diff --git a/.github/workflows/run-pretrained-transducer-stateless-modified-2-aishell.yml b/.github/workflows/run-pretrained-transducer-stateless-modified-2-aishell.yml new file mode 100644 index 000000000..c27ffc374 --- /dev/null +++ b/.github/workflows/run-pretrained-transducer-stateless-modified-2-aishell.yml @@ -0,0 +1,153 @@ +# Copyright 2021 Fangjun Kuang (csukuangfj@gmail.com) + +# See ../../LICENSE for clarification regarding multiple authors +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +name: run-pre-trained-trandsucer-stateless-modified-2-aishell + +on: + push: + branches: + - master + pull_request: + types: [labeled] + +jobs: + run_pre_trained_transducer_stateless_modified_2_aishell: + if: github.event.label.name == 'ready' || github.event_name == 'push' + runs-on: ${{ matrix.os }} + strategy: + matrix: + os: [ubuntu-18.04] + python-version: [3.7, 3.8, 3.9] + torch: ["1.10.0"] + torchaudio: ["0.10.0"] + k2-version: ["1.9.dev20211101"] + + fail-fast: false + + steps: + - uses: actions/checkout@v2 + with: + fetch-depth: 0 + + - name: Setup Python ${{ matrix.python-version }} + uses: actions/setup-python@v1 + with: + python-version: ${{ matrix.python-version }} + + - name: Install Python dependencies + run: | + python3 -m pip install --upgrade pip pytest + # numpy 1.20.x does not support python 3.6 + pip install numpy==1.19 + pip install torch==${{ matrix.torch }}+cpu torchaudio==${{ matrix.torchaudio }}+cpu -f https://download.pytorch.org/whl/cpu/torch_stable.html + pip install k2==${{ matrix.k2-version }}+cpu.torch${{ matrix.torch }} -f https://k2-fsa.org/nightly/ + + python3 -m pip install git+https://github.com/lhotse-speech/lhotse + python3 -m pip install kaldifeat + # We are in ./icefall and there is a file: requirements.txt in it + pip install -r requirements.txt + + - name: Install graphviz + shell: bash + run: | + python3 -m pip install -qq graphviz + sudo apt-get -qq install graphviz + + - name: Download pre-trained model + shell: bash + run: | + sudo apt-get -qq install git-lfs tree sox + cd egs/aishell/ASR + mkdir tmp + cd tmp + git lfs install + git clone https://huggingface.co/csukuangfj/icefall-aishell-transducer-stateless-modified-2-2022-03-01 + + cd .. + tree tmp + soxi tmp/icefall-aishell-transducer-stateless-modified-2-2022-03-01/test_wavs/*.wav + ls -lh tmp/icefall-aishell-transducer-stateless-modified-2-2022-03-01/test_wavs/*.wav + + - name: Run greedy search decoding (max-sym-per-frame 1) + shell: bash + run: | + export PYTHONPATH=$PWD:PYTHONPATH + cd egs/aishell/ASR + ./transducer_stateless_modified-2/pretrained.py \ + --method greedy_search \ + --max-sym-per-frame 1 \ + --checkpoint ./tmp/icefall-aishell-transducer-stateless-modified-2-2022-03-01/exp/pretrained.pt \ + --lang-dir ./tmp/icefall-aishell-transducer-stateless-modified-2-2022-03-01/data/lang_char \ + ./tmp/icefall-aishell-transducer-stateless-modified-2-2022-03-01/test_wavs/BAC009S0764W0121.wav \ + ./tmp/icefall-aishell-transducer-stateless-modified-2-2022-03-01/test_wavs/BAC009S0764W0122.wav \ + ./tmp/icefall-aishell-transducer-stateless-modified-2-2022-03-01/test_wavs/BAC009S0764W0123.wav + + - name: Run greedy search decoding (max-sym-per-frame 2) + shell: bash + run: | + export PYTHONPATH=$PWD:PYTHONPATH + cd egs/aishell/ASR + ./transducer_stateless_modified-2/pretrained.py \ + --method greedy_search \ + --max-sym-per-frame 2 \ + --checkpoint ./tmp/icefall-aishell-transducer-stateless-modified-2-2022-03-01/exp/pretrained.pt \ + --lang-dir ./tmp/icefall-aishell-transducer-stateless-modified-2-2022-03-01/data/lang_char \ + ./tmp/icefall-aishell-transducer-stateless-modified-2-2022-03-01/test_wavs/BAC009S0764W0121.wav \ + ./tmp/icefall-aishell-transducer-stateless-modified-2-2022-03-01/test_wavs/BAC009S0764W0122.wav \ + ./tmp/icefall-aishell-transducer-stateless-modified-2-2022-03-01/test_wavs/BAC009S0764W0123.wav + + - name: Run greedy search decoding (max-sym-per-frame 3) + shell: bash + run: | + export PYTHONPATH=$PWD:PYTHONPATH + cd egs/aishell/ASR + ./transducer_stateless_modified-2/pretrained.py \ + --method greedy_search \ + --max-sym-per-frame 3 \ + --checkpoint ./tmp/icefall-aishell-transducer-stateless-modified-2-2022-03-01/exp/pretrained.pt \ + --lang-dir ./tmp/icefall-aishell-transducer-stateless-modified-2-2022-03-01/data/lang_char \ + ./tmp/icefall-aishell-transducer-stateless-modified-2-2022-03-01/test_wavs/BAC009S0764W0121.wav \ + ./tmp/icefall-aishell-transducer-stateless-modified-2-2022-03-01/test_wavs/BAC009S0764W0122.wav \ + ./tmp/icefall-aishell-transducer-stateless-modified-2-2022-03-01/test_wavs/BAC009S0764W0123.wav + + - name: Run beam search decoding + shell: bash + run: | + export PYTHONPATH=$PWD:$PYTHONPATH + cd egs/aishell/ASR + ./transducer_stateless_modified-2/pretrained.py \ + --method beam_search \ + --beam-size 4 \ + --checkpoint ./tmp/icefall-aishell-transducer-stateless-modified-2-2022-03-01/exp/pretrained.pt \ + --lang-dir ./tmp/icefall-aishell-transducer-stateless-modified-2-2022-03-01/data/lang_char \ + ./tmp/icefall-aishell-transducer-stateless-modified-2-2022-03-01/test_wavs/BAC009S0764W0121.wav \ + ./tmp/icefall-aishell-transducer-stateless-modified-2-2022-03-01/test_wavs/BAC009S0764W0122.wav \ + ./tmp/icefall-aishell-transducer-stateless-modified-2-2022-03-01/test_wavs/BAC009S0764W0123.wav + + + - name: Run modified beam search decoding + shell: bash + run: | + export PYTHONPATH=$PWD:$PYTHONPATH + cd egs/aishell/ASR + ./transducer_stateless_modified-2/pretrained.py \ + --method modified_beam_search \ + --beam-size 4 \ + --checkpoint ./tmp/icefall-aishell-transducer-stateless-modified-2-2022-03-01/exp/pretrained.pt \ + --lang-dir ./tmp/icefall-aishell-transducer-stateless-modified-2-2022-03-01/data/lang_char \ + ./tmp/icefall-aishell-transducer-stateless-modified-2-2022-03-01/test_wavs/BAC009S0764W0121.wav \ + ./tmp/icefall-aishell-transducer-stateless-modified-2-2022-03-01/test_wavs/BAC009S0764W0122.wav \ + ./tmp/icefall-aishell-transducer-stateless-modified-2-2022-03-01/test_wavs/BAC009S0764W0123.wav diff --git a/.github/workflows/run-pretrained-transducer-stateless-modified-aishell.yml b/.github/workflows/run-pretrained-transducer-stateless-modified-aishell.yml new file mode 100644 index 000000000..2e38abb5a --- /dev/null +++ b/.github/workflows/run-pretrained-transducer-stateless-modified-aishell.yml @@ -0,0 +1,153 @@ +# Copyright 2021 Fangjun Kuang (csukuangfj@gmail.com) + +# See ../../LICENSE for clarification regarding multiple authors +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +name: run-pre-trained-trandsucer-stateless-modified-aishell + +on: + push: + branches: + - master + pull_request: + types: [labeled] + +jobs: + run_pre_trained_transducer_stateless_modified_aishell: + if: github.event.label.name == 'ready' || github.event_name == 'push' + runs-on: ${{ matrix.os }} + strategy: + matrix: + os: [ubuntu-18.04] + python-version: [3.7, 3.8, 3.9] + torch: ["1.10.0"] + torchaudio: ["0.10.0"] + k2-version: ["1.9.dev20211101"] + + fail-fast: false + + steps: + - uses: actions/checkout@v2 + with: + fetch-depth: 0 + + - name: Setup Python ${{ matrix.python-version }} + uses: actions/setup-python@v1 + with: + python-version: ${{ matrix.python-version }} + + - name: Install Python dependencies + run: | + python3 -m pip install --upgrade pip pytest + # numpy 1.20.x does not support python 3.6 + pip install numpy==1.19 + pip install torch==${{ matrix.torch }}+cpu torchaudio==${{ matrix.torchaudio }}+cpu -f https://download.pytorch.org/whl/cpu/torch_stable.html + pip install k2==${{ matrix.k2-version }}+cpu.torch${{ matrix.torch }} -f https://k2-fsa.org/nightly/ + + python3 -m pip install git+https://github.com/lhotse-speech/lhotse + python3 -m pip install kaldifeat + # We are in ./icefall and there is a file: requirements.txt in it + pip install -r requirements.txt + + - name: Install graphviz + shell: bash + run: | + python3 -m pip install -qq graphviz + sudo apt-get -qq install graphviz + + - name: Download pre-trained model + shell: bash + run: | + sudo apt-get -qq install git-lfs tree sox + cd egs/aishell/ASR + mkdir tmp + cd tmp + git lfs install + git clone https://huggingface.co/csukuangfj/icefall-aishell-transducer-stateless-modified-2022-03-01 + + cd .. + tree tmp + soxi tmp/icefall-aishell-transducer-stateless-modified-2022-03-01/test_wavs/*.wav + ls -lh tmp/icefall-aishell-transducer-stateless-modified-2022-03-01/test_wavs/*.wav + + - name: Run greedy search decoding (max-sym-per-frame 1) + shell: bash + run: | + export PYTHONPATH=$PWD:PYTHONPATH + cd egs/aishell/ASR + ./transducer_stateless_modified/pretrained.py \ + --method greedy_search \ + --max-sym-per-frame 1 \ + --checkpoint ./tmp/icefall-aishell-transducer-stateless-modified-2022-03-01/exp/pretrained.pt \ + --lang-dir ./tmp/icefall-aishell-transducer-stateless-modified-2022-03-01/data/lang_char \ + ./tmp/icefall-aishell-transducer-stateless-modified-2022-03-01/test_wavs/BAC009S0764W0121.wav \ + ./tmp/icefall-aishell-transducer-stateless-modified-2022-03-01/test_wavs/BAC009S0764W0122.wav \ + ./tmp/icefall-aishell-transducer-stateless-modified-2022-03-01/test_wavs/BAC009S0764W0123.wav + + - name: Run greedy search decoding (max-sym-per-frame 2) + shell: bash + run: | + export PYTHONPATH=$PWD:PYTHONPATH + cd egs/aishell/ASR + ./transducer_stateless_modified/pretrained.py \ + --method greedy_search \ + --max-sym-per-frame 2 \ + --checkpoint ./tmp/icefall-aishell-transducer-stateless-modified-2022-03-01/exp/pretrained.pt \ + --lang-dir ./tmp/icefall-aishell-transducer-stateless-modified-2022-03-01/data/lang_char \ + ./tmp/icefall-aishell-transducer-stateless-modified-2022-03-01/test_wavs/BAC009S0764W0121.wav \ + ./tmp/icefall-aishell-transducer-stateless-modified-2022-03-01/test_wavs/BAC009S0764W0122.wav \ + ./tmp/icefall-aishell-transducer-stateless-modified-2022-03-01/test_wavs/BAC009S0764W0123.wav + + - name: Run greedy search decoding (max-sym-per-frame 3) + shell: bash + run: | + export PYTHONPATH=$PWD:PYTHONPATH + cd egs/aishell/ASR + ./transducer_stateless_modified/pretrained.py \ + --method greedy_search \ + --max-sym-per-frame 3 \ + --checkpoint ./tmp/icefall-aishell-transducer-stateless-modified-2022-03-01/exp/pretrained.pt \ + --lang-dir ./tmp/icefall-aishell-transducer-stateless-modified-2022-03-01/data/lang_char \ + ./tmp/icefall-aishell-transducer-stateless-modified-2022-03-01/test_wavs/BAC009S0764W0121.wav \ + ./tmp/icefall-aishell-transducer-stateless-modified-2022-03-01/test_wavs/BAC009S0764W0122.wav \ + ./tmp/icefall-aishell-transducer-stateless-modified-2022-03-01/test_wavs/BAC009S0764W0123.wav + + - name: Run beam search decoding + shell: bash + run: | + export PYTHONPATH=$PWD:$PYTHONPATH + cd egs/aishell/ASR + ./transducer_stateless_modified/pretrained.py \ + --method beam_search \ + --beam-size 4 \ + --checkpoint ./tmp/icefall-aishell-transducer-stateless-modified-2022-03-01/exp/pretrained.pt \ + --lang-dir ./tmp/icefall-aishell-transducer-stateless-modified-2022-03-01/data/lang_char \ + ./tmp/icefall-aishell-transducer-stateless-modified-2022-03-01/test_wavs/BAC009S0764W0121.wav \ + ./tmp/icefall-aishell-transducer-stateless-modified-2022-03-01/test_wavs/BAC009S0764W0122.wav \ + ./tmp/icefall-aishell-transducer-stateless-modified-2022-03-01/test_wavs/BAC009S0764W0123.wav + + + - name: Run modified beam search decoding + shell: bash + run: | + export PYTHONPATH=$PWD:$PYTHONPATH + cd egs/aishell/ASR + ./transducer_stateless_modified/pretrained.py \ + --method modified_beam_search \ + --beam-size 4 \ + --checkpoint ./tmp/icefall-aishell-transducer-stateless-modified-2022-03-01/exp/pretrained.pt \ + --lang-dir ./tmp/icefall-aishell-transducer-stateless-modified-2022-03-01/data/lang_char \ + ./tmp/icefall-aishell-transducer-stateless-modified-2022-03-01/test_wavs/BAC009S0764W0121.wav \ + ./tmp/icefall-aishell-transducer-stateless-modified-2022-03-01/test_wavs/BAC009S0764W0122.wav \ + ./tmp/icefall-aishell-transducer-stateless-modified-2022-03-01/test_wavs/BAC009S0764W0123.wav diff --git a/README.md b/README.md index ec9d7e69c..a49b30df0 100644 --- a/README.md +++ b/README.md @@ -113,7 +113,7 @@ The best CER we currently have is: | | test | |-----|------| -| CER | 5.05 | +| CER | 4.68 | We provide a Colab notebook to run a pre-trained TransducerStateless model: [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/drive/14XaT2MhnBkK-3_RqqWq3K90Xlbin-GZC?usp=sharing) diff --git a/egs/aishell/ASR/README.md b/egs/aishell/ASR/README.md index 3fd177376..1b3c5a2e3 100644 --- a/egs/aishell/ASR/README.md +++ b/egs/aishell/ASR/README.md @@ -1,3 +1,20 @@ +# Introduction + Please refer to for how to run models in this recipe. + +# Transducers + +There are various folders containing the name `transducer` in this folder. +The following table lists the differences among them. + +| | Encoder | Decoder | Comment | +|------------------------------------|-----------|--------------------|-----------------------------------------------------------------------------------| +| `transducer_stateless` | Conformer | Embedding + Conv1d | with `k2.rnnt_loss` | +| `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 | + +The decoder in `transducer_stateless` 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/aishell/ASR/RESULTS.md b/egs/aishell/ASR/RESULTS.md index 53cc394a1..ecc93c21b 100644 --- a/egs/aishell/ASR/RESULTS.md +++ b/egs/aishell/ASR/RESULTS.md @@ -1,12 +1,153 @@ ## Results ### Aishell training result(Transducer-stateless) + +#### 2022-03-01 + +[./transducer_stateless_modified-2](./transducer_stateless_modified-2) + +Stateless transducer + modified transducer + using [aidatatang_200zh](http://www.openslr.org/62/) as extra training data. + + +| | test |comment | +|------------------------|------|----------------------------------------------------------------| +| greedy search | 4.94 |--epoch 89, --avg 38, --max-duration 100, --max-sym-per-frame 1 | +| modified beam search | 4.68 |--epoch 89, --avg 38, --max-duration 100 --beam-size 4 | + +The training commands are: + +```bash +cd egs/aishell/ASR +./prepare.sh --stop-stage 6 +./prepare_aidatatang_200zh.sh + +export CUDA_VISIBLE_DEVICES="0,1,2" + +./transducer_stateless_modified-2/train.py \ + --world-size 3 \ + --num-epochs 90 \ + --start-epoch 0 \ + --exp-dir transducer_stateless_modified-2/exp-2 \ + --max-duration 250 \ + --lr-factor 2.0 \ + --context-size 2 \ + --modified-transducer-prob 0.25 \ + --datatang-prob 0.2 +``` + +The tensorboard log is available at + + +The commands for decoding are + +```bash +# greedy search +for epoch in 89; do + for avg in 38; do + ./transducer_stateless_modified-2/decode.py \ + --epoch $epoch \ + --avg $avg \ + --exp-dir transducer_stateless_modified-2/exp-2 \ + --max-duration 100 \ + --context-size 2 \ + --decoding-method greedy_search \ + --max-sym-per-frame 1 + done +done + +# modified beam search +for epoch in 89; do + for avg in 38; do + ./transducer_stateless_modified-2/decode.py \ + --epoch $epoch \ + --avg $avg \ + --exp-dir transducer_stateless_modified-2/exp-2 \ + --max-duration 100 \ + --context-size 2 \ + --decoding-method modified_beam_search \ + --beam-size 4 + done +done +``` + +You can find a pre-trained model, decoding logs, and decoding results at + + +#### 2022-03-01 + +[./transducer_stateless_modified](./transducer_stateless_modified) + +Stateless transducer + modified transducer. + +| | test |comment | +|------------------------|------|----------------------------------------------------------------| +| greedy search | 5.22 |--epoch 64, --avg 33, --max-duration 100, --max-sym-per-frame 1 | +| modified beam search | 5.02 |--epoch 64, --avg 33, --max-duration 100 --beam-size 4 | + +The training commands are: + +```bash +cd egs/aishell/ASR +./prepare.sh --stop-stage 6 + +export CUDA_VISIBLE_DEVICES="0,1,2" + +./transducer_stateless_modified/train.py \ + --world-size 3 \ + --num-epochs 90 \ + --start-epoch 0 \ + --exp-dir transducer_stateless_modified/exp-4 \ + --max-duration 250 \ + --lr-factor 2.0 \ + --context-size 2 \ + --modified-transducer-prob 0.25 +``` + +The tensorboard log is available at + + +The commands for decoding are + +```bash +# greedy search +for epoch in 64; do + for avg in 33; do + ./transducer_stateless_modified/decode.py \ + --epoch $epoch \ + --avg $avg \ + --exp-dir transducer_stateless_modified/exp-4 \ + --max-duration 100 \ + --context-size 2 \ + --decoding-method greedy_search \ + --max-sym-per-frame 1 + done +done + +# modified beam search +for epoch in 64; do + for avg in 33; do + ./transducer_stateless_modified/decode.py \ + --epoch $epoch \ + --avg $avg \ + --exp-dir transducer_stateless_modified/exp-4 \ + --max-duration 100 \ + --context-size 2 \ + --decoding-method modified_beam_search \ + --beam-size 4 + done +done +``` + +You can find a pre-trained model, decoding logs, and decoding results at + + + #### 2022-2-19 (Duo Ma): The tensorboard log for training is available at https://tensorboard.dev/experiment/25PmX3MxSVGTdvIdhOwllw/#scalars You can find a pretrained model by visiting https://huggingface.co/shuanguanma/icefall_aishell_transducer_stateless_context_size2_epoch60_2022_2_19 | | test |comment | |---------------------------|------|-----------------------------------------| -| greedy search | 5.4 |--epoch 59, --avg 10, --max-duration 100 | -| beam search | 5.05|--epoch 59, --avg 10, --max-duration 100 | +| greedy search | 5.4 |--epoch 59, --avg 10, --max-duration 100 | +| beam search | 5.05|--epoch 59, --avg 10, --max-duration 100 | You can use the following commands to reproduce our results: @@ -23,7 +164,7 @@ python3 ./transducer_stateless/train.py \ lang_dir=data/lang_char dir=exp/transducer_stateless_context_size2 -python3 ./transducer_stateless/decode.py\ +python3 ./transducer_stateless/decode.py \ --epoch 59 \ --avg 10 \ --exp-dir $dir \ @@ -35,8 +176,8 @@ python3 ./transducer_stateless/decode.py\ lang_dir=data/lang_char dir=exp/transducer_stateless_context_size2 python3 ./transducer_stateless/decode.py \ - --epoch 59\ - --avg 10\ + --epoch 59 \ + --avg 10 \ --exp-dir $dir \ --lang-dir $lang_dir \ --decoding-method beam_search \ diff --git a/egs/aishell/ASR/local/display_manifest_statistics.py b/egs/aishell/ASR/local/display_manifest_statistics.py index 5e8b5cd3a..0ae731a1d 100755 --- a/egs/aishell/ASR/local/display_manifest_statistics.py +++ b/egs/aishell/ASR/local/display_manifest_statistics.py @@ -31,7 +31,10 @@ from lhotse import load_manifest def main(): # path = "./data/fbank/cuts_train.json.gz" # path = "./data/fbank/cuts_test.json.gz" - path = "./data/fbank/cuts_dev.json.gz" + # path = "./data/fbank/cuts_dev.json.gz" + # path = "./data/fbank/aidatatang_200zh/cuts_train_raw.jsonl.gz" + # path = "./data/fbank/aidatatang_200zh/cuts_test_raw.jsonl.gz" + path = "./data/fbank/aidatatang_200zh/cuts_dev_raw.jsonl.gz" cuts = load_manifest(path) cuts.describe() @@ -115,4 +118,79 @@ min 1.6 99.5% 8.9 99.9% 10.3 max 12.5 + +## aidatatang_200zh (train) +Cuts count: 164905 +Total duration (hours): 139.9 +Speech duration (hours): 139.9 (100.0%) +*** +Duration statistics (seconds): +mean 3.1 +std 1.1 +min 1.1 +0.1% 1.5 +0.5% 1.7 +1% 1.8 +5% 2.0 +10% 2.1 +10% 2.1 +25% 2.3 +50% 2.7 +75% 3.4 +90% 4.6 +95% 5.4 +99% 7.1 +99.5% 7.8 +99.9% 9.1 +max 16.3 + +## aidatatang_200zh (test) +Cuts count: 48144 +Total duration (hours): 40.2 +Speech duration (hours): 40.2 (100.0%) +*** +Duration statistics (seconds): +mean 3.0 +std 1.1 +min 0.9 +0.1% 1.5 +0.5% 1.8 +1% 1.8 +5% 2.0 +10% 2.1 +10% 2.1 +25% 2.3 +50% 2.6 +75% 3.4 +90% 4.4 +95% 5.2 +99% 6.9 +99.5% 7.5 +99.9% 9.0 +max 21.8 + +## aidatatang_200zh (dev) +Cuts count: 24216 +Total duration (hours): 20.2 +Speech duration (hours): 20.2 (100.0%) +*** +Duration statistics (seconds): +mean 3.0 +std 1.0 +min 1.2 +0.1% 1.6 +0.5% 1.7 +1% 1.8 +5% 2.0 +10% 2.1 +10% 2.1 +25% 2.3 +50% 2.7 +75% 3.4 +90% 4.4 +95% 5.1 +99% 6.7 +99.5% 7.3 +99.9% 8.8 +max 11.3 """ diff --git a/egs/aishell/ASR/local/process_aidatatang_200zh.py b/egs/aishell/ASR/local/process_aidatatang_200zh.py new file mode 100755 index 000000000..2c6951d42 --- /dev/null +++ b/egs/aishell/ASR/local/process_aidatatang_200zh.py @@ -0,0 +1,72 @@ +#!/usr/bin/env python3 +# Copyright 2022 Xiaomi Corp. (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. + +import logging +from pathlib import Path + +from lhotse import CutSet +from lhotse.recipes.utils import read_manifests_if_cached + + +def preprocess_aidatatang_200zh(): + src_dir = Path("data/manifests/aidatatang_200zh") + output_dir = Path("data/fbank/aidatatang_200zh") + output_dir.mkdir(exist_ok=True, parents=True) + + dataset_parts = ( + "train", + "test", + "dev", + ) + + logging.info("Loading manifest") + manifests = read_manifests_if_cached( + dataset_parts=dataset_parts, + output_dir=src_dir, + ) + assert len(manifests) > 0 + + for partition, m in manifests.items(): + logging.info(f"Processing {partition}") + raw_cuts_path = output_dir / f"cuts_{partition}_raw.jsonl.gz" + if raw_cuts_path.is_file(): + logging.info(f"{partition} already exists - skipping") + continue + + for sup in m["supervisions"]: + sup.custom = {"origin": "aidatatang_200zh"} + + cut_set = CutSet.from_manifests( + recordings=m["recordings"], + supervisions=m["supervisions"], + ) + + logging.info(f"Saving to {raw_cuts_path}") + cut_set.to_file(raw_cuts_path) + + +def main(): + formatter = ( + "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s" + ) + logging.basicConfig(format=formatter, level=logging.INFO) + + preprocess_aidatatang_200zh() + + +if __name__ == "__main__": + main() diff --git a/egs/aishell/ASR/prepare_aidatatang_200zh.sh b/egs/aishell/ASR/prepare_aidatatang_200zh.sh new file mode 100755 index 000000000..60b2060ec --- /dev/null +++ b/egs/aishell/ASR/prepare_aidatatang_200zh.sh @@ -0,0 +1,59 @@ +#!/usr/bin/env bash + +set -eou pipefail + +stage=-1 +stop_stage=100 + +# We assume dl_dir (download dir) contains the following +# directories and files. If not, they will be downloaded +# by this script automatically. +# +# - $dl_dir/aidatatang_200zh +# You can find "corpus" and "transcript" inside it. +# You can download it at +# https://openslr.org/62/ + +dl_dir=$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 [ ! -f $dl_dir/aidatatang_200zh/transcript/aidatatang_200_zh_transcript.txt ]; then + lhotse download aidatatang-200zh $dl_dir + fi +fi + +if [ $stage -le 1 ] && [ $stop_stage -ge 1 ]; then + log "Stage 1: Prepare manifest" + # We assume that you have downloaded the aidatatang_200zh corpus + # to $dl_dir/aidatatang_200zh + if [ ! -f data/manifests/aidatatang_200zh/.manifests.done ]; then + mkdir -p data/manifests/aidatatang_200zh + lhotse prepare aidatatang-200zh $dl_dir data/manifests/aidatatang_200zh + touch data/manifests/aidatatang_200zh/.manifests.done + fi +fi + +if [ $stage -le 2 ] && [ $stop_stage -ge 2 ]; then + log "Stage 2: Process aidatatang_200zh" + if [ ! -f data/fbank/aidatatang_200zh/.fbank.done ]; then + mkdir -p data/fbank/aidatatang_200zh + lhotse prepare aidatatang-200zh $dl_dir data/manifests/aidatatang_200zh + touch data/fbank/aidatatang_200zh/.fbank.done + fi +fi diff --git a/egs/aishell/ASR/transducer_stateless/model.py b/egs/aishell/ASR/transducer_stateless/model.py index c19325a15..994305fc1 100644 --- a/egs/aishell/ASR/transducer_stateless/model.py +++ b/egs/aishell/ASR/transducer_stateless/model.py @@ -14,15 +14,9 @@ # See the License for the specific language governing permissions and # limitations under the License. -""" -Note we use `rnnt_loss` from torchaudio, which exists only in -torchaudio >= v0.10.0. It also means you have to use torch >= v1.10.0 -""" import k2 import torch import torch.nn as nn -import torchaudio -import torchaudio.functional from encoder_interface import EncoderInterface from icefall.utils import add_sos @@ -115,11 +109,6 @@ class Transducer(nn.Module): boundary[:, 2] = y_lens boundary[:, 3] = x_lens - assert hasattr(torchaudio.functional, "rnnt_loss"), ( - f"Current torchaudio version: {torchaudio.__version__}\n" - "Please install a version >= 0.10.0" - ) - loss = k2.rnnt_loss(logits, y_padded, blank_id, boundary) return loss diff --git a/egs/aishell/ASR/transducer_stateless_modified-2/README.md b/egs/aishell/ASR/transducer_stateless_modified-2/README.md new file mode 100644 index 000000000..b3c539670 --- /dev/null +++ b/egs/aishell/ASR/transducer_stateless_modified-2/README.md @@ -0,0 +1,59 @@ +## Introduction + +The decoder, i.e., the prediction network, is from +https://ieeexplore.ieee.org/stamp/stamp.jsp?arnumber=9054419 +(Rnn-Transducer with Stateless Prediction Network) + +Different from `../transducer_stateless_modified`, this folder +uses extra data, i.e., http://www.openslr.org/62/, during training. + +You can use the following command to start the training: + +```bash +cd egs/aishell/ASR +./prepare.sh --stop-stage 6 +./prepare_aidatatang_200zh.sh + +export CUDA_VISIBLE_DEVICES="0,1,2" + +./transducer_stateless_modified-2/train.py \ + --world-size 3 \ + --num-epochs 90 \ + --start-epoch 0 \ + --exp-dir transducer_stateless_modified-2/exp-2 \ + --max-duration 250 \ + --lr-factor 2.0 \ + --context-size 2 \ + --modified-transducer-prob 0.25 \ + --datatang-prob 0.2 +``` + +To decode, you can use + +```bash +for epoch in 89; do + for avg in 30 38; do + ./transducer_stateless_modified-2/decode.py \ + --epoch $epoch \ + --avg $avg \ + --exp-dir transducer_stateless_modified-2/exp-2 \ + --max-duration 100 \ + --context-size 2 \ + --decoding-method greedy_search \ + --max-sym-per-frame 1 + done +done + +for epoch in 89; do + for avg in 38; do + ./transducer_stateless_modified-2/decode.py \ + --epoch $epoch \ + --avg $avg \ + --exp-dir transducer_stateless_modified-2/exp-2 \ + --max-duration 100 \ + --context-size 2 \ + --decoding-method modified_beam_search \ + --beam-size 4 + done +done +``` diff --git a/egs/aishell/ASR/transducer_stateless_modified-2/__init__.py b/egs/aishell/ASR/transducer_stateless_modified-2/__init__.py new file mode 100644 index 000000000..e69de29bb diff --git a/egs/aishell/ASR/transducer_stateless_modified-2/aidatatang_200zh.py b/egs/aishell/ASR/transducer_stateless_modified-2/aidatatang_200zh.py new file mode 100644 index 000000000..84ca64c89 --- /dev/null +++ b/egs/aishell/ASR/transducer_stateless_modified-2/aidatatang_200zh.py @@ -0,0 +1,53 @@ +# Copyright 2021 Piotr Żelasko +# 2022 Xiaomi Corp. (authors: Fangjun Kuang) +# +# See ../../../../LICENSE for clarification regarding multiple authors +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import logging +from pathlib import Path + +from lhotse import CutSet, load_manifest + + +class AIDatatang200zh: + def __init__(self, manifest_dir: str): + """ + Args: + manifest_dir: + It is expected to contain the following files:: + + - cuts_dev_raw.jsonl.gz + - cuts_train_raw.jsonl.gz + - cuts_test_raw.jsonl.gz + """ + self.manifest_dir = Path(manifest_dir) + + def train_cuts(self) -> CutSet: + f = self.manifest_dir / "cuts_train_raw.jsonl.gz" + logging.info(f"About to get train cuts from {f}") + cuts_train = load_manifest(f) + return cuts_train + + def valid_cuts(self) -> CutSet: + f = self.manifest_dir / "cuts_valid_raw.jsonl.gz" + logging.info(f"About to get valid cuts from {f}") + cuts_valid = load_manifest(f) + return cuts_valid + + def test_cuts(self) -> CutSet: + f = self.manifest_dir / "cuts_test_raw.jsonl.gz" + logging.info(f"About to get test cuts from {f}") + cuts_test = load_manifest(f) + return cuts_test diff --git a/egs/aishell/ASR/transducer_stateless_modified-2/aishell.py b/egs/aishell/ASR/transducer_stateless_modified-2/aishell.py new file mode 100644 index 000000000..94d1da066 --- /dev/null +++ b/egs/aishell/ASR/transducer_stateless_modified-2/aishell.py @@ -0,0 +1,53 @@ +# Copyright 2021 Piotr Żelasko +# 2022 Xiaomi Corp. (authors: Fangjun Kuang) +# +# See ../../../../LICENSE for clarification regarding multiple authors +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import logging +from pathlib import Path + +from lhotse import CutSet, load_manifest + + +class AIShell: + def __init__(self, manifest_dir: str): + """ + Args: + manifest_dir: + It is expected to contain the following files:: + + - cuts_dev.json.gz + - cuts_train.json.gz + - cuts_test.json.gz + """ + self.manifest_dir = Path(manifest_dir) + + def train_cuts(self) -> CutSet: + f = self.manifest_dir / "cuts_train.json.gz" + logging.info(f"About to get train cuts from {f}") + cuts_train = load_manifest(f) + return cuts_train + + def valid_cuts(self) -> CutSet: + f = self.manifest_dir / "cuts_dev.json.gz" + logging.info(f"About to get valid cuts from {f}") + cuts_valid = load_manifest(f) + return cuts_valid + + def test_cuts(self) -> CutSet: + f = self.manifest_dir / "cuts_test.json.gz" + logging.info(f"About to get test cuts from {f}") + cuts_test = load_manifest(f) + return cuts_test diff --git a/egs/aishell/ASR/transducer_stateless_modified-2/asr_datamodule.py b/egs/aishell/ASR/transducer_stateless_modified-2/asr_datamodule.py new file mode 100644 index 000000000..fe0d0a872 --- /dev/null +++ b/egs/aishell/ASR/transducer_stateless_modified-2/asr_datamodule.py @@ -0,0 +1,304 @@ +# Copyright 2021 Piotr Żelasko +# 2022 Xiaomi Corp. (authors: Fangjun Kuang) +# +# See ../../../../LICENSE for clarification regarding multiple authors +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import argparse +import logging +from pathlib import Path +from typing import Optional + +from lhotse import CutSet, Fbank, FbankConfig +from lhotse.dataset import ( + BucketingSampler, + CutMix, + DynamicBucketingSampler, + K2SpeechRecognitionDataset, + SpecAugment, +) +from lhotse.dataset.input_strategies import ( + OnTheFlyFeatures, + PrecomputedFeatures, +) +from torch.utils.data import DataLoader + +from icefall.utils import str2bool + + +class AsrDataModule: + 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( + "--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 BucketingSampler " + "and DynamicBucketingSampler." + "(you might want to increase it for larger datasets).", + ) + + group.add_argument( + "--shuffle", + type=str2bool, + default=True, + help="When enabled (=default), the examples will be " + "shuffled for each epoch.", + ) + + 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. ", + ) + + group.add_argument( + "--manifest-dir", + type=Path, + default=Path("data/fbank"), + help="Path to directory with train/valid/test cuts.", + ) + + 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. Used only in dev/test CutSet", + ) + + def train_dataloaders( + self, + cuts_train: CutSet, + dynamic_bucketing: bool, + on_the_fly_feats: bool, + cuts_musan: Optional[CutSet] = None, + ) -> DataLoader: + """ + Args: + cuts_train: + Cuts for training. + cuts_musan: + If not None, it is the cuts for mixing. + dynamic_bucketing: + True to use DynamicBucketingSampler; + False to use BucketingSampler. + on_the_fly_feats: + True to use OnTheFlyFeatures; + False to use PrecomputedFeatures. + """ + transforms = [] + if cuts_musan is not None: + logging.info("Enable MUSAN") + transforms.append( + CutMix( + cuts=cuts_musan, prob=0.5, snr=(10, 20), preserve_id=True + ) + ) + else: + logging.info("Disable MUSAN") + + 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}" + ) + input_transforms.append( + SpecAugment( + time_warp_factor=self.args.spec_aug_time_warp_factor, + num_frame_masks=2, + 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, + ) + + # 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))) + if on_the_fly_feats + else PrecomputedFeatures() + ), + input_transforms=input_transforms, + return_cuts=self.args.return_cuts, + ) + + if dynamic_bucketing: + 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, + drop_last=True, + ) + else: + logging.info("Using BucketingSampler.") + train_sampler = BucketingSampler( + cuts_train, + max_duration=self.args.max_duration, + shuffle=self.args.shuffle, + num_buckets=self.args.num_buckets, + bucket_method="equal_duration", + drop_last=True, + ) + + logging.info("About to create train dataloader") + 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 = [] + + 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 = BucketingSampler( + 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.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(), + return_cuts=self.args.return_cuts, + ) + sampler = BucketingSampler( + 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/aishell/ASR/transducer_stateless_modified-2/beam_search.py b/egs/aishell/ASR/transducer_stateless_modified-2/beam_search.py new file mode 120000 index 000000000..e188617a8 --- /dev/null +++ b/egs/aishell/ASR/transducer_stateless_modified-2/beam_search.py @@ -0,0 +1 @@ +../../../librispeech/ASR/transducer_stateless/beam_search.py \ No newline at end of file diff --git a/egs/aishell/ASR/transducer_stateless_modified-2/conformer.py b/egs/aishell/ASR/transducer_stateless_modified-2/conformer.py new file mode 120000 index 000000000..88975988f --- /dev/null +++ b/egs/aishell/ASR/transducer_stateless_modified-2/conformer.py @@ -0,0 +1 @@ +../transducer_stateless_modified/conformer.py \ No newline at end of file diff --git a/egs/aishell/ASR/transducer_stateless_modified-2/decode.py b/egs/aishell/ASR/transducer_stateless_modified-2/decode.py new file mode 100755 index 000000000..8b851bd17 --- /dev/null +++ b/egs/aishell/ASR/transducer_stateless_modified-2/decode.py @@ -0,0 +1,491 @@ +#!/usr/bin/env python3 +# +# Copyright 2021 Xiaomi Corporation (Author: Fangjun Kuang) +# +# See ../../../../LICENSE for clarification regarding multiple authors +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +""" +Usage: +(1) greedy search +./transducer_stateless_modified-2/decode.py \ + --epoch 89 \ + --avg 38 \ + --exp-dir ./transducer_stateless_modified-2/exp \ + --max-duration 100 \ + --decoding-method greedy_search + +(2) beam search +./transducer_stateless_modified/decode.py \ + --epoch 89 \ + --avg 38 \ + --exp-dir ./transducer_stateless_modified-2/exp \ + --max-duration 100 \ + --decoding-method beam_search \ + --beam-size 4 + +(3) modified beam search +./transducer_stateless_modified-2/decode.py \ + --epoch 89 \ + --avg 38 \ + --exp-dir ./transducer_stateless_modified/exp \ + --max-duration 100 \ + --decoding-method modified_beam_search \ + --beam-size 4 +""" + +import argparse +import logging +from collections import defaultdict +from pathlib import Path +from typing import Dict, List, Tuple + +import torch +import torch.nn as nn +from aishell import AIShell +from asr_datamodule import AsrDataModule +from beam_search import beam_search, greedy_search, modified_beam_search +from conformer import Conformer +from decoder import Decoder +from joiner import Joiner +from model import Transducer + +from icefall.checkpoint import average_checkpoints, load_checkpoint +from icefall.env import get_env_info +from icefall.lexicon import Lexicon +from icefall.utils import ( + AttributeDict, + setup_logger, + store_transcripts, + write_error_stats, +) + + +def get_parser(): + parser = argparse.ArgumentParser( + formatter_class=argparse.ArgumentDefaultsHelpFormatter + ) + + parser.add_argument( + "--epoch", + type=int, + default=30, + help="It specifies the checkpoint to use for decoding." + "Note: Epoch counts from 0.", + ) + parser.add_argument( + "--avg", + type=int, + default=10, + help="Number of checkpoints to average. Automatically select " + "consecutive checkpoints before the checkpoint specified by " + "'--epoch'. ", + ) + + parser.add_argument( + "--exp-dir", + type=str, + default="transducer_stateless_modified-2/exp", + help="The experiment dir", + ) + + parser.add_argument( + "--lang-dir", + type=str, + default="data/lang_char", + help="The lang dir", + ) + + parser.add_argument( + "--decoding-method", + type=str, + default="greedy_search", + help="""Possible values are: + - greedy_search + - beam_search + - modified_beam_search + """, + ) + + parser.add_argument( + "--beam-size", + type=int, + default=4, + help="Used only when --decoding-method is beam_search " + "and modified_beam_search", + ) + + parser.add_argument( + "--context-size", + type=int, + default=2, + help="The context size in the decoder. 1 means bigram; " + "2 means tri-gram", + ) + parser.add_argument( + "--max-sym-per-frame", + type=int, + default=3, + help="Maximum number of symbols per frame", + ) + return parser + + +def get_params() -> AttributeDict: + params = AttributeDict( + { + # parameters for conformer + "feature_dim": 80, + "encoder_out_dim": 512, + "subsampling_factor": 4, + "attention_dim": 512, + "nhead": 8, + "dim_feedforward": 2048, + "num_encoder_layers": 12, + "vgg_frontend": False, + "env_info": get_env_info(), + } + ) + return params + + +def get_encoder_model(params: AttributeDict): + # TODO: We can add an option to switch between Conformer and Transformer + encoder = Conformer( + num_features=params.feature_dim, + output_dim=params.encoder_out_dim, + subsampling_factor=params.subsampling_factor, + d_model=params.attention_dim, + nhead=params.nhead, + dim_feedforward=params.dim_feedforward, + num_encoder_layers=params.num_encoder_layers, + vgg_frontend=params.vgg_frontend, + ) + return encoder + + +def get_decoder_model(params: AttributeDict): + decoder = Decoder( + vocab_size=params.vocab_size, + embedding_dim=params.encoder_out_dim, + blank_id=params.blank_id, + context_size=params.context_size, + ) + return decoder + + +def get_joiner_model(params: AttributeDict): + joiner = Joiner( + input_dim=params.encoder_out_dim, + output_dim=params.vocab_size, + ) + return joiner + + +def get_transducer_model(params: AttributeDict): + encoder = get_encoder_model(params) + decoder = get_decoder_model(params) + joiner = get_joiner_model(params) + + model = Transducer( + encoder=encoder, + decoder=decoder, + joiner=joiner, + ) + return model + + +def decode_one_batch( + params: AttributeDict, + model: nn.Module, + lexicon: Lexicon, + 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 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`. + lexicon: + It contains the token symbol table and the word symbol table. + Returns: + Return the decoding result. See above description for the format of + the returned dict. + """ + device = model.device + feature = batch["inputs"] + assert feature.ndim == 3 + + feature = feature.to(device) + # at entry, feature is (N, T, C) + + supervisions = batch["supervisions"] + feature_lens = supervisions["num_frames"].to(device) + + encoder_out, encoder_out_lens = model.encoder( + x=feature, x_lens=feature_lens + ) + hyps = [] + 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 + ) + elif params.decoding_method == "modified_beam_search": + hyp = modified_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([lexicon.token_table[i] for i in hyp]) + + if params.decoding_method == "greedy_search": + return {"greedy_search": hyps} + else: + return {f"beam_{params.beam_size}": hyps} + + +def decode_dataset( + dl: torch.utils.data.DataLoader, + params: AttributeDict, + model: nn.Module, + lexicon: Lexicon, +) -> 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. + Returns: + Return a dict, whose key may be "greedy_search" if greedy search + is used, or it may be "beam_7" if beam size of 7 is used. + Its value is a list of tuples. Each tuple contains two elements: + The first is the reference transcript, and the second is the + predicted result. + """ + num_cuts = 0 + + try: + num_batches = len(dl) + except TypeError: + num_batches = "?" + + if params.decoding_method == "greedy_search": + log_interval = 100 + else: + log_interval = 2 + + results = defaultdict(list) + for batch_idx, batch in enumerate(dl): + texts = batch["supervisions"]["text"] + + hyps_dict = decode_one_batch( + params=params, + model=model, + lexicon=lexicon, + batch=batch, + ) + + for name, hyps in hyps_dict.items(): + this_batch = [] + assert len(hyps) == len(texts) + for hyp_words, ref_text in zip(hyps, texts): + ref_words = ref_text.split() + this_batch.append((ref_words, hyp_words)) + + results[name].extend(this_batch) + + num_cuts += len(texts) + + if batch_idx % log_interval == 0: + batch_str = f"{batch_idx}/{num_batches}" + + logging.info( + f"batch {batch_str}, cuts processed until now is {num_cuts}" + ) + return results + + +def save_results( + params: AttributeDict, + test_set_name: str, + results_dict: Dict[str, List[Tuple[List[int], List[int]]]], +): + test_set_wers = dict() + for key, results in results_dict.items(): + recog_path = ( + params.res_dir / f"recogs-{test_set_name}-{key}-{params.suffix}.txt" + ) + store_transcripts(filename=recog_path, texts=results) + + # 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" + ) + # we compute CER for aishell dataset. + results_char = [] + for res in results: + results_char.append((list("".join(res[0])), list("".join(res[1])))) + with open(errs_filename, "w") as f: + wer = write_error_stats( + f, f"{test_set_name}-{key}", results_char, 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\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) + args.lang_dir = Path(args.lang_dir) + + params = get_params() + params.update(vars(args)) + + assert params.decoding_method in ( + "greedy_search", + "beam_search", + "modified_beam_search", + ) + params.res_dir = params.exp_dir / params.decoding_method + + params.suffix = f"epoch-{params.epoch}-avg-{params.avg}" + if "beam_search" in params.decoding_method: + params.suffix += f"-beam-{params.beam_size}" + else: + params.suffix += f"-context-{params.context_size}" + params.suffix += f"-max-sym-per-frame-{params.max_sym_per_frame}" + + 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 = 0 + params.vocab_size = max(lexicon.tokens) + 1 + + logging.info(params) + + logging.info("About to create model") + model = get_transducer_model(params) + + if 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), strict=False + ) + + model.to(device) + model.eval() + model.device = device + + num_param = sum([p.numel() for p in model.parameters()]) + logging.info(f"Number of model parameters: {num_param}") + + asr_datamodule = AsrDataModule(args) + aishell = AIShell(manifest_dir=args.manifest_dir) + test_cuts = aishell.test_cuts() + test_dl = asr_datamodule.test_dataloaders(test_cuts) + + test_sets = ["test"] + test_dls = [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, + ) + + 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/aishell/ASR/transducer_stateless_modified-2/decoder.py b/egs/aishell/ASR/transducer_stateless_modified-2/decoder.py new file mode 120000 index 000000000..bdfcea5c2 --- /dev/null +++ b/egs/aishell/ASR/transducer_stateless_modified-2/decoder.py @@ -0,0 +1 @@ +../transducer_stateless_modified/decoder.py \ No newline at end of file diff --git a/egs/aishell/ASR/transducer_stateless_modified-2/encoder_interface.py b/egs/aishell/ASR/transducer_stateless_modified-2/encoder_interface.py new file mode 120000 index 000000000..a2a5f22cf --- /dev/null +++ b/egs/aishell/ASR/transducer_stateless_modified-2/encoder_interface.py @@ -0,0 +1 @@ +../transducer_stateless_modified/encoder_interface.py \ No newline at end of file diff --git a/egs/aishell/ASR/transducer_stateless_modified-2/export.py b/egs/aishell/ASR/transducer_stateless_modified-2/export.py new file mode 100755 index 000000000..d009de603 --- /dev/null +++ b/egs/aishell/ASR/transducer_stateless_modified-2/export.py @@ -0,0 +1,246 @@ +#!