -79,10 +79,10 @@ It will generate the following 3 files inside $repo/exp: import argparse import logging -from icefall import is_module_available +import torch from onnx_pretrained import OnnxModel -import torch +from icefall import is_module_available def get_parser(): diff --git a/egs/commonvoice/ASR/pruned_transducer_stateless7/train.py b/egs/commonvoice/ASR/pruned_transducer_stateless7/train.py index 4aedeffe4..4957c0c31 100755 --- a/egs/commonvoice/ASR/pruned_transducer_stateless7/train.py +++ b/egs/commonvoice/ASR/pruned_transducer_stateless7/train.py @@ -79,6 +79,7 @@ from icefall.checkpoint import ( ) from icefall.dist import cleanup_dist, setup_dist from icefall.env import get_env_info +from icefall.err import raise_grad_scale_is_too_small_error from icefall.hooks import register_inf_check_hooks from icefall.utils import ( AttributeDict, @@ -871,9 +872,7 @@ def train_one_epoch( if cur_grad_scale < 0.01: logging.warning(f"Grad scale is small: {cur_grad_scale}") if cur_grad_scale < 1.0e-05: - raise RuntimeError( - f"grad_scale is too small, exiting: {cur_grad_scale}" - ) + raise_grad_scale_is_too_small_error(cur_grad_scale) if batch_idx % params.log_interval == 0: cur_lr = scheduler.get_last_lr()[0] diff --git a/egs/commonvoice/ASR/pruned_transducer_stateless7_streaming/commonvoice_fr.py b/egs/commonvoice/ASR/pruned_transducer_stateless7_streaming/commonvoice_fr.py index 79cf86b84..91220bd11 100644 --- a/egs/commonvoice/ASR/pruned_transducer_stateless7_streaming/commonvoice_fr.py +++ b/egs/commonvoice/ASR/pruned_transducer_stateless7_streaming/commonvoice_fr.py @@ -31,7 +31,7 @@ from lhotse.dataset import ( # noqa F401 for PrecomputedFeatures DynamicBucketingSampler, K2SpeechRecognitionDataset, PrecomputedFeatures, - SingleCutSampler, + SimpleCutSampler, SpecAugment, ) from lhotse.dataset.input_strategies import ( # noqa F401 For AudioSamples @@ -232,7 +232,7 @@ class CommonVoiceAsrDataModule: logging.info("About to get Musan cuts") cuts_musan = load_manifest(self.args.manifest_dir / "musan_cuts.jsonl.gz") transforms.append( - CutMix(cuts=cuts_musan, prob=0.5, snr=(10, 20), preserve_id=True) + CutMix(cuts=cuts_musan, p=0.5, snr=(10, 20), preserve_id=True) ) else: logging.info("Disable MUSAN") @@ -315,8 +315,8 @@ class CommonVoiceAsrDataModule: drop_last=self.args.drop_last, ) else: - logging.info("Using SingleCutSampler.") - train_sampler = SingleCutSampler( + logging.info("Using SimpleCutSampler.") + train_sampler = SimpleCutSampler( cuts_train, max_duration=self.args.max_duration, shuffle=self.args.shuffle, @@ -383,9 +383,11 @@ class CommonVoiceAsrDataModule: def test_dataloaders(self, cuts: CutSet) -> DataLoader: logging.debug("About to create test dataset") test = K2SpeechRecognitionDataset( - input_strategy=OnTheFlyFeatures(Fbank(FbankConfig(num_mel_bins=80))) - if self.args.on_the_fly_feats - else eval(self.args.input_strategy)(), + input_strategy=( + OnTheFlyFeatures(Fbank(FbankConfig(num_mel_bins=80))) + if self.args.on_the_fly_feats + else eval(self.args.input_strategy)() + ), return_cuts=self.args.return_cuts, ) sampler = DynamicBucketingSampler( diff --git a/egs/commonvoice/ASR/pruned_transducer_stateless7_streaming/do_not_use_it_directly.py b/egs/commonvoice/ASR/pruned_transducer_stateless7_streaming/do_not_use_it_directly.py index 0426bc9a3..8e16cc0bf 100755 --- a/egs/commonvoice/ASR/pruned_transducer_stateless7_streaming/do_not_use_it_directly.py +++ b/egs/commonvoice/ASR/pruned_transducer_stateless7_streaming/do_not_use_it_directly.py @@ -889,9 +889,7 @@ def train_one_epoch( if cur_grad_scale < 0.01: logging.warning(f"Grad scale is small: {cur_grad_scale}") if cur_grad_scale < 1.0e-05: - raise RuntimeError( - f"grad_scale is too small, exiting: {cur_grad_scale}" - ) + raise RuntimeError(f", exiting: {cur_grad_scale}") if batch_idx % params.log_interval == 0: cur_lr = scheduler.get_last_lr()[0] diff --git a/egs/commonvoice/ASR/pruned_transducer_stateless7_streaming/finetune.py b/egs/commonvoice/ASR/pruned_transducer_stateless7_streaming/finetune.py index 3a10c5d81..47c4ed312 100755 --- a/egs/commonvoice/ASR/pruned_transducer_stateless7_streaming/finetune.py +++ b/egs/commonvoice/ASR/pruned_transducer_stateless7_streaming/finetune.py @@ -81,6 +81,7 @@ from icefall.checkpoint import ( ) from icefall.dist import cleanup_dist, setup_dist from icefall.env import get_env_info +from icefall.err import raise_grad_scale_is_too_small_error from icefall.hooks import register_inf_check_hooks from icefall.utils import ( AttributeDict, @@ -965,9 +966,7 @@ def train_one_epoch( if cur_grad_scale < 0.01: logging.warning(f"Grad scale is small: {cur_grad_scale}") if cur_grad_scale < 1.0e-05: - raise RuntimeError( - f"grad_scale is too small, exiting: {cur_grad_scale}" - ) + raise_grad_scale_is_too_small_error(cur_grad_scale) if batch_idx % params.log_interval == 0: cur_lr = scheduler.get_last_lr()[0] diff --git a/egs/commonvoice/ASR/pruned_transducer_stateless7_streaming/train.py b/egs/commonvoice/ASR/pruned_transducer_stateless7_streaming/train.py index a9bc9c2a2..dc5d0a858 100755 --- a/egs/commonvoice/ASR/pruned_transducer_stateless7_streaming/train.py +++ b/egs/commonvoice/ASR/pruned_transducer_stateless7_streaming/train.py @@ -78,6 +78,7 @@ from icefall.checkpoint import ( ) from icefall.dist import cleanup_dist, setup_dist from icefall.env import get_env_info +from icefall.err import raise_grad_scale_is_too_small_error from icefall.hooks import register_inf_check_hooks from icefall.utils import AttributeDict, MetricsTracker, setup_logger, str2bool @@ -888,9 +889,7 @@ def train_one_epoch( if cur_grad_scale < 0.01: logging.warning(f"Grad scale is small: {cur_grad_scale}") if cur_grad_scale < 1.0e-05: - raise RuntimeError( - f"grad_scale is too small, exiting: {cur_grad_scale}" - ) + raise_grad_scale_is_too_small_error(cur_grad_scale) if batch_idx % params.log_interval == 0: cur_lr = scheduler.get_last_lr()[0] diff --git a/egs/csj/ASR/pruned_transducer_stateless7_streaming/do_not_use_it_directly.py b/egs/csj/ASR/pruned_transducer_stateless7_streaming/do_not_use_it_directly.py index 685f6ece6..6d256308c 100755 --- a/egs/csj/ASR/pruned_transducer_stateless7_streaming/do_not_use_it_directly.py +++ b/egs/csj/ASR/pruned_transducer_stateless7_streaming/do_not_use_it_directly.py @@ -81,6 +81,7 @@ from icefall.checkpoint import ( ) from icefall.dist import cleanup_dist, setup_dist from icefall.env import get_env_info +from icefall.err import raise_grad_scale_is_too_small_error from icefall.hooks import register_inf_check_hooks from icefall.utils import AttributeDict, MetricsTracker, setup_logger, str2bool @@ -909,9 +910,7 @@ def train_one_epoch( if cur_grad_scale < 0.01: logging.warning(f"Grad scale is small: {cur_grad_scale}") if cur_grad_scale < 1.0e-05: - raise RuntimeError( - f"grad_scale is too small, exiting: {cur_grad_scale}" - ) + raise_grad_scale_is_too_small_error(cur_grad_scale) if batch_idx % params.log_interval == 0: cur_lr = scheduler.get_last_lr()[0] diff --git a/egs/csj/ASR/pruned_transducer_stateless7_streaming/export-for-ncnn.py b/egs/csj/ASR/pruned_transducer_stateless7_streaming/export-for-ncnn.py index b210430c6..06a0fa96b 100755 --- a/egs/csj/ASR/pruned_transducer_stateless7_streaming/export-for-ncnn.py +++ b/egs/csj/ASR/pruned_transducer_stateless7_streaming/export-for-ncnn.py @@ -70,9 +70,9 @@ import logging from pathlib import Path import torch +from do_not_use_it_directly import add_model_arguments, get_params, get_transducer_model from scaling_converter import convert_scaled_to_non_scaled from tokenizer import Tokenizer -from do_not_use_it_directly import add_model_arguments, get_params, get_transducer_model from icefall.checkpoint import ( average_checkpoints, diff --git a/egs/csj/ASR/pruned_transducer_stateless7_streaming/train.py b/egs/csj/ASR/pruned_transducer_stateless7_streaming/train.py index 73fcd67aa..ef7ea9013 100755 --- a/egs/csj/ASR/pruned_transducer_stateless7_streaming/train.py +++ b/egs/csj/ASR/pruned_transducer_stateless7_streaming/train.py @@ -81,6 +81,7 @@ from icefall.checkpoint import ( ) from icefall.dist import cleanup_dist, setup_dist from icefall.env import get_env_info +from icefall.err import raise_grad_scale_is_too_small_error from icefall.hooks import register_inf_check_hooks from icefall.utils import AttributeDict, MetricsTracker, setup_logger, str2bool @@ -908,9 +909,7 @@ def train_one_epoch( if cur_grad_scale < 0.01: logging.warning(f"Grad scale is small: {cur_grad_scale}") if cur_grad_scale < 1.0e-05: - raise RuntimeError( - f"grad_scale is too small, exiting: {cur_grad_scale}" - ) + raise_grad_scale_is_too_small_error(cur_grad_scale) if batch_idx % params.log_interval == 0: cur_lr = scheduler.get_last_lr()[0] diff --git a/egs/gigaspeech/ASR/local/preprocess_gigaspeech.py b/egs/gigaspeech/ASR/local/preprocess_gigaspeech.py index b6603f80d..a31685211 100755 --- a/egs/gigaspeech/ASR/local/preprocess_gigaspeech.py +++ b/egs/gigaspeech/ASR/local/preprocess_gigaspeech.py @@ -23,6 +23,7 @@ from pathlib import Path from lhotse import CutSet, SupervisionSegment from lhotse.recipes.utils import read_manifests_if_cached + from icefall.utils import str2bool # Similar text filtering and normalization procedure as in: diff --git a/egs/gigaspeech/ASR/pruned_transducer_stateless2/decode.py b/egs/gigaspeech/ASR/pruned_transducer_stateless2/decode.py index 72f74c968..ef430302d 100755 --- a/egs/gigaspeech/ASR/pruned_transducer_stateless2/decode.py +++ b/egs/gigaspeech/ASR/pruned_transducer_stateless2/decode.py @@ -76,6 +76,7 @@ from beam_search import ( ) from gigaspeech_scoring import asr_text_post_processing from train import get_params, get_transducer_model + from icefall.checkpoint import ( average_checkpoints, average_checkpoints_with_averaged_model, diff --git a/egs/gigaspeech/ASR/zipformer/ctc_decode.py b/egs/gigaspeech/ASR/zipformer/ctc_decode.py index aa51036d5..651f20cb6 100755 --- a/egs/gigaspeech/ASR/zipformer/ctc_decode.py +++ b/egs/gigaspeech/ASR/zipformer/ctc_decode.py @@ -88,7 +88,7 @@ import sentencepiece as spm import torch import torch.nn as nn from asr_datamodule import GigaSpeechAsrDataModule -from train import add_model_arguments, get_params, get_model +from train import add_model_arguments, get_model, get_params from icefall.checkpoint import ( average_checkpoints, diff --git a/egs/gigaspeech/ASR/zipformer/streaming_decode.py b/egs/gigaspeech/ASR/zipformer/streaming_decode.py index 7cada8c9d..cb3fd0dc7 100755 --- a/egs/gigaspeech/ASR/zipformer/streaming_decode.py +++ b/egs/gigaspeech/ASR/zipformer/streaming_decode.py @@ -51,7 +51,7 @@ from streaming_beam_search import ( ) from torch import Tensor, nn from torch.nn.utils.rnn import pad_sequence -from train import add_model_arguments, get_params, get_model +from train import add_model_arguments, get_model, get_params from icefall.checkpoint import ( average_checkpoints, diff --git a/egs/gigaspeech/ASR/zipformer/train.py b/egs/gigaspeech/ASR/zipformer/train.py index c5335562c..f0ad98147 100755 --- a/egs/gigaspeech/ASR/zipformer/train.py +++ b/egs/gigaspeech/ASR/zipformer/train.py @@ -89,6 +89,7 @@ from icefall.checkpoint import ( ) from icefall.dist import cleanup_dist, setup_dist from icefall.env import get_env_info +from icefall.err import raise_grad_scale_is_too_small_error from icefall.hooks import register_inf_check_hooks from icefall.utils import ( AttributeDict, @@ -1031,9 +1032,7 @@ def train_one_epoch( logging.warning(f"Grad scale is small: {cur_grad_scale}") if cur_grad_scale < 1.0e-05: save_bad_model() - raise RuntimeError( - f"grad_scale is too small, exiting: {cur_grad_scale}" - ) + raise_grad_scale_is_too_small_error(cur_grad_scale) if batch_idx % params.log_interval == 0: cur_lr = max(scheduler.get_last_lr()) diff --git a/egs/gigaspeech/KWS/zipformer/decode.py b/egs/gigaspeech/KWS/zipformer/decode.py index 98b003937..0df2ec356 100755 --- a/egs/gigaspeech/KWS/zipformer/decode.py +++ b/egs/gigaspeech/KWS/zipformer/decode.py @@ -42,12 +42,10 @@ import sentencepiece as spm import torch import torch.nn as nn from asr_datamodule import GigaSpeechAsrDataModule -from beam_search import ( - keywords_search, -) +from beam_search import keywords_search +from lhotse.cut import Cut from train import add_model_arguments, get_model, get_params -from lhotse.cut import Cut from icefall import ContextGraph from icefall.checkpoint import ( average_checkpoints, diff --git a/egs/gigaspeech/KWS/zipformer/finetune.py b/egs/gigaspeech/KWS/zipformer/finetune.py index b8e8802cb..a7ba56127 100755 --- a/egs/gigaspeech/KWS/zipformer/finetune.py +++ b/egs/gigaspeech/KWS/zipformer/finetune.py @@ -76,25 +76,6 @@ from torch import Tensor from torch.cuda.amp import GradScaler from torch.nn.parallel import DistributedDataParallel as DDP from torch.utils.tensorboard import SummaryWriter - -from icefall import diagnostics -from icefall.checkpoint import remove_checkpoints -from icefall.checkpoint import save_checkpoint as save_checkpoint_impl -from icefall.checkpoint import ( - save_checkpoint_with_global_batch_idx, - update_averaged_model, -) -from icefall.dist import cleanup_dist, setup_dist -from icefall.env import get_env_info -from icefall.hooks import register_inf_check_hooks -from icefall.utils import ( - AttributeDict, - MetricsTracker, - get_parameter_groups_with_lrs, - setup_logger, - str2bool, -) - from train import ( add_model_arguments, add_training_arguments, @@ -110,6 +91,25 @@ from train import ( set_batch_count, ) +from icefall import diagnostics +from icefall.checkpoint import remove_checkpoints +from icefall.checkpoint import save_checkpoint as save_checkpoint_impl +from icefall.checkpoint import ( + save_checkpoint_with_global_batch_idx, + update_averaged_model, +) +from icefall.dist import cleanup_dist, setup_dist +from icefall.env import get_env_info +from icefall.err import raise_grad_scale_is_too_small_error +from icefall.hooks import register_inf_check_hooks +from icefall.utils import ( + AttributeDict, + MetricsTracker, + get_parameter_groups_with_lrs, + setup_logger, + str2bool, +) + LRSchedulerType = Union[torch.optim.lr_scheduler._LRScheduler, optim.LRScheduler] @@ -372,9 +372,7 @@ def train_one_epoch( logging.warning(f"Grad scale is small: {cur_grad_scale}") if cur_grad_scale < 1.0e-05: save_bad_model() - raise RuntimeError( - f"grad_scale is too small, exiting: {cur_grad_scale}" - ) + raise_grad_scale_is_too_small_error(cur_grad_scale) if batch_idx % params.log_interval == 0: cur_lr = max(scheduler.get_last_lr()) diff --git a/egs/gigaspeech/KWS/zipformer/train.py b/egs/gigaspeech/KWS/zipformer/train.py index e7387dd39..a4d670169 100755 --- a/egs/gigaspeech/KWS/zipformer/train.py +++ b/egs/gigaspeech/KWS/zipformer/train.py @@ -89,6 +89,7 @@ from icefall.checkpoint import ( ) from icefall.dist import cleanup_dist, setup_dist from icefall.env import get_env_info +from icefall.err import raise_grad_scale_is_too_small_error from icefall.hooks import register_inf_check_hooks from icefall.utils import ( AttributeDict, @@ -1034,9 +1035,7 @@ def train_one_epoch( logging.warning(f"Grad scale is small: {cur_grad_scale}") if cur_grad_scale < 1.0e-05: save_bad_model() - raise RuntimeError( - f"grad_scale is too small, exiting: {cur_grad_scale}" - ) + raise_grad_scale_is_too_small_error(cur_grad_scale) if batch_idx % params.log_interval == 0: cur_lr = max(scheduler.get_last_lr()) diff --git a/egs/libricss/SURT/dprnn_zipformer/train.py b/egs/libricss/SURT/dprnn_zipformer/train.py index 6598f8b5d..90d742e7c 100755 --- a/egs/libricss/SURT/dprnn_zipformer/train.py +++ b/egs/libricss/SURT/dprnn_zipformer/train.py @@ -85,6 +85,7 @@ from icefall.checkpoint import ( ) from icefall.dist import cleanup_dist, setup_dist from icefall.env import get_env_info +from icefall.err import raise_grad_scale_is_too_small_error from icefall.utils import AttributeDict, MetricsTracker, setup_logger, str2bool LRSchedulerType = Union[torch.optim.lr_scheduler._LRScheduler, optim.LRScheduler] @@ -1169,9 +1170,7 @@ def train_one_epoch( if cur_grad_scale < 0.01: logging.warning(f"Grad scale is small: {cur_grad_scale}") if cur_grad_scale < 1.0e-05: - raise RuntimeError( - f"grad_scale is too small, exiting: {cur_grad_scale}" - ) + raise_grad_scale_is_too_small_error(cur_grad_scale) if batch_idx % params.log_interval == 0: cur_lr = scheduler.get_last_lr()[0] diff --git a/egs/libricss/SURT/dprnn_zipformer/train_adapt.py b/egs/libricss/SURT/dprnn_zipformer/train_adapt.py index 1c1b0c28c..8c37430ec 100755 --- a/egs/libricss/SURT/dprnn_zipformer/train_adapt.py +++ b/egs/libricss/SURT/dprnn_zipformer/train_adapt.py @@ -81,6 +81,7 @@ from icefall.checkpoint import ( ) from icefall.dist import cleanup_dist, setup_dist from icefall.env import get_env_info +from icefall.err import raise_grad_scale_is_too_small_error from icefall.utils import AttributeDict, MetricsTracker, setup_logger, str2bool LRSchedulerType = Union[torch.optim.lr_scheduler._LRScheduler, optim.LRScheduler] @@ -1056,9 +1057,7 @@ def train_one_epoch( if cur_grad_scale < 0.01: logging.warning(f"Grad scale is small: {cur_grad_scale}") if cur_grad_scale < 1.0e-05: - raise RuntimeError( - f"grad_scale is too small, exiting: {cur_grad_scale}" - ) + raise_grad_scale_is_too_small_error(cur_grad_scale) if batch_idx % params.log_interval == 0: cur_lr = scheduler.get_last_lr()[0] diff --git a/egs/libriheavy/ASR/zipformer/asr_datamodule.py b/egs/libriheavy/ASR/zipformer/asr_datamodule.py index e23c9b1b7..4985f3f4c 100644 --- a/egs/libriheavy/ASR/zipformer/asr_datamodule.py +++ b/egs/libriheavy/ASR/zipformer/asr_datamodule.py @@ -232,7 +232,7 @@ class LibriHeavyAsrDataModule: logging.info("About to get Musan cuts") cuts_musan = load_manifest(self.args.manifest_dir / "musan_cuts.jsonl.gz") transforms.append( - CutMix(cuts=cuts_musan, prob=0.5, snr=(10, 20), preserve_id=True) + CutMix(cuts=cuts_musan, p=0.5, snr=(10, 20), preserve_id=True) ) else: logging.info("Disable MUSAN") diff --git a/egs/libriheavy/ASR/zipformer/train.py b/egs/libriheavy/ASR/zipformer/train.py index c97da4a11..8d4d9d067 100644 --- a/egs/libriheavy/ASR/zipformer/train.py +++ b/egs/libriheavy/ASR/zipformer/train.py @@ -93,6 +93,7 @@ from icefall.checkpoint import ( ) from icefall.dist import cleanup_dist, setup_dist from icefall.env import get_env_info +from icefall.err import raise_grad_scale_is_too_small_error from icefall.hooks import register_inf_check_hooks from icefall.utils import ( AttributeDict, @@ -1036,9 +1037,7 @@ def train_one_epoch( logging.warning(f"Grad scale is small: {cur_grad_scale}") if cur_grad_scale < 1.0e-05: save_bad_model() - raise RuntimeError( - f"grad_scale is too small, exiting: {cur_grad_scale}" - ) + raise_grad_scale_is_too_small_error(cur_grad_scale) if batch_idx % params.log_interval == 0: cur_lr = max(scheduler.get_last_lr()) diff --git a/egs/libriheavy/ASR/zipformer_prompt_asr/asr_datamodule.py b/egs/libriheavy/ASR/zipformer_prompt_asr/asr_datamodule.py index 1a4c9a532..552f63905 100644 --- a/egs/libriheavy/ASR/zipformer_prompt_asr/asr_datamodule.py +++ b/egs/libriheavy/ASR/zipformer_prompt_asr/asr_datamodule.py @@ -425,9 +425,11 @@ class LibriHeavyAsrDataModule: def test_dataloaders(self, cuts: CutSet) -> DataLoader: logging.debug("About to create test dataset") test = K2SpeechRecognitionDataset( - input_strategy=OnTheFlyFeatures(Fbank(FbankConfig(num_mel_bins=80))) - if self.args.on_the_fly_feats - else PrecomputedFeatures(), + input_strategy=( + OnTheFlyFeatures(Fbank(FbankConfig(num_mel_bins=80))) + if self.args.on_the_fly_feats + else PrecomputedFeatures() + ), return_cuts=self.args.return_cuts, ) sampler = DynamicBucketingSampler( diff --git a/egs/libriheavy/ASR/zipformer_prompt_asr/train_baseline.py b/egs/libriheavy/ASR/zipformer_prompt_asr/train_baseline.py index c8b20d021..93f7e1248 100644 --- a/egs/libriheavy/ASR/zipformer_prompt_asr/train_baseline.py +++ b/egs/libriheavy/ASR/zipformer_prompt_asr/train_baseline.py @@ -103,6 +103,7 @@ from icefall.checkpoint import ( ) from icefall.dist import cleanup_dist, setup_dist from icefall.env import get_env_info +from icefall.err import raise_grad_scale_is_too_small_error from icefall.hooks import register_inf_check_hooks from icefall.utils import ( AttributeDict, @@ -1051,9 +1052,7 @@ def train_one_epoch( logging.warning(f"Grad scale is small: {cur_grad_scale}") if cur_grad_scale < 1.0e-05: save_bad_model() - raise RuntimeError( - f"grad_scale is too small, exiting: {cur_grad_scale}" - ) + raise_grad_scale_is_too_small_error(cur_grad_scale) if batch_idx % params.log_interval == 0: cur_lr = max(scheduler.get_last_lr()) diff --git a/egs/libriheavy/ASR/zipformer_prompt_asr/train_bert_encoder.py b/egs/libriheavy/ASR/zipformer_prompt_asr/train_bert_encoder.py index 9822b99c1..2a2c206aa 100755 --- a/egs/libriheavy/ASR/zipformer_prompt_asr/train_bert_encoder.py +++ b/egs/libriheavy/ASR/zipformer_prompt_asr/train_bert_encoder.py @@ -117,6 +117,7 @@ from icefall.checkpoint import ( ) from icefall.dist import cleanup_dist, setup_dist from icefall.env import get_env_info +from icefall.err import raise_grad_scale_is_too_small_error from icefall.hooks import register_inf_check_hooks from icefall.utils import ( AttributeDict, @@ -855,9 +856,9 @@ def get_joiner_model(params: AttributeDict) -> nn.Module: decoder_dim=params.decoder_dim, joiner_dim=params.joiner_dim, vocab_size=params.vocab_size, - context_dim=4 * 768 - if params.context_injection - else -1, # the output dim of text encoder + context_dim=( + 4 * 768 if params.context_injection else -1 + ), # the output dim of text encoder context_injection=params.context_injection, ) return joiner @@ -1398,9 +1399,7 @@ def train_one_epoch( logging.warning(f"Grad scale is small: {cur_grad_scale}") if cur_grad_scale < 1.0e-05: save_bad_model() - raise RuntimeError( - f"grad_scale is too small, exiting: {cur_grad_scale}" - ) + raise_grad_scale_is_too_small_error(cur_grad_scale) if batch_idx % params.log_interval == 0: cur_lr = max(scheduler.get_last_lr()) diff --git a/egs/librispeech/ASR/README.md b/egs/librispeech/ASR/README.md index edad41c60..5c5a76917 100644 --- a/egs/librispeech/ASR/README.md +++ b/egs/librispeech/ASR/README.md @@ -35,8 +35,7 @@ The following table lists the differences among them. | `lstm_transducer_stateless2` | LSTM | Embedding + Conv1d | Using LSTM with mechanisms in reworked model + gigaspeech (multi-dataset setup) | | `lstm_transducer_stateless3` | LSTM | Embedding + Conv1d | Using LSTM with mechanisms in reworked model + gradient filter + delay penalty | | `zipformer` | Upgraded Zipformer | Embedding + Conv1d | The latest recipe | -| `zipformer_adapter` | Upgraded Zipformer | Embedding + Conv1d | Finetune `zipformer` with adapter | -| `zipformer_lora` | Upgraded Zipformer | Embedding + Conv1d | Finetune `zipformer` with LoRA | +| `zipformer_adapter` | Upgraded Zipformer | Embedding + Conv1d | It supports domain adaptation of Zipformer using parameter efficient adapters | The decoder in `transducer_stateless` is modified from the paper [Rnn-Transducer with Stateless Prediction Network](https://ieeexplore.ieee.org/document/9054419/). diff --git a/egs/librispeech/ASR/conformer_ctc3/test_model.py b/egs/librispeech/ASR/conformer_ctc3/test_model.py index b97b7eed8..aa12d6f83 100755 --- a/egs/librispeech/ASR/conformer_ctc3/test_model.py +++ b/egs/librispeech/ASR/conformer_ctc3/test_model.py @@ -24,8 +24,7 @@ To run this file, do: """ import torch - -from train import get_params, get_ctc_model +from train import get_ctc_model, get_params def test_model(): diff --git a/egs/librispeech/ASR/conv_emformer_transducer_stateless2/export-onnx.py b/egs/librispeech/ASR/conv_emformer_transducer_stateless2/export-onnx.py index 1e59e0858..79728afa4 100755 --- a/egs/librispeech/ASR/conv_emformer_transducer_stateless2/export-onnx.py +++ b/egs/librispeech/ASR/conv_emformer_transducer_stateless2/export-onnx.py @@ -59,9 +59,9 @@ import onnx import torch import torch.nn as nn from decoder import Decoder +from do_not_use_it_directly import add_model_arguments, get_params, get_transducer_model from emformer import Emformer from scaling_converter import convert_scaled_to_non_scaled -from do_not_use_it_directly import add_model_arguments, get_params, get_transducer_model from icefall.checkpoint import ( average_checkpoints, diff --git a/egs/librispeech/ASR/conv_emformer_transducer_stateless2/jit_pretrained.py b/egs/librispeech/ASR/conv_emformer_transducer_stateless2/jit_pretrained.py index 58f587c91..1deecbfc7 100755 --- a/egs/librispeech/ASR/conv_emformer_transducer_stateless2/jit_pretrained.py +++ b/egs/librispeech/ASR/conv_emformer_transducer_stateless2/jit_pretrained.py @@ -39,7 +39,7 @@ Usage of this script: import argparse import logging import math -from typing import List +from typing import List, Optional import kaldifeat import sentencepiece as spm @@ -47,7 +47,6 @@ import torch import torchaudio from kaldifeat import FbankOptions, OnlineFbank, OnlineFeature from torch.nn.utils.rnn import pad_sequence -from typing import Optional, List def get_parser(): diff --git a/egs/librispeech/ASR/long_file_recog/recognize.py b/egs/librispeech/ASR/long_file_recog/recognize.py index 466253446..f4008c23b 100755 --- a/egs/librispeech/ASR/long_file_recog/recognize.py +++ b/egs/librispeech/ASR/long_file_recog/recognize.py @@ -31,28 +31,28 @@ https://huggingface.co/csukuangfj/icefall-asr-librispeech-pruned-transducer-stat """ import argparse -import torch.multiprocessing as mp -import torch -import torch.nn as nn import logging from concurrent.futures import ThreadPoolExecutor -from typing import List, Optional, Tuple - from pathlib import Path +from typing import List, Optional, Tuple import k2 import sentencepiece as spm +import torch +import torch.multiprocessing as mp +import torch.nn as nn from asr_datamodule import AsrDataModule from beam_search import ( fast_beam_search_one_best, greedy_search_batch, modified_beam_search, ) -from icefall.utils import AttributeDict, convert_timestamp, setup_logger from lhotse import CutSet, load_manifest_lazy from lhotse.cut import Cut -from lhotse.supervision import AlignmentItem from lhotse.serialization import SequentialJsonlWriter +from lhotse.supervision import AlignmentItem + +from icefall.utils import AttributeDict, convert_timestamp, setup_logger def get_parser(): diff --git a/egs/librispeech/ASR/lstm_transducer_stateless2/onnx_check.py b/egs/librispeech/ASR/lstm_transducer_stateless2/onnx_check.py index c83f38b2a..85e0648d3 100755 --- a/egs/librispeech/ASR/lstm_transducer_stateless2/onnx_check.py +++ b/egs/librispeech/ASR/lstm_transducer_stateless2/onnx_check.py @@ -73,12 +73,11 @@ It will generate the following 3 files inside $repo/exp: import argparse import logging +import torch from onnx_pretrained import OnnxModel from icefall import is_module_available -import torch - def get_parser(): parser = argparse.ArgumentParser( diff --git a/egs/librispeech/ASR/pruned_transducer_stateless/my_profile.py b/egs/librispeech/ASR/pruned_transducer_stateless/my_profile.py index b844ba613..9762d878c 100755 --- a/egs/librispeech/ASR/pruned_transducer_stateless/my_profile.py +++ b/egs/librispeech/ASR/pruned_transducer_stateless/my_profile.py @@ -22,11 +22,12 @@ Usage: ./pruned_transducer_stateless/my_profile.py import argparse import logging + import sentencepiece as spm import torch +from train import add_model_arguments, get_encoder_model, get_params from icefall.profiler import get_model_profile -from train import get_encoder_model, add_model_arguments, get_params def get_parser(): diff --git a/egs/librispeech/ASR/pruned_transducer_stateless/onnx_decode.py b/egs/librispeech/ASR/pruned_transducer_stateless/onnx_decode.py index 8134d43f8..a235d7b13 100755 --- a/egs/librispeech/ASR/pruned_transducer_stateless/onnx_decode.py +++ b/egs/librispeech/ASR/pruned_transducer_stateless/onnx_decode.py @@ -75,8 +75,7 @@ import sentencepiece as spm import torch import torch.nn as nn from asr_datamodule import LibriSpeechAsrDataModule - -from onnx_pretrained import greedy_search, OnnxModel +from onnx_pretrained import OnnxModel, greedy_search from icefall.utils import setup_logger, store_transcripts, write_error_stats diff --git a/egs/librispeech/ASR/pruned_transducer_stateless3/onnx_check.py b/egs/librispeech/ASR/pruned_transducer_stateless3/onnx_check.py index 5ca4173c1..e2c1d6b5b 100755 --- a/egs/librispeech/ASR/pruned_transducer_stateless3/onnx_check.py +++ b/egs/librispeech/ASR/pruned_transducer_stateless3/onnx_check.py @@ -78,10 +78,10 @@ It will generate the following 3 files inside $repo/exp: import argparse import logging -from icefall import is_module_available +import torch from onnx_pretrained import OnnxModel -import torch +from icefall import is_module_available def get_parser(): diff --git a/egs/librispeech/ASR/pruned_transducer_stateless3/onnx_decode.py b/egs/librispeech/ASR/pruned_transducer_stateless3/onnx_decode.py index 3b1c72cf1..f8fed9519 100755 --- a/egs/librispeech/ASR/pruned_transducer_stateless3/onnx_decode.py +++ b/egs/librispeech/ASR/pruned_transducer_stateless3/onnx_decode.py @@ -76,8 +76,7 @@ import torch import torch.nn as nn from asr_datamodule import AsrDataModule from librispeech import LibriSpeech - -from onnx_pretrained import greedy_search, OnnxModel +from onnx_pretrained import OnnxModel, greedy_search from icefall.utils import setup_logger, store_transcripts, write_error_stats diff --git a/egs/librispeech/ASR/pruned_transducer_stateless4/my_profile.py b/egs/librispeech/ASR/pruned_transducer_stateless4/my_profile.py index 4bf773918..cf0598ca3 100755 --- a/egs/librispeech/ASR/pruned_transducer_stateless4/my_profile.py +++ b/egs/librispeech/ASR/pruned_transducer_stateless4/my_profile.py @@ -22,15 +22,15 @@ Usage: ./pruned_transducer_stateless4/my_profile.py import argparse import logging +from typing import Tuple + import sentencepiece as spm import torch - -from typing import Tuple +from scaling import BasicNorm, DoubleSwish from torch import Tensor, nn +from train import add_model_arguments, get_encoder_model, get_joiner_model, get_params from icefall.profiler import get_model_profile -from scaling import BasicNorm, DoubleSwish -from train import get_encoder_model, get_joiner_model, add_model_arguments, get_params def get_parser(): diff --git a/egs/librispeech/ASR/pruned_transducer_stateless5/onnx_decode.py b/egs/librispeech/ASR/pruned_transducer_stateless5/onnx_decode.py index 6f26e34b5..b0f76317b 100755 --- a/egs/librispeech/ASR/pruned_transducer_stateless5/onnx_decode.py +++ b/egs/librispeech/ASR/pruned_transducer_stateless5/onnx_decode.py @@ -82,8 +82,7 @@ import sentencepiece as spm import torch import torch.nn as nn from asr_datamodule import LibriSpeechAsrDataModule - -from onnx_pretrained import greedy_search, OnnxModel +from onnx_pretrained import OnnxModel, greedy_search from icefall.utils import setup_logger, store_transcripts, write_error_stats diff --git a/egs/librispeech/ASR/pruned_transducer_stateless7/alignment.py b/egs/librispeech/ASR/pruned_transducer_stateless7/alignment.py index bfb5fe609..ee8196c3f 100644 --- a/egs/librispeech/ASR/pruned_transducer_stateless7/alignment.py +++ b/egs/librispeech/ASR/pruned_transducer_stateless7/alignment.py @@ -20,7 +20,6 @@ from typing import List import k2 import torch - from beam_search import Hypothesis, HypothesisList, get_hyps_shape # The force alignment problem can be formulated as finding diff --git a/egs/librispeech/ASR/pruned_transducer_stateless7/decode_gigaspeech.py b/egs/librispeech/ASR/pruned_transducer_stateless7/decode_gigaspeech.py index b0e4be0d1..7095c3cc8 100644 --- a/egs/librispeech/ASR/pruned_transducer_stateless7/decode_gigaspeech.py +++ b/egs/librispeech/ASR/pruned_transducer_stateless7/decode_gigaspeech.py @@ -107,9 +107,6 @@ import k2 import sentencepiece as spm import torch import torch.nn as nn - -# from asr_datamodule import LibriSpeechAsrDataModule -from gigaspeech import GigaSpeechAsrDataModule from beam_search import ( beam_search, fast_beam_search_nbest, @@ -120,6 +117,9 @@ from beam_search import ( greedy_search_batch, modified_beam_search, ) + +# from asr_datamodule import LibriSpeechAsrDataModule +from gigaspeech import GigaSpeechAsrDataModule from gigaspeech_scoring import asr_text_post_processing from train import add_model_arguments, get_params, get_transducer_model diff --git a/egs/librispeech/ASR/pruned_transducer_stateless7/finetune.py b/egs/librispeech/ASR/pruned_transducer_stateless7/finetune.py index a7a8ef149..e7546ec45 100755 --- a/egs/librispeech/ASR/pruned_transducer_stateless7/finetune.py +++ b/egs/librispeech/ASR/pruned_transducer_stateless7/finetune.py @@ -80,6 +80,7 @@ from icefall.checkpoint import ( ) from icefall.dist import cleanup_dist, setup_dist from icefall.env import get_env_info +from icefall.err import raise_grad_scale_is_too_small_error from icefall.hooks import register_inf_check_hooks from icefall.utils import ( AttributeDict, @@ -976,9 +977,7 @@ def train_one_epoch( if cur_grad_scale < 0.01: logging.warning(f"Grad scale is small: {cur_grad_scale}") if cur_grad_scale < 1.0e-05: - raise RuntimeError( - f"grad_scale is too small, exiting: {cur_grad_scale}" - ) + raise_grad_scale_is_too_small_error(cur_grad_scale) if batch_idx % params.log_interval == 0: cur_lr = scheduler.get_last_lr()[0] diff --git a/egs/librispeech/ASR/pruned_transducer_stateless7/generate_model_from_checkpoint.py b/egs/librispeech/ASR/pruned_transducer_stateless7/generate_model_from_checkpoint.py index 37edc0390..3fd14aa47 100755 --- a/egs/librispeech/ASR/pruned_transducer_stateless7/generate_model_from_checkpoint.py +++ b/egs/librispeech/ASR/pruned_transducer_stateless7/generate_model_from_checkpoint.py @@ -65,16 +65,15 @@ from typing import Dict, List import sentencepiece as spm import torch - from train import add_model_arguments, get_params, get_transducer_model -from icefall.utils import str2bool from icefall.checkpoint import ( average_checkpoints, average_checkpoints_with_averaged_model, find_checkpoints, load_checkpoint, ) +from icefall.utils import str2bool def get_parser(): diff --git a/egs/librispeech/ASR/pruned_transducer_stateless7/my_profile.py b/egs/librispeech/ASR/pruned_transducer_stateless7/my_profile.py index 5a068b3b6..1416c6828 100755 --- a/egs/librispeech/ASR/pruned_transducer_stateless7/my_profile.py +++ b/egs/librispeech/ASR/pruned_transducer_stateless7/my_profile.py @@ -22,15 +22,15 @@ Usage: ./pruned_transducer_stateless7/my_profile.py import argparse import logging +from typing import Tuple + import sentencepiece as spm import torch - -from typing import Tuple +from scaling import BasicNorm, DoubleSwish from torch import Tensor, nn +from train import add_model_arguments, get_encoder_model, get_joiner_model, get_params from icefall.profiler import get_model_profile -from scaling import BasicNorm, DoubleSwish -from train import get_encoder_model, get_joiner_model, add_model_arguments, get_params def get_parser(): diff --git a/egs/librispeech/ASR/pruned_transducer_stateless7/onnx_decode.py b/egs/librispeech/ASR/pruned_transducer_stateless7/onnx_decode.py index 67585ee47..e00281239 100755 --- a/egs/librispeech/ASR/pruned_transducer_stateless7/onnx_decode.py +++ b/egs/librispeech/ASR/pruned_transducer_stateless7/onnx_decode.py @@ -75,8 +75,7 @@ import sentencepiece as spm import torch import torch.nn as nn from asr_datamodule import LibriSpeechAsrDataModule - -from onnx_pretrained import greedy_search, OnnxModel +from onnx_pretrained import OnnxModel, greedy_search from icefall.utils import setup_logger, store_transcripts, write_error_stats diff --git a/egs/librispeech/ASR/pruned_transducer_stateless7/test_model.py b/egs/librispeech/ASR/pruned_transducer_stateless7/test_model.py index cdf914df3..1f50eb309 100755 --- a/egs/librispeech/ASR/pruned_transducer_stateless7/test_model.py +++ b/egs/librispeech/ASR/pruned_transducer_stateless7/test_model.py @@ -24,7 +24,6 @@ To run this file, do: """ import torch - from scaling_converter import convert_scaled_to_non_scaled from train import get_params, get_transducer_model diff --git a/egs/librispeech/ASR/pruned_transducer_stateless7/train.py b/egs/librispeech/ASR/pruned_transducer_stateless7/train.py index fac3706d2..436ec53b4 100755 --- a/egs/librispeech/ASR/pruned_transducer_stateless7/train.py +++ b/egs/librispeech/ASR/pruned_transducer_stateless7/train.py @@ -81,6 +81,7 @@ from icefall.checkpoint import ( ) from icefall.dist import cleanup_dist, setup_dist from icefall.env import get_env_info +from icefall.err import raise_grad_scale_is_too_small_error from icefall.hooks import register_inf_check_hooks from icefall.utils import ( AttributeDict, @@ -878,9 +879,7 @@ def train_one_epoch( if cur_grad_scale < 0.01: logging.warning(f"Grad scale is small: {cur_grad_scale}") if cur_grad_scale < 1.0e-05: - raise RuntimeError( - f"grad_scale is too small, exiting: {cur_grad_scale}" - ) + raise_grad_scale_is_too_small_error(cur_grad_scale) if batch_idx % params.log_interval == 0: cur_lr = scheduler.get_last_lr()[0] diff --git a/egs/librispeech/ASR/pruned_transducer_stateless7_ctc/train.py b/egs/librispeech/ASR/pruned_transducer_stateless7_ctc/train.py index d8fa08372..b35e56abc 100755 --- a/egs/librispeech/ASR/pruned_transducer_stateless7_ctc/train.py +++ b/egs/librispeech/ASR/pruned_transducer_stateless7_ctc/train.py @@ -81,6 +81,7 @@ from icefall.checkpoint import ( ) from icefall.dist import cleanup_dist, setup_dist from icefall.env import get_env_info +from icefall.err import raise_grad_scale_is_too_small_error from icefall.hooks import register_inf_check_hooks from icefall.utils import ( AttributeDict, @@ -902,9 +903,7 @@ def train_one_epoch( if cur_grad_scale < 0.01: logging.warning(f"Grad scale is small: {cur_grad_scale}") if cur_grad_scale < 1.0e-05: - raise RuntimeError( - f"grad_scale is too small, exiting: {cur_grad_scale}" - ) + raise_grad_scale_is_too_small_error(cur_grad_scale) if batch_idx % params.log_interval == 0: cur_lr = scheduler.get_last_lr()[0] diff --git a/egs/librispeech/ASR/pruned_transducer_stateless7_ctc_bs/ctc_guide_decode_bs.py b/egs/librispeech/ASR/pruned_transducer_stateless7_ctc_bs/ctc_guide_decode_bs.py index 01ba7b711..e2f08abc6 100755 --- a/egs/librispeech/ASR/pruned_transducer_stateless7_ctc_bs/ctc_guide_decode_bs.py +++ b/egs/librispeech/ASR/pruned_transducer_stateless7_ctc_bs/ctc_guide_decode_bs.py @@ -118,8 +118,8 @@ from beam_search import ( greedy_search_batch, modified_beam_search, ) -from train import add_model_arguments, get_params, get_transducer_model from torch.nn.utils.rnn import pad_sequence +from train import add_model_arguments, get_params, get_transducer_model from icefall.checkpoint import ( average_checkpoints, diff --git a/egs/librispeech/ASR/pruned_transducer_stateless7_ctc_bs/lconv.py b/egs/librispeech/ASR/pruned_transducer_stateless7_ctc_bs/lconv.py index a902358ae..2faec7ade 100644 --- a/egs/librispeech/ASR/pruned_transducer_stateless7_ctc_bs/lconv.py +++ b/egs/librispeech/ASR/pruned_transducer_stateless7_ctc_bs/lconv.py @@ -18,10 +18,7 @@ from typing import List, Optional, Tuple, Union import torch import torch.nn as nn -from scaling import ( - ActivationBalancer, - ScaledConv1d, -) +from scaling import ActivationBalancer, ScaledConv1d class LConv(nn.Module): diff --git a/egs/librispeech/ASR/pruned_transducer_stateless7_ctc_bs/onnx_pretrained.py b/egs/librispeech/ASR/pruned_transducer_stateless7_ctc_bs/onnx_pretrained.py index 0ff110370..3a16985bc 100755 --- a/egs/librispeech/ASR/pruned_transducer_stateless7_ctc_bs/onnx_pretrained.py +++ b/egs/librispeech/ASR/pruned_transducer_stateless7_ctc_bs/onnx_pretrained.py @@ -52,7 +52,7 @@ import onnxruntime as ort import sentencepiece as spm import torch import torchaudio -from torch.nn.utils.rnn import pad_sequence, pack_padded_sequence +from torch.nn.utils.rnn import pack_padded_sequence, pad_sequence from icefall.utils import make_pad_mask diff --git a/egs/librispeech/ASR/pruned_transducer_stateless7_ctc_bs/onnx_wrapper.py b/egs/librispeech/ASR/pruned_transducer_stateless7_ctc_bs/onnx_wrapper.py index 247da0949..07e97bbdb 100755 --- a/egs/librispeech/ASR/pruned_transducer_stateless7_ctc_bs/onnx_wrapper.py +++ b/egs/librispeech/ASR/pruned_transducer_stateless7_ctc_bs/onnx_wrapper.py @@ -14,6 +14,7 @@ import torch from torch import nn + from icefall.utils import make_pad_mask diff --git a/egs/librispeech/ASR/pruned_transducer_stateless7_ctc_bs/train.py b/egs/librispeech/ASR/pruned_transducer_stateless7_ctc_bs/train.py index 25a1aa674..c2d877a93 100755 --- a/egs/librispeech/ASR/pruned_transducer_stateless7_ctc_bs/train.py +++ b/egs/librispeech/ASR/pruned_transducer_stateless7_ctc_bs/train.py @@ -77,6 +77,7 @@ from icefall.checkpoint import ( ) from icefall.dist import cleanup_dist, setup_dist from icefall.env import get_env_info +from icefall.err import raise_grad_scale_is_too_small_error from icefall.hooks import register_inf_check_hooks from icefall.utils import ( AttributeDict, @@ -891,9 +892,7 @@ def train_one_epoch( if cur_grad_scale < 0.01: logging.warning(f"Grad scale is small: {cur_grad_scale}") if cur_grad_scale < 1.0e-05: - raise RuntimeError( - f"grad_scale is too small, exiting: {cur_grad_scale}" - ) + raise_grad_scale_is_too_small_error(cur_grad_scale) if batch_idx % params.log_interval == 0: cur_lr = scheduler.get_last_lr()[0] diff --git a/egs/librispeech/ASR/pruned_transducer_stateless7_streaming/do_not_use_it_directly.py b/egs/librispeech/ASR/pruned_transducer_stateless7_streaming/do_not_use_it_directly.py index 9a6d2155b..8e239e322 100755 --- a/egs/librispeech/ASR/pruned_transducer_stateless7_streaming/do_not_use_it_directly.py +++ b/egs/librispeech/ASR/pruned_transducer_stateless7_streaming/do_not_use_it_directly.py @@ -80,6 +80,7 @@ from icefall.checkpoint import ( ) from icefall.dist import cleanup_dist, setup_dist from icefall.env import get_env_info +from icefall.err import raise_grad_scale_is_too_small_error from icefall.hooks import register_inf_check_hooks from icefall.utils import AttributeDict, MetricsTracker, setup_logger, str2bool @@ -880,9 +881,7 @@ def train_one_epoch( if cur_grad_scale < 0.01: logging.warning(f"Grad scale is small: {cur_grad_scale}") if cur_grad_scale < 1.0e-05: - raise RuntimeError( - f"grad_scale is too small, exiting: {cur_grad_scale}" - ) + raise_grad_scale_is_too_small_error(cur_grad_scale) if batch_idx % params.log_interval == 0: cur_lr = scheduler.get_last_lr()[0] diff --git a/egs/librispeech/ASR/pruned_transducer_stateless7_streaming/ncnn_custom_layer.py b/egs/librispeech/ASR/pruned_transducer_stateless7_streaming/ncnn_custom_layer.py index 442a0a8af..451c35332 100644 --- a/egs/librispeech/ASR/pruned_transducer_stateless7_streaming/ncnn_custom_layer.py +++ b/egs/librispeech/ASR/pruned_transducer_stateless7_streaming/ncnn_custom_layer.py @@ -4,7 +4,6 @@ import ncnn import numpy as np - layer_list = [] diff --git a/egs/librispeech/ASR/pruned_transducer_stateless7_streaming/streaming-ncnn-decode.py b/egs/librispeech/ASR/pruned_transducer_stateless7_streaming/streaming-ncnn-decode.py index 999f7e0b4..06127607d 100755 --- a/egs/librispeech/ASR/pruned_transducer_stateless7_streaming/streaming-ncnn-decode.py +++ b/egs/librispeech/ASR/pruned_transducer_stateless7_streaming/streaming-ncnn-decode.py @@ -42,7 +42,6 @@ import ncnn import torch import torchaudio from kaldifeat import FbankOptions, OnlineFbank, OnlineFeature - from ncnn_custom_layer import RegisterCustomLayers diff --git a/egs/librispeech/ASR/pruned_transducer_stateless7_streaming/train.py b/egs/librispeech/ASR/pruned_transducer_stateless7_streaming/train.py index e1bdce49d..8bd00bbef 100755 --- a/egs/librispeech/ASR/pruned_transducer_stateless7_streaming/train.py +++ b/egs/librispeech/ASR/pruned_transducer_stateless7_streaming/train.py @@ -80,6 +80,7 @@ from icefall.checkpoint import ( ) from icefall.dist import cleanup_dist, setup_dist from icefall.env import get_env_info +from icefall.err import raise_grad_scale_is_too_small_error from icefall.hooks import register_inf_check_hooks from icefall.utils import AttributeDict, MetricsTracker, setup_logger, str2bool @@ -879,9 +880,7 @@ def train_one_epoch( if cur_grad_scale < 0.01: logging.warning(f"Grad scale is small: {cur_grad_scale}") if cur_grad_scale < 1.0e-05: - raise RuntimeError( - f"grad_scale is too small, exiting: {cur_grad_scale}" - ) + raise_grad_scale_is_too_small_error(cur_grad_scale) if batch_idx % params.log_interval == 0: cur_lr = scheduler.get_last_lr()[0] diff --git a/egs/librispeech/ASR/pruned_transducer_stateless7_streaming_multi/train.py b/egs/librispeech/ASR/pruned_transducer_stateless7_streaming_multi/train.py index 1642ef4b7..da5e144c9 100755 --- a/egs/librispeech/ASR/pruned_transducer_stateless7_streaming_multi/train.py +++ b/egs/librispeech/ASR/pruned_transducer_stateless7_streaming_multi/train.py @@ -84,6 +84,7 @@ from icefall.checkpoint import ( ) from icefall.dist import cleanup_dist, setup_dist from icefall.env import get_env_info +from icefall.err import raise_grad_scale_is_too_small_error from icefall.hooks import register_inf_check_hooks from icefall.utils import AttributeDict, MetricsTracker, setup_logger, str2bool @@ -946,9 +947,7 @@ def train_one_epoch( if cur_grad_scale < 0.01: logging.warning(f"Grad scale is small: {cur_grad_scale}") if cur_grad_scale < 1.0e-05: - raise RuntimeError( - f"grad_scale is too small, exiting: {cur_grad_scale}" - ) + raise_grad_scale_is_too_small_error(cur_grad_scale) if batch_idx % params.log_interval == 0: cur_lr = scheduler.get_last_lr()[0] diff --git a/egs/librispeech/ASR/pruned_transducer_stateless8/train.py b/egs/librispeech/ASR/pruned_transducer_stateless8/train.py index 3f271c5b4..646f30ca1 100755 --- a/egs/librispeech/ASR/pruned_transducer_stateless8/train.py +++ b/egs/librispeech/ASR/pruned_transducer_stateless8/train.py @@ -89,6 +89,7 @@ from icefall.checkpoint import ( ) from icefall.dist import cleanup_dist, setup_dist from icefall.env import get_env_info +from icefall.err import raise_grad_scale_is_too_small_error from icefall.hooks import register_inf_check_hooks from icefall.utils import AttributeDict, MetricsTracker, setup_logger, str2bool @@ -946,9 +947,7 @@ def train_one_epoch( if cur_grad_scale < 0.01: logging.warning(f"Grad scale is small: {cur_grad_scale}") if cur_grad_scale < 1.0e-05: - raise RuntimeError( - f"grad_scale is too small, exiting: {cur_grad_scale}" - ) + raise_grad_scale_is_too_small_error(cur_grad_scale) if batch_idx % params.log_interval == 0: cur_lr = scheduler.get_last_lr()[0] diff --git a/egs/librispeech/ASR/tiny_transducer_ctc/decode.py b/egs/librispeech/ASR/tiny_transducer_ctc/decode.py index 6c2bf9ea1..cc4471e2b 100644 --- a/egs/librispeech/ASR/tiny_transducer_ctc/decode.py +++ b/egs/librispeech/ASR/tiny_transducer_ctc/decode.py @@ -1,10 +1,11 @@ import argparse import logging import math +import pprint from collections import defaultdict from pathlib import Path from typing import Dict, List, Optional, Tuple -import pprint + import k2 import sentencepiece as spm import torch diff --git a/egs/librispeech/ASR/tiny_transducer_ctc/train.py b/egs/librispeech/ASR/tiny_transducer_ctc/train.py index 8920764cd..1bfd071de 100644 --- a/egs/librispeech/ASR/tiny_transducer_ctc/train.py +++ b/egs/librispeech/ASR/tiny_transducer_ctc/train.py @@ -66,6 +66,7 @@ from icefall.checkpoint import ( ) from icefall.dist import cleanup_dist, setup_dist from icefall.env import get_env_info +from icefall.err import raise_grad_scale_is_too_small_error from icefall.hooks import register_inf_check_hooks from icefall.lexicon import UniqLexicon from icefall.utils import ( @@ -883,9 +884,7 @@ def train_one_epoch( if cur_grad_scale < 0.01: logging.warning(f"Grad scale is small: {cur_grad_scale}") if cur_grad_scale < 1.0e-05: - raise RuntimeError( - f"grad_scale is too small, exiting: {cur_grad_scale}" - ) + raise_grad_scale_is_too_small_error(cur_grad_scale) if batch_idx % params.log_interval == 0: cur_lr = scheduler.get_last_lr()[0] diff --git a/egs/librispeech/ASR/zipformer/ctc_decode.py b/egs/librispeech/ASR/zipformer/ctc_decode.py index 4db50b981..1f0f9bfac 100755 --- a/egs/librispeech/ASR/zipformer/ctc_decode.py +++ b/egs/librispeech/ASR/zipformer/ctc_decode.py @@ -88,7 +88,7 @@ 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_model +from train import add_model_arguments, get_model, get_params from icefall.checkpoint import ( average_checkpoints, diff --git a/egs/librispeech/ASR/zipformer/finetune.py b/egs/librispeech/ASR/zipformer/finetune.py index 843d103cc..2f7ec0c17 100755 --- a/egs/librispeech/ASR/zipformer/finetune.py +++ b/egs/librispeech/ASR/zipformer/finetune.py @@ -92,6 +92,7 @@ from icefall.checkpoint import ( ) from icefall.dist import cleanup_dist, setup_dist from icefall.env import get_env_info +from icefall.err import raise_grad_scale_is_too_small_error from icefall.hooks import register_inf_check_hooks from icefall.utils import ( AttributeDict, @@ -1122,9 +1123,7 @@ def train_one_epoch( logging.warning(f"Grad scale is small: {cur_grad_scale}") if cur_grad_scale < 1.0e-05: save_bad_model() - raise RuntimeError( - f"grad_scale is too small, exiting: {cur_grad_scale}" - ) + raise_grad_scale_is_too_small_error(cur_grad_scale) if batch_idx % params.log_interval == 0: cur_lr = max(scheduler.get_last_lr()) diff --git a/egs/librispeech/ASR/zipformer/model.py b/egs/librispeech/ASR/zipformer/model.py index 73009d35c..86da3ab29 100644 --- a/egs/librispeech/ASR/zipformer/model.py +++ b/egs/librispeech/ASR/zipformer/model.py @@ -22,9 +22,9 @@ import k2 import torch import torch.nn as nn from encoder_interface import EncoderInterface +from scaling import ScaledLinear from icefall.utils import add_sos, make_pad_mask -from scaling import ScaledLinear class AsrModel(nn.Module): diff --git a/egs/librispeech/ASR/zipformer/my_profile.py b/egs/librispeech/ASR/zipformer/my_profile.py index ca20956fb..7e1fd777a 100755 --- a/egs/librispeech/ASR/zipformer/my_profile.py +++ b/egs/librispeech/ASR/zipformer/my_profile.py @@ -22,24 +22,24 @@ Usage: ./zipformer/my_profile.py import argparse import logging +from typing import Tuple + import sentencepiece as spm import torch - -from typing import Tuple -from torch import Tensor, nn - -from icefall.utils import make_pad_mask -from icefall.profiler import get_model_profile from scaling import BiasNorm +from torch import Tensor, nn from train import ( + add_model_arguments, get_encoder_embed, get_encoder_model, get_joiner_model, - add_model_arguments, get_params, ) from zipformer import BypassModule +from icefall.profiler import get_model_profile +from icefall.utils import make_pad_mask + def get_parser(): parser = argparse.ArgumentParser( diff --git a/egs/librispeech/ASR/zipformer/onnx_decode.py b/egs/librispeech/ASR/zipformer/onnx_decode.py index 356c2a830..449294444 100755 --- a/egs/librispeech/ASR/zipformer/onnx_decode.py +++ b/egs/librispeech/ASR/zipformer/onnx_decode.py @@ -77,11 +77,10 @@ from typing import List, Tuple import torch import torch.nn as nn from asr_datamodule import LibriSpeechAsrDataModule - -from onnx_pretrained import greedy_search, OnnxModel +from k2 import SymbolTable +from onnx_pretrained import OnnxModel, greedy_search from icefall.utils import setup_logger, store_transcripts, write_error_stats -from k2 import SymbolTable def get_parser(): diff --git a/egs/librispeech/ASR/zipformer/onnx_pretrained_ctc_H.py b/egs/librispeech/ASR/zipformer/onnx_pretrained_ctc_H.py index a77c3bf2a..114490599 100755 --- a/egs/librispeech/ASR/zipformer/onnx_pretrained_ctc_H.py +++ b/egs/librispeech/ASR/zipformer/onnx_pretrained_ctc_H.py @@ -27,11 +27,10 @@ https://huggingface.co/csukuangfj/sherpa-onnx-zipformer-ctc-en-2023-10-02 import argparse import logging import math -from typing import List, Tuple +from typing import Dict, List, Tuple import k2 import kaldifeat -from typing import Dict import kaldifst import onnxruntime as ort import torch diff --git a/egs/librispeech/ASR/zipformer/onnx_pretrained_ctc_HL.py b/egs/librispeech/ASR/zipformer/onnx_pretrained_ctc_HL.py index 6ef944514..f7d3e5253 100755 --- a/egs/librispeech/ASR/zipformer/onnx_pretrained_ctc_HL.py +++ b/egs/librispeech/ASR/zipformer/onnx_pretrained_ctc_HL.py @@ -27,11 +27,10 @@ https://huggingface.co/csukuangfj/sherpa-onnx-zipformer-ctc-en-2023-10-02 import argparse import logging import math -from typing import List, Tuple +from typing import Dict, List, Tuple import k2 import kaldifeat -from typing import Dict import kaldifst import onnxruntime as ort import torch diff --git a/egs/librispeech/ASR/zipformer/onnx_pretrained_ctc_HLG.py b/egs/librispeech/ASR/zipformer/onnx_pretrained_ctc_HLG.py index ccb3107ea..ebd385364 100755 --- a/egs/librispeech/ASR/zipformer/onnx_pretrained_ctc_HLG.py +++ b/egs/librispeech/ASR/zipformer/onnx_pretrained_ctc_HLG.py @@ -27,11 +27,10 @@ https://huggingface.co/csukuangfj/sherpa-onnx-zipformer-ctc-en-2023-10-02 import argparse import logging import math -from typing import List, Tuple +from typing import Dict, List, Tuple import k2 import kaldifeat -from typing import Dict import kaldifst import onnxruntime as ort import torch diff --git a/egs/librispeech/ASR/zipformer/scaling.py b/egs/librispeech/ASR/zipformer/scaling.py index c0f1e3087..29ac33c02 100644 --- a/egs/librispeech/ASR/zipformer/scaling.py +++ b/egs/librispeech/ASR/zipformer/scaling.py @@ -15,15 +15,16 @@ # limitations under the License. -from typing import Optional, Tuple, Union import logging -import k2 -from torch.cuda.amp import custom_fwd, custom_bwd -import random -import torch import math +import random +from typing import Optional, Tuple, Union + +import k2 +import torch import torch.nn as nn from torch import Tensor +from torch.cuda.amp import custom_bwd, custom_fwd def logaddexp_onnx(x: Tensor, y: Tensor) -> Tensor: diff --git a/egs/librispeech/ASR/zipformer/streaming_decode.py b/egs/librispeech/ASR/zipformer/streaming_decode.py index 8087c1460..360523b8e 100755 --- a/egs/librispeech/ASR/zipformer/streaming_decode.py +++ b/egs/librispeech/ASR/zipformer/streaming_decode.py @@ -51,7 +51,7 @@ from streaming_beam_search import ( ) from torch import Tensor, nn from torch.nn.utils.rnn import pad_sequence -from train import add_model_arguments, get_params, get_model +from train import add_model_arguments, get_model, get_params from icefall.checkpoint import ( average_checkpoints, diff --git a/egs/librispeech/ASR/zipformer/subsampling.py b/egs/librispeech/ASR/zipformer/subsampling.py index d16d87bac..b2f769d3f 100644 --- a/egs/librispeech/ASR/zipformer/subsampling.py +++ b/egs/librispeech/ASR/zipformer/subsampling.py @@ -16,11 +16,10 @@ # See the License for the specific language governing permissions and # limitations under the License. -from typing import Tuple import warnings +from typing import Tuple import torch -from torch import Tensor, nn from scaling import ( Balancer, BiasNorm, @@ -34,6 +33,7 @@ from scaling import ( SwooshR, Whiten, ) +from torch import Tensor, nn class ConvNeXt(nn.Module): diff --git a/egs/librispeech/ASR/zipformer/train.py b/egs/librispeech/ASR/zipformer/train.py index 3ccf7d2f1..1111d32ab 100755 --- a/egs/librispeech/ASR/zipformer/train.py +++ b/egs/librispeech/ASR/zipformer/train.py @@ -90,6 +90,7 @@ from icefall.checkpoint import ( ) from icefall.dist import cleanup_dist, setup_dist from icefall.env import get_env_info +from icefall.err import raise_grad_scale_is_too_small_error from icefall.hooks import register_inf_check_hooks from icefall.utils import ( AttributeDict, @@ -1021,9 +1022,7 @@ def train_one_epoch( logging.warning(f"Grad scale is small: {cur_grad_scale}") if cur_grad_scale < 1.0e-05: save_bad_model() - raise RuntimeError( - f"grad_scale is too small, exiting: {cur_grad_scale}" - ) + raise_grad_scale_is_too_small_error(cur_grad_scale) if batch_idx % params.log_interval == 0: cur_lr = max(scheduler.get_last_lr()) diff --git a/egs/librispeech/ASR/zipformer_adapter/asr_datamodule.py b/egs/librispeech/ASR/zipformer_adapter/asr_datamodule.py new file mode 120000 index 000000000..fa1b8cca3 --- /dev/null +++ b/egs/librispeech/ASR/zipformer_adapter/asr_datamodule.py @@ -0,0 +1 @@ +../tdnn_lstm_ctc/asr_datamodule.py \ No newline at end of file diff --git a/egs/librispeech/ASR/zipformer_adapter/beam_search.py b/egs/librispeech/ASR/zipformer_adapter/beam_search.py new file mode 120000 index 000000000..8554e44cc --- /dev/null +++ b/egs/librispeech/ASR/zipformer_adapter/beam_search.py @@ -0,0 +1 @@ +../pruned_transducer_stateless2/beam_search.py \ No newline at end of file diff --git a/egs/librispeech/ASR/zipformer_adapter/decode.py b/egs/librispeech/ASR/zipformer_adapter/decode.py new file mode 100755 index 000000000..91533be8d --- /dev/null +++ b/egs/librispeech/ASR/zipformer_adapter/decode.py @@ -0,0 +1,1070 @@ +#!/usr/bin/env python3 +# +# Copyright 2021-2023 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: + +- To activate the adapter (test on the target domain) +set --use-adapter True + +- To deactivate the adapter (test on the original domain) +set --use-adapter False + +(1) greedy search +./zipformer/decode.py \ + --epoch 28 \ + --avg 15 \ + --exp-dir ./zipformer/exp \ + --max-duration 600 \ + --decoding-method greedy_search + +(2) beam search (not recommended) +./zipformer/decode.py \ + --epoch 28 \ + --avg 15 \ + --exp-dir ./zipformer/exp \ + --max-duration 600 \ + --decoding-method beam_search \ + --beam-size 4 + +(3) modified beam search +./zipformer/decode.py \ + --epoch 28 \ + --avg 15 \ + --exp-dir ./zipformer/exp \ + --max-duration 600 \ + --decoding-method modified_beam_search \ + --beam-size 4 + +(4) fast beam search (one best) +./zipformer/decode.py \ + --epoch 28 \ + --avg 15 \ + --exp-dir ./zipformer/exp \ + --max-duration 600 \ + --decoding-method fast_beam_search \ + --beam 20.0 \ + --max-contexts 8 \ + --max-states 64 + +(5) fast beam search (nbest) +./zipformer/decode.py \ + --epoch 28 \ + --avg 15 \ + --exp-dir ./zipformer/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) +./zipformer/decode.py \ + --epoch 28 \ + --avg 15 \ + --exp-dir ./zipformer/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) +./zipformer/decode.py \ + --epoch 28 \ + --avg 15 \ + --exp-dir ./zipformer/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 +import os +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, + modified_beam_search_lm_rescore, + modified_beam_search_lm_rescore_LODR, + modified_beam_search_lm_shallow_fusion, + modified_beam_search_LODR, +) +from train import add_model_arguments, get_model, get_params + +from icefall import ContextGraph, LmScorer, NgramLm +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, + make_pad_mask, + 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="zipformer/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 + - modified_beam_search_LODR + - 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( + "--use-shallow-fusion", + type=str2bool, + default=False, + help="""Use neural network LM for shallow fusion. + If you want to use LODR, you will also need to set this to true + """, + ) + + parser.add_argument( + "--lm-type", + type=str, + default="rnn", + help="Type of NN lm", + choices=["rnn", "transformer"], + ) + + parser.add_argument( + "--lm-scale", + type=float, + default=0.3, + help="""The scale of the neural network LM + Used only when `--use-shallow-fusion` is set to True. + """, + ) + + parser.add_argument( + "--tokens-ngram", + type=int, + default=2, + help="""The order of the ngram lm. + """, + ) + + parser.add_argument( + "--backoff-id", + type=int, + default=500, + help="ID of the backoff symbol in the ngram LM", + ) + + parser.add_argument( + "--context-score", + type=float, + default=2, + help=""" + The bonus score of each token for the context biasing words/phrases. + Used only when --decoding-method is modified_beam_search and + modified_beam_search_LODR. + """, + ) + + parser.add_argument( + "--context-file", + type=str, + default="", + help=""" + The path of the context biasing lists, one word/phrase each line + Used only when --decoding-method is modified_beam_search and + modified_beam_search_LODR. + """, + ) + 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, + context_graph: Optional[ContextGraph] = None, + LM: Optional[LmScorer] = None, + ngram_lm=None, + ngram_lm_scale: float = 0.0, +) -> 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. + LM: + A neural network language model. + ngram_lm: + A ngram language model + ngram_lm_scale: + The scale for the ngram language model. + 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.causal: + # this seems to cause insertions at the end of the utterance if used with zipformer. + pad_len = 30 + feature_lens += pad_len + feature = torch.nn.functional.pad( + feature, + pad=(0, 0, 0, pad_len), + value=LOG_EPS, + ) + + encoder_out, encoder_out_lens = model.forward_encoder(feature, 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, + context_graph=context_graph, + ) + for hyp in sp.decode(hyp_tokens): + hyps.append(hyp.split()) + elif params.decoding_method == "modified_beam_search_lm_shallow_fusion": + hyp_tokens = modified_beam_search_lm_shallow_fusion( + model=model, + encoder_out=encoder_out, + encoder_out_lens=encoder_out_lens, + beam=params.beam_size, + LM=LM, + ) + for hyp in sp.decode(hyp_tokens): + hyps.append(hyp.split()) + elif params.decoding_method == "modified_beam_search_LODR": + hyp_tokens = modified_beam_search_LODR( + model=model, + encoder_out=encoder_out, + encoder_out_lens=encoder_out_lens, + beam=params.beam_size, + LODR_lm=ngram_lm, + LODR_lm_scale=ngram_lm_scale, + LM=LM, + context_graph=context_graph, + ) + for hyp in sp.decode(hyp_tokens): + hyps.append(hyp.split()) + elif params.decoding_method == "modified_beam_search_lm_rescore": + lm_scale_list = [0.01 * i for i in range(10, 50)] + ans_dict = modified_beam_search_lm_rescore( + model=model, + encoder_out=encoder_out, + encoder_out_lens=encoder_out_lens, + beam=params.beam_size, + LM=LM, + lm_scale_list=lm_scale_list, + ) + elif params.decoding_method == "modified_beam_search_lm_rescore_LODR": + lm_scale_list = [0.02 * i for i in range(2, 30)] + ans_dict = modified_beam_search_lm_rescore_LODR( + model=model, + encoder_out=encoder_out, + encoder_out_lens=encoder_out_lens, + beam=params.beam_size, + LM=LM, + LODR_lm=ngram_lm, + sp=sp, + lm_scale_list=lm_scale_list, + ) + 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} + elif "modified_beam_search" in params.decoding_method: + prefix = f"beam_size_{params.beam_size}" + if params.decoding_method in ( + "modified_beam_search_lm_rescore", + "modified_beam_search_lm_rescore_LODR", + ): + ans = dict() + assert ans_dict is not None + for key, hyps in ans_dict.items(): + hyps = [sp.decode(hyp).split() for hyp in hyps] + ans[f"{prefix}_{key}"] = hyps + return ans + else: + if params.has_contexts: + prefix += f"-context-score-{params.context_score}" + return {prefix: 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, + context_graph: Optional[ContextGraph] = None, + LM: Optional[LmScorer] = None, + ngram_lm=None, + ngram_lm_scale: float = 0.0, +) -> 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, + context_graph=context_graph, + word_table=word_table, + batch=batch, + LM=LM, + ngram_lm=ngram_lm, + ngram_lm_scale=ngram_lm_scale, + ) + + 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) + LmScorer.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", + "modified_beam_search_LODR", + "modified_beam_search_lm_shallow_fusion", + "modified_beam_search_lm_rescore", + "modified_beam_search_lm_rescore_LODR", + ) + params.res_dir = params.exp_dir / params.decoding_method + + if os.path.exists(params.context_file): + params.has_contexts = True + else: + params.has_contexts = False + + 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.causal: + assert ( + "," not in params.chunk_size + ), "chunk_size should be one value in decoding." + assert ( + "," not in params.left_context_frames + ), "left_context_frames should be one value in decoding." + params.suffix += f"-chunk-{params.chunk_size}" + params.suffix += f"-left-context-{params.left_context_frames}" + + 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}" + if params.decoding_method in ( + "modified_beam_search", + "modified_beam_search_LODR", + ): + if params.has_contexts: + params.suffix += f"-context-score-{params.context_score}" + else: + params.suffix += f"-context-{params.context_size}" + params.suffix += f"-max-sym-per-frame-{params.max_sym_per_frame}" + + if params.use_shallow_fusion: + params.suffix += f"-{params.lm_type}-lm-scale-{params.lm_scale}" + + if "LODR" in params.decoding_method: + params.suffix += ( + f"-LODR-{params.tokens_ngram}gram-scale-{params.ngram_lm_scale}" + ) + + 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) + + # and are defined in local/train_bpe_model.py + params.blank_id = sp.piece_to_id("") + params.unk_id = sp.piece_to_id("") + params.vocab_size = sp.get_piece_size() + + logging.info(params) + + logging.info("About to create model") + model = get_model(params) + import pdb + + pdb.set_trace() + + 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), strict=False + ) + elif params.avg == 1: + load_checkpoint( + f"{params.exp_dir}/epoch-{params.epoch}.pt", model, strict=False + ) + 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), strict=False + ) + 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, + ), + strict=False, + ) + 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, + ), + strict=False, + ) + + model.to(device) + model.eval() + + # only load the neural network LM if required + if params.use_shallow_fusion or params.decoding_method in ( + "modified_beam_search_lm_rescore", + "modified_beam_search_lm_rescore_LODR", + "modified_beam_search_lm_shallow_fusion", + "modified_beam_search_LODR", + ): + LM = LmScorer( + lm_type=params.lm_type, + params=params, + device=device, + lm_scale=params.lm_scale, + ) + LM.to(device) + LM.eval() + else: + LM = None + + # only load N-gram LM when needed + if params.decoding_method == "modified_beam_search_lm_rescore_LODR": + try: + import kenlm + except ImportError: + print("Please install kenlm first. You can use") + print(" pip install https://github.com/kpu/kenlm/archive/master.zip") + print("to install it") + import sys + + sys.exit(-1) + ngram_file_name = str(params.lang_dir / f"{params.tokens_ngram}gram.arpa") + logging.info(f"lm filename: {ngram_file_name}") + ngram_lm = kenlm.Model(ngram_file_name) + ngram_lm_scale = None # use a list to search + + elif params.decoding_method == "modified_beam_search_LODR": + lm_filename = f"{params.tokens_ngram}gram.fst.txt" + logging.info(f"Loading token level lm: {lm_filename}") + ngram_lm = NgramLm( + str(params.lang_dir / lm_filename), + backoff_id=params.backoff_id, + is_binary=False, + ) + logging.info(f"num states: {ngram_lm.lm.num_states}") + ngram_lm_scale = params.ngram_lm_scale + else: + ngram_lm = None + ngram_lm_scale = None + + 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 + + if "modified_beam_search" in params.decoding_method: + if os.path.exists(params.context_file): + contexts = [] + for line in open(params.context_file).readlines(): + contexts.append((sp.encode(line.strip()), 0.0)) + context_graph = ContextGraph(params.context_score) + context_graph.build(contexts) + else: + context_graph = None + else: + context_graph = 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, + context_graph=context_graph, + LM=LM, + ngram_lm=ngram_lm, + ngram_lm_scale=ngram_lm_scale, + ) + + save_results( + params=params, + test_set_name=test_set, + results_dict=results_dict, + ) + + logging.info("Done!") + + +if __name__ == "__main__": + main() diff --git a/egs/librispeech/ASR/zipformer_adapter/decode_gigaspeech.py b/egs/librispeech/ASR/zipformer_adapter/decode_gigaspeech.py new file mode 100755 index 000000000..bbc582f50 --- /dev/null +++ b/egs/librispeech/ASR/zipformer_adapter/decode_gigaspeech.py @@ -0,0 +1,1115 @@ +#!/usr/bin/env python3 +# +# Copyright 2021-2023 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 +./zipformer/decode.py \ + --epoch 28 \ + --avg 15 \ + --exp-dir ./zipformer/exp \ + --max-duration 600 \ + --decoding-method greedy_search + +(2) beam search (not recommended) +./zipformer/decode.py \ + --epoch 28 \ + --avg 15 \ + --exp-dir ./zipformer/exp \ + --max-duration 600 \ + --decoding-method beam_search \ + --beam-size 4 + +(3) modified beam search +./zipformer/decode.py \ + --epoch 28 \ + --avg 15 \ + --exp-dir ./zipformer/exp \ + --max-duration 600 \ + --decoding-method modified_beam_search \ + --beam-size 4 + +(4) fast beam search (one best) +./zipformer/decode.py \ + --epoch 28 \ + --avg 15 \ + --exp-dir ./zipformer/exp \ + --max-duration 600 \ + --decoding-method fast_beam_search \ + --beam 20.0 \ + --max-contexts 8 \ + --max-states 64 + +(5) fast beam search (nbest) +./zipformer/decode.py \ + --epoch 28 \ + --avg 15 \ + --exp-dir ./zipformer/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) +./zipformer/decode.py \ + --epoch 28 \ + --avg 15 \ + --exp-dir ./zipformer/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) +./zipformer/decode.py \ + --epoch 28 \ + --avg 15 \ + --exp-dir ./zipformer/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 +import os +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, + modified_beam_search_lm_rescore, + modified_beam_search_lm_rescore_LODR, + modified_beam_search_lm_shallow_fusion, + modified_beam_search_LODR, +) +from train import add_finetune_arguments, add_model_arguments, get_model, get_params + +from icefall import ContextGraph, LmScorer, NgramLm +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, + make_pad_mask, + setup_logger, + store_transcripts, + str2bool, + write_error_stats, +) + +LOG_EPS = math.log(1e-10) + +conversational_filler = [ + "UH", + "UHH", + "UM", + "EH", + "MM", + "HM", + "AH", + "HUH", + "HA", + "ER", + "OOF", + "HEE", + "ACH", + "EEE", + "EW", +] +unk_tags = ["", ""] +gigaspeech_punctuations = [ + "", + "", + "", + "", +] +gigaspeech_garbage_utterance_tags = ["", "", "", ""] +non_scoring_words = ( + conversational_filler + + unk_tags + + gigaspeech_punctuations + + gigaspeech_garbage_utterance_tags +) + + +def asr_text_post_processing(text: str) -> str: + # 1. convert to uppercase + text = text.upper() + + # 2. remove hyphen + # "E-COMMERCE" -> "E COMMERCE", "STATE-OF-THE-ART" -> "STATE OF THE ART" + text = text.replace("-", " ") + + # 3. remove non-scoring words from evaluation + remaining_words = [] + for word in text.split(): + if word in non_scoring_words: + continue + remaining_words.append(word) + + return " ".join(remaining_words) + + +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="zipformer/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 + - modified_beam_search_LODR + - 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( + "--use-shallow-fusion", + type=str2bool, + default=False, + help="""Use neural network LM for shallow fusion. + If you want to use LODR, you will also need to set this to true + """, + ) + + parser.add_argument( + "--lm-type", + type=str, + default="rnn", + help="Type of NN lm", + choices=["rnn", "transformer"], + ) + + parser.add_argument( + "--lm-scale", + type=float, + default=0.3, + help="""The scale of the neural network LM + Used only when `--use-shallow-fusion` is set to True. + """, + ) + + parser.add_argument( + "--tokens-ngram", + type=int, + default=2, + help="""The order of the ngram lm. + """, + ) + + parser.add_argument( + "--backoff-id", + type=int, + default=500, + help="ID of the backoff symbol in the ngram LM", + ) + + parser.add_argument( + "--context-score", + type=float, + default=2, + help=""" + The bonus score of each token for the context biasing words/phrases. + Used only when --decoding-method is modified_beam_search and + modified_beam_search_LODR. + """, + ) + + parser.add_argument( + "--context-file", + type=str, + default="", + help=""" + The path of the context biasing lists, one word/phrase each line + Used only when --decoding-method is modified_beam_search and + modified_beam_search_LODR. + """, + ) + add_model_arguments(parser) + add_finetune_arguments(parser) + + return parser + + +def post_processing( + results: List[Tuple[str, List[str], List[str]]], +) -> List[Tuple[str, List[str], List[str]]]: + new_results = [] + for key, ref, hyp in results: + new_ref = asr_text_post_processing(" ".join(ref)).split() + new_hyp = asr_text_post_processing(" ".join(hyp)).split() + new_results.append((key, new_ref, new_hyp)) + return new_results + + +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, + context_graph: Optional[ContextGraph] = None, + LM: Optional[LmScorer] = None, + ngram_lm=None, + ngram_lm_scale: float = 0.0, +) -> 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. + LM: + A neural network language model. + ngram_lm: + A ngram language model + ngram_lm_scale: + The scale for the ngram language model. + 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.causal: + # this seems to cause insertions at the end of the utterance if used with zipformer. + pad_len = 30 + feature_lens += pad_len + feature = torch.nn.functional.pad( + feature, + pad=(0, 0, 0, pad_len), + value=LOG_EPS, + ) + + encoder_out, encoder_out_lens = model.forward_encoder(feature, 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, + context_graph=context_graph, + ) + for hyp in sp.decode(hyp_tokens): + hyps.append(hyp.split()) + elif params.decoding_method == "modified_beam_search_lm_shallow_fusion": + hyp_tokens = modified_beam_search_lm_shallow_fusion( + model=model, + encoder_out=encoder_out, + encoder_out_lens=encoder_out_lens, + beam=params.beam_size, + LM=LM, + ) + for hyp in sp.decode(hyp_tokens): + hyps.append(hyp.split()) + elif params.decoding_method == "modified_beam_search_LODR": + hyp_tokens = modified_beam_search_LODR( + model=model, + encoder_out=encoder_out, + encoder_out_lens=encoder_out_lens, + beam=params.beam_size, + LODR_lm=ngram_lm, + LODR_lm_scale=ngram_lm_scale, + LM=LM, + context_graph=context_graph, + ) + for hyp in sp.decode(hyp_tokens): + hyps.append(hyp.split()) + elif params.decoding_method == "modified_beam_search_lm_rescore": + lm_scale_list = [0.01 * i for i in range(10, 50)] + ans_dict = modified_beam_search_lm_rescore( + model=model, + encoder_out=encoder_out, + encoder_out_lens=encoder_out_lens, + beam=params.beam_size, + LM=LM, + lm_scale_list=lm_scale_list, + ) + elif params.decoding_method == "modified_beam_search_lm_rescore_LODR": + lm_scale_list = [0.02 * i for i in range(2, 30)] + ans_dict = modified_beam_search_lm_rescore_LODR( + model=model, + encoder_out=encoder_out, + encoder_out_lens=encoder_out_lens, + beam=params.beam_size, + LM=LM, + LODR_lm=ngram_lm, + sp=sp, + lm_scale_list=lm_scale_list, + ) + 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} + elif "modified_beam_search" in params.decoding_method: + prefix = f"beam_size_{params.beam_size}" + if params.decoding_method in ( + "modified_beam_search_lm_rescore", + "modified_beam_search_lm_rescore_LODR", + ): + ans = dict() + assert ans_dict is not None + for key, hyps in ans_dict.items(): + hyps = [sp.decode(hyp).split() for hyp in hyps] + ans[f"{prefix}_{key}"] = hyps + return ans + else: + if params.has_contexts: + prefix += f"-context-score-{params.context_score}" + return {prefix: 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, + context_graph: Optional[ContextGraph] = None, + LM: Optional[LmScorer] = None, + ngram_lm=None, + ngram_lm_scale: float = 0.0, +) -> 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, + context_graph=context_graph, + word_table=word_table, + batch=batch, + LM=LM, + ngram_lm=ngram_lm, + ngram_lm_scale=ngram_lm_scale, + ) + + 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 = post_processing(results) + 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) + LmScorer.