From ac84220de91dee10c00e8f4223287f937b1930b6 Mon Sep 17 00:00:00 2001 From: Fangjun Kuang Date: Fri, 29 Apr 2022 15:40:30 +0800 Subject: [PATCH] Modified conformer with multi datasets (#312) * Copy files for editing. * Use librispeech + gigaspeech with modified conformer. * Support specifying number of workers for on-the-fly feature extraction. * Feature extraction code for GigaSpeech. * Combine XL splits lazily during training. * Fix warnings in decoding. * Add decoding code for GigaSpeech. * Fix decoding the gigaspeech dataset. We have to use the decoder/joiner networks for the GigaSpeech dataset. * Disable speed perturbe for XL subset. * Compute the Nbest oracle WER for RNN-T decoding. * Minor fixes. * Minor fixes. * Add results. * Update results. * Update CI. * Update results. * Fix style issues. * Update results. * Fix style issues. --- .flake8 | 2 + ...pruned-transducer-stateless2-2022-04-29.sh | 51 + ...pruned-transducer-stateless3-2022-04-29.sh | 51 + .../workflows/run-librispeech-2022-04-29.yml | 85 ++ README.md | 17 + egs/librispeech/ASR/.gitignore | 1 + egs/librispeech/ASR/README.md | 20 +- egs/librispeech/ASR/RESULTS.md | 154 +++ .../compute_fbank_gigaspeech_dev_test.py | 92 ++ .../local/compute_fbank_gigaspeech_splits.py | 169 +++ .../ASR/local/preprocess_gigaspeech.py | 29 +- .../ASR/local/test_load_XL_split.py | 51 + egs/librispeech/ASR/prepare_giga_speech.sh | 33 + .../ASR/pruned_transducer_stateless/decode.py | 2 +- .../beam_search.py | 181 ++- .../pruned_transducer_stateless2/decode.py | 4 +- .../pretrained.py | 306 +++++ .../asr_datamodule.py | 314 +++++ .../beam_search.py | 1 + .../pruned_transducer_stateless3/conformer.py | 1 + .../decode-giga.py | 567 +++++++++ .../pruned_transducer_stateless3/decode.py | 626 +++++++++ .../pruned_transducer_stateless3/decoder.py | 1 + .../encoder_interface.py | 1 + .../pruned_transducer_stateless3/export.py | 183 +++ .../gigaspeech.py | 94 ++ .../gigaspeech_scoring.py | 1 + .../pruned_transducer_stateless3/joiner.py | 1 + .../librispeech.py | 74 ++ .../ASR/pruned_transducer_stateless3/model.py | 235 ++++ .../ASR/pruned_transducer_stateless3/optim.py | 1 + .../pretrained.py | 306 +++++ .../pruned_transducer_stateless3/scaling.py | 1 + .../ASR/pruned_transducer_stateless3/train.py | 1125 +++++++++++++++++ requirements-ci.txt | 2 +- 35 files changed, 4737 insertions(+), 45 deletions(-) create mode 100755 .github/scripts/run-librispeech-pruned-transducer-stateless2-2022-04-29.sh create mode 100755 .github/scripts/run-librispeech-pruned-transducer-stateless3-2022-04-29.sh create mode 100644 .github/workflows/run-librispeech-2022-04-29.yml create mode 100644 egs/librispeech/ASR/.gitignore create mode 100644 egs/librispeech/ASR/local/compute_fbank_gigaspeech_dev_test.py create mode 100644 egs/librispeech/ASR/local/compute_fbank_gigaspeech_splits.py create mode 100755 egs/librispeech/ASR/local/test_load_XL_split.py create mode 100755 egs/librispeech/ASR/pruned_transducer_stateless2/pretrained.py create mode 100644 egs/librispeech/ASR/pruned_transducer_stateless3/asr_datamodule.py create mode 120000 egs/librispeech/ASR/pruned_transducer_stateless3/beam_search.py create mode 120000 egs/librispeech/ASR/pruned_transducer_stateless3/conformer.py create mode 100755 egs/librispeech/ASR/pruned_transducer_stateless3/decode-giga.py create mode 100755 egs/librispeech/ASR/pruned_transducer_stateless3/decode.py create mode 120000 egs/librispeech/ASR/pruned_transducer_stateless3/decoder.py create mode 120000 egs/librispeech/ASR/pruned_transducer_stateless3/encoder_interface.py create mode 100755 egs/librispeech/ASR/pruned_transducer_stateless3/export.py create mode 100644 egs/librispeech/ASR/pruned_transducer_stateless3/gigaspeech.py create mode 120000 egs/librispeech/ASR/pruned_transducer_stateless3/gigaspeech_scoring.py create mode 120000 egs/librispeech/ASR/pruned_transducer_stateless3/joiner.py create mode 100644 egs/librispeech/ASR/pruned_transducer_stateless3/librispeech.py create mode 100644 egs/librispeech/ASR/pruned_transducer_stateless3/model.py create mode 120000 egs/librispeech/ASR/pruned_transducer_stateless3/optim.py create mode 100755 egs/librispeech/ASR/pruned_transducer_stateless3/pretrained.py create mode 120000 egs/librispeech/ASR/pruned_transducer_stateless3/scaling.py create mode 100755 egs/librispeech/ASR/pruned_transducer_stateless3/train.py diff --git a/.flake8 b/.flake8 index cd55ded73..67fbbdfad 100644 --- a/.flake8 +++ b/.flake8 @@ -9,6 +9,8 @@ per-file-ignores = egs/tedlium3/ASR/*/conformer.py: E501, egs/gigaspeech/ASR/*/conformer.py: E501, egs/librispeech/ASR/pruned_transducer_stateless2/*.py: E501, + egs/librispeech/ASR/*/optim.py: E501, + egs/librispeech/ASR/*/scaling.py: E501, # invalid escape sequence (cause by tex formular), W605 icefall/utils.py: E501, W605 diff --git a/.github/scripts/run-librispeech-pruned-transducer-stateless2-2022-04-29.sh b/.github/scripts/run-librispeech-pruned-transducer-stateless2-2022-04-29.sh new file mode 100755 index 000000000..ee8610996 --- /dev/null +++ b/.github/scripts/run-librispeech-pruned-transducer-stateless2-2022-04-29.sh @@ -0,0 +1,51 @@ +#!/usr/bin/env bash + +log() { + # This function is from espnet + local fname=${BASH_SOURCE[1]##*/} + echo -e "$(date '+%Y-%m-%d %H:%M:%S') (${fname}:${BASH_LINENO[0]}:${FUNCNAME[1]}) $*" +} + +cd egs/librispeech/ASR + +repo_url=https://huggingface.co/csukuangfj/icefall-asr-librispeech-pruned-transducer-stateless2-2022-04-29 + +log "Downloading pre-trained model from $repo_url" +git lfs install +git clone $repo_url +repo=$(basename $repo_url) + +log "Display test files" +tree $repo/ +soxi $repo/test_wavs/*.wav +ls -lh $repo/test_wavs/*.wav + +pushd $repo/exp +ln -s pretrained-epoch-38-avg-10.pt pretrained.pt +popd + +for sym in 1 2 3; do + log "Greedy search with --max-sym-per-frame $sym" + + ./pruned_transducer_stateless2/pretrained.py \ + --method greedy_search \ + --max-sym-per-frame $sym \ + --checkpoint $repo/exp/pretrained.pt \ + --bpe-model $repo/data/lang_bpe_500/bpe.model \ + $repo/test_wavs/1089-134686-0001.wav \ + $repo/test_wavs/1221-135766-0001.wav \ + $repo/test_wavs/1221-135766-0002.wav +done + +for method in modified_beam_search beam_search fast_beam_search; do + log "$method" + + ./pruned_transducer_stateless2/pretrained.py \ + --method $method \ + --beam-size 4 \ + --checkpoint $repo/exp/pretrained.pt \ + --bpe-model $repo/data/lang_bpe_500/bpe.model \ + $repo/test_wavs/1089-134686-0001.wav \ + $repo/test_wavs/1221-135766-0001.wav \ + $repo/test_wavs/1221-135766-0002.wav +done diff --git a/.github/scripts/run-librispeech-pruned-transducer-stateless3-2022-04-29.sh b/.github/scripts/run-librispeech-pruned-transducer-stateless3-2022-04-29.sh new file mode 100755 index 000000000..d28e888e7 --- /dev/null +++ b/.github/scripts/run-librispeech-pruned-transducer-stateless3-2022-04-29.sh @@ -0,0 +1,51 @@ +#!/usr/bin/env bash + +log() { + # This function is from espnet + local fname=${BASH_SOURCE[1]##*/} + echo -e "$(date '+%Y-%m-%d %H:%M:%S') (${fname}:${BASH_LINENO[0]}:${FUNCNAME[1]}) $*" +} + +cd egs/librispeech/ASR + +repo_url=https://huggingface.co/csukuangfj/icefall-asr-librispeech-pruned-transducer-stateless3-2022-04-29 + +log "Downloading pre-trained model from $repo_url" +git lfs install +git clone $repo_url +repo=$(basename $repo_url) + +log "Display test files" +tree $repo/ +soxi $repo/test_wavs/*.wav +ls -lh $repo/test_wavs/*.wav + +pushd $repo/exp +ln -s pretrained-epoch-25-avg-6.pt pretrained.pt +popd + +for sym in 1 2 3; do + log "Greedy search with --max-sym-per-frame $sym" + + ./pruned_transducer_stateless3/pretrained.py \ + --method greedy_search \ + --max-sym-per-frame $sym \ + --checkpoint $repo/exp/pretrained.pt \ + --bpe-model $repo/data/lang_bpe_500/bpe.model \ + $repo/test_wavs/1089-134686-0001.wav \ + $repo/test_wavs/1221-135766-0001.wav \ + $repo/test_wavs/1221-135766-0002.wav +done + +for method in modified_beam_search beam_search fast_beam_search; do + log "$method" + + ./pruned_transducer_stateless3/pretrained.py \ + --method $method \ + --beam-size 4 \ + --checkpoint $repo/exp/pretrained.pt \ + --bpe-model $repo/data/lang_bpe_500/bpe.model \ + $repo/test_wavs/1089-134686-0001.wav \ + $repo/test_wavs/1221-135766-0001.wav \ + $repo/test_wavs/1221-135766-0002.wav +done diff --git a/.github/workflows/run-librispeech-2022-04-29.yml b/.github/workflows/run-librispeech-2022-04-29.yml new file mode 100644 index 000000000..129e30698 --- /dev/null +++ b/.github/workflows/run-librispeech-2022-04-29.yml @@ -0,0 +1,85 @@ +# 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-librispeech-2022-04-29 +# stateless pruned transducer (reworked model) + giga speech + +on: + push: + branches: + - master + pull_request: + types: [labeled] + +jobs: + run_librispeech_2022_04_29: + 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] + + fail-fast: false + + steps: + - uses: actions/checkout@v2 + with: + fetch-depth: 0 + + - name: Setup Python ${{ matrix.python-version }} + uses: actions/setup-python@v2 + with: + python-version: ${{ matrix.python-version }} + cache: 'pip' + cache-dependency-path: '**/requirements-ci.txt' + + - name: Install Python dependencies + run: | + grep -v '^#' ./requirements-ci.txt | xargs -n 1 -L 1 pip install + + - name: Cache kaldifeat + id: my-cache + uses: actions/cache@v2 + with: + path: | + ~/tmp/kaldifeat + key: cache-tmp-${{ matrix.python-version }} + + - name: Install kaldifeat + if: steps.my-cache.outputs.cache-hit != 'true' + shell: bash + run: | + mkdir -p ~/tmp + cd ~/tmp + git clone https://github.com/csukuangfj/kaldifeat + cd kaldifeat + mkdir build + cd build + cmake -DCMAKE_BUILD_TYPE=Release .. + make -j2 _kaldifeat + + - name: Inference with pre-trained model + shell: bash + run: | + sudo apt-get -qq install git-lfs tree sox + export PYTHONPATH=$PWD:$PYTHONPATH + export PYTHONPATH=~/tmp/kaldifeat/kaldifeat/python:$PYTHONPATH + export PYTHONPATH=~/tmp/kaldifeat/build/lib:$PYTHONPATH + + .github/scripts/run-librispeech-pruned-transducer-stateless2-2022-04-29.sh + + .github/scripts/run-librispeech-pruned-transducer-stateless3-2022-04-29.sh diff --git a/README.md b/README.md index 6adba4955..188ca013b 100644 --- a/README.md +++ b/README.md @@ -35,6 +35,9 @@ We do provide a Colab notebook for this recipe. ### LibriSpeech +Please see +for the **latest** results. + We provide 4 models for this recipe: - [conformer CTC model][LibriSpeech_conformer_ctc] @@ -92,6 +95,20 @@ in the decoding. We provide a Colab notebook to run a pre-trained transducer conformer + stateless decoder model: [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/drive/1CO1bXJ-2khDckZIW8zjOPHGSKLHpTDlp?usp=sharing) + +#### k2 pruned RNN-T + +| | test-clean | test-other | +|-----|------------|------------| +| WER | 2.57 | 5.95 | + +#### k2 pruned RNN-T + GigaSpeech + +| | test-clean | test-other | +|-----|------------|------------| +| WER | 2.19 | 4.97 | + + ### Aishell We provide two models for this recipe: [conformer CTC model][Aishell_conformer_ctc] diff --git a/egs/librispeech/ASR/.gitignore b/egs/librispeech/ASR/.gitignore new file mode 100644 index 000000000..5592679cc --- /dev/null +++ b/egs/librispeech/ASR/.gitignore @@ -0,0 +1 @@ +log-* diff --git a/egs/librispeech/ASR/README.md b/egs/librispeech/ASR/README.md index de9d6d50a..8d299a28b 100644 --- a/egs/librispeech/ASR/README.md +++ b/egs/librispeech/ASR/README.md @@ -1,23 +1,19 @@ - -# Introduction - -Please refer to -for how to run models in this recipe. - +# 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` | Conformer | LSTM | | +|---------------------------------------|---------------------|--------------------|---------------------------------------------------| +| `transducer` | Conformer | LSTM | | | `transducer_stateless` | Conformer | Embedding + Conv1d | Using optimized_transducer from computing RNN-T loss | | `transducer_stateless2` | Conformer | Embedding + Conv1d | Using torchaudio for computing RNN-T loss | -| `transducer_lstm` | LSTM | LSTM | | -| `transducer_stateless_multi_datasets` | Conformer | Embedding + Conv1d | Using data from GigaSpeech as extra training data | -| `pruned_transducer_stateless` | Conformer | Embedding + Conv1d | Using k2 pruned RNN-T loss | -| `pruned_transducer_stateless2` | Conformer(modified) | Embedding + Conv1d | Using k2 pruned RNN-T loss | +| `transducer_lstm` | LSTM | LSTM | | +| `transducer_stateless_multi_datasets` | Conformer | Embedding + Conv1d | Using data from GigaSpeech as extra training data | +| `pruned_transducer_stateless` | Conformer | Embedding + Conv1d | Using k2 pruned RNN-T loss | +| `pruned_transducer_stateless2` | Conformer(modified) | Embedding + Conv1d | Using k2 pruned RNN-T loss | +| `pruned_transducer_stateless3` | Conformer(modified) | Embedding + Conv1d | Using k2 pruned RNN-T loss + using GigaSpeech as extra training data | The decoder in `transducer_stateless` is modified from the paper diff --git a/egs/librispeech/ASR/RESULTS.md b/egs/librispeech/ASR/RESULTS.md index 2fed7feed..b190e13e8 100644 --- a/egs/librispeech/ASR/RESULTS.md +++ b/egs/librispeech/ASR/RESULTS.md @@ -1,5 +1,155 @@ ## Results +### LibriSpeech BPE training results (Pruned Transducer 3) + +[pruned_transducer_stateless3](./pruned_transducer_stateless3) +Same as `Pruned Transducer 2` but using the XL subset from +[GigaSpeech](https://github.com/SpeechColab/GigaSpeech) as extra training data. + +During training, it selects either a batch from GigaSpeech with prob `giga_prob` +or a batch from LibriSpeech with prob `1 - giga_prob`. All utterances within +a batch comes from the same dataset. + +See + +The WERs are: + +| | test-clean | test-other | comment | +|-------------------------------------|------------|------------|----------------------------------------| +| greedy search (max sym per frame 1) | 2.21 | 5.09 | --epoch 27 --avg 2 --max-duration 600 | +| greedy search (max sym per frame 1) | 2.25 | 5.02 | --epoch 27 --avg 12 --max-duration 600 | +| modified beam search | 2.19 | 5.03 | --epoch 25 --avg 6 --max-duration 600 | +| modified beam search | 2.23 | 4.94 | --epoch 27 --avg 10 --max-duration 600 | +| beam search | 2.16 | 4.95 | --epoch 25 --avg 7 --max-duration 600 | +| fast beam search | 2.21 | 4.96 | --epoch 27 --avg 10 --max-duration 600 | +| fast beam search | 2.19 | 4.97 | --epoch 27 --avg 12 --max-duration 600 | + +The training commands are: + +```bash +./prepare.sh +./prepare_giga_speech.sh + +export CUDA_VISIBLE_DEVICES="0,1,2,3,4,5,6,7" + +./pruned_transducer_stateless3/train.py \ + --world-size 8 \ + --num-epochs 30 \ + --start-epoch 0 \ + --full-libri 1 \ + --exp-dir pruned_transducer_stateless3/exp \ + --max-duration 300 \ + --use-fp16 1 \ + --lr-epochs 4 \ + --num-workers 2 \ + --giga-prob 0.8 +``` + +The tensorboard log can be found at + +(Note: The training process is killed manually at `epoch-28.pt`.) + +Pretrained models, training logs, decoding logs, and decoding results +are available at + + +Decoding commands are: + +```bash + +# greedy search +./pruned_transducer_stateless3/decode.py \ + --epoch 27 \ + --avg 2 \ + --exp-dir ./pruned_transducer_stateless3/exp \ + --max-duration 600 \ + --decoding-method greedy_search \ + --max-sym-per-frame 1 + +# modified beam search +./pruned_transducer_stateless3/decode.py \ + --epoch 25 \ + --avg 6 \ + --exp-dir ./pruned_transducer_stateless3/exp \ + --max-duration 600 \ + --decoding-method modified_beam_search \ + --max-sym-per-frame 1 + +# beam search +./pruned_transducer_stateless3/decode.py \ + --epoch 25 \ + --avg 7 \ + --exp-dir ./pruned_transducer_stateless3/exp \ + --max-duration 600 \ + --decoding-method beam_search \ + --max-sym-per-frame 1 + +# fast beam search +for epoch in 27; do + for avg in 10 12; do + ./pruned_transducer_stateless3/decode.py \ + --epoch $epoch \ + --avg $avg \ + --exp-dir ./pruned_transducer_stateless3/exp \ + --max-duration 600 \ + --decoding-method fast_beam_search \ + --max-states 32 \ + --beam 8 + done +done +``` + +The following table shows the +[Nbest oracle WER](http://kaldi-asr.org/doc/lattices.html#lattices_operations_oracle) +for fast beam search. +| epoch | avg | num_paths | nbest_scale | test-clean | test-other | +|-------|-----|-----------|-------------|------------|------------| +| 27 | 10 | 50 | 0.5 | 0.91 | 2.74 | +| 27 | 10 | 50 | 0.8 | 0.94 | 2.82 | +| 27 | 10 | 50 | 1.0 | 1.06 | 2.88 | +| 27 | 10 | 100 | 0.5 | 0.82 | 2.58 | +| 27 | 10 | 100 | 0.8 | 0.92 | 2.65 | +| 27 | 10 | 100 | 1.0 | 0.95 | 2.77 | +| 27 | 10 | 200 | 0.5 | 0.81 | 2.50 | +| 27 | 10 | 200 | 0.8 | 0.85 | 2.56 | +| 27 | 10 | 200 | 1.0 | 0.91 | 2.64 | +| 27 | 10 | 400 | 0.5 | N/A | N/A | +| 27 | 10 | 400 | 0.8 | 0.81 | 2.49 | +| 27 | 10 | 400 | 1.0 | 0.85 | 2.54 | + +The Nbest oracle WER is computed using the following steps: + + - 1. Use `fast_beam_search` to produce a lattice. + - 2. Extract `N` paths from the lattice using [k2.random_path](https://k2-fsa.github.io/k2/python_api/api.html#random-paths) + - 3. [Unique](https://k2-fsa.github.io/k2/python_api/api.html#unique) paths so that each path + has a distinct sequence of tokens + - 4. Compute the edit distance of each path with the ground truth + - 5. The path with the lowest edit distance is the final output and is used to + compute the WER + +The command to compute the Nbest oracle WER is: + +```bash +for epoch in 27; do + for avg in 10 ; do + for num_paths in 50 100 200 400; do + for nbest_scale in 0.5 0.8 1.0; do + ./pruned_transducer_stateless3/decode.py \ + --epoch $epoch \ + --avg $avg \ + --exp-dir ./pruned_transducer_stateless3/exp \ + --max-duration 600 \ + --decoding-method fast_beam_search_nbest_oracle \ + --num-paths $num_paths \ + --max-states 32 \ + --beam 8 \ + --nbest-scale $nbest_scale + done + done + done +done +``` + ### LibriSpeech BPE training results (Pruned Transducer 2) [pruned_transducer_stateless2](./pruned_transducer_stateless2) @@ -33,6 +183,10 @@ and: The Tensorboard log is at (apologies, log starts only from epoch 3). +The pretrained models, training logs, decoding logs, and decoding results +can be found at + + #### Training on train-clean-100: diff --git a/egs/librispeech/ASR/local/compute_fbank_gigaspeech_dev_test.py b/egs/librispeech/ASR/local/compute_fbank_gigaspeech_dev_test.py new file mode 100644 index 000000000..9f1039893 --- /dev/null +++ b/egs/librispeech/ASR/local/compute_fbank_gigaspeech_dev_test.py @@ -0,0 +1,92 @@ +#!/usr/bin/env python3 +# Copyright 2021 Johns Hopkins University (Piotr Żelasko) +# Copyright 2021 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 + +import torch +from lhotse import ( + CutSet, + KaldifeatFbank, + KaldifeatFbankConfig, +) + +# Torch's multithreaded behavior needs to be disabled or +# it wastes a lot of CPU and slow things down. +# Do this outside of main() in case it needs to take effect +# even when we are not invoking the main (e.g. when spawning subprocesses). +torch.set_num_threads(1) +torch.set_num_interop_threads(1) + + +def compute_fbank_gigaspeech_dev_test(): + in_out_dir = Path("data/fbank") + # number of workers in dataloader + num_workers = 20 + + # number of seconds in a batch + batch_duration = 600 + + subsets = ("DEV", "TEST") + + device = torch.device("cpu") + if torch.cuda.is_available(): + device = torch.device("cuda", 0) + extractor = KaldifeatFbank(KaldifeatFbankConfig(device=device)) + + logging.info(f"device: {device}") + + for partition in subsets: + cuts_path = in_out_dir / f"cuts_{partition}.jsonl.gz" + if cuts_path.is_file(): + logging.info(f"{cuts_path} exists - skipping") + continue + + raw_cuts_path = in_out_dir / f"cuts_{partition}_raw.jsonl.gz" + + logging.info(f"Loading {raw_cuts_path}") + cut_set = CutSet.from_file(raw_cuts_path) + + logging.info("Computing features") + + cut_set = cut_set.compute_and_store_features_batch( + extractor=extractor, + storage_path=f"{in_out_dir}/feats_{partition}", + num_workers=num_workers, + batch_duration=batch_duration, + ) + cut_set = cut_set.trim_to_supervisions( + keep_overlapping=False, min_duration=None + ) + + logging.info(f"Saving to {cuts_path}") + cut_set.to_file(cuts_path) + logging.info(f"Saved to {cuts_path}") + + +def main(): + formatter = ( + "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s" + ) + logging.basicConfig(format=formatter, level=logging.INFO) + + compute_fbank_gigaspeech_dev_test() + + +if __name__ == "__main__": + main() diff --git a/egs/librispeech/ASR/local/compute_fbank_gigaspeech_splits.py b/egs/librispeech/ASR/local/compute_fbank_gigaspeech_splits.py new file mode 100644 index 000000000..a7ed2467d --- /dev/null +++ b/egs/librispeech/ASR/local/compute_fbank_gigaspeech_splits.py @@ -0,0 +1,169 @@ +#!