/usr/bin/env python3 +# +# Copyright 2021 Xiaomi Corporation (Author: Fangjun Kuang) +# +# See ../../../../LICENSE for clarification regarding multiple authors +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +# This script converts several saved checkpoints +# to a single one using model averaging. +""" +Usage: +./transducer_stateless_modified-2/export.py \ + --exp-dir ./transducer_stateless_modified-2/exp \ + --epoch 89 \ + --avg 38 + +It will generate a file exp_dir/pretrained.pt + +To use the generated file with `transducer_stateless_modified-2/decode.py`, +you can do:: + + cd /path/to/exp_dir + ln -s pretrained.pt epoch-9999.pt + + cd /path/to/egs/aishell/ASR + ./transducer_stateless_modified-2/decode.py \ + --exp-dir ./transducer_stateless_modified-2/exp \ + --epoch 9999 \ + --avg 1 \ + --max-duration 100 \ + --lang-dir data/lang_char +""" + +import argparse +import logging +from pathlib import Path + +import torch +import torch.nn as nn +from conformer import Conformer +from decoder import Decoder +from joiner import Joiner +from model import Transducer + +from icefall.checkpoint import average_checkpoints, load_checkpoint +from icefall.env import get_env_info +from icefall.lexicon import Lexicon +from icefall.utils import AttributeDict, str2bool + + +def get_parser(): + parser = argparse.ArgumentParser( + formatter_class=argparse.ArgumentDefaultsHelpFormatter + ) + + parser.add_argument( + "--epoch", + type=int, + default=20, + help="It specifies the checkpoint to use for decoding." + "Note: Epoch counts from 0.", + ) + + parser.add_argument( + "--avg", + type=int, + default=10, + help="Number of checkpoints to average. Automatically select " + "consecutive checkpoints before the checkpoint specified by " + "'--epoch'. ", + ) + + parser.add_argument( + "--exp-dir", + type=Path, + default=Path("transducer_stateless_modified-2/exp"), + help="""It specifies the directory where all training related + files, e.g., checkpoints, log, etc, are saved + """, + ) + + parser.add_argument( + "--jit", + type=str2bool, + default=False, + help="""True to save a model after applying torch.jit.script. + """, + ) + + parser.add_argument( + "--lang-dir", + type=Path, + default=Path("data/lang_char"), + help="The lang dir", + ) + + parser.add_argument( + "--context-size", + type=int, + default=2, + help="The context size in the decoder. 1 means bigram; " + "2 means tri-gram", + ) + + return parser + + +def get_params() -> AttributeDict: + params = AttributeDict( + { + # parameters for conformer + "feature_dim": 80, + "encoder_out_dim": 512, + "subsampling_factor": 4, + "attention_dim": 512, + "nhead": 8, + "dim_feedforward": 2048, + "num_encoder_layers": 12, + "vgg_frontend": False, + "env_info": get_env_info(), + } + ) + return params + + +def get_encoder_model(params: AttributeDict) -> nn.Module: + encoder = Conformer( + num_features=params.feature_dim, + output_dim=params.encoder_out_dim, + subsampling_factor=params.subsampling_factor, + d_model=params.attention_dim, + nhead=params.nhead, + dim_feedforward=params.dim_feedforward, + num_encoder_layers=params.num_encoder_layers, + vgg_frontend=params.vgg_frontend, + ) + return encoder + + +def get_decoder_model(params: AttributeDict) -> nn.Module: + decoder = Decoder( + vocab_size=params.vocab_size, + embedding_dim=params.encoder_out_dim, + blank_id=params.blank_id, + context_size=params.context_size, + ) + return decoder + + +def get_joiner_model(params: AttributeDict) -> nn.Module: + joiner = Joiner( + input_dim=params.encoder_out_dim, + output_dim=params.vocab_size, + ) + return joiner + + +def get_transducer_model(params: AttributeDict) -> nn.Module: + encoder = get_encoder_model(params) + decoder = get_decoder_model(params) + joiner = get_joiner_model(params) + + model = Transducer( + encoder=encoder, + decoder=decoder, + joiner=joiner, + ) + return model + + +def main(): + args = get_parser().parse_args() + + assert args.jit is False, "torchscript support will be added later" + + 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}") + + lexicon = Lexicon(params.lang_dir) + + params.blank_id = 0 + params.vocab_size = max(lexicon.tokens) + 1 + + logging.info(params) + + logging.info("About to create model") + model = get_transducer_model(params) + + model.to(device) + + if 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), strict=False + ) + + model.to("cpu") + model.eval() + + if params.jit: + logging.info("Using torch.jit.script") + model = torch.jit.script(model) + filename = params.exp_dir / "cpu_jit.pt" + model.save(str(filename)) + logging.info(f"Saved to {filename}") + else: + logging.info("Not using torch.jit.script") + # Save it using a format so that it can be loaded + # by :func:`load_checkpoint` + filename = params.exp_dir / "pretrained.pt" + torch.save({"model": model.state_dict()}, str(filename)) + logging.info(f"Saved to {filename}") + + +if __name__ == "__main__": + formatter = ( + "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s" + ) + + logging.basicConfig(format=formatter, level=logging.INFO) + main() diff --git a/egs/aishell/ASR/transducer_stateless_modified-2/joiner.py b/egs/aishell/ASR/transducer_stateless_modified-2/joiner.py new file mode 120000 index 000000000..e9e435ecd --- /dev/null +++ b/egs/aishell/ASR/transducer_stateless_modified-2/joiner.py @@ -0,0 +1 @@ +../transducer_stateless_modified/joiner.py \ No newline at end of file diff --git a/egs/aishell/ASR/transducer_stateless_modified-2/model.py b/egs/aishell/ASR/transducer_stateless_modified-2/model.py new file mode 100644 index 000000000..086957d0b --- /dev/null +++ b/egs/aishell/ASR/transducer_stateless_modified-2/model.py @@ -0,0 +1,163 @@ +# 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. + +import random +from typing import Optional + +import k2 +import torch +import torch.nn as nn +from encoder_interface import EncoderInterface + +from icefall.utils import add_sos + + +class Transducer(nn.Module): + """It implements https://arxiv.org/pdf/1211.3711.pdf + "Sequence Transduction with Recurrent Neural Networks" + """ + + def __init__( + self, + encoder: EncoderInterface, + decoder: nn.Module, + joiner: nn.Module, + decoder_datatang: Optional[nn.Module] = None, + joiner_datatang: Optional[nn.Module] = None, + ): + """ + Args: + encoder: + It is the transcription network in the paper. Its accepts + two inputs: `x` of (N, T, C) and `x_lens` of shape (N,). + It returns two tensors: `logits` of shape (N, T, C) and + `logit_lens` of shape (N,). + decoder: + It is the prediction network in the paper. Its input shape + is (N, U) and its output shape is (N, U, C). It should contain + one attribute: `blank_id`. + joiner: + It has two inputs with shapes: (N, T, C) and (N, U, C). Its + output shape is (N, T, U, C). Note that its output contains + unnormalized probs, i.e., not processed by log-softmax. + decoder_datatang: + The decoder for the aidatatang_200zh dataset. + joiner_datatang: + The joiner for the aidatatang_200zh dataset. + """ + super().__init__() + assert isinstance(encoder, EncoderInterface), type(encoder) + assert hasattr(decoder, "blank_id") + if decoder_datatang is not None: + assert hasattr(decoder_datatang, "blank_id") + + self.encoder = encoder + self.decoder = decoder + self.joiner = joiner + + self.decoder_datatang = decoder_datatang + self.joiner_datatang = joiner_datatang + + def forward( + self, + x: torch.Tensor, + x_lens: torch.Tensor, + y: k2.RaggedTensor, + aishell: bool = True, + modified_transducer_prob: float = 0.0, + ) -> torch.Tensor: + """ + Args: + x: + A 3-D tensor of shape (N, T, C). + x_lens: + A 1-D tensor of shape (N,). It contains the number of frames in `x` + before padding. + y: + A ragged tensor with 2 axes [utt][label]. It contains labels of each + utterance. + modified_transducer_prob: + The probability to use modified transducer loss. + Returns: + Return the transducer loss. + """ + assert x.ndim == 3, x.shape + assert x_lens.ndim == 1, x_lens.shape + assert y.num_axes == 2, y.num_axes + + assert x.size(0) == x_lens.size(0) == y.dim0 + + encoder_out, x_lens = self.encoder(x, x_lens) + assert torch.all(x_lens > 0) + + # Now for the decoder, i.e., the prediction network + row_splits = y.shape.row_splits(1) + y_lens = row_splits[1:] - row_splits[:-1] + + blank_id = self.decoder.blank_id + sos_y = add_sos(y, sos_id=blank_id) + + sos_y_padded = sos_y.pad(mode="constant", padding_value=blank_id) + sos_y_padded = sos_y_padded.to(torch.int64) + + if aishell: + decoder = self.decoder + joiner = self.joiner + else: + decoder = self.decoder_datatang + joiner = self.joiner_datatang + + decoder_out = decoder(sos_y_padded) + + # +1 here since a blank is prepended to each utterance. + logits = joiner( + encoder_out=encoder_out, + decoder_out=decoder_out, + encoder_out_len=x_lens, + decoder_out_len=y_lens + 1, + ) + + # rnnt_loss requires 0 padded targets + # Note: y does not start with SOS + y_padded = y.pad(mode="constant", padding_value=0) + + # We don't put this `import` at the beginning of the file + # as it is required only in the training, not during the + # reference stage + import optimized_transducer + + assert 0 <= modified_transducer_prob <= 1 + + if modified_transducer_prob == 0: + one_sym_per_frame = False + elif random.random() < modified_transducer_prob: + # random.random() returns a float in the range [0, 1) + one_sym_per_frame = True + else: + one_sym_per_frame = False + + loss = optimized_transducer.transducer_loss( + logits=logits, + targets=y_padded, + logit_lengths=x_lens, + target_lengths=y_lens, + blank=blank_id, + reduction="sum", + one_sym_per_frame=one_sym_per_frame, + from_log_softmax=False, + ) + + return loss diff --git a/egs/aishell/ASR/transducer_stateless_modified-2/pretrained.py b/egs/aishell/ASR/transducer_stateless_modified-2/pretrained.py new file mode 100755 index 000000000..31bab122c --- /dev/null +++ b/egs/aishell/ASR/transducer_stateless_modified-2/pretrained.py @@ -0,0 +1,331 @@ +#!/usr/bin/env python3 +# Copyright 2021 Xiaomi Corp. (authors: Fangjun Kuang, +# Wei Kang) +# +# See ../../../../LICENSE for clarification regarding multiple authors +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +""" +Usage: + +# greedy search +./transducer_stateless_modified-2/pretrained.py \ + --checkpoint /path/to/pretrained.pt \ + --lang-dir /path/to/lang_char \ + --method greedy_search \ + /path/to/foo.wav \ + /path/to/bar.wav + +# beam search +./transducer_stateless_modified-2/pretrained.py \ + --checkpoint /path/to/pretrained.pt \ + --lang-dir /path/to/lang_char \ + --method beam_search \ + --beam-size 4 \ + /path/to/foo.wav \ + /path/to/bar.wav + +# modified beam search +./transducer_stateless_modified-2/pretrained.py \ + --checkpoint /path/to/pretrained.pt \ + --lang-dir /path/to/lang_char \ + --method modified_beam_search \ + --beam-size 4 \ + /path/to/foo.wav \ + /path/to/bar.wav + +""" + +import argparse +import logging +import math +from pathlib import Path +from typing import List + +import kaldifeat +import torch +import torch.nn as nn +import torchaudio +from beam_search import beam_search, greedy_search, modified_beam_search +from conformer import Conformer +from decoder import Decoder +from joiner import Joiner +from model import Transducer +from torch.nn.utils.rnn import pad_sequence + +from icefall.env import get_env_info +from icefall.lexicon import Lexicon +from icefall.utils import AttributeDict + + +def get_parser(): + parser = argparse.ArgumentParser( + formatter_class=argparse.ArgumentDefaultsHelpFormatter + ) + + parser.add_argument( + "--checkpoint", + type=str, + required=True, + help="Path to the checkpoint. " + "The checkpoint is assumed to be saved by " + "icefall.checkpoint.save_checkpoint().", + ) + + parser.add_argument( + "--lang-dir", + type=Path, + default=Path("data/lang_char"), + help="The lang dir", + ) + + parser.add_argument( + "--method", + type=str, + default="greedy_search", + help="""Possible values are: + - greedy_search + - beam_search + - modified_beam_search + """, + ) + + parser.add_argument( + "sound_files", + type=str, + nargs="+", + help="The input sound file(s) to transcribe. " + "Supported formats are those supported by torchaudio.load(). " + "For example, wav and flac are supported. " + "The sample rate has to be 16kHz.", + ) + + parser.add_argument( + "--beam-size", + type=int, + default=4, + help="Used only when --method is beam_search and modified_beam_search", + ) + + parser.add_argument( + "--context-size", + type=int, + default=2, + help="The context size in the decoder. 1 means bigram; " + "2 means tri-gram", + ) + parser.add_argument( + "--max-sym-per-frame", + type=int, + default=3, + help="Maximum number of symbols per frame. " + "Use only when --method is greedy_search", + ) + return parser + + return parser + + +def get_params() -> AttributeDict: + params = AttributeDict( + { + # parameters for conformer + "feature_dim": 80, + "encoder_out_dim": 512, + "subsampling_factor": 4, + "attention_dim": 512, + "nhead": 8, + "dim_feedforward": 2048, + "num_encoder_layers": 12, + "vgg_frontend": False, + "env_info": get_env_info(), + "sample_rate": 16000, + } + ) + return params + + +def get_encoder_model(params: AttributeDict) -> nn.Module: + encoder = Conformer( + num_features=params.feature_dim, + output_dim=params.encoder_out_dim, + subsampling_factor=params.subsampling_factor, + d_model=params.attention_dim, + nhead=params.nhead, + dim_feedforward=params.dim_feedforward, + num_encoder_layers=params.num_encoder_layers, + vgg_frontend=params.vgg_frontend, + ) + return encoder + + +def get_decoder_model(params: AttributeDict) -> nn.Module: + decoder = Decoder( + vocab_size=params.vocab_size, + embedding_dim=params.encoder_out_dim, + blank_id=params.blank_id, + context_size=params.context_size, + ) + return decoder + + +def get_joiner_model(params: AttributeDict) -> nn.Module: + joiner = Joiner( + input_dim=params.encoder_out_dim, + output_dim=params.vocab_size, + ) + return joiner + + +def get_transducer_model(params: AttributeDict) -> nn.Module: + encoder = get_encoder_model(params) + decoder = get_decoder_model(params) + joiner = get_joiner_model(params) + + model = Transducer( + encoder=encoder, + decoder=decoder, + joiner=joiner, + ) + return model + + +def read_sound_files( + filenames: List[str], expected_sample_rate: float +) -> List[torch.Tensor]: + """Read a list of sound files into a list 1-D float32 torch tensors. + Args: + filenames: + A list of sound filenames. + expected_sample_rate: + The expected sample rate of the sound files. + Returns: + Return a list of 1-D float32 torch tensors. + """ + ans = [] + for f in filenames: + wave, sample_rate = torchaudio.load(f) + assert sample_rate == expected_sample_rate, ( + f"expected sample rate: {expected_sample_rate}. " + f"Given: {sample_rate}" + ) + # We use only the first channel + ans.append(wave[0]) + return ans + + +def main(): + parser = get_parser() + args = parser.parse_args() + + 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}") + + lexicon = Lexicon(params.lang_dir) + + params.blank_id = 0 + params.vocab_size = max(lexicon.tokens) + 1 + + logging.info(params) + + logging.info("About to create model") + model = get_transducer_model(params) + + checkpoint = torch.load(args.checkpoint, map_location="cpu") + model.load_state_dict(checkpoint["model"]) + model.to(device) + model.eval() + model.device = device + + logging.info("Constructing Fbank computer") + opts = kaldifeat.FbankOptions() + opts.device = device + opts.frame_opts.dither = 0 + opts.frame_opts.snip_edges = False + opts.frame_opts.samp_freq = params.sample_rate + opts.mel_opts.num_bins = params.feature_dim + + fbank = kaldifeat.Fbank(opts) + + logging.info(f"Reading sound files: {params.sound_files}") + waves = read_sound_files( + filenames=params.sound_files, expected_sample_rate=params.sample_rate + ) + waves = [w.to(device) for w in waves] + + logging.info("Decoding started") + features = fbank(waves) + feature_lens = [f.size(0) for f in features] + feature_lens = torch.tensor(feature_lens, device=device) + + features = pad_sequence( + features, batch_first=True, padding_value=math.log(1e-10) + ) + + hyps = [] + with torch.no_grad(): + encoder_out, encoder_out_lens = model.encoder( + x=features, x_lens=feature_lens + ) + + for i in range(encoder_out.size(0)): + # fmt: off + encoder_out_i = encoder_out[i:i+1, :encoder_out_lens[i]] + # fmt: on + if params.method == "greedy_search": + hyp = greedy_search( + model=model, + encoder_out=encoder_out_i, + max_sym_per_frame=params.max_sym_per_frame, + ) + elif params.method == "beam_search": + hyp = beam_search( + model=model, + encoder_out=encoder_out_i, + beam=params.beam_size, + ) + elif params.method == "modified_beam_search": + hyp = modified_beam_search( + model=model, + encoder_out=encoder_out_i, + beam=params.beam_size, + ) + else: + raise ValueError( + f"Unsupported decoding method: {params.method}" + ) + hyps.append([lexicon.token_table[i] for i in hyp]) + + s = "\n" + for filename, hyp in zip(params.sound_files, hyps): + words = " ".join(hyp) + s += f"{filename}:\n{words}\n\n" + logging.info(s) + + logging.info("Decoding Done") + + +if __name__ == "__main__": + formatter = ( + "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s" + ) + + logging.basicConfig(format=formatter, level=logging.INFO) + main() diff --git a/egs/aishell/ASR/transducer_stateless_modified-2/subsampling.py b/egs/aishell/ASR/transducer_stateless_modified-2/subsampling.py new file mode 120000 index 000000000..6fee09e58 --- /dev/null +++ b/egs/aishell/ASR/transducer_stateless_modified-2/subsampling.py @@ -0,0 +1 @@ +../conformer_ctc/subsampling.py \ No newline at end of file diff --git a/egs/aishell/ASR/transducer_stateless_modified-2/test_decoder.py b/egs/aishell/ASR/transducer_stateless_modified-2/test_decoder.py new file mode 120000 index 000000000..fbe1679ea --- /dev/null +++ b/egs/aishell/ASR/transducer_stateless_modified-2/test_decoder.py @@ -0,0 +1 @@ +../transducer_stateless_modified/test_decoder.py \ No newline at end of file diff --git a/egs/aishell/ASR/transducer_stateless_modified-2/train.py b/egs/aishell/ASR/transducer_stateless_modified-2/train.py new file mode 100755 index 000000000..53d4e455f --- /dev/null +++ b/egs/aishell/ASR/transducer_stateless_modified-2/train.py @@ -0,0 +1,875 @@ +#!/usr/bin/env python3 +# Copyright 2021 Xiaomi Corp. (authors: Fangjun Kuang, +# Wei Kang +# Mingshuang Luo) +# Copyright 2021 (Pingfeng 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. + +""" +Usage: +./prepare.sh +./prepare_aidatatang_200zh.sh + +export CUDA_VISIBLE_DEVICES="0,1,2" + +./transducer_stateless_modified-2/train.py \ + --world-size 3 \ + --num-epochs 90 \ + --start-epoch 0 \ + --exp-dir transducer_stateless_modified-2/exp-2 \ + --max-duration 250 \ + --lr-factor 2.0 \ + --context-size 2 \ + --modified-transducer-prob 0.25 \ + --datatang-prob 0.2 +""" + + +import argparse +import logging +import random +from pathlib import Path +from shutil import copyfile +from typing import Optional, Tuple + +import k2 +import torch +import torch.multiprocessing as mp +import torch.nn as nn +from aidatatang_200zh import AIDatatang200zh +from aishell import AIShell +from asr_datamodule import AsrDataModule +from conformer import Conformer +from decoder import Decoder +from joiner import Joiner +from lhotse import CutSet, load_manifest +from lhotse.cut import Cut +from lhotse.utils import fix_random_seed +from model import Transducer +from torch import Tensor +from torch.nn.parallel import DistributedDataParallel as DDP +from torch.nn.utils import clip_grad_norm_ +from torch.utils.tensorboard import SummaryWriter +from transformer import Noam + +from icefall.char_graph_compiler import CharCtcTrainingGraphCompiler +from icefall.checkpoint import load_checkpoint +from icefall.checkpoint import save_checkpoint as save_checkpoint_impl +from icefall.dist import cleanup_dist, setup_dist +from icefall.env import get_env_info +from icefall.lexicon import Lexicon +from icefall.utils import AttributeDict, MetricsTracker, setup_logger, str2bool + + +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=0, + help="""Resume training from from this epoch. + If it is positive, it will load checkpoint from + transducer_stateless/exp/epoch-{start_epoch-1}.pt + """, + ) + + parser.add_argument( + "--exp-dir", + type=str, + default="transducer_stateless_modified-2/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( + "--lr-factor", + type=float, + default=5.0, + help="The lr_factor for Noam optimizer", + ) + + 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( + "--modified-transducer-prob", + type=float, + default=0.25, + help="""The probability to use modified transducer loss. + In modified transduer, it limits the maximum number of symbols + per frame to 1. See also the option --max-sym-per-frame in + transducer_stateless/decode.py + """, + ) + + parser.add_argument( + "--datatang-prob", + type=float, + default=0.2, + help="The probability to select a batch from the " + "aidatatang_200zh dataset", + ) + + 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. + + - attention_dim: Hidden dim for multi-head attention model. + + - num_decoder_layers: Number of decoder layer of transformer decoder. + + - warm_step: The warm_step for Noam optimizer. + """ + 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": 800, # For the 100h subset, use 800 + # parameters for conformer + "feature_dim": 80, + "encoder_out_dim": 512, + "subsampling_factor": 4, + "attention_dim": 512, + "nhead": 8, + "dim_feedforward": 2048, + "num_encoder_layers": 12, + "vgg_frontend": False, + # parameters for Noam + "warm_step": 80000, # For the 100h subset, use 8k + "env_info": get_env_info(), + } + ) + + return params + + +def get_encoder_model(params: AttributeDict) -> nn.Module: + # TODO: We can add an option to switch between Conformer and Transformer + encoder = Conformer( + num_features=params.feature_dim, + output_dim=params.encoder_out_dim, + subsampling_factor=params.subsampling_factor, + d_model=params.attention_dim, + nhead=params.nhead, + dim_feedforward=params.dim_feedforward, + num_encoder_layers=params.num_encoder_layers, + vgg_frontend=params.vgg_frontend, + ) + return encoder + + +def get_decoder_model(params: AttributeDict) -> nn.Module: + decoder = Decoder( + vocab_size=params.vocab_size, + embedding_dim=params.encoder_out_dim, + blank_id=params.blank_id, + context_size=params.context_size, + ) + return decoder + + +def get_joiner_model(params: AttributeDict) -> nn.Module: + joiner = Joiner( + input_dim=params.encoder_out_dim, + output_dim=params.vocab_size, + ) + return joiner + + +def get_transducer_model(params: AttributeDict) -> nn.Module: + encoder = get_encoder_model(params) + decoder = get_decoder_model(params) + joiner = get_joiner_model(params) + + decoder_datatang = get_decoder_model(params) + joiner_datatang = get_joiner_model(params) + + model = Transducer( + encoder=encoder, + decoder=decoder, + joiner=joiner, + decoder_datatang=decoder_datatang, + joiner_datatang=joiner_datatang, + ) + return model + + +def load_checkpoint_if_available( + params: AttributeDict, + model: nn.Module, + optimizer: Optional[torch.optim.Optimizer] = None, + scheduler: Optional[torch.optim.lr_scheduler._LRScheduler] = None, +) -> None: + """Load checkpoint from file. + + If params.start_epoch is positive, it will load the checkpoint from + `params.start_epoch - 1`. Otherwise, this function does nothing. + + Apart from loading state dict for `model`, `optimizer` and `scheduler`, + 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. + optimizer: + The optimizer that we are using. + scheduler: + The learning rate scheduler we are using. + Returns: + Return None. + """ + if params.start_epoch <= 0: + return + + filename = params.exp_dir / f"epoch-{params.start_epoch-1}.pt" + saved_params = load_checkpoint( + filename, + model=model, + 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] + + return saved_params + + +def save_checkpoint( + params: AttributeDict, + model: nn.Module, + optimizer: Optional[torch.optim.Optimizer] = None, + scheduler: Optional[torch.optim.lr_scheduler._LRScheduler] = 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. + """ + if rank != 0: + return + filename = params.exp_dir / f"epoch-{params.cur_epoch}.pt" + save_checkpoint_impl( + filename=filename, + model=model, + params=params, + optimizer=optimizer, + scheduler=scheduler, + 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 is_aishell(c: Cut) -> bool: + """Return True if this cut is from the AIShell dataset. + + Note: + During data preparation, we set the custom field in + the supervision segment of aidatatang_200zh to + dict(origin='aidatatang_200zh') + See ../local/process_aidatatang_200zh.py. + """ + return c.supervisions[0].custom is None + + +def compute_loss( + params: AttributeDict, + model: nn.Module, + 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 Conformer 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. + """ + device = model.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) + + aishell = is_aishell(supervisions["cut"][0]) + + texts = batch["supervisions"]["text"] + y = graph_compiler.texts_to_ids(texts) + y = k2.RaggedTensor(y).to(device) + + with torch.set_grad_enabled(is_training): + loss = model( + x=feature, + x_lens=feature_lens, + y=y, + aishell=aishell, + modified_transducer_prob=params.modified_transducer_prob, + ) + + assert loss.requires_grad == is_training + + info = MetricsTracker() + 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, + model: nn.Module, + 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: nn.Module, + optimizer: torch.optim.Optimizer, + graph_compiler: CharCtcTrainingGraphCompiler, + train_dl: torch.utils.data.DataLoader, + datatang_train_dl: torch.utils.data.DataLoader, + valid_dl: torch.utils.data.DataLoader, + rng: random.Random, + tb_writer: Optional[SummaryWriter] = None, + world_size: int = 1, +) -> 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. + train_dl: + Dataloader for the training dataset. + datatang_train_dl: + Dataloader for the aidatatang_200zh training dataset. + valid_dl: + Dataloader for the validation dataset. + tb_writer: + Writer to write log messages to tensorboard. + world_size: + Number of nodes in DDP training. If it is 1, DDP is disabled. + """ + model.train() + + aishell_tot_loss = MetricsTracker() + datatang_tot_loss = MetricsTracker() + tot_loss = MetricsTracker() + + # index 0: for LibriSpeech + # index 1: for GigaSpeech + # This sets the probabilities for choosing which datasets + dl_weights = [1 - params.datatang_prob, params.datatang_prob] + + iter_aishell = iter(train_dl) + iter_datatang = iter(datatang_train_dl) + + batch_idx = 0 + + while True: + idx = rng.choices((0, 1), weights=dl_weights, k=1)[0] + dl = iter_aishell if idx == 0 else iter_datatang + + try: + batch = next(dl) + except StopIteration: + break + batch_idx += 1 + + params.batch_idx_train += 1 + batch_size = len(batch["supervisions"]["text"]) + + aishell = is_aishell(batch["supervisions"]["cut"][0]) + + 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 + if aishell: + aishell_tot_loss = ( + aishell_tot_loss * (1 - 1 / params.reset_interval) + ) + loss_info + prefix = "aishell" # for logging only + else: + datatang_tot_loss = ( + datatang_tot_loss * (1 - 1 / params.reset_interval) + ) + loss_info + prefix = "datatang" + + # NOTE: We use reduction==sum and loss is computed over utterances + # in the batch and there is no normalization to it so far. + + optimizer.zero_grad() + loss.backward() + clip_grad_norm_(model.parameters(), 5.0, 2.0) + optimizer.step() + + if batch_idx % params.log_interval == 0: + logging.info( + f"Epoch {params.cur_epoch}, " + f"batch {batch_idx}, {prefix}_loss[{loss_info}], " + f"tot_loss[{tot_loss}], batch size: {batch_size}, " + f"aishell_tot_loss[{aishell_tot_loss}], " + f"datatang_tot_loss[{datatang_tot_loss}], " + f"batch size: {batch_size}" + ) + + if batch_idx % params.log_interval == 0: + if tb_writer is not None: + loss_info.write_summary( + tb_writer, + f"train/current_{prefix}_", + params.batch_idx_train, + ) + tot_loss.write_summary( + tb_writer, "train/tot_", params.batch_idx_train + ) + aishell_tot_loss.write_summary( + tb_writer, "train/aishell_tot_", params.batch_idx_train + ) + datatang_tot_loss.write_summary( + tb_writer, "train/datatang_tot_", params.batch_idx_train + ) + + if batch_idx > 0 and batch_idx % params.valid_interval == 0: + 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}") + 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 filter_short_and_long_utterances(cuts: CutSet) -> CutSet: + def remove_short_and_long_utt(c: Cut): + # Keep only utterances with duration between 1 second and 12 seconds + return 1.0 <= c.duration <= 12.0 + + num_in_total = len(cuts) + cuts = cuts.filter(remove_short_and_long_utt) + + num_left = len(cuts) + num_removed = num_in_total - num_left + removed_percent = num_removed / num_in_total * 100 + + logging.info(f"Before removing short and long utterances: {num_in_total}") + logging.info(f"After removing short and long utterances: {num_left}") + logging.info(f"Removed {num_removed} utterances ({removed_percent:.5f}%)") + + return cuts + + +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)) + + seed = 42 + fix_random_seed(seed) + rng = random.Random(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, + oov="", + ) + + params.blank_id = 0 + params.vocab_size = max(lexicon.tokens) + 1 + + logging.info(params) + + logging.info("About to create model") + model = get_transducer_model(params) + + num_param = sum([p.numel() for p in model.parameters()]) + logging.info(f"Number of model parameters: {num_param}") + + checkpoints = load_checkpoint_if_available(params=params, model=model) + + model.to(device) + if