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", + "modified_beam_search_LODR", + "modified_beam_search_lm_shallow_fusion", + "modified_beam_search_lm_rescore", + "modified_beam_search_lm_rescore_LODR", + ) + params.res_dir = params.exp_dir / (params.decoding_method + "_giga") + + if os.path.exists(params.context_file): + params.has_contexts = True + else: + params.has_contexts = False + + 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.causal: + assert ( + "," not in params.chunk_size + ), "chunk_size should be one value in decoding." + assert ( + "," not in params.left_context_frames + ), "left_context_frames should be one value in decoding." + params.suffix += f"-chunk-{params.chunk_size}" + params.suffix += f"-left-context-{params.left_context_frames}" + + 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}" + if params.decoding_method in ( + "modified_beam_search", + "modified_beam_search_LODR", + ): + if params.has_contexts: + params.suffix += f"-context-score-{params.context_score}" + else: + params.suffix += f"-context-{params.context_size}" + params.suffix += f"-max-sym-per-frame-{params.max_sym_per_frame}" + + if params.use_shallow_fusion: + params.suffix += f"-{params.lm_type}-lm-scale-{params.lm_scale}" + + if "LODR" in params.decoding_method: + params.suffix += ( + f"-LODR-{params.tokens_ngram}gram-scale-{params.ngram_lm_scale}" + ) + + 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) + + # and are defined in local/train_bpe_model.py + params.blank_id = sp.piece_to_id("") + params.unk_id = sp.piece_to_id("") + params.vocab_size = sp.get_piece_size() + + logging.info(params) + + logging.info("About to create model") + model = get_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() + + # only load the neural network LM if required + if params.use_shallow_fusion or params.decoding_method in ( + "modified_beam_search_lm_rescore", + "modified_beam_search_lm_rescore_LODR", + "modified_beam_search_lm_shallow_fusion", + "modified_beam_search_LODR", + ): + LM = LmScorer( + lm_type=params.lm_type, + params=params, + device=device, + lm_scale=params.lm_scale, + ) + LM.to(device) + LM.eval() + else: + LM = None + + # only load N-gram LM when needed + if params.decoding_method == "modified_beam_search_lm_rescore_LODR": + try: + import kenlm + except ImportError: + print("Please install kenlm first. You can use") + print(" pip install https://github.com/kpu/kenlm/archive/master.zip") + print("to install it") + import sys + + sys.exit(-1) + ngram_file_name = str(params.lang_dir / f"{params.tokens_ngram}gram.arpa") + logging.info(f"lm filename: {ngram_file_name}") + ngram_lm = kenlm.Model(ngram_file_name) + ngram_lm_scale = None # use a list to search + + elif params.decoding_method == "modified_beam_search_LODR": + lm_filename = f"{params.tokens_ngram}gram.fst.txt" + logging.info(f"Loading token level lm: {lm_filename}") + ngram_lm = NgramLm( + str(params.lang_dir / lm_filename), + backoff_id=params.backoff_id, + is_binary=False, + ) + logging.info(f"num states: {ngram_lm.lm.num_states}") + ngram_lm_scale = params.ngram_lm_scale + else: + ngram_lm = None + ngram_lm_scale = None + + 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 + + if "modified_beam_search" in params.decoding_method: + if os.path.exists(params.context_file): + contexts = [] + for line in open(params.context_file).readlines(): + contexts.append((sp.encode(line.strip()), 0.0)) + context_graph = ContextGraph(params.context_score) + context_graph.build(contexts) + else: + context_graph = None + else: + context_graph = 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) + + gigaspeech_dev_cuts = librispeech.gigaspeech_dev_cuts() + gigaspeech_test_cuts = librispeech.gigaspeech_test_cuts() + + dev_dl = librispeech.test_dataloaders(gigaspeech_dev_cuts) + test_dl = librispeech.test_dataloaders(gigaspeech_test_cuts) + + test_sets = ["dev", "test"] + test_dl = [dev_dl, test_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, + context_graph=context_graph, + LM=LM, + ngram_lm=ngram_lm, + ngram_lm_scale=ngram_lm_scale, + ) + + save_results( + params=params, + test_set_name=test_set, + results_dict=results_dict, + ) + + logging.info("Done!") + + +if __name__ == "__main__": + main() diff --git a/egs/librispeech/ASR/zipformer_adapter/decoder.py b/egs/librispeech/ASR/zipformer_adapter/decoder.py new file mode 120000 index 000000000..cab465d2b --- /dev/null +++ b/egs/librispeech/ASR/zipformer_adapter/decoder.py @@ -0,0 +1 @@ +../zipformer/decoder.py \ No newline at end of file diff --git a/egs/librispeech/ASR/zipformer_adapter/encoder_interface.py b/egs/librispeech/ASR/zipformer_adapter/encoder_interface.py new file mode 120000 index 000000000..aa5d0217a --- /dev/null +++ b/egs/librispeech/ASR/zipformer_adapter/encoder_interface.py @@ -0,0 +1 @@ +../transducer_stateless/encoder_interface.py \ No newline at end of file diff --git a/egs/librispeech/ASR/zipformer_adapter/export-onnx.py b/egs/librispeech/ASR/zipformer_adapter/export-onnx.py new file mode 100755 index 000000000..ea29e8159 --- /dev/null +++ b/egs/librispeech/ASR/zipformer_adapter/export-onnx.py @@ -0,0 +1,621 @@ +#!/usr/bin/env python3 +# +# Copyright 2023 Xiaomi Corporation (Author: Fangjun Kuang, Wei Kang) +# Copyright 2023 Danqing Fu (danqing.fu@gmail.com) + +""" +This script exports a transducer model from PyTorch to ONNX. + +We use the pre-trained model from +https://huggingface.co/Zengwei/icefall-asr-librispeech-zipformer-2023-05-15 +as an example to show how to use this file. + +1. Download the pre-trained model + +cd egs/librispeech/ASR + +repo_url=https://huggingface.co/Zengwei/icefall-asr-librispeech-zipformer-2023-05-15 +GIT_LFS_SKIP_SMUDGE=1 git clone $repo_url +repo=$(basename $repo_url) + +pushd $repo +git lfs pull --include "exp/pretrained.pt" + +cd exp +ln -s pretrained.pt epoch-99.pt +popd + +2. Export the model to ONNX + +./zipformer/export-onnx.py \ + --tokens $repo/data/lang_bpe_500/tokens.txt \ + --use-averaged-model 0 \ + --epoch 99 \ + --avg 1 \ + --exp-dir $repo/exp \ + --num-encoder-layers "2,2,3,4,3,2" \ + --downsampling-factor "1,2,4,8,4,2" \ + --feedforward-dim "512,768,1024,1536,1024,768" \ + --num-heads "4,4,4,8,4,4" \ + --encoder-dim "192,256,384,512,384,256" \ + --query-head-dim 32 \ + --value-head-dim 12 \ + --pos-head-dim 4 \ + --pos-dim 48 \ + --encoder-unmasked-dim "192,192,256,256,256,192" \ + --cnn-module-kernel "31,31,15,15,15,31" \ + --decoder-dim 512 \ + --joiner-dim 512 \ + --causal False \ + --chunk-size "16,32,64,-1" \ + --left-context-frames "64,128,256,-1" + +It will generate the following 3 files inside $repo/exp: + + - encoder-epoch-99-avg-1.onnx + - decoder-epoch-99-avg-1.onnx + - joiner-epoch-99-avg-1.onnx + +See ./onnx_pretrained.py and ./onnx_check.py for how to +use the exported ONNX models. +""" + +import argparse +import logging +from pathlib import Path +from typing import Dict, Tuple + +import k2 +import onnx +import torch +import torch.nn as nn +from decoder import Decoder +from onnxruntime.quantization import QuantType, quantize_dynamic +from scaling_converter import convert_scaled_to_non_scaled +from train import add_finetune_arguments, add_model_arguments, get_model, get_params +from zipformer import Zipformer2 + +from icefall.checkpoint import ( + average_checkpoints, + average_checkpoints_with_averaged_model, + find_checkpoints, + load_checkpoint, +) +from icefall.utils import make_pad_mask, num_tokens, str2bool + + +def get_parser(): + parser = argparse.ArgumentParser( + formatter_class=argparse.ArgumentDefaultsHelpFormatter + ) + + parser.add_argument( + "--epoch", + type=int, + default=28, + help="""It specifies the checkpoint to use for averaging. + 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' 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="zipformer/exp", + help="""It specifies the directory where all training related + files, e.g., checkpoints, log, etc, are saved + """, + ) + + parser.add_argument( + "--tokens", + type=str, + default="data/lang_bpe_500/tokens.txt", + help="Path to the tokens.txt", + ) + + 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) + add_finetune_arguments(parser) + + return parser + + +def add_meta_data(filename: str, meta_data: Dict[str, str]): + """Add meta data to an ONNX model. It is changed in-place. + + Args: + filename: + Filename of the ONNX model to be changed. + meta_data: + Key-value pairs. + """ + model = onnx.load(filename) + for key, value in meta_data.items(): + meta = model.metadata_props.add() + meta.key = key + meta.value = value + + onnx.save(model, filename) + + +class OnnxEncoder(nn.Module): + """A wrapper for Zipformer and the encoder_proj from the joiner""" + + def __init__( + self, encoder: Zipformer2, encoder_embed: nn.Module, encoder_proj: nn.Linear + ): + """ + Args: + encoder: + A Zipformer encoder. + encoder_proj: + The projection layer for encoder from the joiner. + """ + super().__init__() + self.encoder = encoder + self.encoder_embed = encoder_embed + self.encoder_proj = encoder_proj + + def forward( + self, + x: torch.Tensor, + x_lens: torch.Tensor, + ) -> Tuple[torch.Tensor, torch.Tensor]: + """Please see the help information of Zipformer.forward + + Args: + x: + A 3-D tensor of shape (N, T, C) + x_lens: + A 1-D tensor of shape (N,). Its dtype is torch.int64 + Returns: + Return a tuple containing: + - encoder_out, A 3-D tensor of shape (N, T', joiner_dim) + - encoder_out_lens, A 1-D tensor of shape (N,) + """ + x, x_lens = self.encoder_embed(x, x_lens) + src_key_padding_mask = make_pad_mask(x_lens) + x = x.permute(1, 0, 2) + encoder_out, encoder_out_lens = self.encoder(x, x_lens, src_key_padding_mask) + encoder_out = encoder_out.permute(1, 0, 2) + encoder_out = self.encoder_proj(encoder_out) + # Now encoder_out is of shape (N, T, joiner_dim) + + return encoder_out, encoder_out_lens + + +class OnnxDecoder(nn.Module): + """A wrapper for Decoder and the decoder_proj from the joiner""" + + def __init__(self, decoder: Decoder, decoder_proj: nn.Linear): + super().__init__() + self.decoder = decoder + self.decoder_proj = decoder_proj + + def forward(self, y: torch.Tensor) -> torch.Tensor: + """ + Args: + y: + A 2-D tensor of shape (N, context_size). + Returns + Return a 2-D tensor of shape (N, joiner_dim) + """ + need_pad = False + decoder_output = self.decoder(y, need_pad=need_pad) + decoder_output = decoder_output.squeeze(1) + output = self.decoder_proj(decoder_output) + + return output + + +class OnnxJoiner(nn.Module): + """A wrapper for the joiner""" + + def __init__(self, output_linear: nn.Linear): + super().__init__() + self.output_linear = output_linear + + def forward( + self, + encoder_out: torch.Tensor, + decoder_out: torch.Tensor, + ) -> torch.Tensor: + """ + Args: + encoder_out: + A 2-D tensor of shape (N, joiner_dim) + decoder_out: + A 2-D tensor of shape (N, joiner_dim) + Returns: + Return a 2-D tensor of shape (N, vocab_size) + """ + logit = encoder_out + decoder_out + logit = self.output_linear(torch.tanh(logit)) + return logit + + +def export_encoder_model_onnx( + encoder_model: OnnxEncoder, + encoder_filename: str, + opset_version: int = 11, +) -> None: + """Export the given encoder model to ONNX format. + The exported model has two inputs: + + - x, a tensor of shape (N, T, C); dtype is torch.float32 + - x_lens, a tensor of shape (N,); dtype is torch.int64 + + and it has two outputs: + + - encoder_out, a tensor of shape (N, T', joiner_dim) + - encoder_out_lens, a tensor of shape (N,) + + Args: + encoder_model: + The input encoder model + encoder_filename: + The filename to save the exported ONNX model. + opset_version: + The opset version to use. + """ + x = torch.zeros(1, 100, 80, dtype=torch.float32) + x_lens = torch.tensor([100], dtype=torch.int64) + + encoder_model = torch.jit.trace(encoder_model, (x, x_lens)) + + torch.onnx.export( + encoder_model, + (x, x_lens), + encoder_filename, + verbose=False, + opset_version=opset_version, + input_names=["x", "x_lens"], + output_names=["encoder_out", "encoder_out_lens"], + dynamic_axes={ + "x": {0: "N", 1: "T"}, + "x_lens": {0: "N"}, + "encoder_out": {0: "N", 1: "T"}, + "encoder_out_lens": {0: "N"}, + }, + ) + + meta_data = { + "model_type": "zipformer2", + "version": "1", + "model_author": "k2-fsa", + "comment": "non-streaming zipformer2", + } + logging.info(f"meta_data: {meta_data}") + + add_meta_data(filename=encoder_filename, meta_data=meta_data) + + +def export_decoder_model_onnx( + decoder_model: OnnxDecoder, + decoder_filename: str, + opset_version: int = 11, +) -> None: + """Export the decoder model to ONNX format. + + The exported model has one input: + + - y: a torch.int64 tensor of shape (N, decoder_model.context_size) + + and has one output: + + - decoder_out: a torch.float32 tensor of shape (N, joiner_dim) + + Args: + decoder_model: + The decoder model to be exported. + decoder_filename: + Filename to save the exported ONNX model. + opset_version: + The opset version to use. + """ + context_size = decoder_model.decoder.context_size + vocab_size = decoder_model.decoder.vocab_size + + y = torch.zeros(10, context_size, dtype=torch.int64) + decoder_model = torch.jit.script(decoder_model) + torch.onnx.export( + decoder_model, + y, + decoder_filename, + verbose=False, + opset_version=opset_version, + input_names=["y"], + output_names=["decoder_out"], + dynamic_axes={ + "y": {0: "N"}, + "decoder_out": {0: "N"}, + }, + ) + + meta_data = { + "context_size": str(context_size), + "vocab_size": str(vocab_size), + } + add_meta_data(filename=decoder_filename, meta_data=meta_data) + + +def export_joiner_model_onnx( + joiner_model: nn.Module, + joiner_filename: str, + opset_version: int = 11, +) -> None: + """Export the joiner model to ONNX format. + The exported joiner model has two inputs: + + - encoder_out: a tensor of shape (N, joiner_dim) + - decoder_out: a tensor of shape (N, joiner_dim) + + and produces one output: + + - logit: a tensor of shape (N, vocab_size) + """ + joiner_dim = joiner_model.output_linear.weight.shape[1] + logging.info(f"joiner dim: {joiner_dim}") + + projected_encoder_out = torch.rand(11, joiner_dim, dtype=torch.float32) + projected_decoder_out = torch.rand(11, joiner_dim, dtype=torch.float32) + + torch.onnx.export( + joiner_model, + (projected_encoder_out, projected_decoder_out), + joiner_filename, + verbose=False, + opset_version=opset_version, + input_names=[ + "encoder_out", + "decoder_out", + ], + output_names=["logit"], + dynamic_axes={ + "encoder_out": {0: "N"}, + "decoder_out": {0: "N"}, + "logit": {0: "N"}, + }, + ) + meta_data = { + "joiner_dim": str(joiner_dim), + } + add_meta_data(filename=joiner_filename, meta_data=meta_data) + + +@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}") + + token_table = k2.SymbolTable.from_file(params.tokens) + params.blank_id = token_table[""] + params.vocab_size = num_tokens(token_table) + 1 + + logging.info(params) + + logging.info("About to create model") + model = get_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() + + convert_scaled_to_non_scaled(model, inplace=True, is_onnx=True) + + encoder = OnnxEncoder( + encoder=model.encoder, + encoder_embed=model.encoder_embed, + encoder_proj=model.joiner.encoder_proj, + ) + + decoder = OnnxDecoder( + decoder=model.decoder, + decoder_proj=model.joiner.decoder_proj, + ) + + joiner = OnnxJoiner(output_linear=model.joiner.output_linear) + + encoder_num_param = sum([p.numel() for p in encoder.parameters()]) + decoder_num_param = sum([p.numel() for p in decoder.parameters()]) + joiner_num_param = sum([p.numel() for p in joiner.parameters()]) + total_num_param = encoder_num_param + decoder_num_param + joiner_num_param + logging.info(f"encoder parameters: {encoder_num_param}") + logging.info(f"decoder parameters: {decoder_num_param}") + logging.info(f"joiner parameters: {joiner_num_param}") + logging.info(f"total parameters: {total_num_param}") + + if params.iter > 0: + suffix = f"iter-{params.iter}" + else: + suffix = f"epoch-{params.epoch}" + + suffix += f"-avg-{params.avg}" + + opset_version = 13 + + logging.info("Exporting encoder") + encoder_filename = params.exp_dir / f"encoder-{suffix}.onnx" + export_encoder_model_onnx( + encoder, + encoder_filename, + opset_version=opset_version, + ) + logging.info(f"Exported encoder to {encoder_filename}") + + logging.info("Exporting decoder") + decoder_filename = params.exp_dir / f"decoder-{suffix}.onnx" + export_decoder_model_onnx( + decoder, + decoder_filename, + opset_version=opset_version, + ) + logging.info(f"Exported decoder to {decoder_filename}") + + logging.info("Exporting joiner") + joiner_filename = params.exp_dir / f"joiner-{suffix}.onnx" + export_joiner_model_onnx( + joiner, + joiner_filename, + opset_version=opset_version, + ) + logging.info(f"Exported joiner to {joiner_filename}") + + # Generate int8 quantization models + # See https://onnxruntime.ai/docs/performance/model-optimizations/quantization.html#data-type-selection + + logging.info("Generate int8 quantization models") + + encoder_filename_int8 = params.exp_dir / f"encoder-{suffix}.int8.onnx" + quantize_dynamic( + model_input=encoder_filename, + model_output=encoder_filename_int8, + op_types_to_quantize=["MatMul"], + weight_type=QuantType.QInt8, + ) + + decoder_filename_int8 = params.exp_dir / f"decoder-{suffix}.int8.onnx" + quantize_dynamic( + model_input=decoder_filename, + model_output=decoder_filename_int8, + op_types_to_quantize=["MatMul", "Gather"], + weight_type=QuantType.QInt8, + ) + + joiner_filename_int8 = params.exp_dir / f"joiner-{suffix}.int8.onnx" + quantize_dynamic( + model_input=joiner_filename, + model_output=joiner_filename_int8, + op_types_to_quantize=["MatMul"], + weight_type=QuantType.QInt8, + ) + + +if __name__ == "__main__": + formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s" + logging.basicConfig(format=formatter, level=logging.INFO) + main() diff --git a/egs/librispeech/ASR/zipformer_adapter/joiner.py b/egs/librispeech/ASR/zipformer_adapter/joiner.py new file mode 120000 index 000000000..444cb5f15 --- /dev/null +++ b/egs/librispeech/ASR/zipformer_adapter/joiner.py @@ -0,0 +1 @@ +../zipformer/joiner.py \ No newline at end of file diff --git a/egs/librispeech/ASR/zipformer_adapter/model.py b/egs/librispeech/ASR/zipformer_adapter/model.py new file mode 120000 index 000000000..0c6fe6112 --- /dev/null +++ b/egs/librispeech/ASR/zipformer_adapter/model.py @@ -0,0 +1 @@ +../zipformer/model.py \ No newline at end of file diff --git a/egs/librispeech/ASR/zipformer_adapter/onnx_decode.py b/egs/librispeech/ASR/zipformer_adapter/onnx_decode.py new file mode 100755 index 000000000..e3f7ce85a --- /dev/null +++ b/egs/librispeech/ASR/zipformer_adapter/onnx_decode.py @@ -0,0 +1,386 @@ +#!/usr/bin/env python3 +# +# Copyright 2021-2023 Xiaomi Corporation (Author: Fangjun Kuang, +# Zengwei Yao, +# Xiaoyu Yang) +# +# 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 ONNX exported models and uses them to decode the test sets. + +We use the pre-trained model from +https://huggingface.co/Zengwei/icefall-asr-librispeech-zipformer-2023-05-15 +as an example to show how to use this file. + +1. Download the pre-trained model + +cd egs/librispeech/ASR + +repo_url=https://huggingface.co/Zengwei/icefall-asr-librispeech-zipformer-2023-05-15 +GIT_LFS_SKIP_SMUDGE=1 git clone $repo_url +repo=$(basename $repo_url) + +pushd $repo +git lfs pull --include "data/lang_bpe_500/bpe.model" +git lfs pull --include "exp/pretrained.pt" + +cd exp +ln -s pretrained.pt epoch-99.pt +popd + +2. Export the model to ONNX + +./zipformer/export-onnx.py \ + --tokens $repo/data/lang_bpe_500/tokens.txt \ + --use-averaged-model 0 \ + --epoch 99 \ + --avg 1 \ + --exp-dir $repo/exp \ + --causal False + +It will generate the following 3 files inside $repo/exp: + + - encoder-epoch-99-avg-1.onnx + - decoder-epoch-99-avg-1.onnx + - joiner-epoch-99-avg-1.onnx + +2. Run this file + +./zipformer/onnx_decode.py \ + --exp-dir $repo/exp \ + --max-duration 600 \ + --encoder-model-filename $repo/exp/encoder-epoch-99-avg-1.onnx \ + --decoder-model-filename $repo/exp/decoder-epoch-99-avg-1.onnx \ + --joiner-model-filename $repo/exp/joiner-epoch-99-avg-1.onnx \ + --tokens $repo/data/lang_bpe_500/tokens.txt \ +""" + + +import argparse +import logging +import time +from pathlib import Path +from typing import List, Tuple + +import torch +import torch.nn as nn +from asr_datamodule import LibriSpeechAsrDataModule +from k2 import SymbolTable +from onnx_pretrained import OnnxModel, greedy_search + +from icefall.utils import setup_logger, store_transcripts, write_error_stats + +conversational_filler = [ + "UH", + "UHH", + "UM", + "EH", + "MM", + "HM", + "AH", + "HUH", + "HA", + "ER", + "OOF", + "HEE", + "ACH", + "EEE", + "EW", +] +unk_tags = ["", ""] +gigaspeech_punctuations = [ + "", + "", + "", + "", +] +gigaspeech_garbage_utterance_tags = ["", "", "", ""] +non_scoring_words = ( + conversational_filler + + unk_tags + + gigaspeech_punctuations + + gigaspeech_garbage_utterance_tags +) + + +def asr_text_post_processing(text: str) -> str: + # 1. convert to uppercase + text = text.upper() + + # 2. remove hyphen + # "E-COMMERCE" -> "E COMMERCE", "STATE-OF-THE-ART" -> "STATE OF THE ART" + text = text.replace("-", " ") + + # 3. remove non-scoring words from evaluation + remaining_words = [] + for word in text.split(): + if word in non_scoring_words: + continue + remaining_words.append(word) + + return " ".join(remaining_words) + + +def get_parser(): + parser = argparse.ArgumentParser( + formatter_class=argparse.ArgumentDefaultsHelpFormatter + ) + + parser.add_argument( + "--encoder-model-filename", + type=str, + required=True, + help="Path to the encoder onnx model. ", + ) + + parser.add_argument( + "--decoder-model-filename", + type=str, + required=True, + help="Path to the decoder onnx model. ", + ) + + parser.add_argument( + "--joiner-model-filename", + type=str, + required=True, + help="Path to the joiner onnx model. ", + ) + + parser.add_argument( + "--exp-dir", + type=str, + default="zipformer/exp", + help="The experiment dir", + ) + + parser.add_argument( + "--tokens", + type=str, + help="""Path to tokens.txt.""", + ) + + parser.add_argument( + "--decoding-method", + type=str, + default="greedy_search", + help="Valid values are greedy_search and modified_beam_search", + ) + + return parser + + +def post_processing( + results: List[Tuple[str, List[str], List[str]]], +) -> List[Tuple[str, List[str], List[str]]]: + new_results = [] + for key, ref, hyp in results: + new_ref = asr_text_post_processing(" ".join(ref)).split() + new_hyp = asr_text_post_processing(" ".join(hyp)).split() + new_results.append((key, new_ref, new_hyp)) + return new_results + + +def decode_one_batch( + model: OnnxModel, token_table: SymbolTable, batch: dict +) -> List[List[str]]: + """Decode one batch and return the result. + Currently it only greedy_search is supported. + + Args: + model: + The neural model. + token_table: + The token table. + batch: + It is the return value from iterating + `lhotse.dataset.K2SpeechRecognitionDataset`. See its documentation + for the format of the `batch`. + + Returns: + Return the decoded results for each utterance. + """ + feature = batch["inputs"] + assert feature.ndim == 3 + # at entry, feature is (N, T, C) + + supervisions = batch["supervisions"] + feature_lens = supervisions["num_frames"].to(dtype=torch.int64) + + encoder_out, encoder_out_lens = model.run_encoder(x=feature, x_lens=feature_lens) + + hyps = greedy_search( + model=model, encoder_out=encoder_out, encoder_out_lens=encoder_out_lens + ) + + def token_ids_to_words(token_ids: List[int]) -> str: + text = "" + for i in token_ids: + text += token_table[i] + return text.replace("▁", " ").strip() + + hyps = [token_ids_to_words(h).split() for h in hyps] + return hyps + + +def decode_dataset( + dl: torch.utils.data.DataLoader, + model: nn.Module, + token_table: SymbolTable, +) -> Tuple[List[Tuple[str, List[str], List[str]]], float]: + """Decode dataset. + + Args: + dl: + PyTorch's dataloader containing the dataset to decode. + model: + The neural model. + token_table: + The token table. + + Returns: + - A list of tuples. Each tuple contains three elements: + - cut_id, + - reference transcript, + - predicted result. + - The total duration (in seconds) of the dataset. + """ + num_cuts = 0 + + try: + num_batches = len(dl) + except TypeError: + num_batches = "?" + + log_interval = 10 + total_duration = 0 + + results = [] + for batch_idx, batch in enumerate(dl): + texts = batch["supervisions"]["text"] + cut_ids = [cut.id for cut in batch["supervisions"]["cut"]] + total_duration += sum([cut.duration for cut in batch["supervisions"]["cut"]]) + + hyps = decode_one_batch(model=model, token_table=token_table, batch=batch) + + 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.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, total_duration + + +def save_results( + res_dir: Path, + test_set_name: str, + results: List[Tuple[str, List[str], List[str]]], +): + recog_path = res_dir / f"recogs-{test_set_name}.txt" + results = post_processing(results) + 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 = res_dir / f"errs-{test_set_name}.txt" + with open(errs_filename, "w") as f: + wer = write_error_stats(f, f"{test_set_name}", results, enable_log=True) + + logging.info("Wrote detailed error stats to {}".format(errs_filename)) + + errs_info = res_dir / f"wer-summary-{test_set_name}.txt" + with open(errs_info, "w") as f: + print("WER", file=f) + print(wer, file=f) + + s = "\nFor {}, WER is {}:\n".format(test_set_name, wer) + logging.info(s) + + +@torch.no_grad() +def main(): + parser = get_parser() + LibriSpeechAsrDataModule.add_arguments(parser) + args = parser.parse_args() + + assert ( + args.decoding_method == "greedy_search" + ), "Only supports greedy_search currently." + res_dir = Path(args.exp_dir) / f"onnx-{args.decoding_method}" + + setup_logger(f"{res_dir}/log-decode") + logging.info("Decoding started") + + device = torch.device("cpu") + logging.info(f"Device: {device}") + + token_table = SymbolTable.from_file(args.tokens) + + logging.info(vars(args)) + + logging.info("About to create model") + model = OnnxModel( + encoder_model_filename=args.encoder_model_filename, + decoder_model_filename=args.decoder_model_filename, + joiner_model_filename=args.joiner_model_filename, + ) + + # we need cut ids to display recognition results. + args.return_cuts = True + librispeech = LibriSpeechAsrDataModule(args) + + gigaspeech_dev_cuts = librispeech.gigaspeech_dev_cuts() + gigaspeech_test_cuts = librispeech.gigaspeech_test_cuts() + + dev_dl = librispeech.test_dataloaders(gigaspeech_dev_cuts) + test_dl = librispeech.test_dataloaders(gigaspeech_test_cuts) + + test_sets = ["dev", "test"] + test_dl = [dev_dl, test_dl] + + for test_set, test_dl in zip(test_sets, test_dl): + start_time = time.time() + results, total_duration = decode_dataset( + dl=test_dl, model=model, token_table=token_table + ) + end_time = time.time() + elapsed_seconds = end_time - start_time + rtf = elapsed_seconds / total_duration + + logging.info(f"Elapsed time: {elapsed_seconds:.3f} s") + logging.info(f"Wave duration: {total_duration:.3f} s") + logging.info( + f"Real time factor (RTF): {elapsed_seconds:.3f}/{total_duration:.3f} = {rtf:.3f}" + ) + + save_results(res_dir=res_dir, test_set_name=test_set, results=results) + + logging.info("Done!") + + +if __name__ == "__main__": + main() diff --git a/egs/librispeech/ASR/zipformer_adapter/onnx_pretrained.py b/egs/librispeech/ASR/zipformer_adapter/onnx_pretrained.py new file mode 120000 index 000000000..a085def83 --- /dev/null +++ b/egs/librispeech/ASR/zipformer_adapter/onnx_pretrained.py @@ -0,0 +1 @@ +../zipformer/onnx_pretrained.py \ No newline at end of file diff --git a/egs/librispeech/ASR/zipformer_adapter/optim.py b/egs/librispeech/ASR/zipformer_adapter/optim.py new file mode 120000 index 000000000..207eecfcd --- /dev/null +++ b/egs/librispeech/ASR/zipformer_adapter/optim.py @@ -0,0 +1 @@ +../zipformer/optim.py \ No newline at end of file diff --git a/egs/librispeech/ASR/zipformer_adapter/scaling.py b/egs/librispeech/ASR/zipformer_adapter/scaling.py new file mode 120000 index 000000000..58e4b0a0f --- /dev/null +++ b/egs/librispeech/ASR/zipformer_adapter/scaling.py @@ -0,0 +1 @@ +../zipformer/scaling.py \ No newline at end of file diff --git a/egs/librispeech/ASR/zipformer_adapter/scaling_converter.py b/egs/librispeech/ASR/zipformer_adapter/scaling_converter.py new file mode 120000 index 000000000..bc7c7b5e3 --- /dev/null +++ b/egs/librispeech/ASR/zipformer_adapter/scaling_converter.py @@ -0,0 +1 @@ +../zipformer/scaling_converter.py \ No newline at end of file diff --git a/egs/librispeech/ASR/zipformer_adapter/subsampling.py b/egs/librispeech/ASR/zipformer_adapter/subsampling.py new file mode 120000 index 000000000..d178adc2e --- /dev/null +++ b/egs/librispeech/ASR/zipformer_adapter/subsampling.py @@ -0,0 +1 @@ +../zipformer/subsampling.py \ No newline at end of file diff --git a/egs/librispeech/ASR/zipformer_adapter/train.py b/egs/librispeech/ASR/zipformer_adapter/train.py new file mode 100755 index 000000000..6c55896a8 --- /dev/null +++ b/egs/librispeech/ASR/zipformer_adapter/train.py @@ -0,0 +1,1544 @@ +#!/usr/bin/env python3 +# Copyright 2021-2024 Xiaomi Corp. (authors: Fangjun Kuang, +# Wei Kang, +# Mingshuang Luo, +# Zengwei Yao, +# Daniel Povey, +# Xiaoyu Yang) +# +# 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: + +export CUDA_VISIBLE_DEVICES="0,1,2,3" + +# Finetune non-streaming model using adapters: + +./zipformer/finetune.py \ + --world-size 4 \ + --num-epochs 30 \ + --start-epoch 1 \ + --use-fp16 1 \ + --do-finetune 1 \ + --use-mux 0 \ + --use-adapters 1 \ + --adapter-dim 16 \ + --finetune-ckpt icefall-asr-librispeech-zipformer-2023-05-15/exp/pretrained.pt \ + --exp-dir zipformer/exp \ + --max-duration 1000 + +""" + + +import argparse +import copy +import logging +import warnings +from pathlib import Path +from shutil import copyfile +from typing import Any, Dict, List, Optional, Tuple, Union + +import k2 +import optim +import sentencepiece as spm +import torch +import torch.multiprocessing as mp +import torch.nn as nn +from asr_datamodule import LibriSpeechAsrDataModule +from decoder import Decoder +from joiner import Joiner +from lhotse.cut import Cut, CutSet +from lhotse.dataset.sampling.base import CutSampler +from lhotse.utils import fix_random_seed +from model import AsrModel +from optim import Eden, ScaledAdam +from scaling import ScheduledFloat +from subsampling import Conv2dSubsampling +from torch import Tensor +from torch.cuda.amp import GradScaler +from torch.nn.parallel import DistributedDataParallel as DDP +from torch.utils.tensorboard import SummaryWriter +from zipformer import Zipformer2 + +from icefall import diagnostics +from icefall.checkpoint import load_checkpoint, remove_checkpoints +from icefall.checkpoint import save_checkpoint as save_checkpoint_impl +from icefall.checkpoint import ( + save_checkpoint_with_global_batch_idx, + update_averaged_model, +) +from icefall.dist import cleanup_dist, setup_dist +from icefall.env import get_env_info +from icefall.err import raise_grad_scale_is_too_small_error +from icefall.hooks import register_inf_check_hooks +from icefall.utils import ( + AttributeDict, + MetricsTracker, + get_parameter_groups_with_lrs, + setup_logger, + str2bool, +) + +LRSchedulerType = Union[torch.optim.lr_scheduler._LRScheduler, optim.LRScheduler] + + +def get_adjusted_batch_count(params: AttributeDict) -> float: + # returns the number of batches we would have used so far if we had used the reference + # duration. This is for purposes of set_batch_count(). + return ( + params.batch_idx_train + * (params.max_duration * params.world_size) + / params.ref_duration + ) + 1000000 + + +def set_batch_count(model: Union[nn.Module, DDP], batch_count: float) -> None: + if isinstance(model, DDP): + # get underlying nn.Module + model = model.module + for name, module in model.named_modules(): + if hasattr(module, "batch_count"): + module.batch_count = batch_count + if hasattr(module, "name"): + module.name = name + + +def add_finetune_arguments(parser: argparse.ArgumentParser): + parser.add_argument( + "--do-finetune", + type=str2bool, + default=True, + help="If true, finetune from a pre-trained checkpoint", + ) + + parser.add_argument( + "--use-mux", + type=str2bool, + default=False, + help=""" + Whether to adapt. If true, we will mix 5% of the new data + with 95% of the original data to fine-tune. This is useful + if you want to maintain the performance on the original domain + """, + ) + + parser.add_argument( + "--use-adapters", + type=str2bool, + default=True, + help="If use adapter to finetune the model", + ) + + parser.add_argument( + "--adapter-dim", + type=int, + default=16, + help="The bottleneck dimension of the adapter", + ) + + parser.add_argument( + "--init-modules", + type=str, + default=None, + help=""" + Modules to be initialized. It matches all parameters starting with + a specific key. The keys are given with Comma seperated. If None, + all modules will be initialised. For example, if you only want to + initialise all parameters staring with "encoder", use "encoder"; + if you want to initialise parameters starting with encoder or decoder, + use "encoder,joiner". + """, + ) + + parser.add_argument( + "--finetune-ckpt", + type=str, + default=None, + help="Fine-tuning from which checkpoint (path to a .pt file)", + ) + + +def add_model_arguments(parser: argparse.ArgumentParser): + parser.add_argument( + "--num-encoder-layers", + type=str, + default="2,2,3,4,3,2", + help="Number of zipformer encoder layers per stack, comma separated.", + ) + + parser.add_argument( + "--downsampling-factor", + type=str, + default="1,2,4,8,4,2", + help="Downsampling factor for each stack of encoder layers.", + ) + + parser.add_argument( + "--feedforward-dim", + type=str, + default="512,768,1024,1536,1024,768", + help="Feedforward dimension of the zipformer encoder layers, per stack, comma separated.", + ) + + parser.add_argument( + "--num-heads", + type=str, + default="4,4,4,8,4,4", + help="Number of attention heads in the zipformer encoder layers: a single int or comma-separated list.", + ) + + parser.add_argument( + "--encoder-dim", + type=str, + default="192,256,384,512,384,256", + help="Embedding dimension in encoder stacks: a single int or comma-separated list.", + ) + + parser.add_argument( + "--query-head-dim", + type=str, + default="32", + help="Query/key dimension per head in encoder stacks: a single int or comma-separated list.", + ) + + parser.add_argument( + "--value-head-dim", + type=str, + default="12", + help="Value dimension per head in encoder stacks: a single int or comma-separated list.", + ) + + parser.add_argument( + "--pos-head-dim", + type=str, + default="4", + help="Positional-encoding dimension per head in encoder stacks: a single int or comma-separated list.", + ) + + parser.add_argument( + "--pos-dim", + type=int, + default="48", + help="Positional-encoding embedding dimension", + ) + + parser.add_argument( + "--encoder-unmasked-dim", + type=str, + default="192,192,256,256,256,192", + help="Unmasked dimensions in the encoders, relates to augmentation during training. " + "A single int or comma-separated list. Must be <= each corresponding encoder_dim.", + ) + + parser.add_argument( + "--cnn-module-kernel", + type=str, + default="31,31,15,15,15,31", + help="Sizes of convolutional kernels in convolution modules in each encoder stack: " + "a single int or comma-separated list.", + ) + + parser.add_argument( + "--decoder-dim", + type=int, + default=512, + help="Embedding dimension in the decoder model.", + ) + + parser.add_argument( + "--joiner-dim", + type=int, + default=512, + help="""Dimension used in the joiner model. + Outputs from the encoder and decoder model are projected + to this dimension before adding. + """, + ) + + parser.add_argument( + "--causal", + type=str2bool, + default=False, + help="If True, use causal version of model.", + ) + + parser.add_argument( + "--chunk-size", + type=str, + default="16,32,64,-1", + help="Chunk sizes (at 50Hz frame rate) will be chosen randomly from this list during training. " + " Must be just -1 if --causal=False", + ) + + parser.add_argument( + "--left-context-frames", + type=str, + default="64,128,256,-1", + help="Maximum left-contexts for causal training, measured in frames which will " + "be converted to a number of chunks. If splitting into chunks, " + "chunk left-context frames will be chosen randomly from this list; else not relevant.", + ) + + parser.add_argument( + "--use-transducer", + type=str2bool, + default=True, + help="If True, use Transducer head.", + ) + + parser.add_argument( + "--use-ctc", + type=str2bool, + default=False, + help="If True, use CTC head.", + ) + + +def get_parser(): + parser = argparse.ArgumentParser( + formatter_class=argparse.ArgumentDefaultsHelpFormatter + ) + + parser.add_argument( + "--world-size", + type=int, + default=1, + help="Number of GPUs for DDP training.", + ) + + parser.add_argument( + "--master-port", + type=int, + default=12354, + help="Master port to use for DDP training.", + ) + + parser.add_argument( + "--tensorboard", + type=str2bool, + default=True, + help="Should various information be logged in tensorboard.", + ) + + parser.add_argument( + "--num-epochs", + type=int, + default=30, + help="Number of epochs to train.", + ) + + parser.add_argument( + "--start-epoch", + type=int, + default=1, + help="""Resume training from this epoch. It should be positive. + If larger than 1, it will load checkpoint from + exp-dir/epoch-{start_epoch-1}.pt + """, + ) + + parser.add_argument( + "--start-batch", + type=int, + default=0, + help="""If positive, --start-epoch is ignored and + it loads the checkpoint from exp-dir/checkpoint-{start_batch}.pt + """, + ) + + parser.add_argument( + "--exp-dir", + type=str, + default="zipformer/exp", + help="""The experiment dir. + 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( + "--base-lr", type=float, default=0.045, help="The base learning rate." + ) + + parser.add_argument( + "--lr-batches", + type=float, + default=7500, + help="""Number of steps that affects how rapidly the learning rate + decreases. We suggest not to change this.""", + ) + + parser.add_argument( + "--lr-epochs", + type=float, + default=3.5, + help="""Number of epochs that affects how rapidly the learning rate decreases. + """, + ) + + parser.add_argument( + "--ref-duration", + type=float, + default=600, + help="Reference batch duration for purposes of adjusting batch counts for setting various " + "schedules inside the model", + ) + + 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( + "--prune-range", + type=int, + default=5, + help="The prune range for rnnt loss, it means how many symbols(context)" + "we are using to compute the loss", + ) + + parser.add_argument( + "--lm-scale", + type=float, + default=0.25, + help="The scale to smooth the loss with lm " + "(output of prediction network) part.", + ) + + parser.add_argument( + "--am-scale", + type=float, + default=0.0, + help="The scale to smooth the loss with am (output of encoder network)" "part.", + ) + + parser.add_argument( + "--simple-loss-scale", + type=float, + default=0.5, + help="To get pruning ranges, we will calculate a simple version" + "loss(joiner is just addition), this simple loss also uses for" + "training (as a regularization item). We will scale the simple loss" + "with this parameter before adding to the final loss.", + ) + + parser.add_argument( + "--ctc-loss-scale", + type=float, + default=0.2, + help="Scale for CTC loss.", + ) + + parser.add_argument( + "--seed", + type=int, + default=42, + help="The seed for random generators intended for reproducibility", + ) + + parser.add_argument( + "--print-diagnostics", + type=str2bool, + default=False, + help="Accumulate stats on activations, print them and exit.", + ) + + parser.add_argument( + "--inf-check", + type=str2bool, + default=False, + help="Add hooks to check for infinite module outputs and gradients.", + ) + + parser.add_argument( + "--save-every-n", + type=int, + default=4000, + help="""Save checkpoint after processing this number of batches" + periodically. We save checkpoint to exp-dir/ whenever + params.batch_idx_train % save_every_n == 0. The checkpoint filename + has the form: f'exp-dir/checkpoint-{params.batch_idx_train}.pt' + Note: It also saves checkpoint to `exp-dir/epoch-xxx.pt` at the + end of each epoch where `xxx` is the epoch number counting from 1. + """, + ) + + parser.add_argument( + "--keep-last-k", + type=int, + default=30, + help="""Only keep this number of checkpoints on disk. + For instance, if it is 3, there are only 3 checkpoints + in the exp-dir with filenames `checkpoint-xxx.pt`. + It does not affect checkpoints with name `epoch-xxx.pt`. + """, + ) + + parser.add_argument( + "--average-period", + type=int, + default=200, + help="""Update the averaged model, namely `model_avg`, after processing + this number of batches. `model_avg` is a separate version of model, + in which each floating-point parameter is the average of all the + parameters from the start of training. Each time we take the average, + we do: `model_avg = model * (average_period / batch_idx_train) + + model_avg * ((batch_idx_train - average_period) / batch_idx_train)`. + """, + ) + + parser.add_argument( + "--use-fp16", + type=str2bool, + default=False, + help="Whether to use half precision training.", + ) + + add_model_arguments(parser) + add_finetune_arguments(parser) + + return parser + + +def get_params() -> AttributeDict: + """Return a dict containing training parameters. + + All training related parameters that are not passed from the commandline + are saved in the variable `params`. + + Commandline options are merged into `params` after they are parsed, so + you can also access them via `params`. + + Explanation of options saved in `params`: + + - best_train_loss: Best training loss so far. It is used to select + the model that has the lowest training loss. It is + updated during the training. + + - best_valid_loss: Best validation loss so far. It is used to select + the model that has the lowest validation loss. It is + updated during the training. + + - best_train_epoch: It is the epoch that has the best training loss. + + - best_valid_epoch: It is the epoch that has the best validation loss. + + - batch_idx_train: Used to writing statistics to tensorboard. It + contains number of batches trained so far across + epochs. + + - log_interval: Print training loss if batch_idx % log_interval` is 0 + + - reset_interval: Reset statistics if batch_idx % reset_interval is 0 + + - valid_interval: Run validation if batch_idx % valid_interval is 0 + + - feature_dim: The model input dim. It has to match the one used + in computing features. + + - subsampling_factor: The subsampling factor for the model. + + - encoder_dim: Hidden dim for multi-head attention model. + + - num_decoder_layers: Number of decoder layer of transformer decoder. + + - warm_step: The warmup period that dictates the decay of the + scale on "simple" (un-pruned) loss. + """ + params = AttributeDict( + { + "best_train_loss": float("inf"), + "best_valid_loss": float("inf"), + "best_train_epoch": -1, + "best_valid_epoch": -1, + "batch_idx_train": 0, + "log_interval": 50, + "reset_interval": 200, + "valid_interval": 3000, # For the 100h subset, use 800 + # parameters for zipformer + "feature_dim": 80, + "subsampling_factor": 4, # not passed in, this is fixed. + "warm_step": 2000, + "env_info": get_env_info(), + } + ) + + return params + + +def _to_int_tuple(s: str): + return tuple(map(int, s.split(","))) + + +def get_encoder_embed(params: AttributeDict) -> nn.Module: + # encoder_embed converts the input of shape (N, T, num_features) + # to the shape (N, (T - 7) // 2, encoder_dims). + # That is, it does two things simultaneously: + # (1) subsampling: T -> (T - 7) // 2 + # (2) embedding: num_features -> encoder_dims + # In the normal configuration, we will downsample once more at the end + # by a factor of 2, and most of the encoder stacks will run at a lower + # sampling rate. + encoder_embed = Conv2dSubsampling( + in_channels=params.feature_dim, + out_channels=_to_int_tuple(params.encoder_dim)[0], + dropout=ScheduledFloat((0.0, 0.3), (20000.0, 0.1)), + ) + return encoder_embed + + +def get_encoder_model(params: AttributeDict) -> nn.Module: + encoder = Zipformer2( + output_downsampling_factor=2, + downsampling_factor=_to_int_tuple(params.downsampling_factor), + num_encoder_layers=_to_int_tuple(params.num_encoder_layers), + encoder_dim=_to_int_tuple(params.encoder_dim), + encoder_unmasked_dim=_to_int_tuple(params.encoder_unmasked_dim), + query_head_dim=_to_int_tuple(params.query_head_dim), + pos_head_dim=_to_int_tuple(params.pos_head_dim), + value_head_dim=_to_int_tuple(params.value_head_dim), + pos_dim=params.pos_dim, + num_heads=_to_int_tuple(params.num_heads), + feedforward_dim=_to_int_tuple(params.feedforward_dim), + cnn_module_kernel=_to_int_tuple(params.cnn_module_kernel), + dropout=ScheduledFloat((0.0, 0.3), (20000.0, 0.1)), + warmup_batches=4000.0, + causal=params.causal, + chunk_size=_to_int_tuple(params.chunk_size), + left_context_frames=_to_int_tuple(params.left_context_frames), + use_adapters=params.use_adapters, + adapter_dim=params.adapter_dim, + ) + return encoder + + +def get_decoder_model(params: AttributeDict) -> nn.Module: + decoder = Decoder( + vocab_size=params.vocab_size, + decoder_dim=params.decoder_dim, + blank_id=params.blank_id, + context_size=params.context_size, + ) + return decoder + + +def get_joiner_model(params: AttributeDict) -> nn.Module: + joiner = Joiner( + encoder_dim=max(_to_int_tuple(params.encoder_dim)), + decoder_dim=params.decoder_dim, + joiner_dim=params.joiner_dim, + vocab_size=params.vocab_size, + ) + return joiner + + +def get_model(params: AttributeDict) -> nn.Module: + assert params.use_transducer or params.use_ctc, ( + f"At least one of them should be True, " + f"but got params.use_transducer={params.use_transducer}, " + f"params.use_ctc={params.use_ctc}" + ) + + encoder_embed = get_encoder_embed(params) + encoder = get_encoder_model(params) + + if params.use_transducer: + decoder = get_decoder_model(params) + joiner = get_joiner_model(params) + else: + decoder = None + joiner = None + + model = AsrModel( + encoder_embed=encoder_embed, + encoder=encoder, + decoder=decoder, + joiner=joiner, + encoder_dim=max(_to_int_tuple(params.encoder_dim)), + decoder_dim=params.decoder_dim, + vocab_size=params.vocab_size, + use_transducer=params.use_transducer, + use_ctc=params.use_ctc, + ) + return model + + +def load_checkpoint_if_available( + params: AttributeDict, + model: nn.Module, + model_avg: nn.Module = None, + optimizer: Optional[torch.optim.Optimizer] = None, + scheduler: Optional[LRSchedulerType] = None, +) -> Optional[Dict[str, Any]]: + """Load checkpoint from file. + + If params.start_batch is positive, it will load the checkpoint from + `params.exp_dir/checkpoint-{params.start_batch}.pt`. Otherwise, if + params.start_epoch is larger than 1, it will load the checkpoint from + `params.start_epoch - 1`. + + Apart from loading state dict for `model` and `optimizer` it also updates + `best_train_epoch`, `best_train_loss`, `best_valid_epoch`, + and `best_valid_loss` in `params`. + + Args: + params: + The return value of :func:`get_params`. + model: + The training model. + model_avg: + The stored model averaged from the start of training. + optimizer: + The optimizer that we are using. + scheduler: + The scheduler that we are using. + Returns: + Return a dict containing previously saved training info. + """ + if params.start_batch > 0: + filename = params.exp_dir / f"checkpoint-{params.start_batch}.pt" + elif params.start_epoch > 1: + filename = params.exp_dir / f"epoch-{params.start_epoch-1}.pt" + else: + return None + + assert filename.is_file(), f"{filename} does not exist!" + + saved_params = load_checkpoint( + filename, + model=model, + model_avg=model_avg, + optimizer=optimizer, + scheduler=scheduler, + ) + + keys = [ + "best_train_epoch", + "best_valid_epoch", + "batch_idx_train", + "best_train_loss", + "best_valid_loss", + ] + for k in keys: + params[k] = saved_params[k] + + if params.start_batch > 0: + if "cur_epoch" in saved_params: + params["start_epoch"] = saved_params["cur_epoch"] + + return saved_params + + +def load_model_params( + ckpt: str, model: nn.Module, init_modules: List[str] = None, strict: bool = True +): + """Load model params from checkpoint + + Args: + ckpt (str): Path to the checkpoint + model (nn.Module): model to be loaded + init_modules (list[str]): List of modules to be initialized + + """ + logging.info(f"Loading checkpoint from {ckpt}") + checkpoint = torch.load(ckpt, map_location="cpu") + + # if module list is empty, load the whole model from ckpt + if not init_modules: + if next(iter(checkpoint["model"])).startswith("module."): + logging.info("Loading checkpoint saved by DDP") + + dst_state_dict = model.state_dict() + src_state_dict = checkpoint["model"] + for key in dst_state_dict.keys(): + src_key = "{}.{}".format("module", key) + dst_state_dict[key] = src_state_dict.pop(src_key) + assert len(src_state_dict) == 0 + model.load_state_dict(dst_state_dict, strict=strict) + else: + model.load_state_dict(checkpoint["model"], strict=strict) + else: + src_state_dict = checkpoint["model"] + dst_state_dict = model.state_dict() + for module in init_modules: + logging.info(f"Loading parameters starting with prefix {module}") + src_keys = [ + k for k in src_state_dict.keys() if k.startswith(module.strip() + ".") + ] + dst_keys = [ + k for k in dst_state_dict.keys() if k.startswith(module.strip() + ".") + ] + assert set(src_keys) == set(dst_keys) # two sets should match exactly + for key in src_keys: + dst_state_dict[key] = src_state_dict.pop(key) + + model.load_state_dict(dst_state_dict, strict=strict) + + return None + + +def save_checkpoint( + params: AttributeDict, + model: Union[nn.Module, DDP], + model_avg: Optional[nn.Module] = None, + optimizer: Optional[torch.optim.Optimizer] = None, + scheduler: Optional[LRSchedulerType] = None, + sampler: Optional[CutSampler] = None, + scaler: Optional[GradScaler] = None, + rank: int = 0, +) -> None: + """Save model, optimizer, scheduler and training stats to file. + + Args: + params: + It is returned by :func:`get_params`. + model: + The training model. + model_avg: + The stored model averaged from the start of training. + optimizer: + The optimizer used in the training. + sampler: + The sampler for the training dataset. + scaler: + The scaler used for mix precision training. + """ + if rank != 0: + return + filename = params.exp_dir / f"epoch-{params.cur_epoch}.pt" + save_checkpoint_impl( + filename=filename, + model=model, + model_avg=model_avg, + params=params, + optimizer=optimizer, + scheduler=scheduler, + sampler=sampler, + scaler=scaler, + rank=rank, + ) + + if params.best_train_epoch == params.cur_epoch: + best_train_filename = params.exp_dir / "best-train-loss.pt" + copyfile(src=filename, dst=best_train_filename) + + if params.best_valid_epoch == params.cur_epoch: + best_valid_filename = params.exp_dir / "best-valid-loss.pt" + copyfile(src=filename, dst=best_valid_filename) + + +def compute_loss( + params: AttributeDict, + model: Union[nn.Module, DDP], + sp: spm.SentencePieceProcessor, + batch: dict, + is_training: bool, +) -> Tuple[Tensor, MetricsTracker]: + """ + Compute loss given the model and its inputs. + + Args: + params: + Parameters for training. See :func:`get_params`. + model: + The model for training. It is an instance of Zipformer in our case. + batch: + A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()` + for the content in it. + is_training: + True for training. False for validation. When it is True, this + function enables autograd during computation; when it is False, it + disables autograd. + warmup: a floating point value which increases throughout training; + values >= 1.0 are fully warmed up and have all modules present. + """ + device = model.device if isinstance(model, DDP) else next(model.parameters()).device + feature = batch["inputs"] + # at entry, feature is (N, T, C) + assert feature.ndim == 3 + feature = feature.to(device) + + supervisions = batch["supervisions"] + feature_lens = supervisions["num_frames"].to(device) + + batch_idx_train = params.batch_idx_train + warm_step = params.warm_step + + texts = batch["supervisions"]["text"] + y = sp.encode(texts, out_type=int) + y = k2.RaggedTensor(y) + + with torch.set_grad_enabled(is_training): + simple_loss, pruned_loss, ctc_loss = model( + x=feature, + x_lens=feature_lens, + y=y, + prune_range=params.prune_range, + am_scale=params.am_scale, + lm_scale=params.lm_scale, + ) + + loss = 0.0 + + if params.use_transducer: + s = params.simple_loss_scale + # take down the scale on the simple loss from 1.0 at the start + # to params.simple_loss scale by warm_step. + simple_loss_scale = ( + s + if batch_idx_train >= warm_step + else 1.0 - (batch_idx_train / warm_step) * (1.0 - s) + ) + pruned_loss_scale = ( + 1.0 + if batch_idx_train >= warm_step + else 0.1 + 0.9 * (batch_idx_train / warm_step) + ) + loss += simple_loss_scale * simple_loss + pruned_loss_scale * pruned_loss + + if params.use_ctc: + loss += params.ctc_loss_scale * ctc_loss + + assert loss.requires_grad == is_training + + info = MetricsTracker() + with warnings.catch_warnings(): + warnings.simplefilter("ignore") + info["frames"] = (feature_lens // params.subsampling_factor).sum().item() + + # Note: We use reduction=sum while computing the loss. + info["loss"] = loss.detach().cpu().item() + if params.use_transducer: + info["simple_loss"] = simple_loss.detach().cpu().item() + info["pruned_loss"] = pruned_loss.detach().cpu().item() + if params.use_ctc: + info["ctc_loss"] = ctc_loss.detach().cpu().item() + + return loss, info + + +def compute_validation_loss( + params: AttributeDict, + model: Union[nn.Module, DDP], + sp: spm.SentencePieceProcessor, + valid_dl: torch.utils.data.DataLoader, + world_size: int = 1, +) -> MetricsTracker: + """Run the validation process.""" + model.eval() + + tot_loss = MetricsTracker() + + for batch_idx, batch in enumerate(valid_dl): + loss, loss_info = compute_loss( + params=params, + model=model, + sp=sp, + batch=batch, + is_training=False, + ) + assert loss.requires_grad is False + tot_loss = tot_loss + loss_info + + if world_size > 1: + tot_loss.reduce(loss.device) + + loss_value = tot_loss["loss"] / tot_loss["frames"] + if loss_value < params.best_valid_loss: + params.best_valid_epoch = params.cur_epoch + params.best_valid_loss = loss_value + + return tot_loss + + +def train_one_epoch( + params: AttributeDict, + model: Union[nn.Module, DDP], + optimizer: torch.optim.Optimizer, + scheduler: LRSchedulerType, + sp: spm.SentencePieceProcessor, + train_dl: torch.utils.data.DataLoader, + valid_dls: torch.utils.data.DataLoader, + valid_sets: List[str], + scaler: GradScaler, + model_avg: Optional[nn.Module] = None, + tb_writer: Optional[SummaryWriter] = None, + world_size: int = 1, + rank: int = 0, +) -> None: + """Train the model for one epoch. + + The training loss from the mean of all frames is saved in + `params.train_loss`. It runs the validation process every + `params.valid_interval` batches. + + Args: + params: + It is returned by :func:`get_params`. + model: + The model for training. + optimizer: + The optimizer we are using. + scheduler: + The learning rate scheduler, we call step() every step. + train_dl: + Dataloader for the training dataset. + valid_dl: + Dataloader for the validation dataset. + scaler: + The scaler used for mix precision training. + model_avg: + The stored model averaged from the start of training. + tb_writer: + Writer to write log messages to tensorboard. + world_size: + Number of nodes in DDP training. If it is 1, DDP is disabled. + rank: + The rank of the node in DDP training. If no DDP is used, it should + be set to 0. + """ + model.train() + # set modules except adapters to eval mode + for name, m in model.named_modules(): + if "adapter" in name: + m.training = True + else: + m.training = False + + tot_loss = MetricsTracker() + + saved_bad_model = False + + def save_bad_model(suffix: str = ""): + save_checkpoint_impl( + filename=params.exp_dir / f"bad-model{suffix}-{rank}.pt", + model=model, + model_avg=model_avg, + params=params, + optimizer=optimizer, + scheduler=scheduler, + sampler=train_dl.sampler, + scaler=scaler, + rank=0, + ) + + for batch_idx, batch in enumerate(train_dl): + if batch_idx % 10 == 0: + set_batch_count(model, get_adjusted_batch_count(params)) + + params.batch_idx_train += 1 + batch_size = len(batch["supervisions"]["text"]) + + 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, + ) + # 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() + except: # noqa + save_bad_model() + display_and_save_batch(batch, params=params, sp=sp) + raise + + if params.print_diagnostics and batch_idx == 5: + return + + if ( + rank == 0 + and params.batch_idx_train > 0 + and params.batch_idx_train % params.average_period == 0 + ): + update_averaged_model( + params=params, + model_cur=model, + model_avg=model_avg, + ) + + if ( + params.batch_idx_train > 0 + and params.batch_idx_train % params.save_every_n == 0 + ): + save_checkpoint_with_global_batch_idx( + out_dir=params.exp_dir, + global_batch_idx=params.batch_idx_train, + model=model, + model_avg=model_avg, + params=params, + optimizer=optimizer, + scheduler=scheduler, + sampler=train_dl.sampler, + scaler=scaler, + rank=rank, + ) + remove_checkpoints( + out_dir=params.exp_dir, + topk=params.keep_last_k, + rank=rank, + ) + + if batch_idx % 100 == 0 and params.use_fp16: + # If the grad scale was less than 1, try increasing it. The _growth_interval + # of the grad scaler is configurable, but we can't configure it to have different + # behavior depending on the current grad scale. + cur_grad_scale = scaler._scale.item() + + if cur_grad_scale < 8.0 or (cur_grad_scale < 32.0 and batch_idx % 400 == 0): + scaler.update(cur_grad_scale * 2.0) + if cur_grad_scale < 0.01: + if not saved_bad_model: + save_bad_model(suffix="-first-warning") + saved_bad_model = True + logging.warning(f"Grad scale is small: {cur_grad_scale}") + if cur_grad_scale < 1.0e-05: + save_bad_model() + raise_grad_scale_is_too_small_error(cur_grad_scale) + + if batch_idx % params.log_interval == 0: + cur_lr = max(scheduler.get_last_lr()) + cur_grad_scale = scaler._scale.item() if params.use_fp16 else 1.0 + + logging.info( + f"Epoch {params.cur_epoch}, " + f"batch {batch_idx}, loss[{loss_info}], " + f"tot_loss[{tot_loss}], batch size: {batch_size}, " + f"lr: {cur_lr:.2e}, " + + (f"grad_scale: {scaler._scale.item()}" if params.use_fp16 else "") + ) + + if tb_writer is not None: + tb_writer.add_scalar( + "train/learning_rate", cur_lr, params.batch_idx_train + ) + + loss_info.write_summary( + tb_writer, "train/current_", params.batch_idx_train + ) + tot_loss.write_summary(tb_writer, "train/tot_", params.batch_idx_train) + if params.use_fp16: + tb_writer.add_scalar( + "train/grad_scale", cur_grad_scale, params.batch_idx_train + ) + + if batch_idx % params.valid_interval == 0 and not params.print_diagnostics: + for valid_set, valid_dl in zip(valid_sets, valid_dls): + logging.info(f"Computing validation loss on {valid_set}") + valid_info = compute_validation_loss( + params=params, + model=model, + sp=sp, + valid_dl=valid_dl, + world_size=world_size, + ) + logging.info( + f"Validation on {valid_set}: Epoch {params.cur_epoch}, validation: {valid_info}" + ) + logging.info( + f"Maximum memory allocated so far is {torch.cuda.max_memory_allocated()//1000000}MB" + ) + if tb_writer is not None: + valid_info.write_summary( + tb_writer, f"train/{valid_set}_valid_", params.batch_idx_train + ) + model.train() + # set modules except adapters to eval mode + for name, m in model.named_modules(): + if "adapter" in name: + m.training = True + else: + m.training = False + + loss_value = tot_loss["loss"] / tot_loss["frames"] + params.train_loss = loss_value + if params.train_loss < params.best_train_loss: + params.best_train_epoch = params.cur_epoch + params.best_train_loss = params.train_loss + + +def run(rank, world_size, args): + """ + Args: + rank: + It is a value between 0 and `world_size-1`, which is + passed automatically by `mp.spawn()` in :func:`main`. + The node with rank 0 is responsible for saving checkpoint. + world_size: + Number of GPUs for DDP training. + args: + The return value of get_parser().parse_args() + """ + params = get_params() + params.update(vars(args)) + + fix_random_seed(params.seed) + if world_size > 1: + setup_dist(rank, world_size, params.master_port) + + setup_logger(f"{params.exp_dir}/log/log-train") + logging.info("Training started") + + if args.tensorboard and rank == 0: + tb_writer = SummaryWriter(log_dir=f"{params.exp_dir}/tensorboard") + else: + tb_writer = None + + device = torch.device("cpu") + if torch.cuda.is_available(): + device = torch.device("cuda", rank) + logging.info(f"Device: {device}") + + sp = spm.SentencePieceProcessor() + sp.load(params.bpe_model) + + # is defined in local/train_bpe_model.py + params.blank_id = sp.piece_to_id("") + params.vocab_size = sp.get_piece_size() + + if not params.use_transducer: + params.ctc_loss_scale = 1.0 + + logging.info(params) + + logging.info("About to create model") + model = get_model(params) + + num_param = sum([p.numel() for p in model.parameters()]) + logging.info(f"Number of model parameters: {num_param}") + + assert params.save_every_n >= params.average_period + model_avg: Optional[nn.Module] = None + if rank == 0: + # model_avg is only used with rank 0 + model_avg = copy.deepcopy(model).to(torch.float64) + + # load model parameters for model fine-tuning + if params.do_finetune: + assert params.start_epoch == 1, "Fine-tune must start from epoch 1" + modules = params.init_modules.split(",") if params.init_modules else None + checkpoints = load_model_params( + ckpt=params.finetune_ckpt, model=model, init_modules=modules, strict=False + ) + # Need to update the model_avg if use initialisation + if rank == 0: + # model_avg is only used with rank 0 + model_avg = copy.deepcopy(model).to(torch.float64) + else: + # resuming training + assert params.start_epoch > 1, params.start_epoch + checkpoints = load_checkpoint_if_available( + params=params, model=model, model_avg=model_avg + ) + + # keep the original model untouched, only update the adapters + num_trainable = 0 + for name, p in model.named_parameters(): + if "adapter" in name: + p.requires_grad = True + num_trainable += p.numel() + else: + p.requires_grad = False + + logging.info( + "A total of {} trainable parameters ({:.3f}% of the whole model)".format( + num_trainable, num_trainable / num_param * 100 + ) + ) + + model.to(device) + if world_size > 1: + logging.info("Using DDP") + model = DDP(model, device_ids=[rank], find_unused_parameters=True) + + optimizer = ScaledAdam( + get_parameter_groups_with_lrs(model, lr=params.base_lr, include_names=True), + lr=params.base_lr, # should have no effect + clipping_scale=2.0, + ) + + scheduler = Eden(optimizer, params.lr_batches, params.lr_epochs) + + if checkpoints and "optimizer" in checkpoints: + logging.info("Loading optimizer state dict") + optimizer.load_state_dict(checkpoints["optimizer"]) + + if ( + checkpoints + and "scheduler" in checkpoints + and checkpoints["scheduler"] is not None + ): + logging.info("Loading scheduler state dict") + scheduler.load_state_dict(checkpoints["scheduler"]) + + if params.print_diagnostics: + opts = diagnostics.TensorDiagnosticOptions( + 512 + ) # allow 4 megabytes per sub-module + diagnostic = diagnostics.attach_diagnostics(model, opts) + + if params.inf_check: + register_inf_check_hooks(model) + + librispeech = LibriSpeechAsrDataModule(args) + + gigaspeech_cuts = librispeech.gigaspeech_subset_small_cuts() + if params.use_mux: + librispeech_cuts = librispeech.train_all_shuf_cuts() + train_cuts = CutSet.mux( + gigaspeech_cuts, # num cuts = 688182 + librispeech_cuts, # num cuts = 843723 + weights=[688182, 843723], + stop_early=True, + ) + else: + train_cuts = gigaspeech_cuts + logging.info(train_cuts) + + def remove_short_and_long_utt(c: Cut): + # Keep only utterances with duration between 1 second and 20 seconds + # + # Caution: There is a reason to select 20.0 here. Please see + # ../local/display_manifest_statistics.py + # + # You should use ../local/display_manifest_statistics.py to get + # an utterance duration distribution for your dataset to select + # the threshold + if c.duration < 1.0 or c.duration > 20.0: + # logging.warning( + # f"Exclude cut with ID {c.id} from training. Duration: {c.duration}" + # ) + return False + + # In pruned RNN-T, we require that T >= S + # where T is the number of feature frames after subsampling + # and S is the number of tokens in the utterance + + # In ./zipformer.py, the conv module uses the following expression + # for subsampling + T = ((c.num_frames - 7) // 2 + 1) // 2 + tokens = sp.encode(c.supervisions[0].text, out_type=str) + + if T < len(tokens): + logging.warning( + f"Exclude cut with ID {c.id} from training. " + f"Number of frames (before subsampling): {c.num_frames}. " + f"Number of frames (after subsampling): {T}. " + f"Text: {c.supervisions[0].text}. " + f"Tokens: {tokens}. " + f"Number of tokens: {len(tokens)}" + ) + return False + + return True + + train_cuts = train_cuts.filter(remove_short_and_long_utt) + + if params.start_batch > 0 and checkpoints and "sampler" in checkpoints: + # We only load the sampler's state dict when it loads a checkpoint + # saved in the middle of an epoch + sampler_state_dict = checkpoints["sampler"] + else: + sampler_state_dict = None + + train_dl = librispeech.train_dataloaders( + train_cuts, sampler_state_dict=sampler_state_dict + ) + + valid_cuts = librispeech.dev_clean_cuts() + valid_cuts += librispeech.dev_other_cuts() + gigaspeech_dev_cuts = librispeech.gigaspeech_dev_cuts() + + valid_sets = ["librispeech", "gigaspeech"] + valid_dls = [ + librispeech.valid_dataloaders(valid_cuts), + librispeech.valid_dataloaders(gigaspeech_dev_cuts), + ] + + if not params.print_diagnostics: + scan_pessimistic_batches_for_oom( + model=model, + train_dl=train_dl, + optimizer=optimizer, + sp=sp, + params=params, + ) + + scaler = GradScaler(enabled=params.use_fp16, init_scale=1.0) + if checkpoints and "grad_scaler" in checkpoints: + logging.info("Loading grad scaler state dict") + scaler.load_state_dict(checkpoints["grad_scaler"]) + + for epoch in range(params.start_epoch, params.num_epochs + 1): + scheduler.step_epoch(epoch - 1) + fix_random_seed(params.seed + epoch - 1) + train_dl.sampler.set_epoch(epoch - 1) + + if tb_writer is not None: + tb_writer.add_scalar("train/epoch", epoch, params.batch_idx_train) + + params.cur_epoch = epoch + + train_one_epoch( + params=params, + model=model, + model_avg=model_avg, + optimizer=optimizer, + scheduler=scheduler, + sp=sp, + train_dl=train_dl, + valid_dls=valid_dls, + valid_sets=valid_sets, + scaler=scaler, + tb_writer=tb_writer, + world_size=world_size, + rank=rank, + ) + + if params.print_diagnostics: + diagnostic.print_diagnostics() + break + + save_checkpoint( + params=params, + model=model, + model_avg=model_avg, + optimizer=optimizer, + scheduler=scheduler, + sampler=train_dl.sampler, + scaler=scaler, + rank=rank, + ) + + logging.info("Done!") + + if world_size > 1: + torch.distributed.barrier() + 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: Union[nn.Module, DDP], + train_dl: torch.utils.data.DataLoader, + optimizer: torch.optim.Optimizer, + sp: spm.SentencePieceProcessor, + params: AttributeDict, +): + from lhotse.dataset import find_pessimistic_batches + + logging.info( + "Sanity check -- see if any of the batches in epoch 1 would cause OOM." + ) + batches, crit_values = find_pessimistic_batches(train_dl.sampler) + for criterion, cuts in batches.items(): + batch = train_dl.dataset[cuts] + try: + with torch.cuda.amp.autocast(enabled=params.use_fp16): + loss, _ = compute_loss( + params=params, + model=model, + sp=sp, + batch=batch, + is_training=True, + ) + loss.backward() + optimizer.zero_grad() + except Exception as e: + if "CUDA out of memory" in str(e): + logging.error( + "Your GPU ran out of memory with the current " + "max_duration setting. We recommend decreasing " + "max_duration and trying again.\n" + f"Failing criterion: {criterion} " + f"(={crit_values[criterion]}) ..." + ) + display_and_save_batch(batch, params=params, sp=sp) + raise + logging.info( + f"Maximum memory allocated so far is {torch.cuda.max_memory_allocated()//1000000}MB" + ) + + +def main(): + parser = get_parser() + LibriSpeechAsrDataModule.add_arguments(parser) + args = parser.parse_args() + args.exp_dir = Path(args.exp_dir) + + world_size = args.world_size + assert world_size >= 1 + if world_size > 1: + mp.spawn(run, args=(world_size, args), nprocs=world_size, join=True) + else: + run(rank=0, world_size=1, args=args) + + +torch.set_num_threads(1) +torch.set_num_interop_threads(1) + +if __name__ == "__main__": + main() diff --git a/egs/librispeech/ASR/zipformer_adapter/zipformer.py b/egs/librispeech/ASR/zipformer_adapter/zipformer.py new file mode 100644 index 000000000..4e4695fa5 --- /dev/null +++ b/egs/librispeech/ASR/zipformer_adapter/zipformer.py @@ -0,0 +1,2515 @@ +#!/usr/bin/env python3 +# Copyright 2022-2023 Xiaomi Corp. (authors: Daniel Povey, +# Zengwei Yao, +# Xiaoyu Yang) +# +# 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. + +import copy +import logging +import math +import random +import warnings +from typing import List, Optional, Tuple, Union + +import torch +from encoder_interface import EncoderInterface +from scaling import ( + Identity, # more friendly to backward hooks than nn.Identity(), for diagnostic reasons. +) +from scaling import ( + ScaledLinear, # not as in other dirs.. just scales down initial parameter values. +) +from scaling import ( + ActivationDropoutAndLinear, + Balancer, + BiasNorm, + ChunkCausalDepthwiseConv1d, + Dropout2, + FloatLike, + ScheduledFloat, + SwooshL, + SwooshR, + Whiten, + convert_num_channels, + limit_param_value, + penalize_abs_values_gt, + softmax, +) +from torch import Tensor, nn + + +class Zipformer2(EncoderInterface): + """ + Args: + + Note: all "int or Tuple[int]" arguments below will be treated as lists of the same length + as downsampling_factor if they are single ints or one-element tuples. The length of + downsampling_factor defines the number of stacks. + + output_downsampling_factor (int): how much to downsample at the output. Note: + we also downsample by a factor of 2 in the Conv2dSubsampling encoder. + You should probably leave this at 2. + downsampling_factor (Tuple[int]): downsampling factor for each encoder stack. + Note: this is in addition to the downsampling factor of 2 that is applied in + the frontend (self.encoder_embed). + encoder_dim (Tuple[int]): embedding dimension of each of the encoder stacks, one per + encoder stack. + num_encoder_layers (int or Tuple[int])): number of encoder layers for each stack + encoder_unmasked_dim (int or Tuple[int]): unmasked dimension in each of + the encoder stacks for purposes of per-frame dropout (recommend 256 for + now). + query_head_dim (int or Tuple[int]): dimension of query and key per attention + head: per stack, if a tuple.. + pos_head_dim (int or Tuple[int]): dimension of positional-encoding projection per + attention head + value_head_dim (int or Tuple[int]): dimension of value in each attention head + num_heads: (int or Tuple[int]): number of heads in the self-attention mechanism. + Must be at least 4. + feedforward_dim (int or Tuple[int]): hidden dimension in feedforward modules + cnn_module_kernel (int or Tuple[int])): Kernel size of convolution module + + pos_dim (int): the dimension of each positional-encoding vector prior to projection, + e.g. 128. + + dropout (float): dropout rate + warmup_batches (float): number of batches to warm up over; this controls + dropout of encoder layers. + causal (bool): if True, support chunkwise causal convolution. This should + not hurt WER as no modeling power is lost, but the convolution modules will be + slightly slower and use more memory. Enables use of the chunk_size and + left_context_chunks options in forward(), which simulates streaming + decoding. + chunk_size: (list of int): only set this to other than [-1] if causal; + the chunk size will be randomly chosen from this list. -1 means no chunking. + left_context_frames: (list of int): determines the number of left- + context chunks for causal training; will be rounded to a number of + chunks. Must not be less than cnn_module_kernel (after factoring in + rounding and downsampling); an error will be thrown if this is violated. + use_adapters: insert adapters in the zipformer encoder + adapter_dim: the dimension of the adapters + """ + + def __init__( + self, + output_downsampling_factor: int = 2, + downsampling_factor: Tuple[int] = (2, 4), + encoder_dim: Union[int, Tuple[int]] = 384, + num_encoder_layers: Union[int, Tuple[int]] = 4, + encoder_unmasked_dim: Union[int, Tuple[int]] = 256, + query_head_dim: Union[int, Tuple[int]] = 24, + pos_head_dim: Union[int, Tuple[int]] = 4, + value_head_dim: Union[int, Tuple[int]] = 12, + num_heads: Union[int, Tuple[int]] = 8, + feedforward_dim: Union[int, Tuple[int]] = 1536, + cnn_module_kernel: Union[int, Tuple[int]] = 31, + pos_dim: int = 192, + dropout: FloatLike = None, # see code below for default + warmup_batches: float = 4000.0, + causal: bool = False, + chunk_size: Tuple[int] = [-1], + left_context_frames: Tuple[int] = [-1], + use_adapters: bool = False, + adapter_dim: int = 16, + ) -> None: + super(Zipformer2, self).