/usr/bin/env python3 +# Copyright 2021 Johns Hopkins University (Piotr Żelasko) +# Copyright 2021 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 argparse +import logging +import os +from datetime import datetime +from pathlib import Path + +import torch +from lhotse import CutSet, KaldifeatFbank, KaldifeatFbankConfig + +# Torch's multithreaded behavior needs to be disabled or +# it wastes a lot of CPU and slow things down. +# Do this outside of main() in case it needs to take effect +# even when we are not invoking the main (e.g. when spawning subprocesses). +torch.set_num_threads(1) +torch.set_num_interop_threads(1) + + +def get_parser(): + parser = argparse.ArgumentParser( + formatter_class=argparse.ArgumentDefaultsHelpFormatter + ) + + parser.add_argument( + "--num-workers", + type=int, + default=20, + help="Number of dataloading workers used for reading the audio.", + ) + parser.add_argument( + "--batch-duration", + type=float, + default=600.0, + help="The maximum number of audio seconds in a batch." + "Determines batch size dynamically.", + ) + + parser.add_argument( + "--num-splits", + type=int, + required=True, + help="The number of splits of the XL subset", + ) + + parser.add_argument( + "--start", + type=int, + default=0, + help="Process pieces starting from this number (inclusive).", + ) + + parser.add_argument( + "--stop", + type=int, + default=-1, + help="Stop processing pieces until this number (exclusive).", + ) + return parser + + +def compute_fbank_gigaspeech_splits(args): + num_splits = args.num_splits + output_dir = f"data/fbank/XL_split_{num_splits}" + output_dir = Path(output_dir) + assert output_dir.exists(), f"{output_dir} does not exist!" + + num_digits = len(str(num_splits)) + + start = args.start + stop = args.stop + if stop < start: + stop = num_splits + + stop = min(stop, num_splits) + + device = torch.device("cpu") + if torch.cuda.is_available(): + device = torch.device("cuda", 0) + extractor = KaldifeatFbank(KaldifeatFbankConfig(device=device)) + logging.info(f"device: {device}") + + num_digits = 8 # num_digits is fixed by lhotse split-lazy + for i in range(start, stop): + idx = f"{i + 1}".zfill(num_digits) + logging.info(f"Processing {idx}/{num_splits}") + + cuts_path = output_dir / f"cuts_XL.{idx}.jsonl.gz" + if cuts_path.is_file(): + logging.info(f"{cuts_path} exists - skipping") + continue + + raw_cuts_path = output_dir / f"cuts_XL_raw.{idx}.jsonl.gz" + if not raw_cuts_path.is_file(): + logging.info(f"{raw_cuts_path} does not exist - skipping it") + continue + + logging.info(f"Loading {raw_cuts_path}") + cut_set = CutSet.from_file(raw_cuts_path) + + logging.info("Computing features") + if (output_dir / f"feats_XL_{idx}.lca").exists(): + logging.info(f"Removing {output_dir}/feats_XL_{idx}.lca") + os.remove(output_dir / f"feats_XL_{idx}.lca") + + cut_set = cut_set.compute_and_store_features_batch( + extractor=extractor, + storage_path=f"{output_dir}/feats_XL_{idx}", + num_workers=args.num_workers, + batch_duration=args.batch_duration, + ) + + logging.info("About to split cuts into smaller chunks.") + cut_set = cut_set.trim_to_supervisions( + keep_overlapping=False, min_duration=None + ) + + logging.info(f"Saving to {cuts_path}") + cut_set.to_file(cuts_path) + logging.info(f"Saved to {cuts_path}") + + +def main(): + now = datetime.now() + date_time = now.strftime("%Y-%m-%d-%H-%M-%S") + + log_filename = "log-compute_fbank_gigaspeech_splits" + formatter = ( + "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s" + ) + log_filename = f"{log_filename}-{date_time}" + + logging.basicConfig( + filename=log_filename, + format=formatter, + level=logging.INFO, + filemode="w", + ) + + console = logging.StreamHandler() + console.setLevel(logging.INFO) + console.setFormatter(logging.Formatter(formatter)) + logging.getLogger("").addHandler(console) + + parser = get_parser() + args = parser.parse_args() + logging.info(vars(args)) + + compute_fbank_gigaspeech_splits(args) + + +if __name__ == "__main__": + main() diff --git a/egs/librispeech/ASR/local/preprocess_gigaspeech.py b/egs/librispeech/ASR/local/preprocess_gigaspeech.py index 4168a7185..cd1345904 100644 --- a/egs/librispeech/ASR/local/preprocess_gigaspeech.py +++ b/egs/librispeech/ASR/local/preprocess_gigaspeech.py @@ -91,21 +91,20 @@ def preprocess_giga_speech(): ) # Run data augmentation that needs to be done in the # time domain. - if partition not in ["DEV", "TEST"]: - logging.info( - f"Speed perturb for {partition} with factors 0.9 and 1.1 " - "(Perturbing may take 8 minutes and saving may take 20 minutes)" - ) - cut_set = ( - cut_set - + cut_set.perturb_speed(0.9) - + cut_set.perturb_speed(1.1) - ) - - logging.info("About to split cuts into smaller chunks.") - cut_set = cut_set.trim_to_supervisions( - keep_overlapping=False, min_duration=None - ) + # if partition not in ["DEV", "TEST"]: + # logging.info( + # f"Speed perturb for {partition} with factors 0.9 and 1.1 " + # "(Perturbing may take 8 minutes and saving may" + # " take 20 minutes)" + # ) + # cut_set = ( + # cut_set + # + cut_set.perturb_speed(0.9) + # + cut_set.perturb_speed(1.1) + # ) + # + # Note: No need to perturb the training subset as not all of the + # data is going to be used in the training. logging.info(f"Saving to {raw_cuts_path}") cut_set.to_file(raw_cuts_path) diff --git a/egs/librispeech/ASR/local/test_load_XL_split.py b/egs/librispeech/ASR/local/test_load_XL_split.py new file mode 100755 index 000000000..3982a7157 --- /dev/null +++ b/egs/librispeech/ASR/local/test_load_XL_split.py @@ -0,0 +1,51 @@ +#!/usr/bin/env python3 +# Copyright 2022 Xiaomi Corp. (authors: Fangjun Kuang) +# +# See ../../../../LICENSE for clarification regarding multiple authors +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +""" +This file can be used to check if any split is corrupted. +""" + +import glob +import re + +import lhotse + + +def main(): + d = "data/fbank/XL_split_2000" + filenames = list(glob.glob(f"{d}/cuts_XL.*.jsonl.gz")) + + pattern = re.compile(r"cuts_XL.([0-9]+).jsonl.gz") + + idx_filenames = [(int(pattern.search(c).group(1)), c) for c in filenames] + + idx_filenames = sorted(idx_filenames, key=lambda x: x[0]) + + print(f"Loading {len(idx_filenames)} splits") + + s = 0 + for i, f in idx_filenames: + cuts = lhotse.load_manifest_lazy(f) + print(i, "filename", f) + for i, c in enumerate(cuts): + s += c.features.load().shape[0] + if i > 5: + break + + +if __name__ == "__main__": + main() diff --git a/egs/librispeech/ASR/prepare_giga_speech.sh b/egs/librispeech/ASR/prepare_giga_speech.sh index 49124c4d7..26b921eab 100755 --- a/egs/librispeech/ASR/prepare_giga_speech.sh +++ b/egs/librispeech/ASR/prepare_giga_speech.sh @@ -24,6 +24,15 @@ stop_stage=100 # DEV 12 hours # Test 40 hours +# Split XL subset to this number of pieces +# This is to avoid OOM during feature extraction. +num_splits=2000 +# We use lazy split from lhotse. +# The XL subset (10k hours) contains 37956 cuts without speed perturbing. +# We want to split it into 2000 splits, so each split +# contains about 37956 / 2000 = 19 cuts. As a result, there will be 1998 splits. +chunk_size=19 # number of cuts in each split. The last split may contain fewer cuts. + dl_dir=$PWD/download . shared/parse_options.sh || exit 1 @@ -107,3 +116,27 @@ if [ $stage -le 2 ] && [ $stop_stage -ge 2 ]; then touch data/fbank/.preprocess_complete fi fi + +if [ $stage -le 3 ] && [ $stop_stage -ge 3 ]; then + log "Stage 3: Compute features for DEV and TEST subsets of GigaSpeech (may take 2 minutes)" + python3 ./local/compute_fbank_gigaspeech_dev_test.py +fi + +if [ $stage -le 4 ] && [ $stop_stage -ge 4 ]; then + log "Stage 4: Split XL subset into ${num_splits} pieces" + split_dir=data/fbank/XL_split_${num_splits} + if [ ! -f $split_dir/.split_completed ]; then + lhotse split-lazy ./data/fbank/cuts_XL_raw.jsonl.gz $split_dir $chunk_size + touch $split_dir/.split_completed + fi +fi + +if [ $stage -le 5 ] && [ $stop_stage -ge 5 ]; then + log "Stage 5: Compute features for XL" + # Note: The script supports --start and --stop options. + # You can use several machines to compute the features in parallel. + python3 ./local/compute_fbank_gigaspeech_splits.py \ + --num-workers $nj \ + --batch-duration 600 \ + --num-splits $num_splits +fi diff --git a/egs/librispeech/ASR/pruned_transducer_stateless/decode.py b/egs/librispeech/ASR/pruned_transducer_stateless/decode.py index 5082ab71a..349e4c281 100755 --- a/egs/librispeech/ASR/pruned_transducer_stateless/decode.py +++ b/egs/librispeech/ASR/pruned_transducer_stateless/decode.py @@ -174,7 +174,7 @@ def get_parser(): "--beam-size", type=int, default=4, - help="""An interger indicating how many candidates we will keep for each + help="""An integer indicating how many candidates we will keep for each frame. Used only when --decoding-method is beam_search or modified_beam_search.""", ) diff --git a/egs/librispeech/ASR/pruned_transducer_stateless2/beam_search.py b/egs/librispeech/ASR/pruned_transducer_stateless2/beam_search.py index 2e9bf3e0b..ad492aaa5 100644 --- a/egs/librispeech/ASR/pruned_transducer_stateless2/beam_search.py +++ b/egs/librispeech/ASR/pruned_transducer_stateless2/beam_search.py @@ -14,6 +14,7 @@ # See the License for the specific language governing permissions and # limitations under the License. +import warnings from dataclasses import dataclass from typing import Dict, List, Optional @@ -21,11 +22,11 @@ import k2 import torch from model import Transducer -from icefall.decode import one_best_decoding +from icefall.decode import Nbest, one_best_decoding from icefall.utils import get_texts -def fast_beam_search( +def fast_beam_search_one_best( model: Transducer, decoding_graph: k2.Fsa, encoder_out: torch.Tensor, @@ -36,6 +37,9 @@ def fast_beam_search( ) -> List[List[int]]: """It limits the maximum number of symbols per frame to 1. + A lattice is first obtained using modified beam search, and then + the shortest path within the lattice is used as the final output. + Args: model: An instance of `Transducer`. @@ -55,6 +59,148 @@ def fast_beam_search( Returns: Return the decoded result. """ + lattice = fast_beam_search( + model=model, + decoding_graph=decoding_graph, + encoder_out=encoder_out, + encoder_out_lens=encoder_out_lens, + beam=beam, + max_states=max_states, + max_contexts=max_contexts, + ) + + best_path = one_best_decoding(lattice) + hyps = get_texts(best_path) + return hyps + + +def fast_beam_search_nbest_oracle( + model: Transducer, + decoding_graph: k2.Fsa, + encoder_out: torch.Tensor, + encoder_out_lens: torch.Tensor, + beam: float, + max_states: int, + max_contexts: int, + num_paths: int, + ref_texts: List[List[int]], + use_double_scores: bool = True, + nbest_scale: float = 0.5, +) -> List[List[int]]: + """It limits the maximum number of symbols per frame to 1. + + A lattice is first obtained using modified beam search, and then + we select `num_paths` linear paths from the lattice. The path + that has the minimum edit distance with the given reference transcript + is used as the output. + + This is the best result we can achieve for any nbest based rescoring + methods. + + Args: + model: + An instance of `Transducer`. + decoding_graph: + Decoding graph used for decoding, may be a TrivialGraph or a HLG. + encoder_out: + A tensor of shape (N, T, C) from the encoder. + encoder_out_lens: + A tensor of shape (N,) containing the number of frames in `encoder_out` + before padding. + beam: + Beam value, similar to the beam used in Kaldi.. + max_states: + Max states per stream per frame. + max_contexts: + Max contexts pre stream per frame. + num_paths: + Number of paths to extract from the decoded lattice. + ref_texts: + A list-of-list of integers containing the reference transcripts. + If the decoding_graph is a trivial_graph, the integer ID is the + BPE token ID. + use_double_scores: + True to use double precision for computation. False to use + single precision. + nbest_scale: + It's the scale applied to the lattice.scores. A smaller value + yields more unique paths. + + Returns: + Return the decoded result. + """ + lattice = fast_beam_search( + model=model, + decoding_graph=decoding_graph, + encoder_out=encoder_out, + encoder_out_lens=encoder_out_lens, + beam=beam, + max_states=max_states, + max_contexts=max_contexts, + ) + + nbest = Nbest.from_lattice( + lattice=lattice, + num_paths=num_paths, + use_double_scores=use_double_scores, + nbest_scale=nbest_scale, + ) + + hyps = nbest.build_levenshtein_graphs() + refs = k2.levenshtein_graph(ref_texts, device=hyps.device) + + levenshtein_alignment = k2.levenshtein_alignment( + refs=refs, + hyps=hyps, + hyp_to_ref_map=nbest.shape.row_ids(1), + sorted_match_ref=True, + ) + + tot_scores = levenshtein_alignment.get_tot_scores( + use_double_scores=False, log_semiring=False + ) + ragged_tot_scores = k2.RaggedTensor(nbest.shape, tot_scores) + + max_indexes = ragged_tot_scores.argmax() + + best_path = k2.index_fsa(nbest.fsa, max_indexes) + + hyps = get_texts(best_path) + return hyps + + +def fast_beam_search( + model: Transducer, + decoding_graph: k2.Fsa, + encoder_out: torch.Tensor, + encoder_out_lens: torch.Tensor, + beam: float, + max_states: int, + max_contexts: int, +) -> k2.Fsa: + """It limits the maximum number of symbols per frame to 1. + + Args: + model: + An instance of `Transducer`. + decoding_graph: + Decoding graph used for decoding, may be a TrivialGraph or a HLG. + encoder_out: + A tensor of shape (N, T, C) from the encoder. + encoder_out_lens: + A tensor of shape (N,) containing the number of frames in `encoder_out` + before padding. + beam: + Beam value, similar to the beam used in Kaldi.. + max_states: + Max states per stream per frame. + max_contexts: + Max contexts pre stream per frame. + Returns: + Return an FsaVec with axes [utt][state][arc] containing the decoded + lattice. Note: When the input graph is a TrivialGraph, the returned + lattice is actually an acceptor. + """ assert encoder_out.ndim == 3 context_size = model.decoder.context_size @@ -103,9 +249,7 @@ def fast_beam_search( decoding_streams.terminate_and_flush_to_streams() lattice = decoding_streams.format_output(encoder_out_lens.tolist()) - best_path = one_best_decoding(lattice) - hyps = get_texts(best_path) - return hyps + return lattice def greedy_search( @@ -130,6 +274,7 @@ def greedy_search( blank_id = model.decoder.blank_id context_size = model.decoder.context_size + unk_id = getattr(model, "unk_id", blank_id) device = model.device @@ -170,7 +315,7 @@ def greedy_search( # logits is (1, 1, 1, vocab_size) y = logits.argmax().item() - if y != blank_id: + if y not in (blank_id, unk_id): hyp.append(y) decoder_input = torch.tensor( [hyp[-context_size:]], device=device @@ -211,6 +356,7 @@ def greedy_search_batch( T = encoder_out.size(1) blank_id = model.decoder.blank_id + unk_id = getattr(model, "unk_id", blank_id) context_size = model.decoder.context_size hyps = [[blank_id] * context_size for _ in range(batch_size)] @@ -239,7 +385,7 @@ def greedy_search_batch( y = logits.argmax(dim=1).tolist() emitted = False for i, v in enumerate(y): - if v != blank_id: + if v not in (blank_id, unk_id): hyps[i].append(v) emitted = True if emitted: @@ -432,6 +578,7 @@ def modified_beam_search( T = encoder_out.size(1) blank_id = model.decoder.blank_id + unk_id = getattr(model, "unk_id", blank_id) context_size = model.decoder.context_size device = model.device B = [HypothesisList() for _ in range(batch_size)] @@ -503,8 +650,10 @@ def modified_beam_search( for i in range(batch_size): topk_log_probs, topk_indexes = ragged_log_probs[i].topk(beam) - topk_hyp_indexes = (topk_indexes // vocab_size).tolist() - topk_token_indexes = (topk_indexes % vocab_size).tolist() + with warnings.catch_warnings(): + warnings.simplefilter("ignore") + topk_hyp_indexes = (topk_indexes // vocab_size).tolist() + topk_token_indexes = (topk_indexes % vocab_size).tolist() for k in range(len(topk_hyp_indexes)): hyp_idx = topk_hyp_indexes[k] @@ -512,7 +661,7 @@ def modified_beam_search( new_ys = hyp.ys[:] new_token = topk_token_indexes[k] - if new_token != blank_id: + if new_token not in (blank_id, unk_id): new_ys.append(new_token) new_log_prob = topk_log_probs[k] @@ -553,6 +702,7 @@ def _deprecated_modified_beam_search( # support only batch_size == 1 for now assert encoder_out.size(0) == 1, encoder_out.size(0) blank_id = model.decoder.blank_id + unk_id = getattr(model, "unk_id", blank_id) context_size = model.decoder.context_size device = model.device @@ -614,14 +764,16 @@ def _deprecated_modified_beam_search( topk_hyp_indexes = topk_indexes // logits.size(-1) topk_token_indexes = topk_indexes % logits.size(-1) - topk_hyp_indexes = topk_hyp_indexes.tolist() - topk_token_indexes = topk_token_indexes.tolist() + with warnings.catch_warnings(): + warnings.simplefilter("ignore") + topk_hyp_indexes = topk_hyp_indexes.tolist() + topk_token_indexes = topk_token_indexes.tolist() for i in range(len(topk_hyp_indexes)): hyp = A[topk_hyp_indexes[i]] new_ys = hyp.ys[:] new_token = topk_token_indexes[i] - if new_token != blank_id: + if new_token not in (blank_id, unk_id): new_ys.append(new_token) new_log_prob = topk_log_probs[i] new_hyp = Hypothesis(ys=new_ys, log_prob=new_log_prob) @@ -658,6 +810,7 @@ def beam_search( # support only batch_size == 1 for now assert encoder_out.size(0) == 1, encoder_out.size(0) blank_id = model.decoder.blank_id + unk_id = getattr(model, "unk_id", blank_id) context_size = model.decoder.context_size device = model.device @@ -743,7 +896,7 @@ def beam_search( # Second, process other non-blank labels values, indices = log_prob.topk(beam + 1) for i, v in zip(indices.tolist(), values.tolist()): - if i == blank_id: + if i in (blank_id, unk_id): continue new_ys = y_star.ys + [i] new_log_prob = y_star.log_prob + v diff --git a/egs/librispeech/ASR/pruned_transducer_stateless2/decode.py b/egs/librispeech/ASR/pruned_transducer_stateless2/decode.py index d581e8672..5d946003a 100755 --- a/egs/librispeech/ASR/pruned_transducer_stateless2/decode.py +++ b/egs/librispeech/ASR/pruned_transducer_stateless2/decode.py @@ -69,7 +69,7 @@ import torch.nn as nn from asr_datamodule import LibriSpeechAsrDataModule from beam_search import ( beam_search, - fast_beam_search, + fast_beam_search_one_best, greedy_search, greedy_search_batch, modified_beam_search, @@ -252,7 +252,7 @@ def decode_one_batch( hyps = [] if params.decoding_method == "fast_beam_search": - hyp_tokens = fast_beam_search( + hyp_tokens = fast_beam_search_one_best( model=model, decoding_graph=decoding_graph, encoder_out=encoder_out, diff --git a/egs/librispeech/ASR/pruned_transducer_stateless2/pretrained.py b/egs/librispeech/ASR/pruned_transducer_stateless2/pretrained.py new file mode 100755 index 000000000..bcafe68d6 --- /dev/null +++ b/egs/librispeech/ASR/pruned_transducer_stateless2/pretrained.py @@ -0,0 +1,306 @@ +#!/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. +""" +Usage: + +(1) greedy search +./pruned_transducer_stateless2/pretrained.py \ + --checkpoint ./pruned_transducer_stateless2/exp/pretrained.pt \ + --bpe-model ./data/lang_bpe_500/bpe.model \ + --method greedy_search \ + /path/to/foo.wav \ + /path/to/bar.wav \ + +(1) beam search +./pruned_transducer_stateless2/pretrained.py \ + --checkpoint ./pruned_transducer_stateless2/exp/pretrained.pt \ + --bpe-model ./data/lang_bpe_500/bpe.model \ + --method beam_search \ + --beam-size 4 \ + /path/to/foo.wav \ + /path/to/bar.wav \ + +You can also use `./pruned_transducer_stateless2/exp/epoch-xx.pt`. + +Note: ./pruned_transducer_stateless2/exp/pretrained.pt is generated by +./pruned_transducer_stateless2/export.py +""" + + +import argparse +import logging +import math +from typing import List + +import k2 +import kaldifeat +import sentencepiece as spm +import torch +import torchaudio +from beam_search import ( + beam_search, + fast_beam_search_one_best, + greedy_search, + greedy_search_batch, + modified_beam_search, +) +from torch.nn.utils.rnn import pad_sequence +from train import get_params, get_transducer_model + + +def get_parser(): + parser = argparse.ArgumentParser( + formatter_class=argparse.ArgumentDefaultsHelpFormatter + ) + + parser.add_argument( + "--checkpoint", + type=str, + required=True, + help="Path to the checkpoint. " + "The checkpoint is assumed to be saved by " + "icefall.checkpoint.save_checkpoint().", + ) + + parser.add_argument( + "--bpe-model", + type=str, + help="""Path to bpe.model. + Used only when method is ctc-decoding. + """, + ) + + parser.add_argument( + "--method", + type=str, + default="greedy_search", + help="""Possible values are: + - greedy_search + - beam_search + - modified_beam_search + - fast_beam_search + """, + ) + + parser.add_argument( + "sound_files", + type=str, + nargs="+", + help="The input sound file(s) to transcribe. " + "Supported formats are those supported by torchaudio.load(). " + "For example, wav and flac are supported. " + "The sample rate has to be 16kHz.", + ) + + parser.add_argument( + "--sample-rate", + type=int, + default=16000, + help="The sample rate