world_size > 1: + logging.info("Using DDP") + model = DDP(model, device_ids=[rank], find_unused_parameters=True) + model.device = device + + optimizer = Noam( + model.parameters(), + model_size=params.attention_dim, + factor=params.lr_factor, + warm_step=params.warm_step, + ) + + if checkpoints and "optimizer" in checkpoints: + logging.info("Loading optimizer state dict") + optimizer.load_state_dict(checkpoints["optimizer"]) + + aishell = AIShell(manifest_dir=args.manifest_dir) + + train_cuts = aishell.train_cuts() + train_cuts = filter_short_and_long_utterances(train_cuts) + + datatang = AIDatatang200zh( + manifest_dir=f"{args.manifest_dir}/aidatatang_200zh" + ) + train_datatang_cuts = datatang.train_cuts() + train_datatang_cuts = filter_short_and_long_utterances(train_datatang_cuts) + + if args.enable_musan: + cuts_musan = load_manifest( + Path(args.manifest_dir) / "cuts_musan.json.gz" + ) + else: + cuts_musan = None + + asr_datamodule = AsrDataModule(args) + + train_dl = asr_datamodule.train_dataloaders( + train_cuts, + dynamic_bucketing=False, + on_the_fly_feats=False, + cuts_musan=cuts_musan, + ) + + datatang_train_dl = asr_datamodule.train_dataloaders( + train_datatang_cuts, + dynamic_bucketing=True, + on_the_fly_feats=True, + cuts_musan=cuts_musan, + ) + + valid_cuts = aishell.valid_cuts() + valid_dl = asr_datamodule.valid_dataloaders(valid_cuts) + + for dl in [train_dl, datatang_train_dl]: + scan_pessimistic_batches_for_oom( + model=model, + train_dl=dl, + optimizer=optimizer, + graph_compiler=graph_compiler, + params=params, + ) + + for epoch in range(params.start_epoch, params.num_epochs): + train_dl.sampler.set_epoch(epoch) + datatang_train_dl.sampler.set_epoch(epoch) + + cur_lr = optimizer._rate + if tb_writer is not None: + tb_writer.add_scalar( + "train/learning_rate", cur_lr, params.batch_idx_train + ) + tb_writer.add_scalar("train/epoch", epoch, params.batch_idx_train) + + if rank == 0: + logging.info("epoch {}, learning rate {}".format(epoch, cur_lr)) + + params.cur_epoch = epoch + + train_one_epoch( + params=params, + model=model, + optimizer=optimizer, + graph_compiler=graph_compiler, + train_dl=train_dl, + datatang_train_dl=datatang_train_dl, + valid_dl=valid_dl, + rng=rng, + tb_writer=tb_writer, + world_size=world_size, + ) + + save_checkpoint( + params=params, + model=model, + optimizer=optimizer, + rank=rank, + ) + + logging.info("Done!") + + if world_size > 1: + torch.distributed.barrier() + cleanup_dist() + + +def scan_pessimistic_batches_for_oom( + model: nn.Module, + 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 0 would cause OOM." + ) + batches, crit_values = find_pessimistic_batches(train_dl.sampler) + for criterion, cuts in batches.items(): + batch = train_dl.dataset[cuts] + try: + optimizer.zero_grad() + loss, _ = compute_loss( + params=params, + model=model, + graph_compiler=graph_compiler, + batch=batch, + is_training=True, + ) + loss.backward() + clip_grad_norm_(model.parameters(), 5.0, 2.0) + optimizer.step() + except RuntimeError 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]}) ..." + ) + raise + + +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) + + assert 0 <= args.datatang_prob < 1, args.datatang_prob + + 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/aishell/ASR/transducer_stateless_modified-2/transformer.py b/egs/aishell/ASR/transducer_stateless_modified-2/transformer.py new file mode 120000 index 000000000..4320d1105 --- /dev/null +++ b/egs/aishell/ASR/transducer_stateless_modified-2/transformer.py @@ -0,0 +1 @@ +../transducer_stateless_modified/transformer.py \ No newline at end of file diff --git a/egs/aishell/ASR/transducer_stateless_modified/README.md b/egs/aishell/ASR/transducer_stateless_modified/README.md new file mode 100644 index 000000000..9709eb9a0 --- /dev/null +++ b/egs/aishell/ASR/transducer_stateless_modified/README.md @@ -0,0 +1,21 @@ +## Introduction + +The decoder, i.e., the prediction network, is from +https://ieeexplore.ieee.org/stamp/stamp.jsp?arnumber=9054419 +(Rnn-Transducer with Stateless Prediction Network) + +You can use the following command to start the training: + +```bash +cd egs/aishell/ASR + +export CUDA_VISIBLE_DEVICES="0,1,2,3,4,5,6,7" + +./transducer_stateless_modified/train.py \ + --world-size 8 \ + --num-epochs 30 \ + --start-epoch 0 \ + --exp-dir transducer_stateless_modified/exp \ + --max-duration 250 \ + --lr-factor 2.5 +``` diff --git a/egs/aishell/ASR/transducer_stateless_modified/__init__.py b/egs/aishell/ASR/transducer_stateless_modified/__init__.py new file mode 100644 index 000000000..e69de29bb diff --git a/egs/aishell/ASR/transducer_stateless_modified/asr_datamodule.py b/egs/aishell/ASR/transducer_stateless_modified/asr_datamodule.py new file mode 120000 index 000000000..a73848de9 --- /dev/null +++ b/egs/aishell/ASR/transducer_stateless_modified/asr_datamodule.py @@ -0,0 +1 @@ +../conformer_ctc/asr_datamodule.py \ No newline at end of file diff --git a/egs/aishell/ASR/transducer_stateless_modified/beam_search.py b/egs/aishell/ASR/transducer_stateless_modified/beam_search.py new file mode 120000 index 000000000..e188617a8 --- /dev/null +++ b/egs/aishell/ASR/transducer_stateless_modified/beam_search.py @@ -0,0 +1 @@ +../../../librispeech/ASR/transducer_stateless/beam_search.py \ No newline at end of file diff --git a/egs/aishell/ASR/transducer_stateless_modified/conformer.py b/egs/aishell/ASR/transducer_stateless_modified/conformer.py new file mode 120000 index 000000000..8be0dc864 --- /dev/null +++ b/egs/aishell/ASR/transducer_stateless_modified/conformer.py @@ -0,0 +1 @@ +../../../librispeech/ASR/transducer_stateless/conformer.py \ No newline at end of file diff --git a/egs/aishell/ASR/transducer_stateless_modified/decode.py b/egs/aishell/ASR/transducer_stateless_modified/decode.py new file mode 100755 index 000000000..5b5fe6ffa --- /dev/null +++ b/egs/aishell/ASR/transducer_stateless_modified/decode.py @@ -0,0 +1,486 @@ +#!/usr/bin/env python3 +# +# Copyright 2021 Xiaomi Corporation (Author: Fangjun Kuang) +# +# See ../../../../LICENSE for clarification regarding multiple authors +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +""" +Usage: +(1) greedy search +./transducer_stateless_modified/decode.py \ + --epoch 64 \ + --avg 33 \ + --exp-dir ./transducer_stateless_modified/exp \ + --max-duration 100 \ + --decoding-method greedy_search + +(2) beam search +./transducer_stateless_modified/decode.py \ + --epoch 14 \ + --avg 7 \ + --exp-dir ./transducer_stateless_modified/exp \ + --max-duration 100 \ + --decoding-method beam_search \ + --beam-size 4 + +(3) modified beam search +./transducer_stateless_modified/decode.py \ + --epoch 14 \ + --avg 7 \ + --exp-dir ./transducer_stateless_modified/exp \ + --max-duration 100 \ + --decoding-method modified_beam_search \ + --beam-size 4 +""" + +import argparse +import logging +from collections import defaultdict +from pathlib import Path +from typing import Dict, List, Tuple + +import torch +import torch.nn as nn +from asr_datamodule import AishellAsrDataModule +from beam_search import beam_search, greedy_search, modified_beam_search +from conformer import Conformer +from decoder import Decoder +from joiner import Joiner +from model import Transducer + +from icefall.checkpoint import average_checkpoints, load_checkpoint +from icefall.env import get_env_info +from icefall.lexicon import Lexicon +from icefall.utils import ( + AttributeDict, + setup_logger, + store_transcripts, + write_error_stats, +) + + +def get_parser(): + parser = argparse.ArgumentParser( + formatter_class=argparse.ArgumentDefaultsHelpFormatter + ) + + parser.add_argument( + "--epoch", + type=int, + default=30, + help="It specifies the checkpoint to use for decoding." + "Note: Epoch counts from 0.", + ) + parser.add_argument( + "--avg", + type=int, + default=10, + help="Number of checkpoints to average. Automatically select " + "consecutive checkpoints before the checkpoint specified by " + "'--epoch'. ", + ) + + parser.add_argument( + "--exp-dir", + type=str, + default="transducer_stateless_modified/exp", + help="The experiment dir", + ) + + parser.add_argument( + "--lang-dir", + type=str, + default="data/lang_char", + help="The lang dir", + ) + + parser.add_argument( + "--decoding-method", + type=str, + default="greedy_search", + help="""Possible values are: + - greedy_search + - beam_search + - modified_beam_search + """, + ) + + parser.add_argument( + "--beam-size", + type=int, + default=4, + help="Used only when --decoding-method is beam_search", + ) + + parser.add_argument( + "--context-size", + type=int, + default=2, + help="The context size in the decoder. 1 means bigram; " + "2 means tri-gram", + ) + parser.add_argument( + "--max-sym-per-frame", + type=int, + default=3, + help="Maximum number of symbols per frame", + ) + return parser + + +def get_params() -> AttributeDict: + params = AttributeDict( + { + # parameters for conformer + "feature_dim": 80, + "encoder_out_dim": 512, + "subsampling_factor": 4, + "attention_dim": 512, + "nhead": 8, + "dim_feedforward": 2048, + "num_encoder_layers": 12, + "vgg_frontend": False, + "env_info": get_env_info(), + } + ) + return params + + +def get_encoder_model(params: AttributeDict): + # TODO: We can add an option to switch between Conformer and Transformer + encoder = Conformer( + num_features=params.feature_dim, + output_dim=params.encoder_out_dim, + subsampling_factor=params.subsampling_factor, + d_model=params.attention_dim, + nhead=params.nhead, + dim_feedforward=params.dim_feedforward, + num_encoder_layers=params.num_encoder_layers, + vgg_frontend=params.vgg_frontend, + ) + return encoder + + +def get_decoder_model(params: AttributeDict): + decoder = Decoder( + vocab_size=params.vocab_size, + embedding_dim=params.encoder_out_dim, + blank_id=params.blank_id, + context_size=params.context_size, + ) + return decoder + + +def get_joiner_model(params: AttributeDict): + joiner = Joiner( + input_dim=params.encoder_out_dim, + output_dim=params.vocab_size, + ) + return joiner + + +def get_transducer_model(params: AttributeDict): + encoder = get_encoder_model(params) + decoder = get_decoder_model(params) + joiner = get_joiner_model(params) + + model = Transducer( + encoder=encoder, + decoder=decoder, + joiner=joiner, + ) + return model + + +def decode_one_batch( + params: AttributeDict, + model: nn.Module, + lexicon: Lexicon, + 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 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`. + lexicon: + It contains the token symbol table and the word symbol table. + Returns: + Return the decoding result. See above description for the format of + the returned dict. + """ + device = model.device + feature = batch["inputs"] + assert feature.ndim == 3 + + feature = feature.to(device) + # at entry, feature is (N, T, C) + + supervisions = batch["supervisions"] + feature_lens = supervisions["num_frames"].to(device) + + encoder_out, encoder_out_lens = model.encoder( + x=feature, x_lens=feature_lens + ) + hyps = [] + 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 + ) + elif params.decoding_method == "modified_beam_search": + hyp = modified_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([lexicon.token_table[i] for i in hyp]) + + if params.decoding_method == "greedy_search": + return {"greedy_search": hyps} + else: + return {f"beam_{params.beam_size}": hyps} + + +def decode_dataset( + dl: torch.utils.data.DataLoader, + params: AttributeDict, + model: nn.Module, + lexicon: Lexicon, +) -> 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. + Returns: + Return a dict, whose key may be "greedy_search" if greedy search + is used, or it may be "beam_7" if beam size of 7 is used. + Its value is a list of tuples. Each tuple contains two elements: + The first is the reference transcript, and the second is the + predicted result. + """ + num_cuts = 0 + + try: + num_batches = len(dl) + except TypeError: + num_batches = "?" + + if params.decoding_method == "greedy_search": + log_interval = 100 + else: + log_interval = 2 + + results = defaultdict(list) + for batch_idx, batch in enumerate(dl): + texts = batch["supervisions"]["text"] + + hyps_dict = decode_one_batch( + params=params, + model=model, + lexicon=lexicon, + batch=batch, + ) + + for name, hyps in hyps_dict.items(): + this_batch = [] + assert len(hyps) == len(texts) + for hyp_words, ref_text in zip(hyps, texts): + ref_words = ref_text.split() + this_batch.append((ref_words, hyp_words)) + + results[name].extend(this_batch) + + num_cuts += len(texts) + + if batch_idx % log_interval == 0: + batch_str = f"{batch_idx}/{num_batches}" + + logging.info( + f"batch {batch_str}, cuts processed until now is {num_cuts}" + ) + return results + + +def save_results( + params: AttributeDict, + test_set_name: str, + results_dict: Dict[str, List[Tuple[List[int], List[int]]]], +): + test_set_wers = dict() + for key, results in results_dict.items(): + recog_path = ( + params.res_dir / f"recogs-{test_set_name}-{key}-{params.suffix}.txt" + ) + store_transcripts(filename=recog_path, texts=results) + + # 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" + ) + # we compute CER for aishell dataset. + results_char = [] + for res in results: + results_char.append((list("".join(res[0])), list("".join(res[1])))) + with open(errs_filename, "w") as f: + wer = write_error_stats( + f, f"{test_set_name}-{key}", results_char, 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\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() + AishellAsrDataModule.add_arguments(parser) + args = parser.parse_args() + args.exp_dir = Path(args.exp_dir) + args.lang_dir = Path(args.lang_dir) + + params = get_params() + params.update(vars(args)) + + assert params.decoding_method in ( + "greedy_search", + "beam_search", + "modified_beam_search", + ) + params.res_dir = params.exp_dir / params.decoding_method + + params.suffix = f"epoch-{params.epoch}-avg-{params.avg}" + if "beam_search" in params.decoding_method: + params.suffix += f"-beam-{params.beam_size}" + else: + params.suffix += f"-context-{params.context_size}" + params.suffix += f"-max-sym-per-frame-{params.max_sym_per_frame}" + + 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 = 0 + params.vocab_size = max(lexicon.tokens) + 1 + + logging.info(params) + + logging.info("About to create model") + model = get_transducer_model(params) + + if 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)) + + model.to(device) + model.eval() + model.device = device + + num_param = sum([p.numel() for p in model.parameters()]) + logging.info(f"Number of model parameters: {num_param}") + + aishell = AishellAsrDataModule(args) + test_cuts = aishell.test_cuts() + test_dl = aishell.test_dataloaders(test_cuts) + + test_sets = ["test"] + test_dls = [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, + ) + + 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/aishell/ASR/transducer_stateless_modified/decoder.py b/egs/aishell/ASR/transducer_stateless_modified/decoder.py new file mode 120000 index 000000000..82337f7ef --- /dev/null +++ b/egs/aishell/ASR/transducer_stateless_modified/decoder.py @@ -0,0 +1 @@ +../../../librispeech/ASR/transducer_stateless/decoder.py \ No newline at end of file diff --git a/egs/aishell/ASR/transducer_stateless_modified/encoder_interface.py b/egs/aishell/ASR/transducer_stateless_modified/encoder_interface.py new file mode 120000 index 000000000..653c5b09a --- /dev/null +++ b/egs/aishell/ASR/transducer_stateless_modified/encoder_interface.py @@ -0,0 +1 @@ +../../../librispeech/ASR/transducer_stateless/encoder_interface.py \ No newline at end of file diff --git a/egs/aishell/ASR/transducer_stateless_modified/export.py b/egs/aishell/ASR/transducer_stateless_modified/export.py new file mode 100755 index 000000000..9a20fab6f --- /dev/null +++ b/egs/aishell/ASR/transducer_stateless_modified/export.py @@ -0,0 +1,246 @@ +#!