__init__() + + if dropout is None: + dropout = ScheduledFloat((0.0, 0.3), (20000.0, 0.1)) + + def _to_tuple(x): + """Converts a single int or a 1-tuple of an int to a tuple with the same length + as downsampling_factor""" + if isinstance(x, int): + x = (x,) + if len(x) == 1: + x = x * len(downsampling_factor) + else: + assert len(x) == len(downsampling_factor) and isinstance(x[0], int) + return x + + self.output_downsampling_factor = output_downsampling_factor # int + self.downsampling_factor = downsampling_factor # tuple + self.encoder_dim = encoder_dim = _to_tuple(encoder_dim) # tuple + self.encoder_unmasked_dim = encoder_unmasked_dim = _to_tuple( + encoder_unmasked_dim + ) # tuple + num_encoder_layers = _to_tuple(num_encoder_layers) + self.num_encoder_layers = num_encoder_layers + self.query_head_dim = query_head_dim = _to_tuple(query_head_dim) + self.value_head_dim = value_head_dim = _to_tuple(value_head_dim) + pos_head_dim = _to_tuple(pos_head_dim) + self.num_heads = num_heads = _to_tuple(num_heads) + feedforward_dim = _to_tuple(feedforward_dim) + self.cnn_module_kernel = cnn_module_kernel = _to_tuple(cnn_module_kernel) + + self.causal = causal + self.chunk_size = chunk_size + self.left_context_frames = left_context_frames + self.use_adapters = use_adapters + + for u, d in zip(encoder_unmasked_dim, encoder_dim): + assert u <= d + + # each one will be Zipformer2Encoder or DownsampledZipformer2Encoder + encoders = [] + + num_encoders = len(downsampling_factor) + for i in range(num_encoders): + encoder_layer = Zipformer2EncoderLayer( + embed_dim=encoder_dim[i], + pos_dim=pos_dim, + num_heads=num_heads[i], + query_head_dim=query_head_dim[i], + pos_head_dim=pos_head_dim[i], + value_head_dim=value_head_dim[i], + feedforward_dim=feedforward_dim[i], + dropout=dropout, + cnn_module_kernel=cnn_module_kernel[i], + causal=causal, + use_adapters=use_adapters, + adapter_dim=adapter_dim, + ) + + # For the segment of the warmup period, we let the Conv2dSubsampling + # layer learn something. Then we start to warm up the other encoders. + encoder = Zipformer2Encoder( + encoder_layer, + num_encoder_layers[i], + pos_dim=pos_dim, + dropout=dropout, + warmup_begin=warmup_batches * (i + 1) / (num_encoders + 1), + warmup_end=warmup_batches * (i + 2) / (num_encoders + 1), + final_layerdrop_rate=0.035 * (downsampling_factor[i] ** 0.5), + ) + + if downsampling_factor[i] != 1: + encoder = DownsampledZipformer2Encoder( + encoder, + dim=encoder_dim[i], + downsample=downsampling_factor[i], + dropout=dropout, + ) + + encoders.append(encoder) + + self.encoders = nn.ModuleList(encoders) + + self.downsample_output = SimpleDownsample( + max(encoder_dim), downsample=output_downsampling_factor, dropout=dropout + ) + + def get_feature_masks(self, x: Tensor) -> Union[List[float], List[Tensor]]: + """ + In eval mode, returns [1.0] * num_encoders; in training mode, returns a number of + randomized feature masks, one per encoder. + On e.g. 15% of frames, these masks will zero out all enocder dims larger than + some supplied number, e.g. >256, so in effect on those frames we are using + a smaller encoer dim. + + We generate the random masks at this level because we want the 2 masks to 'agree' + all the way up the encoder stack. This will mean that the 1st mask will have + mask values repeated self.zipformer_subsampling_factor times. + + Args: + x: the embeddings (needed for the shape and dtype and device), of shape + (1, batch_size, encoder_dims0) + """ + num_encoders = len(self.encoder_dim) + if not self.training: + return [1.0] * num_encoders + + (num_frames0, batch_size, _encoder_dims0) = x.shape + + assert self.encoder_dim[0] == _encoder_dims0, ( + self.encoder_dim[0], + _encoder_dims0, + ) + + feature_mask_dropout_prob = 0.125 + + # mask1 shape: (1, batch_size, 1) + mask1 = ( + torch.rand(1, batch_size, 1, device=x.device) > feature_mask_dropout_prob + ).to(x.dtype) + + # mask2 has additional sequences masked, about twice the number. + mask2 = torch.logical_and( + mask1, + ( + torch.rand(1, batch_size, 1, device=x.device) + > feature_mask_dropout_prob + ).to(x.dtype), + ) + + # dim: (1, batch_size, 2) + mask = torch.cat((mask1, mask2), dim=-1) + + feature_masks = [] + for i in range(num_encoders): + channels = self.encoder_dim[i] + feature_mask = torch.ones( + 1, batch_size, channels, dtype=x.dtype, device=x.device + ) + u1 = self.encoder_unmasked_dim[i] + u2 = u1 + (channels - u1) // 2 + + feature_mask[:, :, u1:u2] *= mask[..., 0:1] + feature_mask[:, :, u2:] *= mask[..., 1:2] + + feature_masks.append(feature_mask) + + return feature_masks + + def get_chunk_info(self) -> Tuple[int, int]: + """ + Returns chunk_size and left_context_chunks. + """ + if not self.causal: + return -1, -1 + + if torch.jit.is_scripting() or torch.jit.is_tracing(): + assert len(self.chunk_size) == 1, self.chunk_size + chunk_size = self.chunk_size[0] + else: + chunk_size = random.choice(self.chunk_size) + + if chunk_size == -1: + left_context_chunks = -1 + else: + if torch.jit.is_scripting() or torch.jit.is_tracing(): + assert len(self.left_context_frames) == 1, self.left_context_frames + left_context_frames = self.left_context_frames[0] + else: + left_context_frames = random.choice(self.left_context_frames) + # Note: in Python, -1 // n == -1 for n > 0 + left_context_chunks = left_context_frames // chunk_size + if left_context_chunks == 0: + left_context_chunks = 1 + + return chunk_size, left_context_chunks + + def forward( + self, + x: Tensor, + x_lens: Tensor, + src_key_padding_mask: Optional[Tensor] = None, + ) -> Tuple[Tensor, Tensor]: + """ + Args: + x: + The input tensor. Its shape is (seq_len, batch_size, feature_dim). + x_lens: + A tensor of shape (batch_size,) containing the number of frames in + `x` before padding. + src_key_padding_mask: + The mask for padding, of shape (batch_size, seq_len); True means + masked position. May be None. + Returns: + Return a tuple containing 2 tensors: + - embeddings: its shape is (output_seq_len, batch_size, max(encoder_dim)) + - lengths, a tensor of shape (batch_size,) containing the number + of frames in `embeddings` before padding. + """ + outputs = [] + if torch.jit.is_scripting() or torch.jit.is_tracing(): + feature_masks = [1.0] * len(self.encoder_dim) + else: + feature_masks = self.get_feature_masks(x) + + chunk_size, left_context_chunks = self.get_chunk_info() + + if torch.jit.is_scripting() or torch.jit.is_tracing(): + # Not support exporting a model for simulating streaming decoding + attn_mask = None + else: + attn_mask = self._get_attn_mask(x, chunk_size, left_context_chunks) + + for i, module in enumerate(self.encoders): + ds = self.downsampling_factor[i] + x = convert_num_channels(x, self.encoder_dim[i]) + + x = module( + x, + chunk_size=chunk_size, + feature_mask=feature_masks[i], + src_key_padding_mask=( + None + if src_key_padding_mask is None + else src_key_padding_mask[..., ::ds] + ), + attn_mask=attn_mask, + ) + outputs.append(x) + + # if the last output has the largest dimension, x will be unchanged, + # it will be the same as outputs[-1]. Otherwise it will be concatenated + # from different pieces of 'outputs', taking each dimension from the + # most recent output that has it present. + x = self._get_full_dim_output(outputs) + x = self.downsample_output(x) + # class Downsample has this rounding behavior.. + assert self.output_downsampling_factor == 2, self.output_downsampling_factor + if torch.jit.is_scripting() or torch.jit.is_tracing(): + lengths = (x_lens + 1) // 2 + else: + with warnings.catch_warnings(): + warnings.simplefilter("ignore") + lengths = (x_lens + 1) // 2 + + return x, lengths + + def _get_attn_mask( + self, x: Tensor, chunk_size: int, left_context_chunks: int + ) -> Optional[Tensor]: + """ + Return None if chunk_size == -1, else return attention mask of shape + (seq_len, seq_len), interpreted as (tgt_seq_len, src_seq_len). True + means a masked position. + Args: + x: embeddings after self.encoder_embed(), of shape (seq_len, batch_size, embed_dim). + chunk_size: chunk size, must divide + """ + if chunk_size <= 0: + return None + assert all(chunk_size % d == 0 for d in self.downsampling_factor) + if left_context_chunks >= 0: + num_encoders = len(self.encoder_dim) + assert all( + chunk_size * left_context_chunks + >= (self.cnn_module_kernel[i] // 2) * self.downsampling_factor[i] + for i in range(num_encoders) + ) + else: + left_context_chunks = 1000000 + + seq_len = x.shape[0] + + # t is frame index, shape (seq_len,) + t = torch.arange(seq_len, dtype=torch.int32, device=x.device) + # c is chunk index for each frame, shape (seq_len,) + if torch.jit.is_scripting() or torch.jit.is_tracing(): + c = t // chunk_size + else: + with warnings.catch_warnings(): + warnings.simplefilter("ignore") + c = t // chunk_size + src_c = c + tgt_c = c.unsqueeze(-1) + + attn_mask = torch.logical_or(src_c > tgt_c, src_c < tgt_c - left_context_chunks) + if __name__ == "__main__": + logging.info(f"attn_mask = {attn_mask}") + return attn_mask + + def _get_full_dim_output(self, outputs: List[Tensor]): + num_encoders = len(self.encoder_dim) + assert len(outputs) == num_encoders + output_dim = max(self.encoder_dim) + output_pieces = [outputs[-1]] + cur_dim = self.encoder_dim[-1] + for i in range(num_encoders - 2, -1, -1): + d = self.encoder_dim[i] + if d > cur_dim: + this_output = outputs[i] + output_pieces.append(this_output[..., cur_dim:d]) + cur_dim = d + assert cur_dim == output_dim + return torch.cat(output_pieces, dim=-1) + + def streaming_forward( + self, + x: Tensor, + x_lens: Tensor, + states: List[Tensor], + src_key_padding_mask: Tensor, + ) -> Tuple[Tensor, Tensor, List[Tensor]]: + """ + Args: + x: + The input tensor. Its shape is (seq_len, batch_size, feature_dim). + x_lens: + A tensor of shape (batch_size,) containing the number of frames in + `x` before padding. + states: list of cached tensors of all encoder layers. For layer-i, + states[i*6:(i+1)*6] is (cached_key, cached_nonlin_attn, cached_val1, cached_val2, + cached_conv1, cached_conv2). + src_key_padding_mask: + The mask for padding, of shape (batch_size, seq_len); True means + masked position. May be None. + Returns: + Return a tuple containing 2 tensors: + - embeddings: its shape is (output_seq_len, batch_size, max(encoder_dim)) + - lengths, a tensor of shape (batch_size,) containing the number + of frames in `embeddings` before padding. + - updated states + """ + outputs = [] + new_states = [] + layer_offset = 0 + + for i, module in enumerate(self.encoders): + num_layers = module.num_layers + ds = self.downsampling_factor[i] + x = convert_num_channels(x, self.encoder_dim[i]) + + x, new_layer_states = module.streaming_forward( + x, + states=states[layer_offset * 6 : (layer_offset + num_layers) * 6], + left_context_len=self.left_context_frames[0] // ds, + src_key_padding_mask=src_key_padding_mask[..., ::ds], + ) + layer_offset += num_layers + outputs.append(x) + new_states += new_layer_states + + # if the last output has the largest dimension, x will be unchanged, + # it will be the same as outputs[-1]. Otherwise it will be concatenated + # from different pieces of 'outputs', taking each dimension from the + # most recent output that has it present. + x = self._get_full_dim_output(outputs) + x = self.downsample_output(x) + # class Downsample has this rounding behavior.. + assert self.output_downsampling_factor == 2 + if torch.jit.is_scripting() or torch.jit.is_tracing(): + lengths = (x_lens + 1) // 2 + else: + with warnings.catch_warnings(): + warnings.simplefilter("ignore") + lengths = (x_lens + 1) // 2 + + return x, lengths, new_states + + @torch.jit.export + def get_init_states( + self, + batch_size: int = 1, + device: torch.device = torch.device("cpu"), + ) -> List[Tensor]: + """Get initial states. + + A list of cached tensors of all encoder layers. For layer-i, states[i*6:(i+1)*6] + is (cached_key, cached_nonlin_attn, cached_val1, cached_val2, cached_conv1, cached_conv2). + """ + states = [] + for i, module in enumerate(self.encoders): + num_layers = module.num_layers + embed_dim = self.encoder_dim[i] + ds = self.downsampling_factor[i] + num_heads = self.num_heads[i] + key_dim = self.query_head_dim[i] * num_heads + value_dim = self.value_head_dim[i] * num_heads + downsample_left = self.left_context_frames[0] // ds + nonlin_attn_head_dim = 3 * embed_dim // 4 + conv_left_pad = self.cnn_module_kernel[i] // 2 + for layer in range(num_layers): + cached_key = torch.zeros(downsample_left, batch_size, key_dim).to( + device + ) + cached_nonlin_attn = torch.zeros( + 1, batch_size, downsample_left, nonlin_attn_head_dim + ).to(device) + cached_val1 = torch.zeros(downsample_left, batch_size, value_dim).to( + device + ) + cached_val2 = torch.zeros(downsample_left, batch_size, value_dim).to( + device + ) + cached_conv1 = torch.zeros(batch_size, embed_dim, conv_left_pad).to( + device + ) + cached_conv2 = torch.zeros(batch_size, embed_dim, conv_left_pad).to( + device + ) + states += [ + cached_key, + cached_nonlin_attn, + cached_val1, + cached_val2, + cached_conv1, + cached_conv2, + ] + + return states + + +def _whitening_schedule(x: float, ratio: float = 2.0) -> ScheduledFloat: + return ScheduledFloat((0.0, x), (20000.0, ratio * x), default=x) + + +def _balancer_schedule(min_prob: float): + return ScheduledFloat((0.0, 0.4), (8000.0, min_prob)) + + +class Zipformer2EncoderLayer(nn.Module): + """ + Args: + embed_dim: the number of expected features in the input (required). + nhead: the number of heads in the multiheadattention models (required). + feedforward_dim: the dimension of the feedforward network model (default=2048). + dropout: the dropout value (default=0.1). + cnn_module_kernel (int): Kernel size of convolution module. + use_adapters: insert adapters in each layer + adapter_dim: the bottleneck dimension of the adapter + + Examples:: + >>> encoder_layer = Zipformer2EncoderLayer(embed_dim=512, nhead=8) + >>> src = torch.rand(10, 32, 512) + >>> pos_emb = torch.rand(32, 19, 512) + >>> out = encoder_layer(src, pos_emb) + """ + + def __init__( + self, + embed_dim: int, + pos_dim: int, + num_heads: int, + query_head_dim: int, + pos_head_dim: int, + value_head_dim: int, + feedforward_dim: int, + dropout: FloatLike = 0.1, + cnn_module_kernel: int = 31, + causal: bool = False, + attention_skip_rate: FloatLike = ScheduledFloat( + (0.0, 0.2), (4000.0, 0.05), (16000, 0.0), default=0 + ), + conv_skip_rate: FloatLike = ScheduledFloat( + (0.0, 0.2), (4000.0, 0.05), (16000, 0.0), default=0 + ), + const_attention_rate: FloatLike = ScheduledFloat( + (0.0, 0.25), (4000.0, 0.025), default=0 + ), + ff2_skip_rate: FloatLike = ScheduledFloat( + (0.0, 0.1), (4000.0, 0.01), (50000.0, 0.0) + ), + ff3_skip_rate: FloatLike = ScheduledFloat( + (0.0, 0.1), (4000.0, 0.01), (50000.0, 0.0) + ), + bypass_skip_rate: FloatLike = ScheduledFloat( + (0.0, 0.5), (4000.0, 0.02), default=0 + ), + use_adapters: bool = False, + adapter_dim: int = 16, + ) -> None: + super(Zipformer2EncoderLayer, self).__init__() + self.embed_dim = embed_dim + + # self.bypass implements layer skipping as well as bypass; see its default values. + self.bypass = BypassModule( + embed_dim, skip_rate=bypass_skip_rate, straight_through_rate=0 + ) + # bypass_mid is bypass used in the middle of the layer. + self.bypass_mid = BypassModule(embed_dim, straight_through_rate=0) + + # skip probability for dynamic modules (meaning: anything but feedforward). + self.attention_skip_rate = copy.deepcopy(attention_skip_rate) + # an additional skip probability that applies to ConvModule to stop it from + # contributing too much early on. + self.conv_skip_rate = copy.deepcopy(conv_skip_rate) + + # ff2_skip_rate is to prevent the ff2 module from having output that's too big + # compared to its residual. + self.ff2_skip_rate = copy.deepcopy(ff2_skip_rate) + self.ff3_skip_rate = copy.deepcopy(ff3_skip_rate) + + self.const_attention_rate = copy.deepcopy(const_attention_rate) + + self.self_attn_weights = RelPositionMultiheadAttentionWeights( + embed_dim, + pos_dim=pos_dim, + num_heads=num_heads, + query_head_dim=query_head_dim, + pos_head_dim=pos_head_dim, + dropout=0.0, + ) + + self.self_attn1 = SelfAttention(embed_dim, num_heads, value_head_dim) + + self.self_attn2 = SelfAttention(embed_dim, num_heads, value_head_dim) + + self.feed_forward1 = FeedforwardModule( + embed_dim, (feedforward_dim * 3) // 4, dropout + ) + + self.feed_forward2 = FeedforwardModule(embed_dim, feedforward_dim, dropout) + + self.feed_forward3 = FeedforwardModule( + embed_dim, (feedforward_dim * 5) // 4, dropout + ) + + self.nonlin_attention = NonlinAttention( + embed_dim, hidden_channels=3 * embed_dim // 4 + ) + + self.conv_module1 = ConvolutionModule( + embed_dim, cnn_module_kernel, causal=causal + ) + + self.conv_module2 = ConvolutionModule( + embed_dim, cnn_module_kernel, causal=causal + ) + + # TODO: remove it + self.bypass_scale = nn.Parameter(torch.full((embed_dim,), 0.5)) + + self.norm = BiasNorm(embed_dim) + + self.balancer1 = Balancer( + embed_dim, + channel_dim=-1, + min_positive=0.45, + max_positive=0.55, + min_abs=0.2, + max_abs=4.0, + ) + + # balancer for output of NonlinAttentionModule + self.balancer_na = Balancer( + embed_dim, + channel_dim=-1, + min_positive=0.3, + max_positive=0.7, + min_abs=ScheduledFloat((0.0, 0.004), (4000.0, 0.02)), + prob=0.05, # out of concern for memory usage + ) + + # balancer for output of feedforward2, prevent it from staying too + # small. give this a very small probability, even at the start of + # training, it's to fix a rare problem and it's OK to fix it slowly. + self.balancer_ff2 = Balancer( + embed_dim, + channel_dim=-1, + min_positive=0.3, + max_positive=0.7, + min_abs=ScheduledFloat((0.0, 0.0), (4000.0, 0.1), default=0.0), + max_abs=2.0, + prob=0.05, + ) + + self.balancer_ff3 = Balancer( + embed_dim, + channel_dim=-1, + min_positive=0.3, + max_positive=0.7, + min_abs=ScheduledFloat((0.0, 0.0), (4000.0, 0.2), default=0.0), + max_abs=4.0, + prob=0.05, + ) + + self.whiten = Whiten( + num_groups=1, + whitening_limit=_whitening_schedule(4.0, ratio=3.0), + prob=(0.025, 0.25), + grad_scale=0.01, + ) + + self.balancer2 = Balancer( + embed_dim, + channel_dim=-1, + min_positive=0.45, + max_positive=0.55, + min_abs=0.1, + max_abs=4.0, + ) + + self.use_adapters = use_adapters + if use_adapters: + self.mid_adapter = AdapterModule( + embed_dim=embed_dim, + bottleneck_dim=adapter_dim, + ) + + # placed after the 1st self-attn module + self.post_sa_adapter = AdapterModule( + embed_dim=embed_dim, + bottleneck_dim=adapter_dim, + ) + + # placed after the 2nd convolution module + self.post_conv_adapter = AdapterModule( + embed_dim=embed_dim, + bottleneck_dim=adapter_dim, + ) + + # at the end of each layer + self.adapter = AdapterModule( + embed_dim=embed_dim, + bottleneck_dim=adapter_dim, + ) + else: + self.mid_adapter = None + self.post_sa_adapter = None + self.post_conv_adapter = None + self.adapter = None + + def get_sequence_dropout_mask( + self, x: Tensor, dropout_rate: float + ) -> Optional[Tensor]: + if ( + dropout_rate == 0.0 + or not self.training + or torch.jit.is_scripting() + or torch.jit.is_tracing() + ): + return None + batch_size = x.shape[1] + mask = (torch.rand(batch_size, 1, device=x.device) > dropout_rate).to(x.dtype) + return mask + + def sequence_dropout(self, x: Tensor, dropout_rate: float) -> Tensor: + """ + Apply sequence-level dropout to x. + x shape: (seq_len, batch_size, embed_dim) + """ + dropout_mask = self.get_sequence_dropout_mask(x, dropout_rate) + if dropout_mask is None: + return x + else: + return x * dropout_mask + + def forward( + self, + src: Tensor, + pos_emb: Tensor, + chunk_size: int = -1, + attn_mask: Optional[Tensor] = None, + src_key_padding_mask: Optional[Tensor] = None, + ) -> Tensor: + """ + Pass the input through the encoder layer. + Args: + src: the sequence to the encoder (required): shape (seq_len, batch_size, embedding_dim). + pos_emb: (1, 2*seq_len-1, pos_emb_dim) or (batch_size, 2*seq_len-1, pos_emb_dim) + chunk_size: the number of frames per chunk, of >= 0; if -1, no chunking. + feature_mask: something that broadcasts with src, that we'll multiply `src` + by at every layer: if a Tensor, likely of shape (seq_len, batch_size, embedding_dim) + attn_mask: the attention mask, of shape (batch_size, seq_len, seq_len) or (seq_len, seq_len), + interpreted as (batch_size, tgt_seq_len, src_seq_len) or (tgt_seq_len, src_seq_len). + True means masked position. May be None. + src_key_padding_mask: the mask for padding, of shape (batch_size, seq_len); True means + masked position. May be None. + + Returns: + A tensor which has the same shape as src + """ + src_orig = src + + # dropout rate for non-feedforward submodules + if torch.jit.is_scripting() or torch.jit.is_tracing(): + attention_skip_rate = 0.0 + else: + attention_skip_rate = ( + float(self.attention_skip_rate) if self.training else 0.0 + ) + + # attn_weights: (num_heads, batch_size, seq_len, seq_len) + attn_weights = self.self_attn_weights( + src, + pos_emb=pos_emb, + attn_mask=attn_mask, + key_padding_mask=src_key_padding_mask, + ) + + src = src + self.feed_forward1(src) + + self_attn_dropout_mask = self.get_sequence_dropout_mask( + src, attention_skip_rate + ) + + selected_attn_weights = attn_weights[0:1] + if torch.jit.is_scripting() or torch.jit.is_tracing(): + pass + elif not self.training and random.random() < float(self.const_attention_rate): + # Make attention weights constant. The intention is to + # encourage these modules to do something similar to an + # averaging-over-time operation. + # only need the mask, can just use the 1st one and expand later + selected_attn_weights = selected_attn_weights[0:1] + selected_attn_weights = (selected_attn_weights > 0.0).to( + selected_attn_weights.dtype + ) + selected_attn_weights = selected_attn_weights * ( + 1.0 / selected_attn_weights.sum(dim=-1, keepdim=True) + ) + + na = self.balancer_na(self.nonlin_attention(src, selected_attn_weights)) + + src = src + ( + na if self_attn_dropout_mask is None else na * self_attn_dropout_mask + ) + + self_attn = self.self_attn1(src, attn_weights) + + src = src + ( + self_attn + if self_attn_dropout_mask is None + else self_attn * self_attn_dropout_mask + ) + + if self.use_adapters and self.post_sa_adapter is not None: + src = self.post_sa_adapter(src) + + if torch.jit.is_scripting() or torch.jit.is_tracing(): + conv_skip_rate = 0.0 + else: + conv_skip_rate = float(self.conv_skip_rate) if self.training else 0.0 + src = src + self.sequence_dropout( + self.conv_module1( + src, chunk_size=chunk_size, src_key_padding_mask=src_key_padding_mask + ), + conv_skip_rate, + ) + + if torch.jit.is_scripting() or torch.jit.is_tracing(): + ff2_skip_rate = 0.0 + else: + ff2_skip_rate = float(self.ff2_skip_rate) if self.training else 0.0 + src = src + self.sequence_dropout( + self.balancer_ff2(self.feed_forward2(src)), ff2_skip_rate + ) + + # bypass in the middle of the layer. + src = self.bypass_mid(src_orig, src) + + if self.use_adapters and self.mid_adapter is not None: + src = self.mid_adapter(src) + + self_attn = self.self_attn2(src, attn_weights) + + src = src + ( + self_attn + if self_attn_dropout_mask is None + else self_attn * self_attn_dropout_mask + ) + + if torch.jit.is_scripting() or torch.jit.is_tracing(): + conv_skip_rate = 0.0 + else: + conv_skip_rate = float(self.conv_skip_rate) if self.training else 0.0 + src = src + self.sequence_dropout( + self.conv_module2( + src, chunk_size=chunk_size, src_key_padding_mask=src_key_padding_mask + ), + conv_skip_rate, + ) + + if self.use_adapters and self.post_conv_adapter is not None: + src = self.post_conv_adapter(src) + + if torch.jit.is_scripting() or torch.jit.is_tracing(): + ff3_skip_rate = 0.0 + else: + ff3_skip_rate = float(self.ff3_skip_rate) if self.training else 0.0 + src = src + self.sequence_dropout( + self.balancer_ff3(self.feed_forward3(src)), ff3_skip_rate + ) + + src = self.balancer1(src) + src = self.norm(src) + + src = self.bypass(src_orig, src) + + src = self.balancer2(src) + src = self.whiten(src) + + if self.use_adapters and self.adapter is not None: + src = self.adapter(src) + + return src + + def streaming_forward( + self, + src: Tensor, + pos_emb: Tensor, + cached_key: Tensor, + cached_nonlin_attn: Tensor, + cached_val1: Tensor, + cached_val2: Tensor, + cached_conv1: Tensor, + cached_conv2: Tensor, + left_context_len: int, + src_key_padding_mask: Tensor, + ) -> Tuple[Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor]: + """Pass the input through the encoder layer in streaming forward mode. + + Args: + src: the sequence to the encoder (required): shape (seq_len, batch_size, embedding_dim). + pos_emb: (1, left_context_len+2*seq_len-1, pos_emb_dim) or + (batch_size, left_context_len+2*seq_len-1, pos_emb_dim) + cached_key: cached attention key tensor of left context, + of shape (left_context_len, batch_size, key_dim) + cached_nonlin_attn: left context for nonlin_attention module, a Tensor of shape + (num_heads, batch_size, left_context_len, head_dim) + cached_val1: cached left context for the first attention module, + of shape (left_context_len, batch_size, value_dim) + cached_val2: cached left context for the second attention module, + of shape (left_context_len, batch_size, value_dim) + cached_conv1: cached left context for the first convolution module, + of shape (batch_size, channels, left_pad) + cached_conv2: cached left context for the second convolution module, + of shape (batch_size, channels, left_pad) + left_context_len: number of left context frames. + src_key_padding_mask: the mask for padding, of shape + (batch_size, left_context_len + seq_len); True means masked position. + May be None. + + Returns: + - x, with the same shape as src + - updated cached_key + - updated cached_nonlin_attn + - updated cached_val1 + - updated cached_val2 + - updated cached_conv1 + - updated cached_conv2 + """ + src_orig = src + + # attn_weights: (num_heads, batch_size, seq_len, seq_len) + attn_weights, cached_key = self.self_attn_weights.streaming_forward( + src, + pos_emb=pos_emb, + cached_key=cached_key, + left_context_len=left_context_len, + key_padding_mask=src_key_padding_mask, + ) + + src = src + self.feed_forward1(src) + + na, cached_nonlin_attn = self.nonlin_attention.streaming_forward( + src, + attn_weights[0:1], + cached_x=cached_nonlin_attn, + left_context_len=left_context_len, + ) + src = src + na + + self_attn, cached_val1 = self.self_attn1.streaming_forward( + src, + attn_weights=attn_weights, + cached_val=cached_val1, + left_context_len=left_context_len, + ) + src = src + self_attn + + src_conv, cached_conv1 = self.conv_module1.streaming_forward( + src, + cache=cached_conv1, + src_key_padding_mask=src_key_padding_mask[:, left_context_len:], + ) + src = src + src_conv + + src = src + self.feed_forward2(src) + + # bypass in the middle of the layer. + src = self.bypass_mid(src_orig, src) + + self_attn, cached_val2 = self.self_attn2.streaming_forward( + src, + attn_weights=attn_weights, + cached_val=cached_val2, + left_context_len=left_context_len, + ) + src = src + self_attn + + src_conv, cached_conv2 = self.conv_module2.streaming_forward( + src, + cache=cached_conv2, + src_key_padding_mask=src_key_padding_mask[:, left_context_len:], + ) + src = src + src_conv + + src = src + self.feed_forward3(src) + + src = self.norm(src) + + src = self.bypass(src_orig, src) + + return ( + src, + cached_key, + cached_nonlin_attn, + cached_val1, + cached_val2, + cached_conv1, + cached_conv2, + ) + + +class Zipformer2Encoder(nn.Module): + r"""Zipformer2Encoder is a stack of N encoder layers + + Args: + encoder_layer: an instance of the Zipformer2EncoderLayer() class (required). + num_layers: the number of sub-encoder-layers in the encoder (required). + pos_dim: the dimension for the relative positional encoding + + Examples:: + >>> encoder_layer = Zipformer2EncoderLayer(embed_dim=512, nhead=8) + >>> zipformer_encoder = Zipformer2Encoder(encoder_layer, num_layers=6) + >>> src = torch.rand(10, 32, 512) + >>> out = zipformer_encoder(src) + """ + + def __init__( + self, + encoder_layer: nn.Module, + num_layers: int, + pos_dim: int, + dropout: float, + warmup_begin: float, + warmup_end: float, + initial_layerdrop_rate: float = 0.5, + final_layerdrop_rate: float = 0.05, + ) -> None: + super().__init__() + self.encoder_pos = CompactRelPositionalEncoding( + pos_dim, dropout_rate=0.15, length_factor=1.0 + ) + + self.layers = nn.ModuleList( + [copy.deepcopy(encoder_layer) for i in range(num_layers)] + ) + self.num_layers = num_layers + + assert 0 <= warmup_begin <= warmup_end + + delta = (1.0 / num_layers) * (warmup_end - warmup_begin) + cur_begin = warmup_begin # interpreted as a training batch index + for i in range(num_layers): + cur_end = cur_begin + delta + self.layers[i].bypass.skip_rate = ScheduledFloat( + (cur_begin, initial_layerdrop_rate), + (cur_end, final_layerdrop_rate), + default=0.0, + ) + cur_begin = cur_end + + def forward( + self, + src: Tensor, + chunk_size: int = -1, + feature_mask: Union[Tensor, float] = 1.0, + attn_mask: Optional[Tensor] = None, + src_key_padding_mask: Optional[Tensor] = None, + ) -> Tensor: + r"""Pass the input through the encoder layers in turn. + + Args: + src: the sequence to the encoder (required): shape (seq_len, batch_size, embedding_dim). + chunk_size: the number of frames per chunk, of >= 0; if -1, no chunking. + feature_mask: something that broadcasts with src, that we'll multiply `src` + by at every layer: if a Tensor, likely of shape (seq_len, batch_size, embedding_dim) + attn_mask: the attention mask, of shape (batch_size, seq_len, seq_len) or (seq_len, seq_len), + interpreted as (batch_size, tgt_seq_len, src_seq_len) or (tgt_seq_len, src_seq_len). + True means masked position. May be None. + src_key_padding_mask: the mask for padding, of shape (batch_size, seq_len); True means + masked position. May be None. + + Returns: a Tensor with the same shape as src. + """ + pos_emb = self.encoder_pos(src) + output = src + + if not torch.jit.is_scripting() and not torch.jit.is_tracing(): + output = output * feature_mask + + for i, mod in enumerate(self.layers): + output = mod( + output, + pos_emb, + chunk_size=chunk_size, + attn_mask=attn_mask, + src_key_padding_mask=src_key_padding_mask, + ) + + if not torch.jit.is_scripting() and not torch.jit.is_tracing(): + output = output * feature_mask + + return output + + def streaming_forward( + self, + src: Tensor, + states: List[Tensor], + left_context_len: int, + src_key_padding_mask: Tensor, + ) -> Tuple[Tensor, List[Tensor]]: + r"""Pass the input through the encoder layers in turn. + + Args: + src: the sequence to the encoder (required): shape (seq_len, batch_size, embedding_dim). + states: list of cached tensors of N encoder layers. For layer-i, states[i*6:(i+1)*6] is + (cached_key, cached_nonlin_attn, cached_val1, cached_val2, cached_conv1, cached_conv2). + left_context_len: Number of left context frames. + src_key_padding_mask: the mask for padding, of shape + (batch_size, left_context_len + seq_len); True means masked position. + May be None. + + Returns: + - output, a Tensor with the same shape as src. + - updated states + """ + pos_emb = self.encoder_pos(src, left_context_len) + output = src + + new_states = [] + for i, mod in enumerate(self.layers): + ( + cached_key, + cached_nonlin_attn, + cached_val1, + cached_val2, + cached_conv1, + cached_conv2, + ) = states[i * 6 : (i + 1) * 6] + ( + output, + new_cached_key, + new_cached_nonlin_attn, + new_cached_val1, + new_cached_val2, + new_cached_conv1, + new_cached_conv2, + ) = mod.streaming_forward( + output, + pos_emb, + cached_key=cached_key, + cached_nonlin_attn=cached_nonlin_attn, + cached_val1=cached_val1, + cached_val2=cached_val2, + cached_conv1=cached_conv1, + cached_conv2=cached_conv2, + left_context_len=left_context_len, + src_key_padding_mask=src_key_padding_mask, + ) + new_states += [ + new_cached_key, + new_cached_nonlin_attn, + new_cached_val1, + new_cached_val2, + new_cached_conv1, + new_cached_conv2, + ] + + return output, new_states + + +class BypassModule(nn.Module): + """ + An nn.Module that implements a learnable bypass scale, and also randomized per-sequence + layer-skipping. The bypass is limited during early stages of training to be close to + "straight-through", i.e. to not do the bypass operation much initially, in order to + force all the modules to learn something. + """ + + def __init__( + self, + embed_dim: int, + skip_rate: FloatLike = 0.0, + straight_through_rate: FloatLike = 0.0, + scale_min: FloatLike = ScheduledFloat((0.0, 0.9), (20000.0, 0.2), default=0), + scale_max: FloatLike = 1.0, + ): + super().__init__() + self.bypass_scale = nn.Parameter(torch.full((embed_dim,), 0.5)) + self.skip_rate = copy.deepcopy(skip_rate) + self.straight_through_rate = copy.deepcopy(straight_through_rate) + self.scale_min = copy.deepcopy(scale_min) + self.scale_max = copy.deepcopy(scale_max) + + def _get_bypass_scale(self, batch_size: int): + # returns bypass-scale of shape (num_channels,), + # or (batch_size, num_channels,). This is actually the + # scale on the non-residual term, so 0 correponds to bypassing + # this module. + if torch.jit.is_scripting() or torch.jit.is_tracing() or not self.training: + return self.bypass_scale + else: + ans = limit_param_value( + self.bypass_scale, min=float(self.scale_min), max=float(self.scale_max) + ) + skip_rate = float(self.skip_rate) + if skip_rate != 0.0: + mask = torch.rand((batch_size, 1), device=ans.device) > skip_rate + ans = ans * mask + # now ans is of shape (batch_size, num_channels), and is zero for sequences + # on which we have randomly chosen to do layer-skipping. + straight_through_rate = float(self.straight_through_rate) + if straight_through_rate != 0.0: + mask = ( + torch.rand((batch_size, 1), device=ans.device) + < straight_through_rate + ) + ans = torch.maximum(ans, mask.to(ans.dtype)) + return ans + + def forward(self, src_orig: Tensor, src: Tensor): + """ + Args: src_orig and src are both of shape (seq_len, batch_size, num_channels) + Returns: something with the same shape as src and src_orig + """ + bypass_scale = self._get_bypass_scale(src.shape[1]) + return src_orig + (src - src_orig) * bypass_scale + + +class DownsampledZipformer2Encoder(nn.Module): + r""" + DownsampledZipformer2Encoder is a zipformer encoder evaluated at a reduced frame rate, + after convolutional downsampling, and then upsampled again at the output, and combined + with the origin input, so that the output has the same shape as the input. + """ + + def __init__( + self, encoder: nn.Module, dim: int, downsample: int, dropout: FloatLike + ): + super(DownsampledZipformer2Encoder, self).__init__() + self.downsample_factor = downsample + self.downsample = SimpleDownsample(dim, downsample, dropout) + self.num_layers = encoder.num_layers + self.encoder = encoder + self.upsample = SimpleUpsample(dim, downsample) + self.out_combiner = BypassModule(dim, straight_through_rate=0) + + def forward( + self, + src: Tensor, + chunk_size: int = -1, + feature_mask: Union[Tensor, float] = 1.0, + attn_mask: Optional[Tensor] = None, + src_key_padding_mask: Optional[Tensor] = None, + ) -> Tensor: + r"""Downsample, go through encoder, upsample. + + Args: + src: the sequence to the encoder (required): shape (seq_len, batch_size, embedding_dim). + feature_mask: something that broadcasts with src, that we'll multiply `src` + by at every layer: if a Tensor, likely of shape (seq_len, batch_size, embedding_dim) + attn_mask: the attention mask, of shape (batch_size, seq_len, seq_len) or (seq_len, seq_len), + interpreted as (batch_size, tgt_seq_len, src_seq_len) or (tgt_seq_len, src_seq_len). + True means masked position. May be None. + src_key_padding_mask: the mask for padding, of shape (batch_size, seq_len); True means + masked position. May be None. + + Returns: a Tensor with the same shape as src. + """ + src_orig = src + src = self.downsample(src) + ds = self.downsample_factor + if attn_mask is not None: + attn_mask = attn_mask[::ds, ::ds] + + src = self.encoder( + src, + chunk_size=chunk_size // ds, + feature_mask=feature_mask, + attn_mask=attn_mask, + src_key_padding_mask=src_key_padding_mask, + ) + src = self.upsample(src) + # remove any extra frames that are not a multiple of downsample_factor + src = src[: src_orig.shape[0]] + + return self.out_combiner(src_orig, src) + + def streaming_forward( + self, + src: Tensor, + states: List[Tensor], + left_context_len: int, + src_key_padding_mask: Tensor, + ) -> Tuple[Tensor, List[Tensor]]: + r"""Downsample, go through encoder, upsample, in streaming forward mode. + + Args: + src: the sequence to the encoder (required): shape (seq_len, batch_size, embedding_dim). + states: list of cached tensors of N encoder layers. For layer-i, states[i*6:(i+1)*6] is + (cached_key, cached_nonlin_attn, cached_val1, cached_val2, cached_conv1, cached_conv2). + left_context_len: Number of left context frames. + src_key_padding_mask: the mask for padding, of shape (batch_size, left_context_len+seq_len); + True means masked position. May be None. + + Returns: + - output, a Tensor with the same shape as src. + - updated states + """ + src_orig = src + src = self.downsample(src) + + src, new_states = self.encoder.streaming_forward( + src, + states=states, + left_context_len=left_context_len, + src_key_padding_mask=src_key_padding_mask, + ) + src = self.upsample(src) + # remove any extra frames that are not a multiple of downsample_factor + src = src[: src_orig.shape[0]] + + return self.out_combiner(src_orig, src), new_states + + +class SimpleDownsample(torch.nn.Module): + """ + Does downsampling with attention, by weighted sum, and a projection.. + """ + + def __init__(self, channels: int, downsample: int, dropout: FloatLike): + super(SimpleDownsample, self).__init__() + + self.bias = nn.Parameter(torch.zeros(downsample)) + + self.name = None # will be set from training code + self.dropout = copy.deepcopy(dropout) + + self.downsample = downsample + + def forward(self, src: Tensor) -> Tensor: + """ + x: (seq_len, batch_size, in_channels) + Returns a tensor of shape + ( (seq_len+downsample-1)//downsample, batch_size, channels) + """ + (seq_len, batch_size, in_channels) = src.shape + ds = self.downsample + d_seq_len = (seq_len + ds - 1) // ds + + # Pad to an exact multiple of self.downsample + # right-pad src, repeating the last element. + pad = d_seq_len * ds - seq_len + src_extra = src[src.shape[0] - 1 :].expand(pad, src.shape[1], src.shape[2]) + src = torch.cat((src, src_extra), dim=0) + assert src.shape[0] == d_seq_len * ds + + src = src.reshape(d_seq_len, ds, batch_size, in_channels) + + weights = self.bias.softmax(dim=0) + # weights: (downsample, 1, 1) + weights = weights.unsqueeze(-1).unsqueeze(-1) + + # ans1 is the first `in_channels` channels of the output + ans = (src * weights).sum(dim=1) + + return ans + + +class SimpleUpsample(torch.nn.Module): + """ + A very simple form of upsampling that mostly just repeats the input, but + also adds a position-specific bias. + """ + + def __init__(self, num_channels: int, upsample: int): + super(SimpleUpsample, self).__init__() + self.upsample = upsample + + def forward(self, src: Tensor) -> Tensor: + """ + x: (seq_len, batch_size, num_channels) + Returns a tensor of shape + ( (seq_len*upsample), batch_size, num_channels) + """ + upsample = self.upsample + (seq_len, batch_size, num_channels) = src.shape + src = src.unsqueeze(1).expand(seq_len, upsample, batch_size, num_channels) + src = src.reshape(seq_len * upsample, batch_size, num_channels) + return src + + +class CompactRelPositionalEncoding(torch.nn.Module): + """ + Relative positional encoding module. This version is "compact" meaning it is able to encode + the important information about the relative position in a relatively small number of dimensions. + The goal is to make it so that small differences between large relative offsets (e.g. 1000 vs. 1001) + make very little difference to the embedding. Such differences were potentially important + when encoding absolute position, but not important when encoding relative position because there + is now no need to compare two large offsets with each other. + + Our embedding works done by projecting the interval [-infinity,infinity] to a finite interval + using the atan() function, before doing the fourier transform of that fixed interval. The + atan() function would compress the "long tails" too small, + making it hard to distinguish between different magnitudes of large offsets, so we use a logarithmic + function to compress large offsets to a smaller range before applying atan(). + Scalings are chosen in such a way that the embedding can clearly distinguish invidual offsets as long + as they are quite close to the origin, e.g. abs(offset) <= about sqrt(embedding_dim) + + + Args: + embed_dim: Embedding dimension. + dropout_rate: Dropout rate. + max_len: Maximum input length: just a heuristic for initialization. + length_factor: a heuristic scale (should be >= 1.0) which, if larger, gives + less weight to small differences of offset near the origin. + """ + + def __init__( + self, + embed_dim: int, + dropout_rate: FloatLike, + max_len: int = 1000, + length_factor: float = 1.0, + ) -> None: + """Construct a CompactRelPositionalEncoding object.""" + super(CompactRelPositionalEncoding, self).__init__() + self.embed_dim = embed_dim + assert embed_dim % 2 == 0 + self.dropout = Dropout2(dropout_rate) + self.pe = None + assert length_factor >= 1.0 + self.length_factor = length_factor + self.extend_pe(torch.tensor(0.0).expand(max_len)) + + def extend_pe(self, x: Tensor, left_context_len: int = 0) -> None: + """Reset the positional encodings.""" + T = x.size(0) + left_context_len + + if self.pe is not None: + # self.pe contains both positive and negative parts + # the length of self.pe is 2 * input_len - 1 + if self.pe.size(0) >= T * 2 - 1: + self.pe = self.pe.to(dtype=x.dtype, device=x.device) + return + + # if T == 4, x would contain [ -3, -2, 1, 0, 1, 2, 3 ] + x = torch.arange(-(T - 1), T, device=x.device).to(torch.float32).unsqueeze(1) + + freqs = 1 + torch.arange(self.embed_dim // 2, device=x.device) + + # `compression_length` this is arbitrary/heuristic, if it is larger we have more resolution + # for small time offsets but less resolution for large time offsets. + compression_length = self.embed_dim**0.5 + # x_compressed, like X, goes from -infinity to infinity as T goes from -infinity to infinity; + # but it does so more slowly than T for large absolute values of T. + # The formula is chosen so that d(x_compressed )/dx is 1 around x == 0, which + # is important. + x_compressed = ( + compression_length + * x.sign() + * ((x.abs() + compression_length).log() - math.log(compression_length)) + ) + + # if self.length_factor == 1.0, then length_scale is chosen so that the + # FFT can exactly separate points close to the origin (T == 0). So this + # part of the formulation is not really heuristic. + # But empirically, for ASR at least, length_factor > 1.0 seems to work better. + length_scale = self.length_factor * self.embed_dim / (2.0 * math.pi) + + # note for machine implementations: if atan is not available, we can use: + # x.sign() * ((1 / (x.abs() + 1)) - 1) * (-math.pi/2) + # check on wolframalpha.com: plot(sign(x) * (1 / ( abs(x) + 1) - 1 ) * -pi/2 , atan(x)) + x_atan = (x_compressed / length_scale).atan() # results between -pi and pi + + cosines = (x_atan * freqs).cos() + sines = (x_atan * freqs).sin() + + pe = torch.zeros(x.shape[0], self.embed_dim, device=x.device) + pe[:, 0::2] = cosines + pe[:, 1::2] = sines + pe[:, -1] = 1.0 # for bias. + + self.pe = pe.to(dtype=x.dtype) + + def forward(self, x: Tensor, left_context_len: int = 0) -> Tensor: + """Create positional encoding. + + Args: + x (Tensor): Input tensor (time, batch, `*`). + left_context_len: (int): Length of cached left context. + + Returns: + positional embedding, of shape (batch, left_context_len + 2*time-1, `*`). + """ + self.extend_pe(x, left_context_len) + x_size_left = x.size(0) + left_context_len + # length of positive side: x.size(0) + left_context_len + # length of negative side: x.size(0) + pos_emb = self.pe[ + self.pe.size(0) // 2 + - x_size_left + + 1 : self.pe.size(0) // 2 # noqa E203 + + x.size(0), + :, + ] + pos_emb = pos_emb.unsqueeze(0) + return self.dropout(pos_emb) + + +class RelPositionMultiheadAttentionWeights(nn.Module): + r"""Module that computes multi-head attention weights with relative position encoding. + Various other modules consume the resulting attention weights: see, for example, the + SimpleAttention module which allows you to compute conventional attention. + + This is a quite heavily modified from: "Transformer-XL: Attentive Language Models Beyond a Fixed-Length Context", + we have to write up the differences. + + + Args: + embed_dim: number of channels at the input to this module, e.g. 256 + pos_dim: dimension of the positional encoding vectors, e.g. 128. + num_heads: number of heads to compute weights for, e.g. 8 + query_head_dim: dimension of the query (and key), per head. e.g. 24. + pos_head_dim: dimension of the projected positional encoding per head, e.g. 4. + dropout: dropout probability for attn_output_weights. Default: 0.0. + pos_emb_skip_rate: probability for skipping the pos_emb part of the scores on + any given call to forward(), in training time. + """ + + def __init__( + self, + embed_dim: int, + pos_dim: int, + num_heads: int, + query_head_dim: int, + pos_head_dim: int, + dropout: float = 0.0, + pos_emb_skip_rate: FloatLike = ScheduledFloat((0.0, 0.5), (4000.0, 0.0)), + ) -> None: + super().__init__() + self.embed_dim = embed_dim + self.num_heads = num_heads + self.query_head_dim = query_head_dim + self.pos_head_dim = pos_head_dim + self.dropout = dropout + self.pos_emb_skip_rate = copy.deepcopy(pos_emb_skip_rate) + self.name = None # will be overwritten in training code; for diagnostics. + + key_head_dim = query_head_dim + in_proj_dim = (query_head_dim + key_head_dim + pos_head_dim) * num_heads + + # the initial_scale is supposed to take over the "scaling" factor of + # head_dim ** -0.5 that has been used in previous forms of attention, + # dividing it between the query and key. Note: this module is intended + # to be used with the ScaledAdam optimizer; with most other optimizers, + # it would be necessary to apply the scaling factor in the forward function. + self.in_proj = ScaledLinear( + embed_dim, in_proj_dim, bias=True, initial_scale=query_head_dim**-0.25 + ) + + self.whiten_keys = Whiten( + num_groups=num_heads, + whitening_limit=_whitening_schedule(3.0), + prob=(0.025, 0.25), + grad_scale=0.025, + ) + + # add a balancer for the keys that runs with very small probability, and + # tries to enforce that all dimensions have mean around zero. The + # weights produced by this module are invariant to adding a constant to + # the keys, so the derivative of the bias is mathematically zero; but + # due to how Adam/ScaledAdam work, it can learn a fairly large nonzero + # bias because the small numerical roundoff tends to have a non-random + # sign. This module is intended to prevent that. Use a very small + # probability; that should be suffixient to fix the problem. + self.balance_keys = Balancer( + key_head_dim * num_heads, + channel_dim=-1, + min_positive=0.4, + max_positive=0.6, + min_abs=0.0, + max_abs=100.0, + prob=0.025, + ) + + # linear transformation for positional encoding. + self.linear_pos = ScaledLinear( + pos_dim, num_heads * pos_head_dim, bias=False, initial_scale=0.05 + ) + + # the following are for diagnosics only, see --print-diagnostics option + self.copy_pos_query = Identity() + self.copy_query = Identity() + + def forward( + self, + x: Tensor, + pos_emb: Tensor, + key_padding_mask: Optional[Tensor] = None, + attn_mask: Optional[Tensor] = None, + ) -> Tensor: + r""" + Args: + x: input of shape (seq_len, batch_size, embed_dim) + pos_emb: Positional embedding tensor, of shape (1, 2*seq_len - 1, pos_dim) + key_padding_mask: a bool tensor of shape (batch_size, seq_len). Positions that + are True in this mask will be ignored as sources in the attention weighting. + attn_mask: mask of shape (seq_len, seq_len) or (batch_size, seq_len, seq_len), + interpreted as ([batch_size,] tgt_seq_len, src_seq_len) + saying which positions are allowed to attend to which other positions. + Returns: + a tensor of attention weights, of shape (hum_heads, batch_size, seq_len, seq_len) + interpreted as (hum_heads, batch_size, tgt_seq_len, src_seq_len). + """ + x = self.in_proj(x) + query_head_dim = self.query_head_dim + pos_head_dim = self.pos_head_dim + num_heads = self.num_heads + + seq_len, batch_size, _ = x.shape + + query_dim = query_head_dim * num_heads + + # self-attention + q = x[..., 0:query_dim] + k = x[..., query_dim : 2 * query_dim] + # p is the position-encoding query + p = x[..., 2 * query_dim :] + assert p.shape[-1] == num_heads * pos_head_dim + + q = self.copy_query(q) # for diagnostics only, does nothing. + k = self.whiten_keys(self.balance_keys(k)) # does nothing in the forward pass. + p = self.copy_pos_query(p) # for diagnostics only, does nothing. + + q = q.reshape(seq_len, batch_size, num_heads, query_head_dim) + p = p.reshape(seq_len, batch_size, num_heads, pos_head_dim) + k = k.reshape(seq_len, batch_size, num_heads, query_head_dim) + + # time1 refers to target, time2 refers to source. + q = q.permute(2, 1, 0, 3) # (head, batch, time1, query_head_dim) + p = p.permute(2, 1, 0, 3) # (head, batch, time1, pos_head_dim) + k = k.permute(2, 1, 3, 0) # (head, batch, d_k, time2) + + attn_scores = torch.matmul(q, k) + + use_pos_scores = False + if torch.jit.is_scripting() or torch.jit.is_tracing(): + # We can't put random.random() in the same line + use_pos_scores = True + elif not self.training or random.random() >= float(self.pos_emb_skip_rate): + use_pos_scores = True + + if use_pos_scores: + pos_emb = self.linear_pos(pos_emb) + seq_len2 = 2 * seq_len - 1 + pos_emb = pos_emb.reshape(-1, seq_len2, num_heads, pos_head_dim).permute( + 2, 0, 3, 1 + ) + # pos shape now: (head, {1 or batch_size}, pos_dim, seq_len2) + + # (head, batch, time1, pos_dim) x (head, 1, pos_dim, seq_len2) -> (head, batch, time1, seq_len2) + # [where seq_len2 represents relative position.] + pos_scores = torch.matmul(p, pos_emb) + # the following .as_strided() expression converts the last axis of pos_scores from relative + # to absolute position. I don't know whether I might have got the time-offsets backwards or + # not, but let this code define which way round it is supposed to be. + if torch.jit.is_tracing(): + (num_heads, batch_size, time1, n) = pos_scores.shape + rows = torch.arange(start=time1 - 1, end=-1, step=-1) + cols = torch.arange(seq_len) + rows = rows.repeat(batch_size * num_heads).unsqueeze(-1) + indexes = rows + cols + pos_scores = pos_scores.reshape(-1, n) + pos_scores = torch.gather(pos_scores, dim=1, index=indexes) + pos_scores = pos_scores.reshape(num_heads, batch_size, time1, seq_len) + else: + pos_scores = pos_scores.as_strided( + (num_heads, batch_size, seq_len, seq_len), + ( + pos_scores.stride(0), + pos_scores.stride(1), + pos_scores.stride(2) - pos_scores.stride(3), + pos_scores.stride(3), + ), + storage_offset=pos_scores.stride(3) * (seq_len - 1), + ) + + attn_scores = attn_scores + pos_scores + + if torch.jit.is_scripting() or torch.jit.is_tracing(): + pass + elif self.training and random.random() < 0.1: + # This is a harder way of limiting the attention scores to not be + # too large. It incurs a penalty if any of them has an absolute + # value greater than 50.0. this should be outside the normal range + # of the attention scores. We use this mechanism instead of, say, + # something added to the loss function involving the entropy, + # because once the entropy gets very small gradients through the + # softmax can become very small, and we'd get zero derivatives. The + # choices of 1.0e-04 as the scale on the penalty makes this + # mechanism vulnerable to the absolute scale of the loss function, + # but we view this as a failsafe to avoid "implausible" parameter + # values rather than a regularization method that should be active + # under normal circumstances. + attn_scores = penalize_abs_values_gt( + attn_scores, limit=25.0, penalty=1.0e-04, name=self.name + ) + + assert attn_scores.shape == (num_heads, batch_size, seq_len, seq_len) + + if attn_mask is not None: + assert attn_mask.dtype == torch.bool + # use -1000 to avoid nan's where attn_mask and key_padding_mask make + # all scores zero. It's important that this be large enough that exp(-1000) + # is exactly zero, for reasons related to const_attention_rate, it + # compares the final weights with zero. + attn_scores = attn_scores.masked_fill(attn_mask, -1000) + + if key_padding_mask is not None: + assert key_padding_mask.shape == ( + batch_size, + seq_len, + ), key_padding_mask.shape + attn_scores = attn_scores.masked_fill( + key_padding_mask.unsqueeze(1), + -1000, + ) + + # We use our own version of softmax, defined in scaling.py, which should + # save a little of the memory used in backprop by, if we are in + # automatic mixed precision mode (amp / autocast), by only storing the + # half-precision output for backprop purposes. + attn_weights = softmax(attn_scores, dim=-1) + + if torch.jit.is_scripting() or torch.jit.is_tracing(): + pass + elif random.random() < 0.001 and not self.training: + self._print_attn_entropy(attn_weights) + + attn_weights = nn.functional.dropout( + attn_weights, p=self.dropout, training=self.training + ) + + return attn_weights + + def streaming_forward( + self, + x: Tensor, + pos_emb: Tensor, + cached_key: Tensor, + left_context_len: int, + key_padding_mask: Tensor, + ) -> Tuple[Tensor, Tensor]: + r""" + Args: + x: input of shape (seq_len, batch_size, embed_dim) + pos_emb: Positional embedding tensor, of shape (1, left_context_len+2*seq_len-1, pos_dim) + cached_key: cached attention key tensor of left context, + of shape (left_context_len, batch_size, key_dim) + left_context_len: number of left context frames. + key_padding_mask: a bool tensor of shape (batch_size, seq_len). Positions that + are True in this mask will be ignored as sources in the attention weighting. + + Returns: + - attention weights, of shape (hum_heads, batch_size, seq_len, seq_len2), + interpreted as (hum_heads, batch_size, tgt_seq_len, src_seq_len). + - updated cached attention key tensor of left context. + """ + x = self.in_proj(x) + query_head_dim = self.query_head_dim + pos_head_dim = self.pos_head_dim + num_heads = self.num_heads + + seq_len, batch_size, _ = x.shape + + query_dim = query_head_dim * num_heads + + # self-attention + q = x[..., 0:query_dim] + k = x[..., query_dim : 2 * query_dim] + # p is the position-encoding query + p = x[..., 2 * query_dim :] + assert p.shape[-1] == num_heads * pos_head_dim + + # Pad cached left contexts + assert cached_key.shape[0] == left_context_len, ( + cached_key.shape[0], + left_context_len, + ) + k = torch.cat([cached_key, k], dim=0) + # Update cached left contexts + cached_key = k[-left_context_len:, ...] + + # The length of key + k_len = k.shape[0] + + q = q.reshape(seq_len, batch_size, num_heads, query_head_dim) + p = p.reshape(seq_len, batch_size, num_heads, pos_head_dim) + k = k.reshape(k_len, batch_size, num_heads, query_head_dim) + + # time1 refers to target, time2 refers to source. + q = q.permute(2, 1, 0, 3) # (head, batch, time1, query_head_dim) + p = p.permute(2, 1, 0, 3) # (head, batch, time1, pos_head_dim) + k = k.permute(2, 1, 3, 0) # (head, batch, d_k, time2) + + attn_scores = torch.matmul(q, k) + + pos_emb = self.linear_pos(pos_emb) + seq_len2 = 2 * seq_len - 1 + left_context_len + pos_emb = pos_emb.reshape(-1, seq_len2, num_heads, pos_head_dim).permute( + 2, 0, 3, 1 + ) + # pos shape now: (head, {1 or batch_size}, pos_dim, seq_len2) + + # (head, batch, time1, pos_dim) x (head, 1, pos_dim, seq_len2) -> (head, batch, time1, seq_len2) + # [where seq_len2 represents relative position.] + pos_scores = torch.matmul(p, pos_emb) + + if torch.jit.is_tracing(): + (num_heads, batch_size, time1, n) = pos_scores.shape + rows = torch.arange(start=time1 - 1, end=-1, step=-1) + cols = torch.arange(k_len) + rows = rows.repeat(batch_size * num_heads).unsqueeze(-1) + indexes = rows + cols + pos_scores = pos_scores.reshape(-1, n) + pos_scores = torch.gather(pos_scores, dim=1, index=indexes) + pos_scores = pos_scores.reshape(num_heads, batch_size, time1, k_len) + # the following .as_strided() expression converts the last axis of pos_scores from relative + # to absolute position. I don't know whether I might have got the time-offsets backwards or + # not, but let this code define which way round it is supposed to be. + else: + pos_scores = pos_scores.as_strided( + (num_heads, batch_size, seq_len, k_len), + ( + pos_scores.stride(0), + pos_scores.stride(1), + pos_scores.stride(2) - pos_scores.stride(3), + pos_scores.stride(3), + ), + storage_offset=pos_scores.stride(3) * (seq_len - 1), + ) + + attn_scores = attn_scores + pos_scores + + assert attn_scores.shape == ( + num_heads, + batch_size, + seq_len, + k_len, + ), attn_scores.shape + + if key_padding_mask is not None: + assert key_padding_mask.shape == (batch_size, k_len), key_padding_mask.shape + attn_scores = attn_scores.masked_fill( + key_padding_mask.unsqueeze(1), + -1000, + ) + + attn_weights = attn_scores.softmax(dim=-1) + + return attn_weights, cached_key + + def _print_attn_entropy(self, attn_weights: Tensor): + # attn_weights: (num_heads, batch_size, seq_len, seq_len) + (num_heads, batch_size, seq_len, seq_len) = attn_weights.shape + + with torch.no_grad(): + with torch.cuda.amp.autocast(enabled=False): + attn_weights = attn_weights.to(torch.float32) + attn_weights_entropy = ( + -((attn_weights + 1.0e-20).log() * attn_weights) + .sum(dim=-1) + .mean(dim=(1, 2)) + ) + logging.info( + f"name={self.name}, attn_weights_entropy = {attn_weights_entropy}" + ) + + +class SelfAttention(nn.Module): + """ + The simplest possible attention module. This one works with already-computed attention + weights, e.g. as computed by RelPositionMultiheadAttentionWeights. + + Args: + embed_dim: the input and output embedding dimension + num_heads: the number of attention heads + value_head_dim: the value dimension per head + """ + + def __init__( + self, + embed_dim: int, + num_heads: int, + value_head_dim: int, + ) -> None: + super().__init__() + self.in_proj = nn.Linear(embed_dim, num_heads * value_head_dim, bias=True) + + self.out_proj = ScaledLinear( + num_heads * value_head_dim, embed_dim, bias=True, initial_scale=0.05 + ) + + self.whiten = Whiten( + num_groups=1, + whitening_limit=_whitening_schedule(7.5, ratio=3.0), + prob=(0.025, 0.25), + grad_scale=0.01, + ) + + def forward( + self, + x: Tensor, + attn_weights: Tensor, + ) -> Tensor: + """ + Args: + x: input tensor, of shape (seq_len, batch_size, embed_dim) + attn_weights: a tensor of shape (num_heads, batch_size, seq_len, seq_len), + with seq_len being interpreted as (tgt_seq_len, src_seq_len). Expect + attn_weights.sum(dim=-1) == 1. + Returns: + a tensor with the same shape as x. + """ + (seq_len, batch_size, embed_dim) = x.shape + num_heads = attn_weights.shape[0] + assert attn_weights.shape == (num_heads, batch_size, seq_len, seq_len) + + x = self.in_proj(x) # (seq_len, batch_size, num_heads * value_head_dim) + x = x.reshape(seq_len, batch_size, num_heads, -1).permute(2, 1, 0, 3) + # now x: (num_heads, batch_size, seq_len, value_head_dim) + value_head_dim = x.shape[-1] + + # todo: see whether there is benefit in overriding matmul + x = torch.matmul(attn_weights, x) + # v: (num_heads, batch_size, seq_len, value_head_dim) + + x = ( + x.permute(2, 1, 0, 3) + .contiguous() + .view(seq_len, batch_size, num_heads * value_head_dim) + ) + + # returned value is of shape (seq_len, batch_size, embed_dim), like the input. + x = self.out_proj(x) + x = self.whiten(x) + + return x + + def streaming_forward( + self, + x: Tensor, + attn_weights: Tensor, + cached_val: Tensor, + left_context_len: int, + ) -> Tuple[Tensor, Tensor]: + """ + Args: + x: input tensor, of shape (seq_len, batch_size, embed_dim) + attn_weights: a tensor of shape (num_heads, batch_size, seq_len, seq_len), + with seq_len being interpreted as (tgt_seq_len, src_seq_len). Expect + attn_weights.sum(dim=-1) == 1. + cached_val: cached attention value tensor of left context, + of shape (left_context_len, batch_size, value_dim) + left_context_len: number of left context frames. + + Returns: + - attention weighted output, a tensor with the same shape as x. + - updated cached attention value tensor of left context. + """ + (seq_len, batch_size, embed_dim) = x.shape + num_heads = attn_weights.shape[0] + seq_len2 = seq_len + left_context_len + assert attn_weights.shape == (num_heads, batch_size, seq_len, seq_len2) + + x = self.in_proj(x) # (seq_len, batch_size, num_heads * value_head_dim) + + # Pad cached left contexts + assert cached_val.shape[0] == left_context_len, ( + cached_val.shape[0], + left_context_len, + ) + x = torch.cat([cached_val, x], dim=0) + # Update cached left contexts + cached_val = x[-left_context_len:, ...] + + x = x.reshape(seq_len2, batch_size, num_heads, -1).permute(2, 1, 0, 3) + # now x: (num_heads, batch_size, seq_len, value_head_dim) + value_head_dim = x.shape[-1] + + # todo: see whether there is benefit in overriding matmul + x = torch.matmul(attn_weights, x) + # v: (num_heads, batch_size, seq_len, value_head_dim) + + x = ( + x.permute(2, 1, 0, 3) + .contiguous() + .view(seq_len, batch_size, num_heads * value_head_dim) + ) + + # returned value is of shape (seq_len, batch_size, embed_dim), like the input. + x = self.out_proj(x) + + return x, cached_val + + +class FeedforwardModule(nn.Module): + """Feedforward module in Zipformer2 model.""" + + def __init__(self, embed_dim: int, feedforward_dim: int, dropout: FloatLike): + super(FeedforwardModule, self).__init__() + self.in_proj = nn.Linear(embed_dim, feedforward_dim) + + self.hidden_balancer = Balancer( + feedforward_dim, + channel_dim=-1, + min_positive=0.3, + max_positive=1.0, + min_abs=0.75, + max_abs=5.0, + ) + + # shared_dim=0 means we share the dropout mask along the time axis + self.out_proj = ActivationDropoutAndLinear( + feedforward_dim, + embed_dim, + activation="SwooshL", + dropout_p=dropout, + dropout_shared_dim=0, + bias=True, + initial_scale=0.1, + ) + + self.out_whiten = Whiten( + num_groups=1, + whitening_limit=_whitening_schedule(7.5), + prob=(0.025, 0.25), + grad_scale=0.01, + ) + + def forward(self, x: Tensor): + x = self.in_proj(x) + x = self.hidden_balancer(x) + # out_proj contains SwooshL activation, then dropout, then linear. + x = self.out_proj(x) + x = self.out_whiten(x) + return x + + +class NonlinAttention(nn.Module): + """This is like the ConvolutionModule, but refactored so that we use multiplication by attention weights (borrowed + from the attention module) in place of actual convolution. We also took out the second nonlinearity, the + one after the attention mechanism. + + Args: + channels (int): The number of channels of conv layers. + """ + + def __init__( + self, + channels: int, + hidden_channels: int, + ) -> None: + super().__init__() + + self.hidden_channels = hidden_channels + + self.in_proj = nn.Linear(channels, hidden_channels * 3, bias=True) + + # balancer that goes before the sigmoid. Have quite a large min_abs value, at 2.0, + # because we noticed that well-trained instances of this module have abs-value before the sigmoid + # starting from about 3, and poorly-trained instances of the module have smaller abs values + # before the sigmoid. + self.balancer = Balancer( + hidden_channels, + channel_dim=-1, + min_positive=ScheduledFloat((0.0, 0.25), (20000.0, 0.05)), + max_positive=ScheduledFloat((0.0, 0.75), (20000.0, 0.95)), + min_abs=0.5, + max_abs=5.0, + ) + self.tanh = nn.Tanh() + + self.identity1 = Identity() # for diagnostics. + self.identity2 = Identity() # for diagnostics. + self.identity3 = Identity() # for diagnostics. + + self.out_proj = ScaledLinear( + hidden_channels, channels, bias=True, initial_scale=0.05 + ) + + self.whiten1 = Whiten( + num_groups=1, + whitening_limit=_whitening_schedule(5.0), + prob=(0.025, 0.25), + grad_scale=0.01, + ) + + self.whiten2 = Whiten( + num_groups=1, + whitening_limit=_whitening_schedule(5.0, ratio=3.0), + prob=(0.025, 0.25), + grad_scale=0.01, + ) + + def forward( + self, + x: Tensor, + attn_weights: Tensor, + ) -> Tensor: + """. + Args: + x: a Tensor of shape (seq_len, batch_size, num_channels) + attn_weights: a Tensor of shape (num_heads, batch_size, seq_len, seq_len) + Returns: + a Tensor with the same shape as x + """ + x = self.in_proj(x) + + (seq_len, batch_size, _) = x.shape + hidden_channels = self.hidden_channels + + s, x, y = x.chunk(3, dim=2) + + # s will go through tanh. + + s = self.balancer(s) + s = self.tanh(s) + + s = s.unsqueeze(-1).reshape(seq_len, batch_size, hidden_channels) + x = self.whiten1(x) + x = x * s + x = self.identity1(x) # diagnostics only, it's the identity. + + (seq_len, batch_size, embed_dim) = x.shape + num_heads = attn_weights.shape[0] + assert attn_weights.shape == (num_heads, batch_size, seq_len, seq_len) + + x = x.reshape(seq_len, batch_size, num_heads, -1).permute(2, 1, 0, 3) + # now x: (num_heads, batch_size, seq_len, head_dim) + x = torch.matmul(attn_weights, x) + # now x: (num_heads, batch_size, seq_len, head_dim) + x = x.permute(2, 1, 0, 3).reshape(seq_len, batch_size, -1) + + y = self.identity2(y) + x = x * y + x = self.identity3(x) + + x = self.out_proj(x) + x = self.whiten2(x) + return x + + def streaming_forward( + self, + x: Tensor, + attn_weights: Tensor, + cached_x: Tensor, + left_context_len: int, + ) -> Tuple[Tensor, Tensor]: + """. + Args: + x: a Tensor of shape (seq_len, batch_size, num_channels) + attn_weights: a Tensor of shape (num_heads, batch_size, seq_len, seq_len) + cached_x: left context, a Tensor of shape + (num_heads, batch_size, left_context_len, head_dim) + left_context_len: number of left context frames. + Returns: + - a Tensor with the same shape as x + - updated left context with same shape as cached_x + """ + x = self.in_proj(x) + + (seq_len, batch_size, _) = x.shape + hidden_channels = self.hidden_channels + + s, x, y = x.chunk(3, dim=2) + + # s will go through tanh. + s = self.tanh(s) + + s = s.unsqueeze(-1).reshape(seq_len, batch_size, hidden_channels) + x = x * s + + (seq_len, batch_size, embed_dim) = x.shape + num_heads = attn_weights.shape[0] + assert attn_weights.shape == ( + num_heads, + batch_size, + seq_len, + left_context_len + seq_len, + ) + + x = x.reshape(seq_len, batch_size, num_heads, -1).permute(2, 1, 0, 3) + # now x: (num_heads, batch_size, seq_len, head_dim) + + # Pad cached tensor + assert cached_x.shape[2] == left_context_len, ( + cached_x.shape[2], + left_context_len, + ) + x_pad = torch.cat([cached_x, x], dim=2) + # Update cached tensor + cached_x = x_pad[:, :, -left_context_len:, :] + + x = torch.matmul(attn_weights, x_pad) + # now x: (num_heads, batch_size, seq_len, head_dim) + x = x.permute(2, 1, 0, 3).reshape(seq_len, batch_size, -1) + + x = x * y + + x = self.out_proj(x) + return x, cached_x + + +class ConvolutionModule(nn.Module): + """ConvolutionModule in Zipformer2 model. + Modified from https://github.com/espnet/espnet/blob/master/espnet/nets/pytorch_backend/zipformer/convolution.py + + Args: + channels (int): The number of channels of conv layers. + kernel_size (int): Kernerl size of conv layers. + bias (bool): Whether to use bias in conv layers (default=True). + + """ + + def __init__( + self, + channels: int, + kernel_size: int, + causal: bool, + ) -> None: + """Construct a ConvolutionModule object.""" + super(ConvolutionModule, self).__init__() + # kernerl_size should be a odd number for 'SAME' padding + assert (kernel_size - 1) % 2 == 0 + + bottleneck_dim = channels + self.causal = causal + + self.in_proj = nn.Linear( + channels, + 2 * bottleneck_dim, + ) + # the gradients on in_proj are a little noisy, likely to do with the + # sigmoid in glu. + + # after in_proj we put x through a gated linear unit (nn.functional.glu). + # For most layers the normal rms value of channels of x seems to be in the range 1 to 4, + # but sometimes, for some reason, for layer 0 the rms ends up being very large, + # between 50 and 100 for different channels. This will cause very peaky and + # sparse derivatives for the sigmoid gating function, which will tend to make + # the loss function not learn effectively. (for most layers the average absolute values + # are in the range 0.5..9.0, and the average p(x>0), i.e. positive proportion, + # at the output of pointwise_conv1.output is around 0.35 to 0.45 for different + # layers, which likely breaks down as 0.5 for the "linear" half and + # 0.2 to 0.3 for the part that goes into the sigmoid. The idea is that if we + # constrain the rms values to a reasonable range via a constraint of max_abs=10.0, + # it will be in a better position to start learning something, i.e. to latch onto + # the correct range. + self.balancer1 = Balancer( + bottleneck_dim, + channel_dim=-1, + min_positive=ScheduledFloat((0.0, 0.05), (8000.0, 0.025)), + max_positive=1.0, + min_abs=1.5, + max_abs=ScheduledFloat((0.0, 5.0), (8000.0, 10.0), default=1.0), + ) + + self.activation1 = Identity() # for diagnostics + + self.sigmoid = nn.Sigmoid() + + self.activation2 = Identity() # for diagnostics + + assert kernel_size % 2 == 1 + + self.depthwise_conv = ( + ChunkCausalDepthwiseConv1d(channels=bottleneck_dim, kernel_size=kernel_size) + if causal + else nn.Conv1d( + in_channels=bottleneck_dim, + out_channels=bottleneck_dim, + groups=bottleneck_dim, + kernel_size=kernel_size, + padding=kernel_size // 2, + ) + ) + + self.balancer2 = Balancer( + bottleneck_dim, + channel_dim=1, + min_positive=ScheduledFloat((0.0, 0.1), (8000.0, 0.05)), + max_positive=1.0, + min_abs=ScheduledFloat((0.0, 0.2), (20000.0, 0.5)), + max_abs=10.0, + ) + + self.whiten = Whiten( + num_groups=1, + whitening_limit=_whitening_schedule(7.5), + prob=(0.025, 0.25), + grad_scale=0.01, + ) + + self.out_proj = ActivationDropoutAndLinear( + bottleneck_dim, + channels, + activation="SwooshR", + dropout_p=0.0, + initial_scale=0.05, + ) + + def forward( + self, + x: Tensor, + src_key_padding_mask: Optional[Tensor] = None, + chunk_size: int = -1, + ) -> Tensor: + """Compute convolution module. + + Args: + x: Input tensor (#time, batch, channels). + src_key_padding_mask: the mask for the src keys per batch (optional): + (batch, #time), contains True in masked positions. + + Returns: + Tensor: Output tensor (#time, batch, channels). + + """ + + x = self.in_proj(x) # (time, batch, 2*channels) + + x, s = x.chunk(2, dim=2) + s = self.balancer1(s) + s = self.sigmoid(s) + x = self.activation1(x) # identity. + x = x * s + x = self.activation2(x) # identity + + # (time, batch, channels) + + # exchange the temporal dimension and the feature dimension + x = x.permute(1, 2, 0) # (#batch, channels, time). + + if src_key_padding_mask is not None: + x = x.masked_fill(src_key_padding_mask.unsqueeze(1).expand_as(x), 0.0) + + if ( + not torch.jit.is_scripting() + and not torch.jit.is_tracing() + and chunk_size >= 0 + ): + # Not support exporting a model for simulated streaming decoding + assert ( + self.causal + ), "Must initialize model with causal=True if you use chunk_size" + x = self.depthwise_conv(x, chunk_size=chunk_size) + else: + x = self.depthwise_conv(x) + + x = self.balancer2(x) + x = x.permute(2, 0, 1) # (time, batch, channels) + + x = self.whiten(x) # (time, batch, channels) + x = self.out_proj(x) # (time, batch, channels) + + return x + + def streaming_forward( + self, + x: Tensor, + cache: Tensor, + src_key_padding_mask: Tensor, + ) -> Tuple[Tensor, Tensor]: + """Compute convolution module in streaming forward mode. + + Args: + x: Input tensor (#time, batch, channels). + cache: cached left context for depthwise_conv of shape + (#batch, channels, left_pad) + src_key_padding_mask: the mask for the src keys per batch (optional): + (batch, #time), contains True in masked positions. + + Returns: + - Output tensor (#time, batch, channels). + - Updated cache (#batch, channels, left_pad) + """ + + x = self.in_proj(x) # (time, batch, 2*channels) + + x, s = x.chunk(2, dim=2) + s = self.sigmoid(s) + x = x * s + # (time, batch, channels) + + # exchange the temporal dimension and the feature dimension + x = x.permute(1, 2, 0) # (#batch, channels, time). + + if src_key_padding_mask is not None: + x = x.masked_fill(src_key_padding_mask.unsqueeze(1).expand_as(x), 0.0) + + x, cache = self.depthwise_conv.streaming_forward(x, cache=cache) + + x = x.permute(2, 0, 1) # (time, batch, channels) + + x = self.out_proj(x) # (time, batch, channels) + + return x, cache + + +class ScalarMultiply(nn.Module): + def __init__(self, scale: float): + super().__init__() + self.scale = scale + + def forward(self, x): + return x * self.scale + + +def _test_zipformer_main(causal: bool = False): + batch_size = 5 + seq_len = 20 + # Just make sure the forward pass runs. + + c = Zipformer2( + encoder_dim=(64, 96), + encoder_unmasked_dim=(48, 64), + num_heads=(4, 4), + causal=causal, + chunk_size=(4,) if causal else (-1,), + left_context_frames=(64,), + ) + batch_size = 5 + seq_len = 20 + # Just make sure the forward pass runs. + f = c( + torch.randn(seq_len, batch_size, 64), + torch.full((batch_size,), seq_len, dtype=torch.int64), + ) + f[0].sum().backward() + c.eval() + f = c( + torch.randn(seq_len, batch_size, 64), + torch.full((batch_size,), seq_len, dtype=torch.int64), + ) + f # to remove flake8 warnings + + +class AdapterModule(nn.Module): + def __init__( + self, + embed_dim: int = 384, + bottleneck_dim: int = 16, + ): + # The simplest adapter + super(AdapterModule, self).__init__() + self.embed_dim = embed_dim + self.bottleneck_dim = bottleneck_dim + self.activation = SwooshL() + + self.in_proj = nn.Linear(embed_dim, bottleneck_dim) + self.out_proj = nn.Linear(bottleneck_dim, embed_dim) + + def forward(self, x): + x_orig = x + x = self.activation(self.in_proj(x)) + x = self.out_proj(x) + return x_orig + x + + +if __name__ == "__main__": + logging.getLogger().setLevel(logging.INFO) + torch.set_num_threads(1) + torch.set_num_interop_threads(1) + _test_zipformer_main(False) + _test_zipformer_main(True) diff --git a/egs/librispeech/ASR/zipformer_ctc/train.py b/egs/librispeech/ASR/zipformer_ctc/train.py index 60990456d..60112a84e 100755 --- a/egs/librispeech/ASR/zipformer_ctc/train.py +++ b/egs/librispeech/ASR/zipformer_ctc/train.py @@ -62,6 +62,7 @@ from icefall.checkpoint import ( ) from icefall.dist import cleanup_dist, setup_dist from icefall.env import get_env_info +from icefall.err import raise_grad_scale_is_too_small_error from icefall.hooks import register_inf_check_hooks from icefall.lexicon import Lexicon from icefall.utils import AttributeDict, MetricsTracker, setup_logger, str2bool @@ -797,9 +798,7 @@ def train_one_epoch( if cur_grad_scale < 0.01: logging.warning(f"Grad scale is small: {cur_grad_scale}") if cur_grad_scale < 1.0e-05: - raise RuntimeError( - f"grad_scale is too small, exiting: {cur_grad_scale}" - ) + raise_grad_scale_is_too_small_error(cur_grad_scale) if batch_idx % params.log_interval == 0: cur_lr = scheduler.get_last_lr()[0] diff --git a/egs/librispeech/ASR/zipformer_mmi/train.py b/egs/librispeech/ASR/zipformer_mmi/train.py index dd8949523..c1785a328 100755 --- a/egs/librispeech/ASR/zipformer_mmi/train.py +++ b/egs/librispeech/ASR/zipformer_mmi/train.py @@ -79,6 +79,7 @@ from icefall.checkpoint import ( ) from icefall.dist import cleanup_dist, setup_dist from icefall.env import get_env_info +from icefall.err import raise_grad_scale_is_too_small_error from icefall.hooks import register_inf_check_hooks from icefall.lexicon import Lexicon, UniqLexicon from icefall.mmi import LFMMILoss @@ -816,9 +817,7 @@ def train_one_epoch( if cur_grad_scale < 0.01: logging.warning(f"Grad scale is small: {cur_grad_scale}") if cur_grad_scale < 1.0e-05: - raise RuntimeError( - f"grad_scale is too small, exiting: {cur_grad_scale}" - ) + raise_grad_scale_is_too_small_error(cur_grad_scale) if batch_idx % params.log_interval == 0: cur_lr = scheduler.get_last_lr()[0] diff --git a/egs/librispeech/WSASR/conformer_ctc2/train.py b/egs/librispeech/WSASR/conformer_ctc2/train.py index fe6c5af91..82c68803f 100755 --- a/egs/librispeech/WSASR/conformer_ctc2/train.py +++ b/egs/librispeech/WSASR/conformer_ctc2/train.py @@ -31,6 +31,7 @@ export CUDA_VISIBLE_DEVICES="0,1,2,3" --exp-dir conformer_ctc2/exp \ --lang-dir data/lang_bpe_200 \ --otc-token "" \ + --feature-dim 768 \ --allow-bypass-arc true \ --allow-self-loop-arc true \ --initial-bypass-weight -19 \ @@ -160,6 +161,14 @@ def get_parser(): """, ) + parser.add_argument( + "--feature-dim", + type=int, + default=768, + help="""Number of features extracted in feature extraction stage.last dimension of feature vector. + 80 when using fbank features and 768 or 1024 whn using wave2vec""", + ) + parser.add_argument( "--initial-lr", type=float, @@ -385,7 +394,6 @@ def get_params() -> AttributeDict: "valid_interval": 800, # For the 100h subset, use 800 "alignment_interval": 25, # parameters for conformer - "feature_dim": 768, "subsampling_factor": 2, "encoder_dim": 512, "nhead": 8, diff --git a/egs/ljspeech/TTS/README.md b/egs/ljspeech/TTS/README.md new file mode 100644 index 000000000..7b112c12c --- /dev/null +++ b/egs/ljspeech/TTS/README.md @@ -0,0 +1,103 @@ +# Introduction + +This is a public domain speech dataset consisting of 13,100 short audio clips of a single speaker reading passages from 7 non-fiction books. +A transcription is provided for each clip. +Clips vary in length from 1 to 10 seconds and have a total length of approximately 24 hours. + +The texts were published between 1884 and 1964, and are in the public domain. +The audio was recorded in 2016-17 by the [LibriVox](https://librivox.org/) project and is also in the public domain. + +The above information is from the [LJSpeech website](https://keithito.com/LJ-Speech-Dataset/). + +# VITS + +This recipe provides a VITS model trained on the LJSpeech dataset. + +Pretrained model can be found [here](https://huggingface.co/Zengwei/icefall-tts-ljspeech-vits-2024-02-28). + +For tutorial and more details, please refer to the [VITS documentation](https://k2-fsa.github.io/icefall/recipes/TTS/ljspeech/vits.html). + +The training command is given below: +``` +export CUDA_VISIBLE_DEVICES=0,1,2,3 +./vits/train.py \ + --world-size 4 \ + --num-epochs 1000 \ + --start-epoch 1 \ + --use-fp16 1 \ + --exp-dir vits/exp \ + --max-duration 500 +``` + +To inference, use: +``` +./vits/infer.py \ + --exp-dir vits/exp \ + --epoch 1000 \ + --tokens data/tokens.txt +``` + +## Quality vs speed + +If you feel that the trained model is slow at runtime, you can specify the +argument `--model-type` during training. Possible values are: + + - `low`, means **low** quality. The resulting model is very small in file size + and runs very fast. The following is a wave file generatd by a `low` quality model + + https://github.com/k2-fsa/icefall/assets/5284924/d5758c24-470d-40ee-b089-e57fcba81633 + + The text is `Ask not what your country can do for you; ask what you can do for your country.` + + The exported onnx model has a file size of ``26.8 MB`` (float32). + + - `medium`, means **medium** quality. + The following is a wave file generatd by a `medium` quality model + + https://github.com/k2-fsa/icefall/assets/5284924/b199d960-3665-4d0d-9ae9-a1bb69cbc8ac + + The text is `Ask not what your country can do for you; ask what you can do for your country.` + + The exported onnx model has a file size of ``70.9 MB`` (float32). + + - `high`, means **high** quality. This is the default value. + + The following is a wave file generatd by a `high` quality model + + https://github.com/k2-fsa/icefall/assets/5284924/b39f3048-73a6-4267-bf95-df5abfdb28fc + + The text is `Ask not what your country can do for you; ask what you can do for your country.` + + The exported onnx model has a file size of ``113 MB`` (float32). + + +A pre-trained `low` model trained using 4xV100 32GB GPU with the following command can be found at + + +```bash +export CUDA_VISIBLE_DEVICES=0,1,2,3 +./vits/train.py \ + --world-size 4 \ + --num-epochs 1601 \ + --start-epoch 1 \ + --use-fp16 1 \ + --exp-dir vits/exp \ + --model-type low \ + --max-duration 800 +``` + +A pre-trained `medium` model trained using 4xV100 32GB GPU with the following command can be found at + +```bash +export CUDA_VISIBLE_DEVICES=4,5,6,7 +./vits/train.py \ + --world-size 4 \ + --num-epochs 1000 \ + --start-epoch 1 \ + --use-fp16 1 \ + --exp-dir vits/exp-medium \ + --model-type medium \ + --max-duration 500 + +# (Note it is killed after `epoch-820.pt`) +``` diff --git a/egs/ljspeech/TTS/local/prepare_token_file.py b/egs/ljspeech/TTS/local/prepare_token_file.py index df976804a..5b048b600 100755 --- a/egs/ljspeech/TTS/local/prepare_token_file.py +++ b/egs/ljspeech/TTS/local/prepare_token_file.py @@ -17,7 +17,7 @@ """ -This file reads the texts in given manifest and generates the file that maps tokens to IDs. +This file generates the file that maps tokens to IDs. """ import argparse @@ -25,80 +25,38 @@ import logging from pathlib import Path from typing import Dict -from lhotse import load_manifest +from piper_phonemize import get_espeak_map def get_args(): parser = argparse.ArgumentParser() - parser.add_argument( - "--manifest-file", - type=Path, - default=Path("data/spectrogram/ljspeech_cuts_train.jsonl.gz"), - help="Path to the manifest file", - ) - parser.add_argument( "--tokens", type=Path, default=Path("data/tokens.txt"), - help="Path to the tokens", + help="Path to the dict that maps the text tokens to IDs", ) return parser.parse_args() -def write_mapping(filename: str, sym2id: Dict[str, int]) -> None: - """Write a symbol to ID mapping to a file. +def get_token2id(filename: Path) -> Dict[str, int]: + """Get a dict that maps token to IDs, and save it to the given filename.""" + all_tokens = get_espeak_map() # token: [token_id] + all_tokens = {token: token_id[0] for token, token_id in all_tokens.items()} + # sort by token_id + all_tokens = sorted(all_tokens.items(), key=lambda x: x[1]) - Note: - No need to implement `read_mapping` as it can be done - through :func:`k2.SymbolTable.from_file`. - - Args: - filename: - Filename to save the mapping. - sym2id: - A dict mapping symbols to IDs. - Returns: - Return None. - """ with open(filename, "w", encoding="utf-8") as f: - for sym, i in sym2id.items(): - f.write(f"{sym} {i}\n") - - -def get_token2id(manifest_file: Path) -> Dict[str, int]: - """Return a dict that maps token to IDs.""" - extra_tokens = [ - "", # 0 for blank - "", # 1 for sos and eos symbols. - "", # 2 for OOV - ] - all_tokens = set() - - cut_set = load_manifest(manifest_file) - - for cut in cut_set: - # Each cut only contain one supervision - assert len(cut.supervisions) == 1, len(cut.supervisions) - for t in cut.tokens: - all_tokens.add(t) - - all_tokens = extra_tokens + list(all_tokens) - - token2id: Dict[str, int] = {token: i for i, token in enumerate(all_tokens)} - return token2id + for token, token_id in all_tokens: + f.write(f"{token} {token_id}\n") if __name__ == "__main__": formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s" - logging.basicConfig(format=formatter, level=logging.INFO) args = get_args() - manifest_file = Path(args.manifest_file) out_file = Path(args.tokens) - - token2id = get_token2id(manifest_file) - write_mapping(out_file, token2id) + get_token2id(out_file) diff --git a/egs/ljspeech/TTS/local/prepare_tokens_ljspeech.py b/egs/ljspeech/TTS/local/prepare_tokens_ljspeech.py index fcd0137a0..4ba88604c 100755 --- a/egs/ljspeech/TTS/local/prepare_tokens_ljspeech.py +++ b/egs/ljspeech/TTS/local/prepare_tokens_ljspeech.py @@ -23,9 +23,13 @@ This file reads the texts in given manifest and save the new cuts with phoneme t import logging from pathlib import Path -import g2p_en -import tacotron_cleaner.cleaners +try: + import tacotron_cleaner.cleaners +except ModuleNotFoundError as ex: + raise RuntimeError(f"{ex}\nPlease run\n pip install espnet_tts_frontend\n") + from lhotse import CutSet, load_manifest +from piper_phonemize import phonemize_espeak def prepare_tokens_ljspeech(): @@ -35,17 +39,20 @@ def prepare_tokens_ljspeech(): partition = "all" cut_set = load_manifest(output_dir / f"{prefix}_cuts_{partition}.