of the input sound file", + ) + + parser.add_argument( + "--beam-size", + type=int, + default=4, + help="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=1, + help="""Maximum number of symbols per frame. Used only when + --method is greedy_search. + """, + ) + + return parser + + +def read_sound_files( + filenames: List[str], expected_sample_rate: float +) -> List[torch.Tensor]: + """Read a list of sound files into a list 1-D float32 torch tensors. + Args: + filenames: + A list of sound filenames. + expected_sample_rate: + The expected sample rate of the sound files. + Returns: + Return a list of 1-D float32 torch tensors. + """ + ans = [] + for f in filenames: + wave, sample_rate = torchaudio.load(f) + assert sample_rate == expected_sample_rate, ( + f"expected sample rate: {expected_sample_rate}. " + f"Given: {sample_rate}" + ) + # We use only the first channel + ans.append(wave[0]) + return ans + + +@torch.no_grad() +def main(): + parser = get_parser() + args = parser.parse_args() + + params = get_params() + + params.update(vars(args)) + + sp = spm.SentencePieceProcessor() + sp.load(params.bpe_model) + + # is defined in local/train_bpe_model.py + params.blank_id = sp.piece_to_id("") + params.unk_id = sp.piece_to_id("") + params.vocab_size = sp.get_piece_size() + + logging.info(f"{params}") + + device = torch.device("cpu") + if torch.cuda.is_available(): + device = torch.device("cuda", 0) + + logging.info(f"device: {device}") + + logging.info("Creating model") + model = get_transducer_model(params) + + num_param = sum([p.numel() for p in model.parameters()]) + logging.info(f"Number of model parameters: {num_param}") + + checkpoint = torch.load(args.checkpoint, map_location="cpu") + model.load_state_dict(checkpoint["model"], strict=False) + model.to(device) + model.eval() + model.device = device + + logging.info("Constructing Fbank computer") + opts = kaldifeat.FbankOptions() + opts.device = device + opts.frame_opts.dither = 0 + opts.frame_opts.snip_edges = False + opts.frame_opts.samp_freq = params.sample_rate + opts.mel_opts.num_bins = params.feature_dim + + fbank = kaldifeat.Fbank(opts) + + logging.info(f"Reading sound files: {params.sound_files}") + waves = read_sound_files( + filenames=params.sound_files, expected_sample_rate=params.sample_rate + ) + waves = [w.to(device) for w in waves] + + logging.info("Decoding started") + features = fbank(waves) + feature_lengths = [f.size(0) for f in features] + + features = pad_sequence( + features, batch_first=True, padding_value=math.log(1e-10) + ) + + feature_lengths = torch.tensor(feature_lengths, device=device) + + encoder_out, encoder_out_lens = model.encoder( + x=features, x_lens=feature_lengths + ) + + num_waves = encoder_out.size(0) + hyps = [] + msg = f"Using {params.method}" + if params.method == "beam_search": + msg += f" with beam size {params.beam_size}" + logging.info(msg) + + if params.method == "fast_beam_search": + decoding_graph = k2.trivial_graph(params.vocab_size - 1, device=device) + hyp_tokens = fast_beam_search_one_best( + model=model, + decoding_graph=decoding_graph, + encoder_out=encoder_out, + encoder_out_lens=encoder_out_lens, + beam=8.0, + max_contexts=32, + max_states=8, + ) + for hyp in sp.decode(hyp_tokens): + hyps.append(hyp.split()) + elif params.method == "modified_beam_search": + hyp_tokens = modified_beam_search( + model=model, + encoder_out=encoder_out, + beam=params.beam_size, + ) + + for hyp in sp.decode(hyp_tokens): + hyps.append(hyp.split()) + elif params.method == "greedy_search" and params.max_sym_per_frame == 1: + hyp_tokens = greedy_search_batch( + model=model, + encoder_out=encoder_out, + ) + for hyp in sp.decode(hyp_tokens): + hyps.append(hyp.split()) + else: + for i in range(num_waves): + # fmt: off + encoder_out_i = encoder_out[i:i+1, :encoder_out_lens[i]] + # fmt: on + if params.method == "greedy_search": + hyp = greedy_search( + model=model, + encoder_out=encoder_out_i, + max_sym_per_frame=params.max_sym_per_frame, + ) + elif params.method == "beam_search": + hyp = beam_search( + model=model, + encoder_out=encoder_out_i, + beam=params.beam_size, + ) + else: + raise ValueError(f"Unsupported method: {params.method}") + + hyps.append(sp.decode(hyp).split()) + + s = "\n" + for filename, hyp in zip(params.sound_files, hyps): + words = " ".join(hyp) + s += f"{filename}:\n{words}\n\n" + logging.info(s) + + logging.info("Decoding Done") + + +if __name__ == "__main__": + formatter = ( + "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s" + ) + + logging.basicConfig(format=formatter, level=logging.INFO) + main() diff --git a/egs/librispeech/ASR/pruned_transducer_stateless3/asr_datamodule.py b/egs/librispeech/ASR/pruned_transducer_stateless3/asr_datamodule.py new file mode 100644 index 000000000..df1e52202 --- /dev/null +++ b/egs/librispeech/ASR/pruned_transducer_stateless3/asr_datamodule.py @@ -0,0 +1,314 @@ +# 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( + "--on-the-fly-num-workers", + type=int, + default=0, + help="The number of workers for on-the-fly feature extraction", + ) + + 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( + extractor=Fbank(FbankConfig(num_mel_bins=80)), + num_workers=self.args.on_the_fly_num_workers, + ) + 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/librispeech/ASR/pruned_transducer_stateless3/beam_search.py b/egs/librispeech/ASR/pruned_transducer_stateless3/beam_search.py new file mode 120000 index 000000000..8554e44cc --- /dev/null +++ b/egs/librispeech/ASR/pruned_transducer_stateless3/beam_search.py @@ -0,0 +1 @@ +../pruned_transducer_stateless2/beam_search.py \ No newline at end of file diff --git a/egs/librispeech/ASR/pruned_transducer_stateless3/conformer.py b/egs/librispeech/ASR/pruned_transducer_stateless3/conformer.py new file mode 120000 index 000000000..3b84b9573 --- /dev/null +++ b/egs/librispeech/ASR/pruned_transducer_stateless3/conformer.py @@ -0,0 +1 @@ +../pruned_transducer_stateless2/conformer.py \ No newline at end of file diff --git a/egs/librispeech/ASR/pruned_transducer_stateless3/decode-giga.py b/egs/librispeech/ASR/pruned_transducer_stateless3/decode-giga.py new file mode 100755 index 000000000..e6a9a0aee --- /dev/null +++ b/egs/librispeech/ASR/pruned_transducer_stateless3/decode-giga.py @@ -0,0 +1,567 @@ +#!/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 +./pruned_transducer_stateless3/decode-giga.py \ + --epoch 28 \ + --avg 15 \ + --exp-dir ./pruned_transducer_stateless3/exp \ + --max-duration 100 \ + --decoding-method greedy_search + +(2) beam search +./pruned_transducer_stateless3/decode-giga.py \ + --epoch 28 \ + --avg 15 \ + --exp-dir ./pruned_transducer_stateless3/exp \ + --max-duration 100 \ + --decoding-method beam_search \ + --beam-size 4 + +(3) modified beam search +./pruned_transducer_stateless3/decode-giga.py \ + --epoch 28 \ + --avg 15 \ + --exp-dir ./pruned_transducer_stateless3/exp \ + --max-duration 100 \ + --decoding-method modified_beam_search \ + --beam-size 4 + +(4) fast beam search +./pruned_transducer_stateless3/decode-giga.py \ + --epoch 28 \ + --avg 15 \ + --exp-dir ./pruned_transducer_stateless3/exp \ + --max-duration 1500 \ + --decoding-method fast_beam_search \ + --beam 4 \ + --max-contexts 4 \ + --max-states 8 +""" + + +import argparse +import logging +from collections import defaultdict +from pathlib import Path +from typing import Dict, List, Optional, Tuple + +import k2 +import sentencepiece as spm +import torch +import torch.nn as nn +from asr_datamodule import AsrDataModule +from beam_search import ( + beam_search, + fast_beam_search, + greedy_search, + greedy_search_batch, + modified_beam_search, +) +from gigaspeech import GigaSpeech +from gigaspeech_scoring import asr_text_post_processing +from train import get_params, get_transducer_model + +from icefall.checkpoint import ( + average_checkpoints, + find_checkpoints, + load_checkpoint, +) +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=28, + help="It specifies the checkpoint to use for decoding." + "Note: Epoch counts from 0.", + ) + parser.add_argument( + "--avg", + type=int, + default=15, + help="Number of checkpoints to average. Automatically select " + "consecutive checkpoints before the checkpoint specified by " + "'--epoch'. ", + ) + + parser.add_argument( + "--avg-last-n", + type=int, + default=0, + help="""If positive, --epoch and --avg are ignored and it + will use the last n checkpoints exp_dir/checkpoint-xxx.pt + where xxx is the number of processed batches while + saving that checkpoint. + """, + ) + + parser.add_argument( + "--exp-dir", + type=str, + default="pruned_transducer_stateless3/exp", + help="The experiment dir", + ) + + parser.add_argument( + "--bpe-model", + type=str, + default="data/lang_bpe_500/bpe.model", + help="Path to the BPE model", + ) + + parser.add_argument( + "--decoding-method", + type=str, + default="greedy_search", + help="""Possible values are: + - greedy_search + - beam_search + - modified_beam_search + - fast_beam_search + """, + ) + + parser.add_argument( + "--beam-size", + type=int, + default=4, + help="""An integer indicating how many candidates we will keep for each + frame. Used only when --decoding-method is beam_search or + modified_beam_search.""", + ) + + parser.add_argument( + "--beam", + type=float, + default=4, + help="""A floating point value to calculate the cutoff score during beam + search (i.e., `cutoff = max-score - beam`), which is the same as the + `beam` in Kaldi. + Used only when --decoding-method is fast_beam_search""", + ) + + parser.add_argument( + "--max-contexts", + type=int, + default=4, + help="""Used only when --decoding-method is + fast_beam_search""", + ) + + parser.add_argument( + "--max-states", + type=int, + default=8, + help="""Used only when --decoding-method is + fast_beam_search""", + ) + + parser.add_argument( + "--context-size", + type=int, + default=2, + help="The context size in the decoder. 