/usr/bin/env python3 +# +# Copyright 2021 Xiaomi Corporation (Author: Fangjun Kuang) +# +# See ../../../../LICENSE for clarification regarding multiple authors +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +# This script converts several saved checkpoints +# to a single one using model averaging. +""" +Usage: +./transducer_stateless_modified/export.py \ + --exp-dir ./transducer_stateless_modified/exp \ + --epoch 64 \ + --avg 33 + +It will generate a file exp_dir/pretrained.pt + +To use the generated file with `transducer_stateless_modified/decode.py`, +you can do:: + + cd /path/to/exp_dir + ln -s pretrained.pt epoch-9999.pt + + cd /path/to/egs/aishell/ASR + ./transducer_stateless_modified/decode.py \ + --exp-dir ./transducer_stateless_modified/exp \ + --epoch 9999 \ + --avg 1 \ + --max-duration 100 \ + --lang-dir data/lang_char +""" + +import argparse +import logging +from pathlib import Path + +import torch +import torch.nn as nn +from conformer import Conformer +from decoder import Decoder +from joiner import Joiner +from model import Transducer + +from icefall.checkpoint import average_checkpoints, load_checkpoint +from icefall.env import get_env_info +from icefall.lexicon import Lexicon +from icefall.utils import AttributeDict, str2bool + + +def get_parser(): + parser = argparse.ArgumentParser( + formatter_class=argparse.ArgumentDefaultsHelpFormatter + ) + + parser.add_argument( + "--epoch", + type=int, + default=20, + help="It specifies the checkpoint to use for decoding." + "Note: Epoch counts from 0.", + ) + + parser.add_argument( + "--avg", + type=int, + default=10, + help="Number of checkpoints to average. Automatically select " + "consecutive checkpoints before the checkpoint specified by " + "'--epoch'. ", + ) + + parser.add_argument( + "--exp-dir", + type=Path, + default=Path("transducer_stateless_modified/exp"), + help="""It specifies the directory where all training related + files, e.g., checkpoints, log, etc, are saved + """, + ) + + parser.add_argument( + "--jit", + type=str2bool, + default=False, + help="""True to save a model after applying torch.jit.script. + """, + ) + + parser.add_argument( + "--lang-dir", + type=Path, + default=Path("data/lang_char"), + help="The lang dir", + ) + + parser.add_argument( + "--context-size", + type=int, + default=2, + help="The context size in the decoder. 1 means bigram; " + "2 means tri-gram", + ) + + return parser + + +def get_params() -> AttributeDict: + params = AttributeDict( + { + # parameters for conformer + "feature_dim": 80, + "encoder_out_dim": 512, + "subsampling_factor": 4, + "attention_dim": 512, + "nhead": 8, + "dim_feedforward": 2048, + "num_encoder_layers": 12, + "vgg_frontend": False, + "env_info": get_env_info(), + } + ) + return params + + +def get_encoder_model(params: AttributeDict) -> nn.Module: + encoder = Conformer( + num_features=params.feature_dim, + output_dim=params.encoder_out_dim, + subsampling_factor=params.subsampling_factor, + d_model=params.attention_dim, + nhead=params.nhead, + dim_feedforward=params.dim_feedforward, + num_encoder_layers=params.num_encoder_layers, + vgg_frontend=params.vgg_frontend, + ) + return encoder + + +def get_decoder_model(params: AttributeDict) -> nn.Module: + decoder = Decoder( + vocab_size=params.vocab_size, + embedding_dim=params.encoder_out_dim, + blank_id=params.blank_id, + context_size=params.context_size, + ) + return decoder + + +def get_joiner_model(params: AttributeDict) -> nn.Module: + joiner = Joiner( + input_dim=params.encoder_out_dim, + output_dim=params.vocab_size, + ) + return joiner + + +def get_transducer_model(params: AttributeDict) -> nn.Module: + encoder = get_encoder_model(params) + decoder = get_decoder_model(params) + joiner = get_joiner_model(params) + + model = Transducer( + encoder=encoder, + decoder=decoder, + joiner=joiner, + ) + return model + + +def main(): + args = get_parser().parse_args() + + assert args.jit is False, "torchscript support will be added later" + + 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}") + + lexicon = Lexicon(params.lang_dir) + + params.blank_id = 0 + params.vocab_size = max(lexicon.tokens) + 1 + + logging.info(params) + + logging.info("About to create model") + model = get_transducer_model(params) + + model.to(device) + + if 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), strict=False + ) + + model.to("cpu") + model.eval() + + if params.jit: + logging.info("Using torch.jit.script") + model = torch.jit.script(model) + filename = params.exp_dir / "cpu_jit.pt" + model.save(str(filename)) + logging.info(f"Saved to {filename}") + else: + logging.info("Not using torch.jit.script") + # Save it using a format so that it can be loaded + # by :func:`load_checkpoint` + filename = params.exp_dir / "pretrained.pt" + torch.save({"model": model.state_dict()}, str(filename)) + logging.info(f"Saved to {filename}") + + +if __name__ == "__main__": + formatter = ( + "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s" + ) + + logging.basicConfig(format=formatter, level=logging.INFO) + main() diff --git a/egs/aishell/ASR/transducer_stateless_modified/joiner.py b/egs/aishell/ASR/transducer_stateless_modified/joiner.py new file mode 120000 index 000000000..1aec6bfaf --- /dev/null +++ b/egs/aishell/ASR/transducer_stateless_modified/joiner.py @@ -0,0 +1 @@ +../../../librispeech/ASR/transducer_stateless/joiner.py \ No newline at end of file diff --git a/egs/aishell/ASR/transducer_stateless_modified/model.py b/egs/aishell/ASR/transducer_stateless_modified/model.py new file mode 120000 index 000000000..16ddd93f0 --- /dev/null +++ b/egs/aishell/ASR/transducer_stateless_modified/model.py @@ -0,0 +1 @@ +../../../librispeech/ASR/transducer_stateless/model.py \ No newline at end of file diff --git a/egs/aishell/ASR/transducer_stateless_modified/pretrained.py b/egs/aishell/ASR/transducer_stateless_modified/pretrained.py new file mode 100755 index 000000000..698594e92 --- /dev/null +++ b/egs/aishell/ASR/transducer_stateless_modified/pretrained.py @@ -0,0 +1,331 @@ +#!/usr/bin/env python3 +# Copyright 2021 Xiaomi Corp. (authors: Fangjun Kuang, +# Wei Kang) +# +# See ../../../../LICENSE for clarification regarding multiple authors +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +""" +Usage: + +# greedy search +./transducer_stateless_modified/pretrained.py \ + --checkpoint /path/to/pretrained.pt \ + --lang-dir /path/to/lang_char \ + --method greedy_search \ + /path/to/foo.wav \ + /path/to/bar.wav + +# beam search +./transducer_stateless_modified/pretrained.py \ + --checkpoint /path/to/pretrained.pt \ + --lang-dir /path/to/lang_char \ + --method beam_search \ + --beam-size 4 \ + /path/to/foo.wav \ + /path/to/bar.wav + +# modified beam search +./transducer_stateless_modified/pretrained.py \ + --checkpoint /path/to/pretrained.pt \ + --lang-dir /path/to/lang_char \ + --method modified_beam_search \ + --beam-size 4 \ + /path/to/foo.wav \ + /path/to/bar.wav + +""" + +import argparse +import logging +import math +from pathlib import Path +from typing import List + +import kaldifeat +import torch +import torch.nn as nn +import torchaudio +from beam_search import beam_search, greedy_search, modified_beam_search +from conformer import Conformer +from decoder import Decoder +from joiner import Joiner +from model import Transducer +from torch.nn.utils.rnn import pad_sequence + +from icefall.env import get_env_info +from icefall.lexicon import Lexicon +from icefall.utils import AttributeDict + + +def get_parser(): + parser = argparse.ArgumentParser( + formatter_class=argparse.ArgumentDefaultsHelpFormatter + ) + + parser.add_argument( + "--checkpoint", + type=str, + required=True, + help="Path to the checkpoint. " + "The checkpoint is assumed to be saved by " + "icefall.checkpoint.save_checkpoint().", + ) + + parser.add_argument( + "--lang-dir", + type=Path, + default=Path("data/lang_char"), + help="The lang dir", + ) + + parser.add_argument( + "--method", + type=str, + default="greedy_search", + help="""Possible values are: + - greedy_search + - beam_search + - modified_beam_search + """, + ) + + parser.add_argument( + "sound_files", + type=str, + nargs="+", + help="The input sound file(s) to transcribe. " + "Supported formats are those supported by torchaudio.load(). " + "For example, wav and flac are supported. " + "The sample rate has to be 16kHz.", + ) + + parser.add_argument( + "--beam-size", + type=int, + default=4, + help="Used only when --method is beam_search and modified_beam_search", + ) + + parser.add_argument( + "--context-size", + type=int, + default=2, + help="The context size in the decoder. 1 means bigram; " + "2 means tri-gram", + ) + parser.add_argument( + "--max-sym-per-frame", + type=int, + default=3, + help="Maximum number of symbols per frame. " + "Use only when --method is greedy_search", + ) + return parser + + return parser + + +def get_params() -> AttributeDict: + params = AttributeDict( + { + # parameters for conformer + "feature_dim": 80, + "encoder_out_dim": 512, + "subsampling_factor": 4, + "attention_dim": 512, + "nhead": 8, + "dim_feedforward": 2048, + "num_encoder_layers": 12, + "vgg_frontend": False, + "env_info": get_env_info(), + "sample_rate": 16000, + } + ) + return params + + +def get_encoder_model(params: AttributeDict) -> nn.Module: + encoder = Conformer( + num_features=params.feature_dim, + output_dim=params.encoder_out_dim, + subsampling_factor=params.subsampling_factor, + d_model=params.attention_dim, + nhead=params.nhead, + dim_feedforward=params.dim_feedforward, + num_encoder_layers=params.num_encoder_layers, + vgg_frontend=params.vgg_frontend, + ) + return encoder + + +def get_decoder_model(params: AttributeDict) -> nn.Module: + decoder = Decoder( + vocab_size=params.vocab_size, + embedding_dim=params.encoder_out_dim, + blank_id=params.blank_id, + context_size=params.context_size, + ) + return decoder + + +def get_joiner_model(params: AttributeDict) -> nn.Module: + joiner = Joiner( + input_dim=params.encoder_out_dim, + output_dim=params.vocab_size, + ) + return joiner + + +def get_transducer_model(params: AttributeDict) -> nn.Module: + encoder = get_encoder_model(params) + decoder = get_decoder_model(params) + joiner = get_joiner_model(params) + + model = Transducer( + encoder=encoder, + decoder=decoder, + joiner=joiner, + ) + return model + + +def read_sound_files( + filenames: List[str], expected_sample_rate: float +) -> List[torch.Tensor]: + """Read a list of sound files into a list 1-D float32 torch tensors. + Args: + filenames: + A list of sound filenames. + expected_sample_rate: + The expected sample rate of the sound files. + Returns: + Return a list of 1-D float32 torch tensors. + """ + ans = [] + for f in filenames: + wave, sample_rate = torchaudio.load(f) + assert sample_rate == expected_sample_rate, ( + f"expected sample rate: {expected_sample_rate}. " + f"Given: {sample_rate}" + ) + # We use only the first channel + ans.append(wave[0]) + return ans + + +def main(): + parser = get_parser() + args = parser.parse_args() + + 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}") + + lexicon = Lexicon(params.lang_dir) + + params.blank_id = 0 + params.vocab_size = max(lexicon.tokens) + 1 + + logging.info(params) + + logging.info("About to create model") + model = get_transducer_model(params) + + checkpoint = torch.load(args.checkpoint, map_location="cpu") + model.load_state_dict(checkpoint["model"]) + model.to(device) + model.eval() + model.device = device + + logging.info("Constructing Fbank computer") + opts = kaldifeat.FbankOptions() + opts.device = device + opts.frame_opts.dither = 0 + opts.frame_opts.snip_edges = False + opts.frame_opts.samp_freq = params.sample_rate + opts.mel_opts.num_bins = params.feature_dim + + fbank = kaldifeat.Fbank(opts) + + logging.info(f"Reading sound files: {params.sound_files}") + waves = read_sound_files( + filenames=params.sound_files, expected_sample_rate=params.sample_rate + ) + waves = [w.to(device) for w in waves] + + logging.info("Decoding started") + features = fbank(waves) + feature_lens = [f.size(0) for f in features] + feature_lens = torch.tensor(feature_lens, device=device) + + features = pad_sequence( + features, batch_first=True, padding_value=math.log(1e-10) + ) + + hyps = [] + with torch.no_grad(): + encoder_out, encoder_out_lens = model.encoder( + x=features, x_lens=feature_lens + ) + + for i in range(encoder_out.size(0)): + # fmt: off + encoder_out_i = encoder_out[i:i+1, :encoder_out_lens[i]] + # fmt: on + if params.method == "greedy_search": + hyp = greedy_search( + model=model, + encoder_out=encoder_out_i, + max_sym_per_frame=params.max_sym_per_frame, + ) + elif params.method == "beam_search": + hyp = beam_search( + model=model, + encoder_out=encoder_out_i, + beam=params.beam_size, + ) + elif params.method == "modified_beam_search": + hyp = modified_beam_search( + model=model, + encoder_out=encoder_out_i, + beam=params.beam_size, + ) + else: + raise ValueError( + f"Unsupported decoding method: {params.method}" + ) + hyps.append([lexicon.token_table[i] for i in hyp]) + + s = "\n" + for filename, hyp in zip(params.sound_files, hyps): + words = " ".join(hyp) + s += f"{filename}:\n{words}\n\n" + logging.info(s) + + logging.info("Decoding Done") + + +if __name__ == "__main__": + formatter = ( + "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s" + ) + + logging.basicConfig(format=formatter, level=logging.INFO) + main() diff --git a/egs/aishell/ASR/transducer_stateless_modified/subsampling.py b/egs/aishell/ASR/transducer_stateless_modified/subsampling.py new file mode 120000 index 000000000..6fee09e58 --- /dev/null +++ b/egs/aishell/ASR/transducer_stateless_modified/subsampling.py @@ -0,0 +1 @@ +../conformer_ctc/subsampling.py \ No newline at end of file diff --git a/egs/aishell/ASR/transducer_stateless_modified/test_decoder.py b/egs/aishell/ASR/transducer_stateless_modified/test_decoder.py new file mode 100755 index 000000000..fe0bdee70 --- /dev/null +++ b/egs/aishell/ASR/transducer_stateless_modified/test_decoder.py @@ -0,0 +1,58 @@ +#!