{suffix}") - g2p = g2p_en.G2p() new_cuts = [] for cut in cut_set: # Each cut only contains one supervision - assert len(cut.supervisions) == 1, len(cut.supervisions) + assert len(cut.supervisions) == 1, (len(cut.supervisions), cut) text = cut.supervisions[0].normalized_text # Text normalization text = tacotron_cleaner.cleaners.custom_english_cleaners(text) # Convert to phonemes - cut.tokens = g2p(text) + tokens_list = phonemize_espeak(text, "en-us") + tokens = [] + for t in tokens_list: + tokens.extend(t) + cut.tokens = tokens new_cuts.append(cut) new_cut_set = CutSet.from_cuts(new_cuts) diff --git a/egs/ljspeech/TTS/prepare.sh b/egs/ljspeech/TTS/prepare.sh index ed0a07f5e..9ed0f93fd 100755 --- a/egs/ljspeech/TTS/prepare.sh +++ b/egs/ljspeech/TTS/prepare.sh @@ -28,9 +28,9 @@ if [ $stage -le -1 ] && [ $stop_stage -ge -1 ]; then log "Stage -1: build monotonic_align lib" if [ ! -d vits/monotonic_align/build ]; then cd vits/monotonic_align - python setup.py build_ext --inplace + python3 setup.py build_ext --inplace cd ../../ - else + else log "monotonic_align lib already built" fi fi @@ -54,7 +54,7 @@ fi if [ $stage -le 1 ] && [ $stop_stage -ge 1 ]; then log "Stage 1: Prepare LJSpeech manifest" # We assume that you have downloaded the LJSpeech corpus - # to $dl_dir/LJSpeech + # to $dl_dir/LJSpeech-1.1 mkdir -p data/manifests if [ ! -e data/manifests/.ljspeech.done ]; then lhotse prepare ljspeech $dl_dir/LJSpeech-1.1 data/manifests @@ -80,6 +80,10 @@ fi if [ $stage -le 3 ] && [ $stop_stage -ge 3 ]; then log "Stage 3: Prepare phoneme tokens for LJSpeech" + # We assume you have installed piper_phonemize and espnet_tts_frontend. + # If not, please install them with: + # - piper_phonemize: pip install piper_phonemize -f https://k2-fsa.github.io/icefall/piper_phonemize.html, + # - espnet_tts_frontend, `pip install espnet_tts_frontend`, refer to https://github.com/espnet/espnet_tts_frontend/ if [ ! -e data/spectrogram/.ljspeech_with_token.done ]; then ./local/prepare_tokens_ljspeech.py mv data/spectrogram/ljspeech_cuts_with_tokens_all.jsonl.gz \ @@ -113,13 +117,12 @@ fi if [ $stage -le 5 ] && [ $stop_stage -ge 5 ]; then log "Stage 5: Generate token file" - # We assume you have installed g2p_en and espnet_tts_frontend. + # We assume you have installed piper_phonemize and espnet_tts_frontend. # If not, please install them with: - # - g2p_en: `pip install g2p_en`, refer to https://github.com/Kyubyong/g2p + # - piper_phonemize: refer to https://github.com/rhasspy/piper-phonemize, + # could install the pre-built wheels from https://github.com/csukuangfj/piper-phonemize/releases/tag/2023.12.5 # - espnet_tts_frontend, `pip install espnet_tts_frontend`, refer to https://github.com/espnet/espnet_tts_frontend/ if [ ! -e data/tokens.txt ]; then - ./local/prepare_token_file.py \ - --manifest-file data/spectrogram/ljspeech_cuts_train.jsonl.gz \ - --tokens data/tokens.txt + ./local/prepare_token_file.py --tokens data/tokens.txt fi fi diff --git a/egs/ljspeech/TTS/vits/README.md b/egs/ljspeech/TTS/vits/README.md index 1141326b9..f2deed588 100644 --- a/egs/ljspeech/TTS/vits/README.md +++ b/egs/ljspeech/TTS/vits/README.md @@ -1,3 +1,4 @@ See https://k2-fsa.github.io/icefall/recipes/TTS/ljspeech/vits.html for detailed tutorials. -Training logs, Tensorboard logs, and checkpoints are uploaded to https://huggingface.co/Zengwei/icefall-tts-ljspeech-vits-2023-11-29. +Training logs, Tensorboard logs, and checkpoints are uploaded to +https://huggingface.co/Zengwei/icefall-tts-ljspeech-vits-2024-02-28 diff --git a/egs/ljspeech/TTS/vits/export-onnx.py b/egs/ljspeech/TTS/vits/export-onnx.py index f82f9dbe9..0740757c0 100755 --- a/egs/ljspeech/TTS/vits/export-onnx.py +++ b/egs/ljspeech/TTS/vits/export-onnx.py @@ -25,9 +25,8 @@ Export the model to ONNX: --exp-dir vits/exp \ --tokens data/tokens.txt -It will generate two files inside vits/exp: +It will generate one file inside vits/exp: - vits-epoch-1000.onnx - - vits-epoch-1000.int8.onnx (quantizated model) See ./test_onnx.py for how to use the exported ONNX models. """ @@ -40,7 +39,6 @@ from typing import Dict, Tuple import onnx import torch import torch.nn as nn -from onnxruntime.quantization import QuantType, quantize_dynamic from tokenizer import Tokenizer from train import get_model, get_params @@ -75,6 +73,16 @@ def get_parser(): help="""Path to vocabulary.""", ) + parser.add_argument( + "--model-type", + type=str, + default="high", + choices=["low", "medium", "high"], + help="""If not empty, valid values are: low, medium, high. + It controls the model size. low -> runs faster. + """, + ) + return parser @@ -91,7 +99,7 @@ def add_meta_data(filename: str, meta_data: Dict[str, str]): for key, value in meta_data.items(): meta = model.metadata_props.add() meta.key = key - meta.value = value + meta.value = str(value) onnx.save(model, filename) @@ -136,7 +144,7 @@ class OnnxModel(nn.Module): Return a tuple containing: - audio, generated wavform tensor, (B, T_wav) """ - audio, _, _ = self.model.inference( + audio, _, _ = self.model.generator.inference( text=tokens, text_lengths=tokens_lens, noise_scale=noise_scale, @@ -198,11 +206,21 @@ def export_model_onnx( }, ) + if model.model.spks is None: + num_speakers = 1 + else: + num_speakers = model.model.spks + meta_data = { - "model_type": "VITS", + "model_type": "vits", "version": "1", "model_author": "k2-fsa", - "comment": "VITS generator", + "comment": "icefall", # must be icefall for models from icefall + "language": "English", + "voice": "en-us", # Choose your language appropriately + "has_espeak": 1, + "n_speakers": num_speakers, + "sample_rate": model.model.sampling_rate, # Must match the real sample rate } logging.info(f"meta_data: {meta_data}") @@ -218,8 +236,7 @@ def main(): params.update(vars(args)) tokenizer = Tokenizer(params.tokens) - params.blank_id = tokenizer.blank_id - params.oov_id = tokenizer.oov_id + params.blank_id = tokenizer.pad_id params.vocab_size = tokenizer.vocab_size logging.info(params) @@ -229,14 +246,13 @@ def main(): load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model) - model = model.generator model.to("cpu") model.eval() model = OnnxModel(model=model) num_param = sum([p.numel() for p in model.parameters()]) - logging.info(f"generator parameters: {num_param}") + logging.info(f"generator parameters: {num_param}, or {num_param/1000/1000} M") suffix = f"epoch-{params.epoch}" @@ -252,20 +268,149 @@ def main(): ) logging.info(f"Exported generator to {model_filename}") - # Generate int8 quantization models - # See https://onnxruntime.ai/docs/performance/model-optimizations/quantization.html#data-type-selection - - logging.info("Generate int8 quantization models") - - model_filename_int8 = params.exp_dir / f"vits-{suffix}.int8.onnx" - quantize_dynamic( - model_input=model_filename, - model_output=model_filename_int8, - weight_type=QuantType.QUInt8, - ) - if __name__ == "__main__": formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s" logging.basicConfig(format=formatter, level=logging.INFO) main() + +""" +Supported languages. + +LJSpeech is using "en-us" from the second column. + +Pty Language Age/Gender VoiceName File Other Languages + 5 af --/M Afrikaans gmw/af + 5 am --/M Amharic sem/am + 5 an --/M Aragonese roa/an + 5 ar --/M Arabic sem/ar + 5 as --/M Assamese inc/as + 5 az --/M Azerbaijani trk/az + 5 ba --/M Bashkir trk/ba + 5 be --/M Belarusian zle/be + 5 bg --/M Bulgarian zls/bg + 5 bn --/M Bengali inc/bn + 5 bpy --/M Bishnupriya_Manipuri inc/bpy + 5 bs --/M Bosnian zls/bs + 5 ca --/M Catalan roa/ca + 5 chr-US-Qaaa-x-west --/M Cherokee_ iro/chr + 5 cmn --/M Chinese_(Mandarin,_latin_as_English) sit/cmn (zh-cmn 5)(zh 5) + 5 cmn-latn-pinyin --/M Chinese_(Mandarin,_latin_as_Pinyin) sit/cmn-Latn-pinyin (zh-cmn 5)(zh 5) + 5 cs --/M Czech zlw/cs + 5 cv --/M Chuvash trk/cv + 5 cy --/M Welsh cel/cy + 5 da --/M Danish gmq/da + 5 de --/M German gmw/de + 5 el --/M Greek grk/el + 5 en-029 --/M English_(Caribbean) gmw/en-029 (en 10) + 2 en-gb --/M English_(Great_Britain) gmw/en (en 2) + 5 en-gb-scotland --/M English_(Scotland) gmw/en-GB-scotland (en 4) + 5 en-gb-x-gbclan --/M English_(Lancaster) gmw/en-GB-x-gbclan (en-gb 3)(en 5) + 5 en-gb-x-gbcwmd --/M English_(West_Midlands) gmw/en-GB-x-gbcwmd (en-gb 9)(en 9) + 5 en-gb-x-rp --/M English_(Received_Pronunciation) gmw/en-GB-x-rp (en-gb 4)(en 5) + 2 en-us --/M English_(America) gmw/en-US (en 3) + 5 en-us-nyc --/M English_(America,_New_York_City) gmw/en-US-nyc + 5 eo --/M Esperanto art/eo + 5 es --/M Spanish_(Spain) roa/es + 5 es-419 --/M Spanish_(Latin_America) roa/es-419 (es-mx 6) + 5 et --/M Estonian urj/et + 5 eu --/M Basque eu + 5 fa --/M Persian ira/fa + 5 fa-latn --/M Persian_(Pinglish) ira/fa-Latn + 5 fi --/M Finnish urj/fi + 5 fr-be --/M French_(Belgium) roa/fr-BE (fr 8) + 5 fr-ch --/M French_(Switzerland) roa/fr-CH (fr 8) + 5 fr-fr --/M French_(France) roa/fr (fr 5) + 5 ga --/M Gaelic_(Irish) cel/ga + 5 gd --/M Gaelic_(Scottish) cel/gd + 5 gn --/M Guarani sai/gn + 5 grc --/M Greek_(Ancient) grk/grc + 5 gu --/M Gujarati inc/gu + 5 hak --/M Hakka_Chinese sit/hak + 5 haw --/M Hawaiian map/haw + 5 he --/M Hebrew sem/he + 5 hi --/M Hindi inc/hi + 5 hr --/M Croatian zls/hr (hbs 5) + 5 ht --/M Haitian_Creole roa/ht + 5 hu --/M Hungarian urj/hu + 5 hy --/M Armenian_(East_Armenia) ine/hy (hy-arevela 5) + 5 hyw --/M Armenian_(West_Armenia) ine/hyw (hy-arevmda 5)(hy 8) + 5 ia --/M Interlingua art/ia + 5 id --/M Indonesian poz/id + 5 io --/M Ido art/io + 5 is --/M Icelandic gmq/is + 5 it --/M Italian roa/it + 5 ja --/M Japanese jpx/ja + 5 jbo --/M Lojban art/jbo + 5 ka --/M Georgian ccs/ka + 5 kk --/M Kazakh trk/kk + 5 kl --/M Greenlandic esx/kl + 5 kn --/M Kannada dra/kn + 5 ko --/M Korean ko + 5 kok --/M Konkani inc/kok + 5 ku --/M Kurdish ira/ku + 5 ky --/M Kyrgyz trk/ky + 5 la --/M Latin itc/la + 5 lb --/M Luxembourgish gmw/lb + 5 lfn --/M Lingua_Franca_Nova art/lfn + 5 lt --/M Lithuanian bat/lt + 5 ltg --/M Latgalian bat/ltg + 5 lv --/M Latvian bat/lv + 5 mi --/M Māori poz/mi + 5 mk --/M Macedonian zls/mk + 5 ml --/M Malayalam dra/ml + 5 mr --/M Marathi inc/mr + 5 ms --/M Malay poz/ms + 5 mt --/M Maltese sem/mt + 5 mto --/M Totontepec_Mixe miz/mto + 5 my --/M Myanmar_(Burmese) sit/my + 5 nb --/M Norwegian_Bokmål gmq/nb (no 5) + 5 nci --/M Nahuatl_(Classical) azc/nci + 5 ne --/M Nepali inc/ne + 5 nl --/M Dutch gmw/nl + 5 nog --/M Nogai trk/nog + 5 om --/M Oromo cus/om + 5 or --/M Oriya inc/or + 5 pa --/M Punjabi inc/pa + 5 pap --/M Papiamento roa/pap + 5 piqd --/M Klingon art/piqd + 5 pl --/M Polish zlw/pl + 5 pt --/M Portuguese_(Portugal) roa/pt (pt-pt 5) + 5 pt-br --/M Portuguese_(Brazil) roa/pt-BR (pt 6) + 5 py --/M Pyash art/py + 5 qdb --/M Lang_Belta art/qdb + 5 qu --/M Quechua qu + 5 quc --/M K'iche' myn/quc + 5 qya --/M Quenya art/qya + 5 ro --/M Romanian roa/ro + 5 ru --/M Russian zle/ru + 5 ru-cl --/M Russian_(Classic) zle/ru-cl + 2 ru-lv --/M Russian_(Latvia) zle/ru-LV + 5 sd --/M Sindhi inc/sd + 5 shn --/M Shan_(Tai_Yai) tai/shn + 5 si --/M Sinhala inc/si + 5 sjn --/M Sindarin art/sjn + 5 sk --/M Slovak zlw/sk + 5 sl --/M Slovenian zls/sl + 5 smj --/M Lule_Saami urj/smj + 5 sq --/M Albanian ine/sq + 5 sr --/M Serbian zls/sr + 5 sv --/M Swedish gmq/sv + 5 sw --/M Swahili bnt/sw + 5 ta --/M Tamil dra/ta + 5 te --/M Telugu dra/te + 5 th --/M Thai tai/th + 5 tk --/M Turkmen trk/tk + 5 tn --/M Setswana bnt/tn + 5 tr --/M Turkish trk/tr + 5 tt --/M Tatar trk/tt + 5 ug --/M Uyghur trk/ug + 5 uk --/M Ukrainian zle/uk + 5 ur --/M Urdu inc/ur + 5 uz --/M Uzbek trk/uz + 5 vi --/M Vietnamese_(Northern) aav/vi + 5 vi-vn-x-central --/M Vietnamese_(Central) aav/vi-VN-x-central + 5 vi-vn-x-south --/M Vietnamese_(Southern) aav/vi-VN-x-south + 5 yue --/M Chinese_(Cantonese) sit/yue (zh-yue 5)(zh 8) + 5 yue --/M Chinese_(Cantonese,_latin_as_Jyutping) sit/yue-Latn-jyutping (zh-yue 5)(zh 8) +""" diff --git a/egs/ljspeech/TTS/vits/generator.py b/egs/ljspeech/TTS/vits/generator.py index 66c8cedb1..b9add9e82 100644 --- a/egs/ljspeech/TTS/vits/generator.py +++ b/egs/ljspeech/TTS/vits/generator.py @@ -189,7 +189,7 @@ class VITSGenerator(torch.nn.Module): self.upsample_factor = int(np.prod(decoder_upsample_scales)) self.spks = None if spks is not None and spks > 1: - assert global_channels > 0 + assert global_channels > 0, global_channels self.spks = spks self.global_emb = torch.nn.Embedding(spks, global_channels) self.spk_embed_dim = None diff --git a/egs/ljspeech/TTS/vits/infer.py b/egs/ljspeech/TTS/vits/infer.py index cf0d20ae2..7be76e315 100755 --- a/egs/ljspeech/TTS/vits/infer.py +++ b/egs/ljspeech/TTS/vits/infer.py @@ -72,6 +72,16 @@ def get_parser(): help="""Path to vocabulary.""", ) + parser.add_argument( + "--model-type", + type=str, + default="high", + choices=["low", "medium", "high"], + help="""If not empty, valid values are: low, medium, high. + It controls the model size. low -> runs faster. + """, + ) + return parser @@ -94,6 +104,7 @@ def infer_dataset( tokenizer: Used to convert text to phonemes. """ + # Background worker save audios to disk. def _save_worker( batch_size: int, @@ -130,14 +141,16 @@ def infer_dataset( batch_size = len(batch["tokens"]) tokens = batch["tokens"] - tokens = tokenizer.tokens_to_token_ids(tokens) + tokens = tokenizer.tokens_to_token_ids( + tokens, intersperse_blank=True, add_sos=True, add_eos=True + ) tokens = k2.RaggedTensor(tokens) row_splits = tokens.shape.row_splits(1) tokens_lens = row_splits[1:] - row_splits[:-1] tokens = tokens.to(device) tokens_lens = tokens_lens.to(device) # tensor of shape (B, T) - tokens = tokens.pad(mode="constant", padding_value=tokenizer.blank_id) + tokens = tokens.pad(mode="constant", padding_value=tokenizer.pad_id) audio = batch["audio"] audio_lens = batch["audio_lens"].tolist() @@ -201,8 +214,7 @@ def main(): device = torch.device("cuda", 0) tokenizer = Tokenizer(params.tokens) - params.blank_id = tokenizer.blank_id - params.oov_id = tokenizer.oov_id + params.blank_id = tokenizer.pad_id params.vocab_size = tokenizer.vocab_size logging.info(f"Device: {device}") diff --git a/egs/ljspeech/TTS/vits/monotonic_align/__init__.py b/egs/ljspeech/TTS/vits/monotonic_align/__init__.py index 2b35654f5..5dc3641e5 100644 --- a/egs/ljspeech/TTS/vits/monotonic_align/__init__.py +++ b/egs/ljspeech/TTS/vits/monotonic_align/__init__.py @@ -10,7 +10,11 @@ import warnings import numpy as np import torch -from numba import njit, prange + +try: + from numba import njit, prange +except ModuleNotFoundError as ex: + raise RuntimeError(f"{ex}/nPlease run\n pip install numba") try: from .core import maximum_path_c diff --git a/egs/ljspeech/TTS/vits/test_model.py b/egs/ljspeech/TTS/vits/test_model.py new file mode 100755 index 000000000..1de10f012 --- /dev/null +++ b/egs/ljspeech/TTS/vits/test_model.py @@ -0,0 +1,50 @@ +#!/usr/bin/env python3 +# Copyright 2024 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. + + +from tokenizer import Tokenizer +from train import get_model, get_params +from vits import VITS + + +def test_model_type(model_type): + tokens = "./data/tokens.txt" + + params = get_params() + + tokenizer = Tokenizer(tokens) + params.blank_id = tokenizer.pad_id + params.vocab_size = tokenizer.vocab_size + params.model_type = model_type + + model = get_model(params) + generator = model.generator + + num_param = sum([p.numel() for p in generator.parameters()]) + print( + f"{model_type}: generator parameters: {num_param}, or {num_param/1000/1000} M" + ) + + +def main(): + test_model_type("high") # 35.63 M + test_model_type("low") # 7.55 M + test_model_type("medium") # 23.61 M + + +if __name__ == "__main__": + main() diff --git a/egs/ljspeech/TTS/vits/test_onnx.py b/egs/ljspeech/TTS/vits/test_onnx.py index fcbc1d663..b3805fadb 100755 --- a/egs/ljspeech/TTS/vits/test_onnx.py +++ b/egs/ljspeech/TTS/vits/test_onnx.py @@ -54,6 +54,20 @@ def get_parser(): help="""Path to vocabulary.""", ) + parser.add_argument( + "--text", + type=str, + default="Ask not what your country can do for you; ask what you can do for your country.", + help="Text to generate speech for", + ) + + parser.add_argument( + "--output-filename", + type=str, + default="test_onnx.wav", + help="Filename to save the generated wave file.", + ) + return parser @@ -61,7 +75,7 @@ class OnnxModel: def __init__(self, model_filename: str): session_opts = ort.SessionOptions() session_opts.inter_op_num_threads = 1 - session_opts.intra_op_num_threads = 4 + session_opts.intra_op_num_threads = 1 self.session_opts = session_opts @@ -72,6 +86,9 @@ class OnnxModel: ) logging.info(f"{self.model.get_modelmeta().custom_metadata_map}") + metadata = self.model.get_modelmeta().custom_metadata_map + self.sample_rate = int(metadata["sample_rate"]) + def __call__(self, tokens: torch.Tensor, tokens_lens: torch.Tensor) -> torch.Tensor: """ Args: @@ -101,20 +118,24 @@ class OnnxModel: def main(): args = get_parser().parse_args() + logging.info(vars(args)) tokenizer = Tokenizer(args.tokens) logging.info("About to create onnx model") model = OnnxModel(args.model_filename) - text = "I went there to see the land, the people and how their system works, end quote." - tokens = tokenizer.texts_to_token_ids([text]) + text = args.text + tokens = tokenizer.texts_to_token_ids( + [text], intersperse_blank=True, add_sos=True, add_eos=True + ) tokens = torch.tensor(tokens) # (1, T) tokens_lens = torch.tensor([tokens.shape[1]], dtype=torch.int64) # (1, T) audio = model(tokens, tokens_lens) # (1, T') - torchaudio.save(str("test_onnx.wav"), audio, sample_rate=22050) - logging.info("Saved to test_onnx.wav") + output_filename = args.output_filename + torchaudio.save(output_filename, audio, sample_rate=model.sample_rate) + logging.info(f"Saved to {output_filename}") if __name__ == "__main__": diff --git a/egs/ljspeech/TTS/vits/text_encoder.py b/egs/ljspeech/TTS/vits/text_encoder.py index fcbae7103..9b21ed9cb 100644 --- a/egs/ljspeech/TTS/vits/text_encoder.py +++ b/egs/ljspeech/TTS/vits/text_encoder.py @@ -92,9 +92,9 @@ class TextEncoder(torch.nn.Module): x_lengths (Tensor): Length tensor (B,). Returns: - Tensor: Encoded hidden representation (B, attention_dim, T_text). - Tensor: Projected mean tensor (B, attention_dim, T_text). - Tensor: Projected scale tensor (B, attention_dim, T_text). + Tensor: Encoded hidden representation (B, embed_dim, T_text). + Tensor: Projected mean tensor (B, embed_dim, T_text). + Tensor: Projected scale tensor (B, embed_dim, T_text). Tensor: Mask tensor for input tensor (B, 1, T_text). """ @@ -108,6 +108,7 @@ class TextEncoder(torch.nn.Module): # encoder assume the channel last (B, T_text, embed_dim) x = self.encoder(x, key_padding_mask=pad_mask) + # Note: attention_dim == embed_dim # convert the channel first (B, embed_dim, T_text) x = x.transpose(1, 2) diff --git a/egs/ljspeech/TTS/vits/tokenizer.py b/egs/ljspeech/TTS/vits/tokenizer.py index b0afc6a04..3c9046add 100644 --- a/egs/ljspeech/TTS/vits/tokenizer.py +++ b/egs/ljspeech/TTS/vits/tokenizer.py @@ -1,4 +1,4 @@ -# Copyright 2023 Xiaomi Corp. (authors: Zengwei Yao) +# Copyright 2023-2024 Xiaomi Corp. (authors: Zengwei Yao) # # See ../../LICENSE for clarification regarding multiple authors # @@ -14,10 +14,19 @@ # See the License for the specific language governing permissions and # limitations under the License. +import logging from typing import Dict, List -import g2p_en import tacotron_cleaner.cleaners + +try: + from piper_phonemize import phonemize_espeak +except Exception as ex: + raise RuntimeError( + f"{ex}\nPlease run\n" + "pip install piper_phonemize -f https://k2-fsa.github.io/icefall/piper_phonemize.html" + ) + from utils import intersperse @@ -38,21 +47,37 @@ class Tokenizer(object): id = int(info[0]) else: token, id = info[0], int(info[1]) + assert token not in self.token2id, token self.token2id[token] = id - self.blank_id = self.token2id[""] - self.oov_id = self.token2id[""] + # Refer to https://github.com/rhasspy/piper/blob/master/TRAINING.md + self.pad_id = self.token2id["_"] # padding + self.sos_id = self.token2id["^"] # beginning of an utterance (bos) + self.eos_id = self.token2id["$"] # end of an utterance (eos) + self.space_id = self.token2id[" "] # word separator (whitespace) + self.vocab_size = len(self.token2id) - self.g2p = g2p_en.G2p() - - def texts_to_token_ids(self, texts: List[str], intersperse_blank: bool = True): + def texts_to_token_ids( + self, + texts: List[str], + intersperse_blank: bool = True, + add_sos: bool = False, + add_eos: bool = False, + lang: str = "en-us", + ) -> List[List[int]]: """ Args: texts: A list of transcripts. intersperse_blank: Whether to intersperse blanks in the token sequence. + add_sos: + Whether to add sos token at the start. + add_eos: + Whether to add eos token at the end. + lang: + Language argument passed to phonemize_espeak(). Returns: Return a list of token id list [utterance][token_id] @@ -63,30 +88,46 @@ class Tokenizer(object): # Text normalization text = tacotron_cleaner.cleaners.custom_english_cleaners(text) # Convert to phonemes - tokens = self.g2p(text) + tokens_list = phonemize_espeak(text, lang) + tokens = [] + for t in tokens_list: + tokens.extend(t) + token_ids = [] for t in tokens: - if t in self.token2id: - token_ids.append(self.token2id[t]) - else: - token_ids.append(self.oov_id) + if t not in self.token2id: + logging.warning(f"Skip OOV {t}") + continue + token_ids.append(self.token2id[t]) if intersperse_blank: - token_ids = intersperse(token_ids, self.blank_id) + token_ids = intersperse(token_ids, self.pad_id) + if add_sos: + token_ids = [self.sos_id] + token_ids + if add_eos: + token_ids = token_ids + [self.eos_id] token_ids_list.append(token_ids) return token_ids_list def tokens_to_token_ids( - self, tokens_list: List[str], intersperse_blank: bool = True - ): + self, + tokens_list: List[str], + intersperse_blank: bool = True, + add_sos: bool = False, + add_eos: bool = False, + ) -> List[List[int]]: """ Args: tokens_list: A list of token list, each corresponding to one utterance. intersperse_blank: Whether to intersperse blanks in the token sequence. + add_sos: + Whether to add sos token at the start. + add_eos: + Whether to add eos token at the end. Returns: Return a list of token id list [utterance][token_id] @@ -96,13 +137,17 @@ class Tokenizer(object): for tokens in tokens_list: token_ids = [] for t in tokens: - if t in self.token2id: - token_ids.append(self.token2id[t]) - else: - token_ids.append(self.oov_id) + if t not in self.token2id: + logging.warning(f"Skip OOV {t}") + continue + token_ids.append(self.token2id[t]) if intersperse_blank: - token_ids = intersperse(token_ids, self.blank_id) + token_ids = intersperse(token_ids, self.pad_id) + if add_sos: + token_ids = [self.sos_id] + token_ids + if add_eos: + token_ids = token_ids + [self.eos_id] token_ids_list.append(token_ids) diff --git a/egs/ljspeech/TTS/vits/train.py b/egs/ljspeech/TTS/vits/train.py index 71c4224fa..34b943765 100755 --- a/egs/ljspeech/TTS/vits/train.py +++ b/egs/ljspeech/TTS/vits/train.py @@ -153,6 +153,16 @@ def get_parser(): help="Whether to use half precision training.", ) + parser.add_argument( + "--model-type", + type=str, + default="high", + choices=["low", "medium", "high"], + help="""If not empty, valid values are: low, medium, high. + It controls the model size. low -> runs faster. + """, + ) + return parser @@ -189,15 +199,6 @@ def get_params() -> AttributeDict: - feature_dim: The model input dim. It has to match the one used in computing features. - - - subsampling_factor: The subsampling factor for the model. - - - encoder_dim: Hidden dim for multi-head attention model. - - - num_decoder_layers: Number of decoder layer of transformer decoder. - - - warm_step: The warmup period that dictates the decay of the - scale on "simple" (un-pruned) loss. """ params = AttributeDict( { @@ -278,6 +279,7 @@ def get_model(params: AttributeDict) -> nn.Module: vocab_size=params.vocab_size, feature_dim=params.feature_dim, sampling_rate=params.sampling_rate, + model_type=params.model_type, mel_loss_params=mel_loss_params, lambda_adv=params.lambda_adv, lambda_mel=params.lambda_mel, @@ -296,14 +298,16 @@ def prepare_input(batch: dict, tokenizer: Tokenizer, device: torch.device): features_lens = batch["features_lens"].to(device) tokens = batch["tokens"] - tokens = tokenizer.tokens_to_token_ids(tokens) + tokens = tokenizer.tokens_to_token_ids( + tokens, intersperse_blank=True, add_sos=True, add_eos=True + ) tokens = k2.RaggedTensor(tokens) row_splits = tokens.shape.row_splits(1) tokens_lens = row_splits[1:] - row_splits[:-1] tokens = tokens.to(device) tokens_lens = tokens_lens.to(device) # a tensor of shape (B, T) - tokens = tokens.pad(mode="constant", padding_value=tokenizer.blank_id) + tokens = tokens.pad(mode="constant", padding_value=tokenizer.pad_id) return audio, audio_lens, features, features_lens, tokens, tokens_lens @@ -361,7 +365,7 @@ def train_one_epoch( model.train() device = model.device if isinstance(model, DDP) else next(model.parameters()).device - # used to summary the stats over iterations in one epoch + # used to track the stats over iterations in one epoch tot_loss = MetricsTracker() saved_bad_model = False @@ -742,8 +746,7 @@ def run(rank, world_size, args): logging.info(f"Device: {device}") tokenizer = Tokenizer(params.tokens) - params.blank_id = tokenizer.blank_id - params.oov_id = tokenizer.oov_id + params.blank_id = tokenizer.pad_id params.vocab_size = tokenizer.vocab_size logging.info(params) diff --git a/egs/ljspeech/TTS/vits/tts_datamodule.py b/egs/ljspeech/TTS/vits/tts_datamodule.py index 8ff868bc8..e1a9c7b3c 100644 --- a/egs/ljspeech/TTS/vits/tts_datamodule.py +++ b/egs/ljspeech/TTS/vits/tts_datamodule.py @@ -255,6 +255,7 @@ class LJSpeechTtsDataModule: valid_sampler = DynamicBucketingSampler( cuts_valid, max_duration=self.args.max_duration, + num_buckets=self.args.num_buckets, shuffle=False, ) logging.info("About to create valid dataloader") @@ -294,6 +295,7 @@ class LJSpeechTtsDataModule: test_sampler = DynamicBucketingSampler( cuts, max_duration=self.args.max_duration, + num_buckets=self.args.num_buckets, shuffle=False, ) logging.info("About to create test dataloader") diff --git a/egs/ljspeech/TTS/vits/vits.py b/egs/ljspeech/TTS/vits/vits.py index b4f0c21e6..0b9575cbd 100644 --- a/egs/ljspeech/TTS/vits/vits.py +++ b/egs/ljspeech/TTS/vits/vits.py @@ -5,6 +5,7 @@ """VITS module for GAN-TTS task.""" +import copy from typing import Any, Dict, Optional, Tuple import torch @@ -38,6 +39,36 @@ AVAILABLE_DISCRIMINATORS = { "hifigan_multi_scale_multi_period_discriminator": HiFiGANMultiScaleMultiPeriodDiscriminator, # NOQA } +LOW_CONFIG = { + "hidden_channels": 96, + "decoder_upsample_scales": (8, 8, 4), + "decoder_channels": 256, + "decoder_upsample_kernel_sizes": (16, 16, 8), + "decoder_resblock_kernel_sizes": (3, 5, 7), + "decoder_resblock_dilations": ((1, 2), (2, 6), (3, 12)), + "text_encoder_cnn_module_kernel": 3, +} + +MEDIUM_CONFIG = { + "hidden_channels": 192, + "decoder_upsample_scales": (8, 8, 4), + "decoder_channels": 256, + "decoder_upsample_kernel_sizes": (16, 16, 8), + "decoder_resblock_kernel_sizes": (3, 5, 7), + "decoder_resblock_dilations": ((1, 2), (2, 6), (3, 12)), + "text_encoder_cnn_module_kernel": 3, +} + +HIGH_CONFIG = { + "hidden_channels": 192, + "decoder_upsample_scales": (8, 8, 2, 2), + "decoder_channels": 512, + "decoder_upsample_kernel_sizes": (16, 16, 4, 4), + "decoder_resblock_kernel_sizes": (3, 7, 11), + "decoder_resblock_dilations": ((1, 3, 5), (1, 3, 5), (1, 3, 5)), + "text_encoder_cnn_module_kernel": 5, +} + class VITS(nn.Module): """Implement VITS, `Conditional Variational Autoencoder with Adversarial Learning for End-to-End Text-to-Speech`""" @@ -49,6 +80,7 @@ class VITS(nn.Module): feature_dim: int = 513, sampling_rate: int = 22050, generator_type: str = "vits_generator", + model_type: str = "", generator_params: Dict[str, Any] = { "hidden_channels": 192, "spks": None, @@ -155,12 +187,13 @@ class VITS(nn.Module): """Initialize VITS module. Args: - idim (int): Input vocabrary size. + idim (int): Input vocabulary size. odim (int): Acoustic feature dimension. The actual output channels will be 1 since VITS is the end-to-end text-to-wave model but for the compatibility odim is used to indicate the acoustic feature dimension. sampling_rate (int): Sampling rate, not used for the training but it will be referred in saving waveform during the inference. + model_type (str): If not empty, must be one of: low, medium, high generator_type (str): Generator type. generator_params (Dict[str, Any]): Parameter dict for generator. discriminator_type (str): Discriminator type. @@ -181,6 +214,24 @@ class VITS(nn.Module): """ super().__init__() + generator_params = copy.deepcopy(generator_params) + discriminator_params = copy.deepcopy(discriminator_params) + generator_adv_loss_params = copy.deepcopy(generator_adv_loss_params) + discriminator_adv_loss_params = copy.deepcopy(discriminator_adv_loss_params) + feat_match_loss_params = copy.deepcopy(feat_match_loss_params) + mel_loss_params = copy.deepcopy(mel_loss_params) + + if model_type != "": + assert model_type in ("low", "medium", "high"), model_type + if model_type == "low": + generator_params.update(LOW_CONFIG) + elif model_type == "medium": + generator_params.update(MEDIUM_CONFIG) + elif model_type == "high": + generator_params.update(HIGH_CONFIG) + else: + raise ValueError(f"Unknown model_type: ${model_type}") + # define modules generator_class = AVAILABLE_GENERATERS[generator_type] if generator_type == "vits_generator": diff --git a/egs/mdcc/ASR/README.md b/egs/mdcc/ASR/README.md new file mode 100644 index 000000000..112845b73 --- /dev/null +++ b/egs/mdcc/ASR/README.md @@ -0,0 +1,19 @@ +# Introduction + +Multi-Domain Cantonese Corpus (MDCC), consists of 73.6 hours of clean read speech paired with +transcripts, collected from Cantonese audiobooks from Hong Kong. It comprises philosophy, +politics, education, culture, lifestyle and family domains, covering a wide range of topics. + +Manuscript can be found at: https://arxiv.org/abs/2201.02419 + +# Transducers + + + +| | Encoder | Decoder | Comment | +|---------------------------------------|---------------------|--------------------|-----------------------------| +| `zipformer` | Upgraded Zipformer | Embedding + Conv1d | The latest recipe with context-size set to 1 | + +The decoder is modified from the paper +[Rnn-Transducer with Stateless Prediction Network](https://ieeexplore.ieee.org/document/9054419/). +We place an additional Conv1d layer right after the input embedding layer. diff --git a/egs/mdcc/ASR/RESULTS.md b/egs/mdcc/ASR/RESULTS.md new file mode 100644 index 000000000..ff7ddc957 --- /dev/null +++ b/egs/mdcc/ASR/RESULTS.md @@ -0,0 +1,41 @@ +## Results + +#### Zipformer + +See + +[./zipformer](./zipformer) + +##### normal-scaled model, number of model parameters: 74470867, i.e., 74.47 M + +| | test | valid | comment | +|------------------------|------|-------|-----------------------------------------| +| greedy search | 7.45 | 7.51 | --epoch 45 --avg 35 | +| modified beam search | 6.68 | 6.73 | --epoch 45 --avg 35 | +| fast beam search | 7.22 | 7.28 | --epoch 45 --avg 35 | + +The training command: + +``` +export CUDA_VISIBLE_DEVICES="0,1,2,3" + +./zipformer/train.py \ + --world-size 4 \ + --start-epoch 1 \ + --num-epochs 50 \ + --use-fp16 1 \ + --exp-dir ./zipformer/exp \ + --max-duration 1000 +``` + +The decoding command: + +``` + ./zipformer/decode.py \ + --epoch 45 \ + --avg 35 \ + --exp-dir ./zipformer/exp \ + --decoding-method greedy_search # modified_beam_search +``` + +The pretrained model is available at: https://huggingface.co/zrjin/icefall-asr-mdcc-zipformer-2024-03-11/ \ No newline at end of file diff --git a/egs/mdcc/ASR/local/compile_hlg.py b/egs/mdcc/ASR/local/compile_hlg.py new file mode 120000 index 000000000..471aa7fb4 --- /dev/null +++ b/egs/mdcc/ASR/local/compile_hlg.py @@ -0,0 +1 @@ +../../../librispeech/ASR/local/compile_hlg.py \ No newline at end of file diff --git a/egs/mdcc/ASR/local/compile_hlg_using_openfst.py b/egs/mdcc/ASR/local/compile_hlg_using_openfst.py new file mode 120000 index 000000000..d34edd7f3 --- /dev/null +++ b/egs/mdcc/ASR/local/compile_hlg_using_openfst.py @@ -0,0 +1 @@ +../../../librispeech/ASR/local/compile_hlg_using_openfst.py \ No newline at end of file diff --git a/egs/mdcc/ASR/local/compile_lg.py b/egs/mdcc/ASR/local/compile_lg.py new file mode 120000 index 000000000..462d6d3fb --- /dev/null +++ b/egs/mdcc/ASR/local/compile_lg.py @@ -0,0 +1 @@ +../../../librispeech/ASR/local/compile_lg.py \ No newline at end of file diff --git a/egs/mdcc/ASR/local/compute_fbank_mdcc.py b/egs/mdcc/ASR/local/compute_fbank_mdcc.py new file mode 100755 index 000000000..647b21127 --- /dev/null +++ b/egs/mdcc/ASR/local/compute_fbank_mdcc.py @@ -0,0 +1,157 @@ +#!/usr/bin/env python3 +# Copyright 2021-2024 Xiaomi Corp. (authors: Fangjun Kuang, +# Zengrui Jin,) +# +# 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 file computes fbank features of the aishell dataset. +It looks for manifests in the directory data/manifests. + +The generated fbank features are saved in data/fbank. +""" + +import argparse +import logging +import os +from pathlib import Path + +import torch +from lhotse import ( + CutSet, + Fbank, + FbankConfig, + LilcomChunkyWriter, + WhisperFbank, + WhisperFbankConfig, +) +from lhotse.recipes.utils import read_manifests_if_cached + +from icefall.utils import get_executor, str2bool + +# Torch's multithreaded behavior needs to be disabled or +# it wastes a lot of CPU and slow things down. +# Do this outside of main() in case it needs to take effect +# even when we are not invoking the main (e.g. when spawning subprocesses). +torch.set_num_threads(1) +torch.set_num_interop_threads(1) + + +def compute_fbank_mdcc( + num_mel_bins: int = 80, + perturb_speed: bool = False, + whisper_fbank: bool = False, + output_dir: str = "data/fbank", +): + src_dir = Path("data/manifests") + output_dir = Path(output_dir) + num_jobs = min(15, os.cpu_count()) + + dataset_parts = ( + "train", + "valid", + "test", + ) + prefix = "mdcc" + suffix = "jsonl.gz" + manifests = read_manifests_if_cached( + dataset_parts=dataset_parts, + output_dir=src_dir, + prefix=prefix, + suffix=suffix, + ) + assert manifests is not None + + assert len(manifests) == len(dataset_parts), ( + len(manifests), + len(dataset_parts), + list(manifests.keys()), + dataset_parts, + ) + if whisper_fbank: + extractor = WhisperFbank( + WhisperFbankConfig(num_filters=num_mel_bins, device="cuda") + ) + else: + extractor = Fbank(FbankConfig(num_mel_bins=num_mel_bins)) + + with get_executor() as ex: # Initialize the executor only once. + for partition, m in manifests.items(): + if (output_dir / f"{prefix}_cuts_{partition}.{suffix}").is_file(): + logging.info(f"{partition} already exists - skipping.") + continue + logging.info(f"Processing {partition}") + cut_set = CutSet.from_manifests( + recordings=m["recordings"], + supervisions=m["supervisions"], + ) + if "train" in partition and perturb_speed: + logging.info("Doing speed perturb") + cut_set = ( + cut_set + cut_set.perturb_speed(0.9) + cut_set.perturb_speed(1.1) + ) + cut_set = cut_set.compute_and_store_features( + extractor=extractor, + storage_path=f"{output_dir}/{prefix}_feats_{partition}", + # when an executor is specified, make more partitions + num_jobs=num_jobs if ex is None else 80, + executor=ex, + storage_type=LilcomChunkyWriter, + ) + cut_set.to_file(output_dir / f"{prefix}_cuts_{partition}.{suffix}") + + +def get_args(): + parser = argparse.ArgumentParser() + parser.add_argument( + "--num-mel-bins", + type=int, + default=80, + help="""The number of mel bins for Fbank""", + ) + parser.add_argument( + "--perturb-speed", + type=str2bool, + default=False, + help="Enable 0.9 and 1.1 speed perturbation for data augmentation. Default: False.", + ) + parser.add_argument( + "--whisper-fbank", + type=str2bool, + default=False, + help="Use WhisperFbank instead of Fbank. Default: False.", + ) + parser.add_argument( + "--output-dir", + type=str, + default="data/fbank", + help="Output directory. Default: data/fbank.", + ) + return parser.parse_args() + + +if __name__ == "__main__": + formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s" + + logging.basicConfig(format=formatter, level=logging.INFO) + + args = get_args() + compute_fbank_mdcc( + num_mel_bins=args.num_mel_bins, + perturb_speed=args.perturb_speed, + whisper_fbank=args.whisper_fbank, + output_dir=args.output_dir, + ) diff --git a/egs/mdcc/ASR/local/display_manifest_statistics.py b/egs/mdcc/ASR/local/display_manifest_statistics.py new file mode 100755 index 000000000..27cf8c943 --- /dev/null +++ b/egs/mdcc/ASR/local/display_manifest_statistics.py @@ -0,0 +1,144 @@ +#!/usr/bin/env python3 +# Copyright 2021-2024 Xiaomi Corp. (authors: Fangjun Kuang, +# Zengrui Jin,) +# +# 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 file displays duration statistics of utterances in a manifest. +You can use the displayed value to choose minimum/maximum duration +to remove short and long utterances during the training. + +See the function `remove_short_and_long_utt()` in transducer/train.py +for usage. +""" + + +from lhotse import load_manifest_lazy + + +def main(): + path = "./data/fbank/mdcc_cuts_train.jsonl.gz" + path = "./data/fbank/mdcc_cuts_valid.jsonl.gz" + path = "./data/fbank/mdcc_cuts_test.jsonl.gz" + + cuts = load_manifest_lazy(path) + cuts.describe(full=True) + + +if __name__ == "__main__": + main() + +""" +data/fbank/mdcc_cuts_train.jsonl.gz (with speed perturbation) +_________________________________________ +_ Cuts count: _ 195360 +_________________________________________ +_ Total duration (hh:mm:ss) _ 173:44:59 +_________________________________________ +_ mean _ 3.2 +_________________________________________ +_ std _ 2.1 +_________________________________________ +_ min _ 0.2 +_________________________________________ +_ 25% _ 1.8 +_________________________________________ +_ 50% _ 2.7 +_________________________________________ +_ 75% _ 4.0 +_________________________________________ +_ 99% _ 11.0 _ +_________________________________________ +_ 99.5% _ 12.4 _ +_________________________________________ +_ 99.9% _ 14.8 _ +_________________________________________ +_ max _ 16.7 _ +_________________________________________ +_ Recordings available: _ 195360 _ +_________________________________________ +_ Features available: _ 195360 _ +_________________________________________ +_ Supervisions available: _ 195360 _ +_________________________________________ + +data/fbank/mdcc_cuts_valid.jsonl.gz +________________________________________ +_ Cuts count: _ 5663 _ +________________________________________ +_ Total duration (hh:mm:ss) _ 05:03:12 _ +________________________________________ +_ mean _ 3.2 _ +________________________________________ +_ std _ 2.0 _ +________________________________________ +_ min _ 0.3 _ +________________________________________ +_ 25% _ 1.8 _ +________________________________________ +_ 50% _ 2.7 _ +________________________________________ +_ 75% _ 4.0 _ +________________________________________ +_ 99% _ 10.9 _ +________________________________________ +_ 99.5% _ 12.3 _ +________________________________________ +_ 99.9% _ 14.4 _ +________________________________________ +_ max _ 14.8 _ +________________________________________ +_ Recordings available: _ 5663 _ +________________________________________ +_ Features available: _ 5663 _ +________________________________________ +_ Supervisions available: _ 5663 _ +________________________________________ + +data/fbank/mdcc_cuts_test.jsonl.gz +________________________________________ +_ Cuts count: _ 12492 _ +________________________________________ +_ Total duration (hh:mm:ss) _ 11:00:31 _ +________________________________________ +_ mean _ 3.2 _ +________________________________________ +_ std _ 2.0 _ +________________________________________ +_ min _ 0.2 _ +________________________________________ +_ 25% _ 1.8 _ +________________________________________ +_ 50% _ 2.7 _ +________________________________________ +_ 75% _ 4.0 _ +________________________________________ +_ 99% _ 10.5 _ +________________________________________ +_ 99.5% _ 12.1 _ +________________________________________ +_ 99.9% _ 14.0 _ +________________________________________ +_ max _ 14.8 _ +________________________________________ +_ Recordings available: _ 12492 _ +________________________________________ +_ Features available: _ 12492 _ +________________________________________ +_ Supervisions available: _ 12492 _ +________________________________________ + +""" diff --git a/egs/mdcc/ASR/local/prepare_char.py b/egs/mdcc/ASR/local/prepare_char.py new file mode 120000 index 000000000..42743b544 --- /dev/null +++ b/egs/mdcc/ASR/local/prepare_char.py @@ -0,0 +1 @@ +../../../aishell/ASR/local/prepare_char.py \ No newline at end of file diff --git a/egs/mdcc/ASR/local/prepare_char_lm_training_data.py b/egs/mdcc/ASR/local/prepare_char_lm_training_data.py new file mode 120000 index 000000000..2374cafdd --- /dev/null +++ b/egs/mdcc/ASR/local/prepare_char_lm_training_data.py @@ -0,0 +1 @@ +../../../aishell/ASR/local/prepare_char_lm_training_data.py \ No newline at end of file diff --git a/egs/mdcc/ASR/local/prepare_lang.py b/egs/mdcc/ASR/local/prepare_lang.py new file mode 120000 index 000000000..bee8d5f03 --- /dev/null +++ b/egs/mdcc/ASR/local/prepare_lang.py @@ -0,0 +1 @@ +../../../aishell/ASR/local/prepare_lang.py \ No newline at end of file diff --git a/egs/mdcc/ASR/local/prepare_lang_fst.py b/egs/mdcc/ASR/local/prepare_lang_fst.py new file mode 120000 index 000000000..c5787c534 --- /dev/null +++ b/egs/mdcc/ASR/local/prepare_lang_fst.py @@ -0,0 +1 @@ +../../../librispeech/ASR/local/prepare_lang_fst.py \ No newline at end of file diff --git a/egs/mdcc/ASR/local/preprocess_mdcc.py b/egs/mdcc/ASR/local/preprocess_mdcc.py new file mode 100755 index 000000000..cd0dc7de8 --- /dev/null +++ b/egs/mdcc/ASR/local/preprocess_mdcc.py @@ -0,0 +1,157 @@ +#!/usr/bin/env python3 +# Copyright 2024 Xiaomi Corp. (authors: Zengrui Jin) +# +# 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 takes a text file "data/lang_char/text" as input, the file consist of +lines each containing a transcript, applies text norm and generates the following +files in the directory "data/lang_char": + - text_norm + - words.txt + - words_no_ids.txt + - text_words_segmentation +""" + +import argparse +import logging +from pathlib import Path +from typing import List + +import pycantonese +from tqdm.auto import tqdm + +from icefall.utils import is_cjk + + +def get_parser(): + parser = argparse.ArgumentParser( + description="Prepare char lexicon", + formatter_class=argparse.ArgumentDefaultsHelpFormatter, + ) + parser.add_argument( + "--input-file", + "-i", + default="data/lang_char/text", + type=str, + help="The input text file", + ) + parser.add_argument( + "--output-dir", + "-o", + default="data/lang_char", + type=str, + help="The output directory", + ) + return parser + + +def get_norm_lines(lines: List[str]) -> List[str]: + def _text_norm(text: str) -> str: + # to cope with the protocol for transcription: + # When taking notes, the annotators adhere to the following guidelines: + # 1) If the audio contains pure music, the annotators mark the label + # "(music)" in the file name of its transcript. 2) If the utterance + # contains one or several sentences with background music or noise, the + # annotators mark the label "(music)" before each sentence in the transcript. + # 3) The annotators use {} symbols to enclose words they are uncertain + # about, for example, {梁佳佳}，我是{}人. + + # here we manually fix some errors in the transcript + + return ( + text.strip() + .replace("(music)", "") + .replace("(music", "") + .replace("{", "") + .replace("}", "") + .replace("BB所以就指腹為親喇", "BB 所以就指腹為親喇") + .upper() + ) + + return [_text_norm(line) for line in lines] + + +def get_word_segments(lines: List[str]) -> List[str]: + # the current pycantonese segmenter does not handle the case when the input + # is code switching, so we need to handle it separately + + new_lines = [] + + for line in tqdm(lines, desc="Segmenting lines"): + try: + # code switching + if len(line.strip().split(" ")) > 1: + segments = [] + for segment in line.strip().split(" "): + if segment.strip() == "": + continue + try: + if not is_cjk(segment[0]): # en segment + segments.append(segment) + else: # zh segment + segments.extend(pycantonese.segment(segment)) + except Exception as e: + logging.error(f"Failed to process segment: {segment}") + raise e + new_lines.append(" ".join(segments) + "\n") + # not code switching + else: + new_lines.append(" ".join(pycantonese.segment(line)) + "\n") + except Exception as e: + logging.error(f"Failed to process line: {line}") + raise e + return new_lines + + +def get_words(lines: List[str]) -> List[str]: + words = set() + for line in tqdm(lines, desc="Getting words"): + words.update(line.strip().split(" ")) + return list(words) + + +if __name__ == "__main__": + parser = get_parser() + args = parser.parse_args() + + input_file = Path(args.input_file) + output_dir = Path(args.output_dir) + + assert output_dir.is_dir(), f"{output_dir} does not exist" + assert input_file.is_file(), f"{input_file} does not exist" + + lines = input_file.read_text(encoding="utf-8").strip().split("\n") + + norm_lines = get_norm_lines(lines) + with open(output_dir / "text_norm", "w+", encoding="utf-8") as f: + f.writelines([line + "\n" for line in norm_lines]) + + text_words_segments = get_word_segments(norm_lines) + with open(output_dir / "text_words_segmentation", "w+", encoding="utf-8") as f: + f.writelines(text_words_segments) + + words = get_words(text_words_segments)[1:] # remove "\n" from words + with open(output_dir / "words_no_ids.txt", "w+", encoding="utf-8") as f: + f.writelines([word + "\n" for word in sorted(words)]) + + words = ( + ["", "!SIL", "", ""] + + sorted(words) + + ["#0", "", "<\s>"] + ) + + with open(output_dir / "words.txt", "w+", encoding="utf-8") as f: + f.writelines([f"{word} {i}\n" for i, word in enumerate(words)]) diff --git a/egs/mdcc/ASR/local/text2segments.py b/egs/mdcc/ASR/local/text2segments.py new file mode 100755 index 000000000..8ce7ab7e5 --- /dev/null +++ b/egs/mdcc/ASR/local/text2segments.py @@ -0,0 +1,86 @@ +#!/usr/bin/env python +# -*- coding: utf-8 -*- + +# Copyright 2021 Xiaomi Corp. (authors: Mingshuang Luo) +# 2022 Xiaomi Corp. (authors: Weiji Zhuang) +# 2024 Xiaomi Corp. (authors: Zengrui Jin) +# +# 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 takes as input "text", which refers to the transcript file for +MDCC: + - text +and generates the output file text_word_segmentation which is implemented +with word segmenting: + - text_words_segmentation +""" + +import argparse +from typing import List + +import pycantonese +from tqdm.auto import tqdm + + +def get_parser(): + parser = argparse.ArgumentParser( + description="Cantonese Word Segmentation for text", + formatter_class=argparse.ArgumentDefaultsHelpFormatter, + ) + parser.add_argument( + "--input-file", + "-i", + default="data/lang_char/text", + type=str, + help="the input text file for MDCC", + ) + parser.add_argument( + "--output-file", + "-o", + default="data/lang_char/text_words_segmentation", + type=str, + help="the text implemented with words segmenting for MDCC", + ) + + return parser + + +def get_word_segments(lines: List[str]) -> List[str]: + return [ + " ".join(pycantonese.segment(line)) + "\n" + for line in tqdm(lines, desc="Segmenting lines") + ] + + +def main(): + parser = get_parser() + args = parser.parse_args() + + input_file = args.input_file + output_file = args.output_file + + with open(input_file, "r", encoding="utf-8") as fr: + lines = fr.readlines() + + new_lines = get_word_segments(lines) + + with open(output_file, "w", encoding="utf-8") as fw: + fw.writelines(new_lines) + + +if __name__ == "__main__": + main() diff --git a/egs/mdcc/ASR/local/text2token.py b/egs/mdcc/ASR/local/text2token.py new file mode 120000 index 000000000..81e459d69 --- /dev/null +++ b/egs/mdcc/ASR/local/text2token.py @@ -0,0 +1 @@ +../../../aidatatang_200zh/ASR/local/text2token.py \ No newline at end of file diff --git a/egs/mdcc/ASR/prepare.sh b/egs/mdcc/ASR/prepare.sh new file mode 100755 index 000000000..f4d9bc47e --- /dev/null +++ b/egs/mdcc/ASR/prepare.sh @@ -0,0 +1,308 @@ +#!/usr/bin/env bash + +# fix segmentation fault reported in https://github.com/k2-fsa/icefall/issues/674 +export PROTOCOL_BUFFERS_PYTHON_IMPLEMENTATION=python + +set -eou pipefail + +stage=-1 +stop_stage=100 +perturb_speed=true + + +# We assume dl_dir (download dir) contains the following +# directories and files. If not, they will be downloaded +# by this script automatically. +# +# - $dl_dir/mdcc +# |-- README.md +# |-- audio/ +# |-- clip_info_rthk.csv +# |-- cnt_asr_metadata_full.csv +# |-- cnt_asr_test_metadata.csv +# |-- cnt_asr_train_metadata.csv +# |-- cnt_asr_valid_metadata.csv +# |-- data_statistic.py +# |-- length +# |-- podcast_447_2021.csv +# |-- test.txt +# |-- transcription/ +# `-- words_length +# You can download them from: +# https://drive.google.com/file/d/1epfYMMhXdBKA6nxPgUugb2Uj4DllSxkn/view?usp=drive_link +# +# - $dl_dir/musan +# This directory contains the following directories downloaded from +# http://www.openslr.org/17/ +# +# - music +# - noise +# - speech + +dl_dir=$PWD/download + +. shared/parse_options.sh || exit 1 + +# All files generated by this script are saved in "data". +# You can safely remove "data" and rerun this script to regenerate it. +mkdir -p data + +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]}) $*" +} + +log "dl_dir: $dl_dir" + +if [ $stage -le 0 ] && [ $stop_stage -ge 0 ]; then + log "stage 0: Download data" + + # If you have pre-downloaded it to /path/to/mdcc, + # you can create a symlink + # + # ln -sfv /path/to/mdcc $dl_dir/mdcc + # + # The directory structure is + # mdcc/ + # |-- README.md + # |-- audio/ + # |-- clip_info_rthk.csv + # |-- cnt_asr_metadata_full.csv + # |-- cnt_asr_test_metadata.csv + # |-- cnt_asr_train_metadata.csv + # |-- cnt_asr_valid_metadata.csv + # |-- data_statistic.py + # |-- length + # |-- podcast_447_2021.csv + # |-- test.txt + # |-- transcription/ + # `-- words_length + + if [ ! -d $dl_dir/mdcc/audio ]; then + lhotse download mdcc $dl_dir + + # this will download and unzip dataset.zip to $dl_dir/ + + mv $dl_dir/dataset $dl_dir/mdcc + fi + + # If you have pre-downloaded it to /path/to/musan, + # you can create a symlink + # + # ln -sfv /path/to/musan $dl_dir/musan + # + if [ ! -d $dl_dir/musan ]; then + lhotse download musan $dl_dir + fi +fi + +if [ $stage -le 1 ] && [ $stop_stage -ge 1 ]; then + log "Stage 1: Prepare MDCC manifest" + # We assume that you have downloaded the MDCC corpus + # to $dl_dir/mdcc + if [ ! -f data/manifests/.mdcc_manifests.done ]; then + log "Might take 40 minutes to traverse the directory." + mkdir -p data/manifests + lhotse prepare mdcc $dl_dir/mdcc data/manifests + touch data/manifests/.mdcc_manifests.done + fi +fi + +if [ $stage -le 2 ] && [ $stop_stage -ge 2 ]; then + log "Stage 2: Prepare musan manifest" + # We assume that you have downloaded the musan corpus + # to data/musan + if [ ! -f data/manifests/.musan_manifests.done ]; then + log "It may take 6 minutes" + mkdir -p data/manifests + lhotse prepare musan $dl_dir/musan data/manifests + touch data/manifests/.musan_manifests.done + fi +fi + +if [ $stage -le 3 ] && [ $stop_stage -ge 3 ]; then + log "Stage 3: Compute fbank for MDCC" + if [ ! -f data/fbank/.mdcc.done ]; then + mkdir -p data/fbank + ./local/compute_fbank_mdcc.py --perturb-speed ${perturb_speed} + touch data/fbank/.mdcc.done + fi +fi + +if [ $stage -le 4 ] && [ $stop_stage -ge 4 ]; then + log "Stage 4: Compute fbank for musan" + if [ ! -f data/fbank/.msuan.done ]; then + mkdir -p data/fbank + ./local/compute_fbank_musan.py + touch data/fbank/.msuan.done + fi +fi + +lang_char_dir=data/lang_char +if [ $stage -le 5 ] && [ $stop_stage -ge 5 ]; then + log "Stage 5: Prepare char based lang" + mkdir -p $lang_char_dir + + # Prepare text. + # Note: in Linux, you can install jq with the following command: + # 1. wget -O jq https://github.com/stedolan/jq/releases/download/jq-1.6/jq-linux64 + # 2. chmod +x ./jq + # 3. cp jq /usr/bin + if [ ! -f $lang_char_dir/text ]; then + gunzip -c data/manifests/mdcc_supervisions_train.jsonl.gz \ + |jq '.text' | sed 's/"//g' | ./local/text2token.py -t "char" \ + > $lang_char_dir/train_text + + cat $lang_char_dir/train_text > $lang_char_dir/text + + gunzip -c data/manifests/mdcc_supervisions_test.jsonl.gz \ + |jq '.text' | sed 's/"//g' | ./local/text2token.py -t "char" \ + > $lang_char_dir/valid_text + + cat $lang_char_dir/valid_text >> $lang_char_dir/text + + gunzip -c data/manifests/mdcc_supervisions_valid.jsonl.gz \ + |jq '.text' | sed 's/"//g' | ./local/text2token.py -t "char" \ + > $lang_char_dir/test_text + + cat $lang_char_dir/test_text >> $lang_char_dir/text + fi + + if [ ! -f $lang_char_dir/text_words_segmentation ]; then + ./local/preprocess_mdcc.py --input-file $lang_char_dir/text \ + --output-dir $lang_char_dir + + mv $lang_char_dir/text $lang_char_dir/_text + cp $lang_char_dir/text_words_segmentation $lang_char_dir/text + fi + + if [ ! -f $lang_char_dir/tokens.txt ]; then + ./local/prepare_char.py --lang-dir $lang_char_dir + fi +fi + +if [ $stage -le 6 ] && [ $stop_stage -ge 6 ]; then + log "Stage 6: Prepare G" + + mkdir -p data/lm + + # Train LM on transcripts + if [ ! -f data/lm/3-gram.unpruned.arpa ]; then + python3 ./shared/make_kn_lm.py \ + -ngram-order 3 \ + -text $lang_char_dir/text_words_segmentation \ + -lm data/lm/3-gram.unpruned.arpa + fi + + # We assume you have installed kaldilm, if not, please install + # it using: pip install kaldilm + if [ ! -f data/lm/G_3_gram_char.fst.txt ]; then + # It is used in building HLG + python3 -m kaldilm \ + --read-symbol-table="$lang_char_dir/words.txt" \ + --disambig-symbol='#0' \ + --max-order=3 \ + data/lm/3-gram.unpruned.arpa > data/lm/G_3_gram_char.fst.txt + fi + + if [ ! -f $lang_char_dir/HLG.fst ]; then + ./local/prepare_lang_fst.py \ + --lang-dir $lang_char_dir \ + --ngram-G ./data/lm/G_3_gram_char.fst.txt + fi +fi + +if [ $stage -le 7 ] && [ $stop_stage -ge 7 ]; then + log "Stage 7: Compile LG & HLG" + + ./local/compile_hlg.py --lang-dir $lang_char_dir --lm G_3_gram_char + ./local/compile_lg.py --lang-dir $lang_char_dir --lm G_3_gram_char +fi + +if [ $stage -le 8 ] && [ $stop_stage -ge 8 ]; then + log "Stage 8: Generate LM training data" + + log "Processing char based data" + out_dir=data/lm_training_char + mkdir -p $out_dir $dl_dir/lm + + if [ ! -f $dl_dir/lm/mdcc-train-word.txt ]; then + ./local/text2segments.py --input-file $lang_char_dir/train_text \ + --output-file $dl_dir/lm/mdcc-train-word.txt + fi + + # training words + ./local/prepare_char_lm_training_data.py \ + --lang-char data/lang_char \ + --lm-data $dl_dir/lm/mdcc-train-word.txt \ + --lm-archive $out_dir/lm_data.pt + + # valid words + if [ ! -f $dl_dir/lm/mdcc-valid-word.txt ]; then + ./local/text2segments.py --input-file $lang_char_dir/valid_text \ + --output-file $dl_dir/lm/mdcc-valid-word.txt + fi + + ./local/prepare_char_lm_training_data.py \ + --lang-char data/lang_char \ + --lm-data $dl_dir/lm/mdcc-valid-word.txt \ + --lm-archive $out_dir/lm_data_valid.pt + + # test words + if [ ! -f $dl_dir/lm/mdcc-test-word.txt ]; then + ./local/text2segments.py --input-file $lang_char_dir/test_text \ + --output-file $dl_dir/lm/mdcc-test-word.txt + fi + + ./local/prepare_char_lm_training_data.py \ + --lang-char data/lang_char \ + --lm-data $dl_dir/lm/mdcc-test-word.txt \ + --lm-archive $out_dir/lm_data_test.pt +fi + +if [ $stage -le 9 ] && [ $stop_stage -ge 9 ]; then + log "Stage 9: Sort LM training data" + # Sort LM training data by sentence length in descending order + # for ease of training. + # + # Sentence length equals to the number of tokens + # in a sentence. + + out_dir=data/lm_training_char + mkdir -p $out_dir + ln -snf ../../../librispeech/ASR/local/sort_lm_training_data.py local/ + + ./local/sort_lm_training_data.py \ + --in-lm-data $out_dir/lm_data.pt \ + --out-lm-data $out_dir/sorted_lm_data.pt \ + --out-statistics $out_dir/statistics.txt + + ./local/sort_lm_training_data.py \ + --in-lm-data $out_dir/lm_data_valid.pt \ + --out-lm-data $out_dir/sorted_lm_data-valid.pt \ + --out-statistics $out_dir/statistics-valid.txt + + ./local/sort_lm_training_data.py \ + --in-lm-data $out_dir/lm_data_test.pt \ + --out-lm-data $out_dir/sorted_lm_data-test.pt \ + --out-statistics $out_dir/statistics-test.txt +fi + +if [ $stage -le 12 ] && [ $stop_stage -ge 12 ]; then + log "Stage 12: Train RNN LM model" + python ../../../icefall/rnn_lm/train.py \ + --start-epoch 0 \ + --world-size 1 \ + --num-epochs 20 \ + --use-fp16 0 \ + --embedding-dim 512 \ + --hidden-dim 512 \ + --num-layers 2 \ + --batch-size 400 \ + --exp-dir rnnlm_char/exp \ + --lm-data $out_dir/sorted_lm_data.pt \ + --lm-data-valid $out_dir/sorted_lm_data-valid.pt \ + --vocab-size 4336 \ + --master-port 12345 +fi diff --git a/egs/mdcc/ASR/shared b/egs/mdcc/ASR/shared new file mode 120000 index 000000000..4c5e91438 --- /dev/null +++ b/egs/mdcc/ASR/shared @@ -0,0 +1 @@ +../../../icefall/shared/ \ No newline at end of file diff --git a/egs/mdcc/ASR/zipformer/__init__.py b/egs/mdcc/ASR/zipformer/__init__.py new file mode 100644 index 000000000..e69de29bb diff --git a/egs/mdcc/ASR/zipformer/asr_datamodule.py b/egs/mdcc/ASR/zipformer/asr_datamodule.py new file mode 100644 index 000000000..1f49b6520 --- /dev/null +++ b/egs/mdcc/ASR/zipformer/asr_datamodule.py @@ -0,0 +1,382 @@ +# Copyright 2021 Piotr Żelasko +# Copyright 2022 Xiaomi Corporation (Author: Mingshuang Luo) +# Copyright 2024 Xiaomi Corporation (Author: Zengrui Jin) +# +# 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. + + +import argparse +import inspect +import logging +from functools import lru_cache +from pathlib import Path +from typing import Any, Dict, List, Optional + +from lhotse import CutSet, Fbank, FbankConfig, load_manifest, load_manifest_lazy +from lhotse.dataset import ( + CutConcatenate, + CutMix, + DynamicBucketingSampler, + K2SpeechRecognitionDataset, + PrecomputedFeatures, + SimpleCutSampler, + SpecAugment, +) +from lhotse.dataset.input_strategies import OnTheFlyFeatures +from torch.utils.data import DataLoader + +from icefall.utils import str2bool + + +class MdccAsrDataModule: + """ + DataModule for k2 ASR experiments. + It assumes there is always one train and valid dataloader, + but there can be multiple test dataloaders (e.g. LibriSpeech test-clean + and test-other). + + It contains all the common data pipeline modules used in ASR + experiments, e.g.: + - dynamic batch size, + - bucketing samplers, + - cut concatenation, + - augmentation, + - on-the-fly feature extraction + + This class should be derived for specific corpora used in ASR tasks. + """ + + def __init__(self, args: argparse.Namespace): + self.args = args + + @classmethod + def add_arguments(cls, parser: argparse.ArgumentParser): + group = parser.add_argument_group( + title="ASR data related options", + description="These options are used for the preparation of " + "PyTorch DataLoaders from Lhotse CutSet's -- they control the " + "effective batch sizes, sampling strategies, applied data " + "augmentations, etc.", + ) + group.add_argument( + "--manifest-dir", + type=Path, + default=Path("data/fbank"), + help="Path to directory with train/valid/test cuts.", + ) + group.add_argument( + "--max-duration", + type=int, + default=200.0, + help="Maximum pooled recordings duration (seconds) in a " + "single batch. You can reduce it if it causes CUDA OOM.", + ) + group.add_argument( + "--bucketing-sampler", + type=str2bool, + default=True, + help="When enabled, the batches will come from buckets of " + "similar duration (saves padding frames).", + ) + group.add_argument( + "--num-buckets", + type=int, + default=30, + help="The number of buckets for the DynamicBucketingSampler" + "(you might want to increase it for larger datasets).", + ) + group.add_argument( + "--concatenate-cuts", + type=str2bool, + default=False, + help="When enabled, utterances (cuts) will be concatenated " + "to minimize the amount of padding.", + ) + group.add_argument( + "--duration-factor", + type=float, + default=1.0, + help="Determines the maximum duration of a concatenated cut " + "relative to the duration of the longest cut in a batch.", + ) + group.add_argument( + "--gap", + type=float, + default=1.0, + help="The amount of padding (in seconds) inserted between " + "concatenated cuts. This padding is filled with noise when " + "noise augmentation is used.", + ) + group.add_argument( + "--on-the-fly-feats", + type=str2bool, + default=False, + help="When enabled, use on-the-fly cut mixing and feature " + "extraction. Will drop existing precomputed feature manifests " + "if available.", + ) + group.add_argument( + "--shuffle", + type=str2bool, + default=True, + help="When enabled (=default), the examples will be " + "shuffled for each epoch.", + ) + group.add_argument( + "--drop-last", + type=str2bool, + default=True, + help="Whether to drop last batch. Used by sampler.", + ) + group.add_argument( + "--return-cuts", + type=str2bool, + default=True, + help="When enabled, each batch will have the " + "field: batch['supervisions']['cut'] with the cuts that " + "were used to construct it.", + ) + + group.add_argument( + "--num-workers", + type=int, + default=2, + help="The number of training dataloader workers that " + "collect the batches.", + ) + + group.add_argument( + "--enable-spec-aug", + type=str2bool, + default=True, + help="When enabled, use SpecAugment for training dataset.", + ) + + group.add_argument( + "--spec-aug-time-warp-factor", + type=int, + default=80, + help="Used only when --enable-spec-aug is True. " + "It specifies the factor for time warping in SpecAugment. " + "Larger values mean more warping. " + "A value less than 1 means to disable time warp.", + ) + + group.add_argument( + "--enable-musan", + type=str2bool, + default=True, + help="When enabled, select noise from MUSAN and mix it" + "with training dataset. ", + ) + + def train_dataloaders( + self, cuts_train: CutSet, sampler_state_dict: Optional[Dict[str, Any]] = None + ) -> DataLoader: + """ + Args: + cuts_train: + CutSet for training. + sampler_state_dict: + The state dict for the training sampler. + """ + logging.info("About to get Musan cuts") + cuts_musan = load_manifest(self.args.manifest_dir / "musan_cuts.jsonl.gz") + + transforms = [] + if self.args.enable_musan: + logging.info("Enable MUSAN") + transforms.append( + CutMix(cuts=cuts_musan, p=0.5, snr=(10, 20), preserve_id=True) + ) + else: + logging.info("Disable MUSAN") + + if self.args.concatenate_cuts: + logging.info( + f"Using cut concatenation with duration factor " + f"{self.args.duration_factor} and gap {self.args.gap}." + ) + # Cut concatenation should be the first transform in the list, + # so that if we e.g. mix noise in, it will fill the gaps between + # different utterances. + transforms = [ + CutConcatenate( + duration_factor=self.args.duration_factor, gap=self.args.gap + ) + ] + transforms + + input_transforms = [] + if self.args.enable_spec_aug: + logging.info("Enable SpecAugment") + logging.info(f"Time warp factor: {self.args.spec_aug_time_warp_factor}") + # Set the value of num_frame_masks according to Lhotse's version. + # In different Lhotse's versions, the default of num_frame_masks is + # different. + num_frame_masks = 10 + num_frame_masks_parameter = inspect.signature( + SpecAugment.__init__ + ).parameters["num_frame_masks"] + if num_frame_masks_parameter.default == 1: + num_frame_masks = 2 + logging.info(f"Num frame mask: {num_frame_masks}") + input_transforms.append( + SpecAugment( + time_warp_factor=self.args.spec_aug_time_warp_factor, + num_frame_masks=num_frame_masks, + features_mask_size=27, + num_feature_masks=2, + frames_mask_size=100, + ) + ) + else: + logging.info("Disable SpecAugment") + + logging.info("About to create train dataset") + train = K2SpeechRecognitionDataset( + cut_transforms=transforms, + input_transforms=input_transforms, + return_cuts=self.args.return_cuts, + ) + + if self.args.on_the_fly_feats: + # NOTE: the PerturbSpeed transform should be added only if we + # remove it from data prep stage. + # Add on-the-fly speed perturbation; since originally it would + # have increased epoch size by 3, we will apply prob 2/3 and use + # 3x more epochs. + # Speed perturbation probably should come first before + # concatenation, but in principle the transforms order doesn't have + # to be strict (e.g. could be randomized) + # transforms = [PerturbSpeed(factors=[0.9, 1.1], p=2/3)] + transforms # noqa + # Drop feats to be on the safe side. + train = K2SpeechRecognitionDataset( + cut_transforms=transforms, + input_strategy=OnTheFlyFeatures(Fbank(FbankConfig(num_mel_bins=80))), + input_transforms=input_transforms, + return_cuts=self.args.return_cuts, + ) + + if self.args.bucketing_sampler: + logging.info("Using DynamicBucketingSampler.") + train_sampler = DynamicBucketingSampler( + cuts_train, + max_duration=self.args.max_duration, + shuffle=self.args.shuffle, + num_buckets=self.args.num_buckets, + buffer_size=self.args.num_buckets * 2000, + shuffle_buffer_size=self.args.num_buckets * 5000, + drop_last=self.args.drop_last, + ) + else: + logging.info("Using SimpleCutSampler.") + train_sampler = SimpleCutSampler( + cuts_train, + max_duration=self.args.max_duration, + shuffle=self.args.shuffle, + ) + logging.info("About to create train dataloader") + + if sampler_state_dict is not None: + logging.info("Loading sampler state dict") + train_sampler.load_state_dict(sampler_state_dict) + + train_dl = DataLoader( + train, + sampler=train_sampler, + batch_size=None, + num_workers=self.args.num_workers, + persistent_workers=False, + ) + + return train_dl + + def valid_dataloaders(self, cuts_valid: CutSet) -> DataLoader: + transforms = [] + if self.args.concatenate_cuts: + transforms = [ + CutConcatenate( + duration_factor=self.args.duration_factor, gap=self.args.gap + ) + ] + transforms + + logging.info("About to create dev dataset") + if self.args.on_the_fly_feats: + validate = K2SpeechRecognitionDataset( + cut_transforms=transforms, + input_strategy=OnTheFlyFeatures(Fbank(FbankConfig(num_mel_bins=80))), + return_cuts=self.args.return_cuts, + ) + else: + validate = K2SpeechRecognitionDataset( + cut_transforms=transforms, + return_cuts=self.args.return_cuts, + ) + valid_sampler = DynamicBucketingSampler( + cuts_valid, + max_duration=self.args.max_duration, + shuffle=False, + ) + logging.info("About to create dev dataloader") + valid_dl = DataLoader( + validate, + sampler=valid_sampler, + batch_size=None, + num_workers=2, + persistent_workers=False, + ) + + return valid_dl + + def test_dataloaders(self, cuts: CutSet) -> DataLoader: + logging.info("About to create test dataset") + test = K2SpeechRecognitionDataset( + input_strategy=( + OnTheFlyFeatures(Fbank(FbankConfig(num_mel_bins=80))) + if self.args.on_the_fly_feats + else PrecomputedFeatures() + ), + return_cuts=self.args.return_cuts, + ) + sampler = DynamicBucketingSampler( + cuts, + max_duration=self.args.max_duration, + shuffle=False, + ) + test_dl = DataLoader( + test, + batch_size=None, + sampler=sampler, + num_workers=self.args.num_workers, + ) + return test_dl + + @lru_cache() + def train_cuts(self) -> CutSet: + logging.info("About to get train cuts") + cuts_train = load_manifest_lazy( + self.args.manifest_dir / "mdcc_cuts_train.jsonl.gz" + ) + return cuts_train + + @lru_cache() + def valid_cuts(self) -> CutSet: + logging.info("About to get valid cuts") + return load_manifest_lazy(self.args.manifest_dir / "mdcc_cuts_valid.jsonl.gz") + + @lru_cache() + def test_cuts(self) -> List[CutSet]: + logging.info("About to get test cuts") + return load_manifest_lazy(self.args.manifest_dir / "mdcc_cuts_test.jsonl.gz") diff --git a/egs/mdcc/ASR/zipformer/beam_search.py b/egs/mdcc/ASR/zipformer/beam_search.py new file mode 120000 index 000000000..e24eca39f --- /dev/null +++ b/egs/mdcc/ASR/zipformer/beam_search.py @@ -0,0 +1 @@ +../../../librispeech/ASR/pruned_transducer_stateless2/beam_search.py \ No newline at end of file diff --git a/egs/mdcc/ASR/zipformer/decode.py b/egs/mdcc/ASR/zipformer/decode.py new file mode 100755 index 000000000..ce104baf7 --- /dev/null +++ b/egs/mdcc/ASR/zipformer/decode.py @@ -0,0 +1,813 @@ +#!/usr/bin/env python3 +# +# Copyright 2021-2024 Xiaomi Corporation (Author: Fangjun Kuang, +# Zengwei Yao, +# Mingshuang Luo, +# Zengrui Jin,) +# +# 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 +./zipformer/decode.py \ + --epoch 35 \ + --avg 15 \ + --exp-dir ./zipformer/exp \ + --lang-dir data/lang_char \ + --max-duration 600 \ + --decoding-method greedy_search + +(2) modified beam search +./zipformer/decode.py \ + --epoch 35 \ + --avg 15 \ + --exp-dir ./zipformer/exp \ + --lang-dir data/lang_char \ + --max-duration 600 \ + --decoding-method modified_beam_search \ + --beam-size 4 + +(3) fast beam search (trivial_graph) +./zipformer/decode.py \ + --epoch 35 \ + --avg 15 \ + --exp-dir ./zipformer/exp \ + --lang-dir data/lang_char \ + --max-duration 600 \ + --decoding-method fast_beam_search \ + --beam 20.0 \ + --max-contexts 8 \ + --max-states 64 + +(4) fast beam search (LG) +./zipformer/decode.py \ + --epoch 30 \ + --avg 15 \ + --exp-dir ./zipformer/exp \ + --lang-dir data/lang_char \ + --max-duration 600 \ + --decoding-method fast_beam_search_LG \ + --beam 20.0 \ + --max-contexts 8 \ + --max-states 64 + +(5) fast beam search (nbest oracle WER) +./zipformer/decode.py \ + --epoch 35 \ + --avg 15 \ + --exp-dir ./zipformer/exp \ + --lang-dir data/lang_char \ + --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 +""" + + +import argparse +import logging +import math +from collections import defaultdict +from pathlib import Path +from typing import Dict, List, Optional, Tuple + +import k2 +import torch +import torch.nn as nn +from asr_datamodule import MdccAsrDataModule +from beam_search import ( + beam_search, + fast_beam_search_nbest_oracle, + fast_beam_search_one_best, + greedy_search, + greedy_search_batch, + modified_beam_search, +) +from lhotse.cut import Cut +from train import add_model_arguments, get_model, get_params + +from icefall.char_graph_compiler import CharCtcTrainingGraphCompiler +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, + make_pad_mask, + 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="zipformer/exp", + help="The experiment dir", + ) + + parser.add_argument( + "--lang-dir", + type=Path, + default="data/lang_char", + 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 + - modified_beam_search + - fast_beam_search + - fast_beam_search_LG + - fast_beam_search_nbest_oracle + If you use fast_beam_search_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, fast_beam_search_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_LG. + It specifies the scale for n-gram LM scores. + """, + ) + + parser.add_argument( + "--ilme-scale", + type=float, + default=0.2, + help=""" + Used only when --decoding_method is fast_beam_search_LG. + It specifies the scale for the internal language model estimation. + """, + ) + + parser.add_argument( + "--max-contexts", + type=int, + default=8, + help="""Used only when --decoding-method is + fast_beam_search, fast_beam_search, fast_beam_search_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, fast_beam_search_LG, + and fast_beam_search_nbest_oracle""", + ) + + parser.add_argument( + "--context-size", + type=int, + default=1, + 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_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 and fast_beam_search_nbest_oracle""", + ) + + parser.add_argument( + "--blank-penalty", + type=float, + default=0.0, + help=""" + The penalty applied on blank symbol during decoding. + Note: It is a positive value that would be applied to logits like + this `logits[:, 0] -= blank_penalty` (suppose logits.shape is + [batch_size, vocab] and blank id is 0). + """, + ) + + add_model_arguments(parser) + + return parser + + +def decode_one_batch( + params: AttributeDict, + model: nn.Module, + lexicon: Lexicon, + graph_compiler: CharCtcTrainingGraphCompiler, + 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. + 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 LG, 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.causal: + # this seems to cause insertions at the end of the utterance if used with zipformer. + pad_len = 30 + feature_lens += pad_len + feature = torch.nn.functional.pad( + feature, + pad=(0, 0, 0, pad_len), + value=LOG_EPS, + ) + + x, x_lens = model.encoder_embed(feature, feature_lens) + + src_key_padding_mask = make_pad_mask(x_lens) + x = x.permute(1, 0, 2) # (N, T, C) -> (T, N, C) + + encoder_out, encoder_out_lens = model.encoder(x, x_lens, src_key_padding_mask) + encoder_out = encoder_out.permute(1, 0, 2) # (T, N, C) ->(N, T, C) + + 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, + blank_penalty=params.blank_penalty, + ) + for i in range(encoder_out.size(0)): + hyps.append([lexicon.token_table[idx] for idx in hyp_tokens[i]]) + elif params.decoding_method == "fast_beam_search_LG": + 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, + blank_penalty=params.blank_penalty, + ilme_scale=params.ilme_scale, + ) + for hyp in hyp_tokens: + sentence = "".join([lexicon.word_table[i] for i in hyp]) + hyps.append(list(sentence)) + 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=graph_compiler.texts_to_ids(supervisions["text"]), + nbest_scale=params.nbest_scale, + blank_penalty=params.blank_penalty, + ) + for i in range(encoder_out.size(0)): + hyps.append([lexicon.token_table[idx] for idx in hyp_tokens[i]]) + 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, + blank_penalty=params.blank_penalty, + ) + for i in range(encoder_out.size(0)): + hyps.append([lexicon.token_table[idx] for idx in hyp_tokens[i]]) + elif params.decoding_method == "modified_beam_search": + hyp_tokens = modified_beam_search( + model=model, + encoder_out=encoder_out, + encoder_out_lens=encoder_out_lens, + blank_penalty=params.blank_penalty, + beam=params.beam_size, + ) + for i in range(encoder_out.size(0)): + hyps.append([lexicon.token_table[idx] for idx in hyp_tokens[i]]) + 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, + blank_penalty=params.blank_penalty, + ) + elif params.decoding_method == "beam_search": + hyp = beam_search( + model=model, + encoder_out=encoder_out_i, + beam=params.beam_size, + blank_penalty=params.blank_penalty, + ) + else: + raise ValueError( + f"Unsupported decoding method: {params.decoding_method}" + ) + hyps.append([lexicon.token_table[idx] for idx in hyp]) + + key = f"blank_penalty_{params.blank_penalty}" + if params.decoding_method == "greedy_search": + return {"greedy_search_" + key: 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"_ilme_scale_{params.ilme_scale}" + key += f"_ngram_lm_scale_{params.ngram_lm_scale}" + + return {key: hyps} + else: + return {f"beam_size_{params.beam_size}_" + key: hyps} + + +def decode_dataset( + dl: torch.utils.data.DataLoader, + params: AttributeDict, + model: nn.Module, + lexicon: Lexicon, + graph_compiler: CharCtcTrainingGraphCompiler, + 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. + decoding_graph: + The decoding graph. Can be either a `k2.trivial_graph` or LG, 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"] + texts = [list("".join(text.split())) for text in texts] + cut_ids = [cut.id for cut in batch["supervisions"]["cut"]] + + hyps_dict = decode_one_batch( + params=params, + model=model, + lexicon=lexicon, + graph_compiler=graph_compiler, + decoding_graph=decoding_graph, + 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): + this_batch.append((cut_id, ref_text, 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" + ) + 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() + MdccAsrDataModule.