1 means bigram; " + "2 means tri-gram", + ) + parser.add_argument( + "--max-sym-per-frame", + type=int, + default=1, + help="""Maximum number of symbols per frame. + Used only when --decoding_method is greedy_search""", + ) + + return parser + + +def post_processing( + results: List[Tuple[List[List[str]], List[List[str]]]], +) -> List[Tuple[List[List[str]], List[List[str]]]]: + new_results = [] + for ref, hyp in results: + new_ref = asr_text_post_processing(" ".join(ref)).split() + new_hyp = asr_text_post_processing(" ".join(hyp)).split() + new_results.append((new_ref, new_hyp)) + return new_results + + +def decode_one_batch( + params: AttributeDict, + model: nn.Module, + sp: spm.SentencePieceProcessor, + batch: dict, + decoding_graph: Optional[k2.Fsa] = None, +) -> Dict[str, List[List[str]]]: + """Decode one batch and return the result in a dict. The dict has the + following format: + + - key: It indicates the setting used for decoding. For example, + if greedy_search is used, it would be "greedy_search" + If beam search with a beam size of 7 is used, it would be + "beam_7" + - value: It contains the decoding result. `len(value)` equals to + batch size. `value[i]` is the decoding result for the i-th + utterance in the given batch. + Args: + params: + It's the return value of :func:`get_params`. + model: + The neural model. + sp: + The BPE model. + batch: + It is the return value from iterating + `lhotse.dataset.K2SpeechRecognitionDataset`. See its documentation + for the format of the `batch`. + decoding_graph: + The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used + only when --decoding_method is fast_beam_search. + Returns: + Return the decoding result. See above description for the format of + the returned dict. + """ + device = 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 = [] + + if params.decoding_method == "fast_beam_search": + hyp_tokens = fast_beam_search( + model=model, + decoding_graph=decoding_graph, + encoder_out=encoder_out, + encoder_out_lens=encoder_out_lens, + beam=params.beam, + max_contexts=params.max_contexts, + max_states=params.max_states, + ) + for hyp in sp.decode(hyp_tokens): + hyps.append(hyp.split()) + elif ( + params.decoding_method == "greedy_search" + and params.max_sym_per_frame == 1 + ): + hyp_tokens = greedy_search_batch( + model=model, + encoder_out=encoder_out, + ) + for hyp in sp.decode(hyp_tokens): + hyps.append(hyp.split()) + elif params.decoding_method == "modified_beam_search": + hyp_tokens = modified_beam_search( + model=model, + encoder_out=encoder_out, + beam=params.beam_size, + ) + for hyp in sp.decode(hyp_tokens): + hyps.append(hyp.split()) + else: + batch_size = encoder_out.size(0) + + for i in range(batch_size): + # fmt: off + encoder_out_i = encoder_out[i:i+1, :encoder_out_lens[i]] + # fmt: on + if params.decoding_method == "greedy_search": + hyp = greedy_search( + model=model, + encoder_out=encoder_out_i, + max_sym_per_frame=params.max_sym_per_frame, + ) + elif params.decoding_method == "beam_search": + hyp = beam_search( + model=model, + encoder_out=encoder_out_i, + beam=params.beam_size, + ) + else: + raise ValueError( + f"Unsupported decoding method: {params.decoding_method}" + ) + hyps.append(sp.decode(hyp).split()) + + if params.decoding_method == "greedy_search": + return {"greedy_search": hyps} + elif params.decoding_method == "fast_beam_search": + return { + ( + f"beam_{params.beam}_" + f"max_contexts_{params.max_contexts}_" + f"max_states_{params.max_states}" + ): hyps + } + else: + return {f"beam_size_{params.beam_size}": hyps} + + +def decode_dataset( + dl: torch.utils.data.DataLoader, + params: AttributeDict, + model: nn.Module, + sp: spm.SentencePieceProcessor, + decoding_graph: Optional[k2.Fsa] = None, +) -> Dict[str, List[Tuple[List[str], List[str]]]]: + """Decode dataset. + + Args: + dl: + PyTorch's dataloader containing the dataset to decode. + params: + It is returned by :func:`get_params`. + model: + The neural model. + sp: + The BPE model. + decoding_graph: + The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used + only when --decoding_method is fast_beam_search. + Returns: + Return a dict, whose key may be "greedy_search" if greedy search + is used, or it may be "beam_7" if beam size of 7 is used. + Its value is a list of tuples. Each tuple contains two elements: + The first is the reference transcript, and the second is the + predicted result. + """ + num_cuts = 0 + + try: + num_batches = len(dl) + except TypeError: + num_batches = "?" + + if params.decoding_method == "greedy_search": + log_interval = 100 + else: + log_interval = 2 + + results = defaultdict(list) + for batch_idx, batch in enumerate(dl): + texts = batch["supervisions"]["text"] + + hyps_dict = decode_one_batch( + params=params, + model=model, + sp=sp, + decoding_graph=decoding_graph, + batch=batch, + ) + + for name, hyps in hyps_dict.items(): + this_batch = [] + assert len(hyps) == len(texts) + for hyp_words, ref_text in zip(hyps, texts): + ref_words = ref_text.split() + this_batch.append((ref_words, hyp_words)) + + results[name].extend(this_batch) + + num_cuts += len(texts) + + if batch_idx % log_interval == 0: + batch_str = f"{batch_idx}/{num_batches}" + + logging.info( + f"batch {batch_str}, cuts processed until now is {num_cuts}" + ) + return results + + +def save_results( + params: AttributeDict, + test_set_name: str, + results_dict: Dict[str, List[Tuple[List[str], List[str]]]], +): + test_set_wers = dict() + for key, results in results_dict.items(): + recog_path = ( + params.res_dir / f"recogs-{test_set_name}-{key}-{params.suffix}.txt" + ) + results = post_processing(results) + store_transcripts(filename=recog_path, texts=results) + logging.info(f"The transcripts are stored in {recog_path}") + + # The following prints out WERs, per-word error statistics and aligned + # ref/hyp pairs. + errs_filename = ( + params.res_dir / f"errs-{test_set_name}-{key}-{params.suffix}.txt" + ) + with open(errs_filename, "w") as f: + wer = write_error_stats( + f, f"{test_set_name}-{key}", results, enable_log=True + ) + test_set_wers[key] = wer + + logging.info("Wrote detailed error stats to {}".format(errs_filename)) + + test_set_wers = sorted(test_set_wers.items(), key=lambda x: x[1]) + errs_info = ( + params.res_dir + / f"wer-summary-{test_set_name}-{key}-{params.suffix}.txt" + ) + with open(errs_info, "w") as f: + print("settings\tWER", file=f) + for key, val in test_set_wers: + print("{}\t{}".format(key, val), file=f) + + s = "\nFor {}, WER of different settings are:\n".format(test_set_name) + note = "\tbest for {}".format(test_set_name) + for key, val in test_set_wers: + s += "{}\t{}{}\n".format(key, val, note) + note = "" + logging.info(s) + + +@torch.no_grad() +def main(): + parser = get_parser() + AsrDataModule.add_arguments(parser) + args = parser.parse_args() + args.exp_dir = Path(args.exp_dir) + + params = get_params() + params.update(vars(args)) + + assert params.decoding_method in ( + "greedy_search", + "beam_search", + "fast_beam_search", + "modified_beam_search", + ) + params.res_dir = params.exp_dir / "giga" / params.decoding_method + + params.suffix = f"epoch-{params.epoch}-avg-{params.avg}" + if "fast_beam_search" in params.decoding_method: + params.suffix += f"-beam-{params.beam}" + params.suffix += f"-max-contexts-{params.max_contexts}" + params.suffix += f"-max-states-{params.max_states}" + elif "beam_search" in params.decoding_method: + params.suffix += f"-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}") + + sp = spm.SentencePieceProcessor() + sp.load(params.bpe_model) + + # is defined in local/train_bpe_model.py + params.blank_id = sp.piece_to_id("") + params.vocab_size = sp.get_piece_size() + + logging.info(params) + + logging.info("About to create model") + model = get_transducer_model(params) + + if params.avg_last_n > 0: + filenames = find_checkpoints(params.exp_dir)[: params.avg_last_n] + logging.info(f"averaging {filenames}") + model.to(device) + model.load_state_dict(average_checkpoints(filenames, device=device)) + elif params.avg == 1: + load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model) + else: + start = params.epoch - params.avg + 1 + filenames = [] + for i in range(start, params.epoch + 1): + if start >= 0: + filenames.append(f"{params.exp_dir}/epoch-{i}.pt") + logging.info(f"averaging {filenames}") + model.to(device) + model.load_state_dict(average_checkpoints(filenames, device=device)) + + model.to(device) + model.eval() + model.device = device + + # In beam_search.py, we are using model.decoder() and model.joiner(), + # so we have to switch to the branch for the GigaSpeech dataset. + model.decoder = model.decoder_giga + model.joiner = model.joiner_giga + + if params.decoding_method == "fast_beam_search": + decoding_graph = k2.trivial_graph(params.vocab_size - 1, device=device) + else: + decoding_graph = None + + num_param = sum([p.numel() for p in model.parameters()]) + logging.info(f"Number of model parameters: {num_param}") + + asr_datamodule = AsrDataModule(args) + gigaspeech = GigaSpeech(manifest_dir=args.manifest_dir) + + test_cuts = gigaspeech.test_cuts() + dev_cuts = gigaspeech.dev_cuts() + + test_dl = asr_datamodule.test_dataloaders(test_cuts) + dev_dl = asr_datamodule.test_dataloaders(dev_cuts) + + test_sets = ["test", "dev"] + test_sets_dl = [test_dl, dev_dl] + + for test_set, dl in zip(test_sets, test_sets_dl): + results_dict = decode_dataset( + dl=dl, + params=params, + model=model, + sp=sp, + decoding_graph=decoding_graph, + ) + + save_results( + params=params, + test_set_name=test_set, + results_dict=results_dict, + ) + + logging.info("Done!") + + +if __name__ == "__main__": + main() diff --git a/egs/librispeech/ASR/pruned_transducer_stateless3/decode.py b/egs/librispeech/ASR/pruned_transducer_stateless3/decode.py new file mode 100755 index 000000000..9a6b5a117 --- /dev/null +++ b/egs/librispeech/ASR/pruned_transducer_stateless3/decode.py @@ -0,0 +1,626 @@ +#!