/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. + +""" +To run this file, do: + + cd icefall/egs/aishell/ASR + python ./transducer_stateless/test_decoder.py +""" + +import torch +from decoder import Decoder + + +def test_decoder(): + vocab_size = 3 + blank_id = 0 + embedding_dim = 128 + context_size = 4 + + decoder = Decoder( + vocab_size=vocab_size, + embedding_dim=embedding_dim, + blank_id=blank_id, + context_size=context_size, + ) + N = 100 + U = 20 + x = torch.randint(low=0, high=vocab_size, size=(N, U)) + y = decoder(x) + assert y.shape == (N, U, embedding_dim) + + # for inference + x = torch.randint(low=0, high=vocab_size, size=(N, context_size)) + y = decoder(x, need_pad=False) + assert y.shape == (N, 1, embedding_dim) + + +def main(): + test_decoder() + + +if __name__ == "__main__": + main() diff --git a/egs/aishell/ASR/transducer_stateless_modified/train.py b/egs/aishell/ASR/transducer_stateless_modified/train.py new file mode 100755 index 000000000..524854b73 --- /dev/null +++ b/egs/aishell/ASR/transducer_stateless_modified/train.py @@ -0,0 +1,751 @@ +#!/usr/bin/env python3 +# Copyright 2021 Xiaomi Corp. (authors: Fangjun Kuang, +# Wei Kang +# Mingshuang Luo) +# Copyright 2021 (Pingfeng 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. + +""" +Usage: + +export CUDA_VISIBLE_DEVICES="0,1,2" + +./transducer_stateless_modified/train.py \ + --world-size 3 \ + --num-epochs 65 \ + --start-epoch 0 \ + --exp-dir transducer_stateless_modified/exp \ + --max-duration 250 \ + --lr-factor 2.0 \ + --context-size 2 \ + --modified-transducer-prob 0.25 +""" + + +import argparse +import logging +from pathlib import Path +from shutil import copyfile +from typing import Optional, Tuple + +import k2 +import torch +import torch.multiprocessing as mp +import torch.nn as nn +from asr_datamodule import AishellAsrDataModule +from conformer import Conformer +from decoder import Decoder +from joiner import Joiner +from lhotse.cut import Cut +from lhotse.utils import fix_random_seed +from model import Transducer +from torch import Tensor +from torch.nn.parallel import DistributedDataParallel as DDP +from torch.nn.utils import clip_grad_norm_ +from torch.utils.tensorboard import SummaryWriter +from transformer import Noam + +from icefall.char_graph_compiler import CharCtcTrainingGraphCompiler +from icefall.checkpoint import load_checkpoint +from icefall.checkpoint import save_checkpoint as save_checkpoint_impl +from icefall.dist import cleanup_dist, setup_dist +from icefall.env import get_env_info +from icefall.lexicon import Lexicon +from icefall.utils import AttributeDict, MetricsTracker, setup_logger, str2bool + + +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=0, + help="""Resume training from from this epoch. + If it is positive, it will load checkpoint from + transducer_stateless/exp/epoch-{start_epoch-1}.pt + """, + ) + + parser.add_argument( + "--exp-dir", + type=str, + default="transducer_stateless_modified/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( + "--lr-factor", + type=float, + default=5.0, + help="The lr_factor for Noam optimizer", + ) + + 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( + "--modified-transducer-prob", + type=float, + default=0.25, + help="""The probability to use modified transducer loss. + In modified transduer, it limits the maximum number of symbols + per frame to 1. See also the option --max-sym-per-frame in + transducer_stateless/decode.py + """, + ) + + 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. + + - attention_dim: Hidden dim for multi-head attention model. + + - num_decoder_layers: Number of decoder layer of transformer decoder. + + - warm_step: The warm_step for Noam optimizer. + """ + 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": 800, + # parameters for conformer + "feature_dim": 80, + "encoder_out_dim": 512, + "subsampling_factor": 4, + "attention_dim": 512, + "nhead": 8, + "dim_feedforward": 2048, + "num_encoder_layers": 12, + "vgg_frontend": False, + # parameters for Noam + "warm_step": 80000, # For the 100h subset, use 8k + "env_info": get_env_info(), + } + ) + + return params + + +def get_encoder_model(params: AttributeDict) -> nn.Module: + # TODO: We can add an option to switch between Conformer and Transformer + encoder = Conformer( + num_features=params.feature_dim, + output_dim=params.encoder_out_dim, + subsampling_factor=params.subsampling_factor, + d_model=params.attention_dim, + nhead=params.nhead, + dim_feedforward=params.dim_feedforward, + num_encoder_layers=params.num_encoder_layers, + vgg_frontend=params.vgg_frontend, + ) + return encoder + + +def get_decoder_model(params: AttributeDict) -> nn.Module: + decoder = Decoder( + vocab_size=params.vocab_size, + embedding_dim=params.encoder_out_dim, + blank_id=params.blank_id, + context_size=params.context_size, + ) + return decoder + + +def get_joiner_model(params: AttributeDict) -> nn.Module: + joiner = Joiner( + input_dim=params.encoder_out_dim, + output_dim=params.vocab_size, + ) + return joiner + + +def get_transducer_model(params: AttributeDict) -> nn.Module: + encoder = get_encoder_model(params) + decoder = get_decoder_model(params) + joiner = get_joiner_model(params) + + model = Transducer( + encoder=encoder, + decoder=decoder, + joiner=joiner, + ) + return model + + +def load_checkpoint_if_available( + params: AttributeDict, + model: nn.Module, + optimizer: Optional[torch.optim.Optimizer] = None, + scheduler: Optional[torch.optim.lr_scheduler._LRScheduler] = None, +) -> None: + """Load checkpoint from file. + + If params.start_epoch is positive, it will load the checkpoint from + `params.start_epoch - 1`. Otherwise, this function does nothing. + + Apart from loading state dict for `model`, `optimizer` and `scheduler`, + 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. + optimizer: + The optimizer that we are using. + scheduler: + The learning rate scheduler we are using. + Returns: + Return None. + """ + if params.start_epoch <= 0: + return + + filename = params.exp_dir / f"epoch-{params.start_epoch-1}.pt" + saved_params = load_checkpoint( + filename, + model=model, + 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] + + return saved_params + + +def save_checkpoint( + params: AttributeDict, + model: nn.Module, + optimizer: Optional[torch.optim.Optimizer] = None, + scheduler: Optional[torch.optim.lr_scheduler._LRScheduler] = 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. + """ + if rank != 0: + return + filename = params.exp_dir / f"epoch-{params.cur_epoch}.pt" + save_checkpoint_impl( + filename=filename, + model=model, + params=params, + optimizer=optimizer, + scheduler=scheduler, + 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: nn.Module, + 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 Conformer 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. + """ + device = model.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) + + texts = batch["supervisions"]["text"] + y = graph_compiler.texts_to_ids(texts) + y = k2.RaggedTensor(y).to(device) + + with torch.set_grad_enabled(is_training): + loss = model( + x=feature, + x_lens=feature_lens, + y=y, + modified_transducer_prob=params.modified_transducer_prob, + ) + + assert loss.requires_grad == is_training + + info = MetricsTracker() + 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, + model: nn.Module, + 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: nn.Module, + optimizer: torch.optim.Optimizer, + graph_compiler: CharCtcTrainingGraphCompiler, + train_dl: torch.utils.data.DataLoader, + valid_dl: torch.utils.data.DataLoader, + tb_writer: Optional[SummaryWriter] = None, + world_size: int = 1, +) -> 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. + train_dl: + Dataloader for the training dataset. + valid_dl: + Dataloader for the validation dataset. + tb_writer: + Writer to write log messages to tensorboard. + world_size: + Number of nodes in DDP training. If it is 1, DDP is disabled. + """ + model.train() + + tot_loss = MetricsTracker() + + for batch_idx, batch in enumerate(train_dl): + params.batch_idx_train += 1 + batch_size = len(batch["supervisions"]["text"]) + + 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. + + optimizer.zero_grad() + loss.backward() + clip_grad_norm_(model.parameters(), 5.0, 2.0) + optimizer.step() + + if batch_idx % params.log_interval == 0: + logging.info( + f"Epoch {params.cur_epoch}, " + f"batch {batch_idx}, loss[{loss_info}], " + f"tot_loss[{tot_loss}], batch size: {batch_size}" + ) + + if batch_idx % params.log_interval == 0: + + if tb_writer is not None: + 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 batch_idx > 0 and batch_idx % params.valid_interval == 0: + 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}") + 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(42) + 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, + oov="", + ) + + params.blank_id = 0 + params.vocab_size = max(lexicon.tokens) + 1 + + logging.info(params) + + logging.info("About to create model") + model = get_transducer_model(params) + + num_param = sum([p.numel() for p in model.parameters()]) + logging.info(f"Number of model parameters: {num_param}") + + checkpoints = load_checkpoint_if_available(params=params, model=model) + + model.to(device) + if world_size > 1: + logging.info("Using DDP") + model = DDP(model, device_ids=[rank]) + model.device = device + + optimizer = Noam( + model.parameters(), + model_size=params.attention_dim, + factor=params.lr_factor, + warm_step=params.warm_step, + ) + + if checkpoints and "optimizer" in checkpoints: + logging.info("Loading optimizer state dict") + optimizer.load_state_dict(checkpoints["optimizer"]) + + aishell = AishellAsrDataModule(args) + train_cuts = aishell.train_cuts() + + def remove_short_and_long_utt(c: Cut): + # Keep only utterances with duration between 1 second and 12 seconds + return 1.0 <= c.duration <= 12.0 + + num_in_total = len(train_cuts) + + train_cuts = train_cuts.filter(remove_short_and_long_utt) + + num_left = len(train_cuts) + num_removed = num_in_total - num_left + removed_percent = num_removed / num_in_total * 100 + + logging.info(f"Before removing short and long utterances: {num_in_total}") + logging.info(f"After removing short and long utterances: {num_left}") + logging.info(f"Removed {num_removed} utterances ({removed_percent:.5f}%)") + + train_dl = aishell.train_dataloaders(train_cuts) + valid_dl = aishell.valid_dataloaders(aishell.valid_cuts()) + + scan_pessimistic_batches_for_oom( + model=model, + train_dl=train_dl, + optimizer=optimizer, + graph_compiler=graph_compiler, + params=params, + ) + + for epoch in range(params.start_epoch, params.num_epochs): + train_dl.sampler.set_epoch(epoch) + + cur_lr = optimizer._rate + if tb_writer is not None: + tb_writer.add_scalar( + "train/learning_rate", cur_lr, params.batch_idx_train + ) + tb_writer.add_scalar("train/epoch", epoch, params.batch_idx_train) + + if rank == 0: + logging.info("epoch {}, learning rate {}".format(epoch, cur_lr)) + + params.cur_epoch = epoch + + train_one_epoch( + params=params, + model=model, + optimizer=optimizer, + graph_compiler=graph_compiler, + train_dl=train_dl, + valid_dl=valid_dl, + tb_writer=tb_writer, + world_size=world_size, + ) + + save_checkpoint( + params=params, + model=model, + optimizer=optimizer, + rank=rank, + ) + + logging.info("Done!") + + if world_size > 1: + torch.distributed.barrier() + cleanup_dist() + + +def scan_pessimistic_batches_for_oom( + model: nn.Module, + 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 0 would cause OOM." + ) + batches, crit_values = find_pessimistic_batches(train_dl.sampler) + for criterion, cuts in batches.items(): + batch = train_dl.dataset[cuts] + try: + optimizer.zero_grad() + loss, _ = compute_loss( + params=params, + model=model, + graph_compiler=graph_compiler, + batch=batch, + is_training=True, + ) + loss.backward() + clip_grad_norm_(model.parameters(), 5.0, 2.0) + optimizer.step() + except RuntimeError 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]}) ..." + ) + raise + + +def main(): + parser = get_parser() + AishellAsrDataModule.add_arguments(parser) + args = parser.parse_args() + args.exp_dir = Path(args.exp_dir) + args.lang_dir = Path(args.lang_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/aishell/ASR/transducer_stateless_modified/transformer.py b/egs/aishell/ASR/transducer_stateless_modified/transformer.py new file mode 120000 index 000000000..214afed39 --- /dev/null +++ b/egs/aishell/ASR/transducer_stateless_modified/transformer.py @@ -0,0 +1 @@ +../../../librispeech/ASR/transducer_stateless/transformer.py \ No newline at end of file diff --git a/egs/librispeech/ASR/transducer_stateless/decode.py b/egs/librispeech/ASR/transducer_stateless/decode.py index c101d9397..f23a3a300 100755 --- a/egs/librispeech/ASR/transducer_stateless/decode.py +++ b/egs/librispeech/ASR/transducer_stateless/decode.py @@ -33,6 +33,15 @@ Usage: --max-duration 100 \ --decoding-method beam_search \ --beam-size 4 + +(3) modified beam search +./transducer_stateless/decode.py \ + --epoch 14 \ + --avg 7 \ + --exp-dir ./transducer_stateless/exp \ + --max-duration 100 \ + --decoding-method modified_beam_search \ + --beam-size 4 """ diff --git a/egs/librispeech/ASR/transducer_stateless/joiner.py b/egs/librispeech/ASR/transducer_stateless/joiner.py index 9fd9da4f1..55f0a81f1 100644 --- a/egs/librispeech/ASR/transducer_stateless/joiner.py +++ b/egs/librispeech/ASR/transducer_stateless/joiner.py @@ -39,6 +39,12 @@ class Joiner(nn.Module): Output from the encoder. Its shape is (N, T, self.input_dim). decoder_out: Output from the decoder. Its shape is (N, U, self.input_dim). + encoder_out_len: + A 1-D tensor of shape (N,) containing valid number of frames + before padding in `encoder_out`. + decoder_out_len: + A 1-D tensor of shape (N,) containing valid number of frames + before padding in `decoder_out`. Returns: Return a tensor of shape (sum_all_TU, self.output_dim). """