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", + "modified_beam_search", + "fast_beam_search", + "fast_beam_search_LG", + "fast_beam_search_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.causal: + assert ( + "," not in params.chunk_size + ), "chunk_size should be one value in decoding." + assert ( + "," not in params.left_context_frames + ), "left_context_frames should be one value in decoding." + params.suffix += f"-chunk-{params.chunk_size}" + params.suffix += f"-left-context-{params.left_context_frames}" + + 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"_ilme_scale_{params.ilme_scale}" + 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}" + params.suffix += f"-blank-penalty-{params.blank_penalty}" + + 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}") + + lexicon = Lexicon(params.lang_dir) + params.blank_id = lexicon.token_table[""] + params.vocab_size = max(lexicon.tokens) + 1 + + graph_compiler = CharCtcTrainingGraphCompiler( + lexicon=lexicon, + device=device, + ) + + logging.info(params) + + logging.info("About to create model") + model = get_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 "LG" in params.decoding_method: + lexicon = Lexicon(params.lang_dir) + 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: + 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}") + + # we need cut ids to display recognition results. + args.return_cuts = True + mdcc = MdccAsrDataModule(args) + + def remove_short_utt(c: Cut): + T = ((c.num_frames - 7) // 2 + 1) // 2 + if T <= 0: + logging.warning( + f"Exclude cut with ID {c.id} from decoding, num_frames : {c.num_frames}." + ) + return T > 0 + + valid_cuts = mdcc.valid_cuts() + valid_cuts = valid_cuts.filter(remove_short_utt) + valid_dl = mdcc.valid_dataloaders(valid_cuts) + + test_cuts = mdcc.test_cuts() + test_cuts = test_cuts.filter(remove_short_utt) + test_dl = mdcc.test_dataloaders(test_cuts) + + test_sets = ["valid", "test"] + test_dls = [valid_dl, test_dl] + + for test_set, test_dl in zip(test_sets, test_dls): + results_dict = decode_dataset( + dl=test_dl, + params=params, + model=model, + lexicon=lexicon, + graph_compiler=graph_compiler, + decoding_graph=decoding_graph, + ) + + save_results( + params=params, + test_set_name=test_set, + results_dict=results_dict, + ) + + logging.info("Done!") + + +if __name__ == "__main__": + main() diff --git a/egs/mdcc/ASR/zipformer/decode_stream.py b/egs/mdcc/ASR/zipformer/decode_stream.py new file mode 120000 index 000000000..b8d8ddfc4 --- /dev/null +++ b/egs/mdcc/ASR/zipformer/decode_stream.py @@ -0,0 +1 @@ +../../../librispeech/ASR/zipformer/decode_stream.py \ No newline at end of file diff --git a/egs/mdcc/ASR/zipformer/decoder.py b/egs/mdcc/ASR/zipformer/decoder.py new file mode 120000 index 000000000..5a8018680 --- /dev/null +++ b/egs/mdcc/ASR/zipformer/decoder.py @@ -0,0 +1 @@ +../../../librispeech/ASR/zipformer/decoder.py \ No newline at end of file diff --git a/egs/mdcc/ASR/zipformer/encoder_interface.py b/egs/mdcc/ASR/zipformer/encoder_interface.py new file mode 120000 index 000000000..c2eaca671 --- /dev/null +++ b/egs/mdcc/ASR/zipformer/encoder_interface.py @@ -0,0 +1 @@ +../../../librispeech/ASR/zipformer/encoder_interface.py \ No newline at end of file diff --git a/egs/mdcc/ASR/zipformer/export-onnx-ctc.py b/egs/mdcc/ASR/zipformer/export-onnx-ctc.py new file mode 120000 index 000000000..f9d756352 --- /dev/null +++ b/egs/mdcc/ASR/zipformer/export-onnx-ctc.py @@ -0,0 +1 @@ +../../../librispeech/ASR/zipformer/export-onnx-ctc.py \ No newline at end of file diff --git a/egs/mdcc/ASR/zipformer/export-onnx-streaming-ctc.py b/egs/mdcc/ASR/zipformer/export-onnx-streaming-ctc.py new file mode 120000 index 000000000..652346001 --- /dev/null +++ b/egs/mdcc/ASR/zipformer/export-onnx-streaming-ctc.py @@ -0,0 +1 @@ +../../../librispeech/ASR/zipformer/export-onnx-streaming-ctc.py \ No newline at end of file diff --git a/egs/mdcc/ASR/zipformer/export-onnx-streaming.py b/egs/mdcc/ASR/zipformer/export-onnx-streaming.py new file mode 120000 index 000000000..2962eb784 --- /dev/null +++ b/egs/mdcc/ASR/zipformer/export-onnx-streaming.py @@ -0,0 +1 @@ +../../../librispeech/ASR/zipformer/export-onnx-streaming.py \ No newline at end of file diff --git a/egs/mdcc/ASR/zipformer/export-onnx.py b/egs/mdcc/ASR/zipformer/export-onnx.py new file mode 120000 index 000000000..70a15683c --- /dev/null +++ b/egs/mdcc/ASR/zipformer/export-onnx.py @@ -0,0 +1 @@ +../../../librispeech/ASR/zipformer/export-onnx.py \ No newline at end of file diff --git a/egs/mdcc/ASR/zipformer/export.py b/egs/mdcc/ASR/zipformer/export.py new file mode 120000 index 000000000..dfc1bec08 --- /dev/null +++ b/egs/mdcc/ASR/zipformer/export.py @@ -0,0 +1 @@ +../../../librispeech/ASR/zipformer/export.py \ No newline at end of file diff --git a/egs/mdcc/ASR/zipformer/joiner.py b/egs/mdcc/ASR/zipformer/joiner.py new file mode 120000 index 000000000..5b8a36332 --- /dev/null +++ b/egs/mdcc/ASR/zipformer/joiner.py @@ -0,0 +1 @@ +../../../librispeech/ASR/zipformer/joiner.py \ No newline at end of file diff --git a/egs/mdcc/ASR/zipformer/model.py b/egs/mdcc/ASR/zipformer/model.py new file mode 120000 index 000000000..cd7e07d72 --- /dev/null +++ b/egs/mdcc/ASR/zipformer/model.py @@ -0,0 +1 @@ +../../../librispeech/ASR/zipformer/model.py \ No newline at end of file diff --git a/egs/mdcc/ASR/zipformer/onnx_check.py b/egs/mdcc/ASR/zipformer/onnx_check.py new file mode 120000 index 000000000..f3dd42004 --- /dev/null +++ b/egs/mdcc/ASR/zipformer/onnx_check.py @@ -0,0 +1 @@ +../../../librispeech/ASR/zipformer/onnx_check.py \ No newline at end of file diff --git a/egs/mdcc/ASR/zipformer/onnx_decode.py b/egs/mdcc/ASR/zipformer/onnx_decode.py new file mode 100755 index 000000000..1ed4a9fa1 --- /dev/null +++ b/egs/mdcc/ASR/zipformer/onnx_decode.py @@ -0,0 +1,286 @@ +#!/usr/bin/env python3 +# +# Copyright 2021-2023 Xiaomi Corporation (Author: Fangjun Kuang, +# Zengwei Yao, +# Xiaoyu Yang, +# 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. +""" +This script loads ONNX exported models and uses them to decode the test sets. +""" + +import argparse +import logging +import time +from pathlib import Path +from typing import List, Tuple + +import k2 +import torch +import torch.nn as nn +from asr_datamodule import MdccAsrDataModule +from lhotse.cut import Cut +from onnx_pretrained import OnnxModel, greedy_search + +from icefall.utils import setup_logger, store_transcripts, write_error_stats + + +def get_parser(): + parser = argparse.ArgumentParser( + formatter_class=argparse.ArgumentDefaultsHelpFormatter + ) + + parser.add_argument( + "--encoder-model-filename", + type=str, + required=True, + help="Path to the encoder onnx model. ", + ) + + parser.add_argument( + "--decoder-model-filename", + type=str, + required=True, + help="Path to the decoder onnx model. ", + ) + + parser.add_argument( + "--joiner-model-filename", + type=str, + required=True, + help="Path to the joiner onnx model. ", + ) + + parser.add_argument( + "--exp-dir", + type=str, + default="pruned_transducer_stateless7/exp", + help="The experiment dir", + ) + + parser.add_argument( + "--tokens", + type=str, + default="data/lang_char/tokens.txt", + help="Path to the tokens.txt", + ) + + parser.add_argument( + "--decoding-method", + type=str, + default="greedy_search", + help="Valid values are greedy_search and modified_beam_search", + ) + + return parser + + +def decode_one_batch( + model: OnnxModel, token_table: k2.SymbolTable, batch: dict +) -> List[List[str]]: + """Decode one batch and return the result. + Currently it only greedy_search is supported. + + Args: + model: + The neural model. + token_table: + Mapping ids to tokens. + batch: + It is the return value from iterating + `lhotse.dataset.K2SpeechRecognitionDataset`. See its documentation + for the format of the `batch`. + + Returns: + Return the decoded results for each utterance. + """ + feature = batch["inputs"] + assert feature.ndim == 3 + # at entry, feature is (N, T, C) + + supervisions = batch["supervisions"] + feature_lens = supervisions["num_frames"].to(dtype=torch.int64) + + encoder_out, encoder_out_lens = model.run_encoder(x=feature, x_lens=feature_lens) + + hyps = greedy_search( + model=model, encoder_out=encoder_out, encoder_out_lens=encoder_out_lens + ) + + hyps = [[token_table[h] for h in hyp] for hyp in hyps] + return hyps + + +def decode_dataset( + dl: torch.utils.data.DataLoader, + model: nn.Module, + token_table: k2.SymbolTable, +) -> Tuple[List[Tuple[str, List[str], List[str]]], float]: + """Decode dataset. + + Args: + dl: + PyTorch's dataloader containing the dataset to decode. + model: + The neural model. + token_table: + Mapping ids to tokens. + + Returns: + - A list of tuples. Each tuple contains three elements: + - cut_id, + - reference transcript, + - predicted result. + - The total duration (in seconds) of the dataset. + """ + num_cuts = 0 + + try: + num_batches = len(dl) + except TypeError: + num_batches = "?" + + log_interval = 10 + total_duration = 0 + + results = [] + for batch_idx, batch in enumerate(dl): + texts = batch["supervisions"]["text"] + cut_ids = [cut.id for cut in batch["supervisions"]["cut"]] + total_duration += sum([cut.duration for cut in batch["supervisions"]["cut"]]) + + hyps = decode_one_batch(model=model, token_table=token_table, batch=batch) + + this_batch = [] + assert len(hyps) == len(texts) + for cut_id, hyp_words, ref_text in zip(cut_ids, hyps, texts): + ref_words = list(ref_text) + this_batch.append((cut_id, ref_words, hyp_words)) + + results.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, total_duration + + +def save_results( + res_dir: Path, + test_set_name: str, + results: List[Tuple[str, List[str], List[str]]], +): + recog_path = res_dir / f"recogs-{test_set_name}.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 = res_dir / f"errs-{test_set_name}.txt" + with open(errs_filename, "w") as f: + wer = write_error_stats(f, f"{test_set_name}", results, enable_log=True) + + logging.info("Wrote detailed error stats to {}".format(errs_filename)) + + errs_info = res_dir / f"wer-summary-{test_set_name}.txt" + with open(errs_info, "w") as f: + print("WER", file=f) + print(wer, file=f) + + s = "\nFor {}, WER is {}:\n".format(test_set_name, wer) + logging.info(s) + + +@torch.no_grad() +def main(): + parser = get_parser() + MdccAsrDataModule.add_arguments(parser) + args = parser.parse_args() + + assert ( + args.decoding_method == "greedy_search" + ), "Only supports greedy_search currently." + res_dir = Path(args.exp_dir) / f"onnx-{args.decoding_method}" + + setup_logger(f"{res_dir}/log-decode") + logging.info("Decoding started") + + device = torch.device("cpu") + logging.info(f"Device: {device}") + + token_table = k2.SymbolTable.from_file(args.tokens) + assert token_table[0] == "" + + logging.info(vars(args)) + + logging.info("About to create model") + model = OnnxModel( + encoder_model_filename=args.encoder_model_filename, + decoder_model_filename=args.decoder_model_filename, + joiner_model_filename=args.joiner_model_filename, + ) + + # we need cut ids to display recognition results. + args.return_cuts = True + + mdcc = MdccAsrDataModule(args) + + def remove_short_utt(c: Cut): + T = ((c.num_frames - 7) // 2 + 1) // 2 + if T <= 0: + logging.warning( + f"Exclude cut with ID {c.id} from decoding, num_frames : {c.num_frames}." + ) + return T > 0 + + valid_cuts = mdcc.valid_cuts() + valid_cuts = valid_cuts.filter(remove_short_utt) + valid_dl = mdcc.valid_dataloaders(valid_cuts) + + test_cuts = mdcc.test_net_cuts() + test_cuts = test_cuts.filter(remove_short_utt) + test_dl = mdcc.test_dataloaders(test_cuts) + + test_sets = ["valid", "test"] + test_dl = [valid_dl, test_dl] + + for test_set, test_dl in zip(test_sets, test_dl): + start_time = time.time() + results, total_duration = decode_dataset( + dl=test_dl, model=model, token_table=token_table + ) + end_time = time.time() + elapsed_seconds = end_time - start_time + rtf = elapsed_seconds / total_duration + + logging.info(f"Elapsed time: {elapsed_seconds:.3f} s") + logging.info(f"Wave duration: {total_duration:.3f} s") + logging.info( + f"Real time factor (RTF): {elapsed_seconds:.3f}/{total_duration:.3f} = {rtf:.3f}" + ) + + save_results(res_dir=res_dir, test_set_name=test_set, results=results) + + logging.info("Done!") + + +if __name__ == "__main__": + main() diff --git a/egs/mdcc/ASR/zipformer/optim.py b/egs/mdcc/ASR/zipformer/optim.py new file mode 120000 index 000000000..5eaa3cffd --- /dev/null +++ b/egs/mdcc/ASR/zipformer/optim.py @@ -0,0 +1 @@ +../../../librispeech/ASR/zipformer/optim.py \ No newline at end of file diff --git a/egs/mdcc/ASR/zipformer/scaling.py b/egs/mdcc/ASR/zipformer/scaling.py new file mode 120000 index 000000000..6f398f431 --- /dev/null +++ b/egs/mdcc/ASR/zipformer/scaling.py @@ -0,0 +1 @@ +../../../librispeech/ASR/zipformer/scaling.py \ No newline at end of file diff --git a/egs/mdcc/ASR/zipformer/scaling_converter.py b/egs/mdcc/ASR/zipformer/scaling_converter.py new file mode 120000 index 000000000..b0ecee05e --- /dev/null +++ b/egs/mdcc/ASR/zipformer/scaling_converter.py @@ -0,0 +1 @@ +../../../librispeech/ASR/zipformer/scaling_converter.py \ No newline at end of file diff --git a/egs/mdcc/ASR/zipformer/streaming_beam_search.py b/egs/mdcc/ASR/zipformer/streaming_beam_search.py new file mode 120000 index 000000000..b1ed54557 --- /dev/null +++ b/egs/mdcc/ASR/zipformer/streaming_beam_search.py @@ -0,0 +1 @@ +../../../librispeech/ASR/zipformer/streaming_beam_search.py \ No newline at end of file diff --git a/egs/mdcc/ASR/zipformer/streaming_decode.py b/egs/mdcc/ASR/zipformer/streaming_decode.py new file mode 100755 index 000000000..dadb0b55f --- /dev/null +++ b/egs/mdcc/ASR/zipformer/streaming_decode.py @@ -0,0 +1,881 @@ +#!/usr/bin/env python3 +# Copyright 2022-2023 Xiaomi Corporation (Authors: Wei Kang, +# 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: +./zipformer/streaming_decode.py \ + --epoch 28 \ + --avg 15 \ + --causal 1 \ + --chunk-size 16 \ + --left-context-frames 256 \ + --exp-dir ./zipformer/exp \ + --decoding-method greedy_search \ + --num-decode-streams 2000 +""" + +import argparse +import logging +import math +from pathlib import Path +from typing import Dict, List, Optional, Tuple + +import k2 +import numpy as np +import torch +from asr_datamodule import MdccAsrDataModule +from decode_stream import DecodeStream +from kaldifeat import Fbank, FbankOptions +from lhotse import CutSet +from streaming_beam_search import ( + fast_beam_search_one_best, + greedy_search, + modified_beam_search, +) +from torch import Tensor, nn +from torch.nn.utils.rnn import pad_sequence +from train import add_model_arguments, get_model, get_params + +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, + make_pad_mask, + 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=28, + 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="zipformer/exp", + help="The experiment dir", + ) + + parser.add_argument( + "--lang-dir", + type=str, + default="data/lang_char", + help="Path to the lang dir(containing lexicon, tokens, etc.)", + ) + + parser.add_argument( + "--decoding-method", + type=str, + default="greedy_search", + help="""Supported decoding methods are: + greedy_search + modified_beam_search + fast_beam_search + """, + ) + + parser.add_argument( + "--num_active_paths", + type=int, + default=4, + help="""An interger indicating how many candidates we will keep for each + frame. Used only when --decoding-method is 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=32, + help="""Used only when --decoding-method is + fast_beam_search""", + ) + + parser.add_argument( + "--context-size", + type=int, + default=1, + help="The context size in the decoder. 1 means bigram; 2 means tri-gram", + ) + + parser.add_argument( + "--blank-penalty", + type=float, + default=0.0, + help=""" + The penalty applied on blank symbol during decoding. + Note: It is a positive value that would be applied to logits like + this `logits[:, 0] -= blank_penalty` (suppose logits.shape is + [batch_size, vocab] and blank id is 0). + """, + ) + + parser.add_argument( + "--num-decode-streams", + type=int, + default=2000, + help="The number of streams that can be decoded parallel.", + ) + + add_model_arguments(parser) + + return parser + + +def get_init_states( + model: nn.Module, + batch_size: int = 1, + device: torch.device = torch.device("cpu"), +) -> List[torch.Tensor]: + """ + Returns a list of cached tensors of all encoder layers. For layer-i, states[i*6:(i+1)*6] + is (cached_key, cached_nonlin_attn, cached_val1, cached_val2, cached_conv1, cached_conv2). + states[-2] is the cached left padding for ConvNeXt module, + of shape (batch_size, num_channels, left_pad, num_freqs) + states[-1] is processed_lens of shape (batch,), which records the number + of processed frames (at 50hz frame rate, after encoder_embed) for each sample in batch. + """ + states = model.encoder.get_init_states(batch_size, device) + + embed_states = model.encoder_embed.get_init_states(batch_size, device) + states.append(embed_states) + + processed_lens = torch.zeros(batch_size, dtype=torch.int32, device=device) + states.append(processed_lens) + + return states + + +def stack_states(state_list: List[List[torch.Tensor]]) -> List[torch.Tensor]: + """Stack list of zipformer states that correspond to separate utterances + into a single emformer state, so that it can be used as an input for + zipformer when those utterances are formed into a batch. + + Args: + state_list: + Each element in state_list corresponding to the internal state + of the zipformer model for a single utterance. For element-n, + state_list[n] is a list of cached tensors of all encoder layers. For layer-i, + state_list[n][i*6:(i+1)*6] is (cached_key, cached_nonlin_attn, cached_val1, + cached_val2, cached_conv1, cached_conv2). + state_list[n][-2] is the cached left padding for ConvNeXt module, + of shape (batch_size, num_channels, left_pad, num_freqs) + state_list[n][-1] is processed_lens of shape (batch,), which records the number + of processed frames (at 50hz frame rate, after encoder_embed) for each sample in batch. + + Note: + It is the inverse of :func:`unstack_states`. + """ + batch_size = len(state_list) + assert (len(state_list[0]) - 2) % 6 == 0, len(state_list[0]) + tot_num_layers = (len(state_list[0]) - 2) // 6 + + batch_states = [] + for layer in range(tot_num_layers): + layer_offset = layer * 6 + # cached_key: (left_context_len, batch_size, key_dim) + cached_key = torch.cat( + [state_list[i][layer_offset] for i in range(batch_size)], dim=1 + ) + # cached_nonlin_attn: (num_heads, batch_size, left_context_len, head_dim) + cached_nonlin_attn = torch.cat( + [state_list[i][layer_offset + 1] for i in range(batch_size)], dim=1 + ) + # cached_val1: (left_context_len, batch_size, value_dim) + cached_val1 = torch.cat( + [state_list[i][layer_offset + 2] for i in range(batch_size)], dim=1 + ) + # cached_val2: (left_context_len, batch_size, value_dim) + cached_val2 = torch.cat( + [state_list[i][layer_offset + 3] for i in range(batch_size)], dim=1 + ) + # cached_conv1: (#batch, channels, left_pad) + cached_conv1 = torch.cat( + [state_list[i][layer_offset + 4] for i in range(batch_size)], dim=0 + ) + # cached_conv2: (#batch, channels, left_pad) + cached_conv2 = torch.cat( + [state_list[i][layer_offset + 5] for i in range(batch_size)], dim=0 + ) + batch_states += [ + cached_key, + cached_nonlin_attn, + cached_val1, + cached_val2, + cached_conv1, + cached_conv2, + ] + + cached_embed_left_pad = torch.cat( + [state_list[i][-2] for i in range(batch_size)], dim=0 + ) + batch_states.append(cached_embed_left_pad) + + processed_lens = torch.cat([state_list[i][-1] for i in range(batch_size)], dim=0) + batch_states.append(processed_lens) + + return batch_states + + +def unstack_states(batch_states: List[Tensor]) -> List[List[Tensor]]: + """Unstack the zipformer state corresponding to a batch of utterances + into a list of states, where the i-th entry is the state from the i-th + utterance in the batch. + + Note: + It is the inverse of :func:`stack_states`. + + Args: + batch_states: A list of cached tensors of all encoder layers. For layer-i, + states[i*6:(i+1)*6] is (cached_key, cached_nonlin_attn, cached_val1, cached_val2, + cached_conv1, cached_conv2). + state_list[-2] is the cached left padding for ConvNeXt module, + of shape (batch_size, num_channels, left_pad, num_freqs) + states[-1] is processed_lens of shape (batch,), which records the number + of processed frames (at 50hz frame rate, after encoder_embed) for each sample in batch. + + Returns: + state_list: A list of list. Each element in state_list corresponding to the internal state + of the zipformer model for a single utterance. + """ + assert (len(batch_states) - 2) % 6 == 0, len(batch_states) + tot_num_layers = (len(batch_states) - 2) // 6 + + processed_lens = batch_states[-1] + batch_size = processed_lens.shape[0] + + state_list = [[] for _ in range(batch_size)] + + for layer in range(tot_num_layers): + layer_offset = layer * 6 + # cached_key: (left_context_len, batch_size, key_dim) + cached_key_list = batch_states[layer_offset].chunk(chunks=batch_size, dim=1) + # cached_nonlin_attn: (num_heads, batch_size, left_context_len, head_dim) + cached_nonlin_attn_list = batch_states[layer_offset + 1].chunk( + chunks=batch_size, dim=1 + ) + # cached_val1: (left_context_len, batch_size, value_dim) + cached_val1_list = batch_states[layer_offset + 2].chunk( + chunks=batch_size, dim=1 + ) + # cached_val2: (left_context_len, batch_size, value_dim) + cached_val2_list = batch_states[layer_offset + 3].chunk( + chunks=batch_size, dim=1 + ) + # cached_conv1: (#batch, channels, left_pad) + cached_conv1_list = batch_states[layer_offset + 4].chunk( + chunks=batch_size, dim=0 + ) + # cached_conv2: (#batch, channels, left_pad) + cached_conv2_list = batch_states[layer_offset + 5].chunk( + chunks=batch_size, dim=0 + ) + for i in range(batch_size): + state_list[i] += [ + cached_key_list[i], + cached_nonlin_attn_list[i], + cached_val1_list[i], + cached_val2_list[i], + cached_conv1_list[i], + cached_conv2_list[i], + ] + + cached_embed_left_pad_list = batch_states[-2].chunk(chunks=batch_size, dim=0) + for i in range(batch_size): + state_list[i].append(cached_embed_left_pad_list[i]) + + processed_lens_list = batch_states[-1].chunk(chunks=batch_size, dim=0) + for i in range(batch_size): + state_list[i].append(processed_lens_list[i]) + + return state_list + + +def streaming_forward( + features: Tensor, + feature_lens: Tensor, + model: nn.Module, + states: List[Tensor], + chunk_size: int, + left_context_len: int, +) -> Tuple[Tensor, Tensor, List[Tensor]]: + """ + Returns encoder outputs, output lengths, and updated states. + """ + cached_embed_left_pad = states[-2] + (x, x_lens, new_cached_embed_left_pad) = model.encoder_embed.streaming_forward( + x=features, + x_lens=feature_lens, + cached_left_pad=cached_embed_left_pad, + ) + assert x.size(1) == chunk_size, (x.size(1), chunk_size) + + src_key_padding_mask = make_pad_mask(x_lens) + + # processed_mask is used to mask out initial states + processed_mask = torch.arange(left_context_len, device=x.device).expand( + x.size(0), left_context_len + ) + processed_lens = states[-1] # (batch,) + # (batch, left_context_size) + processed_mask = (processed_lens.unsqueeze(1) <= processed_mask).flip(1) + # Update processed lengths + new_processed_lens = processed_lens + x_lens + + # (batch, left_context_size + chunk_size) + src_key_padding_mask = torch.cat([processed_mask, src_key_padding_mask], dim=1) + + x = x.permute(1, 0, 2) # (N, T, C) -> (T, N, C) + encoder_states = states[:-2] + ( + encoder_out, + encoder_out_lens, + new_encoder_states, + ) = model.encoder.streaming_forward( + x=x, + x_lens=x_lens, + states=encoder_states, + src_key_padding_mask=src_key_padding_mask, + ) + encoder_out = encoder_out.permute(1, 0, 2) # (T, N, C) ->(N, T, C) + + new_states = new_encoder_states + [ + new_cached_embed_left_pad, + new_processed_lens, + ] + return encoder_out, encoder_out_lens, new_states + + +def decode_one_chunk( + params: AttributeDict, + model: nn.Module, + decode_streams: List[DecodeStream], +) -> List[int]: + """Decode one chunk frames of features for each decode_streams and + return the indexes of finished streams in a List. + + Args: + params: + It's the return value of :func:`get_params`. + model: + The neural model. + decode_streams: + A List of DecodeStream, each belonging to a utterance. + Returns: + Return a List containing which DecodeStreams are finished. + """ + device = model.device + chunk_size = int(params.chunk_size) + left_context_len = int(params.left_context_frames) + + features = [] + feature_lens = [] + states = [] + processed_lens = [] # Used in fast-beam-search + + for stream in decode_streams: + feat, feat_len = stream.get_feature_frames(chunk_size * 2) + features.append(feat) + feature_lens.append(feat_len) + states.append(stream.states) + processed_lens.append(stream.done_frames) + + feature_lens = torch.tensor(feature_lens, device=device) + features = pad_sequence(features, batch_first=True, padding_value=LOG_EPS) + + # Make sure the length after encoder_embed is at least 1. + # The encoder_embed subsample features (T - 7) // 2 + # The ConvNeXt module needs (7 - 1) // 2 = 3 frames of right padding after subsampling + tail_length = chunk_size * 2 + 7 + 2 * 3 + if features.size(1) < tail_length: + pad_length = tail_length - features.size(1) + feature_lens += pad_length + features = torch.nn.functional.pad( + features, + (0, 0, 0, pad_length), + mode="constant", + value=LOG_EPS, + ) + + states = stack_states(states) + + encoder_out, encoder_out_lens, new_states = streaming_forward( + features=features, + feature_lens=feature_lens, + model=model, + states=states, + chunk_size=chunk_size, + left_context_len=left_context_len, + ) + + encoder_out = model.joiner.encoder_proj(encoder_out) + + if params.decoding_method == "greedy_search": + greedy_search( + model=model, + encoder_out=encoder_out, + streams=decode_streams, + blank_penalty=params.blank_penalty, + ) + elif params.decoding_method == "fast_beam_search": + processed_lens = torch.tensor(processed_lens, device=device) + processed_lens = processed_lens + encoder_out_lens + fast_beam_search_one_best( + model=model, + encoder_out=encoder_out, + processed_lens=processed_lens, + streams=decode_streams, + beam=params.beam, + max_states=params.max_states, + max_contexts=params.max_contexts, + blank_penalty=params.blank_penalty, + ) + elif params.decoding_method == "modified_beam_search": + modified_beam_search( + model=model, + streams=decode_streams, + encoder_out=encoder_out, + num_active_paths=params.num_active_paths, + blank_penalty=params.blank_penalty, + ) + else: + raise ValueError(f"Unsupported decoding method: {params.decoding_method}") + + states = unstack_states(new_states) + + finished_streams = [] + for i in range(len(decode_streams)): + decode_streams[i].states = states[i] + decode_streams[i].done_frames += encoder_out_lens[i] + if decode_streams[i].done: + finished_streams.append(i) + + return finished_streams + + +def decode_dataset( + cuts: CutSet, + params: AttributeDict, + model: nn.Module, + lexicon: Lexicon, + decoding_graph: Optional[k2.Fsa] = None, +) -> Dict[str, List[Tuple[List[str], List[str]]]]: + """Decode dataset. + + Args: + cuts: + Lhotse Cutset containing the dataset to decode. + params: + It is returned by :func:`get_params`. + model: + The neural model. + lexicon: + The Lexicon. + 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. + """ + device = model.device + + opts = 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 + opts.mel_opts.high_freq = -400 + + log_interval = 100 + + decode_results = [] + # Contain decode streams currently running. + decode_streams = [] + for num, cut in enumerate(cuts): + # each utterance has a DecodeStream. + initial_states = get_init_states(model=model, batch_size=1, device=device) + decode_stream = DecodeStream( + params=params, + cut_id=cut.id, + initial_states=initial_states, + decoding_graph=decoding_graph, + device=device, + ) + + audio: np.ndarray = cut.load_audio() + # audio.shape: (1, num_samples) + assert len(audio.shape) == 2 + assert audio.shape[0] == 1, "Should be single channel" + assert audio.dtype == np.float32, audio.dtype + + # The trained model is using normalized samples + # - this is to avoid sending [-32k,+32k] signal in... + # - some lhotse AudioTransform classes can make the signal + # be out of range [-1, 1], hence the tolerance 10 + assert ( + np.abs(audio).max() <= 10 + ), "Should be normalized to [-1, 1], 10 for tolerance..." + + samples = torch.from_numpy(audio).squeeze(0) + + fbank = Fbank(opts) + feature = fbank(samples.to(device)) + decode_stream.set_features(feature, tail_pad_len=30) + decode_stream.ground_truth = cut.supervisions[0].text + + decode_streams.append(decode_stream) + + while len(decode_streams) >= params.num_decode_streams: + finished_streams = decode_one_chunk( + params=params, model=model, decode_streams=decode_streams + ) + for i in sorted(finished_streams, reverse=True): + decode_results.append( + ( + decode_streams[i].id, + list(decode_streams[i].ground_truth.strip()), + [ + lexicon.token_table[idx] + for idx in decode_streams[i].decoding_result() + ], + ) + ) + del decode_streams[i] + + if num % log_interval == 0: + logging.info(f"Cuts processed until now is {num}.") + + # decode final chunks of last sequences + while len(decode_streams): + finished_streams = decode_one_chunk( + params=params, model=model, decode_streams=decode_streams + ) + for i in sorted(finished_streams, reverse=True): + decode_results.append( + ( + decode_streams[i].id, + decode_streams[i].ground_truth.split(), + [ + lexicon.token_table[idx] + for idx in decode_streams[i].decoding_result() + ], + ) + ) + del decode_streams[i] + + key = f"blank_penalty_{params.blank_penalty}" + if params.decoding_method == "greedy_search": + key = f"greedy_search_{key}" + elif params.decoding_method == "fast_beam_search": + key = ( + f"beam_{params.beam}_" + f"max_contexts_{params.max_contexts}_" + f"max_states_{params.max_states}_{key}" + ) + elif params.decoding_method == "modified_beam_search": + key = f"num_active_paths_{params.num_active_paths}_{key}" + else: + raise ValueError(f"Unsupported decoding method: {params.decoding_method}") + return {key: decode_results} + + +def save_results( + params: AttributeDict, + test_set_name: str, + results_dict: Dict[str, List[Tuple[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() + MdccAsrDataModule.add_arguments(parser) + args = parser.parse_args() + args.exp_dir = Path(args.exp_dir) + + params = get_params() + params.update(vars(args)) + + params.res_dir = params.exp_dir / "streaming" / 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}" + + assert params.causal, params.causal + assert "," not in params.chunk_size, "chunk_size should be one value in decoding." + assert ( + "," not in params.left_context_frames + ), "left_context_frames should be one value in decoding." + params.suffix += f"-chunk-{params.chunk_size}" + params.suffix += f"-left-context-{params.left_context_frames}" + params.suffix += f"-blank-penalty-{params.blank_penalty}" + + # for fast_beam_search + 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}" + + 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}") + + lexicon = Lexicon(params.lang_dir) + params.blank_id = lexicon.token_table[""] + params.vocab_size = max(lexicon.tokens) + 1 + + logging.info(params) + + logging.info("About to create model") + model = get_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 start >= 0: + 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() + model.device = device + + decoding_graph = None + if params.decoding_method == "fast_beam_search": + decoding_graph = k2.trivial_graph(params.vocab_size - 1, device=device) + + num_param = sum([p.numel() for p in model.parameters()]) + logging.info(f"Number of model parameters: {num_param}") + + mdcc = MdccAsrDataModule(args) + + valid_cuts = mdcc.valid_cuts() + test_cuts = mdcc.test_cuts() + + test_sets = ["valid", "test"] + test_cuts = [valid_cuts, test_cuts] + + for test_set, test_cut in zip(test_sets, test_cuts): + results_dict = decode_dataset( + cuts=test_cut, + params=params, + model=model, + lexicon=lexicon, + decoding_graph=decoding_graph, + ) + save_results( + params=params, + test_set_name=test_set, + results_dict=results_dict, + ) + + logging.info("Done!") + + +if __name__ == "__main__": + main() diff --git a/egs/mdcc/ASR/zipformer/subsampling.py b/egs/mdcc/ASR/zipformer/subsampling.py new file mode 120000 index 000000000..01ae9002c --- /dev/null +++ b/egs/mdcc/ASR/zipformer/subsampling.py @@ -0,0 +1 @@ +../../../librispeech/ASR/zipformer/subsampling.py \ No newline at end of file diff --git a/egs/mdcc/ASR/zipformer/train.py b/egs/mdcc/ASR/zipformer/train.py new file mode 100755 index 000000000..2fae66844 --- /dev/null +++ b/egs/mdcc/ASR/zipformer/train.py @@ -0,0 +1,1345 @@ +#!/usr/bin/env python3 +# Copyright 2021-2024 Xiaomi Corp. (authors: Fangjun Kuang, +# Wei Kang, +# Mingshuang Luo, +# Zengwei Yao, +# Daniel Povey, +# Zengrui Jin,) +# +# 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: + +export CUDA_VISIBLE_DEVICES="0,1,2,3" + +./zipformer/train.py \ + --world-size 4 \ + --num-epochs 50 \ + --start-epoch 1 \ + --exp-dir zipformer/exp \ + --max-duration 350 + +# For mix precision training: + +./zipformer/train.py \ + --world-size 4 \ + --num-epochs 50 \ + --start-epoch 1 \ + --use-fp16 1 \ + --exp-dir zipformer/exp \ + --max-duration 1000 + +""" + + +import argparse +import copy +import logging +import warnings +from pathlib import Path +from shutil import copyfile +from typing import Any, Dict, Optional, Tuple, Union + +import k2 +import optim +import torch +import torch.multiprocessing as mp +import torch.nn as nn +from asr_datamodule import MdccAsrDataModule +from decoder import Decoder +from joiner import Joiner +from lhotse.cut import Cut +from lhotse.dataset.sampling.base import CutSampler +from lhotse.utils import fix_random_seed +from model import AsrModel +from optim import Eden, ScaledAdam +from scaling import ScheduledFloat +from subsampling import Conv2dSubsampling +from torch import Tensor +from torch.cuda.amp import GradScaler +from torch.nn.parallel import DistributedDataParallel as DDP +from torch.utils.tensorboard import SummaryWriter +from zipformer import Zipformer2 + +from icefall import diagnostics +from icefall.char_graph_compiler import CharCtcTrainingGraphCompiler +from icefall.checkpoint import load_checkpoint, remove_checkpoints +from icefall.checkpoint import save_checkpoint as save_checkpoint_impl +from icefall.checkpoint import ( + save_checkpoint_with_global_batch_idx, + update_averaged_model, +) +from icefall.dist import cleanup_dist, setup_dist +from icefall.env import get_env_info +from icefall.hooks import register_inf_check_hooks +from icefall.lexicon import Lexicon +from icefall.utils import ( + AttributeDict, + MetricsTracker, + get_parameter_groups_with_lrs, + setup_logger, + str2bool, +) + +LRSchedulerType = Union[torch.optim.lr_scheduler._LRScheduler, optim.LRScheduler] + + +def get_adjusted_batch_count(params: AttributeDict) -> float: + # returns the number of batches we would have used so far if we had used the reference + # duration. This is for purposes of set_batch_count(). + return ( + params.batch_idx_train + * (params.max_duration * params.world_size) + / params.ref_duration + ) + + +def set_batch_count(model: Union[nn.Module, DDP], batch_count: float) -> None: + if isinstance(model, DDP): + # get underlying nn.Module + model = model.module + for name, module in model.named_modules(): + if hasattr(module, "batch_count"): + module.batch_count = batch_count + if hasattr(module, "name"): + module.name = name + + +def add_model_arguments(parser: argparse.ArgumentParser): + parser.add_argument( + "--num-encoder-layers", + type=str, + default="2,2,3,4,3,2", + help="Number of zipformer encoder layers per stack, comma separated.", + ) + + parser.add_argument( + "--downsampling-factor", + type=str, + default="1,2,4,8,4,2", + help="Downsampling factor for each stack of encoder layers.", + ) + + parser.add_argument( + "--feedforward-dim", + type=str, + default="512,768,1024,1536,1024,768", + help="""Feedforward dimension of the zipformer encoder layers, per stack, comma separated.""", + ) + + parser.add_argument( + "--num-heads", + type=str, + default="4,4,4,8,4,4", + help="""Number of attention heads in the zipformer encoder layers: a single int or comma-separated list.""", + ) + + parser.add_argument( + "--encoder-dim", + type=str, + default="192,256,384,512,384,256", + help="""Embedding dimension in encoder stacks: a single int or comma-separated list.""", + ) + + parser.add_argument( + "--query-head-dim", + type=str, + default="32", + help="""Query/key dimension per head in encoder stacks: a single int or comma-separated list.""", + ) + + parser.add_argument( + "--value-head-dim", + type=str, + default="12", + help="""Value dimension per head in encoder stacks: a single int or comma-separated list.""", + ) + + parser.add_argument( + "--pos-head-dim", + type=str, + default="4", + help="""Positional-encoding dimension per head in encoder stacks: a single int or comma-separated list.""", + ) + + parser.add_argument( + "--pos-dim", + type=int, + default="48", + help="Positional-encoding embedding dimension", + ) + + parser.add_argument( + "--encoder-unmasked-dim", + type=str, + default="192,192,256,256,256,192", + help="""Unmasked dimensions in the encoders, relates to augmentation during training. A single int or comma-separated list. Must be <= each corresponding encoder_dim.""", + ) + + parser.add_argument( + "--cnn-module-kernel", + type=str, + default="31,31,15,15,15,31", + help="""Sizes of convolutional kernels in convolution modules in each encoder stack: a single int or comma-separated list.""", + ) + + parser.add_argument( + "--decoder-dim", + type=int, + default=512, + help="Embedding dimension in the decoder model.", + ) + + parser.add_argument( + "--joiner-dim", + type=int, + default=512, + help="""Dimension used in the joiner model. + Outputs from the encoder and decoder model are projected + to this dimension before adding. + """, + ) + + parser.add_argument( + "--causal", + type=str2bool, + default=False, + help="If True, use causal version of model.", + ) + + parser.add_argument( + "--chunk-size", + type=str, + default="16,32,64,-1", + help="""Chunk sizes (at 50Hz frame rate) will be chosen randomly from this list during training. Must be just -1 if --causal=False""", + ) + + parser.add_argument( + "--left-context-frames", + type=str, + default="64,128,256,-1", + help="""Maximum left-contexts for causal training, measured in frames which will + be converted to a number of chunks. If splitting into chunks, + chunk left-context frames will be chosen randomly from this list; else not relevant.""", + ) + + +def get_parser(): + parser = argparse.ArgumentParser( + formatter_class=argparse.ArgumentDefaultsHelpFormatter + ) + + parser.add_argument( + "--world-size", + type=int, + default=1, + help="Number of GPUs for DDP training.", + ) + + parser.add_argument( + "--master-port", + type=int, + default=12354, + help="Master port to use for DDP training.", + ) + + parser.add_argument( + "--tensorboard", + type=str2bool, + default=True, + help="Should various information be logged in tensorboard.", + ) + + parser.add_argument( + "--num-epochs", + type=int, + default=30, + help="Number of epochs to train.", + ) + + parser.add_argument( + "--start-epoch", + type=int, + default=1, + help="""Resume training from this epoch. It should be positive. + If larger than 1, it will load checkpoint from + exp-dir/epoch-{start_epoch-1}.pt + """, + ) + + parser.add_argument( + "--start-batch", + type=int, + default=0, + help="""If positive, --start-epoch is ignored and + it loads the checkpoint from exp-dir/checkpoint-{start_batch}.pt + """, + ) + + parser.add_argument( + "--exp-dir", + type=str, + default="zipformer/exp", + help="""The experiment dir. + It specifies the directory where all training related + files, e.g., checkpoints, log, etc, are saved + """, + ) + + parser.add_argument( + "--lang-dir", + type=str, + default="data/lang_char", + help="""The lang dir + It contains language related input files such as + "lexicon.txt" + """, + ) + + parser.add_argument( + "--base-lr", type=float, default=0.045, help="The base learning rate." + ) + + parser.add_argument( + "--lr-batches", + type=float, + default=7500, + help="""Number of steps that affects how rapidly the learning rate + decreases. We suggest not to change this.""", + ) + + parser.add_argument( + "--lr-epochs", + type=float, + default=3.5, + help="""Number of epochs that affects how rapidly the learning rate decreases. + """, + ) + + parser.add_argument( + "--ref-duration", + type=float, + default=600, + help="""Reference batch duration for purposes of adjusting batch counts for setting various schedules inside the model""", + ) + + parser.add_argument( + "--context-size", + type=int, + default=1, + help="""The context size in the decoder. 1 means bigram; 2 means tri-gram""", + ) + + parser.add_argument( + "--prune-range", + type=int, + default=5, + help="""The prune range for rnnt loss, it means how many symbols(context) + we are using to compute the loss""", + ) + + parser.add_argument( + "--lm-scale", + type=float, + default=0.25, + help="""The scale to smooth the loss with lm + (output of prediction network) part.""", + ) + + parser.add_argument( + "--am-scale", + type=float, + default=0.0, + help="""The scale to smooth the loss with am (output of encoder network) part.""", + ) + + parser.add_argument( + "--simple-loss-scale", + type=float, + default=0.5, + help="""To get pruning ranges, we will calculate a simple version + loss(joiner is just addition), this simple loss also uses for + training (as a regularization item). We will scale the simple loss + with this parameter before adding to the final loss.""", + ) + + parser.add_argument( + "--seed", + type=int, + default=42, + help="The seed for random generators intended for reproducibility", + ) + + parser.add_argument( + "--print-diagnostics", + type=str2bool, + default=False, + help="Accumulate stats on activations, print them and exit.", + ) + + parser.add_argument( + "--inf-check", + type=str2bool, + default=False, + help="Add hooks to check for infinite module outputs and gradients.", + ) + + parser.add_argument( + "--save-every-n", + type=int, + default=4000, + help="""Save checkpoint after processing this number of batches" + periodically. We save checkpoint to exp-dir/ whenever + params.batch_idx_train % save_every_n == 0. The checkpoint filename + has the form: f'exp-dir/checkpoint-{params.batch_idx_train}.pt' + Note: It also saves checkpoint to `exp-dir/epoch-xxx.pt` at the + end of each epoch where `xxx` is the epoch number counting from 0. + """, + ) + + parser.add_argument( + "--keep-last-k", + type=int, + default=30, + help="""Only keep this number of checkpoints on disk. + For instance, if it is 3, there are only 3 checkpoints + in the exp-dir with filenames `checkpoint-xxx.pt`. + It does not affect checkpoints with name `epoch-xxx.pt`. + """, + ) + + parser.add_argument( + "--average-period", + type=int, + default=200, + help="""Update the averaged model, namely `model_avg`, after processing + this number of batches. `model_avg` is a separate version of model, + in which each floating-point parameter is the average of all the + parameters from the start of training. Each time we take the average, + we do: `model_avg = model * (average_period / batch_idx_train) + + model_avg * ((batch_idx_train - average_period) / batch_idx_train)`. + """, + ) + + parser.add_argument( + "--use-fp16", + type=str2bool, + default=False, + help="Whether to use half precision training.", + ) + + add_model_arguments(parser) + + return parser + + +def get_params() -> AttributeDict: + """Return a dict containing training parameters. + + All training related parameters that are not passed from the commandline + are saved in the variable `params`. + + Commandline options are merged into `params` after they are parsed, so + you can also access them via `params`. + + Explanation of options saved in `params`: + + - best_train_loss: Best training loss so far. It is used to select + the model that has the lowest training loss. It is + updated during the training. + + - best_valid_loss: Best validation loss so far. It is used to select + the model that has the lowest validation loss. It is + updated during the training. + + - best_train_epoch: It is the epoch that has the best training loss. + + - best_valid_epoch: It is the epoch that has the best validation loss. + + - batch_idx_train: Used to writing statistics to tensorboard. It + contains number of batches trained so far across + epochs. + + - log_interval: Print training loss if batch_idx % log_interval` is 0 + + - reset_interval: Reset statistics if batch_idx % reset_interval is 0 + + - valid_interval: Run validation if batch_idx % valid_interval is 0 + + - feature_dim: The model input dim. It has to match the one used + in computing features. + + - subsampling_factor: The subsampling factor for the model. + + - encoder_dim: Hidden dim for multi-head attention model. + + - num_decoder_layers: Number of decoder layer of transformer decoder. + + - warm_step: The warmup period that dictates the decay of the + scale on "simple" (un-pruned) loss. + """ + params = AttributeDict( + { + "best_train_loss": float("inf"), + "best_valid_loss": float("inf"), + "best_train_epoch": -1, + "best_valid_epoch": -1, + "batch_idx_train": 0, + "log_interval": 50, + "reset_interval": 200, + "valid_interval": 3000, + # parameters for zipformer + "feature_dim": 80, + "subsampling_factor": 4, # not passed in, this is fixed. + "warm_step": 2000, + "env_info": get_env_info(), + } + ) + + return params + + +def _to_int_tuple(s: str): + return tuple(map(int, s.split(","))) + + +def get_encoder_embed(params: AttributeDict) -> nn.Module: + # encoder_embed converts the input of shape (N, T, num_features) + # to the shape (N, (T - 7) // 2, encoder_dims). + # That is, it does two things simultaneously: + # (1) subsampling: T -> (T - 7) // 2 + # (2) embedding: num_features -> encoder_dims + # In the normal configuration, we will downsample once more at the end + # by a factor of 2, and most of the encoder stacks will run at a lower + # sampling rate. + encoder_embed = Conv2dSubsampling( + in_channels=params.feature_dim, + out_channels=_to_int_tuple(params.encoder_dim)[0], + dropout=ScheduledFloat((0.0, 0.3), (20000.0, 0.1)), + ) + return encoder_embed + + +def get_encoder_model(params: AttributeDict) -> nn.Module: + encoder = Zipformer2( + output_downsampling_factor=2, + downsampling_factor=_to_int_tuple(params.downsampling_factor), + num_encoder_layers=_to_int_tuple(params.num_encoder_layers), + encoder_dim=_to_int_tuple(params.encoder_dim), + encoder_unmasked_dim=_to_int_tuple(params.encoder_unmasked_dim), + query_head_dim=_to_int_tuple(params.query_head_dim), + pos_head_dim=_to_int_tuple(params.pos_head_dim), + value_head_dim=_to_int_tuple(params.value_head_dim), + pos_dim=params.pos_dim, + num_heads=_to_int_tuple(params.num_heads), + feedforward_dim=_to_int_tuple(params.feedforward_dim), + cnn_module_kernel=_to_int_tuple(params.cnn_module_kernel), + dropout=ScheduledFloat((0.0, 0.3), (20000.0, 0.1)), + warmup_batches=4000.0, + causal=params.causal, + chunk_size=_to_int_tuple(params.chunk_size), + left_context_frames=_to_int_tuple(params.left_context_frames), + ) + return encoder + + +def get_decoder_model(params: AttributeDict) -> nn.Module: + decoder = Decoder( + vocab_size=params.vocab_size, + decoder_dim=params.decoder_dim, + blank_id=params.blank_id, + context_size=params.context_size, + ) + return decoder + + +def get_joiner_model(params: AttributeDict) -> nn.Module: + joiner = Joiner( + encoder_dim=max(_to_int_tuple(params.encoder_dim)), + decoder_dim=params.decoder_dim, + joiner_dim=params.joiner_dim, + vocab_size=params.vocab_size, + ) + return joiner + + +def get_model(params: AttributeDict) -> nn.Module: + encoder_embed = get_encoder_embed(params) + encoder = get_encoder_model(params) + decoder = get_decoder_model(params) + joiner = get_joiner_model(params) + + model = AsrModel( + encoder_embed=encoder_embed, + encoder=encoder, + decoder=decoder, + joiner=joiner, + encoder_dim=int(max(params.encoder_dim.split(","))), + decoder_dim=params.decoder_dim, + vocab_size=params.vocab_size, + ) + return model + + +def load_checkpoint_if_available( + params: AttributeDict, + model: nn.Module, + model_avg: nn.Module = None, + optimizer: Optional[torch.optim.Optimizer] = None, + scheduler: Optional[LRSchedulerType] = None, +) -> Optional[Dict[str, Any]]: + """Load checkpoint from file. + + If params.start_batch is positive, it will load the checkpoint from + `params.exp_dir/checkpoint-{params.start_batch}.pt`. Otherwise, if + params.start_epoch is larger than 1, it will load the checkpoint from + `params.start_epoch - 1`. + + Apart from loading state dict for `model` and `optimizer` it also updates + `best_train_epoch`, `best_train_loss`, `best_valid_epoch`, + and `best_valid_loss` in `params`. + + Args: + params: + The return value of :func:`get_params`. + model: + The training model. + model_avg: + The stored model averaged from the start of training. + optimizer: + The optimizer that we are using. + scheduler: + The scheduler that we are using. + Returns: + Return a dict containing previously saved training info. + """ + if params.start_batch > 0: + filename = params.exp_dir / f"checkpoint-{params.start_batch}.pt" + elif params.start_epoch > 1: + filename = params.exp_dir / f"epoch-{params.start_epoch-1}.pt" + else: + return None + + assert filename.is_file(), f"{filename} does not exist!" + + saved_params = load_checkpoint( + filename, + model=model, + model_avg=model_avg, + optimizer=optimizer, + scheduler=scheduler, + ) + + keys = [ + "best_train_epoch", + "best_valid_epoch", + "batch_idx_train", + "best_train_loss", + "best_valid_loss", + ] + for k in keys: + params[k] = saved_params[k] + + if params.start_batch > 0: + if "cur_epoch" in saved_params: + params["start_epoch"] = saved_params["cur_epoch"] + + if "cur_batch_idx" in saved_params: + params["cur_batch_idx"] = saved_params["cur_batch_idx"] + + return saved_params + + +def save_checkpoint( + params: AttributeDict, + model: Union[nn.Module, DDP], + model_avg: Optional[nn.Module] = None, + optimizer: Optional[torch.optim.Optimizer] = None, + scheduler: Optional[LRSchedulerType] = None, + sampler: Optional[CutSampler] = None, + scaler: Optional[GradScaler] = None, + rank: int = 0, +) -> None: + """Save model, optimizer, scheduler and training stats to file. + + Args: + params: + It is returned by :func:`get_params`. + model: + The training model. + model_avg: + The stored model averaged from the start of training. + optimizer: + The optimizer used in the training. + sampler: + The sampler for the training dataset. + scaler: + The scaler used for mix precision training. + """ + if rank != 0: + return + filename = params.exp_dir / f"epoch-{params.cur_epoch}.pt" + save_checkpoint_impl( + filename=filename, + model=model, + model_avg=model_avg, + params=params, + optimizer=optimizer, + scheduler=scheduler, + sampler=sampler, + scaler=scaler, + rank=rank, + ) + + if params.best_train_epoch == params.cur_epoch: + best_train_filename = params.exp_dir / "best-train-loss.pt" + copyfile(src=filename, dst=best_train_filename) + + if params.best_valid_epoch == params.cur_epoch: + best_valid_filename = params.exp_dir / "best-valid-loss.pt" + copyfile(src=filename, dst=best_valid_filename) + + +def compute_loss( + params: AttributeDict, + model: Union[nn.Module, DDP], + graph_compiler: CharCtcTrainingGraphCompiler, + batch: dict, + is_training: bool, +) -> Tuple[Tensor, MetricsTracker]: + """ + Compute CTC loss given the model and its inputs. + + Args: + params: + Parameters for training. See :func:`get_params`. + model: + The model for training. It is an instance of Zipformer in our case. + batch: + A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()` + for the content in it. + is_training: + True for training. False for validation. When it is True, this + function enables autograd during computation; when it is False, it + disables autograd. + warmup: a floating point value which increases throughout training; + values >= 1.0 are fully warmed up and have all modules present. + """ + device = model.device if isinstance(model, DDP) else next(model.parameters()).device + feature = batch["inputs"] + # at entry, feature is (N, T, C) + assert feature.ndim == 3 + feature = feature.to(device) + + supervisions = batch["supervisions"] + feature_lens = supervisions["num_frames"].to(device) + + batch_idx_train = params.batch_idx_train + warm_step = params.warm_step + + texts = batch["supervisions"]["text"] + y = graph_compiler.texts_to_ids(texts) + y = k2.RaggedTensor(y).to(device) + + with torch.set_grad_enabled(is_training): + simple_loss, pruned_loss, _ = model( + x=feature, + x_lens=feature_lens, + y=y, + prune_range=params.prune_range, + am_scale=params.am_scale, + lm_scale=params.lm_scale, + ) + + s = params.simple_loss_scale + # take down the scale on the simple loss from 1.0 at the start + # to params.simple_loss scale by warm_step. + simple_loss_scale = ( + s + if batch_idx_train >= warm_step + else 1.0 - (batch_idx_train / warm_step) * (1.0 - s) + ) + pruned_loss_scale = ( + 1.0 + if batch_idx_train >= warm_step + else 0.1 + 0.9 * (batch_idx_train / warm_step) + ) + + loss = simple_loss_scale * simple_loss + pruned_loss_scale * pruned_loss + + assert loss.requires_grad == is_training + + info = MetricsTracker() + with warnings.catch_warnings(): + warnings.simplefilter("ignore") + info["frames"] = (feature_lens // params.subsampling_factor).sum().item() + + # Note: We use reduction=sum while computing the loss. + info["loss"] = loss.detach().cpu().item() + info["simple_loss"] = simple_loss.detach().cpu().item() + info["pruned_loss"] = pruned_loss.detach().cpu().item() + + return loss, info + + +def compute_validation_loss( + params: AttributeDict, + model: Union[nn.Module, DDP], + graph_compiler: CharCtcTrainingGraphCompiler, + valid_dl: torch.utils.data.DataLoader, + world_size: int = 1, +) -> MetricsTracker: + """Run the validation process.""" + model.eval() + + tot_loss = MetricsTracker() + + for batch_idx, batch in enumerate(valid_dl): + loss, loss_info = compute_loss( + params=params, + model=model, + graph_compiler=graph_compiler, + batch=batch, + is_training=False, + ) + assert loss.requires_grad is False + tot_loss = tot_loss + loss_info + + if world_size > 1: + tot_loss.reduce(loss.device) + + loss_value = tot_loss["loss"] / tot_loss["frames"] + if loss_value < params.best_valid_loss: + params.best_valid_epoch = params.cur_epoch + params.best_valid_loss = loss_value + + return tot_loss + + +def train_one_epoch( + params: AttributeDict, + model: Union[nn.Module, DDP], + optimizer: torch.optim.Optimizer, + scheduler: LRSchedulerType, + graph_compiler: CharCtcTrainingGraphCompiler, + train_dl: torch.utils.data.DataLoader, + valid_dl: torch.utils.data.DataLoader, + scaler: GradScaler, + model_avg: Optional[nn.Module] = None, + tb_writer: Optional[SummaryWriter] = None, + world_size: int = 1, + rank: int = 0, +) -> None: + """Train the model for one epoch. + + The training loss from the mean of all frames is saved in + `params.train_loss`. It runs the validation process every + `params.valid_interval` batches. + + Args: + params: + It is returned by :func:`get_params`. + model: + The model for training. + optimizer: + The optimizer we are using. + scheduler: + The learning rate scheduler, we call step() every step. + train_dl: + Dataloader for the training dataset. + valid_dl: + Dataloader for the validation dataset. + scaler: + The scaler used for mix precision training. + model_avg: + The stored model averaged from the start of training. + tb_writer: + Writer to write log messages to tensorboard. + world_size: + Number of nodes in DDP training. If it is 1, DDP is disabled. + rank: + The rank of the node in DDP training. If no DDP is used, it should + be set to 0. + """ + model.train() + + tot_loss = MetricsTracker() + + cur_batch_idx = params.get("cur_batch_idx", 0) + + saved_bad_model = False + + def save_bad_model(suffix: str = ""): + save_checkpoint_impl( + filename=params.exp_dir / f"bad-model{suffix}-{rank}.pt", + model=model, + model_avg=model_avg, + params=params, + optimizer=optimizer, + scheduler=scheduler, + sampler=train_dl.sampler, + scaler=scaler, + rank=0, + ) + + for batch_idx, batch in enumerate(train_dl): + if batch_idx % 10 == 0: + set_batch_count(model, get_adjusted_batch_count(params)) + if batch_idx < cur_batch_idx: + continue + cur_batch_idx = batch_idx + + params.batch_idx_train += 1 + batch_size = len(batch["supervisions"]["text"]) + + try: + with torch.cuda.amp.autocast(enabled=params.use_fp16): + loss, loss_info = compute_loss( + params=params, + model=model, + graph_compiler=graph_compiler, + batch=batch, + is_training=True, + ) + # 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() + except: # noqa + save_bad_model() + display_and_save_batch(batch, params=params, graph_compiler=graph_compiler) + raise + + if params.print_diagnostics and batch_idx == 5: + return + + if ( + rank == 0 + and params.batch_idx_train > 0 + and params.batch_idx_train % params.average_period == 0 + ): + update_averaged_model( + params=params, + model_cur=model, + model_avg=model_avg, + ) + + if ( + params.batch_idx_train > 0 + and params.batch_idx_train % params.save_every_n == 0 + ): + params.cur_batch_idx = batch_idx + save_checkpoint_with_global_batch_idx( + out_dir=params.exp_dir, + global_batch_idx=params.batch_idx_train, + model=model, + model_avg=model_avg, + params=params, + optimizer=optimizer, + scheduler=scheduler, + sampler=train_dl.sampler, + scaler=scaler, + rank=rank, + ) + del params.cur_batch_idx + remove_checkpoints( + out_dir=params.exp_dir, + topk=params.keep_last_k, + rank=rank, + ) + + if batch_idx % 100 == 0 and params.use_fp16: + # If the grad scale was less than 1, try increasing it. The _growth_interval + # of the grad scaler is configurable, but we can't configure it to have different + # behavior depending on the current grad scale. + cur_grad_scale = scaler._scale.item() + + if cur_grad_scale < 8.0 or (cur_grad_scale < 32.0 and batch_idx % 400 == 0): + scaler.update(cur_grad_scale * 2.0) + if cur_grad_scale < 0.01: + if not saved_bad_model: + save_bad_model(suffix="-first-warning") + saved_bad_model = True + logging.warning(f"Grad scale is small: {cur_grad_scale}") + if cur_grad_scale < 1.0e-05: + save_bad_model() + raise RuntimeError( + f"grad_scale is too small, exiting: {cur_grad_scale}" + ) + + if batch_idx % params.log_interval == 0: + cur_lr = max(scheduler.get_last_lr()) + cur_grad_scale = scaler._scale.item() if params.use_fp16 else 1.0 + + logging.info( + f"Epoch {params.cur_epoch}, " + f"batch {batch_idx}, loss[{loss_info}], " + f"tot_loss[{tot_loss}], batch size: {batch_size}, " + f"lr: {cur_lr:.2e}, " + + (f"grad_scale: {scaler._scale.item()}" if params.use_fp16 else "") + ) + + if tb_writer is not None: + tb_writer.add_scalar( + "train/learning_rate", cur_lr, params.batch_idx_train + ) + + loss_info.write_summary( + tb_writer, "train/current_", params.batch_idx_train + ) + tot_loss.write_summary(tb_writer, "train/tot_", params.batch_idx_train) + if params.use_fp16: + tb_writer.add_scalar( + "train/grad_scale", cur_grad_scale, params.batch_idx_train + ) + + if batch_idx % params.valid_interval == 0 and not params.print_diagnostics: + logging.info("Computing validation loss") + valid_info = compute_validation_loss( + params=params, + model=model, + graph_compiler=graph_compiler, + valid_dl=valid_dl, + world_size=world_size, + ) + model.train() + logging.info(f"Epoch {params.cur_epoch}, validation: {valid_info}") + logging.info( + f"Maximum memory allocated so far is {torch.cuda.max_memory_allocated()//1000000}MB" + ) + if tb_writer is not None: + valid_info.write_summary( + tb_writer, "train/valid_", params.batch_idx_train + ) + + loss_value = tot_loss["loss"] / tot_loss["frames"] + params.train_loss = loss_value + if params.train_loss < params.best_train_loss: + params.best_train_epoch = params.cur_epoch + params.best_train_loss = params.train_loss + + +def run(rank, world_size, args): + """ + Args: + rank: + It is a value between 0 and `world_size-1`, which is + passed automatically by `mp.spawn()` in :func:`main`. + The node with rank 0 is responsible for saving checkpoint. + world_size: + Number of GPUs for DDP training. + args: + The return value of get_parser().parse_args() + """ + params = get_params() + params.update(vars(args)) + + fix_random_seed(params.seed) + if world_size > 1: + setup_dist(rank, world_size, params.master_port) + + setup_logger(f"{params.exp_dir}/log/log-train") + logging.info("Training started") + + if args.tensorboard and rank == 0: + tb_writer = SummaryWriter(log_dir=f"{params.exp_dir}/tensorboard") + else: + tb_writer = None + + device = torch.device("cpu") + if torch.cuda.is_available(): + device = torch.device("cuda", rank) + logging.info(f"Device: {device}") + + lexicon = Lexicon(params.lang_dir) + graph_compiler = CharCtcTrainingGraphCompiler( + lexicon=lexicon, + device=device, + ) + + params.blank_id = lexicon.token_table[""] + params.vocab_size = max(lexicon.tokens) + 1 + + logging.info(params) + + logging.info("About to create model") + model = get_model(params) + + num_param = sum([p.numel() for p in model.parameters()]) + logging.info(f"Number of model parameters: {num_param}") + + assert params.save_every_n >= params.average_period + model_avg: Optional[nn.Module] = None + if rank == 0: + # model_avg is only used with rank 0 + model_avg = copy.deepcopy(model).to(torch.float64) + + assert params.start_epoch > 0, params.start_epoch + checkpoints = load_checkpoint_if_available( + params=params, model=model, model_avg=model_avg + ) + + model.to(device) + if world_size > 1: + logging.info("Using DDP") + model = DDP(model, device_ids=[rank], find_unused_parameters=True) + + optimizer = ScaledAdam( + get_parameter_groups_with_lrs(model, lr=params.base_lr, include_names=True), + lr=params.base_lr, # should have no effect + clipping_scale=2.0, + ) + + scheduler = Eden(optimizer, params.lr_batches, params.lr_epochs) + + if checkpoints and "optimizer" in checkpoints: + logging.info("Loading optimizer state dict") + optimizer.load_state_dict(checkpoints["optimizer"]) + + if ( + checkpoints + and "scheduler" in checkpoints + and checkpoints["scheduler"] is not None + ): + logging.info("Loading scheduler state dict") + scheduler.load_state_dict(checkpoints["scheduler"]) + + if params.print_diagnostics: + opts = diagnostics.TensorDiagnosticOptions( + 512 + ) # allow 4 megabytes per sub-module + diagnostic = diagnostics.attach_diagnostics(model, opts) + + if params.inf_check: + register_inf_check_hooks(model) + + mdcc = MdccAsrDataModule(args) + + train_cuts = mdcc.train_cuts() + valid_cuts = mdcc.valid_cuts() + + def remove_short_and_long_utt(c: Cut): + # Keep only utterances with duration between 1 second and 15 seconds + # + # Caution: There is a reason to select 15.0 here. Please see + # ../local/display_manifest_statistics.py + # + # You should use ../local/display_manifest_statistics.py to get + # an utterance duration distribution for your dataset to select + # the threshold + if c.duration < 1.0 or c.duration > 20.0: + # logging.warning( + # f"Exclude cut with ID {c.id} from training. Duration: {c.duration}" + # ) + return False + + # In pruned RNN-T, we require that T >= S + # where T is the number of feature frames after subsampling + # and S is the number of tokens in the utterance + + # In ./zipformer.py, the conv module uses the following expression + # for subsampling + T = ((c.num_frames - 7) // 2 + 1) // 2 + tokens = graph_compiler.texts_to_ids([c.supervisions[0].text])[0] + + if T < len(tokens): + logging.warning( + f"Exclude cut with ID {c.id} from training. " + f"Number of frames (before subsampling): {c.num_frames}. " + f"Number of frames (after subsampling): {T}. " + f"Text: {c.supervisions[0].text}. " + f"Tokens: {tokens}. " + f"Number of tokens: {len(tokens)}" + ) + return False + + return True + + train_cuts = train_cuts.filter(remove_short_and_long_utt) + + if params.start_batch > 0 and checkpoints and "sampler" in checkpoints: + # We only load the sampler's state dict when it loads a checkpoint + # saved in the middle of an epoch + sampler_state_dict = checkpoints["sampler"] + else: + sampler_state_dict = None + + train_dl = mdcc.train_dataloaders(train_cuts, sampler_state_dict=sampler_state_dict) + + valid_dl = mdcc.valid_dataloaders(valid_cuts) + + if False and not params.print_diagnostics: + scan_pessimistic_batches_for_oom( + model=model, + train_dl=train_dl, + optimizer=optimizer, + graph_compiler=graph_compiler, + params=params, + ) + + scaler = GradScaler(enabled=params.use_fp16, init_scale=1.0) + if checkpoints and "grad_scaler" in checkpoints: + logging.info("Loading grad scaler state dict") + scaler.load_state_dict(checkpoints["grad_scaler"]) + + for epoch in range(params.start_epoch, params.num_epochs + 1): + scheduler.step_epoch(epoch - 1) + fix_random_seed(params.seed + epoch - 1) + train_dl.sampler.set_epoch(epoch - 1) + + if tb_writer is not None: + tb_writer.add_scalar("train/epoch", epoch, params.batch_idx_train) + + params.cur_epoch = epoch + + train_one_epoch( + params=params, + model=model, + model_avg=model_avg, + optimizer=optimizer, + scheduler=scheduler, + graph_compiler=graph_compiler, + train_dl=train_dl, + valid_dl=valid_dl, + scaler=scaler, + tb_writer=tb_writer, + world_size=world_size, + rank=rank, + ) + + if params.print_diagnostics: + diagnostic.print_diagnostics() + break + + save_checkpoint( + params=params, + model=model, + model_avg=model_avg, + optimizer=optimizer, + scheduler=scheduler, + sampler=train_dl.sampler, + scaler=scaler, + rank=rank, + ) + + logging.info("Done!") + + if world_size > 1: + torch.distributed.barrier() + cleanup_dist() + + +def display_and_save_batch( + batch: dict, + params: AttributeDict, + graph_compiler: CharCtcTrainingGraphCompiler, +) -> 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`. + graph_compiler: + The compiler to encode texts to ids. + """ + 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}") + + texts = supervisions["text"] + y = graph_compiler.texts_to_ids(texts) + num_tokens = sum(len(i) for i in y) + logging.info(f"num tokens: {num_tokens}") + + +def scan_pessimistic_batches_for_oom( + model: Union[nn.Module, DDP], + train_dl: torch.utils.data.DataLoader, + optimizer: torch.optim.Optimizer, + graph_compiler: CharCtcTrainingGraphCompiler, + params: AttributeDict, +): + from lhotse.dataset import find_pessimistic_batches + + logging.info( + "Sanity check -- see if any of the batches in epoch 1 would cause OOM." + ) + batches, crit_values = find_pessimistic_batches(train_dl.sampler) + for criterion, cuts in batches.items(): + batch = train_dl.dataset[cuts] + try: + with torch.cuda.amp.autocast(enabled=params.use_fp16): + loss, _ = compute_loss( + params=params, + model=model, + graph_compiler=graph_compiler, + batch=batch, + is_training=True, + ) + loss.backward() + optimizer.zero_grad() + except Exception as e: + if "CUDA out of memory" in str(e): + logging.error( + "Your GPU ran out of memory with the current " + "max_duration setting. We recommend decreasing " + "max_duration and trying again.\n" + f"Failing criterion: {criterion} " + f"(={crit_values[criterion]}) ..." + ) + display_and_save_batch(batch, params=params, graph_compiler=graph_compiler) + raise + logging.info( + f"Maximum memory allocated so far is {torch.cuda.max_memory_allocated()//1000000}MB" + ) + + +def main(): + parser = get_parser() + MdccAsrDataModule.add_arguments(parser) + args = parser.parse_args() + args.lang_dir = Path(args.lang_dir) + args.exp_dir = Path(args.exp_dir) + + world_size = args.world_size + assert world_size >= 1 + if world_size > 1: + mp.spawn(run, args=(world_size, args), nprocs=world_size, join=True) + else: + run(rank=0, world_size=1, args=args) + + +torch.set_num_threads(1) +torch.set_num_interop_threads(1) + +if __name__ == "__main__": + main() diff --git a/egs/mdcc/ASR/zipformer/zipformer.py b/egs/mdcc/ASR/zipformer/zipformer.py new file mode 120000 index 000000000..23011dda7 --- /dev/null +++ b/egs/mdcc/ASR/zipformer/zipformer.py @@ -0,0 +1 @@ +../../../librispeech/ASR/zipformer/zipformer.py \ No newline at end of file diff --git a/egs/multi_zh-hans/ASR/README.md b/egs/multi_zh-hans/ASR/README.md index 537816a5d..1e60c733c 100644 --- a/egs/multi_zh-hans/ASR/README.md +++ b/egs/multi_zh-hans/ASR/README.md @@ -36,4 +36,4 @@ This recipe includes scripts for training Zipformer model using multiple Chinese 3. AliMeeting 4. MagicData 5. KeSpeech-ASR -6. WeNetSpeech \ No newline at end of file +6. WeNetSpeech diff --git a/egs/multi_zh-hans/ASR/local/compute_fbank_kespeech_dev_test.py b/egs/multi_zh-hans/ASR/local/compute_fbank_kespeech_dev_test.py index 2581ee42f..6f75dbfa4 100755 --- a/egs/multi_zh-hans/ASR/local/compute_fbank_kespeech_dev_test.py +++ b/egs/multi_zh-hans/ASR/local/compute_fbank_kespeech_dev_test.py @@ -17,11 +17,21 @@ # See the License for the specific language governing permissions and # limitations under the License. +import argparse import logging from pathlib import Path import torch -from lhotse import CutSet, KaldifeatFbank, KaldifeatFbankConfig, LilcomChunkyWriter +from lhotse import ( + CutSet, + KaldifeatFbank, + KaldifeatFbankConfig, + LilcomChunkyWriter, + WhisperFbank, + WhisperFbankConfig, +) + +from icefall.utils import str2bool # Torch's multithreaded behavior needs to be disabled or # it wastes a lot of CPU and slow things down. @@ -31,7 +41,28 @@ torch.set_num_threads(1) torch.set_num_interop_threads(1) -def compute_fbank_kespeech_dev_test(): +def get_parser(): + parser = argparse.ArgumentParser( + formatter_class=argparse.ArgumentDefaultsHelpFormatter + ) + + parser.add_argument( + "--num-mel-bins", + type=int, + default=80, + help="""The number of mel bins for Fbank""", + ) + + parser.add_argument( + "--whisper-fbank", + type=str2bool, + default=False, + help="Use WhisperFbank instead of Fbank. Default: False.", + ) + return parser + + +def compute_fbank_kespeech_dev_test(args): in_out_dir = Path("data/fbank/kespeech") # number of workers in dataloader num_workers = 42 @@ -48,7 +79,12 @@ def compute_fbank_kespeech_dev_test(): device = torch.device("cpu") if torch.cuda.is_available(): device = torch.device("cuda", 0) - extractor = KaldifeatFbank(KaldifeatFbankConfig(device=device)) + if args.whisper_fbank: + extractor = WhisperFbank( + WhisperFbankConfig(num_filters=args.num_mel_bins, device=device) + ) + else: + extractor = KaldifeatFbank(KaldifeatFbankConfig(device=device)) logging.info(f"device: {device}") @@ -86,7 +122,11 @@ def main(): formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s" logging.basicConfig(format=formatter, level=logging.INFO) - compute_fbank_kespeech_dev_test() + parser = get_parser() + args = parser.parse_args() + logging.info(vars(args)) + + compute_fbank_kespeech_dev_test(args) if __name__ == "__main__": diff --git a/egs/multi_zh-hans/ASR/local/compute_fbank_kespeech_splits.py b/egs/multi_zh-hans/ASR/local/compute_fbank_kespeech_splits.py index 8bfbc7b50..c398411f6 100755 --- a/egs/multi_zh-hans/ASR/local/compute_fbank_kespeech_splits.py +++ b/egs/multi_zh-hans/ASR/local/compute_fbank_kespeech_splits.py @@ -28,10 +28,14 @@ from lhotse import ( KaldifeatFbank, KaldifeatFbankConfig, LilcomChunkyWriter, + WhisperFbank, + WhisperFbankConfig, set_audio_duration_mismatch_tolerance, set_caching_enabled, ) +from icefall.utils import str2bool + # Torch's multithreaded behavior needs to be disabled or # it wastes a lot of CPU and slow things down. # Do this outside of main() in case it needs to take effect @@ -88,6 +92,20 @@ def get_parser(): default=-1, help="Stop processing pieces until this number (exclusive).", ) + + parser.add_argument( + "--num-mel-bins", + type=int, + default=80, + help="""The number of mel bins for Fbank""", + ) + + parser.add_argument( + "--whisper-fbank", + type=str2bool, + default=False, + help="Use WhisperFbank instead of Fbank. Default: False.", + ) return parser @@ -111,14 +129,19 @@ def compute_fbank_kespeech_splits(args): device = torch.device("cpu") if torch.cuda.is_available(): device = torch.device("cuda", 0) - extractor = KaldifeatFbank(KaldifeatFbankConfig(device=device)) + if args.whisper_fbank: + extractor = WhisperFbank( + WhisperFbankConfig(num_filters=args.num_mel_bins, device="cuda") + ) + else: + extractor = KaldifeatFbank(KaldifeatFbankConfig(device=device)) logging.info(f"device: {device}") set_audio_duration_mismatch_tolerance(0.01) # 10ms tolerance set_caching_enabled(False) for i in range(start, stop): - idx = f"{i + 1}".zfill(num_digits) - logging.info(f"Processing {idx}/{num_splits}") + idx = f"{i}".zfill(num_digits) + logging.info(f"Processing {i+1}/{num_splits}") cuts_path = output_dir / f"kespeech-asr_cuts_{subset}.{idx}.jsonl.gz" if cuts_path.is_file(): diff --git a/egs/multi_zh-hans/ASR/local/compute_fbank_magicdata.py b/egs/multi_zh-hans/ASR/local/compute_fbank_magicdata.py index 5649d3815..192bffa9f 100755 --- a/egs/multi_zh-hans/ASR/local/compute_fbank_magicdata.py +++ b/egs/multi_zh-hans/ASR/local/compute_fbank_magicdata.py @@ -30,10 +30,17 @@ import os from pathlib import Path import torch -from lhotse import CutSet, Fbank, FbankConfig, LilcomChunkyWriter +from lhotse import ( + CutSet, + Fbank, + FbankConfig, + LilcomChunkyWriter, + WhisperFbank, + WhisperFbankConfig, +) from lhotse.recipes.utils import read_manifests_if_cached -from icefall.utils import get_executor +from icefall.utils import get_executor, str2bool # Torch's multithreaded behavior needs to be disabled or # it wastes a lot of CPU and slow things down. @@ -43,10 +50,33 @@ torch.set_num_threads(1) torch.set_num_interop_threads(1) -def compute_fbank_magicdata(num_mel_bins: int = 80, speed_perturb: bool = False): +def get_parser(): + parser = argparse.ArgumentParser( + formatter_class=argparse.ArgumentDefaultsHelpFormatter + ) + + parser.add_argument( + "--num-mel-bins", + type=int, + default=80, + help="""The number of mel bins for Fbank""", + ) + + parser.add_argument( + "--whisper-fbank", + type=str2bool, + default=False, + help="Use WhisperFbank instead of Fbank. Default: False.", + ) + return parser + + +def compute_fbank_magicdata( + num_mel_bins: int = 80, speed_perturb: bool = False, whisper_fbank: bool = False +): src_dir = Path("data/manifests/magicdata") output_dir = Path("data/fbank") - num_jobs = min(30, os.cpu_count()) + num_jobs = min(8, os.cpu_count()) dataset_parts = ("train", "test", "dev") prefix = "magicdata" @@ -66,7 +96,12 @@ def compute_fbank_magicdata(num_mel_bins: int = 80, speed_perturb: bool = False) dataset_parts, ) - extractor = Fbank(FbankConfig(num_mel_bins=num_mel_bins)) + if args.whisper_fbank: + extractor = WhisperFbank( + WhisperFbankConfig(num_filters=args.num_mel_bins, device="cuda") + ) + else: + extractor = Fbank(FbankConfig(num_mel_bins=num_mel_bins)) with get_executor() as ex: # Initialize the executor only once. for partition, m in manifests.items(): @@ -107,7 +142,12 @@ def get_args(): default=False, help="Enable 0.9 and 1.1 speed perturbation for data augmentation. Default: False.", ) - + parser.add_argument( + "--whisper-fbank", + type=str2bool, + default=False, + help="Use WhisperFbank instead of Fbank. Default: False.", + ) return parser.parse_args() @@ -118,5 +158,7 @@ if __name__ == "__main__": args = get_args() compute_fbank_magicdata( - num_mel_bins=args.num_mel_bins, speed_perturb=args.speed_perturb + num_mel_bins=args.num_mel_bins, + speed_perturb=args.speed_perturb, + whisper_fbank=args.whisper_fbank, ) diff --git a/egs/multi_zh-hans/ASR/local/compute_fbank_primewords.py b/egs/multi_zh-hans/ASR/local/compute_fbank_primewords.py index 303a16580..019b10d24 100755 --- a/egs/multi_zh-hans/ASR/local/compute_fbank_primewords.py +++ b/egs/multi_zh-hans/ASR/local/compute_fbank_primewords.py @@ -30,10 +30,17 @@ import os from pathlib import Path import torch -from lhotse import CutSet, Fbank, FbankConfig, LilcomChunkyWriter +from lhotse import ( + CutSet, + Fbank, + FbankConfig, + LilcomChunkyWriter, + WhisperFbank, + WhisperFbankConfig, +) from lhotse.recipes.utils import read_manifests_if_cached -from icefall.utils import get_executor +from icefall.utils import get_executor, str2bool # Torch's multithreaded behavior needs to be disabled or # it wastes a lot of CPU and slow things down. @@ -43,7 +50,9 @@ torch.set_num_threads(1) torch.set_num_interop_threads(1) -def compute_fbank_primewords(num_mel_bins: int = 80, speed_perturb: bool = False): +def compute_fbank_primewords( + num_mel_bins: int = 80, speed_perturb: bool = False, whisper_fbank: bool = False +): src_dir = Path("data/manifests/primewords") output_dir = Path("data/fbank") num_jobs = min(15, os.cpu_count()) @@ -66,7 +75,12 @@ def compute_fbank_primewords(num_mel_bins: int = 80, speed_perturb: bool = False dataset_parts, ) - extractor = Fbank(FbankConfig(num_mel_bins=num_mel_bins)) + if whisper_fbank: + extractor = WhisperFbank( + WhisperFbankConfig(num_filters=args.num_mel_bins, device="cuda") + ) + else: + extractor = Fbank(FbankConfig(num_mel_bins=num_mel_bins)) with get_executor() as ex: # Initialize the executor only once. for partition, m in manifests.items(): @@ -108,6 +122,13 @@ def get_args(): help="Enable 0.9 and 1.1 speed perturbation for data augmentation. Default: False.", ) + parser.add_argument( + "--whisper-fbank", + type=str2bool, + default=False, + help="Use WhisperFbank instead of Fbank. Default: False.", + ) + return parser.parse_args() @@ -118,5 +139,7 @@ if __name__ == "__main__": args = get_args() compute_fbank_primewords( - num_mel_bins=args.num_mel_bins, speed_perturb=args.speed_perturb + num_mel_bins=args.num_mel_bins, + speed_perturb=args.speed_perturb, + whisper_fbank=args.whisper_fbank, ) diff --git a/egs/multi_zh-hans/ASR/local/compute_fbank_stcmds.py b/egs/multi_zh-hans/ASR/local/compute_fbank_stcmds.py index 730806954..f29ae5a46 100755 --- a/egs/multi_zh-hans/ASR/local/compute_fbank_stcmds.py +++ b/egs/multi_zh-hans/ASR/local/compute_fbank_stcmds.py @@ -30,10 +30,17 @@ import os from pathlib import Path import torch -from lhotse import CutSet, Fbank, FbankConfig, LilcomChunkyWriter +from lhotse import ( + CutSet, + Fbank, + FbankConfig, + LilcomChunkyWriter, + WhisperFbank, + WhisperFbankConfig, +) from lhotse.recipes.utils import read_manifests_if_cached -from icefall.utils import get_executor +from icefall.utils import get_executor, str2bool # Torch's multithreaded behavior needs to be disabled or # it wastes a lot of CPU and slow things down. @@ -43,7 +50,9 @@ torch.set_num_threads(1) torch.set_num_interop_threads(1) -def compute_fbank_stcmds(num_mel_bins: int = 80, speed_perturb: bool = False): +def compute_fbank_stcmds( + num_mel_bins: int = 80, speed_perturb: bool = False, whisper_fbank: bool = False +): src_dir = Path("data/manifests/stcmds") output_dir = Path("data/fbank") num_jobs = min(15, os.cpu_count()) @@ -66,7 +75,12 @@ def compute_fbank_stcmds(num_mel_bins: int = 80, speed_perturb: bool = False): dataset_parts, ) - extractor = Fbank(FbankConfig(num_mel_bins=num_mel_bins)) + if whisper_fbank: + extractor = WhisperFbank( + WhisperFbankConfig(num_filters=args.num_mel_bins, device="cuda") + ) + else: + extractor = Fbank(FbankConfig(num_mel_bins=num_mel_bins)) with get_executor() as ex: # Initialize the executor only once. for partition, m in manifests.items(): @@ -107,6 +121,12 @@ def get_args(): default=False, help="Enable 0.9 and 1.1 speed perturbation for data augmentation. Default: False.", ) + parser.add_argument( + "--whisper-fbank", + type=str2bool, + default=False, + help="Use WhisperFbank instead of Fbank. Default: False.", + ) return parser.parse_args() @@ -117,5 +137,7 @@ if __name__ == "__main__": args = get_args() compute_fbank_stcmds( - num_mel_bins=args.num_mel_bins, speed_perturb=args.speed_perturb + num_mel_bins=args.num_mel_bins, + speed_perturb=args.speed_perturb, + whisper_fbank=args.whisper_fbank, ) diff --git a/egs/multi_zh-hans/ASR/local/compute_fbank_thchs30.py b/egs/multi_zh-hans/ASR/local/compute_fbank_thchs30.py index 58bb8002a..4ad78e0ba 100755 --- a/egs/multi_zh-hans/ASR/local/compute_fbank_thchs30.py +++ b/egs/multi_zh-hans/ASR/local/compute_fbank_thchs30.py @@ -30,10 +30,17 @@ import os from pathlib import Path import torch -from lhotse import CutSet, Fbank, FbankConfig, LilcomChunkyWriter +from lhotse import ( + CutSet, + Fbank, + FbankConfig, + LilcomChunkyWriter, + WhisperFbank, + WhisperFbankConfig, +) from lhotse.recipes.utils import read_manifests_if_cached -from icefall.utils import get_executor +from icefall.utils import get_executor, str2bool # Torch's multithreaded behavior needs to be disabled or # it wastes a lot of CPU and slow things down. @@ -43,7 +50,9 @@ torch.set_num_threads(1) torch.set_num_interop_threads(1) -def compute_fbank_thchs30(num_mel_bins: int = 80, speed_perturb: bool = False): +def compute_fbank_thchs30( + num_mel_bins: int = 80, speed_perturb: bool = False, whisper_fbank: bool = False +): src_dir = Path("data/manifests/thchs30") output_dir = Path("data/fbank") num_jobs = min(15, os.cpu_count()) @@ -70,7 +79,12 @@ def compute_fbank_thchs30(num_mel_bins: int = 80, speed_perturb: bool = False): dataset_parts, ) - extractor = Fbank(FbankConfig(num_mel_bins=num_mel_bins)) + if whisper_fbank: + extractor = WhisperFbank( + WhisperFbankConfig(num_filters=args.num_mel_bins, device="cuda") + ) + else: + extractor = Fbank(FbankConfig(num_mel_bins=num_mel_bins)) with get_executor() as ex: # Initialize the executor only once. for partition, m in manifests.items(): @@ -113,6 +127,12 @@ def get_args(): default=False, help="Enable 0.9 and 1.1 speed perturbation for data augmentation. Default: False.", ) + parser.add_argument( + "--whisper-fbank", + type=str2bool, + default=False, + help="Use WhisperFbank instead of Fbank. Default: False.", + ) return parser.parse_args() @@ -123,5 +143,7 @@ if __name__ == "__main__": args = get_args() compute_fbank_thchs30( - num_mel_bins=args.num_mel_bins, speed_perturb=args.speed_perturb + num_mel_bins=args.num_mel_bins, + speed_perturb=args.speed_perturb, + whisper_fbank=args.whisper_fbank, ) diff --git a/egs/multi_zh-hans/ASR/prepare.sh b/egs/multi_zh-hans/ASR/prepare.sh index c09b9c1de..fa515ed50 100755 --- a/egs/multi_zh-hans/ASR/prepare.sh +++ b/egs/multi_zh-hans/ASR/prepare.sh @@ -60,7 +60,7 @@ if [ $stage -le 2 ] && [ $stop_stage -ge 2 ]; then if [ ! -f data/fbank/.thchs30.done ]; then mkdir -p data/fbank - ./local/compute_fbank_thchs30.py + ./local/compute_fbank_thchs30.py --speed-perturb true touch data/fbank/.thchs30.done fi fi @@ -86,7 +86,7 @@ fi log "Dataset: AISHELL-2" if [ $stage -le 4 ] && [ $stop_stage -ge 4 ]; then log "Stage 4: Prepare AISHELL-2" - if [ -e ../../aishell/ASR/data/fbank/.aishell2.done ]; then + if [ -e ../../aishell2/ASR/data/fbank/.aishell2.done ]; then cd data/fbank ln -svf $(realpath ../../../../aishell2/ASR/data/fbank/aishell2_feats_train) . ln -svf $(realpath ../../../../aishell2/ASR/data/fbank/aishell2_feats_dev) . @@ -95,30 +95,30 @@ if [ $stage -le 4 ] && [ $stop_stage -ge 4 ]; then ln -svf $(realpath ../../../../aishell2/ASR/data/fbank/aishell2_cuts_dev.jsonl.gz) . ln -svf $(realpath ../../../../aishell2/ASR/data/fbank/aishell2_cuts_test.jsonl.gz) . cd ../.. - else + else log "Abort! Please run ../../aishell2/ASR/prepare.sh --stage 3 --stop-stage 3" exit 1 - fi + fi fi log "Dataset: AISHELL-4" if [ $stage -le 5 ] && [ $stop_stage -ge 5 ]; then log "Stage 5: Prepare AISHELL-4" - if [ -e ../../aishell/ASR/data/fbank/.aishell4.done ]; then + if [ -e ../../aishell4/ASR/data/fbank/.fbank.done ]; then cd data/fbank - ln -svf $(realpath ../../../../aishell4/ASR/data/fbank/aishell4_feats_train) . - ln -svf $(realpath ../../../../aishell4/ASR/data/fbank/aishell4_feats_dev) . ln -svf $(realpath ../../../../aishell4/ASR/data/fbank/aishell4_feats_test) . + ln -svf $(realpath ../../../../aishell4/ASR/data/fbank/aishell4_feats_train_L) . + ln -svf $(realpath ../../../../aishell4/ASR/data/fbank/aishell4_feats_train_M) . + ln -svf $(realpath ../../../../aishell4/ASR/data/fbank/aishell4_feats_train_S) . ln -svf $(realpath ../../../../aishell4/ASR/data/fbank/aishell4_cuts_train_L.jsonl.gz) . ln -svf $(realpath ../../../../aishell4/ASR/data/fbank/aishell4_cuts_train_M.jsonl.gz) . ln -svf $(realpath ../../../../aishell4/ASR/data/fbank/aishell4_cuts_train_S.jsonl.gz) . - ln -svf $(realpath ../../../../aishell4/ASR/data/fbank/aishell4_cuts_dev.jsonl.gz) . ln -svf $(realpath ../../../../aishell4/ASR/data/fbank/aishell4_cuts_test.jsonl.gz) . cd ../.. - else + else log "Abort! Please run ../../aishell4/ASR/prepare.sh --stage 3 --stop-stage 3" exit 1 - fi + fi fi log "Dataset: ST-CMDS" @@ -137,7 +137,7 @@ if [ $stage -le 6 ] && [ $stop_stage -ge 6 ]; then if [ ! -f data/fbank/.stcmds.done ]; then mkdir -p data/fbank - ./local/compute_fbank_stcmds.py + ./local/compute_fbank_stcmds.py --speed-perturb true touch data/fbank/.stcmds.done fi fi @@ -151,15 +151,15 @@ if [ $stage -le 7 ] && [ $stop_stage -ge 7 ]; then lhotse download primewords $dl_dir/primewords fi - if [ ! -f data/manifests/.stcmds.done ]; then + if [ ! -f data/manifests/.primewords.done ]; then mkdir -p data/manifests - lhotse prepare stcmds $dl_dir/primewords data/manifests/primewords + lhotse prepare primewords $dl_dir/primewords data/manifests/primewords touch data/manifests/.primewords.done fi if [ ! -f data/fbank/.primewords.done ]; then mkdir -p data/fbank - ./local/compute_fbank_primewords.py + ./local/compute_fbank_primewords.py --speed-perturb true touch data/fbank/.primewords.done fi fi @@ -180,7 +180,7 @@ if [ $stage -le 8 ] && [ $stop_stage -ge 8 ]; then if [ ! -f data/fbank/.magicdata.done ]; then mkdir -p data/fbank - ./local/compute_fbank_magicdata.py + ./local/compute_fbank_magicdata.py --speed-perturb true touch data/fbank/.magicdata.done fi fi @@ -231,7 +231,7 @@ if [ $stage -le 11 ] && [ $stop_stage -ge 11 ]; then ln -svf $(realpath ../../../../wenetspeech/ASR/data/fbank/cuts_TEST_MEETING.jsonl.gz) . ln -svf $(realpath ../../../../wenetspeech/ASR/data/fbank/cuts_TEST_NET.jsonl.gz) . - ln -svf $(realpath ../../../../wenetspeech/ASR/data/fbank/L_split_1000) . + ln -svf $(realpath ../../../../wenetspeech/ASR/data/fbank/L_split_${num_splits}) . ln -svf $(realpath ../../../../wenetspeech/ASR/data/fbank/*.lca) . ln -svf $(realpath ../../../../wenetspeech/ASR/data/fbank/) ./wenetspeech cd ../.. @@ -261,7 +261,7 @@ if [ $stage -le 12 ] && [ $stop_stage -ge 12 ]; then if [ ! -f data/manifests/.kespeech.done ]; then mkdir -p data/manifests - lhotse prepare kespeech -j 16 $dl_dir/KeSpeech data/manifests/kespeech + lhotse prepare kespeech -j 16 $dl_dir/KeSpeech data/manifests/kespeech touch data/manifests/.kespeech.done fi @@ -272,29 +272,29 @@ if [ $stage -le 12 ] && [ $stop_stage -ge 12 ]; then if [ ! -f data/fbank/.kespeech_preprocess_complete ]; then python3 ./local/preprocess_kespeech.py touch data/fbank/.kespeech_preprocess_complete - fi - - if [ -f data/fbank/.kespeech.train_phase1.split.${num_splits}.done ]; then + fi + + if [ ! -f data/fbank/.kespeech.train_phase1.split.${num_splits}.done ]; then log "Spliting KeSpeech train_phase1" lhotse split ${num_splits} \ data/fbank/kespeech/kespeech-asr_cuts_train_phase1_raw.jsonl.gz \ data/fbank/kespeech/train_phase1_split_${num_splits} touch data/fbank/.kespeech.train_phase1.split.${num_splits}.done fi - - if [ -f data/fbank/.kespeech.train_phase2.split.${num_splits}.done ]; then + + if [ ! -f data/fbank/.kespeech.train_phase2.split.${num_splits}.done ]; then log "Spliting KeSpeech train_phase2" lhotse split ${num_splits} \ data/fbank/kespeech/kespeech-asr_cuts_train_phase2_raw.jsonl.gz \ data/fbank/kespeech/train_phase2_split_${num_splits} touch data/fbank/.kespeech.train_phase2.split.${num_splits}.done fi - + log "Compute KeSpeech fbank for train_phase1" - ./local/compute_fbank_kespeech_splits.py --num-splits ${num_splits} --training-subset train_phase1 + ./local/compute_fbank_kespeech_splits.py --speed-perturb true --num-splits ${num_splits} --training-subset train_phase1 log "Compute KeSpeech fbank for train_phase2" - ./local/compute_fbank_kespeech_splits.py --num-splits ${num_splits} --training-subset train_phase2 + ./local/compute_fbank_kespeech_splits.py --speed-perturb true --num-splits ${num_splits} --training-subset train_phase2 log "Compute KeSpeech fbank for test/dev" ./local/compute_fbank_kespeech_dev_test.py @@ -303,13 +303,126 @@ if [ $stage -le 12 ] && [ $stop_stage -ge 12 ]; then fi fi +whisper_mel_bins=80 +if [ $stage -le 120 ] && [ $stop_stage -ge 120 ]; then + log "Stage 120: Prepare KeSpeech for whisper" + if [ ! -d $dl_dir/KeSpeech ]; then + log "Abort! Please download KeSpeech first." + log "KeSpeech download link: https://github.com/KeSpeech/KeSpeech" + exit 1 + fi + + if [ ! -f data/manifests/.kespeech.done ]; then + mkdir -p data/manifests + lhotse prepare kespeech -j 8 $dl_dir/KeSpeech data/manifests/kespeech + touch data/manifests/.kespeech.done + fi + + if [ ! -f data/fbank/.kespeech.done ]; then + mkdir -p data/fbank + + log "Preprocess KeSpeech manifest" + if [ ! -f data/fbank/.kespeech_preprocess_complete ]; then + python3 ./local/preprocess_kespeech.py --speed-perturb true + touch data/fbank/.kespeech_preprocess_complete + fi + + if [ ! -f data/fbank/.kespeech.train_phase1.split.${num_splits}.done ]; then + log "Spliting KeSpeech train_phase1" + lhotse split ${num_splits} \ + data/fbank/kespeech/kespeech-asr_cuts_train_phase1_raw.jsonl.gz \ + data/fbank/kespeech/train_phase1_split_${num_splits} + touch data/fbank/.kespeech.train_phase1.split.${num_splits}.done + fi + + if [ ! -f data/fbank/.kespeech.train_phase2.split.${num_splits}.done ]; then + log "Spliting KeSpeech train_phase2" + lhotse split ${num_splits} \ + data/fbank/kespeech/kespeech-asr_cuts_train_phase2_raw.jsonl.gz \ + data/fbank/kespeech/train_phase2_split_${num_splits} + touch data/fbank/.kespeech.train_phase2.split.${num_splits}.done + fi + + log "Compute KeSpeech fbank for train_phase1" + ./local/compute_fbank_kespeech_splits.py --num-splits ${num_splits} --training-subset train_phase1 --num-mel-bins ${whisper_mel_bins} --whisper-fbank true + + log "Compute KeSpeech fbank for train_phase2" + ./local/compute_fbank_kespeech_splits.py --num-splits ${num_splits} --training-subset train_phase2 --num-mel-bins ${whisper_mel_bins} --whisper-fbank true + + log "Compute KeSpeech fbank for test/dev" + # ./local/compute_fbank_kespeech_dev_test.py --num-mel-bins ${whisper_mel_bins} --whisper-fbank true + + if [ ! -f data/fbank/kespeech/kespeech-asr_cuts_train_phase1.jsonl.gz ]; then + pieces=$(find data/fbank/kespeech/train_phase1_split_${num_splits} -name "kespeech-asr_cuts_train_phase1.*.jsonl.gz") + lhotse combine $pieces data/fbank/kespeech/kespeech-asr_cuts_train_phase1.jsonl.gz + fi + if [ ! -f data/fbank/kespeech/kespeech-asr_cuts_train_phase2.jsonl.gz ]; then + pieces=$(find data/fbank/kespeech/train_phase2_split_${num_splits} -name "kespeech-asr_cuts_train_phase2.*.jsonl.gz") + lhotse combine $pieces data/fbank/kespeech/kespeech-asr_cuts_train_phase2.jsonl.gz + fi + touch data/fbank/.kespeech.done + fi +fi + +if [ $stage -le 121 ] && [ $stop_stage -ge 121 ]; then + log "Stage 121: Prepare MagicData, Primewords, ST-CMDS, THCHS-30 for whisper" + + if [ ! -f data/manifests/.magicdata.done ]; then + mkdir -p data/manifests + lhotse prepare magicdata $dl_dir/magicdata data/manifests/magicdata + touch data/manifests/.magicdata.done + fi + + if [ ! -f data/manifests/.primewords.done ]; then + mkdir -p data/manifests + lhotse prepare primewords $dl_dir/primewords data/manifests/primewords + touch data/manifests/.primewords.done + fi + if [ ! -f data/manifests/.stcmds.done ]; then + mkdir -p data/manifests + lhotse prepare stcmds $dl_dir/stcmds data/manifests/stcmds + touch data/manifests/.stcmds.done + fi + + if [ ! -f data/manifests/.thchs30.done ]; then + mkdir -p data/manifests + lhotse prepare thchs-30 $dl_dir/thchs30 data/manifests/thchs30 + touch data/manifests/.thchs30.done + fi + + if [ ! -f data/fbank/.thchs30.done ]; then + mkdir -p data/fbank + ./local/compute_fbank_thchs30.py --speed-perturb true --num-mel-bins ${whisper_mel_bins} --whisper-fbank true + touch data/fbank/.thchs30.done + fi + + if [ ! -f data/fbank/.stcmds.done ]; then + mkdir -p data/fbank + ./local/compute_fbank_stcmds.py --speed-perturb true --num-mel-bins ${whisper_mel_bins} --whisper-fbank true + touch data/fbank/.stcmds.done + fi + if [ ! -f data/fbank/.magicdata.done ]; then + mkdir -p data/fbank + ./local/compute_fbank_magicdata.py --speed-perturb true --num-mel-bins ${whisper_mel_bins} --whisper-fbank true + touch data/fbank/.magicdata.done + fi + + if [ ! -f data/fbank/.primewords.done ]; then + mkdir -p data/fbank + ./local/compute_fbank_primewords.py --speed-perturb true --num-mel-bins ${whisper_mel_bins} --whisper-fbank true + touch data/fbank/.primewords.done + fi + +fi + + if [ $stage -le 13 ] && [ $stop_stage -ge 13 ]; then log "Stage 13: BPE model training (note that we use transcripts of wenetspeech only for BPE training)" ./local/prepare_for_bpe_model.py --lang-dir ./data/lang_char --text ./data/lang_char/text for vocab_size in ${vocab_sizes[@]}; do lang_dir=data/lang_bpe_${vocab_size} - + mkdir -p $lang_dir if [ ! -f $lang_dir/bpe.model ]; then ./local/train_bpe_model.py \ @@ -329,7 +442,7 @@ if [ $stage -le 13 ] && [ $stop_stage -ge 13 ]; then --lexicon $lang_dir/lexicon.txt \ --bpe-model $lang_dir/bpe.model fi - + if [ ! -f $lang_dir/L.fst ]; then log "Converting L.pt to L.fst" ./shared/convert-k2-to-openfst.py \ @@ -350,7 +463,7 @@ fi if [ $stage -le 14 ] && [ $stop_stage -ge 14 ]; then log "Stage 14: Prepare G (note that we use ngram lm of wenetspeech only for G preparation)" - + if [ -d ../../wenetspeech/ASR/data/lang_char/ ]; then cd data ln -s ../../../../wenetspeech/ASR/data/lm . @@ -369,5 +482,3 @@ if [ $stage -le 15 ] && [ $stop_stage -ge 15 ]; then python ./local/compile_lg.py --lang-dir $lang_dir done fi - - diff --git a/egs/multi_zh-hans/ASR/whisper/asr_datamodule.py b/egs/multi_zh-hans/ASR/whisper/asr_datamodule.py new file mode 120000 index 000000000..3c8b7f2d4 --- /dev/null +++ b/egs/multi_zh-hans/ASR/whisper/asr_datamodule.py @@ -0,0 +1 @@ +../zipformer/asr_datamodule.py \ No newline at end of file diff --git a/egs/multi_zh-hans/ASR/whisper/decode.py b/egs/multi_zh-hans/ASR/whisper/decode.py new file mode 100644 index 000000000..aabb80eaf --- /dev/null +++ b/egs/multi_zh-hans/ASR/whisper/decode.py @@ -0,0 +1,519 @@ +#!/usr/bin/env python3 +# Copyright 2021 Xiaomi Corporation (Author: Liyong Guo, +# Fangjun Kuang, +# Wei Kang) +# 2024 Yuekai Zhang +# +# 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: +# Command for decoding using fine-tuned models: +git lfs install +git clone https://huggingface.co/yuekai/icefall_asr_aishell_whisper +ln -s icefall_asr_aishell_whisper/exp_large_v2/epoch-10-avg6.pt whisper/exp_large_v2/epoch-999.pt + +python3 ./whisper/decode.py \ + --exp-dir whisper/exp_large_v2 \ + --model-name large-v2 \ + --epoch 999 --avg 1 \ + --beam-size 10 --max-duration 50 + +# Command for decoding using pretrained models (before fine-tuning): + +python3 ./whisper/decode.py \ + --exp-dir whisper/exp_large_v2 \ + --model-name large-v2 \ + --epoch -1 --avg 1 \ + --remove-whisper-encoder-input-length-restriction False \ + --beam-size 10 --max-duration 50 + +""" + +import argparse +import logging +import re +from collections import defaultdict +from pathlib import Path +from typing import Dict, List, Optional, Tuple + +import k2 +import torch +import torch.nn as nn +import whisper +from asr_datamodule import AsrDataModule +from lhotse.cut import Cut +from multi_dataset import MultiDataset +from tn.chinese.normalizer import Normalizer +from whisper.normalizers import BasicTextNormalizer +from whisper_encoder_forward_monkey_patch import replace_whisper_encoder_forward +from zhconv import convert + +from icefall.checkpoint import average_checkpoints_with_averaged_model, load_checkpoint +from icefall.env import get_env_info +from icefall.utils import ( + AttributeDict, + setup_logger, + store_transcripts, + str2bool, + write_error_stats, +) + + +def average_checkpoints( + filenames: List[Path], device: torch.device = torch.device("cpu") +) -> dict: + """Average a list of checkpoints. + The function is mainly used for deepspeed converted checkpoint averaging, which only include model state_dict. + + Args: + filenames: + Filenames of the checkpoints to be averaged. We assume all + checkpoints are saved by :func:`save_checkpoint`. + device: + Move checkpoints to this device before averaging. + Returns: + Return a dict (i.e., state_dict) which is the average of all + model state dicts contained in the checkpoints. + """ + n = len(filenames) + + if "model" in torch.load(filenames[0], map_location=device): + avg = torch.load(filenames[0], map_location=device)["model"] + else: + avg = torch.load(filenames[0], map_location=device) + + # 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 avg.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 i in range(1, n): + if "model" in torch.load(filenames[i], map_location=device): + state_dict = torch.load(filenames[i], map_location=device)["model"] + else: + state_dict = torch.load(filenames[i], map_location=device) + for k in uniqued_names: + avg[k] += state_dict[k] + + for k in uniqued_names: + if avg[k].is_floating_point(): + avg[k] /= n + else: + avg[k] //= n + + return avg + + +def remove_punctuation(text: str or List[str]): + """Modified from https://github.com/yeyupiaoling/Whisper-Finetune/blob/master/utils/data_utils.py + + Args: + text: It can be a string or a list of strings. + Returns: + Return a string or a list of strings without any punctuation. + """ + punctuation = "!,.;:?、！，。；：？《》 " + if isinstance(text, str): + text = re.sub(r"[{}]+".format(punctuation), "", text).strip() + return text + elif isinstance(text, list): + result_text = [] + for t in text: + t = re.sub(r"[{}]+".format(punctuation), "", t).strip() + result_text.append(t) + return result_text + else: + raise Exception(f"Not support type {type(text)}") + + +def to_simple(text: str or List[str]): + """Convert traditional Chinese to simplified Chinese. + Args: + text: It can be a string or a list of strings. + Returns: + Return a string or a list of strings converted to simplified Chinese. + """ + if isinstance(text, str): + text = convert(text, "zh-cn") + return text + elif isinstance(text, list): + result_text = [] + for t in text: + t = convert(t, "zh-cn") + result_text.append(t) + return result_text + else: + raise Exception(f"Not support type{type(text)}") + + +def get_parser(): + parser = argparse.ArgumentParser( + formatter_class=argparse.ArgumentDefaultsHelpFormatter + ) + + parser.add_argument( + "--epoch", + type=int, + default=-1, + help="It specifies the checkpoint to use for decoding." + "Note: Epoch counts from 0.", + ) + parser.add_argument( + "--avg", + type=int, + default=1, + help="Number of checkpoints to average. Automatically select " + "consecutive checkpoints before the checkpoint specified by " + "'--epoch'. ", + ) + + parser.add_argument( + "--method", + type=str, + default="beam-search", + help="""Decoding method. + Supported values are: + - beam-search + """, + ) + + parser.add_argument( + "--beam-size", + type=int, + default=1, + help="beam size for beam search decoding", + ) + + parser.add_argument( + "--exp-dir", + type=str, + default="whisper/exp", + help="The experiment dir", + ) + + parser.add_argument( + "--model-name", + type=str, + default="large-v2", + choices=["large-v2", "large-v3", "medium", "small", "tiny"], + help="""The model name to use. + """, + ) + + parser.add_argument( + "--remove-whisper-encoder-input-length-restriction", + type=str2bool, + default=True, + help="replace whisper encoder forward method to remove input length restriction", + ) + + return parser + + +def get_params() -> AttributeDict: + params = AttributeDict( + { + "env_info": get_env_info(), + } + ) + return params + + +def decode_one_batch( + params: AttributeDict, + model: nn.Module, + batch: dict, +) -> Dict[str, List[List[int]]]: + """Decode one batch and return the result in a dict. The dict has the + following format: + + - key: "beam-search" + - value: A list of lists. Each sublist is a list of token IDs. + Args: + params: + It is returned by :func:`get_params`. + model: + The neural model. + batch: + It is returned by :meth:`torch.utils.data.DataLoader.__iter__`. + Returns: + Return a dict, whose key may be "beam-search". + """ + dtype = torch.float16 + device = torch.device("cuda") + + feature = batch["inputs"] + assert feature.ndim == 3 + feature = feature.to(device, dtype=dtype).transpose(1, 2) + if not params.remove_whisper_encoder_input_length_restriction: + T = 3000 + if feature.shape[2] < T: + feature = torch.cat( + [ + feature, + torch.zeros( + feature.shape[0], feature.shape[1], T - feature.shape[2] + ).to(device, dtype=dtype), + ], + 2, + ) + + supervisions = batch["supervisions"] + feature_len = supervisions["num_frames"] + feature_len = feature_len.to(device, dtype=dtype) + results = model.decode(feature, params.decoding_options) + hyps = [result.text for result in results] + + hyps = remove_punctuation(hyps) + hyps = to_simple(hyps) + hyps = [params.normalizer.normalize(hyp) for hyp in hyps] + print(hyps) + return {"beam-search": hyps} + + +def decode_dataset( + dl: torch.utils.data.DataLoader, + params: AttributeDict, + model: nn.Module, +) -> Dict[str, List[Tuple[str, List[str], List[str]]]]: + """Decode dataset. + + Args: + dl: + The dataloader. + params: + It is returned by :func:`get_params`. + model: + The neural model. + Returns: + Return a dict, whose key may be "beam-search". + """ + results = [] + + 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, + batch=batch, + ) + + for lm_scale, 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[lm_scale].extend(this_batch) + + num_cuts += len(batch["supervisions"]["text"]) + + 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]]]], +): + + enable_log = True + test_set_wers = dict() + for key, results in results_dict.items(): + recog_path = ( + params.exp_dir / f"recogs-{test_set_name}-{key}-{params.suffix}.txt" + ) + results = sorted(results) + store_transcripts(filename=recog_path, texts=results) + if enable_log: + 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.exp_dir / f"errs-{test_set_name}-{key}-{params.suffix}.txt" + ) + # we compute CER for aishell dataset. + results_char = [] + for res in results: + results_char.append((res[0], list("".join(res[1])), list("".join(res[2])))) + with open(errs_filename, "w") as f: + wer = write_error_stats( + f, f"{test_set_name}-{key}", results_char, enable_log=enable_log + ) + test_set_wers[key] = wer + + if enable_log: + 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.exp_dir / f"cer-summary-{test_set_name}-{params.suffix}.txt" + with open(errs_info, "w") as f: + print("settings\tCER", file=f) + for key, val in test_set_wers: + print("{}\t{}".format(key, val), file=f) + + s = "\nFor {}, CER 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() + AsrDataModule.add_arguments(parser) + args = parser.parse_args() + args.exp_dir = Path(args.exp_dir) + + params = get_params() + params.update(vars(args)) + params.suffix = f"epoch-{params.epoch}-avg-{params.avg}" + setup_logger( + f"{params.exp_dir}/log-{params.method}-beam{params.beam_size}/log-decode-{params.suffix}" + ) + + options = whisper.DecodingOptions( + task="transcribe", + language="zh", + without_timestamps=True, + beam_size=params.beam_size, + ) + params.decoding_options = options + params.cleaner = BasicTextNormalizer() + params.normalizer = Normalizer() + + logging.info("Decoding started") + logging.info(params) + + device = torch.device("cpu") + if torch.cuda.is_available(): + device = torch.device("cuda") + + logging.info(f"device: {device}") + + if params.remove_whisper_encoder_input_length_restriction: + replace_whisper_encoder_forward() + model = whisper.load_model(params.model_name, "cpu") + if params.epoch > 0: + if params.avg > 1: + start = params.epoch - params.avg + assert start >= 1, start + checkpoint = torch.load( + f"{params.exp_dir}/epoch-{params.epoch}.pt", map_location="cpu" + ) + if "model" not in checkpoint: + # deepspeed converted checkpoint only contains model state_dict + filenames = [ + f"{params.exp_dir}/epoch-{epoch}.pt" + for epoch in range(start, params.epoch + 1) + ] + model.load_state_dict(average_checkpoints(filenames)) + else: + 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, + ) + ) + # save checkpoints + filename = f"{params.exp_dir}/epoch-{params.epoch}-avg-{params.avg}.pt" + torch.save(model.state_dict(), filename) + else: + checkpoint = torch.load( + f"{params.exp_dir}/epoch-{params.epoch}.pt", map_location="cpu" + ) + if "model" not in checkpoint: + model.load_state_dict(checkpoint, strict=True) + else: + load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model) + 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 + + data_module = AsrDataModule(args) + multi_dataset = MultiDataset(args.manifest_dir) + + def remove_long_utt(c: Cut): + # Keep only utterances with duration in 30 seconds + # + if c.duration > 30.0: + # logging.warning( + # f"Exclude cut with ID {c.id} from training. Duration: {c.duration}" + # ) + return False + return True + + test_sets_cuts = multi_dataset.test_cuts() + + test_sets = test_sets_cuts.keys() + test_dls = [ + data_module.test_dataloaders(test_sets_cuts[cuts_name].filter(remove_long_utt)) + for cuts_name in test_sets + ] + + for test_set, test_dl in zip(test_sets, test_dls): + results_dict = decode_dataset( + dl=test_dl, + params=params, + model=model, + ) + + save_results(params=params, test_set_name=test_set, results_dict=results_dict) + + logging.info("Done!") + + +torch.set_num_threads(1) +torch.set_num_interop_threads(1) + +if __name__ == "__main__": + main() diff --git a/egs/multi_zh-hans/ASR/whisper/ds_config_zero1.json b/egs/multi_zh-hans/ASR/whisper/ds_config_zero1.json new file mode 120000 index 000000000..af7162d6c --- /dev/null +++ b/egs/multi_zh-hans/ASR/whisper/ds_config_zero1.json @@ -0,0 +1 @@ +../../../aishell/ASR/whisper/ds_config_zero1.json \ No newline at end of file diff --git a/egs/multi_zh-hans/ASR/whisper/label_smoothing.py b/egs/multi_zh-hans/ASR/whisper/label_smoothing.py new file mode 120000 index 000000000..e9d239fff --- /dev/null +++ b/egs/multi_zh-hans/ASR/whisper/label_smoothing.py @@ -0,0 +1 @@ +../../../librispeech/ASR/conformer_ctc/label_smoothing.py \ No newline at end of file diff --git a/egs/multi_zh-hans/ASR/whisper/multi_dataset.py b/egs/multi_zh-hans/ASR/whisper/multi_dataset.py new file mode 100644 index 000000000..b562e626b --- /dev/null +++ b/egs/multi_zh-hans/ASR/whisper/multi_dataset.py @@ -0,0 +1,296 @@ +# Copyright 2023 Xiaomi Corp. (authors: Zengrui Jin) +# +# 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. + + +import glob +import logging +import re +from pathlib import Path +from typing import Dict, List + +import lhotse +from lhotse import CutSet, load_manifest_lazy + + +class MultiDataset: + def __init__(self, fbank_dir: str): + """ + Args: + manifest_dir: + It is expected to contain the following files: + - aishell_cuts_train.jsonl.gz + - aishell2_cuts_train.jsonl.gz + - aishell4_cuts_train_L.jsonl.gz + - aishell4_cuts_train_M.jsonl.gz + - aishell4_cuts_train_S.jsonl.gz + - alimeeting-far_cuts_train.jsonl.gz + - magicdata_cuts_train.jsonl.gz + - primewords_cuts_train.jsonl.gz + - stcmds_cuts_train.jsonl.gz + - thchs_30_cuts_train.jsonl.gz + - kespeech/kespeech-asr_cuts_train_phase1.jsonl.gz + - kespeech/kespeech-asr_cuts_train_phase2.jsonl.gz + - wenetspeech/cuts_L.jsonl.gz + """ + self.fbank_dir = Path(fbank_dir) + + def train_cuts(self) -> CutSet: + logging.info("About to get multidataset train cuts") + + # THCHS-30 + logging.info("Loading THCHS-30 in lazy mode") + thchs_30_cuts = load_manifest_lazy( + self.fbank_dir / "thchs_30_cuts_train.jsonl.gz" + ) + + # AISHELL-1 + logging.info("Loading Aishell-1 in lazy mode") + aishell_cuts = load_manifest_lazy( + self.fbank_dir / "aishell_cuts_train.jsonl.gz" + ) + + # AISHELL-2 + logging.info("Loading Aishell-2 in lazy mode") + aishell_2_cuts = load_manifest_lazy( + self.fbank_dir / "aishell2_cuts_train.jsonl.gz" + ) + + # AISHELL-4 + logging.info("Loading Aishell-4 in lazy mode") + aishell_4_L_cuts = load_manifest_lazy( + self.fbank_dir / "aishell4_cuts_train_L.jsonl.gz" + ) + aishell_4_M_cuts = load_manifest_lazy( + self.fbank_dir / "aishell4_cuts_train_M.jsonl.gz" + ) + aishell_4_S_cuts = load_manifest_lazy( + self.fbank_dir / "aishell4_cuts_train_S.jsonl.gz" + ) + + # ST-CMDS + logging.info("Loading ST-CMDS in lazy mode") + stcmds_cuts = load_manifest_lazy(self.fbank_dir / "stcmds_cuts_train.jsonl.gz") + + # Primewords + logging.info("Loading Primewords in lazy mode") + primewords_cuts = load_manifest_lazy( + self.fbank_dir / "primewords_cuts_train.jsonl.gz" + ) + + # MagicData + logging.info("Loading MagicData in lazy mode") + magicdata_cuts = load_manifest_lazy( + self.fbank_dir / "magicdata_cuts_train.jsonl.gz" + ) + + # Ali-Meeting + logging.info("Loading Ali-Meeting in lazy mode") + alimeeting_cuts = load_manifest_lazy( + self.fbank_dir / "alimeeting-far_cuts_train.jsonl.gz" + ) + + # WeNetSpeech + logging.info("Loading WeNetSpeech in lazy mode") + wenetspeech_L_cuts = load_manifest_lazy( + self.fbank_dir / "wenetspeech" / "cuts_L.jsonl.gz" + ) + + # KeSpeech + logging.info("Loading KeSpeech in lazy mode") + kespeech_1_cuts = load_manifest_lazy( + self.fbank_dir / "kespeech" / "kespeech-asr_cuts_train_phase1.jsonl.gz" + ) + kespeech_2_cuts = load_manifest_lazy( + self.fbank_dir / "kespeech" / "kespeech-asr_cuts_train_phase2.jsonl.gz" + ) + + return CutSet.mux( + thchs_30_cuts, + aishell_cuts, + aishell_2_cuts, + aishell_4_L_cuts, + aishell_4_M_cuts, + aishell_4_S_cuts, + stcmds_cuts, + primewords_cuts, + magicdata_cuts, + alimeeting_cuts, + wenetspeech_L_cuts, + kespeech_1_cuts, + kespeech_2_cuts, + weights=[ + len(thchs_30_cuts), + len(aishell_cuts), + len(aishell_2_cuts), + len(aishell_4_L_cuts), + len(aishell_4_M_cuts), + len(aishell_4_S_cuts), + len(stcmds_cuts), + len(primewords_cuts), + len(magicdata_cuts), + len(alimeeting_cuts), + len(wenetspeech_L_cuts), + len(kespeech_1_cuts), + len(kespeech_2_cuts), + ], + ) + + def dev_cuts(self) -> CutSet: + logging.info("About to get multidataset dev cuts") + + # AISHELL + logging.info("Loading Aishell DEV set in lazy mode") + aishell_dev_cuts = load_manifest_lazy( + self.fbank_dir / "aishell_cuts_dev.jsonl.gz" + ) + + # AISHELL-2 + logging.info("Loading Aishell-2 DEV set in lazy mode") + aishell2_dev_cuts = load_manifest_lazy( + self.fbank_dir / "aishell2_cuts_dev.jsonl.gz" + ) + + # Ali-Meeting + logging.info("Loading Ali-Meeting DEV set in lazy mode") + alimeeting_dev_cuts = load_manifest_lazy( + self.fbank_dir / "alimeeting-far_cuts_eval.jsonl.gz" + ) + + # MagicData + logging.info("Loading MagicData DEV set in lazy mode") + magicdata_dev_cuts = load_manifest_lazy( + self.fbank_dir / "magicdata_cuts_dev.jsonl.gz" + ) + + # KeSpeech + logging.info("Loading KeSpeech DEV set in lazy mode") + kespeech_dev_phase1_cuts = load_manifest_lazy( + self.fbank_dir / "kespeech" / "kespeech-asr_cuts_dev_phase1.jsonl.gz" + ) + kespeech_dev_phase2_cuts = load_manifest_lazy( + self.fbank_dir / "kespeech" / "kespeech-asr_cuts_dev_phase2.jsonl.gz" + ) + + # WeNetSpeech + logging.info("Loading WeNetSpeech DEV set in lazy mode") + wenetspeech_dev_cuts = load_manifest_lazy( + self.fbank_dir / "wenetspeech" / "cuts_DEV.jsonl.gz" + ) + + return wenetspeech_dev_cuts + # return [ + # aishell_dev_cuts, + # aishell2_dev_cuts, + # alimeeting_dev_cuts, + # magicdata_dev_cuts, + # kespeech_dev_phase1_cuts, + # kespeech_dev_phase2_cuts, + # wenetspeech_dev_cuts, + # ] + + def test_cuts(self) -> Dict[str, CutSet]: + logging.info("About to get multidataset test cuts") + + # AISHELL + logging.info("Loading Aishell set in lazy mode") + aishell_test_cuts = load_manifest_lazy( + self.fbank_dir / "aishell_cuts_test.jsonl.gz" + ) + aishell_dev_cuts = load_manifest_lazy( + self.fbank_dir / "aishell_cuts_dev.jsonl.gz" + ) + + # AISHELL-2 + logging.info("Loading Aishell-2 set in lazy mode") + aishell2_test_cuts = load_manifest_lazy( + self.fbank_dir / "aishell2_cuts_test.jsonl.gz" + ) + aishell2_dev_cuts = load_manifest_lazy( + self.fbank_dir / "aishell2_cuts_dev.jsonl.gz" + ) + + # AISHELL-4 + logging.info("Loading Aishell-4 TEST set in lazy mode") + aishell4_test_cuts = load_manifest_lazy( + self.fbank_dir / "aishell4_cuts_test.jsonl.gz" + ) + + # Ali-Meeting + logging.info("Loading Ali-Meeting set in lazy mode") + alimeeting_test_cuts = load_manifest_lazy( + self.fbank_dir / "alimeeting-far_cuts_test.jsonl.gz" + ) + alimeeting_eval_cuts = load_manifest_lazy( + self.fbank_dir / "alimeeting-far_cuts_eval.jsonl.gz" + ) + + # MagicData + logging.info("Loading MagicData set in lazy mode") + magicdata_test_cuts = load_manifest_lazy( + self.fbank_dir / "magicdata_cuts_test.jsonl.gz" + ) + magicdata_dev_cuts = load_manifest_lazy( + self.fbank_dir / "magicdata_cuts_dev.jsonl.gz" + ) + + # KeSpeech + logging.info("Loading KeSpeech set in lazy mode") + kespeech_test_cuts = load_manifest_lazy( + self.fbank_dir / "kespeech" / "kespeech-asr_cuts_test.jsonl.gz" + ) + kespeech_dev_phase1_cuts = load_manifest_lazy( + self.fbank_dir / "kespeech" / "kespeech-asr_cuts_dev_phase1.jsonl.gz" + ) + kespeech_dev_phase2_cuts = load_manifest_lazy( + self.fbank_dir / "kespeech" / "kespeech-asr_cuts_dev_phase2.jsonl.gz" + ) + + # WeNetSpeech + logging.info("Loading WeNetSpeech set in lazy mode") + wenetspeech_test_meeting_cuts = load_manifest_lazy( + self.fbank_dir / "wenetspeech" / "cuts_TEST_MEETING.jsonl.gz" + ) + wenetspeech_test_net_cuts = load_manifest_lazy( + self.fbank_dir / "wenetspeech" / "cuts_TEST_NET.jsonl.gz" + ) + wenetspeech_dev_cuts = load_manifest_lazy( + self.fbank_dir / "wenetspeech" / "cuts_DEV.jsonl.gz" + ) + + return { + "aishell-2_test": aishell2_test_cuts, + "aishell-4": aishell4_test_cuts, + "magicdata_test": magicdata_test_cuts, + "kespeech-asr_test": kespeech_test_cuts, + } + + # return { + # "alimeeting_test": alimeeting_test_cuts, + # "alimeeting_eval": alimeeting_eval_cuts, + # "aishell_test": aishell_test_cuts, + # "aishell_dev": aishell_dev_cuts, + # "aishell-2_test": aishell2_test_cuts, + # "aishell-2_dev": aishell2_dev_cuts, + # "aishell-4": aishell4_test_cuts, + # "magicdata_test": magicdata_test_cuts, + # "magicdata_dev": magicdata_dev_cuts, + # "kespeech-asr_test": kespeech_test_cuts, + # "kespeech-asr_dev_phase1": kespeech_dev_phase1_cuts, + # "kespeech-asr_dev_phase2": kespeech_dev_phase2_cuts, + # "wenetspeech-meeting_test": wenetspeech_test_meeting_cuts, + # "wenetspeech-net_test": wenetspeech_test_net_cuts, + # "wenetspeech_dev": wenetspeech_dev_cuts, + # } diff --git a/egs/multi_zh-hans/ASR/whisper/optim.py b/egs/multi_zh-hans/ASR/whisper/optim.py new file mode 120000 index 000000000..5eaa3cffd --- /dev/null +++ b/egs/multi_zh-hans/ASR/whisper/optim.py @@ -0,0 +1 @@ +../../../librispeech/ASR/zipformer/optim.py \ No newline at end of file diff --git a/egs/multi_zh-hans/ASR/whisper/requirements.txt b/egs/multi_zh-hans/ASR/whisper/requirements.txt new file mode 120000 index 000000000..744bf8bb6 --- /dev/null +++ b/egs/multi_zh-hans/ASR/whisper/requirements.txt @@ -0,0 +1 @@ +../../../aishell/ASR/whisper/requirements.txt \ No newline at end of file diff --git a/egs/multi_zh-hans/ASR/whisper/train.py b/egs/multi_zh-hans/ASR/whisper/train.py new file mode 100644 index 000000000..11a22eec1 --- /dev/null +++ b/egs/multi_zh-hans/ASR/whisper/train.py @@ -0,0 +1,983 @@ +#!/usr/bin/env python3 +# Copyright 2023 Xiaomi Corp. (authors: Xiaoyu Yang) +# 2024 Yuekai Zhang +# +# 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: + +#fine-tuning with deepspeed zero stage 1 +torchrun --nproc-per-node 8 ./whisper/train.py \ + --max-duration 200 \ + --exp-dir whisper/exp_large_v2 \ + --model-name large-v2 \ + --deepspeed \ + --deepspeed_config ./whisper/ds_config_zero1.json + +# fine-tuning with ddp +torchrun --nproc_per_node 8 ./whisper/train.py \ + --max-duration 200 \ + --exp-dir whisper/exp_medium \ + --base-lr 1e-5 \ + --model-name medium +""" + +import argparse +import copy +import logging +import os +import random +import warnings +from pathlib import Path +from shutil import copyfile +from typing import Any, Dict, List, Optional, Tuple, Union + +import deepspeed +import k2 +import optim +import torch +import torch.multiprocessing as mp +import torch.nn as nn +import whisper +from asr_datamodule import AsrDataModule +from deepspeed.utils.zero_to_fp32 import convert_zero_checkpoint_to_fp32_state_dict +from label_smoothing import LabelSmoothingLoss +from lhotse import CutSet, load_manifest +from lhotse.cut import Cut +from lhotse.dataset.sampling.base import CutSampler +from lhotse.utils import fix_random_seed +from multi_dataset import MultiDataset +from optim import Eden, ScaledAdam +from torch import Tensor +from torch.cuda.amp import GradScaler +from torch.nn.functional import pad as pad_tensor +from torch.nn.parallel import DistributedDataParallel as DDP +from torch.utils.tensorboard import SummaryWriter +from whisper_encoder_forward_monkey_patch import replace_whisper_encoder_forward + +from icefall import diagnostics +from icefall.checkpoint import load_checkpoint, remove_checkpoints +from icefall.checkpoint import save_checkpoint as save_checkpoint_impl +from icefall.checkpoint import update_averaged_model +from icefall.dist import cleanup_dist, get_rank, get_world_size, setup_dist +from icefall.env import get_env_info +from icefall.hooks import register_inf_check_hooks +from icefall.utils import ( + AttributeDict, + MetricsTracker, + filter_uneven_sized_batch, + setup_logger, + str2bool, +) + +LRSchedulerType = Union[torch.optim.lr_scheduler._LRScheduler, optim.LRScheduler] + + +def set_batch_count(model: Union[nn.Module, DDP], batch_count: float) -> None: + if isinstance(model, DDP): + # get underlying nn.Module + model = model.module + for module in model.modules(): + if hasattr(module, "batch_count"): + module.batch_count = batch_count + + +def get_parser(): + parser = argparse.ArgumentParser( + formatter_class=argparse.ArgumentDefaultsHelpFormatter + ) + + parser.add_argument( + "--tensorboard", + type=str2bool, + default=True, + help="Should various information be logged in tensorboard.", + ) + + parser.add_argument( + "--num-epochs", + type=int, + default=10, + help="Number of epochs to train.", + ) + + parser.add_argument( + "--start-epoch", + type=int, + default=1, + help="""Resume training from this epoch. It should be positive. + If larger than 1, it will load checkpoint from + exp-dir/epoch-{start_epoch-1}.pt + """, + ) + + parser.add_argument( + "--start-batch", + type=int, + default=0, + help="""If positive, --start-epoch is ignored and + it loads the checkpoint from exp-dir/checkpoint-{start_batch}.pt + """, + ) + + parser.add_argument( + "--exp-dir", + type=str, + default="whisper/exp", + help="""The experiment dir. + It specifies the directory where all training related + files, e.g., checkpoints, log, etc, are saved + """, + ) + + parser.add_argument( + "--model-name", + type=str, + default="large-v2", + choices=["large-v2", "large-v3", "medium", "small", "tiny"], + help="""The model name to use. + """, + ) + + parser.add_argument( + "--pretrained-model-path", + type=str, + default=None, + help="""The path to the pretrained model if it is not None. Training will + start from this model. e.g. ./wenetspeech/ASR/whisper/exp_large_v2/epoch-4-avg-3.pt + """, + ) + + parser.add_argument( + "--base-lr", type=float, default=1e-5, help="The base learning rate." + ) + + 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.""", + ) + + parser.add_argument( + "--lr-epochs", + type=float, + default=6, + help="""Number of epochs that affects how rapidly the learning rate decreases. + """, + ) + + parser.add_argument( + "--seed", + type=int, + default=42, + help="The seed for random generators intended for reproducibility", + ) + + parser.add_argument( + "--print-diagnostics", + type=str2bool, + default=False, + help="Accumulate stats on activations, print them and exit.", + ) + + parser.add_argument( + "--inf-check", + type=str2bool, + default=False, + help="Add hooks to check for infinite module outputs and gradients.", + ) + + parser.add_argument( + "--keep-last-k", + type=int, + default=30, + help="""Only keep this number of checkpoints on disk. + For instance, if it is 3, there are only 3 checkpoints + in the exp-dir with filenames `checkpoint-xxx.pt`. + It does not affect checkpoints with name `epoch-xxx.pt`. + """, + ) + + parser.add_argument( + "--average-period", + type=int, + default=200, + help="""Update the averaged model, namely `model_avg`, after processing + this number of batches. `model_avg` is a separate version of model, + in which each floating-point parameter is the average of all the + parameters from the start of training. Each time we take the average, + we do: `model_avg = model * (average_period / batch_idx_train) + + model_avg * ((batch_idx_train - average_period) / batch_idx_train)`. + """, + ) + + parser.add_argument( + "--use-fp16", + type=str2bool, + default=True, + help="Whether to use half precision training.", + ) + + parser = deepspeed.add_config_arguments(parser) + + return parser + + +def get_params() -> AttributeDict: + """Return a dict containing training parameters. + + All training related parameters that are not passed from the commandline + are saved in the variable `params`. + + Commandline options are merged into `params` after they are parsed, so + you can also access them via `params`. + + Explanation of options saved in `params`: + + - frame_shift_ms: The frame shift in milliseconds. + - allowed_excess_duration_ratio: The allowed excess duration ratio. + - best_train_loss: The best training loss so far. + - best_valid_loss: The best validation loss so far. + - best_train_epoch: The epoch where the best training loss is achieved. + - best_valid_epoch: The epoch where the best validation loss is achieved. + - batch_idx_train: The batch index of the current batch. + - log_interval: Log training stats every `log_interval` batches. + - reset_interval: Reset the stats every `reset_interval` batches. + - valid_interval: Run validation every `valid_interval` batches. + - env_info: The environment information. + """ + params = AttributeDict( + { + "frame_shift_ms": 10.0, + "subsampling_factor": 2, + "allowed_excess_duration_ratio": 0.1, + "best_train_loss": float("inf"), + "best_valid_loss": float("inf"), + "best_train_epoch": -1, + "best_valid_epoch": -1, + "batch_idx_train": 0, + "log_interval": 50, + "reset_interval": 200, + "valid_interval": 10000, + "env_info": get_env_info(), + } + ) + + return params + + +def load_checkpoint_if_available( + params: AttributeDict, + model: nn.Module, + model_avg: nn.Module = None, + optimizer: Optional[torch.optim.Optimizer] = None, + scheduler: Optional[LRSchedulerType] = None, +) -> Optional[Dict[str, Any]]: + """Load checkpoint from file. + + If params.start_batch is positive, it will load the checkpoint from + `params.exp_dir/checkpoint-{params.start_batch}.pt`. Otherwise, if + params.start_epoch is larger than 1, it will load the checkpoint from + `params.start_epoch - 1`. + + Apart from loading state dict for `model` and `optimizer` it also updates + `best_train_epoch`, `best_train_loss`, `best_valid_epoch`, + and `best_valid_loss` in `params`. + + Args: + params: + The return value of :func:`get_params`. + model: + The training model. + model_avg: + The stored model averaged from the start of training. + optimizer: + The optimizer that we are using. + scheduler: + The scheduler that we are using. + Returns: + Return a dict containing previously saved training info. + """ + if params.start_batch > 0: + filename = params.exp_dir / f"checkpoint-{params.start_batch}.pt" + elif params.start_epoch > 1: + filename = params.exp_dir / f"epoch-{params.start_epoch-1}.pt" + else: + return None + + assert filename.is_file(), f"{filename} does not exist!" + + saved_params = load_checkpoint( + filename, + model=model, + model_avg=model_avg, + optimizer=optimizer, + scheduler=scheduler, + ) + + keys = [ + "best_train_epoch", + "best_valid_epoch", + "batch_idx_train", + "best_train_loss", + "best_valid_loss", + ] + for k in keys: + params[k] = saved_params[k] + + if params.start_batch > 0: + if "cur_epoch" in saved_params: + params["start_epoch"] = saved_params["cur_epoch"] + + return saved_params + + +def save_checkpoint( + params: AttributeDict, + model: Union[nn.Module, DDP], + model_avg: Optional[nn.Module] = None, + optimizer: Optional[torch.optim.Optimizer] = None, + scheduler: Optional[LRSchedulerType] = None, + sampler: Optional[CutSampler] = None, + scaler: Optional[GradScaler] = None, + rank: int = 0, +) -> None: + """Save model, optimizer, scheduler and training stats to file. + + Args: + params: + It is returned by :func:`get_params`. + model: + The training model. + model_avg: + The stored model averaged from the start of training. + optimizer: + The optimizer used in the training. + sampler: + The sampler for the training dataset. + scaler: + The scaler used for mix precision training. + """ + if rank != 0: + return + filename = params.exp_dir / f"epoch-{params.cur_epoch}.pt" + save_checkpoint_impl( + filename=filename, + model=model, + model_avg=model_avg, + params=params, + optimizer=optimizer, + scheduler=scheduler, + sampler=sampler, + scaler=scaler, + rank=rank, + ) + + if params.best_train_epoch == params.cur_epoch: + best_train_filename = params.exp_dir / "best-train-loss.pt" + copyfile(src=filename, dst=best_train_filename) + + if params.best_valid_epoch == params.cur_epoch: + best_valid_filename = params.exp_dir / "best-valid-loss.pt" + copyfile(src=filename, dst=best_valid_filename) + + +def compute_loss( + params: AttributeDict, + tokenizer: whisper.tokenizer.Tokenizer, + model: Union[nn.Module, DDP], + batch: dict, + is_training: bool, +) -> Tuple[Tensor, MetricsTracker]: + """ + Compute the loss for the given batch. + Args: + params: + It is returned by :func:`get_params`. + tokenizer: + The tokenizer used to encode the text. + model: + The model for training. + batch: + A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()` + for the content in it. + is_training: + Whether it is training. + Returns: + Return a tuple of two elements. The first element is the loss tensor. + """ + # For the uneven-sized batch, the total duration after padding would possibly + # cause OOM. Hence, for each batch, which is sorted descendingly by length, + # we simply drop the last few shortest samples, so that the retained total frames + # (after padding) would not exceed `allowed_max_frames`: + # `allowed_max_frames = int(max_frames * (1.0 + allowed_excess_duration_ratio))`, + # where `max_frames = max_duration * 1000 // frame_shift_ms`. + # We set allowed_excess_duration_ratio=0.1. + if isinstance(model, DDP): + # get underlying nn.Module + model = model.module + + def _batch_tensors(tensors: List[Tensor], pad_value: Any) -> Tensor: + padding_size = max(tensor.shape[0] for tensor in tensors) + dims = len(tensors[0].shape) + padded_tensors = [] + for tensor in tensors: + padding = [0] * 2 * dims + padding[-1] = padding_size - tensor.shape[0] + padded_tensors.append(pad_tensor(tensor, padding, "constant", pad_value)) + return torch.stack([tensor for tensor in padded_tensors], dim=0) + + max_frames = params.max_duration * 1000 // params.frame_shift_ms + allowed_max_frames = int(max_frames * (1.0 + params.allowed_excess_duration_ratio)) + batch = filter_uneven_sized_batch(batch, allowed_max_frames) + + device = model.device if isinstance(model, DDP) else next(model.parameters()).device + feature = batch["inputs"] + + assert feature.ndim == 3 + feature = feature.to(device) + feature = feature.transpose(1, 2) # (N, C, T) + + supervisions = batch["supervisions"] + feature_lens = supervisions["num_frames"].to(device) + + batch_idx_train = params.batch_idx_train + + texts = batch["supervisions"]["text"] + # remove spaces in texts + texts = [text.replace(" ", "") for text in texts] + + text_tokens_list = [ + list(tokenizer.sot_sequence_including_notimestamps) + + tokenizer.encode(text) + + [tokenizer.eot] + for text in texts + ] + # convert it to torch tensor + text_tokens_list = [ + torch.LongTensor(text_tokens) for text_tokens in text_tokens_list + ] + + # 50256 is the index of for all whisper models + prev_outputs_tokens = _batch_tensors( + [tokens[:-1] for tokens in text_tokens_list], pad_value=50256 + ) + target_tokens = _batch_tensors( + [tokens[1:] for tokens in text_tokens_list], pad_value=50256 + ) + target_lengths = torch.LongTensor( + [tokens.shape[0] - 1 for tokens in text_tokens_list] + ) + + decoder_criterion = LabelSmoothingLoss( + ignore_index=50256, label_smoothing=0.1, reduction="sum" + ) + + # ignore the first 3 tokens, which are always <|lang_id|>, <|transcibe|>, <|notimestampes|> + ignore_prefix_size = 3 + with torch.set_grad_enabled(is_training): + encoder_out = model.encoder(feature) + text_logits = model.decoder(prev_outputs_tokens.to(device), encoder_out) + text_logits = text_logits[:, ignore_prefix_size:, :] + target_tokens = target_tokens[:, ignore_prefix_size:] + loss = decoder_criterion(text_logits, target_tokens.to(device)) + + assert loss.requires_grad == is_training + + info = MetricsTracker() + with warnings.catch_warnings(): + warnings.simplefilter("ignore") + info["frames"] = (feature_lens // params.subsampling_factor).sum().item() + + # Note: We use reduction=sum while computing the loss. + info["loss"] = loss.detach().cpu().item() + + return loss, info + + +def compute_validation_loss( + params: AttributeDict, + tokenizer: whisper.tokenizer.Tokenizer, + model: Union[nn.Module, DDP], + valid_dl: torch.utils.data.DataLoader, + world_size: int = 1, +) -> MetricsTracker: + """Run the validation process.""" + model.eval() + + tot_loss = MetricsTracker() + + for batch_idx, batch in enumerate(valid_dl): + with torch.cuda.amp.autocast(enabled=params.use_fp16): + loss, loss_info = compute_loss( + params=params, + tokenizer=tokenizer, + model=model, + batch=batch, + is_training=False, + ) + assert loss.requires_grad is False + tot_loss = tot_loss + loss_info + + if world_size > 1: + tot_loss.reduce(loss.device) + + loss_value = tot_loss["loss"] / tot_loss["frames"] + if loss_value < params.best_valid_loss: + params.best_valid_epoch = params.cur_epoch + params.best_valid_loss = loss_value + + return tot_loss + + +def train_one_epoch( + params: AttributeDict, + tokenizer: whisper.tokenizer.Tokenizer, + model: Union[nn.Module, DDP], + optimizer: torch.optim.Optimizer, + scheduler: LRSchedulerType, + train_dl: torch.utils.data.DataLoader, + valid_dl: torch.utils.data.DataLoader, + scaler: GradScaler, + model_avg: Optional[nn.Module] = None, + tb_writer: Optional[SummaryWriter] = None, + world_size: int = 1, + rank: int = 0, +) -> None: + """Train the model for one epoch. + + The training loss from the mean of all frames is saved in + `params.train_loss`. It runs the validation process every + `params.valid_interval` batches. + + Args: + params: + It is returned by :func:`get_params`. + model: + The model for training. + optimizer: + The optimizer we are using. + scheduler: + The learning rate scheduler, we call step() every step. + train_dl: + Dataloader for the training dataset. + valid_dl: + Dataloader for the validation dataset. + scaler: + The scaler used for mix precision training. + model_avg: + The stored model averaged from the start of training. + tb_writer: + Writer to write log messages to tensorboard. + world_size: + Number of nodes in DDP training. If it is 1, DDP is disabled. + rank: + The rank of the node in DDP training. If no DDP is used, it should + be set to 0. + """ + model.train() + + tot_loss = MetricsTracker() + + for batch_idx, batch in enumerate(train_dl): + params.batch_idx_train += 1 + batch_size = len(batch["supervisions"]["text"]) + if batch_idx % params.valid_interval == 0 and not params.print_diagnostics: + logging.info("Computing validation loss") + valid_info = compute_validation_loss( + params=params, + tokenizer=tokenizer, + model=model, + valid_dl=valid_dl, + world_size=world_size, + ) + model.train() + logging.info(f"Epoch {params.cur_epoch}, validation: {valid_info}") + logging.info( + f"Maximum memory allocated so far is {torch.cuda.max_memory_allocated()//1000000}MB" + ) + if tb_writer is not None: + valid_info.write_summary( + tb_writer, "train/valid_", params.batch_idx_train + ) + if params.deepspeed: + model.save_checkpoint( + save_dir=params.exp_dir, + tag=f"epoch-{params.cur_epoch}-checkpoint-{batch_idx}", + client_state={}, + ) + if rank == 0: + convert_zero_checkpoint_to_fp32_state_dict( + params.exp_dir, + f"{params.exp_dir}/epoch-{params.cur_epoch}-checkpoint-{batch_idx}.pt", + tag=f"epoch-{params.cur_epoch}-checkpoint-{batch_idx}", + ) + os.system( + f"rm -rf {params.exp_dir}/epoch-{params.cur_epoch}-checkpoint-{batch_idx}" + ) + + try: + with torch.cuda.amp.autocast(enabled=params.use_fp16): + loss, loss_info = compute_loss( + params=params, + tokenizer=tokenizer, + model=model, + batch=batch, + is_training=True, + ) + # 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. + if params.deepspeed: + # deepspeed's backward() is different from torch's backward() + # in that it does not accept a loss tensor as input. + # It computes the loss internally. + model.backward(loss) + model.step() + else: + scaler.scale(loss).backward() + set_batch_count(model, params.batch_idx_train) + scheduler.step_batch(params.batch_idx_train) + + scaler.step(optimizer) + scaler.update() + optimizer.zero_grad() + except: # noqa + display_and_save_batch(batch, params=params) + raise + + if params.print_diagnostics and batch_idx == 5: + return + + if ( + rank == 0 + and params.batch_idx_train > 0 + and params.batch_idx_train % params.average_period == 0 + and not params.deepspeed + ): + update_averaged_model( + params=params, + model_cur=model, + model_avg=model_avg, + ) + + if batch_idx % 100 == 0 and params.use_fp16 and not params.deepspeed: + # If the grad scale was less than 1, try increasing it. The _growth_interval + # of the grad scaler is configurable, but we can't configure it to have different + # behavior depending on the current grad scale. + cur_grad_scale = scaler._scale.item() + if cur_grad_scale < 1.0 or (cur_grad_scale < 8.0 and batch_idx % 400 == 0): + scaler.update(cur_grad_scale * 2.0) + if cur_grad_scale < 0.01: + logging.warning(f"Grad scale is small: {cur_grad_scale}") + if cur_grad_scale < 1.0e-05: + raise RuntimeError( + f"grad_scale is too small, exiting: {cur_grad_scale}" + ) + if batch_idx % params.log_interval == 0: + try: + cur_lr = scheduler.get_last_lr()[0] + except: # noqa + cur_lr = 0.0 + cur_grad_scale = ( + scaler._scale.item() + if (params.use_fp16 and not params.deepspeed) + else 1.0 + ) + + logging.info( + f"Epoch {params.cur_epoch}, " + f"batch {batch_idx}, loss[{loss_info}], " + f"tot_loss[{tot_loss}], batch size: {batch_size}, " + f"lr: {cur_lr:.2e}, " + + ( + f"grad_scale: {scaler._scale.item()}" + if (params.use_fp16 and not params.deepspeed) + else "" + ) + ) + + if tb_writer is not None: + tb_writer.add_scalar( + "train/learning_rate", cur_lr, params.batch_idx_train + ) + + loss_info.write_summary( + tb_writer, "train/current_", params.batch_idx_train + ) + tot_loss.write_summary(tb_writer, "train/tot_", params.batch_idx_train) + if params.use_fp16: + tb_writer.add_scalar( + "train/grad_scale", + cur_grad_scale, + params.batch_idx_train, + ) + + loss_value = tot_loss["loss"] / tot_loss["frames"] + params.train_loss = loss_value + if params.train_loss < params.best_train_loss: + params.best_train_epoch = params.cur_epoch + params.best_train_loss = params.train_loss + + +def run(rank, world_size, args): + """ + Args: + rank: + It is a value between 0 and `world_size-1`, which is + passed automatically by `mp.spawn()` in :func:`main`. + The node with rank 0 is responsible for saving checkpoint. + world_size: + Number of GPUs for DDP training. + args: + The return value of get_parser().parse_args() + """ + params = get_params() + params.update(vars(args)) + + fix_random_seed(params.seed) + + setup_logger(f"{params.exp_dir}/log/log-train") + logging.info(params) + + logging.info("About to create model") + + replace_whisper_encoder_forward() + model = whisper.load_model(params.model_name, "cpu") + del model.alignment_heads + + if params.pretrained_model_path: + checkpoint = torch.load(params.pretrained_model_path, map_location="cpu") + if "model" not in checkpoint: + model.load_state_dict(checkpoint, strict=True) + else: + load_checkpoint(params.pretrained_model_path, model) + + num_param = sum([p.numel() for p in model.parameters()]) + logging.info(f"Number of model parameters: {num_param}") + + tokenizer = whisper.tokenizer.get_tokenizer( + model.is_multilingual, + num_languages=model.num_languages, + language="zh", + task="transcribe", + ) + + model_avg: Optional[nn.Module] = None + if rank == 0: + # model_avg is only used with rank 0 + model_avg = copy.deepcopy(model).to(torch.float64) + + assert params.start_epoch > 0, params.start_epoch + checkpoints = load_checkpoint_if_available( + params=params, model=model, model_avg=model_avg + ) + + if torch.cuda.is_available(): + device = torch.device("cuda", rank) + else: + device = torch.device("cpu") + logging.info(f"Device: {device}") + model.to(device) + + optimizer = torch.optim.AdamW(model.parameters(), lr=params.base_lr) + scheduler = Eden(optimizer, params.lr_batches, params.lr_epochs) + + if checkpoints and "optimizer" in checkpoints: + logging.info("Loading optimizer state dict") + optimizer.load_state_dict(checkpoints["optimizer"]) + + if ( + checkpoints + and "scheduler" in checkpoints + and checkpoints["scheduler"] is not None + ): + logging.info("Loading scheduler state dict") + scheduler.load_state_dict(checkpoints["scheduler"]) + + if world_size > 1: + if params.deepspeed: + logging.info("Using DeepSpeed") + model, optimizer, _, scheduler = deepspeed.initialize( + args=params, model=model, model_parameters=model.parameters() + ) + else: + logging.info("Using DDP") + setup_dist(use_ddp_launch=True) + model = DDP(model, device_ids=[rank], find_unused_parameters=True) + + if params.print_diagnostics: + opts = diagnostics.TensorDiagnosticOptions( + 2**22 + ) # allow 4 megabytes per sub-module + diagnostic = diagnostics.attach_diagnostics(model, opts) + + if params.inf_check: + register_inf_check_hooks(model) + + data_module = AsrDataModule(args) + multi_dataset = MultiDataset(args.manifest_dir) + + if params.start_batch > 0 and checkpoints and "sampler" in checkpoints: + # We only load the sampler's state dict when it loads a checkpoint + # saved in the middle of an epoch + sampler_state_dict = checkpoints["sampler"] + else: + sampler_state_dict = None + + def remove_short_and_long_utt(c: Cut): + # Keep only utterances with duration between 1 second and 20 seconds + # + # Caution: There is a reason to select 20.0 here. Please see + # ../local/display_manifest_statistics.py + # + # You should use ../local/display_manifest_statistics.py to get + # an utterance duration distribution for your dataset to select + # the threshold + if c.duration < 1.0 or c.duration > 20.0: + # logging.warning( + # f"Exclude cut with ID {c.id} from training. Duration: {c.duration}" + # ) + return False + return True + + train_cuts = multi_dataset.train_cuts() + train_cuts = train_cuts.filter(remove_short_and_long_utt) + + train_dl = data_module.train_dataloaders( + train_cuts, sampler_state_dict=sampler_state_dict + ) + + valid_cuts = multi_dataset.dev_cuts() + valid_dl = data_module.valid_dataloaders(valid_cuts) + + scaler = GradScaler(enabled=params.use_fp16, init_scale=1.0) + if checkpoints and "grad_scaler" in checkpoints: + logging.info("Loading grad scaler state dict") + scaler.load_state_dict(checkpoints["grad_scaler"]) + + if args.tensorboard and rank == 0: + tb_writer = SummaryWriter(log_dir=f"{params.exp_dir}/tensorboard") + else: + tb_writer = None + + logging.info(f"start training from epoch {params.start_epoch}") + for epoch in range(params.start_epoch, params.num_epochs + 1): + if not params.deepspeed: + scheduler.step_epoch(epoch - 1) + fix_random_seed(params.seed + epoch - 1) + train_dl.sampler.set_epoch(epoch - 1) + + if tb_writer is not None: + tb_writer.add_scalar("train/epoch", epoch, params.batch_idx_train) + + params.cur_epoch = epoch + + train_one_epoch( + params=params, + tokenizer=tokenizer, + model=model, + model_avg=model_avg, + optimizer=optimizer, + scheduler=scheduler, + train_dl=train_dl, + valid_dl=valid_dl, + scaler=scaler, + tb_writer=tb_writer, + world_size=world_size, + rank=rank, + ) + + if params.print_diagnostics: + diagnostic.print_diagnostics() + break + + if params.deepspeed: + model.save_checkpoint( + save_dir=params.exp_dir, + tag=f"epoch-{params.cur_epoch}", + client_state={}, + ) + if rank == 0: + convert_zero_checkpoint_to_fp32_state_dict( + params.exp_dir, + f"{params.exp_dir}/epoch-{params.cur_epoch}.pt", + tag=f"epoch-{params.cur_epoch}", + ) + os.system(f"rm -rf {params.exp_dir}/epoch-{params.cur_epoch}") + else: + save_checkpoint( + params=params, + model=model, + model_avg=model_avg, + optimizer=optimizer, + scheduler=scheduler, + sampler=train_dl.sampler, + scaler=scaler, + rank=rank, + ) + + logging.info("Done!") + + if world_size > 1 and not params.deepspeed: + torch.distributed.barrier() + cleanup_dist() + + +def display_and_save_batch( + batch: dict, + params: AttributeDict, +) -> 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`. + """ + 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}") + + +def main(): + parser = get_parser() + AsrDataModule.add_arguments(parser) + args = parser.parse_args() + args.exp_dir = Path(args.exp_dir) + + world_size = get_world_size() + rank = get_rank() + + torch.set_num_threads(1) + torch.set_num_interop_threads(1) + run(rank=rank, world_size=world_size, args=args) + + +if __name__ == "__main__": + main() diff --git a/egs/multi_zh-hans/ASR/whisper/whisper_encoder_forward_monkey_patch.py b/egs/multi_zh-hans/ASR/whisper/whisper_encoder_forward_monkey_patch.py new file mode 120000 index 000000000..2a7808921 --- /dev/null +++ b/egs/multi_zh-hans/ASR/whisper/whisper_encoder_forward_monkey_patch.py @@ -0,0 +1 @@ +../../../aishell/ASR/whisper/whisper_encoder_forward_monkey_patch.py \ No newline at end of file diff --git a/egs/multi_zh-hans/ASR/zipformer/train.py b/egs/multi_zh-hans/ASR/zipformer/train.py index c1bbd2ee8..447ca122f 100755 --- a/egs/multi_zh-hans/ASR/zipformer/train.py +++ b/egs/multi_zh-hans/ASR/zipformer/train.py @@ -89,6 +89,7 @@ from icefall.checkpoint import ( ) from icefall.dist import cleanup_dist, setup_dist from icefall.env import get_env_info +from icefall.err import raise_grad_scale_is_too_small_error from icefall.hooks import register_inf_check_hooks from icefall.utils import ( AttributeDict, @@ -1020,9 +1021,7 @@ def train_one_epoch( logging.warning(f"Grad scale is small: {cur_grad_scale}") if cur_grad_scale < 1.0e-05: save_bad_model() - raise RuntimeError( - f"grad_scale is too small, exiting: {cur_grad_scale}" - ) + raise_grad_scale_is_too_small_error(cur_grad_scale) if batch_idx % params.log_interval == 0: cur_lr = max(scheduler.get_last_lr()) diff --git a/egs/multi_zh_en/ASR/zipformer/asr_datamodule.py b/egs/multi_zh_en/ASR/zipformer/asr_datamodule.py index 662ae01c5..489b38e65 100644 --- a/egs/multi_zh_en/ASR/zipformer/asr_datamodule.py +++ b/egs/multi_zh_en/ASR/zipformer/asr_datamodule.py @@ -216,7 +216,7 @@ class AsrDataModule: logging.info("About to get Musan cuts") cuts_musan = load_manifest(self.args.manifest_dir / "musan_cuts.jsonl.gz") transforms.append( - CutMix(cuts=cuts_musan, prob=0.5, snr=(10, 20), preserve_id=True) + CutMix(cuts=cuts_musan, p=0.5, snr=(10, 20), preserve_id=True) ) else: logging.info("Disable MUSAN") diff --git a/egs/multi_zh_en/ASR/zipformer/train.py b/egs/multi_zh_en/ASR/zipformer/train.py index 310c8fe59..5dba584f7 100755 --- a/egs/multi_zh_en/ASR/zipformer/train.py +++ b/egs/multi_zh_en/ASR/zipformer/train.py @@ -89,6 +89,7 @@ from icefall.checkpoint import ( ) from icefall.dist import cleanup_dist, setup_dist from icefall.env import get_env_info +from icefall.err import raise_grad_scale_is_too_small_error from icefall.hooks import register_inf_check_hooks from icefall.utils import ( AttributeDict, @@ -1042,9 +1043,7 @@ def train_one_epoch( logging.warning(f"Grad scale is small: {cur_grad_scale}") if cur_grad_scale < 1.0e-05: save_bad_model() - raise RuntimeError( - f"grad_scale is too small, exiting: {cur_grad_scale}" - ) + raise_grad_scale_is_too_small_error(cur_grad_scale) if batch_idx % params.log_interval == 0: cur_lr = max(scheduler.get_last_lr()) diff --git a/egs/must_c/ST/local/get_text.py b/egs/must_c/ST/local/get_text.py index 558ab6de8..f7b5816a8 100755 --- a/egs/must_c/ST/local/get_text.py +++ b/egs/must_c/ST/local/get_text.py @@ -5,9 +5,9 @@ This file prints the text field of supervisions from cutset to the console """ import argparse +from pathlib import Path from lhotse import load_manifest_lazy -from pathlib import Path def get_args(): diff --git a/egs/must_c/ST/local/get_words.py b/egs/must_c/ST/local/get_words.py index a61f60860..b32925099 100755 --- a/egs/must_c/ST/local/get_words.py +++ b/egs/must_c/ST/local/get_words.py @@ -5,7 +5,6 @@ This file generates words.txt from the given transcript file. """ import argparse - from pathlib import Path diff --git a/egs/speechio/ASR/README.md b/egs/speechio/ASR/README.md new file mode 100644 index 000000000..2675efd9b --- /dev/null +++ b/egs/speechio/ASR/README.md @@ -0,0 +1,15 @@ + +# Introduction + +This recipe includes some different pretrained ASR models' decoding results with [SpeechIO](https://github.com/SpeechColab/Leaderboard) test sets. + +[./RESULTS.md](./RESULTS.md) contains the latest results. + +# Pretrained Models + +The following table lists the pretrained models. + +| | Huggingface | Comment | +|---------------------------------------|--------------------|-----------------------------| +| `zipformer` | zrjin/icefall-asr-multi-zh-hans-zipformer-ctc-2023-10-24 | Using [multi-hans-zh recipe](../../multi_zh-hans/ASR/zipformer/) training | | +| `whisper` | yuekai/icefall_asr_wenetspeech_whisper | Using [wenetspeech recipe](../../wenetspeech/ASR/whisper/) training | diff --git a/egs/speechio/ASR/RESULTS.md b/egs/speechio/ASR/RESULTS.md new file mode 100644 index 000000000..07649e383 --- /dev/null +++ b/egs/speechio/ASR/RESULTS.md @@ -0,0 +1,92 @@ +## Results + +### SpeechIO Test Set Decoding Results + +##### Decoding results using pretrained [multi-hans-zh zipformer](https://huggingface.co/zrjin/icefall-asr-multi-zh-hans-zipformer-ctc-2023-10-24), [whipser-large-v2](https://github.com/openai/whisper/blob/main/whisper/__init__.py#L27), [whisper-large-v2-wenetspeech-ft](https://huggingface.co/yuekai/icefall_asr_wenetspeech_whisper). + +| | zipformer_transducer | zipformer_transducer_blank_penalty_2 | whisper_large_v2 | whisper_large_v2_wenetspeech | whisper_large_v2_wenetspeech_zipformer_fusion | +|------------------------|----------------------|--------------------------------------|------------------|------------------------------|-----------------------------------------------| +| SPEECHIO_ASR_ZH00000 | 10.04 | 8.04 | 11.4 | 9.88 | 7.78 | +| SPEECHIO_ASR_ZH00001 | 1.67 | 1.51 | 2.49 | 1.57 | 1.38 | +| SPEECHIO_ASR_ZH00002 | 5.89 | 5.27 | 7.89 | 5.65 | 4.99 | +| SPEECHIO_ASR_ZH00003 | 2.66 | 2.79 | 5.94 | 2.27 | 2.33 | +| SPEECHIO_ASR_ZH00004 | 3.6 | 3.34 | 4.57 | 3.62 | 3.26 | +| SPEECHIO_ASR_ZH00005 | 7.54 | 5.81 | 8.39 | 7.26 | 5.43 | +| SPEECHIO_ASR_ZH00006 | 15.59 | 13.34 | 19.07 | 13.64 | 11.96 | +| SPEECHIO_ASR_ZH00007 | 15.9 | 15.05 | 16.7 | 14.06 | 13.73 | +| SPEECHIO_ASR_ZH00008 | 11.07 | 9.68 | 14.69 | 10.34 | 8.87 | +| SPEECHIO_ASR_ZH00009 | 7.38 | 6.23 | 8.32 | 6.74 | 5.96 | +| SPEECHIO_ASR_ZH00010 | 9.19 | 7.33 | 11.2 | 8.85 | 6.97 | +| SPEECHIO_ASR_ZH00011 | 4.16 | 3.84 | 54.56 | 4.09 | 3.72 | +| SPEECHIO_ASR_ZH00012 | 7.61 | 6.58 | 10.53 | 8.35 | 6.27 | +| SPEECHIO_ASR_ZH00013 | 8.72 | 7.66 | 9.32 | 7.26 | 6.7 | +| SPEECHIO_ASR_ZH00014 | 9.69 | 8.71 | 9.03 | 7.03 | 6.59 | +| SPEECHIO_ASR_ZH00015 | 11.94 | 11.37 | 16.58 | 12.02 | 11.11 | +| SPEECHIO_ASR_ZH00016 | 9.79 | 8.79 | 14.1 | 10.19 | 8.15 | +| SPEECHIO_ASR_ZH00017 | 8 | 6.72 | 9.04 | 8.9 | 6.44 | +| SPEECHIO_ASR_ZH00018 | 5.42 | 5.02 | 6.06 | 4.86 | 4.4 | +| SPEECHIO_ASR_ZH00019 | 11.26 | 9.06 | 14.8 | 9.83 | 8.22 | +| SPEECHIO_ASR_ZH00020 | 4.37 | 4.23 | 5.97 | 4.23 | 4.13 | +| SPEECHIO_ASR_ZH00021 | 7.81 | 6.34 | 8.53 | 7.08 | 5.88 | +| SPEECHIO_ASR_ZH00022 | 9.11 | 8.54 | 9.7 | 8.97 | 8.02 | +| SPEECHIO_ASR_ZH00023 | 9.98 | 8.98 | 6.31 | 9.44 | 8.57 | +| SPEECHIO_ASR_ZH00024 | 16.15 | 12.95 | 20.54 | 15.92 | 12.28 | +| SPEECHIO_ASR_ZH00025 | 10.38 | 9.82 | 11.4 | 10.26 | 9.27 | +| SPEECHIO_ASR_ZH00026 | 5.69 | 5.63 | 9.09 | 5.95 | 5.51 | +| Average WER (001-026) | 8.48 | 7.48 | 12.11 | 8.01 | 6.93 | + + + + +Command for decoding using fine-tuned whisper: +```bash +git lfs install +git clone https://huggingface.co/yuekai/icefall_asr_wenetspeech_whisper +ln -s icefall_asr_aishell_whisper/exp_large_v2/epoch-4-avg3.pt whisper/exp_large_v2_wenetspeech/epoch-999.pt + +python3 ./whisper/decode.py \ + --exp-dir whisper/exp_large_v2_wenetspeech \ + --model-name large-v2 \ + --epoch 999 --avg 1 \ + --start-index 0 --end-index 26 \ + --remove-whisper-encoder-input-length-restriction True \ + --manifest-dir data/fbank \ + --beam-size 1 --max-duration 50 +``` +Command for decoding using pretrained zipformer: +```bash +git lfs install +GIT_LFS_SKIP_SMUDGE=1 git clone https://huggingface.co/zrjin/icefall-asr-multi-zh-hans-zipformer-ctc-2023-10-24 +cd icefall-asr-multi-zh-hans-zipformer-ctc-2023-10-24 +git lfs pull --include "exp/pretrained.pt" +git lfs pull --include "data/lang_bpe_2000/*" +ln -s ../icefall-asr-multi-zh-hans-zipformer-ctc-2023-10-24/exp/pretrained.pt zipformer/exp_pretrain/epoch-999.pt +ln -s ../icefall-asr-multi-zh-hans-zipformer-ctc-2023-10-24/data/lang_bpe_2000/ ./data +wget https://huggingface.co/pkufool/icefall-asr-zipformer-wenetspeech-20230615/resolve/main/data/lang_char/words.txt +mv words.txt ./data/lang_bpe_2000/ + +./zipformer/decode.py \ + --epoch 999 \ + --avg 1 \ + --blank-penalty 2.0 \ + --use-averaged-model false \ + --exp-dir ./zipformer/exp_pretrain \ + --max-duration 600 \ + --start-index 0 --end-index 26 \ + --manifest-dir data/fbank_kaldi \ + --decoding-method greedy_search +``` +Command for fusion the above decoding results from whisper and zipformer: +```bash +python local/whisper_zipformer_fusion.py \ + --whisper-log-dir ./whisper/exp_large_v2_wenetspeech \ + --zipformer-log-dir ./zipformer/exp_pretrain/greedy_search \ + --output-log-dir ./results_fusion + +``` + +See why the fusion helps [here](./local/whisper_zipformer_fusion.py). + +SpeechIO fbank features, decoding scripts, logs, and decoding results +are available at + diff --git a/egs/speechio/ASR/local/compute_fbank_speechio.py b/egs/speechio/ASR/local/compute_fbank_speechio.py new file mode 100644 index 000000000..5b3489a9f --- /dev/null +++ b/egs/speechio/ASR/local/compute_fbank_speechio.py @@ -0,0 +1,148 @@ +#!/usr/bin/env python3 +# Copyright 2023 Xiaomi Corp. (authors: Fangjun Kuang +# Zengrui Jin) +# +# 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 file computes fbank features of the ST-CMDS dataset. +It looks for manifests in the directory data/manifests/stcmds. + +The generated fbank features are saved in data/fbank. +""" + +import argparse +import logging +import os +from pathlib import Path + +import torch +from lhotse import ( + CutSet, + Fbank, + FbankConfig, + LilcomChunkyWriter, + WhisperFbank, + WhisperFbankConfig, +) +from lhotse.recipes.utils import read_manifests_if_cached + +from icefall.utils import get_executor, str2bool + +# Torch's multithreaded behavior needs to be disabled or +# it wastes a lot of CPU and slow things down. +# Do this outside of main() in case it needs to take effect +# even when we are not invoking the main (e.g. when spawning subprocesses). +torch.set_num_threads(1) +torch.set_num_interop_threads(1) + +SPEECHIO_TESTSET_INDEX = 26 # Currently, from 0 - 26 test sets are open source. + + +def compute_fbank_speechio( + num_mel_bins: int = 80, + speed_perturb: bool = False, + fbank_dir: str = "data/fbank", + whisper_fbank: bool = False, +): + src_dir = Path("data/manifests") + output_dir = Path(fbank_dir) + num_jobs = min(8, os.cpu_count()) + + dataset_parts = [] + for i in range(SPEECHIO_TESTSET_INDEX + 1): + idx = f"{i}".zfill(2) + dataset_parts.append(f"SPEECHIO_ASR_ZH000{idx}") + + prefix = "speechio" + suffix = "jsonl.gz" + manifests = read_manifests_if_cached( + dataset_parts=dataset_parts, + output_dir=src_dir, + prefix=prefix, + suffix=suffix, + ) + assert manifests is not None + + assert len(manifests) == len(dataset_parts), ( + len(manifests), + len(dataset_parts), + list(manifests.keys()), + dataset_parts, + ) + + if whisper_fbank: + extractor = WhisperFbank( + WhisperFbankConfig(num_filters=args.num_mel_bins, device="cuda") + ) + else: + extractor = Fbank(FbankConfig(num_mel_bins=num_mel_bins)) + + with get_executor() as ex: # Initialize the executor only once. + for partition, m in manifests.items(): + if (output_dir / f"{prefix}_cuts_{partition}.{suffix}").is_file(): + logging.info(f"{partition} already exists - skipping.") + continue + logging.info(f"Processing {partition}") + cut_set = CutSet.from_manifests( + recordings=m["recordings"], + supervisions=m["supervisions"], + ) + cut_set = cut_set.compute_and_store_features( + extractor=extractor, + storage_path=f"{output_dir}/{prefix}_feats_{partition}", + # when an executor is specified, make more partitions + num_jobs=num_jobs if ex is None else 80, + executor=ex, + storage_type=LilcomChunkyWriter, + ) + cut_set.to_file(output_dir / f"{prefix}_cuts_{partition}.{suffix}") + + +def get_args(): + parser = argparse.ArgumentParser() + parser.add_argument( + "--num-mel-bins", + type=int, + default=80, + help="""The number of mel bins for Fbank""", + ) + parser.add_argument( + "--whisper-fbank", + type=str2bool, + default=False, + help="Use WhisperFbank instead of Fbank. Default: False.", + ) + parser.add_argument( + "--fbank-dir", + type=Path, + default=Path("data/fbank"), + help="Path to directory with train/valid/test cuts.", + ) + return parser.parse_args() + + +if __name__ == "__main__": + formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s" + + logging.basicConfig(format=formatter, level=logging.INFO) + + args = get_args() + compute_fbank_speechio( + num_mel_bins=args.num_mel_bins, + fbank_dir=args.fbank_dir, + whisper_fbank=args.whisper_fbank, + ) diff --git a/egs/speechio/ASR/local/display_manifest_statistics.py b/egs/speechio/ASR/local/display_manifest_statistics.py new file mode 100644 index 000000000..0c803bfcd --- /dev/null +++ b/egs/speechio/ASR/local/display_manifest_statistics.py @@ -0,0 +1,1162 @@ +#!/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 file displays duration statistics of utterances in a manifest. +You can use the displayed value to choose minimum/maximum duration +to remove short and long utterances during the training. + +See the function `remove_short_and_long_utt()` in transducer_stateless/train.py +for usage. +""" + +SPEECHIO_TESTSET_INDEX = 26 # Currently, from 0 - 26 test sets are open source. + +from lhotse import load_manifest_lazy + + +def main(): + dataset_parts = [] + for i in range(SPEECHIO_TESTSET_INDEX + 1): + idx = f"{i}".zfill(2) + dataset_parts.append(f"SPEECHIO_ASR_ZH000{idx}") + + prefix = "speechio" + suffix = "jsonl.gz" + + for partition in dataset_parts: + path = f"./data/fbank/{prefix}_cuts_{partition}.{suffix}" + cuts = load_manifest_lazy(path) + print( + f"===================Duration statistics of {partition}===================" + ) + cuts.describe() + + +if __name__ == "__main__": + main() + +""" +===================Duration statistics of SPEECHIO_ASR_ZH00000=================== +Cut statistics: +╒═══════════════════════════╤══════════╕ +│ Cuts count: │ 879 │ +├───────────────────────────┼──────────┤ +│ Total duration (hh:mm:ss) │ 01:36:09 │ +├───────────────────────────┼──────────┤ +│ mean │ 6.6 │ +├───────────────────────────┼──────────┤ +│ std │ 2.0 │ +├───────────────────────────┼──────────┤ +│ min │ 1.7 │ +├───────────────────────────┼──────────┤ +│ 25% │ 5.0 │ +├───────────────────────────┼──────────┤ +│ 50% │ 6.5 │ +├───────────────────────────┼──────────┤ +│ 75% │ 8.1 │ +├───────────────────────────┼──────────┤ +│ 99% │ 11.2 │ +├───────────────────────────┼──────────┤ +│ 99.5% │ 11.6 │ +├───────────────────────────┼──────────┤ +│ 99.9% │ 12.2 │ +├───────────────────────────┼──────────┤ +│ max │ 12.5 │ +├───────────────────────────┼──────────┤ +│ Recordings available: │ 879 │ +├───────────────────────────┼──────────┤ +│ Features available: │ 879 │ +├───────────────────────────┼──────────┤ +│ Supervisions available: │ 879 │ +╘═══════════════════════════╧══════════╛ +Speech duration statistics: +╒══════════════════════════════╤══════════╤══════════════════════╕ +│ Total speech duration │ 01:36:09 │ 100.00% of recording │ +├──────────────────────────────┼──────────┼──────────────────────┤ +│ Total speaking time duration │ 01:36:09 │ 100.00% of recording │ +├──────────────────────────────┼──────────┼──────────────────────┤ +│ Total silence duration │ 00:00:00 │ 0.00% of recording │ +╘══════════════════════════════╧══════════╧══════════════════════╛ +===================Duration statistics of SPEECHIO_ASR_ZH00001=================== +Cut statistics: +╒═══════════════════════════╤══════════╕ +│ Cuts count: │ 5069 │ +├───────────────────────────┼──────────┤ +│ Total duration (hh:mm:ss) │ 08:43:04 │ +├───────────────────────────┼──────────┤ +│ mean │ 6.2 │ +├───────────────────────────┼──────────┤ +│ std │ 2.1 │ +├───────────────────────────┼──────────┤ +│ min │ 0.6 │ +├───────────────────────────┼──────────┤ +│ 25% │ 4.6 │ +├───────────────────────────┼──────────┤ +│ 50% │ 6.2 │ +├───────────────────────────┼──────────┤ +│ 75% │ 7.9 │ +├───────────────────────────┼──────────┤ +│ 99% │ 10.0 │ +├───────────────────────────┼──────────┤ +│ 99.5% │ 10.0 │ +├───────────────────────────┼──────────┤ +│ 99.9% │ 10.7 │ +├───────────────────────────┼──────────┤ +│ max │ 12.5 │ +├───────────────────────────┼──────────┤ +│ Recordings available: │ 5069 │ +├───────────────────────────┼──────────┤ +│ Features available: │ 5069 │ +├───────────────────────────┼──────────┤ +│ Supervisions available: │ 5069 │ +╘═══════════════════════════╧══════════╛ +Speech duration statistics: +╒══════════════════════════════╤══════════╤══════════════════════╕ +│ Total speech duration │ 08:43:04 │ 100.00% of recording │ +├──────────────────────────────┼──────────┼──────────────────────┤ +│ Total speaking time duration │ 08:43:04 │ 100.00% of recording │ +├──────────────────────────────┼──────────┼──────────────────────┤ +│ Total silence duration │ 00:00:00 │ 0.00% of recording │ +╘══════════════════════════════╧══════════╧══════════════════════╛ +===================Duration statistics of SPEECHIO_ASR_ZH00002=================== +Cut statistics: +╒═══════════════════════════╤══════════╕ +│ Cuts count: │ 2993 │ +├───────────────────────────┼──────────┤ +│ Total duration (hh:mm:ss) │ 02:45:09 │ +├───────────────────────────┼──────────┤ +│ mean │ 3.3 │ +├───────────────────────────┼──────────┤ +│ std │ 1.5 │ +├───────────────────────────┼──────────┤ +│ min │ 0.4 │ +├───────────────────────────┼──────────┤ +│ 25% │ 2.2 │ +├───────────────────────────┼──────────┤ +│ 50% │ 3.1 │ +├───────────────────────────┼──────────┤ +│ 75% │ 4.3 │ +├───────────────────────────┼──────────┤ +│ 99% │ 7.3 │ +├───────────────────────────┼──────────┤ +│ 99.5% │ 7.8 │ +├───────────────────────────┼──────────┤ +│ 99.9% │ 9.1 │ +├───────────────────────────┼──────────┤ +│ max │ 11.8 │ +├───────────────────────────┼──────────┤ +│ Recordings available: │ 2993 │ +├───────────────────────────┼──────────┤ +│ Features available: │ 2993 │ +├───────────────────────────┼──────────┤ +│ Supervisions available: │ 2993 │ +╘═══════════════════════════╧══════════╛ +Speech duration statistics: +╒══════════════════════════════╤══════════╤══════════════════════╕ +│ Total speech duration │ 02:45:09 │ 100.00% of recording │ +├──────────────────────────────┼──────────┼──────────────────────┤ +│ Total speaking time duration │ 02:45:09 │ 100.00% of recording │ +├──────────────────────────────┼──────────┼──────────────────────┤ +│ Total silence duration │ 00:00:00 │ 0.00% of recording │ +╘══════════════════════════════╧══════════╧══════════════════════╛ +===================Duration statistics of SPEECHIO_ASR_ZH00003=================== +Cut statistics: +╒═══════════════════════════╤══════════╕ +│ Cuts count: │ 1683 │ +├───────────────────────────┼──────────┤ +│ Total duration (hh:mm:ss) │ 02:23:28 │ +├───────────────────────────┼──────────┤ +│ mean │ 5.1 │ +├───────────────────────────┼──────────┤ +│ std │ 1.4 │ +├───────────────────────────┼──────────┤ +│ min │ 2.4 │ +├───────────────────────────┼──────────┤ +│ 25% │ 4.0 │ +├───────────────────────────┼──────────┤ +│ 50% │ 4.9 │ +├───────────────────────────┼──────────┤ +│ 75% │ 6.0 │ +├───────────────────────────┼──────────┤ +│ 99% │ 9.0 │ +├───────────────────────────┼──────────┤ +│ 99.5% │ 9.4 │ +├───────────────────────────┼──────────┤ +│ 99.9% │ 10.8 │ +├───────────────────────────┼──────────┤ +│ max │ 14.2 │ +├───────────────────────────┼──────────┤ +│ Recordings available: │ 1683 │ +├───────────────────────────┼──────────┤ +│ Features available: │ 1683 │ +├───────────────────────────┼──────────┤ +│ Supervisions available: │ 1683 │ +╘═══════════════════════════╧══════════╛ +Speech duration statistics: +╒══════════════════════════════╤══════════╤══════════════════════╕ +│ Total speech duration │ 02:23:28 │ 100.00% of recording │ +├──────────────────────────────┼──────────┼──────────────────────┤ +│ Total speaking time duration │ 02:23:28 │ 100.00% of recording │ +├──────────────────────────────┼──────────┼──────────────────────┤ +│ Total silence duration │ 00:00:00 │ 0.00% of recording │ +╘══════════════════════════════╧══════════╧══════════════════════╛ +===================Duration statistics of SPEECHIO_ASR_ZH00004=================== +Cut statistics: +╒═══════════════════════════╤══════════╕ +│ Cuts count: │ 1311 │ +├───────────────────────────┼──────────┤ +│ Total duration (hh:mm:ss) │ 02:49:16 │ +├───────────────────────────┼──────────┤ +│ mean │ 7.7 │ +├───────────────────────────┼──────────┤ +│ std │ 2.8 │ +├───────────────────────────┼──────────┤ +│ min │ 0.9 │ +├───────────────────────────┼──────────┤ +│ 25% │ 5.8 │ +├───────────────────────────┼──────────┤ +│ 50% │ 8.1 │ +├───────────────────────────┼──────────┤ +│ 75% │ 9.8 │ +├───────────────────────────┼──────────┤ +│ 99% │ 12.9 │ +├───────────────────────────┼──────────┤ +│ 99.5% │ 13.5 │ +├───────────────────────────┼──────────┤ +│ 99.9% │ 13.8 │ +├───────────────────────────┼──────────┤ +│ max │ 14.4 │ +├───────────────────────────┼──────────┤ +│ Recordings available: │ 1311 │ +├───────────────────────────┼──────────┤ +│ Features available: │ 1311 │ +├───────────────────────────┼──────────┤ +│ Supervisions available: │ 1311 │ +╘═══════════════════════════╧══════════╛ +Speech duration statistics: +╒══════════════════════════════╤══════════╤══════════════════════╕ +│ Total speech duration │ 02:49:16 │ 100.00% of recording │ +├──────────────────────────────┼──────────┼──────────────────────┤ +│ Total speaking time duration │ 02:49:16 │ 100.00% of recording │ +├──────────────────────────────┼──────────┼──────────────────────┤ +│ Total silence duration │ 00:00:00 │ 0.00% of recording │ +╘══════════════════════════════╧══════════╧══════════════════════╛ +===================Duration statistics of SPEECHIO_ASR_ZH00005=================== +Cut statistics: +╒═══════════════════════════╤══════════╕ +│ Cuts count: │ 3148 │ +├───────────────────────────┼──────────┤ +│ Total duration (hh:mm:ss) │ 04:22:47 │ +├───────────────────────────┼──────────┤ +│ mean │ 5.0 │ +├───────────────────────────┼──────────┤ +│ std │ 1.4 │ +├───────────────────────────┼──────────┤ +│ min │ 2.0 │ +├───────────────────────────┼──────────┤ +│ 25% │ 3.9 │ +├───────────────────────────┼──────────┤ +│ 50% │ 4.9 │ +├───────────────────────────┼──────────┤ +│ 75% │ 5.9 │ +├───────────────────────────┼──────────┤ +│ 99% │ 8.8 │ +├───────────────────────────┼──────────┤ +│ 99.5% │ 9.3 │ +├───────────────────────────┼──────────┤ +│ 99.9% │ 10.3 │ +├───────────────────────────┼──────────┤ +│ max │ 11.1 │ +├───────────────────────────┼──────────┤ +│ Recordings available: │ 3148 │ +├───────────────────────────┼──────────┤ +│ Features available: │ 3148 │ +├───────────────────────────┼──────────┤ +│ Supervisions available: │ 3148 │ +╘═══════════════════════════╧══════════╛ +Speech duration statistics: +╒══════════════════════════════╤══════════╤══════════════════════╕ +│ Total speech duration │ 04:22:47 │ 100.00% of recording │ +├──────────────────────────────┼──────────┼──────────────────────┤ +│ Total speaking time duration │ 04:22:47 │ 100.00% of recording │ +├──────────────────────────────┼──────────┼──────────────────────┤ +│ Total silence duration │ 00:00:00 │ 0.00% of recording │ +╘══════════════════════════════╧══════════╧══════════════════════╛ +===================Duration statistics of SPEECHIO_ASR_ZH00006=================== +Cut statistics: +╒═══════════════════════════╤══════════╕ +│ Cuts count: │ 1561 │ +├───────────────────────────┼──────────┤ +│ Total duration (hh:mm:ss) │ 01:39:33 │ +├───────────────────────────┼──────────┤ +│ mean │ 3.8 │ +├───────────────────────────┼──────────┤ +│ std │ 2.2 │ +├───────────────────────────┼──────────┤ +│ min │ 0.4 │ +├───────────────────────────┼──────────┤ +│ 25% │ 2.2 │ +├───────────────────────────┼──────────┤ +│ 50% │ 3.3 │ +├───────────────────────────┼──────────┤ +│ 75% │ 4.9 │ +├───────────────────────────┼──────────┤ +│ 99% │ 10.4 │ +├───────────────────────────┼──────────┤ +│ 99.5% │ 11.3 │ +├───────────────────────────┼──────────┤ +│ 99.9% │ 15.3 │ +├───────────────────────────┼──────────┤ +│ max │ 23.8 │ +├───────────────────────────┼──────────┤ +│ Recordings available: │ 1561 │ +├───────────────────────────┼──────────┤ +│ Features available: │ 1561 │ +├───────────────────────────┼──────────┤ +│ Supervisions available: │ 1561 │ +╘═══════════════════════════╧══════════╛ +Speech duration statistics: +╒══════════════════════════════╤══════════╤══════════════════════╕ +│ Total speech duration │ 01:39:33 │ 100.00% of recording │ +├──────────────────────────────┼──────────┼──────────────────────┤ +│ Total speaking time duration │ 01:39:33 │ 100.00% of recording │ +├──────────────────────────────┼──────────┼──────────────────────┤ +│ Total silence duration │ 00:00:00 │ 0.00% of recording │ +╘══════════════════════════════╧══════════╧══════════════════════╛ +===================Duration statistics of SPEECHIO_ASR_ZH00007=================== +Cut statistics: +╒═══════════════════════════╤══════════╕ +│ Cuts count: │ 770 │ +├───────────────────────────┼──────────┤ +│ Total duration (hh:mm:ss) │ 00:58:57 │ +├───────────────────────────┼──────────┤ +│ mean │ 4.6 │ +├───────────────────────────┼──────────┤ +│ std │ 2.4 │ +├───────────────────────────┼──────────┤ +│ min │ 0.7 │ +├───────────────────────────┼──────────┤ +│ 25% │ 2.7 │ +├───────────────────────────┼──────────┤ +│ 50% │ 4.0 │ +├───────────────────────────┼──────────┤ +│ 75% │ 6.0 │ +├───────────────────────────┼──────────┤ +│ 99% │ 11.8 │ +├───────────────────────────┼──────────┤ +│ 99.5% │ 13.0 │ +├───────────────────────────┼──────────┤ +│ 99.9% │ 15.1 │ +├───────────────────────────┼──────────┤ +│ max │ 18.7 │ +├───────────────────────────┼──────────┤ +│ Recordings available: │ 770 │ +├───────────────────────────┼──────────┤ +│ Features available: │ 770 │ +├───────────────────────────┼──────────┤ +│ Supervisions available: │ 770 │ +╘═══════════════════════════╧══════════╛ +Speech duration statistics: +╒══════════════════════════════╤══════════╤══════════════════════╕ +│ Total speech duration │ 00:58:57 │ 100.00% of recording │ +├──────────────────────────────┼──────────┼──────────────────────┤ +│ Total speaking time duration │ 00:58:57 │ 100.00% of recording │ +├──────────────────────────────┼──────────┼──────────────────────┤ +│ Total silence duration │ 00:00:00 │ 0.00% of recording │ +╘══════════════════════════════╧══════════╧══════════════════════╛ +===================Duration statistics of SPEECHIO_ASR_ZH00008=================== +Cut statistics: +╒═══════════════════════════╤══════════╕ +│ Cuts count: │ 884 │ +├───────────────────────────┼──────────┤ +│ Total duration (hh:mm:ss) │ 01:16:55 │ +├───────────────────────────┼──────────┤ +│ mean │ 5.2 │ +├───────────────────────────┼──────────┤ +│ std │ 2.3 │ +├───────────────────────────┼──────────┤ +│ min │ 1.1 │ +├───────────────────────────┼──────────┤ +│ 25% │ 3.5 │ +├───────────────────────────┼──────────┤ +│ 50% │ 5.0 │ +├───────────────────────────┼──────────┤ +│ 75% │ 6.4 │ +├───────────────────────────┼──────────┤ +│ 99% │ 11.3 │ +├───────────────────────────┼──────────┤ +│ 99.5% │ 12.7 │ +├───────────────────────────┼──────────┤ +│ 99.9% │ 16.2 │ +├───────────────────────────┼──────────┤ +│ max │ 18.5 │ +├───────────────────────────┼──────────┤ +│ Recordings available: │ 884 │ +├───────────────────────────┼──────────┤ +│ Features available: │ 884 │ +├───────────────────────────┼──────────┤ +│ Supervisions available: │ 884 │ +╘═══════════════════════════╧══════════╛ +Speech duration statistics: +╒══════════════════════════════╤══════════╤══════════════════════╕ +│ Total speech duration │ 01:16:55 │ 100.00% of recording │ +├──────────────────────────────┼──────────┼──────────────────────┤ +│ Total speaking time duration │ 01:16:55 │ 100.00% of recording │ +├──────────────────────────────┼──────────┼──────────────────────┤ +│ Total silence duration │ 00:00:00 │ 0.00% of recording │ +╘══════════════════════════════╧══════════╧══════════════════════╛ +===================Duration statistics of SPEECHIO_ASR_ZH00009=================== +Cut statistics: +╒═══════════════════════════╤══════════╕ +│ Cuts count: │ 3466 │ +├───────────────────────────┼──────────┤ +│ Total duration (hh:mm:ss) │ 04:38:13 │ +├───────────────────────────┼──────────┤ +│ mean │ 4.8 │ +├───────────────────────────┼──────────┤ +│ std │ 1.9 │ +├───────────────────────────┼──────────┤ +│ min │ 1.1 │ +├───────────────────────────┼──────────┤ +│ 25% │ 3.4 │ +├───────────────────────────┼──────────┤ +│ 50% │ 4.5 │ +├───────────────────────────┼──────────┤ +│ 75% │ 5.9 │ +├───────────────────────────┼──────────┤ +│ 99% │ 10.5 │ +├───────────────────────────┼──────────┤ +│ 99.5% │ 11.3 │ +├───────────────────────────┼──────────┤ +│ 99.9% │ 12.5 │ +├───────────────────────────┼──────────┤ +│ max │ 13.1 │ +├───────────────────────────┼──────────┤ +│ Recordings available: │ 3466 │ +├───────────────────────────┼──────────┤ +│ Features available: │ 3466 │ +├───────────────────────────┼──────────┤ +│ Supervisions available: │ 3466 │ +╘═══════════════════════════╧══════════╛ +Speech duration statistics: +╒══════════════════════════════╤══════════╤══════════════════════╕ +│ Total speech duration │ 04:38:13 │ 100.00% of recording │ +├──────────────────────────────┼──────────┼──────────────────────┤ +│ Total speaking time duration │ 04:38:13 │ 100.00% of recording │ +├──────────────────────────────┼──────────┼──────────────────────┤ +│ Total silence duration │ 00:00:00 │ 0.00% of recording │ +╘══════════════════════════════╧══════════╧══════════════════════╛ +===================Duration statistics of SPEECHIO_ASR_ZH00010=================== +Cut statistics: +╒═══════════════════════════╤══════════╕ +│ Cuts count: │ 2251 │ +├───────────────────────────┼──────────┤ +│ Total duration (hh:mm:ss) │ 04:12:54 │ +├───────────────────────────┼──────────┤ +│ mean │ 6.7 │ +├───────────────────────────┼──────────┤ +│ std │ 3.0 │ +├───────────────────────────┼──────────┤ +│ min │ 1.4 │ +├───────────────────────────┼──────────┤ +│ 25% │ 4.5 │ +├───────────────────────────┼──────────┤ +│ 50% │ 6.3 │ +├───────────────────────────┼──────────┤ +│ 75% │ 8.5 │ +├───────────────────────────┼──────────┤ +│ 99% │ 14.9 │ +├───────────────────────────┼──────────┤ +│ 99.5% │ 15.5 │ +├───────────────────────────┼──────────┤ +│ 99.9% │ 15.8 │ +├───────────────────────────┼──────────┤ +│ max │ 16.2 │ +├───────────────────────────┼──────────┤ +│ Recordings available: │ 2251 │ +├───────────────────────────┼──────────┤ +│ Features available: │ 2251 │ +├───────────────────────────┼──────────┤ +│ Supervisions available: │ 2251 │ +╘═══════════════════════════╧══════════╛ +Speech duration statistics: +╒══════════════════════════════╤══════════╤══════════════════════╕ +│ Total speech duration │ 04:12:54 │ 100.00% of recording │ +├──────────────────────────────┼──────────┼──────────────────────┤ +│ Total speaking time duration │ 04:12:54 │ 100.00% of recording │ +├──────────────────────────────┼──────────┼──────────────────────┤ +│ Total silence duration │ 00:00:00 │ 0.00% of recording │ +╘══════════════════════════════╧══════════╧══════════════════════╛ +===================Duration statistics of SPEECHIO_ASR_ZH00011=================== +Cut statistics: +╒═══════════════════════════╤══════════╕ +│ Cuts count: │ 1053 │ +├───────────────────────────┼──────────┤ +│ Total duration (hh:mm:ss) │ 03:27:12 │ +├───────────────────────────┼──────────┤ +│ mean │ 11.8 │ +├───────────────────────────┼──────────┤ +│ std │ 3.4 │ +├───────────────────────────┼──────────┤ +│ min │ 1.1 │ +├───────────────────────────┼──────────┤ +│ 25% │ 11.5 │ +├───────────────────────────┼──────────┤ +│ 50% │ 13.0 │ +├───────────────────────────┼──────────┤ +│ 75% │ 13.9 │ +├───────────────────────────┼──────────┤ +│ 99% │ 15.0 │ +├───────────────────────────┼──────────┤ +│ 99.5% │ 15.1 │ +├───────────────────────────┼──────────┤ +│ 99.9% │ 20.7 │ +├───────────────────────────┼──────────┤ +│ max │ 22.2 │ +├───────────────────────────┼──────────┤ +│ Recordings available: │ 1053 │ +├───────────────────────────┼──────────┤ +│ Features available: │ 1053 │ +├───────────────────────────┼──────────┤ +│ Supervisions available: │ 1053 │ +╘═══════════════════════════╧══════════╛ +Speech duration statistics: +╒══════════════════════════════╤══════════╤══════════════════════╕ +│ Total speech duration │ 03:27:12 │ 100.00% of recording │ +├──────────────────────────────┼──────────┼──────────────────────┤ +│ Total speaking time duration │ 03:27:12 │ 100.00% of recording │ +├──────────────────────────────┼──────────┼──────────────────────┤ +│ Total silence duration │ 00:00:00 │ 0.00% of recording │ +╘══════════════════════════════╧══════════╧══════════════════════╛ +===================Duration statistics of SPEECHIO_ASR_ZH00012=================== +Cut statistics: +╒═══════════════════════════╤══════════╕ +│ Cuts count: │ 1170 │ +├───────────────────────────┼──────────┤ +│ Total duration (hh:mm:ss) │ 03:23:34 │ +├───────────────────────────┼──────────┤ +│ mean │ 10.4 │ +├───────────────────────────┼──────────┤ +│ std │ 3.5 │ +├───────────────────────────┼──────────┤ +│ min │ 0.8 │ +├───────────────────────────┼──────────┤ +│ 25% │ 8.0 │ +├───────────────────────────┼──────────┤ +│ 50% │ 11.5 │ +├───────────────────────────┼──────────┤ +│ 75% │ 13.2 │ +├───────────────────────────┼──────────┤ +│ 99% │ 15.0 │ +├───────────────────────────┼──────────┤ +│ 99.5% │ 15.1 │ +├───────────────────────────┼──────────┤ +│ 99.9% │ 15.7 │ +├───────────────────────────┼──────────┤ +│ max │ 20.3 │ +├───────────────────────────┼──────────┤ +│ Recordings available: │ 1170 │ +├───────────────────────────┼──────────┤ +│ Features available: │ 1170 │ +├───────────────────────────┼──────────┤ +│ Supervisions available: │ 1170 │ +╘═══════════════════════════╧══════════╛ +Speech duration statistics: +╒══════════════════════════════╤══════════╤══════════════════════╕ +│ Total speech duration │ 03:23:34 │ 100.00% of recording │ +├──────────────────────────────┼──────────┼──────────────────────┤ +│ Total speaking time duration │ 03:23:34 │ 100.00% of recording │ +├──────────────────────────────┼──────────┼──────────────────────┤ +│ Total silence duration │ 00:00:00 │ 0.00% of recording │ +╘══════════════════════════════╧══════════╧══════════════════════╛ +===================Duration statistics of SPEECHIO_ASR_ZH00013=================== +Cut statistics: +╒═══════════════════════════╤══════════╕ +│ Cuts count: │ 1321 │ +├───────────────────────────┼──────────┤ +│ Total duration (hh:mm:ss) │ 01:46:41 │ +├───────────────────────────┼──────────┤ +│ mean │ 4.8 │ +├───────────────────────────┼──────────┤ +│ std │ 1.5 │ +├───────────────────────────┼──────────┤ +│ min │ 0.9 │ +├───────────────────────────┼──────────┤ +│ 25% │ 3.8 │ +├───────────────────────────┼──────────┤ +│ 50% │ 4.8 │ +├───────────────────────────┼──────────┤ +│ 75% │ 5.8 │ +├───────────────────────────┼──────────┤ +│ 99% │ 8.5 │ +├───────────────────────────┼──────────┤ +│ 99.5% │ 9.1 │ +├───────────────────────────┼──────────┤ +│ 99.9% │ 9.5 │ +├───────────────────────────┼──────────┤ +│ max │ 9.7 │ +├───────────────────────────┼──────────┤ +│ Recordings available: │ 1321 │ +├───────────────────────────┼──────────┤ +│ Features available: │ 1321 │ +├───────────────────────────┼──────────┤ +│ Supervisions available: │ 1321 │ +╘═══════════════════════════╧══════════╛ +Speech duration statistics: +╒══════════════════════════════╤══════════╤══════════════════════╕ +│ Total speech duration │ 01:46:41 │ 100.00% of recording │ +├──────────────────────────────┼──────────┼──────────────────────┤ +│ Total speaking time duration │ 01:46:41 │ 100.00% of recording │ +├──────────────────────────────┼──────────┼──────────────────────┤ +│ Total silence duration │ 00:00:00 │ 0.00% of recording │ +╘══════════════════════════════╧══════════╧══════════════════════╛ +===================Duration statistics of SPEECHIO_ASR_ZH00014=================== +Cut statistics: +╒═══════════════════════════╤══════════╕ +│ Cuts count: │ 856 │ +├───────────────────────────┼──────────┤ +│ Total duration (hh:mm:ss) │ 01:00:39 │ +├───────────────────────────┼──────────┤ +│ mean │ 4.3 │ +├───────────────────────────┼──────────┤ +│ std │ 1.8 │ +├───────────────────────────┼──────────┤ +│ min │ 0.8 │ +├───────────────────────────┼──────────┤ +│ 25% │ 2.9 │ +├───────────────────────────┼──────────┤ +│ 50% │ 4.1 │ +├───────────────────────────┼──────────┤ +│ 75% │ 5.5 │ +├───────────────────────────┼──────────┤ +│ 99% │ 8.5 │ +├───────────────────────────┼──────────┤ +│ 99.5% │ 9.2 │ +├───────────────────────────┼──────────┤ +│ 99.9% │ 10.0 │ +├───────────────────────────┼──────────┤ +│ max │ 11.1 │ +├───────────────────────────┼──────────┤ +│ Recordings available: │ 856 │ +├───────────────────────────┼──────────┤ +│ Features available: │ 856 │ +├───────────────────────────┼──────────┤ +│ Supervisions available: │ 856 │ +╘═══════════════════════════╧══════════╛ +Speech duration statistics: +╒══════════════════════════════╤══════════╤══════════════════════╕ +│ Total speech duration │ 01:00:39 │ 100.00% of recording │ +├──────────────────────────────┼──────────┼──────────────────────┤ +│ Total speaking time duration │ 01:00:39 │ 100.00% of recording │ +├──────────────────────────────┼──────────┼──────────────────────┤ +│ Total silence duration │ 00:00:00 │ 0.00% of recording │ +╘══════════════════════════════╧══════════╧══════════════════════╛ +===================Duration statistics of SPEECHIO_ASR_ZH00015=================== +Cut statistics: +╒═══════════════════════════╤══════════╕ +│ Cuts count: │ 1168 │ +├───────────────────────────┼──────────┤ +│ Total duration (hh:mm:ss) │ 02:08:52 │ +├───────────────────────────┼──────────┤ +│ mean │ 6.6 │ +├───────────────────────────┼──────────┤ +│ std │ 2.0 │ +├───────────────────────────┼──────────┤ +│ min │ 1.2 │ +├───────────────────────────┼──────────┤ +│ 25% │ 5.3 │ +├───────────────────────────┼──────────┤ +│ 50% │ 6.8 │ +├───────────────────────────┼──────────┤ +│ 75% │ 8.2 │ +├───────────────────────────┼──────────┤ +│ 99% │ 9.9 │ +├───────────────────────────┼──────────┤ +│ 99.5% │ 10.0 │ +├───────────────────────────┼──────────┤ +│ 99.9% │ 10.1 │ +├───────────────────────────┼──────────┤ +│ max │ 15.5 │ +├───────────────────────────┼──────────┤ +│ Recordings available: │ 1168 │ +├───────────────────────────┼──────────┤ +│ Features available: │ 1168 │ +├───────────────────────────┼──────────┤ +│ Supervisions available: │ 1168 │ +╘═══════════════════════════╧══════════╛ +Speech duration statistics: +╒══════════════════════════════╤══════════╤══════════════════════╕ +│ Total speech duration │ 02:08:52 │ 100.00% of recording │ +├──────────────────────────────┼──────────┼──────────────────────┤ +│ Total speaking time duration │ 02:08:52 │ 100.00% of recording │ +├──────────────────────────────┼──────────┼──────────────────────┤ +│ Total silence duration │ 00:00:00 │ 0.00% of recording │ +╘══════════════════════════════╧══════════╧══════════════════════╛ +===================Duration statistics of SPEECHIO_ASR_ZH00016=================== +Cut statistics: +╒═══════════════════════════╤══════════╕ +│ Cuts count: │ 1201 │ +├───────────────────────────┼──────────┤ +│ Total duration (hh:mm:ss) │ 01:00:46 │ +├───────────────────────────┼──────────┤ +│ mean │ 3.0 │ +├───────────────────────────┼──────────┤ +│ std │ 2.0 │ +├───────────────────────────┼──────────┤ +│ min │ 0.9 │ +├───────────────────────────┼──────────┤ +│ 25% │ 1.6 │ +├───────────────────────────┼──────────┤ +│ 50% │ 2.3 │ +├───────────────────────────┼──────────┤ +│ 75% │ 3.8 │ +├───────────────────────────┼──────────┤ +│ 99% │ 9.0 │ +├───────────────────────────┼──────────┤ +│ 99.5% │ 9.5 │ +├───────────────────────────┼──────────┤ +│ 99.9% │ 9.7 │ +├───────────────────────────┼──────────┤ +│ max │ 9.9 │ +├───────────────────────────┼──────────┤ +│ Recordings available: │ 1201 │ +├───────────────────────────┼──────────┤ +│ Features available: │ 1201 │ +├───────────────────────────┼──────────┤ +│ Supervisions available: │ 1201 │ +╘═══════════════════════════╧══════════╛ +Speech duration statistics: +╒══════════════════════════════╤══════════╤══════════════════════╕ +│ Total speech duration │ 01:00:46 │ 100.00% of recording │ +├──────────────────────────────┼──────────┼──────────────────────┤ +│ Total speaking time duration │ 01:00:46 │ 100.00% of recording │ +├──────────────────────────────┼──────────┼──────────────────────┤ +│ Total silence duration │ 00:00:00 │ 0.00% of recording │ +╘══════════════════════════════╧══════════╧══════════════════════╛ +===================Duration statistics of SPEECHIO_ASR_ZH00017=================== +Cut statistics: +╒═══════════════════════════╤══════════╕ +│ Cuts count: │ 1271 │ +├───────────────────────────┼──────────┤ +│ Total duration (hh:mm:ss) │ 01:47:57 │ +├───────────────────────────┼──────────┤ +│ mean │ 5.1 │ +├───────────────────────────┼──────────┤ +│ std │ 2.2 │ +├───────────────────────────┼──────────┤ +│ min │ 1.0 │ +├───────────────────────────┼──────────┤ +│ 25% │ 3.3 │ +├───────────────────────────┼──────────┤ +│ 50% │ 4.9 │ +├───────────────────────────┼──────────┤ +│ 75% │ 6.8 │ +├───────────────────────────┼──────────┤ +│ 99% │ 9.7 │ +├───────────────────────────┼──────────┤ +│ 99.5% │ 10.0 │ +├───────────────────────────┼──────────┤ +│ 99.9% │ 10.0 │ +├───────────────────────────┼──────────┤ +│ max │ 10.4 │ +├───────────────────────────┼──────────┤ +│ Recordings available: │ 1271 │ +├───────────────────────────┼──────────┤ +│ Features available: │ 1271 │ +├───────────────────────────┼──────────┤ +│ Supervisions available: │ 1271 │ +╘═══════════════════════════╧══════════╛ +Speech duration statistics: +╒══════════════════════════════╤══════════╤══════════════════════╕ +│ Total speech duration │ 01:47:57 │ 100.00% of recording │ +├──────────────────────────────┼──────────┼──────────────────────┤ +│ Total speaking time duration │ 01:47:57 │ 100.00% of recording │ +├──────────────────────────────┼──────────┼──────────────────────┤ +│ Total silence duration │ 00:00:00 │ 0.00% of recording │ +╘══════════════════════════════╧══════════╧══════════════════════╛ +===================Duration statistics of SPEECHIO_ASR_ZH00018=================== +Cut statistics: +╒═══════════════════════════╤══════════╕ +│ Cuts count: │ 899 │ +├───────────────────────────┼──────────┤ +│ Total duration (hh:mm:ss) │ 00:51:12 │ +├───────────────────────────┼──────────┤ +│ mean │ 3.4 │ +├───────────────────────────┼──────────┤ +│ std │ 1.2 │ +├───────────────────────────┼──────────┤ +│ min │ 1.3 │ +├───────────────────────────┼──────────┤ +│ 25% │ 2.5 │ +├───────────────────────────┼──────────┤ +│ 50% │ 3.1 │ +├───────────────────────────┼──────────┤ +│ 75% │ 4.1 │ +├───────────────────────────┼──────────┤ +│ 99% │ 6.7 │ +├───────────────────────────┼──────────┤ +│ 99.5% │ 7.1 │ +├───────────────────────────┼──────────┤ +│ 99.9% │ 8.2 │ +├───────────────────────────┼──────────┤ +│ max │ 9.2 │ +├───────────────────────────┼──────────┤ +│ Recordings available: │ 899 │ +├───────────────────────────┼──────────┤ +│ Features available: │ 899 │ +├───────────────────────────┼──────────┤ +│ Supervisions available: │ 899 │ +╘═══════════════════════════╧══════════╛ +Speech duration statistics: +╒══════════════════════════════╤══════════╤══════════════════════╕ +│ Total speech duration │ 00:51:12 │ 100.00% of recording │ +├──────────────────────────────┼──────────┼──────────────────────┤ +│ Total speaking time duration │ 00:51:12 │ 100.00% of recording │ +├──────────────────────────────┼──────────┼──────────────────────┤ +│ Total silence duration │ 00:00:00 │ 0.00% of recording │ +╘══════════════════════════════╧══════════╧══════════════════════╛ +===================Duration statistics of SPEECHIO_ASR_ZH00019=================== +Cut statistics: +╒═══════════════════════════╤══════════╕ +│ Cuts count: │ 615 │ +├───────────────────────────┼──────────┤ +│ Total duration (hh:mm:ss) │ 00:41:43 │ +├───────────────────────────┼──────────┤ +│ mean │ 4.1 │ +├───────────────────────────┼──────────┤ +│ std │ 1.5 │ +├───────────────────────────┼──────────┤ +│ min │ 1.3 │ +├───────────────────────────┼──────────┤ +│ 25% │ 2.8 │ +├───────────────────────────┼──────────┤ +│ 50% │ 3.9 │ +├───────────────────────────┼──────────┤ +│ 75% │ 5.2 │ +├───────────────────────────┼──────────┤ +│ 99% │ 7.9 │ +├───────────────────────────┼──────────┤ +│ 99.5% │ 8.1 │ +├───────────────────────────┼──────────┤ +│ 99.9% │ 8.6 │ +├───────────────────────────┼──────────┤ +│ max │ 8.8 │ +├───────────────────────────┼──────────┤ +│ Recordings available: │ 615 │ +├───────────────────────────┼──────────┤ +│ Features available: │ 615 │ +├───────────────────────────┼──────────┤ +│ Supervisions available: │ 615 │ +╘═══════════════════════════╧══════════╛ +Speech duration statistics: +╒══════════════════════════════╤══════════╤══════════════════════╕ +│ Total speech duration │ 00:41:43 │ 100.00% of recording │ +├──────────────────────────────┼──────────┼──────────────────────┤ +│ Total speaking time duration │ 00:41:43 │ 100.00% of recording │ +├──────────────────────────────┼──────────┼──────────────────────┤ +│ Total silence duration │ 00:00:00 │ 0.00% of recording │ +╘══════════════════════════════╧══════════╧══════════════════════╛ +===================Duration statistics of SPEECHIO_ASR_ZH00020=================== +Cut statistics: +╒═══════════════════════════╤══════════╕ +│ Cuts count: │ 1590 │ +├───────────────────────────┼──────────┤ +│ Total duration (hh:mm:ss) │ 02:10:54 │ +├───────────────────────────┼──────────┤ +│ mean │ 4.9 │ +├───────────────────────────┼──────────┤ +│ std │ 1.5 │ +├───────────────────────────┼──────────┤ +│ min │ 1.2 │ +├───────────────────────────┼──────────┤ +│ 25% │ 3.8 │ +├───────────────────────────┼──────────┤ +│ 50% │ 4.9 │ +├───────────────────────────┼──────────┤ +│ 75% │ 6.0 │ +├───────────────────────────┼──────────┤ +│ 99% │ 8.5 │ +├───────────────────────────┼──────────┤ +│ 99.5% │ 8.7 │ +├───────────────────────────┼──────────┤ +│ 99.9% │ 9.2 │ +├───────────────────────────┼──────────┤ +│ max │ 10.4 │ +├───────────────────────────┼──────────┤ +│ Recordings available: │ 1590 │ +├───────────────────────────┼──────────┤ +│ Features available: │ 1590 │ +├───────────────────────────┼──────────┤ +│ Supervisions available: │ 1590 │ +╘═══════════════════════════╧══════════╛ +Speech duration statistics: +╒══════════════════════════════╤══════════╤══════════════════════╕ +│ Total speech duration │ 02:10:54 │ 100.00% of recording │ +├──────────────────────────────┼──────────┼──────────────────────┤ +│ Total speaking time duration │ 02:10:54 │ 100.00% of recording │ +├──────────────────────────────┼──────────┼──────────────────────┤ +│ Total silence duration │ 00:00:00 │ 0.00% of recording │ +╘══════════════════════════════╧══════════╧══════════════════════╛ +===================Duration statistics of SPEECHIO_ASR_ZH00021=================== +Cut statistics: +╒═══════════════════════════╤══════════╕ +│ Cuts count: │ 1035 │ +├───────────────────────────┼──────────┤ +│ Total duration (hh:mm:ss) │ 01:44:07 │ +├───────────────────────────┼──────────┤ +│ mean │ 6.0 │ +├───────────────────────────┼──────────┤ +│ std │ 1.8 │ +├───────────────────────────┼──────────┤ +│ min │ 1.1 │ +├───────────────────────────┼──────────┤ +│ 25% │ 4.7 │ +├───────────────────────────┼──────────┤ +│ 50% │ 5.9 │ +├───────────────────────────┼──────────┤ +│ 75% │ 7.3 │ +├───────────────────────────┼──────────┤ +│ 99% │ 10.4 │ +├───────────────────────────┼──────────┤ +│ 99.5% │ 10.6 │ +├───────────────────────────┼──────────┤ +│ 99.9% │ 11.0 │ +├───────────────────────────┼──────────┤ +│ max │ 11.1 │ +├───────────────────────────┼──────────┤ +│ Recordings available: │ 1035 │ +├───────────────────────────┼──────────┤ +│ Features available: │ 1035 │ +├───────────────────────────┼──────────┤ +│ Supervisions available: │ 1035 │ +╘═══════════════════════════╧══════════╛ +Speech duration statistics: +╒══════════════════════════════╤══════════╤══════════════════════╕ +│ Total speech duration │ 01:44:07 │ 100.00% of recording │ +├──────────────────────────────┼──────────┼──────────────────────┤ +│ Total speaking time duration │ 01:44:07 │ 100.00% of recording │ +├──────────────────────────────┼──────────┼──────────────────────┤ +│ Total silence duration │ 00:00:00 │ 0.00% of recording │ +╘══════════════════════════════╧══════════╧══════════════════════╛ +===================Duration statistics of SPEECHIO_ASR_ZH00022=================== +Cut statistics: +╒═══════════════════════════╤══════════╕ +│ Cuts count: │ 1026 │ +├───────────────────────────┼──────────┤ +│ Total duration (hh:mm:ss) │ 01:40:43 │ +├───────────────────────────┼──────────┤ +│ mean │ 5.9 │ +├───────────────────────────┼──────────┤ +│ std │ 2.2 │ +├───────────────────────────┼──────────┤ +│ min │ 0.9 │ +├───────────────────────────┼──────────┤ +│ 25% │ 4.4 │ +├───────────────────────────┼──────────┤ +│ 50% │ 5.8 │ +├───────────────────────────┼──────────┤ +│ 75% │ 7.1 │ +├───────────────────────────┼──────────┤ +│ 99% │ 12.1 │ +├───────────────────────────┼──────────┤ +│ 99.5% │ 12.7 │ +├───────────────────────────┼──────────┤ +│ 99.9% │ 13.9 │ +├───────────────────────────┼──────────┤ +│ max │ 14.0 │ +├───────────────────────────┼──────────┤ +│ Recordings available: │ 1026 │ +├───────────────────────────┼──────────┤ +│ Features available: │ 1026 │ +├───────────────────────────┼──────────┤ +│ Supervisions available: │ 1026 │ +╘═══════════════════════════╧══════════╛ +Speech duration statistics: +╒══════════════════════════════╤══════════╤══════════════════════╕ +│ Total speech duration │ 01:40:43 │ 100.00% of recording │ +├──────────────────────────────┼──────────┼──────────────────────┤ +│ Total speaking time duration │ 01:40:43 │ 100.00% of recording │ +├──────────────────────────────┼──────────┼──────────────────────┤ +│ Total silence duration │ 00:00:00 │ 0.00% of recording │ +╘══════════════════════════════╧══════════╧══════════════════════╛ +===================Duration statistics of SPEECHIO_ASR_ZH00023=================== +Cut statistics: +╒═══════════════════════════╤══════════╕ +│ Cuts count: │ 1528 │ +├───────────────────────────┼──────────┤ +│ Total duration (hh:mm:ss) │ 02:06:51 │ +├───────────────────────────┼──────────┤ +│ mean │ 5.0 │ +├───────────────────────────┼──────────┤ +│ std │ 2.5 │ +├───────────────────────────┼──────────┤ +│ min │ 0.5 │ +├───────────────────────────┼──────────┤ +│ 25% │ 3.1 │ +├───────────────────────────┼──────────┤ +│ 50% │ 4.5 │ +├───────────────────────────┼──────────┤ +│ 75% │ 6.6 │ +├───────────────────────────┼──────────┤ +│ 99% │ 12.3 │ +├───────────────────────────┼──────────┤ +│ 99.5% │ 13.9 │ +├───────────────────────────┼──────────┤ +│ 99.9% │ 15.8 │ +├───────────────────────────┼──────────┤ +│ max │ 16.8 │ +├───────────────────────────┼──────────┤ +│ Recordings available: │ 1528 │ +├───────────────────────────┼──────────┤ +│ Features available: │ 1528 │ +├───────────────────────────┼──────────┤ +│ Supervisions available: │ 1528 │ +╘═══════════════════════════╧══════════╛ +Speech duration statistics: +╒══════════════════════════════╤══════════╤══════════════════════╕ +│ Total speech duration │ 02:06:51 │ 100.00% of recording │ +├──────────────────────────────┼──────────┼──────────────────────┤ +│ Total speaking time duration │ 02:06:51 │ 100.00% of recording │ +├──────────────────────────────┼──────────┼──────────────────────┤ +│ Total silence duration │ 00:00:00 │ 0.00% of recording │ +╘══════════════════════════════╧══════════╧══════════════════════╛ +===================Duration statistics of SPEECHIO_ASR_ZH00024=================== +Cut statistics: +╒═══════════════════════════╤══════════╕ +│ Cuts count: │ 1930 │ +├───────────────────────────┼──────────┤ +│ Total duration (hh:mm:ss) │ 02:39:02 │ +├───────────────────────────┼──────────┤ +│ mean │ 4.9 │ +├───────────────────────────┼──────────┤ +│ std │ 2.0 │ +├───────────────────────────┼──────────┤ +│ min │ 0.9 │ +├───────────────────────────┼──────────┤ +│ 25% │ 3.4 │ +├───────────────────────────┼──────────┤ +│ 50% │ 4.7 │ +├───────────────────────────┼──────────┤ +│ 75% │ 6.2 │ +├───────────────────────────┼──────────┤ +│ 99% │ 10.3 │ +├───────────────────────────┼──────────┤ +│ 99.5% │ 10.9 │ +├───────────────────────────┼──────────┤ +│ 99.9% │ 12.0 │ +├───────────────────────────┼──────────┤ +│ max │ 12.6 │ +├───────────────────────────┼──────────┤ +│ Recordings available: │ 1930 │ +├───────────────────────────┼──────────┤ +│ Features available: │ 1930 │ +├───────────────────────────┼──────────┤ +│ Supervisions available: │ 1930 │ +╘═══════════════════════════╧══════════╛ +Speech duration statistics: +╒══════════════════════════════╤══════════╤══════════════════════╕ +│ Total speech duration │ 02:39:02 │ 100.00% of recording │ +├──────────────────────────────┼──────────┼──────────────────────┤ +│ Total speaking time duration │ 02:39:02 │ 100.00% of recording │ +├──────────────────────────────┼──────────┼──────────────────────┤ +│ Total silence duration │ 00:00:00 │ 0.00% of recording │ +╘══════════════════════════════╧══════════╧══════════════════════╛ +===================Duration statistics of SPEECHIO_ASR_ZH00025=================== +Cut statistics: +╒═══════════════════════════╤══════════╕ +│ Cuts count: │ 1164 │ +├───────────────────────────┼──────────┤ +│ Total duration (hh:mm:ss) │ 01:24:42 │ +├───────────────────────────┼──────────┤ +│ mean │ 4.4 │ +├───────────────────────────┼──────────┤ +│ std │ 1.9 │ +├───────────────────────────┼──────────┤ +│ min │ 0.9 │ +├───────────────────────────┼──────────┤ +│ 25% │ 2.9 │ +├───────────────────────────┼──────────┤ +│ 50% │ 4.1 │ +├───────────────────────────┼──────────┤ +│ 75% │ 5.6 │ +├───────────────────────────┼──────────┤ +│ 99% │ 10.4 │ +├───────────────────────────┼──────────┤ +│ 99.5% │ 10.9 │ +├───────────────────────────┼──────────┤ +│ 99.9% │ 12.5 │ +├───────────────────────────┼──────────┤ +│ max │ 13.0 │ +├───────────────────────────┼──────────┤ +│ Recordings available: │ 1164 │ +├───────────────────────────┼──────────┤ +│ Features available: │ 1164 │ +├───────────────────────────┼──────────┤ +│ Supervisions available: │ 1164 │ +╘═══════════════════════════╧══════════╛ +Speech duration statistics: +╒══════════════════════════════╤══════════╤══════════════════════╕ +│ Total speech duration │ 01:24:42 │ 100.00% of recording │ +├──────────────────────────────┼──────────┼──────────────────────┤ +│ Total speaking time duration │ 01:24:42 │ 100.00% of recording │ +├──────────────────────────────┼──────────┼──────────────────────┤ +│ Total silence duration │ 00:00:00 │ 0.00% of recording │ +╘══════════════════════════════╧══════════╧══════════════════════╛ +===================Duration statistics of SPEECHIO_ASR_ZH00026=================== +Cut statistics: +╒═══════════════════════════╤══════════╕ +│ Cuts count: │ 1336 │ +├───────────────────────────┼──────────┤ +│ Total duration (hh:mm:ss) │ 02:25:38 │ +├───────────────────────────┼──────────┤ +│ mean │ 6.5 │ +├───────────────────────────┼──────────┤ +│ std │ 2.3 │ +├───────────────────────────┼──────────┤ +│ min │ 0.5 │ +├───────────────────────────┼──────────┤ +│ 25% │ 4.9 │ +├───────────────────────────┼──────────┤ +│ 50% │ 6.8 │ +├───────────────────────────┼──────────┤ +│ 75% │ 8.3 │ +├───────────────────────────┼──────────┤ +│ 99% │ 10.4 │ +├───────────────────────────┼──────────┤ +│ 99.5% │ 11.9 │ +├───────────────────────────┼──────────┤ +│ 99.9% │ 12.9 │ +├───────────────────────────┼──────────┤ +│ max │ 13.3 │ +├───────────────────────────┼──────────┤ +│ Recordings available: │ 1336 │ +├───────────────────────────┼──────────┤ +│ Features available: │ 1336 │ +├───────────────────────────┼──────────┤ +│ Supervisions available: │ 1336 │ +╘═══════════════════════════╧══════════╛ +Speech duration statistics: +╒══════════════════════════════╤══════════╤══════════════════════╕ +│ Total speech duration │ 02:25:38 │ 100.00% of recording │ +├──────────────────────────────┼──────────┼──────────────────────┤ +│ Total speaking time duration │ 02:25:38 │ 100.00% of recording │ +├──────────────────────────────┼──────────┼──────────────────────┤ +│ Total silence duration │ 00:00:00 │ 0.00% of recording │ +╘══════════════════════════════╧══════════╧══════════════════════╛ + +""" diff --git a/egs/speechio/ASR/local/whisper_zipformer_fusion.py b/egs/speechio/ASR/local/whisper_zipformer_fusion.py new file mode 100644 index 000000000..04c5e75f0 --- /dev/null +++ b/egs/speechio/ASR/local/whisper_zipformer_fusion.py @@ -0,0 +1,217 @@ +#!/usr/bin/env python3 +# +# Copyright 2024 Author: Yuekai Zhang +# +# 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 file uses whisper and zipformer decoding results to generate fusion decoding results. +Since whisper model is more likely to make deletion errors and zipformer model is more likely to make substitution and insertion errors, +we trust whisper model when it makes substitution and insertion errors and trust zipformer model when it makes deletion errors. + +Usage: + python whisper_zipformer_fusion.py --whisper-log-dir ./whisper_decoding_log_dir --zipformer-log-dir ./zipformer_decoding_log_dir --output-log-dir ./results_fusion +""" + +import argparse +from pathlib import Path +from typing import Dict, List, Optional, Tuple + +import kaldialign + +from icefall.utils import store_transcripts, write_error_stats + + +def get_parser(): + parser = argparse.ArgumentParser( + formatter_class=argparse.ArgumentDefaultsHelpFormatter + ) + parser.add_argument( + "--whisper-log-dir", + type=str, + default="./recogs_whisper", + help="The directory to store the whisper logs: e.g. recogs-SPEECHIO_ASR_ZH00014-beam-search-epoch--1-avg-1.txt", + ) + parser.add_argument( + "--zipformer-log-dir", + type=str, + default="./recogs_zipformer", + help="The directory to store the zipformer logs", + ) + parser.add_argument( + "--output-log-dir", + type=str, + default="./results_fusion", + help="The directory to store the fusion logs", + ) + return parser + + +def save_results( + res_dir: Path, + test_set_name: str, + results_dict: Dict[str, List[Tuple[str, List[str], List[str]]]], +): + test_set_wers = dict() + + suffix = "epoch-999-avg-1" + + for key, results in results_dict.items(): + recog_path = res_dir / f"recogs-{test_set_name}-{key}-{suffix}.txt" + results = sorted(results) + store_transcripts(filename=recog_path, texts=results) + print(f"The transcripts are stored in {recog_path}") + + # The following prints out WERs, per-word error statistics and aligned + # ref/hyp pairs. + errs_filename = res_dir / f"errs-{test_set_name}-{key}-{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 + + print("Wrote detailed error stats to {}".format(errs_filename)) + + test_set_wers = sorted(test_set_wers.items(), key=lambda x: x[1]) + errs_info = res_dir / f"wer-summary-{test_set_name}-{key}-{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 = "" + print(s) + + +def extract_hyp_ref_wavname(filename): + """ + 0Phqz8RWYuE_0007-5: ref=['R', 'Y', 'Y', 'B', '它最大的优势就是进光量或者说是对光线利用率的提升'] + 0Phqz8RWYuE_0007-5: hyp=而YB它最大的优势是近光量或者说是对光线利用率的提升 + """ + hyps, refs, wav_name = [], [], [] + with open(filename, "r") as f: + for line in f: + if "ref" in line: + ref = line.split("ref=")[1].strip() + ref = ref[2:-2] + list_elements = ref.split("', '") + ref = "".join(list_elements) + refs.append(ref) + elif "hyp" in line: + hyp = line.split("hyp=")[1].strip() + hyps.append(hyp) + wav_name.append(line.split(":")[0]) + return hyps, refs, wav_name + + +def get_pair_filenames( + whisper_log_dir, + zipformer_log_dir, + whisper_suffix="beam-search-epoch-999-avg-1", + zipformer_suffix="greedy_search_blank_penalty_2.0-epoch-999-avg-1-context-2-max-sym-per-frame-1-blank-penalty-2.0", +): + results = [] + start_index, end_index = 0, 26 + dataset_parts = [] + for i in range(start_index, end_index + 1): + idx = f"{i}".zfill(2) + dataset_parts.append(f"SPEECHIO_ASR_ZH000{idx}") + for partition in dataset_parts: + whisper_filename = f"{whisper_log_dir}/recogs-{partition}-{whisper_suffix}.txt" + zipformer_filename = ( + f"{zipformer_log_dir}/recogs-{partition}-{zipformer_suffix}.txt" + ) + results.append((whisper_filename, zipformer_filename)) + return results + + +def fusion_hyps_trust_substituion_insertion( + hyps_whisper, hyps_zipformer, refs, ERR="*" +): + """ + alignment example: + [('我', '你'), ('在', '*'), ('任', '任'), ('的', '的'), ('时', '时'), ('候', '候'), ('*', '呢')] + left is whisper, right is zipformer + for whisper substitution, use left + for whisper insertion, use left + for whisper deletion, use right + """ + hyps_fusion = [] + for hyp_w, hyp_z, ref in zip(hyps_whisper, hyps_zipformer, refs): + ali = kaldialign.align(hyp_w, hyp_z, ERR) + hyp_f = "" + for a in ali: + if a[0] == ERR: + hyp_f += a[1] + else: + hyp_f += a[0] + hyps_fusion.append(hyp_f) + return hyps_fusion + + +def fusion_hyps_trust_substituion(hyps_whisper, hyps_zipformer, refs, ERR="*"): + """ + alignment example: + [('我', '你'), ('在', '*'), ('任', '任'), ('的', '的'), ('时', '时'), ('候', '候'), ('*', '呢')] + left is whisper, right is zipformer + for whisper substitution, use left + for whisper insertion, use right + for whisper deletion, use right + """ + hyps_fusion = [] + for hyp_w, hyp_z, ref in zip(hyps_whisper, hyps_zipformer, refs): + ali = kaldialign.align(hyp_w, hyp_z, ERR) + hyp_f = "" + for a in ali: + if a[0] == ERR: + hyp_f += a[1] + elif a[1] == ERR: + pass + else: + hyp_f += a[0] + hyps_fusion.append(hyp_f) + return hyps_fusion + + +def main(): + parser = get_parser() + args = parser.parse_args() + # mkdir output_log_dir + Path(args.output_log_dir).mkdir(parents=True, exist_ok=True) + pair_logs = get_pair_filenames(args.whisper_log_dir, args.zipformer_log_dir) + for pair in pair_logs: + hyps_whisper, refs, wav_name = extract_hyp_ref_wavname(pair[0]) + hyps_zipformer, _, _ = extract_hyp_ref_wavname(pair[1]) + + hyps_fusion = fusion_hyps_trust_substituion_insertion( + hyps_whisper, hyps_zipformer, refs + ) + + partition_name = pair[0].split("/")[-1].split("-")[1] + save_results( + Path(args.output_log_dir), + partition_name, + {"fusion": list(zip(wav_name, refs, hyps_fusion))}, + ) + + print(f"Processed {partition_name}") + + +if __name__ == "__main__": + main() diff --git a/egs/speechio/ASR/prepare.sh b/egs/speechio/ASR/prepare.sh new file mode 100644 index 000000000..048a66d8f --- /dev/null +++ b/egs/speechio/ASR/prepare.sh @@ -0,0 +1,67 @@ +#!/usr/bin/env bash + +set -eou pipefail + +stage=3 +stop_stage=3 + +# We assume dl_dir (download dir) contains the following +# directories and files. If not, they will be downloaded +# by this script automatically. +# +# - $dl_dir/SPEECHIO_ASR_ZH00000 +# This directory contains the following files downloaded from +# https://github.com/SpeechColab/Leaderboard +# +# - metadata.tsv +# - wav +# - wav.scp +# - trans.txt +# + +dl_dir=$PWD/download + +. shared/parse_options.sh || exit 1 + +# All files generated by this script are saved in "data". +# You can safely remove "data" and rerun this script to regenerate it. +mkdir -p data + +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]}) $*" +} + +log "dl_dir: $dl_dir" + +if [ $stage -le 1 ] && [ $stop_stage -ge 1 ]; then + log "Stage 1: Prepare speechio manifest" + # We assume that you have downloaded the speechio dataset + # to $dl_dir + mkdir -p data/manifests + if [ ! -e data/manifests/.speechio.done ]; then + lhotse prepare speechio $dl_dir data/manifests + touch data/manifests/.speechio.done + fi +fi + +whisper_mel_bins=80 +if [ $stage -le 2 ] && [ $stop_stage -ge 2 ]; then + log "Stage 2: Compute whisper fbank for speechio" + if [ ! -f data/fbank/.speechio.done ]; then + mkdir -p data/fbank + ./local/compute_fbank_speechio.py --num-mel-bins ${whisper_mel_bins} --whisper-fbank true + touch data/fbank/.speechio.done + fi +fi + +if [ $stage -le 3 ] && [ $stop_stage -ge 3 ]; then + log "Stage 3: Compute kaldi fbank for speechio" + if [ ! -f data/fbank/.speechio.kaldi.done ]; then + fbank_dir=data/fbank_kaldi + mkdir -p $fbank_dir + ./local/compute_fbank_speechio.py --fbank-dir $fbank_dir + touch data/fbank/.speechio.kaldi.done + fi +fi diff --git a/egs/speechio/ASR/shared b/egs/speechio/ASR/shared new file mode 120000 index 000000000..9d8803a7d --- /dev/null +++ b/egs/speechio/ASR/shared @@ -0,0 +1 @@ +../../../icefall/shared// \ No newline at end of file diff --git a/egs/speechio/ASR/whisper/asr_datamodule.py b/egs/speechio/ASR/whisper/asr_datamodule.py new file mode 100644 index 000000000..7382fd3f5 --- /dev/null +++ b/egs/speechio/ASR/whisper/asr_datamodule.py @@ -0,0 +1,195 @@ +# Copyright 2021 Piotr Żelasko +# Copyright 2022 Xiaomi Corporation (Author: Mingshuang Luo) +# +# 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. + + +import argparse +import inspect +import logging +from functools import lru_cache +from pathlib import Path +from typing import Any, Dict, Optional + +import torch +from lhotse import CutSet, load_manifest, load_manifest_lazy +from lhotse.dataset import ( # noqa F401 for PrecomputedFeatures + CutConcatenate, + CutMix, + DynamicBucketingSampler, + K2SpeechRecognitionDataset, + PrecomputedFeatures, + SimpleCutSampler, + SpecAugment, +) +from lhotse.dataset.input_strategies import AudioSamples # noqa F401 For AudioSamples +from lhotse.utils import fix_random_seed +from torch.utils.data import DataLoader + +from icefall.utils import str2bool + + +class _SeedWorkers: + def __init__(self, seed: int): + self.seed = seed + + def __call__(self, worker_id: int): + fix_random_seed(self.seed + worker_id) + + +class AsrDataModule: + """ + DataModule for k2 ASR experiments. + There is no train and valid dataloader, for speechio dataset + but there can be multiple test dataloaders. + + It contains all the common data pipeline modules used in ASR + experiments, e.g.: + - dynamic batch size, + - bucketing samplers, + - cut concatenation, + + This class should be derived for specific corpora used in ASR tasks. + """ + + def __init__(self, args: argparse.Namespace): + self.args = args + + @classmethod + def add_arguments(cls, parser: argparse.ArgumentParser): + group = parser.add_argument_group( + title="ASR data related options", + description="These options are used for the preparation of " + "PyTorch DataLoaders from Lhotse CutSet's -- they control the " + "effective batch sizes, sampling strategies, applied data " + "augmentations, etc.", + ) + group.add_argument( + "--manifest-dir", + type=Path, + default=Path("data/fbank"), + help="Path to directory with train/valid/test cuts.", + ) + group.add_argument( + "--max-duration", + type=int, + default=300.0, + help="Maximum pooled recordings duration (seconds) in a " + "single batch. You can reduce it if it causes CUDA OOM.", + ) + group.add_argument( + "--bucketing-sampler", + type=str2bool, + default=True, + help="When enabled, the batches will come from buckets of " + "similar duration (saves padding frames).", + ) + group.add_argument( + "--num-buckets", + type=int, + default=30, + help="The number of buckets for the DynamicBucketingSampler" + "(you might want to increase it for larger datasets).", + ) + group.add_argument( + "--concatenate-cuts", + type=str2bool, + default=False, + help="When enabled, utterances (cuts) will be concatenated " + "to minimize the amount of padding.", + ) + group.add_argument( + "--duration-factor", + type=float, + default=1.0, + help="Determines the maximum duration of a concatenated cut " + "relative to the duration of the longest cut in a batch.", + ) + group.add_argument( + "--gap", + type=float, + default=1.0, + help="The amount of padding (in seconds) inserted between " + "concatenated cuts. This padding is filled with noise when " + "noise augmentation is used.", + ) + group.add_argument( + "--shuffle", + type=str2bool, + default=True, + help="When enabled (=default), the examples will be " + "shuffled for each epoch.", + ) + group.add_argument( + "--drop-last", + type=str2bool, + default=True, + help="Whether to drop last batch. Used by sampler.", + ) + group.add_argument( + "--return-cuts", + type=str2bool, + default=True, + help="When enabled, each batch will have the " + "field: batch['supervisions']['cut'] with the cuts that " + "were used to construct it.", + ) + + group.add_argument( + "--num-workers", + type=int, + default=2, + help="The number of training dataloader workers that " + "collect the batches.", + ) + group.add_argument( + "--input-strategy", + type=str, + default="PrecomputedFeatures", + help="AudioSamples or PrecomputedFeatures", + ) + parser.add_argument( + "--start-index", + type=int, + default=0, + help="Decoding will start from dataset SPEECHIO_ASR_ZH000index", + ) + + parser.add_argument( + "--end-index", + type=int, + default=26, + help="Decoding will end with dataset SPEECHIO_ASR_ZH000index", + ) + + def test_dataloaders(self, cuts: CutSet) -> DataLoader: + logging.debug("About to create test dataset") + test = K2SpeechRecognitionDataset( + input_strategy=eval(self.args.input_strategy)(), + return_cuts=self.args.return_cuts, + ) + sampler = DynamicBucketingSampler( + cuts, + max_duration=self.args.max_duration, + shuffle=False, + ) + logging.debug("About to create test dataloader") + test_dl = DataLoader( + test, + batch_size=None, + sampler=sampler, + num_workers=self.args.num_workers, + ) + return test_dl diff --git a/egs/speechio/ASR/whisper/decode.py b/egs/speechio/ASR/whisper/decode.py new file mode 100644 index 000000000..001367791 --- /dev/null +++ b/egs/speechio/ASR/whisper/decode.py @@ -0,0 +1,520 @@ +#!/usr/bin/env python3 +# Copyright 2021 Xiaomi Corporation (Author: Liyong Guo, +# Fangjun Kuang, +# Wei Kang) +# 2024 Yuekai Zhang +# +# 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: +# Command for decoding using fine-tuned models: +git lfs install +git clone https://huggingface.co/yuekai/icefall_asr_aishell_whisper +ln -s icefall_asr_aishell_whisper/exp_large_v2/epoch-10-avg6.pt whisper/exp_large_v2/epoch-999.pt + +python3 ./whisper/decode.py \ + --exp-dir whisper/exp_large_v2 \ + --model-name large-v2 \ + --epoch 999 --avg 1 \ + --beam-size 10 --max-duration 50 + +# Command for decoding using pretrained models (before fine-tuning): + +python3 ./whisper/decode.py \ + --exp-dir whisper/exp_large_v2_pretrained \ + --model-name large-v2 \ + --epoch -1 --avg 1 \ + --start-index 14 --end-index 15 \ + --remove-whisper-encoder-input-length-restriction False \ + --beam-size 1 --max-duration 50 + +""" + +import argparse +import logging +import re +from collections import defaultdict +from pathlib import Path +from typing import Dict, List, Optional, Tuple + +import k2 +import torch +import torch.nn as nn +import whisper +from asr_datamodule import AsrDataModule +from lhotse.cut import Cut +from multi_dataset import MultiDataset +from tn.chinese.normalizer import Normalizer +from whisper.normalizers import BasicTextNormalizer +from whisper_encoder_forward_monkey_patch import replace_whisper_encoder_forward +from zhconv import convert + +from icefall.checkpoint import average_checkpoints_with_averaged_model, load_checkpoint +from icefall.env import get_env_info +from icefall.utils import ( + AttributeDict, + setup_logger, + store_transcripts, + str2bool, + write_error_stats, +) + + +def average_checkpoints( + filenames: List[Path], device: torch.device = torch.device("cpu") +) -> dict: + """Average a list of checkpoints. + The function is mainly used for deepspeed converted checkpoint averaging, which only include model state_dict. + + Args: + filenames: + Filenames of the checkpoints to be averaged. We assume all + checkpoints are saved by :func:`save_checkpoint`. + device: + Move checkpoints to this device before averaging. + Returns: + Return a dict (i.e., state_dict) which is the average of all + model state dicts contained in the checkpoints. + """ + n = len(filenames) + + if "model" in torch.load(filenames[0], map_location=device): + avg = torch.load(filenames[0], map_location=device)["model"] + else: + avg = torch.load(filenames[0], map_location=device) + + # 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 avg.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 i in range(1, n): + if "model" in torch.load(filenames[i], map_location=device): + state_dict = torch.load(filenames[i], map_location=device)["model"] + else: + state_dict = torch.load(filenames[i], map_location=device) + for k in uniqued_names: + avg[k] += state_dict[k] + + for k in uniqued_names: + if avg[k].is_floating_point(): + avg[k] /= n + else: + avg[k] //= n + + return avg + + +def remove_punctuation(text: str or List[str]): + """Modified from https://github.com/yeyupiaoling/Whisper-Finetune/blob/master/utils/data_utils.py + + Args: + text: It can be a string or a list of strings. + Returns: + Return a string or a list of strings without any punctuation. + """ + punctuation = "!,.;:?、！，。；：？《》 " + if isinstance(text, str): + text = re.sub(r"[{}]+".format(punctuation), "", text).strip() + return text + elif isinstance(text, list): + result_text = [] + for t in text: + t = re.sub(r"[{}]+".format(punctuation), "", t).strip() + result_text.append(t) + return result_text + else: + raise Exception(f"Not support type {type(text)}") + + +def to_simple(text: str or List[str]): + """Convert traditional Chinese to simplified Chinese. + Args: + text: It can be a string or a list of strings. + Returns: + Return a string or a list of strings converted to simplified Chinese. + """ + if isinstance(text, str): + text = convert(text, "zh-cn") + return text + elif isinstance(text, list): + result_text = [] + for t in text: + t = convert(t, "zh-cn") + result_text.append(t) + return result_text + else: + raise Exception(f"Not support type{type(text)}") + + +def get_parser(): + parser = argparse.ArgumentParser( + formatter_class=argparse.ArgumentDefaultsHelpFormatter + ) + + parser.add_argument( + "--epoch", + type=int, + default=-1, + help="It specifies the checkpoint to use for decoding." + "Note: Epoch counts from 0.", + ) + parser.add_argument( + "--avg", + type=int, + default=1, + help="Number of checkpoints to average. Automatically select " + "consecutive checkpoints before the checkpoint specified by " + "'--epoch'. ", + ) + + parser.add_argument( + "--method", + type=str, + default="beam-search", + help="""Decoding method. + Supported values are: + - beam-search + """, + ) + + parser.add_argument( + "--beam-size", + type=int, + default=1, + help="beam size for beam search decoding", + ) + + parser.add_argument( + "--exp-dir", + type=str, + default="whisper/exp", + help="The experiment dir", + ) + + parser.add_argument( + "--model-name", + type=str, + default="large-v2", + choices=["large-v2", "large-v3", "medium", "small", "tiny"], + help="""The model name to use. + """, + ) + + parser.add_argument( + "--remove-whisper-encoder-input-length-restriction", + type=str2bool, + default=True, + help="replace whisper encoder forward method to remove input length restriction", + ) + + return parser + + +def get_params() -> AttributeDict: + params = AttributeDict( + { + "env_info": get_env_info(), + } + ) + return params + + +def decode_one_batch( + params: AttributeDict, + model: nn.Module, + batch: dict, +) -> Dict[str, List[List[int]]]: + """Decode one batch and return the result in a dict. The dict has the + following format: + + - key: "beam-search" + - value: A list of lists. Each sublist is a list of token IDs. + Args: + params: + It is returned by :func:`get_params`. + model: + The neural model. + batch: + It is returned by :meth:`torch.utils.data.DataLoader.__iter__`. + Returns: + Return a dict, whose key may be "beam-search". + """ + dtype = torch.float16 + device = torch.device("cuda") + + feature = batch["inputs"] + assert feature.ndim == 3 + feature = feature.to(device, dtype=dtype).transpose(1, 2) + if not params.remove_whisper_encoder_input_length_restriction: + T = 3000 + if feature.shape[2] < T: + feature = torch.cat( + [ + feature, + torch.zeros( + feature.shape[0], feature.shape[1], T - feature.shape[2] + ).to(device, dtype=dtype), + ], + 2, + ) + + supervisions = batch["supervisions"] + feature_len = supervisions["num_frames"] + feature_len = feature_len.to(device, dtype=dtype) + results = model.decode(feature, params.decoding_options) + hyps = [result.text for result in results] + + hyps = remove_punctuation(hyps) + hyps = to_simple(hyps) + hyps = [params.normalizer.normalize(hyp) for hyp in hyps] + print(hyps) + return {"beam-search": hyps} + + +def decode_dataset( + dl: torch.utils.data.DataLoader, + params: AttributeDict, + model: nn.Module, +) -> Dict[str, List[Tuple[str, List[str], List[str]]]]: + """Decode dataset. + + Args: + dl: + The dataloader. + params: + It is returned by :func:`get_params`. + model: + The neural model. + Returns: + Return a dict, whose key may be "beam-search". + """ + results = [] + + 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, + batch=batch, + ) + + for lm_scale, 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[lm_scale].extend(this_batch) + + num_cuts += len(batch["supervisions"]["text"]) + + 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]]]], +): + + enable_log = True + test_set_wers = dict() + for key, results in results_dict.items(): + recog_path = ( + params.exp_dir / f"recogs-{test_set_name}-{key}-{params.suffix}.txt" + ) + results = sorted(results) + store_transcripts(filename=recog_path, texts=results) + if enable_log: + 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.exp_dir / f"errs-{test_set_name}-{key}-{params.suffix}.txt" + ) + # we compute CER for aishell dataset. + results_char = [] + for res in results: + results_char.append((res[0], list("".join(res[1])), list("".join(res[2])))) + with open(errs_filename, "w") as f: + wer = write_error_stats( + f, f"{test_set_name}-{key}", results_char, enable_log=enable_log + ) + test_set_wers[key] = wer + + if enable_log: + 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.exp_dir / f"cer-summary-{test_set_name}-{params.suffix}.txt" + with open(errs_info, "w") as f: + print("settings\tCER", file=f) + for key, val in test_set_wers: + print("{}\t{}".format(key, val), file=f) + + s = "\nFor {}, CER 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() + AsrDataModule.add_arguments(parser) + args = parser.parse_args() + args.exp_dir = Path(args.exp_dir) + + params = get_params() + params.update(vars(args)) + params.suffix = f"epoch-{params.epoch}-avg-{params.avg}" + setup_logger( + f"{params.exp_dir}/log-{params.method}-beam{params.beam_size}/log-decode-{params.suffix}" + ) + + options = whisper.DecodingOptions( + task="transcribe", + language="zh", + without_timestamps=True, + beam_size=params.beam_size, + ) + params.decoding_options = options + params.cleaner = BasicTextNormalizer() + params.normalizer = Normalizer() + + logging.info("Decoding started") + logging.info(params) + + device = torch.device("cpu") + if torch.cuda.is_available(): + device = torch.device("cuda") + + logging.info(f"device: {device}") + + if params.remove_whisper_encoder_input_length_restriction: + replace_whisper_encoder_forward() + model = whisper.load_model(params.model_name, "cpu") + if params.epoch > 0: + if params.avg > 1: + start = params.epoch - params.avg + assert start >= 1, start + checkpoint = torch.load( + f"{params.exp_dir}/epoch-{params.epoch}.pt", map_location="cpu" + ) + if "model" not in checkpoint: + # deepspeed converted checkpoint only contains model state_dict + filenames = [ + f"{params.exp_dir}/epoch-{epoch}.pt" + for epoch in range(start, params.epoch + 1) + ] + model.load_state_dict(average_checkpoints(filenames)) + else: + 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, + ) + ) + # save checkpoints + filename = f"{params.exp_dir}/epoch-{params.epoch}-avg-{params.avg}.pt" + torch.save(model.state_dict(), filename) + else: + checkpoint = torch.load( + f"{params.exp_dir}/epoch-{params.epoch}.pt", map_location="cpu" + ) + if "model" not in checkpoint: + model.load_state_dict(checkpoint, strict=True) + else: + load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model) + 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 + + data_module = AsrDataModule(args) + multi_dataset = MultiDataset(args.manifest_dir, args.start_index, args.end_index) + + def remove_long_utt(c: Cut): + # Keep only utterances with duration in 30 seconds + # + if c.duration > 30.0: + # logging.warning( + # f"Exclude cut with ID {c.id} from training. Duration: {c.duration}" + # ) + return False + return True + + test_sets_cuts = multi_dataset.test_cuts() + + test_sets = test_sets_cuts.keys() + test_dls = [ + data_module.test_dataloaders(test_sets_cuts[cuts_name].filter(remove_long_utt)) + for cuts_name in test_sets + ] + + for test_set, test_dl in zip(test_sets, test_dls): + results_dict = decode_dataset( + dl=test_dl, + params=params, + model=model, + ) + + save_results(params=params, test_set_name=test_set, results_dict=results_dict) + + logging.info("Done!") + + +torch.set_num_threads(1) +torch.set_num_interop_threads(1) + +if __name__ == "__main__": + main() diff --git a/egs/speechio/ASR/whisper/multi_dataset.py b/egs/speechio/ASR/whisper/multi_dataset.py new file mode 100644 index 000000000..f55d45394 --- /dev/null +++ b/egs/speechio/ASR/whisper/multi_dataset.py @@ -0,0 +1,59 @@ +# Copyright 2023 Xiaomi Corp. (authors: Zengrui Jin) +# +# 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. +import glob +import logging +import re +from pathlib import Path +from typing import Dict, List + +import lhotse +from lhotse import CutSet, load_manifest_lazy + + +class MultiDataset: + def __init__(self, fbank_dir: str, start_index: int = 0, end_index: int = 26): + """ + Args: + manifest_dir: + It is expected to contain the following files: + - speechio_cuts_SPEECHIO_ASR_ZH00000.jsonl.gz + ... + - speechio_cuts_SPEECHIO_ASR_ZH00026.jsonl.gz + """ + self.fbank_dir = Path(fbank_dir) + self.start_index = start_index + self.end_index = end_index + + def test_cuts(self) -> Dict[str, CutSet]: + logging.info("About to get multidataset test cuts") + + dataset_parts = [] + for i in range(self.start_index, self.end_index + 1): + idx = f"{i}".zfill(2) + dataset_parts.append(f"SPEECHIO_ASR_ZH000{idx}") + + prefix = "speechio" + suffix = "jsonl.gz" + + results_dict = {} + for partition in dataset_parts: + path = f"{prefix}_cuts_{partition}.{suffix}" + + logging.info(f"Loading {path} set in lazy mode") + test_cuts = load_manifest_lazy(self.fbank_dir / path) + results_dict[partition] = test_cuts + + return results_dict diff --git a/egs/speechio/ASR/whisper/requirements.txt b/egs/speechio/ASR/whisper/requirements.txt new file mode 120000 index 000000000..744bf8bb6 --- /dev/null +++ b/egs/speechio/ASR/whisper/requirements.txt @@ -0,0 +1 @@ +../../../aishell/ASR/whisper/requirements.txt \ No newline at end of file diff --git a/egs/speechio/ASR/whisper/whisper_encoder_forward_monkey_patch.py b/egs/speechio/ASR/whisper/whisper_encoder_forward_monkey_patch.py new file mode 120000 index 000000000..2a7808921 --- /dev/null +++ b/egs/speechio/ASR/whisper/whisper_encoder_forward_monkey_patch.py @@ -0,0 +1 @@ +../../../aishell/ASR/whisper/whisper_encoder_forward_monkey_patch.py \ No newline at end of file diff --git a/egs/speechio/ASR/zipformer/asr_datamodule.py b/egs/speechio/ASR/zipformer/asr_datamodule.py new file mode 120000 index 000000000..bf446dabe --- /dev/null +++ b/egs/speechio/ASR/zipformer/asr_datamodule.py @@ -0,0 +1 @@ +../whisper/asr_datamodule.py \ No newline at end of file diff --git a/egs/speechio/ASR/zipformer/beam_search.py b/egs/speechio/ASR/zipformer/beam_search.py new file mode 120000 index 000000000..8e2c0a65c --- /dev/null +++ b/egs/speechio/ASR/zipformer/beam_search.py @@ -0,0 +1 @@ +../../../librispeech/ASR/zipformer/beam_search.py \ No newline at end of file diff --git a/egs/speechio/ASR/zipformer/ctc_decode.py b/egs/speechio/ASR/zipformer/ctc_decode.py new file mode 100644 index 000000000..f9d0db993 --- /dev/null +++ b/egs/speechio/ASR/zipformer/ctc_decode.py @@ -0,0 +1,623 @@ +#!/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 +./zipformer/ctc_decode.py \ + --epoch 30 \ + --avg 15 \ + --exp-dir ./zipformer/exp \ + --max-duration 600 \ + --decoding-method ctc-decoding + +""" + + +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 AsrDataModule +from lhotse.cut import Cut +from multi_dataset import MultiDataset +from train import add_model_arguments, get_model, get_params + +from icefall.checkpoint import ( + average_checkpoints, + average_checkpoints_with_averaged_model, + find_checkpoints, + load_checkpoint, +) +from icefall.decode import get_lattice, one_best_decoding +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="zipformer/exp", + help="The experiment dir", + ) + + parser.add_argument( + "--bpe-model", + type=str, + default="data/lang_bpe_2000/bpe.model", + help="Path to the BPE model", + ) + + parser.add_argument( + "--lang-dir", + type=Path, + default="data/lang_bpe_2000", + 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. + """, + ) + + 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=1.0, + 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. + """, + ) + + 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, + H: Optional[k2.Fsa], + bpe_model: Optional[spm.SentencePieceProcessor], + batch: dict, +) -> 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. + 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. + """ + 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) + + if params.causal: + # this seems to cause insertions at the end of the utterance if used with zipformer. + pad_len = 30 + feature_lens += pad_len + feature = torch.nn.functional.pad( + feature, + pad=(0, 0, 0, pad_len), + value=LOG_EPS, + ) + + encoder_out, encoder_out_lens = model.forward_encoder(feature, feature_lens) + ctc_output = model.ctc_output(encoder_out) # (N, T, C) + + supervision_segments = torch.stack( + ( + supervisions["sequence_idx"], + torch.div( + supervisions["start_frame"], + params.subsampling_factor, + rounding_mode="floor", + ), + torch.div( + supervisions["num_frames"], + params.subsampling_factor, + rounding_mode="floor", + ), + ), + 1, + ).to(torch.int32) + + assert bpe_model is not None + decoding_graph = H + + lattice = get_lattice( + nnet_output=ctc_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} + + +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, + H=H, + bpe_model=bpe_model, + 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 = list(ref_text.replace(" ", "")) + hyp_words = list("".join(hyp_words)) + 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}-{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}-{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}-{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() + AsrDataModule.add_arguments(parser) + args = parser.parse_args() + args.exp_dir = Path(args.exp_dir) + args.lang_dir = Path(args.lang_dir) + + params = get_params() + # add decoding params + params.update(get_decoding_params()) + params.update(vars(args)) + + assert params.decoding_method in ("ctc-decoding",) + 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.causal: + assert ( + "," not in params.chunk_size + ), "chunk_size should be one value in decoding." + assert ( + "," not in params.left_context_frames + ), "left_context_frames should be one value in decoding." + params.suffix += f"-chunk-{params.chunk_size}" + params.suffix += f"-left-context-{params.left_context_frames}" + + 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 + # and are defined in local/train_bpe_model.py + params.blank_id = 0 + + HLG = None + H = k2.ctc_topo( + max_token=max_token_id, + modified=True, + device=device, + ) + bpe_model = spm.SentencePieceProcessor() + bpe_model.load(str(params.lang_dir / "bpe.model")) + + G = None + logging.info("About to create model") + model = get_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 + data_module = AsrDataModule(args) + multi_dataset = MultiDataset(args.manifest_dir, args.start_index, args.end_index) + + test_sets_cuts = multi_dataset.test_cuts() + + def remove_short_utt(c: Cut): + T = ((c.num_frames - 7) // 2 + 1) // 2 + if T <= 0: + logging.warning( + f"Excluding cut with ID: {c.id} from decoding, num_frames: {c.num_frames}" + ) + return T > 0 + + test_sets = test_sets_cuts.keys() + test_dl = [ + data_module.test_dataloaders(test_sets_cuts[cuts_name].filter(remove_short_utt)) + for cuts_name in test_sets + ] + + for test_set, test_dl in zip(test_sets, test_dl): + logging.info(f"Start decoding test set: {test_set}") + + 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() diff --git a/egs/speechio/ASR/zipformer/decode.py b/egs/speechio/ASR/zipformer/decode.py new file mode 100644 index 000000000..ffdd7b500 --- /dev/null +++ b/egs/speechio/ASR/zipformer/decode.py @@ -0,0 +1,843 @@ +#!/usr/bin/env python3 +# +# Copyright 2021-2023 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 +./zipformer/decode.py \ + --epoch 28 \ + --avg 15 \ + --exp-dir ./zipformer/exp \ + --max-duration 600 \ + --decoding-method greedy_search + +(2) beam search (not recommended) +./zipformer/decode.py \ + --epoch 28 \ + --avg 15 \ + --exp-dir ./zipformer/exp \ + --max-duration 600 \ + --decoding-method beam_search \ + --beam-size 4 + +(3) modified beam search +./zipformer/decode.py \ + --epoch 28 \ + --avg 15 \ + --exp-dir ./zipformer/exp \ + --max-duration 600 \ + --decoding-method modified_beam_search \ + --beam-size 4 + +(4) fast beam search (one best) +./zipformer/decode.py \ + --epoch 28 \ + --avg 15 \ + --exp-dir ./zipformer/exp \ + --max-duration 600 \ + --decoding-method fast_beam_search \ + --beam 20.0 \ + --max-contexts 8 \ + --max-states 64 + +(5) fast beam search (nbest) +./zipformer/decode.py \ + --epoch 28 \ + --avg 15 \ + --exp-dir ./zipformer/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) +./zipformer/decode.py \ + --epoch 28 \ + --avg 15 \ + --exp-dir ./zipformer/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) +./zipformer/decode.py \ + --epoch 28 \ + --avg 15 \ + --exp-dir ./zipformer/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 AsrDataModule +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 lhotse.cut import Cut +from multi_dataset import MultiDataset +from train import add_model_arguments, get_model, get_params + +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, + make_pad_mask, + 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="zipformer/exp", + help="The experiment dir", + ) + + parser.add_argument( + "--bpe-model", + type=str, + default="data/lang_bpe_2000/bpe.model", + help="Path to the BPE model", + ) + + parser.add_argument( + "--lang-dir", + type=Path, + default="data/lang_bpe_2000", + 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( + "--blank-penalty", + type=float, + default=0.0, + help=""" + The penalty applied on blank symbol during decoding. + Note: It is a positive value that would be applied to logits like + this `logits[:, 0] -= blank_penalty` (suppose logits.shape is + [batch_size, vocab] and blank id is 0). + """, + ) + 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.causal: + # this seems to cause insertions at the end of the utterance if used with zipformer. + pad_len = 30 + feature_lens += pad_len + feature = torch.nn.functional.pad( + feature, + pad=(0, 0, 0, pad_len), + value=LOG_EPS, + ) + + encoder_out, encoder_out_lens = model.forward_encoder(feature, 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, + blank_penalty=params.blank_penalty, + ) + 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, + blank_penalty=params.blank_penalty, + ) + 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()) + + key = f"blank_penalty_{params.blank_penalty}" + if params.decoding_method == "greedy_search": + return {"greedy_search_" + key: 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"] + texts = [list(str(text).replace(" ", "")) for text in texts] + 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): + hyp_text = "".join(hyp_words) + this_batch.append((cut_id, ref_text, hyp_text)) + + 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() + AsrDataModule.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.causal: + assert ( + "," not in params.chunk_size + ), "chunk_size should be one value in decoding." + assert ( + "," not in params.left_context_frames + ), "left_context_frames should be one value in decoding." + params.suffix += f"-chunk-{params.chunk_size}" + params.suffix += f"-left-context-{params.left_context_frames}" + + 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}" + + params.suffix += f"-blank-penalty-{params.blank_penalty}" + 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) + + # and are defined in local/train_bpe_model.py + params.blank_id = sp.piece_to_id("") + params.unk_id = sp.piece_to_id("") + params.vocab_size = sp.get_piece_size() + + logging.info(params) + + logging.info("About to create model") + model = get_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 + data_module = AsrDataModule(args) + multi_dataset = MultiDataset(args.manifest_dir, args.start_index, args.end_index) + + def remove_short_utt(c: Cut): + T = ((c.num_frames - 7) // 2 + 1) // 2 + if T <= 0: + logging.warning( + f"Excluding cut with ID: {c.id} from decoding, num_frames: {c.num_frames}" + ) + return T > 0 + + test_sets_cuts = multi_dataset.test_cuts() + + test_sets = test_sets_cuts.keys() + test_dl = [ + data_module.test_dataloaders(test_sets_cuts[cuts_name].filter(remove_short_utt)) + for cuts_name in test_sets + ] + + for test_set, test_dl in zip(test_sets, test_dl): + logging.info(f"Start decoding test set: {test_set}") + + 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() diff --git a/egs/speechio/ASR/zipformer/decoder.py b/egs/speechio/ASR/zipformer/decoder.py new file mode 120000 index 000000000..5a8018680 --- /dev/null +++ b/egs/speechio/ASR/zipformer/decoder.py @@ -0,0 +1 @@ +../../../librispeech/ASR/zipformer/decoder.py \ No newline at end of file diff --git a/egs/speechio/ASR/zipformer/encoder_interface.py b/egs/speechio/ASR/zipformer/encoder_interface.py new file mode 120000 index 000000000..c2eaca671 --- /dev/null +++ b/egs/speechio/ASR/zipformer/encoder_interface.py @@ -0,0 +1 @@ +../../../librispeech/ASR/zipformer/encoder_interface.py \ No newline at end of file diff --git a/egs/speechio/ASR/zipformer/joiner.py b/egs/speechio/ASR/zipformer/joiner.py new file mode 120000 index 000000000..5b8a36332 --- /dev/null +++ b/egs/speechio/ASR/zipformer/joiner.py @@ -0,0 +1 @@ +../../../librispeech/ASR/zipformer/joiner.py \ No newline at end of file diff --git a/egs/speechio/ASR/zipformer/model.py b/egs/speechio/ASR/zipformer/model.py new file mode 120000 index 000000000..cd7e07d72 --- /dev/null +++ b/egs/speechio/ASR/zipformer/model.py @@ -0,0 +1 @@ +../../../librispeech/ASR/zipformer/model.py \ No newline at end of file diff --git a/egs/speechio/ASR/zipformer/multi_dataset.py b/egs/speechio/ASR/zipformer/multi_dataset.py new file mode 120000 index 000000000..af164667a --- /dev/null +++ b/egs/speechio/ASR/zipformer/multi_dataset.py @@ -0,0 +1 @@ +../whisper/multi_dataset.py \ No newline at end of file diff --git a/egs/speechio/ASR/zipformer/optim.py b/egs/speechio/ASR/zipformer/optim.py new file mode 120000 index 000000000..5eaa3cffd --- /dev/null +++ b/egs/speechio/ASR/zipformer/optim.py @@ -0,0 +1 @@ +../../../librispeech/ASR/zipformer/optim.py \ No newline at end of file diff --git a/egs/speechio/ASR/zipformer/scaling.py b/egs/speechio/ASR/zipformer/scaling.py new file mode 120000 index 000000000..6f398f431 --- /dev/null +++ b/egs/speechio/ASR/zipformer/scaling.py @@ -0,0 +1 @@ +../../../librispeech/ASR/zipformer/scaling.py \ No newline at end of file diff --git a/egs/speechio/ASR/zipformer/scaling_converter.py b/egs/speechio/ASR/zipformer/scaling_converter.py new file mode 120000 index 000000000..b0ecee05e --- /dev/null +++ b/egs/speechio/ASR/zipformer/scaling_converter.py @@ -0,0 +1 @@ +../../../librispeech/ASR/zipformer/scaling_converter.py \ No newline at end of file diff --git a/egs/speechio/ASR/zipformer/subsampling.py b/egs/speechio/ASR/zipformer/subsampling.py new file mode 120000 index 000000000..01ae9002c --- /dev/null +++ b/egs/speechio/ASR/zipformer/subsampling.py @@ -0,0 +1 @@ +../../../librispeech/ASR/zipformer/subsampling.py \ No newline at end of file diff --git a/egs/speechio/ASR/zipformer/train.py b/egs/speechio/ASR/zipformer/train.py new file mode 120000 index 000000000..ad7216cf7 --- /dev/null +++ b/egs/speechio/ASR/zipformer/train.py @@ -0,0 +1 @@ +../../../multi_zh-hans/ASR/zipformer/train.py \ No newline at end of file diff --git a/egs/speechio/ASR/zipformer/zipformer.py b/egs/speechio/ASR/zipformer/zipformer.py new file mode 120000 index 000000000..23011dda7 --- /dev/null +++ b/egs/speechio/ASR/zipformer/zipformer.py @@ -0,0 +1 @@ +../../../librispeech/ASR/zipformer/zipformer.py \ No newline at end of file diff --git a/egs/spgispeech/ASR/zipformer/train.py b/egs/spgispeech/ASR/zipformer/train.py index 1709a2845..ed66ca29b 100755 --- a/egs/spgispeech/ASR/zipformer/train.py +++ b/egs/spgispeech/ASR/zipformer/train.py @@ -89,6 +89,7 @@ from icefall.checkpoint import ( ) from icefall.dist import cleanup_dist, setup_dist from icefall.env import get_env_info +from icefall.err import raise_grad_scale_is_too_small_error from icefall.hooks import register_inf_check_hooks from icefall.utils import ( AttributeDict, @@ -1020,9 +1021,7 @@ def train_one_epoch( logging.warning(f"Grad scale is small: {cur_grad_scale}") if cur_grad_scale < 1.0e-05: save_bad_model() - raise RuntimeError( - f"grad_scale is too small, exiting: {cur_grad_scale}" - ) + raise_grad_scale_is_too_small_error(cur_grad_scale) if batch_idx % params.log_interval == 0: cur_lr = max(scheduler.get_last_lr()) diff --git a/egs/swbd/ASR/conformer_ctc/decode.py b/egs/swbd/ASR/conformer_ctc/decode.py index 2bbade374..52e501ae1 100755 --- a/egs/swbd/ASR/conformer_ctc/decode.py +++ b/egs/swbd/ASR/conformer_ctc/decode.py @@ -29,7 +29,6 @@ import torch import torch.nn as nn from asr_datamodule import SwitchBoardAsrDataModule from conformer import Conformer - from sclite_scoring import asr_text_post_processing from icefall.bpe_graph_compiler import BpeCtcTrainingGraphCompiler diff --git a/egs/swbd/ASR/local/filter_empty_text.py b/egs/swbd/ASR/local/filter_empty_text.py index 6b3316800..13b35980b 100755 --- a/egs/swbd/ASR/local/filter_empty_text.py +++ b/egs/swbd/ASR/local/filter_empty_text.py @@ -16,8 +16,8 @@ # limitations under the License. import argparse -from pathlib import Path import logging +from pathlib import Path from typing import List diff --git a/egs/tal_csasr/ASR/pruned_transducer_stateless7_bbpe/jit_pretrained.py b/egs/tal_csasr/ASR/pruned_transducer_stateless7_bbpe/jit_pretrained.py index 8c966a2f6..503cdf4ed 100755 --- a/egs/tal_csasr/ASR/pruned_transducer_stateless7_bbpe/jit_pretrained.py +++ b/egs/tal_csasr/ASR/pruned_transducer_stateless7_bbpe/jit_pretrained.py @@ -45,6 +45,7 @@ import sentencepiece as spm import torch import torchaudio from torch.nn.utils.rnn import pad_sequence + from icefall import smart_byte_decode diff --git a/egs/tal_csasr/ASR/pruned_transducer_stateless7_bbpe/train.py b/egs/tal_csasr/ASR/pruned_transducer_stateless7_bbpe/train.py index aee3972cd..2108266ec 100755 --- a/egs/tal_csasr/ASR/pruned_transducer_stateless7_bbpe/train.py +++ b/egs/tal_csasr/ASR/pruned_transducer_stateless7_bbpe/train.py @@ -78,6 +78,7 @@ from icefall.checkpoint import ( ) from icefall.dist import cleanup_dist, setup_dist from icefall.env import get_env_info +from icefall.err import raise_grad_scale_is_too_small_error from icefall.hooks import register_inf_check_hooks from icefall.utils import ( AttributeDict, @@ -870,9 +871,7 @@ def train_one_epoch( if cur_grad_scale < 0.01: logging.warning(f"Grad scale is small: {cur_grad_scale}") if cur_grad_scale < 1.0e-05: - raise RuntimeError( - f"grad_scale is too small, exiting: {cur_grad_scale}" - ) + raise_grad_scale_is_too_small_error(cur_grad_scale) if batch_idx % params.log_interval == 0: cur_lr = scheduler.get_last_lr()[0] diff --git a/egs/tedlium3/ASR/zipformer/model.py b/egs/tedlium3/ASR/zipformer/model.py index 90ec7e7aa..65b052ab9 100644 --- a/egs/tedlium3/ASR/zipformer/model.py +++ b/egs/tedlium3/ASR/zipformer/model.py @@ -19,9 +19,9 @@ import k2 import torch import torch.nn as nn from encoder_interface import EncoderInterface +from scaling import ScaledLinear from icefall.utils import add_sos, make_pad_mask -from scaling import ScaledLinear class Transducer(nn.Module): diff --git a/egs/tedlium3/ASR/zipformer/train.py b/egs/tedlium3/ASR/zipformer/train.py index 5ad01df27..14a44efb3 100755 --- a/egs/tedlium3/ASR/zipformer/train.py +++ b/egs/tedlium3/ASR/zipformer/train.py @@ -87,6 +87,7 @@ from icefall.checkpoint import ( ) from icefall.dist import cleanup_dist, setup_dist from icefall.env import get_env_info +from icefall.err import raise_grad_scale_is_too_small_error from icefall.hooks import register_inf_check_hooks from icefall.utils import ( AttributeDict, @@ -985,9 +986,7 @@ def train_one_epoch( logging.warning(f"Grad scale is small: {cur_grad_scale}") if cur_grad_scale < 1.0e-05: save_bad_model() - raise RuntimeError( - f"grad_scale is too small, exiting: {cur_grad_scale}" - ) + raise_grad_scale_is_too_small_error(cur_grad_scale) if batch_idx % params.log_interval == 0: cur_lr = max(scheduler.get_last_lr()) diff --git a/egs/vctk/TTS/README.md b/egs/vctk/TTS/README.md new file mode 100644 index 000000000..c07516b77 --- /dev/null +++ b/egs/vctk/TTS/README.md @@ -0,0 +1,37 @@ +# Introduction + +This CSTR VCTK Corpus includes speech data uttered by 110 English speakers with various accents. Each speaker reads out about 400 sentences, which were selected from a newspaper, the rainbow passage and an elicitation paragraph used for the speech accent archive. +The newspaper texts were taken from Herald Glasgow, with permission from Herald & Times Group. Each speaker has a different set of the newspaper texts selected based a greedy algorithm that increases the contextual and phonetic coverage. +The details of the text selection algorithms are described in the following paper: [C. Veaux, J. Yamagishi and S. King, "The voice bank corpus: Design, collection and data analysis of a large regional accent speech database,"](https://doi.org/10.1109/ICSDA.2013.6709856). + +The above information is from the [CSTR VCTK website](https://datashare.ed.ac.uk/handle/10283/3443). + +# VITS + +This recipe provides a VITS model trained on the VCTK dataset. + +Pretrained model can be found [here](https://huggingface.co/zrjin/icefall-tts-vctk-vits-2023-12-05), note that this model was pretrained on the Edinburgh DataShare VCTK dataset. + +For tutorial and more details, please refer to the [VITS documentation](https://k2-fsa.github.io/icefall/recipes/TTS/vctk/vits.html). + +The training command is given below: +``` +export CUDA_VISIBLE_DEVICES="0,1,2,3" +./vits/train.py \ + --world-size 4 \ + --num-epochs 1000 \ + --start-epoch 1 \ + --use-fp16 1 \ + --exp-dir vits/exp \ + --tokens data/tokens.txt + --max-duration 350 +``` + +To inference, use: +``` +./vits/infer.py \ + --epoch 1000 \ + --exp-dir vits/exp \ + --tokens data/tokens.txt \ + --max-duration 500 +``` \ No newline at end of file diff --git a/egs/wenetspeech/ASR/local/compute_fbank_wenetspeech_dev_test.py b/egs/wenetspeech/ASR/local/compute_fbank_wenetspeech_dev_test.py index 1af08fee2..ac4e92ec5 100755 --- a/egs/wenetspeech/ASR/local/compute_fbank_wenetspeech_dev_test.py +++ b/egs/wenetspeech/ASR/local/compute_fbank_wenetspeech_dev_test.py @@ -16,11 +16,19 @@ # See the License for the specific language governing permissions and # limitations under the License. +import argparse import logging from pathlib import Path import torch -from lhotse import CutSet, KaldifeatFbank, KaldifeatFbankConfig, LilcomChunkyWriter +from lhotse import ( + CutSet, + KaldifeatFbank, + KaldifeatFbankConfig, + LilcomChunkyWriter, + WhisperFbank, + WhisperFbankConfig, +) # Torch's multithreaded behavior needs to be disabled or # it wastes a lot of CPU and slow things down. @@ -30,8 +38,31 @@ torch.set_num_threads(1) torch.set_num_interop_threads(1) torch.multiprocessing.set_sharing_strategy("file_system") +from icefall.utils import str2bool -def compute_fbank_wenetspeech_dev_test(): + +def get_parser(): + parser = argparse.ArgumentParser( + formatter_class=argparse.ArgumentDefaultsHelpFormatter + ) + + parser.add_argument( + "--num-mel-bins", + type=int, + default=80, + help="""The number of mel bins for Fbank""", + ) + + parser.add_argument( + "--whisper-fbank", + type=str2bool, + default=False, + help="Use WhisperFbank instead of Fbank. Default: False.", + ) + return parser + + +def compute_fbank_wenetspeech_dev_test(args): in_out_dir = Path("data/fbank") # number of workers in dataloader num_workers = 42 @@ -44,7 +75,12 @@ def compute_fbank_wenetspeech_dev_test(): device = torch.device("cpu") if torch.cuda.is_available(): device = torch.device("cuda", 0) - extractor = KaldifeatFbank(KaldifeatFbankConfig(device=device)) + if args.whisper_fbank: + extractor = WhisperFbank( + WhisperFbankConfig(num_filters=args.num_mel_bins, device="cuda") + ) + else: + extractor = KaldifeatFbank(KaldifeatFbankConfig(device=device)) logging.info(f"device: {device}") @@ -82,7 +118,11 @@ def main(): formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s" logging.basicConfig(format=formatter, level=logging.INFO) - compute_fbank_wenetspeech_dev_test() + parser = get_parser() + args = parser.parse_args() + logging.info(vars(args)) + + compute_fbank_wenetspeech_dev_test(args) if __name__ == "__main__": diff --git a/egs/wenetspeech/ASR/local/compute_fbank_wenetspeech_splits.py b/egs/wenetspeech/ASR/local/compute_fbank_wenetspeech_splits.py index a87801462..804a302bd 100755 --- a/egs/wenetspeech/ASR/local/compute_fbank_wenetspeech_splits.py +++ b/egs/wenetspeech/ASR/local/compute_fbank_wenetspeech_splits.py @@ -22,15 +22,19 @@ from datetime import datetime from pathlib import Path import torch -from lhotse import ( +from lhotse import ( # KaldifeatWhisperFbank,; KaldifeatWhisperFbankConfig, CutSet, KaldifeatFbank, KaldifeatFbankConfig, LilcomChunkyWriter, + WhisperFbank, + WhisperFbankConfig, set_audio_duration_mismatch_tolerance, set_caching_enabled, ) +from icefall.utils import get_executor, str2bool + # Torch's multithreaded behavior needs to be disabled or # it wastes a lot of CPU and slow things down. # Do this outside of main() in case it needs to take effect @@ -87,6 +91,27 @@ def get_parser(): default=-1, help="Stop processing pieces until this number (excluded).", ) + + parser.add_argument( + "--num-mel-bins", + type=int, + default=80, + help="""The number of mel bins for Fbank""", + ) + + parser.add_argument( + "--whisper-fbank", + type=str2bool, + default=False, + help="Use WhisperFbank instead of Fbank. Default: False.", + ) + + parser.add_argument( + "--output-dir-prefix", + type=str, + default="", + help="Prefix of the output directory.", + ) return parser @@ -96,6 +121,7 @@ def compute_fbank_wenetspeech_splits(args): num_splits = args.num_splits output_dir = f"data/fbank/{subset}_split_{num_splits}" output_dir = Path(output_dir) + output_dir = Path(args.output_dir_prefix) / output_dir assert output_dir.exists(), f"{output_dir} does not exist!" num_digits = len(str(num_splits)) @@ -110,14 +136,21 @@ def compute_fbank_wenetspeech_splits(args): device = torch.device("cpu") if torch.cuda.is_available(): device = torch.device("cuda", 0) - extractor = KaldifeatFbank(KaldifeatFbankConfig(device=device)) + if args.whisper_fbank: + extractor = WhisperFbank( + WhisperFbankConfig(num_filters=args.num_mel_bins, device=device) + ) + # extractor = KaldifeatWhisperFbank(KaldifeatWhisperFbankConfig(num_filters=args.num_mel_bins, device=device)) + else: + extractor = KaldifeatFbank(KaldifeatFbankConfig(device=device)) logging.info(f"device: {device}") set_audio_duration_mismatch_tolerance(0.01) # 10ms tolerance set_caching_enabled(False) + # with get_executor() as ex: # Initialize the executor only once. for i in range(start, stop): - idx = f"{i + 1}".zfill(num_digits) - logging.info(f"Processing {idx}/{num_splits}") + idx = f"{i}".zfill(num_digits) + logging.info(f"Processing {i+1}/{num_splits}") cuts_path = output_dir / f"cuts_{subset}.{idx}.jsonl.gz" if cuts_path.is_file(): @@ -143,7 +176,6 @@ def compute_fbank_wenetspeech_splits(args): storage_type=LilcomChunkyWriter, overwrite=True, ) - logging.info(f"Saving to {cuts_path}") cut_set.to_file(cuts_path) diff --git a/egs/wenetspeech/ASR/local/prepare_dataset_from_kaldi_dir.py b/egs/wenetspeech/ASR/local/prepare_dataset_from_kaldi_dir.py index 334a6d023..52da3d6dc 100644 --- a/egs/wenetspeech/ASR/local/prepare_dataset_from_kaldi_dir.py +++ b/egs/wenetspeech/ASR/local/prepare_dataset_from_kaldi_dir.py @@ -17,10 +17,10 @@ import argparse import logging - -import torch -import lhotse from pathlib import Path + +import lhotse +import torch from lhotse import ( CutSet, Fbank, @@ -29,6 +29,7 @@ from lhotse import ( fix_manifests, validate_recordings_and_supervisions, ) + from icefall.utils import get_executor, str2bool # Torch's multithreaded behavior needs to be disabled or diff --git a/egs/wenetspeech/ASR/local/prepare_pinyin.py b/egs/wenetspeech/ASR/local/prepare_pinyin.py index ae40f1cdd..112b50b79 100755 --- a/egs/wenetspeech/ASR/local/prepare_pinyin.py +++ b/egs/wenetspeech/ASR/local/prepare_pinyin.py @@ -41,6 +41,7 @@ from prepare_lang import ( write_lexicon, write_mapping, ) + from icefall.utils import text_to_pinyin diff --git a/egs/wenetspeech/ASR/prepare.sh b/egs/wenetspeech/ASR/prepare.sh index 543d19ce0..e3e28bd24 100755 --- a/egs/wenetspeech/ASR/prepare.sh +++ b/egs/wenetspeech/ASR/prepare.sh @@ -182,6 +182,43 @@ if [ $stage -le 13 ] && [ $stop_stage -ge 13 ]; then fi fi +whisper_mel_bins=80 +if [ $stage -le 129 ] && [ $stop_stage -ge 129 ]; then + log "Stage 129: compute whisper fbank for dev and test sets" + python3 ./local/compute_fbank_wenetspeech_dev_test.py --num-mel-bins ${whisper_mel_bins} --whisper-fbank true +fi +if [ $stage -le 130 ] && [ $stop_stage -ge 130 ]; then + log "Stage 130: Comute features for whisper training set" + + split_dir=data/fbank/L_split_${num_splits} + if [ ! -f $split_dir/.split_completed ]; then + lhotse split $num_splits ./data/fbank/cuts_L_raw.jsonl.gz $split_dir + touch $split_dir/.split_completed + fi + + python3 ./local/compute_fbank_wenetspeech_splits.py \ + --training-subset L \ + --num-workers 8 \ + --batch-duration 1600 \ + --start 0 \ + --num-mel-bins ${whisper_mel_bins} --whisper-fbank true \ + --num-splits $num_splits + + if [ ! -f data/fbank/cuts_L.jsonl.gz ]; then + pieces=$(find data/fbank/L_split_${num_splits} -name "cuts_L.*.jsonl.gz") + lhotse combine $pieces data/fbank/cuts_L.jsonl.gz + fi +fi + +if [ $stage -le 131 ] && [ $stop_stage -ge 131 ]; then + log "Stage 131: concat feats into train set" + if [ ! -f data/fbank/cuts_L.jsonl.gz ]; then + pieces=$(find data/fbank/L_split_${num_splits} -name "cuts_L.*.jsonl.gz") + lhotse combine $pieces data/fbank/cuts_L.jsonl.gz + fi +fi + + if [ $stage -le 14 ] && [ $stop_stage -ge 14 ]; then log "Stage 14: Compute fbank for musan" mkdir -p data/fbank @@ -272,7 +309,7 @@ if [ $stage -le 19 ] && [ $stop_stage -ge 19 ]; then mkdir -p $text_out_dir log "Genearating training text data" - + if [ ! -f $text_out_dir/lm_data.pt ]; then ./local/prepare_char_lm_training_data.py \ --lang-char data/lang_char \ @@ -281,14 +318,14 @@ if [ $stage -le 19 ] && [ $stop_stage -ge 19 ]; then fi log "Generating DEV text data" - # prepare validation text data + # prepare validation text data if [ ! -f $text_out_dir/valid_text_words_segmentation ]; then valid_text=${text_out_dir}/ gunzip -c data/manifests/wenetspeech_supervisions_DEV.jsonl.gz \ | jq '.text' | sed 's/"//g' \ | ./local/text2token.py -t "char" > $text_out_dir/valid_text - + python3 ./local/text2segments.py \ --num-process $nj \ --input-file $text_out_dir/valid_text \ @@ -300,7 +337,7 @@ if [ $stage -le 19 ] && [ $stop_stage -ge 19 ]; then --lm-data $text_out_dir/valid_text_words_segmentation \ --lm-archive $text_out_dir/lm_data_valid.pt - # prepare TEST text data + # prepare TEST text data if [ ! -f $text_out_dir/TEST_text_words_segmentation ]; then log "Prepare text for test set." for test_set in TEST_MEETING TEST_NET; do @@ -313,7 +350,7 @@ if [ $stage -le 19 ] && [ $stop_stage -ge 19 ]; then --input-file $text_out_dir/${test_set}_text \ --output-file $text_out_dir/${test_set}_text_words_segmentation done - + cat $text_out_dir/TEST_*_text_words_segmentation > $text_out_dir/test_text_words_segmentation fi diff --git a/egs/wenetspeech/ASR/pruned_transducer_stateless5/onnx_check.py b/egs/wenetspeech/ASR/pruned_transducer_stateless5/onnx_check.py index ee8252a90..8c192913e 100755 --- a/egs/wenetspeech/ASR/pruned_transducer_stateless5/onnx_check.py +++ b/egs/wenetspeech/ASR/pruned_transducer_stateless5/onnx_check.py @@ -74,10 +74,10 @@ It will generate the following 3 files inside $repo/exp: import argparse import logging -from icefall import is_module_available +import torch from onnx_pretrained import OnnxModel -import torch +from icefall import is_module_available def get_parser(): diff --git a/egs/wenetspeech/ASR/whisper/asr_datamodule.py b/egs/wenetspeech/ASR/whisper/asr_datamodule.py new file mode 120000 index 000000000..a074d6085 --- /dev/null +++ b/egs/wenetspeech/ASR/whisper/asr_datamodule.py @@ -0,0 +1 @@ +../pruned_transducer_stateless2/asr_datamodule.py \ No newline at end of file diff --git a/egs/wenetspeech/ASR/whisper/decode.py b/egs/wenetspeech/ASR/whisper/decode.py new file mode 100755 index 000000000..103f8d725 --- /dev/null +++ b/egs/wenetspeech/ASR/whisper/decode.py @@ -0,0 +1,526 @@ +#!/usr/bin/env python3 +# Copyright 2021 Xiaomi Corporation (Author: Liyong Guo, +# Fangjun Kuang, +# Wei Kang) +# 2024 Yuekai Zhang +# +# 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: +# Command for decoding using fine-tuned models: +git lfs install +git clone https://huggingface.co/yuekai/icefall_asr_aishell_whisper +ln -s icefall_asr_aishell_whisper/exp_large_v2/epoch-10-avg6.pt whisper/exp_large_v2/epoch-999.pt + +python3 ./whisper/decode.py \ + --exp-dir whisper/exp_large_v2 \ + --model-name large-v2 \ + --epoch 999 --avg 1 \ + --beam-size 10 --max-duration 50 + +# Command for decoding using pretrained models (before fine-tuning): + +python3 ./whisper/decode.py \ + --exp-dir whisper/exp_large_v2 \ + --model-name large-v2 \ + --epoch -1 --avg 1 \ + --remove-whisper-encoder-input-length-restriction False \ + --beam-size 10 --max-duration 50 + +""" + +import argparse +import logging +import re +from collections import defaultdict +from pathlib import Path +from typing import Dict, List, Optional, Tuple + +import k2 +import torch +import torch.nn as nn +import whisper +from asr_datamodule import WenetSpeechAsrDataModule +from lhotse.cut import Cut +from tn.chinese.normalizer import Normalizer +from whisper.normalizers import BasicTextNormalizer +from whisper_encoder_forward_monkey_patch import replace_whisper_encoder_forward +from zhconv import convert + +from icefall.checkpoint import average_checkpoints_with_averaged_model, load_checkpoint +from icefall.env import get_env_info +from icefall.utils import ( + AttributeDict, + setup_logger, + store_transcripts, + str2bool, + write_error_stats, +) + + +def average_checkpoints( + filenames: List[Path], device: torch.device = torch.device("cpu") +) -> dict: + """Average a list of checkpoints. + The function is mainly used for deepspeed converted checkpoint averaging, which only include model state_dict. + + Args: + filenames: + Filenames of the checkpoints to be averaged. We assume all + checkpoints are saved by :func:`save_checkpoint`. + device: + Move checkpoints to this device before averaging. + Returns: + Return a dict (i.e., state_dict) which is the average of all + model state dicts contained in the checkpoints. + """ + n = len(filenames) + + if "model" in torch.load(filenames[0], map_location=device): + avg = torch.load(filenames[0], map_location=device)["model"] + else: + avg = torch.load(filenames[0], map_location=device) + + # 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 avg.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 i in range(1, n): + if "model" in torch.load(filenames[i], map_location=device): + state_dict = torch.load(filenames[i], map_location=device)["model"] + else: + state_dict = torch.load(filenames[i], map_location=device) + for k in uniqued_names: + avg[k] += state_dict[k] + + for k in uniqued_names: + if avg[k].is_floating_point(): + avg[k] /= n + else: + avg[k] //= n + + return avg + + +def remove_punctuation(text: str or List[str]): + """Modified from https://github.com/yeyupiaoling/Whisper-Finetune/blob/master/utils/data_utils.py + + Args: + text: It can be a string or a list of strings. + Returns: + Return a string or a list of strings without any punctuation. + """ + punctuation = "!,.;:?、！，。；：？《》 " + if isinstance(text, str): + text = re.sub(r"[{}]+".format(punctuation), "", text).strip() + return text + elif isinstance(text, list): + result_text = [] + for t in text: + t = re.sub(r"[{}]+".format(punctuation), "", t).strip() + result_text.append(t) + return result_text + else: + raise Exception(f"Not support type {type(text)}") + + +def to_simple(text: str or List[str]): + """Convert traditional Chinese to simplified Chinese. + Args: + text: It can be a string or a list of strings. + Returns: + Return a string or a list of strings converted to simplified Chinese. + """ + if isinstance(text, str): + text = convert(text, "zh-cn") + return text + elif isinstance(text, list): + result_text = [] + for t in text: + t = convert(t, "zh-cn") + result_text.append(t) + return result_text + else: + raise Exception(f"Not support type{type(text)}") + + +def get_parser(): + parser = argparse.ArgumentParser( + formatter_class=argparse.ArgumentDefaultsHelpFormatter + ) + + parser.add_argument( + "--epoch", + type=int, + default=-1, + help="It specifies the checkpoint to use for decoding." + "Note: Epoch counts from 0.", + ) + parser.add_argument( + "--avg", + type=int, + default=1, + help="Number of checkpoints to average. Automatically select " + "consecutive checkpoints before the checkpoint specified by " + "'--epoch'. ", + ) + + parser.add_argument( + "--method", + type=str, + default="beam-search", + help="""Decoding method. + Supported values are: + - beam-search + """, + ) + + parser.add_argument( + "--beam-size", + type=int, + default=1, + help="beam size for beam search decoding", + ) + + parser.add_argument( + "--exp-dir", + type=str, + default="whisper/exp", + help="The experiment dir", + ) + + parser.add_argument( + "--model-name", + type=str, + default="large-v2", + choices=["large-v2", "large-v3", "medium", "small", "tiny"], + help="""The model name to use. + """, + ) + + parser.add_argument( + "--remove-whisper-encoder-input-length-restriction", + type=str2bool, + default=True, + help="replace whisper encoder forward method to remove input length restriction", + ) + + return parser + + +def get_params() -> AttributeDict: + params = AttributeDict( + { + "env_info": get_env_info(), + } + ) + return params + + +def decode_one_batch( + params: AttributeDict, + model: nn.Module, + batch: dict, +) -> Dict[str, List[List[int]]]: + """Decode one batch and return the result in a dict. The dict has the + following format: + + - key: "beam-search" + - value: A list of lists. Each sublist is a list of token IDs. + Args: + params: + It is returned by :func:`get_params`. + model: + The neural model. + batch: + It is returned by :meth:`torch.utils.data.DataLoader.__iter__`. + Returns: + Return a dict, whose key may be "beam-search". + """ + dtype = torch.float16 + device = torch.device("cuda") + + feature = batch["inputs"] + assert feature.ndim == 3 + feature = feature.to(device, dtype=dtype).transpose(1, 2) + if not params.remove_whisper_encoder_input_length_restriction: + T = 3000 + if feature.shape[2] < T: + feature = torch.cat( + [ + feature, + torch.zeros( + feature.shape[0], feature.shape[1], T - feature.shape[2] + ).to(device, dtype=dtype), + ], + 2, + ) + + supervisions = batch["supervisions"] + feature_len = supervisions["num_frames"] + feature_len = feature_len.to(device, dtype=dtype) + results = model.decode(feature, params.decoding_options) + hyps = [result.text for result in results] + + hyps = remove_punctuation(hyps) + hyps = to_simple(hyps) + hyps = [params.normalizer.normalize(hyp) for hyp in hyps] + print(hyps) + return {"beam-search": hyps} + + +def decode_dataset( + dl: torch.utils.data.DataLoader, + params: AttributeDict, + model: nn.Module, +) -> Dict[str, List[Tuple[str, List[str], List[str]]]]: + """Decode dataset. + + Args: + dl: + The dataloader. + params: + It is returned by :func:`get_params`. + model: + The neural model. + Returns: + Return a dict, whose key may be "beam-search". + """ + results = [] + + 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, + batch=batch, + ) + + for lm_scale, 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[lm_scale].extend(this_batch) + + num_cuts += len(batch["supervisions"]["text"]) + + 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]]]], +): + + enable_log = True + test_set_wers = dict() + for key, results in results_dict.items(): + recog_path = ( + params.exp_dir / f"recogs-{test_set_name}-{key}-{params.suffix}.txt" + ) + results = sorted(results) + store_transcripts(filename=recog_path, texts=results) + if enable_log: + 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.exp_dir / f"errs-{test_set_name}-{key}-{params.suffix}.txt" + ) + # we compute CER for aishell dataset. + results_char = [] + for res in results: + results_char.append((res[0], list("".join(res[1])), list("".join(res[2])))) + with open(errs_filename, "w") as f: + wer = write_error_stats( + f, f"{test_set_name}-{key}", results_char, enable_log=enable_log + ) + test_set_wers[key] = wer + + if enable_log: + 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.exp_dir / f"cer-summary-{test_set_name}-{params.suffix}.txt" + with open(errs_info, "w") as f: + print("settings\tCER", file=f) + for key, val in test_set_wers: + print("{}\t{}".format(key, val), file=f) + + s = "\nFor {}, CER 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() + WenetSpeechAsrDataModule.add_arguments(parser) + args = parser.parse_args() + args.exp_dir = Path(args.exp_dir) + + params = get_params() + params.update(vars(args)) + params.suffix = f"epoch-{params.epoch}-avg-{params.avg}" + setup_logger( + f"{params.exp_dir}/log-{params.method}-beam{params.beam_size}/log-decode-{params.suffix}" + ) + + options = whisper.DecodingOptions( + task="transcribe", + language="zh", + without_timestamps=True, + beam_size=params.beam_size, + ) + params.decoding_options = options + params.cleaner = BasicTextNormalizer() + params.normalizer = Normalizer() + + logging.info("Decoding started") + logging.info(params) + + device = torch.device("cpu") + if torch.cuda.is_available(): + device = torch.device("cuda") + + logging.info(f"device: {device}") + + if params.remove_whisper_encoder_input_length_restriction: + replace_whisper_encoder_forward() + model = whisper.load_model(params.model_name, "cpu") + if params.epoch > 0: + if params.avg > 1: + start = params.epoch - params.avg + assert start >= 1, start + checkpoint = torch.load( + f"{params.exp_dir}/epoch-{params.epoch}.pt", map_location="cpu" + ) + if "model" not in checkpoint: + # deepspeed converted checkpoint only contains model state_dict + filenames = [ + f"{params.exp_dir}/epoch-{epoch}.pt" + for epoch in range(start, params.epoch + 1) + ] + model.load_state_dict(average_checkpoints(filenames)) + else: + 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, + ) + ) + # save checkpoints + filename = f"{params.exp_dir}/epoch-{params.epoch}-avg-{params.avg}.pt" + torch.save(model.state_dict(), filename) + else: + checkpoint = torch.load( + f"{params.exp_dir}/epoch-{params.epoch}.pt", map_location="cpu" + ) + if "model" not in checkpoint: + model.load_state_dict(checkpoint, strict=True) + else: + load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model) + 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 + wenetspeech = WenetSpeechAsrDataModule(args) + dev_cuts = wenetspeech.valid_cuts() + dev_dl = wenetspeech.valid_dataloaders(dev_cuts) + + def remove_long_utt(c: Cut): + # Keep only utterances with duration in 30 seconds + # + if c.duration > 30.0: + # logging.warning( + # f"Exclude cut with ID {c.id} from training. Duration: {c.duration}" + # ) + return False + return True + + test_net_cuts = wenetspeech.test_net_cuts() + test_net_cuts = test_net_cuts.filter(remove_long_utt) + test_net_dl = wenetspeech.test_dataloaders(test_net_cuts) + + test_meeting_cuts = wenetspeech.test_meeting_cuts() + test_meeting_dl = wenetspeech.test_dataloaders(test_meeting_cuts) + + # test_sets = ["DEV", "TEST_NET", "TEST_MEETING"] + # test_dls = [dev_dl, test_net_dl, test_meeting_dl] + + test_sets = ["TEST_NET"] + test_dls = [test_net_dl] + + # test_sets = ["TEST_MEETING"] + # test_dls = [test_meeting_dl] + + for test_set, test_dl in zip(test_sets, test_dls): + results_dict = decode_dataset( + dl=test_dl, + params=params, + model=model, + ) + + save_results(params=params, test_set_name=test_set, results_dict=results_dict) + + logging.info("Done!") + + +torch.set_num_threads(1) +torch.set_num_interop_threads(1) + +if __name__ == "__main__": + main() diff --git a/egs/wenetspeech/ASR/whisper/ds_config_zero1.json b/egs/wenetspeech/ASR/whisper/ds_config_zero1.json new file mode 120000 index 000000000..af7162d6c --- /dev/null +++ b/egs/wenetspeech/ASR/whisper/ds_config_zero1.json @@ -0,0 +1 @@ +../../../aishell/ASR/whisper/ds_config_zero1.json \ No newline at end of file diff --git a/egs/wenetspeech/ASR/whisper/label_smoothing.py b/egs/wenetspeech/ASR/whisper/label_smoothing.py new file mode 120000 index 000000000..e9d239fff --- /dev/null +++ b/egs/wenetspeech/ASR/whisper/label_smoothing.py @@ -0,0 +1 @@ +../../../librispeech/ASR/conformer_ctc/label_smoothing.py \ No newline at end of file diff --git a/egs/wenetspeech/ASR/whisper/optim.py b/egs/wenetspeech/ASR/whisper/optim.py new file mode 120000 index 000000000..5eaa3cffd --- /dev/null +++ b/egs/wenetspeech/ASR/whisper/optim.py @@ -0,0 +1 @@ +../../../librispeech/ASR/zipformer/optim.py \ No newline at end of file diff --git a/egs/wenetspeech/ASR/whisper/requirements.txt b/egs/wenetspeech/ASR/whisper/requirements.txt new file mode 120000 index 000000000..744bf8bb6 --- /dev/null +++ b/egs/wenetspeech/ASR/whisper/requirements.txt @@ -0,0 +1 @@ +../../../aishell/ASR/whisper/requirements.txt \ No newline at end of file diff --git a/egs/wenetspeech/ASR/whisper/train.py b/egs/wenetspeech/ASR/whisper/train.py new file mode 100644 index 000000000..4b7c1ca42 --- /dev/null +++ b/egs/wenetspeech/ASR/whisper/train.py @@ -0,0 +1,955 @@ +#!/usr/bin/env python3 +# Copyright 2023 Xiaomi Corp. (authors: Xiaoyu Yang) +# 2024 Yuekai Zhang +# +# 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: + +#fine-tuning with deepspeed zero stage 1 +torchrun --nproc-per-node 8 ./whisper/train.py \ + --max-duration 200 \ + --exp-dir whisper/exp_large_v2 \ + --model-name large-v2 \ + --deepspeed \ + --deepspeed_config ./whisper/ds_config_zero1.json + +# fine-tuning with ddp +torchrun --nproc_per_node 8 ./whisper/train.py \ + --max-duration 200 \ + --exp-dir whisper/exp_medium \ + --base-lr 1e-5 \ + --model-name medium +""" + + +import argparse +import copy +import logging +import random +import warnings +from pathlib import Path +from shutil import copyfile +from typing import Any, Dict, List, Optional, Tuple, Union + +import deepspeed +import k2 +import optim +import torch +import torch.multiprocessing as mp +import torch.nn as nn +import whisper +from asr_datamodule import WenetSpeechAsrDataModule +from deepspeed.utils.zero_to_fp32 import convert_zero_checkpoint_to_fp32_state_dict +from label_smoothing import LabelSmoothingLoss +from lhotse import CutSet, load_manifest +from lhotse.cut import Cut +from lhotse.dataset.sampling.base import CutSampler +from lhotse.utils import fix_random_seed +from optim import Eden, ScaledAdam +from torch import Tensor +from torch.cuda.amp import GradScaler +from torch.nn.functional import pad as pad_tensor +from torch.nn.parallel import DistributedDataParallel as DDP +from torch.utils.tensorboard import SummaryWriter +from whisper_encoder_forward_monkey_patch import replace_whisper_encoder_forward + +from icefall import diagnostics +from icefall.checkpoint import load_checkpoint, remove_checkpoints +from icefall.checkpoint import save_checkpoint as save_checkpoint_impl +from icefall.checkpoint import update_averaged_model +from icefall.dist import cleanup_dist, get_rank, get_world_size, setup_dist +from icefall.env import get_env_info +from icefall.hooks import register_inf_check_hooks +from icefall.utils import ( + AttributeDict, + MetricsTracker, + filter_uneven_sized_batch, + setup_logger, + str2bool, +) + +LRSchedulerType = Union[torch.optim.lr_scheduler._LRScheduler, optim.LRScheduler] + + +def set_batch_count(model: Union[nn.Module, DDP], batch_count: float) -> None: + if isinstance(model, DDP): + # get underlying nn.Module + model = model.module + for module in model.modules(): + if hasattr(module, "batch_count"): + module.batch_count = batch_count + + +def get_parser(): + parser = argparse.ArgumentParser( + formatter_class=argparse.ArgumentDefaultsHelpFormatter + ) + + parser.add_argument( + "--tensorboard", + type=str2bool, + default=True, + help="Should various information be logged in tensorboard.", + ) + + parser.add_argument( + "--num-epochs", + type=int, + default=10, + help="Number of epochs to train.", + ) + + parser.add_argument( + "--start-epoch", + type=int, + default=1, + help="""Resume training from this epoch. It should be positive. + If larger than 1, it will load checkpoint from + exp-dir/epoch-{start_epoch-1}.pt + """, + ) + + parser.add_argument( + "--start-batch", + type=int, + default=0, + help="""If positive, --start-epoch is ignored and + it loads the checkpoint from exp-dir/checkpoint-{start_batch}.pt + """, + ) + + parser.add_argument( + "--exp-dir", + type=str, + default="whisper/exp", + help="""The experiment dir. + It specifies the directory where all training related + files, e.g., checkpoints, log, etc, are saved + """, + ) + + parser.add_argument( + "--model-name", + type=str, + default="large-v2", + choices=["large-v2", "large-v3", "medium", "small", "tiny"], + help="""The model name to use. + """, + ) + + parser.add_argument( + "--base-lr", type=float, default=1e-5, help="The base learning rate." + ) + + 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.""", + ) + + parser.add_argument( + "--lr-epochs", + type=float, + default=6, + help="""Number of epochs that affects how rapidly the learning rate decreases. + """, + ) + + parser.add_argument( + "--seed", + type=int, + default=42, + help="The seed for random generators intended for reproducibility", + ) + + parser.add_argument( + "--print-diagnostics", + type=str2bool, + default=False, + help="Accumulate stats on activations, print them and exit.", + ) + + parser.add_argument( + "--inf-check", + type=str2bool, + default=False, + help="Add hooks to check for infinite module outputs and gradients.", + ) + + parser.add_argument( + "--keep-last-k", + type=int, + default=30, + help="""Only keep this number of checkpoints on disk. + For instance, if it is 3, there are only 3 checkpoints + in the exp-dir with filenames `checkpoint-xxx.pt`. + It does not affect checkpoints with name `epoch-xxx.pt`. + """, + ) + + parser.add_argument( + "--average-period", + type=int, + default=200, + help="""Update the averaged model, namely `model_avg`, after processing + this number of batches. `model_avg` is a separate version of model, + in which each floating-point parameter is the average of all the + parameters from the start of training. Each time we take the average, + we do: `model_avg = model * (average_period / batch_idx_train) + + model_avg * ((batch_idx_train - average_period) / batch_idx_train)`. + """, + ) + + parser.add_argument( + "--use-fp16", + type=str2bool, + default=True, + help="Whether to use half precision training.", + ) + + parser = deepspeed.add_config_arguments(parser) + + return parser + + +def get_params() -> AttributeDict: + """Return a dict containing training parameters. + + All training related parameters that are not passed from the commandline + are saved in the variable `params`. + + Commandline options are merged into `params` after they are parsed, so + you can also access them via `params`. + + Explanation of options saved in `params`: + + - frame_shift_ms: The frame shift in milliseconds. + - allowed_excess_duration_ratio: The allowed excess duration ratio. + - best_train_loss: The best training loss so far. + - best_valid_loss: The best validation loss so far. + - best_train_epoch: The epoch where the best training loss is achieved. + - best_valid_epoch: The epoch where the best validation loss is achieved. + - batch_idx_train: The batch index of the current batch. + - log_interval: Log training stats every `log_interval` batches. + - reset_interval: Reset the stats every `reset_interval` batches. + - valid_interval: Run validation every `valid_interval` batches. + - env_info: The environment information. + """ + params = AttributeDict( + { + "frame_shift_ms": 10.0, + "subsampling_factor": 2, + "allowed_excess_duration_ratio": 0.1, + "best_train_loss": float("inf"), + "best_valid_loss": float("inf"), + "best_train_epoch": -1, + "best_valid_epoch": -1, + "batch_idx_train": 0, + "log_interval": 50, + "reset_interval": 200, + "valid_interval": 10000, + "env_info": get_env_info(), + } + ) + + return params + + +def load_checkpoint_if_available( + params: AttributeDict, + model: nn.Module, + model_avg: nn.Module = None, + optimizer: Optional[torch.optim.Optimizer] = None, + scheduler: Optional[LRSchedulerType] = None, +) -> Optional[Dict[str, Any]]: + """Load checkpoint from file. + + If params.start_batch is positive, it will load the checkpoint from + `params.exp_dir/checkpoint-{params.start_batch}.pt`. Otherwise, if + params.start_epoch is larger than 1, it will load the checkpoint from + `params.start_epoch - 1`. + + Apart from loading state dict for `model` and `optimizer` it also updates + `best_train_epoch`, `best_train_loss`, `best_valid_epoch`, + and `best_valid_loss` in `params`. + + Args: + params: + The return value of :func:`get_params`. + model: + The training model. + model_avg: + The stored model averaged from the start of training. + optimizer: + The optimizer that we are using. + scheduler: + The scheduler that we are using. + Returns: + Return a dict containing previously saved training info. + """ + if params.start_batch > 0: + filename = params.exp_dir / f"checkpoint-{params.start_batch}.pt" + elif params.start_epoch > 1: + filename = params.exp_dir / f"epoch-{params.start_epoch-1}.pt" + else: + return None + + assert filename.is_file(), f"{filename} does not exist!" + + saved_params = load_checkpoint( + filename, + model=model, + model_avg=model_avg, + optimizer=optimizer, + scheduler=scheduler, + ) + + keys = [ + "best_train_epoch", + "best_valid_epoch", + "batch_idx_train", + "best_train_loss", + "best_valid_loss", + ] + for k in keys: + params[k] = saved_params[k] + + if params.start_batch > 0: + if "cur_epoch" in saved_params: + params["start_epoch"] = saved_params["cur_epoch"] + + return saved_params + + +def save_checkpoint( + params: AttributeDict, + model: Union[nn.Module, DDP], + model_avg: Optional[nn.Module] = None, + optimizer: Optional[torch.optim.Optimizer] = None, + scheduler: Optional[LRSchedulerType] = None, + sampler: Optional[CutSampler] = None, + scaler: Optional[GradScaler] = None, + rank: int = 0, +) -> None: + """Save model, optimizer, scheduler and training stats to file. + + Args: + params: + It is returned by :func:`get_params`. + model: + The training model. + model_avg: + The stored model averaged from the start of training. + optimizer: + The optimizer used in the training. + sampler: + The sampler for the training dataset. + scaler: + The scaler used for mix precision training. + """ + if rank != 0: + return + filename = params.exp_dir / f"epoch-{params.cur_epoch}.pt" + save_checkpoint_impl( + filename=filename, + model=model, + model_avg=model_avg, + params=params, + optimizer=optimizer, + scheduler=scheduler, + sampler=sampler, + scaler=scaler, + rank=rank, + ) + + if params.best_train_epoch == params.cur_epoch: + best_train_filename = params.exp_dir / "best-train-loss.pt" + copyfile(src=filename, dst=best_train_filename) + + if params.best_valid_epoch == params.cur_epoch: + best_valid_filename = params.exp_dir / "best-valid-loss.pt" + copyfile(src=filename, dst=best_valid_filename) + + +def compute_loss( + params: AttributeDict, + tokenizer: whisper.tokenizer.Tokenizer, + model: Union[nn.Module, DDP], + batch: dict, + is_training: bool, +) -> Tuple[Tensor, MetricsTracker]: + """ + Compute the loss for the given batch. + Args: + params: + It is returned by :func:`get_params`. + tokenizer: + The tokenizer used to encode the text. + model: + The model for training. + batch: + A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()` + for the content in it. + is_training: + Whether it is training. + Returns: + Return a tuple of two elements. The first element is the loss tensor. + """ + # For the uneven-sized batch, the total duration after padding would possibly + # cause OOM. Hence, for each batch, which is sorted descendingly by length, + # we simply drop the last few shortest samples, so that the retained total frames + # (after padding) would not exceed `allowed_max_frames`: + # `allowed_max_frames = int(max_frames * (1.0 + allowed_excess_duration_ratio))`, + # where `max_frames = max_duration * 1000 // frame_shift_ms`. + # We set allowed_excess_duration_ratio=0.1. + if isinstance(model, DDP): + # get underlying nn.Module + model = model.module + + def _batch_tensors(tensors: List[Tensor], pad_value: Any) -> Tensor: + padding_size = max(tensor.shape[0] for tensor in tensors) + dims = len(tensors[0].shape) + padded_tensors = [] + for tensor in tensors: + padding = [0] * 2 * dims + padding[-1] = padding_size - tensor.shape[0] + padded_tensors.append(pad_tensor(tensor, padding, "constant", pad_value)) + return torch.stack([tensor for tensor in padded_tensors], dim=0) + + max_frames = params.max_duration * 1000 // params.frame_shift_ms + allowed_max_frames = int(max_frames * (1.0 + params.allowed_excess_duration_ratio)) + batch = filter_uneven_sized_batch(batch, allowed_max_frames) + + device = model.device if isinstance(model, DDP) else next(model.parameters()).device + feature = batch["inputs"] + + assert feature.ndim == 3 + feature = feature.to(device) + feature = feature.transpose(1, 2) # (N, C, T) + + supervisions = batch["supervisions"] + feature_lens = supervisions["num_frames"].to(device) + + batch_idx_train = params.batch_idx_train + + texts = batch["supervisions"]["text"] + # remove spaces in texts + texts = [text.replace(" ", "") for text in texts] + + text_tokens_list = [ + list(tokenizer.sot_sequence_including_notimestamps) + + tokenizer.encode(text) + + [tokenizer.eot] + for text in texts + ] + # convert it to torch tensor + text_tokens_list = [ + torch.LongTensor(text_tokens) for text_tokens in text_tokens_list + ] + + # 50256 is the index of for all whisper models + prev_outputs_tokens = _batch_tensors( + [tokens[:-1] for tokens in text_tokens_list], pad_value=50256 + ) + target_tokens = _batch_tensors( + [tokens[1:] for tokens in text_tokens_list], pad_value=50256 + ) + target_lengths = torch.LongTensor( + [tokens.shape[0] - 1 for tokens in text_tokens_list] + ) + + decoder_criterion = LabelSmoothingLoss( + ignore_index=50256, label_smoothing=0.1, reduction="sum" + ) + + # ignore the first 3 tokens, which are always <|lang_id|>, <|transcibe|>, <|notimestampes|> + ignore_prefix_size = 3 + with torch.set_grad_enabled(is_training): + encoder_out = model.encoder(feature) + text_logits = model.decoder(prev_outputs_tokens.to(device), encoder_out) + text_logits = text_logits[:, ignore_prefix_size:, :] + target_tokens = target_tokens[:, ignore_prefix_size:] + loss = decoder_criterion(text_logits, target_tokens.to(device)) + + assert loss.requires_grad == is_training + + info = MetricsTracker() + with warnings.catch_warnings(): + warnings.simplefilter("ignore") + info["frames"] = (feature_lens // params.subsampling_factor).sum().item() + + # Note: We use reduction=sum while computing the loss. + info["loss"] = loss.detach().cpu().item() + + return loss, info + + +def compute_validation_loss( + params: AttributeDict, + tokenizer: whisper.tokenizer.Tokenizer, + model: Union[nn.Module, DDP], + valid_dl: torch.utils.data.DataLoader, + world_size: int = 1, +) -> MetricsTracker: + """Run the validation process.""" + model.eval() + + tot_loss = MetricsTracker() + + for batch_idx, batch in enumerate(valid_dl): + with torch.cuda.amp.autocast(enabled=params.use_fp16): + loss, loss_info = compute_loss( + params=params, + tokenizer=tokenizer, + model=model, + batch=batch, + is_training=False, + ) + assert loss.requires_grad is False + tot_loss = tot_loss + loss_info + + if world_size > 1: + tot_loss.reduce(loss.device) + + loss_value = tot_loss["loss"] / tot_loss["frames"] + if loss_value < params.best_valid_loss: + params.best_valid_epoch = params.cur_epoch + params.best_valid_loss = loss_value + + return tot_loss + + +def train_one_epoch( + params: AttributeDict, + tokenizer: whisper.tokenizer.Tokenizer, + model: Union[nn.Module, DDP], + optimizer: torch.optim.Optimizer, + scheduler: LRSchedulerType, + train_dl: torch.utils.data.DataLoader, + valid_dl: torch.utils.data.DataLoader, + scaler: GradScaler, + model_avg: Optional[nn.Module] = None, + tb_writer: Optional[SummaryWriter] = None, + world_size: int = 1, + rank: int = 0, +) -> None: + """Train the model for one epoch. + + The training loss from the mean of all frames is saved in + `params.train_loss`. It runs the validation process every + `params.valid_interval` batches. + + Args: + params: + It is returned by :func:`get_params`. + model: + The model for training. + optimizer: + The optimizer we are using. + scheduler: + The learning rate scheduler, we call step() every step. + train_dl: + Dataloader for the training dataset. + valid_dl: + Dataloader for the validation dataset. + scaler: + The scaler used for mix precision training. + model_avg: + The stored model averaged from the start of training. + tb_writer: + Writer to write log messages to tensorboard. + world_size: + Number of nodes in DDP training. If it is 1, DDP is disabled. + rank: + The rank of the node in DDP training. If no DDP is used, it should + be set to 0. + """ + model.train() + + tot_loss = MetricsTracker() + + for batch_idx, batch in enumerate(train_dl): + params.batch_idx_train += 1 + batch_size = len(batch["supervisions"]["text"]) + if batch_idx % params.valid_interval == 0 and not params.print_diagnostics: + logging.info("Computing validation loss") + valid_info = compute_validation_loss( + params=params, + tokenizer=tokenizer, + model=model, + valid_dl=valid_dl, + world_size=world_size, + ) + model.train() + logging.info(f"Epoch {params.cur_epoch}, validation: {valid_info}") + logging.info( + f"Maximum memory allocated so far is {torch.cuda.max_memory_allocated()//1000000}MB" + ) + if tb_writer is not None: + valid_info.write_summary( + tb_writer, "train/valid_", params.batch_idx_train + ) + if params.deepspeed: + model.save_checkpoint( + save_dir=params.exp_dir, + tag=f"epoch-{params.cur_epoch}-checkpoint-{batch_idx}", + client_state={}, + ) + if rank == 0: + convert_zero_checkpoint_to_fp32_state_dict( + params.exp_dir, + f"{params.exp_dir}/epoch-{params.cur_epoch}-checkpoint-{batch_idx}.pt", + tag=f"epoch-{params.cur_epoch}-checkpoint-{batch_idx}", + ) + + try: + with torch.cuda.amp.autocast(enabled=params.use_fp16): + loss, loss_info = compute_loss( + params=params, + tokenizer=tokenizer, + model=model, + batch=batch, + is_training=True, + ) + # 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. + if params.deepspeed: + # deepspeed's backward() is different from torch's backward() + # in that it does not accept a loss tensor as input. + # It computes the loss internally. + model.backward(loss) + model.step() + else: + scaler.scale(loss).backward() + set_batch_count(model, params.batch_idx_train) + scheduler.step_batch(params.batch_idx_train) + + scaler.step(optimizer) + scaler.update() + optimizer.zero_grad() + except: # noqa + display_and_save_batch(batch, params=params) + raise + + if params.print_diagnostics and batch_idx == 5: + return + + if ( + rank == 0 + and params.batch_idx_train > 0 + and params.batch_idx_train % params.average_period == 0 + and not params.deepspeed + ): + update_averaged_model( + params=params, + model_cur=model, + model_avg=model_avg, + ) + + if batch_idx % 100 == 0 and params.use_fp16 and not params.deepspeed: + # If the grad scale was less than 1, try increasing it. The _growth_interval + # of the grad scaler is configurable, but we can't configure it to have different + # behavior depending on the current grad scale. + cur_grad_scale = scaler._scale.item() + if cur_grad_scale < 1.0 or (cur_grad_scale < 8.0 and batch_idx % 400 == 0): + scaler.update(cur_grad_scale * 2.0) + if cur_grad_scale < 0.01: + logging.warning(f"Grad scale is small: {cur_grad_scale}") + if cur_grad_scale < 1.0e-05: + raise RuntimeError( + f"grad_scale is too small, exiting: {cur_grad_scale}" + ) + if batch_idx % params.log_interval == 0: + try: + cur_lr = scheduler.get_last_lr()[0] + except: # noqa + cur_lr = 0.0 + cur_grad_scale = ( + scaler._scale.item() + if (params.use_fp16 and not params.deepspeed) + else 1.0 + ) + + logging.info( + f"Epoch {params.cur_epoch}, " + f"batch {batch_idx}, loss[{loss_info}], " + f"tot_loss[{tot_loss}], batch size: {batch_size}, " + f"lr: {cur_lr:.2e}, " + + ( + f"grad_scale: {scaler._scale.item()}" + if (params.use_fp16 and not params.deepspeed) + else "" + ) + ) + + if tb_writer is not None: + tb_writer.add_scalar( + "train/learning_rate", cur_lr, params.batch_idx_train + ) + + loss_info.write_summary( + tb_writer, "train/current_", params.batch_idx_train + ) + tot_loss.write_summary(tb_writer, "train/tot_", params.batch_idx_train) + if params.use_fp16: + tb_writer.add_scalar( + "train/grad_scale", + cur_grad_scale, + params.batch_idx_train, + ) + + loss_value = tot_loss["loss"] / tot_loss["frames"] + params.train_loss = loss_value + if params.train_loss < params.best_train_loss: + params.best_train_epoch = params.cur_epoch + params.best_train_loss = params.train_loss + + +def run(rank, world_size, args): + """ + Args: + rank: + It is a value between 0 and `world_size-1`, which is + passed automatically by `mp.spawn()` in :func:`main`. + The node with rank 0 is responsible for saving checkpoint. + world_size: + Number of GPUs for DDP training. + args: + The return value of get_parser().parse_args() + """ + params = get_params() + params.update(vars(args)) + + fix_random_seed(params.seed) + + setup_logger(f"{params.exp_dir}/log/log-train") + logging.info(params) + + logging.info("About to create model") + + replace_whisper_encoder_forward() + model = whisper.load_model(params.model_name, "cpu") + del model.alignment_heads + num_param = sum([p.numel() for p in model.parameters()]) + logging.info(f"Number of model parameters: {num_param}") + + tokenizer = whisper.tokenizer.get_tokenizer( + model.is_multilingual, + num_languages=model.num_languages, + language="zh", + task="transcribe", + ) + + model_avg: Optional[nn.Module] = None + if rank == 0: + # model_avg is only used with rank 0 + model_avg = copy.deepcopy(model).to(torch.float64) + + assert params.start_epoch > 0, params.start_epoch + checkpoints = load_checkpoint_if_available( + params=params, model=model, model_avg=model_avg + ) + + if torch.cuda.is_available(): + device = torch.device("cuda", rank) + else: + device = torch.device("cpu") + logging.info(f"Device: {device}") + model.to(device) + + optimizer = torch.optim.AdamW(model.parameters(), lr=params.base_lr) + scheduler = Eden(optimizer, params.lr_batches, params.lr_epochs) + + if checkpoints and "optimizer" in checkpoints: + logging.info("Loading optimizer state dict") + optimizer.load_state_dict(checkpoints["optimizer"]) + + if ( + checkpoints + and "scheduler" in checkpoints + and checkpoints["scheduler"] is not None + ): + logging.info("Loading scheduler state dict") + scheduler.load_state_dict(checkpoints["scheduler"]) + + if world_size > 1: + if params.deepspeed: + logging.info("Using DeepSpeed") + model, optimizer, _, scheduler = deepspeed.initialize( + args=params, model=model, model_parameters=model.parameters() + ) + else: + logging.info("Using DDP") + setup_dist(use_ddp_launch=True) + model = DDP(model, device_ids=[rank], find_unused_parameters=True) + + if params.print_diagnostics: + opts = diagnostics.TensorDiagnosticOptions( + 2**22 + ) # allow 4 megabytes per sub-module + diagnostic = diagnostics.attach_diagnostics(model, opts) + + if params.inf_check: + register_inf_check_hooks(model) + + wenetspeech = WenetSpeechAsrDataModule(args) + + if params.start_batch > 0 and checkpoints and "sampler" in checkpoints: + # We only load the sampler's state dict when it loads a checkpoint + # saved in the middle of an epoch + sampler_state_dict = checkpoints["sampler"] + else: + sampler_state_dict = None + + def remove_short_and_long_utt(c: Cut): + # Keep only utterances with duration between 1 second and 15 seconds + # + # Caution: There is a reason to select 15.0 here. Please see + # ../local/display_manifest_statistics.py + # + # You should use ../local/display_manifest_statistics.py to get + # an utterance duration distribution for your dataset to select + # the threshold + if c.duration < 1.0 or c.duration > 15.0: + # logging.warning( + # f"Exclude cut with ID {c.id} from training. Duration: {c.duration}" + # ) + return False + return True + + train_cuts = wenetspeech.train_cuts() + train_cuts = train_cuts.filter(remove_short_and_long_utt) + train_dl = wenetspeech.train_dataloaders(train_cuts) + valid_dl = wenetspeech.valid_dataloaders(wenetspeech.valid_cuts()) + + scaler = GradScaler(enabled=params.use_fp16, init_scale=1.0) + if checkpoints and "grad_scaler" in checkpoints: + logging.info("Loading grad scaler state dict") + scaler.load_state_dict(checkpoints["grad_scaler"]) + + if args.tensorboard and rank == 0: + tb_writer = SummaryWriter(log_dir=f"{params.exp_dir}/tensorboard") + else: + tb_writer = None + + logging.info(f"start training from epoch {params.start_epoch}") + for epoch in range(params.start_epoch, params.num_epochs + 1): + if not params.deepspeed: + scheduler.step_epoch(epoch - 1) + fix_random_seed(params.seed + epoch - 1) + train_dl.sampler.set_epoch(epoch - 1) + + if tb_writer is not None: + tb_writer.add_scalar("train/epoch", epoch, params.batch_idx_train) + + params.cur_epoch = epoch + + train_one_epoch( + params=params, + tokenizer=tokenizer, + model=model, + model_avg=model_avg, + optimizer=optimizer, + scheduler=scheduler, + train_dl=train_dl, + valid_dl=valid_dl, + scaler=scaler, + tb_writer=tb_writer, + world_size=world_size, + rank=rank, + ) + + if params.print_diagnostics: + diagnostic.print_diagnostics() + break + + if params.deepspeed: + model.save_checkpoint( + save_dir=params.exp_dir, + tag=f"epoch-{params.cur_epoch}", + client_state={}, + ) + if rank == 0: + convert_zero_checkpoint_to_fp32_state_dict( + params.exp_dir, + f"{params.exp_dir}/epoch-{params.cur_epoch}.pt", + tag=f"epoch-{params.cur_epoch}", + ) + else: + save_checkpoint( + params=params, + model=model, + model_avg=model_avg, + optimizer=optimizer, + scheduler=scheduler, + sampler=train_dl.sampler, + scaler=scaler, + rank=rank, + ) + + logging.info("Done!") + + if world_size > 1 and not params.deepspeed: + torch.distributed.barrier() + cleanup_dist() + + +def display_and_save_batch( + batch: dict, + params: AttributeDict, +) -> 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`. + """ + 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}") + + +def main(): + parser = get_parser() + WenetSpeechAsrDataModule.add_arguments(parser) + args = parser.parse_args() + args.exp_dir = Path(args.exp_dir) + + world_size = get_world_size() + rank = get_rank() + + torch.set_num_threads(1) + torch.set_num_interop_threads(1) + run(rank=rank, world_size=world_size, args=args) + + +if __name__ == "__main__": + main() diff --git a/egs/wenetspeech/ASR/whisper/whisper_encoder_forward_monkey_patch.py b/egs/wenetspeech/ASR/whisper/whisper_encoder_forward_monkey_patch.py new file mode 120000 index 000000000..2a7808921 --- /dev/null +++ b/egs/wenetspeech/ASR/whisper/whisper_encoder_forward_monkey_patch.py @@ -0,0 +1 @@ +../../../aishell/ASR/whisper/whisper_encoder_forward_monkey_patch.py \ No newline at end of file diff --git a/egs/wenetspeech/ASR/zipformer/train.py b/egs/wenetspeech/ASR/zipformer/train.py index b1557dedb..3d3762916 100755 --- a/egs/wenetspeech/ASR/zipformer/train.py +++ b/egs/wenetspeech/ASR/zipformer/train.py @@ -86,6 +86,7 @@ from icefall.checkpoint import ( ) from icefall.dist import cleanup_dist, setup_dist from icefall.env import get_env_info +from icefall.err import raise_grad_scale_is_too_small_error from icefall.hooks import register_inf_check_hooks from icefall.lexicon import Lexicon from icefall.utils import ( @@ -985,9 +986,7 @@ def train_one_epoch( logging.warning(f"Grad scale is small: {cur_grad_scale}") if cur_grad_scale < 1.0e-05: save_bad_model() - raise RuntimeError( - f"grad_scale is too small, exiting: {cur_grad_scale}" - ) + raise_grad_scale_is_too_small_error(cur_grad_scale) if batch_idx % params.log_interval == 0: cur_lr = max(scheduler.get_last_lr()) diff --git a/egs/wenetspeech/KWS/zipformer/decode.py b/egs/wenetspeech/KWS/zipformer/decode.py index 5ed3c6c2c..340a41231 100755 --- a/egs/wenetspeech/KWS/zipformer/decode.py +++ b/egs/wenetspeech/KWS/zipformer/decode.py @@ -30,9 +30,7 @@ import k2 import torch import torch.nn as nn from asr_datamodule import WenetSpeechAsrDataModule -from beam_search import ( - keywords_search, -) +from beam_search import keywords_search from lhotse.cut import Cut from train import add_model_arguments, get_model, get_params diff --git a/egs/wenetspeech/KWS/zipformer/finetune.py b/egs/wenetspeech/KWS/zipformer/finetune.py index 6f34989e2..3ad16fd11 100755 --- a/egs/wenetspeech/KWS/zipformer/finetune.py +++ b/egs/wenetspeech/KWS/zipformer/finetune.py @@ -87,28 +87,6 @@ from torch import Tensor from torch.cuda.amp import GradScaler from torch.nn.parallel import DistributedDataParallel as DDP from torch.utils.tensorboard import SummaryWriter - -from icefall import diagnostics -from icefall.char_graph_compiler import CharCtcTrainingGraphCompiler -from icefall.checkpoint import load_checkpoint, remove_checkpoints -from icefall.checkpoint import save_checkpoint as save_checkpoint_impl -from icefall.checkpoint import ( - save_checkpoint_with_global_batch_idx, - update_averaged_model, -) -from icefall.dist import cleanup_dist, setup_dist -from icefall.env import get_env_info -from icefall.hooks import register_inf_check_hooks -from icefall.lexicon import Lexicon -from icefall.utils import ( - AttributeDict, - MetricsTracker, - get_parameter_groups_with_lrs, - setup_logger, - str2bool, - text_to_pinyin, -) - from train import ( add_model_arguments, add_training_arguments, @@ -123,6 +101,27 @@ from train import ( set_batch_count, ) +from icefall import diagnostics +from icefall.char_graph_compiler import CharCtcTrainingGraphCompiler +from icefall.checkpoint import load_checkpoint, remove_checkpoints +from icefall.checkpoint import save_checkpoint as save_checkpoint_impl +from icefall.checkpoint import ( + save_checkpoint_with_global_batch_idx, + update_averaged_model, +) +from icefall.dist import cleanup_dist, setup_dist +from icefall.env import get_env_info +from icefall.err import raise_grad_scale_is_too_small_error +from icefall.hooks import register_inf_check_hooks +from icefall.lexicon import Lexicon +from icefall.utils import ( + AttributeDict, + MetricsTracker, + get_parameter_groups_with_lrs, + setup_logger, + str2bool, + text_to_pinyin, +) LRSchedulerType = Union[torch.optim.lr_scheduler._LRScheduler, optim.LRScheduler] @@ -527,9 +526,7 @@ def train_one_epoch( logging.warning(f"Grad scale is small: {cur_grad_scale}") if cur_grad_scale < 1.0e-05: save_bad_model() - raise RuntimeError( - f"grad_scale is too small, exiting: {cur_grad_scale}" - ) + raise_grad_scale_is_too_small_error(cur_grad_scale) if batch_idx % params.log_interval == 0: cur_lr = max(scheduler.get_last_lr()) diff --git a/egs/wenetspeech/KWS/zipformer/train.py b/egs/wenetspeech/KWS/zipformer/train.py index 5be34ed99..b5cf3359a 100755 --- a/egs/wenetspeech/KWS/zipformer/train.py +++ b/egs/wenetspeech/KWS/zipformer/train.py @@ -88,6 +88,7 @@ from icefall.checkpoint import ( ) from icefall.dist import cleanup_dist, setup_dist from icefall.env import get_env_info +from icefall.err import raise_grad_scale_is_too_small_error from icefall.hooks import register_inf_check_hooks from icefall.lexicon import Lexicon from icefall.utils import ( @@ -99,7 +100,6 @@ from icefall.utils import ( text_to_pinyin, ) - LRSchedulerType = Union[torch.optim.lr_scheduler._LRScheduler, optim.LRScheduler] @@ -1043,9 +1043,7 @@ def train_one_epoch( logging.warning(f"Grad scale is small: {cur_grad_scale}") if cur_grad_scale < 1.0e-05: save_bad_model() - raise RuntimeError( - f"grad_scale is too small, exiting: {cur_grad_scale}" - ) + raise_grad_scale_is_too_small_error(cur_grad_scale) if batch_idx % params.log_interval == 0: cur_lr = max(scheduler.get_last_lr()) diff --git a/egs/xbmu_amdo31/ASR/pruned_transducer_stateless7/train.py b/egs/xbmu_amdo31/ASR/pruned_transducer_stateless7/train.py index 8c53972fd..d24c27326 100755 --- a/egs/xbmu_amdo31/ASR/pruned_transducer_stateless7/train.py +++ b/egs/xbmu_amdo31/ASR/pruned_transducer_stateless7/train.py @@ -81,6 +81,7 @@ from icefall.checkpoint import ( ) from icefall.dist import cleanup_dist, setup_dist from icefall.env import get_env_info +from icefall.err import raise_grad_scale_is_too_small_error from icefall.hooks import register_inf_check_hooks from icefall.utils import AttributeDict, MetricsTracker, setup_logger, str2bool @@ -854,9 +855,7 @@ def train_one_epoch( if cur_grad_scale < 0.01: logging.warning(f"Grad scale is small: {cur_grad_scale}") if cur_grad_scale < 1.0e-05: - raise RuntimeError( - f"grad_scale is too small, exiting: {cur_grad_scale}" - ) + raise_grad_scale_is_too_small_error(cur_grad_scale) if batch_idx % params.log_interval == 0: cur_lr = scheduler.get_last_lr()[0] diff --git a/egs/yesno/ASR/tdnn/jit_pretrained.py b/egs/yesno/ASR/tdnn/jit_pretrained.py index e29415ffb..6c643c263 100755 --- a/egs/yesno/ASR/tdnn/jit_pretrained.py +++ b/egs/yesno/ASR/tdnn/jit_pretrained.py @@ -18,9 +18,8 @@ you can use ./export.py --jit 1 import argparse import logging -from typing import List import math - +from typing import List import k2 import kaldifeat diff --git a/icefall/byte_utils.py b/icefall/byte_utils.py index 79c1c7545..5f5cc710b 100644 --- a/icefall/byte_utils.py +++ b/icefall/byte_utils.py @@ -8,7 +8,6 @@ import re import unicodedata - WHITESPACE_NORMALIZER = re.compile(r"\s+") SPACE = chr(32) SPACE_ESCAPE = chr(9601) diff --git a/icefall/ctc/prepare_lang.py b/icefall/ctc/prepare_lang.py index 4801b1beb..0e99e70d8 100644 --- a/icefall/ctc/prepare_lang.py +++ b/icefall/ctc/prepare_lang.py @@ -8,12 +8,12 @@ The lang_dir should contain the following files: """ import math +import re from collections import defaultdict from pathlib import Path from typing import List, Tuple import kaldifst -import re class Lexicon: diff --git a/icefall/diagnostics.py b/icefall/diagnostics.py index 65b6f67b0..a3c480c9c 100644 --- a/icefall/diagnostics.py +++ b/icefall/diagnostics.py @@ -18,7 +18,7 @@ import random from dataclasses import dataclass -from typing import Optional, Tuple, List +from typing import List, Optional, Tuple import torch from torch import Tensor, nn diff --git a/icefall/err.py b/icefall/err.py new file mode 100644 index 000000000..065e2a53d --- /dev/null +++ b/icefall/err.py @@ -0,0 +1,47 @@ +# Copyright 2024 Xiaomi Corp. (authors: Zengrui Jin,) +# +# 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. + + +def raise_grad_scale_is_too_small_error(cur_grad_scale: float): + raise RuntimeError( + f""" + grad_scale is too small, exiting: {cur_grad_scale} + + ========================= NOTE ========================= + If you see this error, it means that the gradient scale is too small. + + The default base_lr is 0.045 / 0.05 (depends on which recipe you are + using), this is an empirical value obtained mostly using 4 * 32GB V100 + GPUs with a max_duration of approx. 1,000. + The proper value of base_lr may vary depending on the number of GPUs + and the value of max-duration you are using. + + To fix this issue, you may need to adjust the value of base_lr accordingly. + + We would suggest you to decrease the value of base_lr by 0.005 (e.g., + from 0.045 to 0.04), and try again. If the error still exists, you may + repeat the process until base_lr hits 0.02. (Note that this will lead to + certain loss of performance, but it should work. You can compensate this by + increasing the num_epochs.) + + If the error still exists, you could try to seek help by raising an issue, + with a detailed description of (a) your computational resources, (b) the + base_lr and (c) the max_duration you are using, (d) detailed configuration + of your model. + + ======================================================== + """ + ) diff --git a/icefall/profiler.py b/icefall/profiler.py index 49e138579..762105c48 100644 --- a/icefall/profiler.py +++ b/icefall/profiler.py @@ -5,14 +5,15 @@ # This is modified from https://github.com/microsoft/DeepSpeed/blob/master/deepspeed/profiling/flops_profiler/profiler.py +from collections import OrderedDict +from functools import partial +from typing import List, Optional + import k2 +import numpy as np import torch import torch.nn as nn import torch.nn.functional as F -from functools import partial -from typing import List, Optional -from collections import OrderedDict -import numpy as np Tensor = torch.Tensor diff --git a/icefall/rnn_lm/export-onnx.py b/icefall/rnn_lm/export-onnx.py index dfede708b..1070d443a 100755 --- a/icefall/rnn_lm/export-onnx.py +++ b/icefall/rnn_lm/export-onnx.py @@ -5,16 +5,16 @@ import argparse import logging from pathlib import Path +from typing import Dict import onnx import torch from model import RnnLmModel from onnxruntime.quantization import QuantType, quantize_dynamic +from train import get_params from icefall.checkpoint import average_checkpoints, find_checkpoints, load_checkpoint from icefall.utils import AttributeDict, str2bool -from typing import Dict -from train import get_params def add_meta_data(filename: str, meta_data: Dict[str, str]): diff --git a/icefall/utils.py b/icefall/utils.py index 7d722b1bc..31f9801d9 100644 --- a/icefall/utils.py +++ b/icefall/utils.py @@ -28,8 +28,6 @@ from contextlib import contextmanager from dataclasses import dataclass from datetime import datetime from pathlib import Path -from pypinyin import pinyin, lazy_pinyin -from pypinyin.contrib.tone_convert import to_initials, to_finals_tone, to_finals from shutil import copyfile from typing import Dict, Iterable, List, Optional, TextIO, Tuple, Union @@ -40,6 +38,8 @@ import sentencepiece as spm import torch import torch.distributed as dist import torch.nn as nn +from pypinyin import lazy_pinyin, pinyin +from pypinyin.contrib.tone_convert import to_finals, to_finals_tone, to_initials from torch.utils.tensorboard import SummaryWriter from icefall.checkpoint import average_checkpoints diff --git a/requirements-tts.txt b/requirements-tts.txt index c30e23d54..eae50ba7b 100644 --- a/requirements-tts.txt +++ b/requirements-tts.txt @@ -3,4 +3,5 @@ matplotlib==3.8.2 cython==3.0.6 numba==0.58.1 g2p_en==2.1.0 -espnet_tts_frontend==0.0.3 \ No newline at end of file +espnet_tts_frontend==0.0.3 +# piper_phonemize: refer to https://github.com/rhasspy/piper-phonemize, could install the pre-built wheels from https://github.com/csukuangfj/piper-phonemize/releases/tag/2023.12.5 diff --git a/requirements.txt b/requirements.txt index dec20c6e4..6bafa6aca 100644 --- a/requirements.txt +++ b/requirements.txt @@ -1,4 +1,4 @@ -kaldifst +kaldifst>1.7.0 kaldilm kaldialign num2words @@ -8,6 +8,13 @@ pypinyin==0.50.0 tensorboard typeguard dill -black==22.3.0 onnx==1.15.0 onnxruntime==1.16.3 + +# style check session: +black==22.3.0 +isort==5.10.1 +flake8==5.0.4 + +# cantonese word segment support +pycantonese==3.4.0 \ No newline at end of file