/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 +./pruned_transducer_stateless3/decode.py \ + --epoch 28 \ + --avg 15 \ + --exp-dir ./pruned_transducer_stateless3/exp \ + --max-duration 100 \ + --decoding-method greedy_search + +(2) beam search +./pruned_transducer_stateless3/decode.py \ + --epoch 28 \ + --avg 15 \ + --exp-dir ./pruned_transducer_stateless3/exp \ + --max-duration 100 \ + --decoding-method beam_search \ + --beam-size 4 + +(3) modified beam search +./pruned_transducer_stateless3/decode.py \ + --epoch 28 \ + --avg 15 \ + --exp-dir ./pruned_transducer_stateless3/exp \ + --max-duration 100 \ + --decoding-method modified_beam_search \ + --beam-size 4 + +(4) fast beam search +./pruned_transducer_stateless3/decode.py \ + --epoch 28 \ + --avg 15 \ + --exp-dir ./pruned_transducer_stateless3/exp \ + --max-duration 1500 \ + --decoding-method fast_beam_search \ + --beam 4 \ + --max-contexts 4 \ + --max-states 8 +""" + + +import argparse +import logging +from collections import defaultdict +from pathlib import Path +from typing import Dict, List, Optional, Tuple + +import k2 +import sentencepiece as spm +import torch +import torch.nn as nn +from asr_datamodule import AsrDataModule +from beam_search import ( + beam_search, + fast_beam_search_nbest_oracle, + fast_beam_search_one_best, + greedy_search, + greedy_search_batch, + modified_beam_search, +) +from librispeech import LibriSpeech +from train import get_params, get_transducer_model + +from icefall.checkpoint import ( + average_checkpoints, + find_checkpoints, + load_checkpoint, +) +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=28, + help="""It specifies the checkpoint to use for decoding. + Note: Epoch counts from 0. + You can specify --avg to use more checkpoints for model averaging.""", + ) + + parser.add_argument( + "--iter", + type=int, + default=0, + help="""If positive, --epoch is ignored and it + will use the checkpoint exp_dir/checkpoint-iter.pt. + You can specify --avg to use more checkpoints for model averaging. + """, + ) + + parser.add_argument( + "--avg", + type=int, + default=15, + help="Number of checkpoints to average. Automatically select " + "consecutive checkpoints before the checkpoint specified by " + "'--epoch' and '--iter'", + ) + + parser.add_argument( + "--exp-dir", + type=str, + default="pruned_transducer_stateless3/exp", + help="The experiment dir", + ) + + parser.add_argument( + "--bpe-model", + type=str, + default="data/lang_bpe_500/bpe.model", + help="Path to the BPE model", + ) + + parser.add_argument( + "--decoding-method", + type=str, + default="greedy_search", + help="""Possible values are: + - greedy_search + - beam_search + - modified_beam_search + - fast_beam_search + - fast_beam_search_nbest_oracle + """, + ) + + parser.add_argument( + "--beam-size", + type=int, + default=4, + help="""An integer indicating how many candidates we will keep for each + frame. Used only when --decoding-method is beam_search or + modified_beam_search.""", + ) + + parser.add_argument( + "--beam", + type=float, + default=4, + help="""A floating point value to calculate the cutoff score during beam + search (i.e., `cutoff = max-score - beam`), which is the same as the + `beam` in Kaldi. + Used only when --decoding-method is + fast_beam_search or fast_beam_search_nbest_oracle""", + ) + + parser.add_argument( + "--max-contexts", + type=int, + default=4, + help="""Used only when --decoding-method is + fast_beam_search or fast_beam_search_nbest_oracle""", + ) + + parser.add_argument( + "--max-states", + type=int, + default=8, + help="""Used only when --decoding-method is + fast_beam_search or fast_beam_search_nbest_oracle""", + ) + + parser.add_argument( + "--context-size", + type=int, + default=2, + help="The context size in the decoder. 1 means bigram; " + "2 means tri-gram", + ) + parser.add_argument( + "--max-sym-per-frame", + type=int, + default=1, + help="""Maximum number of symbols per frame. + Used only when --decoding_method is greedy_search""", + ) + + parser.add_argument( + "--num-paths", + type=int, + default=100, + help="""Number of paths for computed nbest oracle WER + when the decoding method is fast_beam_search_nbest_oracle. + """, + ) + + parser.add_argument( + "--nbest-scale", + type=float, + default=0.5, + help="""Scale applied to lattice scores when computing nbest paths. + Used only when the decoding_method is fast_beam_search_nbest_oracle. + """, + ) + return parser + + +def decode_one_batch( + params: AttributeDict, + model: nn.Module, + sp: spm.SentencePieceProcessor, + batch: dict, + decoding_graph: Optional[k2.Fsa] = None, +) -> Dict[str, List[List[str]]]: + """Decode one batch and return the result in a dict. The dict has the + following format: + + - key: It indicates the setting used for decoding. For example, + if greedy_search is used, it would be "greedy_search" + If beam search with a beam size of 7 is used, it would be + "beam_7" + - value: It contains the decoding result. `len(value)` equals to + batch size. `value[i]` is the decoding result for the i-th + utterance in the given batch. + Args: + params: + It's the return value of :func:`get_params`. + model: + The neural model. + sp: + The BPE model. + batch: + It is the return value from iterating + `lhotse.dataset.K2SpeechRecognitionDataset`. See its documentation + for the format of the `batch`. + decoding_graph: + The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used + only when --decoding_method is + fast_beam_search or fast_beam_search_nbest_oracle. + 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 = [] + + if params.decoding_method == "fast_beam_search": + hyp_tokens = fast_beam_search_one_best( + model=model, + decoding_graph=decoding_graph, + encoder_out=encoder_out, + encoder_out_lens=encoder_out_lens, + beam=params.beam, + max_contexts=params.max_contexts, + max_states=params.max_states, + ) + for hyp in sp.decode(hyp_tokens): + hyps.append(hyp.split()) + elif params.decoding_method == "fast_beam_search_nbest_oracle": + hyp_tokens = fast_beam_search_nbest_oracle( + model=model, + decoding_graph=decoding_graph, + encoder_out=encoder_out, + encoder_out_lens=encoder_out_lens, + beam=params.beam, + max_contexts=params.max_contexts, + max_states=params.max_states, + num_paths=params.num_paths, + ref_texts=sp.encode(supervisions["text"]), + nbest_scale=params.nbest_scale, + ) + for hyp in sp.decode(hyp_tokens): + hyps.append(hyp.split()) + elif ( + params.decoding_method == "greedy_search" + and params.max_sym_per_frame == 1 + ): + hyp_tokens = greedy_search_batch( + model=model, + encoder_out=encoder_out, + ) + for hyp in sp.decode(hyp_tokens): + hyps.append(hyp.split()) + elif params.decoding_method == "modified_beam_search": + hyp_tokens = modified_beam_search( + model=model, + encoder_out=encoder_out, + beam=params.beam_size, + ) + for hyp in sp.decode(hyp_tokens): + hyps.append(hyp.split()) + else: + batch_size = encoder_out.size(0) + + for i in range(batch_size): + # fmt: off + encoder_out_i = encoder_out[i:i+1, :encoder_out_lens[i]] + # fmt: on + if params.decoding_method == "greedy_search": + hyp = greedy_search( + model=model, + encoder_out=encoder_out_i, + max_sym_per_frame=params.max_sym_per_frame, + ) + elif params.decoding_method == "beam_search": + hyp = beam_search( + model=model, + encoder_out=encoder_out_i, + beam=params.beam_size, + ) + else: + raise ValueError( + f"Unsupported decoding method: {params.decoding_method}" + ) + hyps.append(sp.decode(hyp).split()) + + if params.decoding_method == "greedy_search": + return {"greedy_search": hyps} + elif params.decoding_method == "fast_beam_search": + return { + ( + f"beam_{params.beam}_" + f"max_contexts_{params.max_contexts}_" + f"max_states_{params.max_states}" + ): hyps + } + elif params.decoding_method == "fast_beam_search_nbest_oracle": + return { + ( + f"beam_{params.beam}_" + f"max_contexts_{params.max_contexts}_" + f"max_states_{params.max_states}_" + f"num_paths_{params.num_paths}_" + f"nbest_scale_{params.nbest_scale}" + ): hyps + } + else: + return {f"beam_size_{params.beam_size}": hyps} + + +def decode_dataset( + dl: torch.utils.data.DataLoader, + params: AttributeDict, + model: nn.Module, + sp: spm.SentencePieceProcessor, + decoding_graph: Optional[k2.Fsa] = None, +) -> Dict[str, List[Tuple[List[str], List[str]]]]: + """Decode dataset. + + Args: + dl: + PyTorch's dataloader containing the dataset to decode. + params: + It is returned by :func:`get_params`. + model: + The neural model. + sp: + The BPE model. + decoding_graph: + The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used + only when --decoding_method is fast_beam_search. + Returns: + Return a dict, whose key may be "greedy_search" if greedy search + is used, or it may be "beam_7" if beam size of 7 is used. + Its value is a list of tuples. Each tuple contains two elements: + The first is the reference transcript, and the second is the + predicted result. + """ + num_cuts = 0 + + try: + num_batches = len(dl) + except TypeError: + num_batches = "?" + + if params.decoding_method == "greedy_search": + log_interval = 100 + else: + log_interval = 2 + + results = defaultdict(list) + for batch_idx, batch in enumerate(dl): + texts = batch["supervisions"]["text"] + + hyps_dict = decode_one_batch( + params=params, + model=model, + sp=sp, + decoding_graph=decoding_graph, + batch=batch, + ) + + for name, hyps in hyps_dict.items(): + this_batch = [] + assert len(hyps) == len(texts) + for hyp_words, ref_text in zip(hyps, texts): + ref_words = ref_text.split() + this_batch.append((ref_words, hyp_words)) + + results[name].extend(this_batch) + + num_cuts += len(texts) + + if batch_idx % log_interval == 0: + batch_str = f"{batch_idx}/{num_batches}" + + logging.info( + f"batch {batch_str}, cuts processed until now is {num_cuts}" + ) + return results + + +def save_results( + params: AttributeDict, + test_set_name: str, + results_dict: Dict[str, List[Tuple[List[int], List[int]]]], +): + test_set_wers = dict() + for key, results in results_dict.items(): + recog_path = ( + params.res_dir / f"recogs-{test_set_name}-{key}-{params.suffix}.txt" + ) + store_transcripts(filename=recog_path, texts=results) + logging.info(f"The transcripts are stored in {recog_path}") + + # The following prints out WERs, per-word error statistics and aligned + # ref/hyp pairs. + errs_filename = ( + params.res_dir / f"errs-{test_set_name}-{key}-{params.suffix}.txt" + ) + with open(errs_filename, "w") as f: + wer = write_error_stats( + f, f"{test_set_name}-{key}", results, enable_log=True + ) + test_set_wers[key] = wer + + logging.info("Wrote detailed error stats to {}".format(errs_filename)) + + test_set_wers = sorted(test_set_wers.items(), key=lambda x: x[1]) + errs_info = ( + params.res_dir + / f"wer-summary-{test_set_name}-{key}-{params.suffix}.txt" + ) + with open(errs_info, "w") as f: + print("settings\tWER", file=f) + for key, val in test_set_wers: + print("{}\t{}".format(key, val), file=f) + + s = "\nFor {}, WER of different settings are:\n".format(test_set_name) + note = "\tbest for {}".format(test_set_name) + for key, val in test_set_wers: + s += "{}\t{}{}\n".format(key, val, note) + note = "" + logging.info(s) + + +@torch.no_grad() +def main(): + parser = get_parser() + AsrDataModule.add_arguments(parser) + args = parser.parse_args() + args.exp_dir = Path(args.exp_dir) + + params = get_params() + params.update(vars(args)) + + assert params.decoding_method in ( + "greedy_search", + "beam_search", + "fast_beam_search", + "fast_beam_search_nbest_oracle", + "modified_beam_search", + ) + params.res_dir = params.exp_dir / params.decoding_method + + if params.iter > 0: + params.suffix = f"iter-{params.iter}-avg-{params.avg}" + else: + params.suffix = f"epoch-{params.epoch}-avg-{params.avg}" + + if params.decoding_method == "fast_beam_search": + params.suffix += f"-beam-{params.beam}" + params.suffix += f"-max-contexts-{params.max_contexts}" + params.suffix += f"-max-states-{params.max_states}" + elif params.decoding_method == "fast_beam_search_nbest_oracle": + params.suffix += f"-beam-{params.beam}" + params.suffix += f"-max-contexts-{params.max_contexts}" + params.suffix += f"-max-states-{params.max_states}" + params.suffix += f"-num-paths-{params.num_paths}" + params.suffix += f"-nbest-scale-{params.nbest_scale}" + elif "beam_search" in params.decoding_method: + params.suffix += ( + f"-{params.decoding_method}-beam-size-{params.beam_size}" + ) + else: + params.suffix += f"-context-{params.context_size}" + params.suffix += f"-max-sym-per-frame-{params.max_sym_per_frame}" + + setup_logger(f"{params.res_dir}/log-decode-{params.suffix}") + logging.info("Decoding started") + + device = torch.device("cpu") + if torch.cuda.is_available(): + device = torch.device("cuda", 0) + + logging.info(f"Device: {device}") + + sp = spm.SentencePieceProcessor() + sp.load(params.bpe_model) + + # and is defined in local/train_bpe_model.py + params.blank_id = sp.piece_to_id("") + params.unk_id = sp.unk_id() + params.vocab_size = sp.get_piece_size() + + logging.info(params) + + logging.info("About to create model") + model = get_transducer_model(params) + + if params.iter > 0: + filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[ + : params.avg + ] + if len(filenames) == 0: + raise ValueError( + f"No checkpoints found for" + f" --iter {params.iter}, --avg {params.avg}" + ) + elif len(filenames) < params.avg: + raise ValueError( + f"Not enough checkpoints ({len(filenames)}) found for" + f" --iter {params.iter}, --avg {params.avg}" + ) + logging.info(f"averaging {filenames}") + model.to(device) + model.load_state_dict(average_checkpoints(filenames, device=device)) + elif params.avg == 1: + load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model) + else: + start = params.epoch - params.avg + 1 + filenames = [] + for i in range(start, params.epoch + 1): + if start >= 0: + filenames.append(f"{params.exp_dir}/epoch-{i}.pt") + logging.info(f"averaging {filenames}") + model.to(device) + model.load_state_dict(average_checkpoints(filenames, device=device)) + + model.to(device) + model.eval() + model.device = device + model.unk_id = params.unk_id + + if params.decoding_method in ( + "fast_beam_search", + "fast_beam_search_nbest_oracle", + ): + decoding_graph = k2.trivial_graph(params.vocab_size - 1, device=device) + else: + decoding_graph = None + + num_param = sum([p.numel() for p in model.parameters()]) + logging.info(f"Number of model parameters: {num_param}") + + asr_datamodule = AsrDataModule(args) + librispeech = LibriSpeech(manifest_dir=args.manifest_dir) + + test_clean_cuts = librispeech.test_clean_cuts() + test_other_cuts = librispeech.test_other_cuts() + + test_clean_dl = asr_datamodule.test_dataloaders(test_clean_cuts) + test_other_dl = asr_datamodule.test_dataloaders(test_other_cuts) + + test_sets = ["test-clean", "test-other"] + test_dl = [test_clean_dl, test_other_dl] + + for test_set, test_dl in zip(test_sets, test_dl): + results_dict = decode_dataset( + dl=test_dl, + params=params, + model=model, + sp=sp, + decoding_graph=decoding_graph, + ) + + save_results( + params=params, + test_set_name=test_set, + results_dict=results_dict, + ) + + logging.info("Done!") + + +if __name__ == "__main__": + main() diff --git a/egs/librispeech/ASR/pruned_transducer_stateless3/decoder.py b/egs/librispeech/ASR/pruned_transducer_stateless3/decoder.py new file mode 120000 index 000000000..0793c5709 --- /dev/null +++ b/egs/librispeech/ASR/pruned_transducer_stateless3/decoder.py @@ -0,0 +1 @@ +../pruned_transducer_stateless2/decoder.py \ No newline at end of file diff --git a/egs/librispeech/ASR/pruned_transducer_stateless3/encoder_interface.py b/egs/librispeech/ASR/pruned_transducer_stateless3/encoder_interface.py new file mode 120000 index 000000000..b9aa0ae08 --- /dev/null +++ b/egs/librispeech/ASR/pruned_transducer_stateless3/encoder_interface.py @@ -0,0 +1 @@ +../pruned_transducer_stateless2/encoder_interface.py \ No newline at end of file diff --git a/egs/librispeech/ASR/pruned_transducer_stateless3/export.py b/egs/librispeech/ASR/pruned_transducer_stateless3/export.py new file mode 100755 index 000000000..29acc7181 --- /dev/null +++ b/egs/librispeech/ASR/pruned_transducer_stateless3/export.py @@ -0,0 +1,183 @@ +#!/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: +./pruned_transducer_stateless3/export.py \ + --exp-dir ./pruned_transducer_stateless3/exp \ + --bpe-model data/lang_bpe_500/bpe.model \ + --epoch 20 \ + --avg 10 + +It will generate a file exp_dir/pretrained.pt + +To use the generated file with `pruned_transducer_stateless3/decode.py`, +you can do: + + cd /path/to/exp_dir + ln -s pretrained.pt epoch-9999.pt + + cd /path/to/egs/librispeech/ASR + ./pruned_transducer_stateless3/decode.py \ + --exp-dir ./pruned_transducer_stateless3/exp \ + --epoch 9999 \ + --avg 1 \ + --max-duration 600 \ + --decoding-method greedy_search \ + --bpe-model data/lang_bpe_500/bpe.model +""" + +import argparse +import logging +from pathlib import Path + +import sentencepiece as spm +import torch +from train import get_params, get_transducer_model + +from icefall.checkpoint import average_checkpoints, load_checkpoint +from icefall.utils import str2bool + + +def get_parser(): + parser = argparse.ArgumentParser( + formatter_class=argparse.ArgumentDefaultsHelpFormatter + ) + + parser.add_argument( + "--epoch", + type=int, + default=28, + help="It specifies the checkpoint to use for decoding." + "Note: Epoch counts from 0.", + ) + + parser.add_argument( + "--avg", + type=int, + default=15, + help="Number of checkpoints to average. Automatically select " + "consecutive checkpoints before the checkpoint specified by " + "'--epoch'. ", + ) + + parser.add_argument( + "--exp-dir", + type=str, + default="pruned_transducer_stateless3/exp", + help="""It specifies the directory where all training related + files, e.g., checkpoints, log, etc, are saved + """, + ) + + parser.add_argument( + "--bpe-model", + type=str, + default="data/lang_bpe_500/bpe.model", + help="Path to the BPE model", + ) + + parser.add_argument( + "--jit", + type=str2bool, + default=False, + help="""True to save a model after applying torch.jit.script. + """, + ) + + 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 main(): + args = get_parser().parse_args() + args.exp_dir = Path(args.exp_dir) + + assert args.jit is False, "Support torchscript 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}") + + sp = spm.SentencePieceProcessor() + sp.load(params.bpe_model) + + # is defined in local/train_bpe_model.py + params.blank_id = sp.piece_to_id("") + params.vocab_size = sp.get_piece_size() + + 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)) + + model.eval() + + 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/librispeech/ASR/pruned_transducer_stateless3/gigaspeech.py b/egs/librispeech/ASR/pruned_transducer_stateless3/gigaspeech.py new file mode 100644 index 000000000..3f8bf3ba9 --- /dev/null +++ b/egs/librispeech/ASR/pruned_transducer_stateless3/gigaspeech.py @@ -0,0 +1,94 @@ +# 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 glob +import logging +import re +from pathlib import Path + +import lhotse +from lhotse import CutSet, load_manifest + + +class GigaSpeech: + def __init__(self, manifest_dir: str): + """ + Args: + manifest_dir: + It is expected to contain the following files:: + + - XL_split_2000/cuts_XL.*.jsonl.gz + - cuts_L_raw.jsonl.gz + - cuts_M_raw.jsonl.gz + - cuts_S_raw.jsonl.gz + - cuts_XS_raw.jsonl.gz + - cuts_DEV_raw.jsonl.gz + - cuts_TEST_raw.jsonl.gz + """ + self.manifest_dir = Path(manifest_dir) + + def train_XL_cuts(self) -> CutSet: + logging.info("About to get train-XL cuts") + + filenames = list( + glob.glob(f"{self.manifest_dir}/XL_split_2000/cuts_XL.*.jsonl.gz") + ) + + pattern = re.compile(r"cuts_XL.([0-9]+).jsonl.gz") + idx_filenames = [ + (int(pattern.search(f).group(1)), f) for f in filenames + ] + idx_filenames = sorted(idx_filenames, key=lambda x: x[0]) + + sorted_filenames = [f[1] for f in idx_filenames] + + logging.info(f"Loading {len(sorted_filenames)} splits") + + return lhotse.combine( + lhotse.load_manifest_lazy(p) for p in sorted_filenames + ) + + def train_L_cuts(self) -> CutSet: + f = self.manifest_dir / "cuts_L_raw.jsonl.gz" + logging.info(f"About to get train-L cuts from {f}") + return CutSet.from_jsonl_lazy(f) + + def train_M_cuts(self) -> CutSet: + f = self.manifest_dir / "cuts_M_raw.jsonl.gz" + logging.info(f"About to get train-M cuts from {f}") + return CutSet.from_jsonl_lazy(f) + + def train_S_cuts(self) -> CutSet: + f = self.manifest_dir / "cuts_S_raw.jsonl.gz" + logging.info(f"About to get train-S cuts from {f}") + return CutSet.from_jsonl_lazy(f) + + def train_XS_cuts(self) -> CutSet: + f = self.manifest_dir / "cuts_XS_raw.jsonl.gz" + logging.info(f"About to get train-XS cuts from {f}") + return CutSet.from_jsonl_lazy(f) + + def test_cuts(self) -> CutSet: + f = self.manifest_dir / "cuts_TEST.jsonl.gz" + logging.info(f"About to get TEST cuts from {f}") + return load_manifest(f) + + def dev_cuts(self) -> CutSet: + f = self.manifest_dir / "cuts_DEV.jsonl.gz" + logging.info(f"About to get DEV cuts from {f}") + return load_manifest(f) diff --git a/egs/librispeech/ASR/pruned_transducer_stateless3/gigaspeech_scoring.py b/egs/librispeech/ASR/pruned_transducer_stateless3/gigaspeech_scoring.py new file mode 120000 index 000000000..b89dd0a38 --- /dev/null +++ b/egs/librispeech/ASR/pruned_transducer_stateless3/gigaspeech_scoring.py @@ -0,0 +1 @@ +../../../gigaspeech/ASR/conformer_ctc/gigaspeech_scoring.py \ No newline at end of file diff --git a/egs/librispeech/ASR/pruned_transducer_stateless3/joiner.py b/egs/librispeech/ASR/pruned_transducer_stateless3/joiner.py new file mode 120000 index 000000000..815fd4bb6 --- /dev/null +++ b/egs/librispeech/ASR/pruned_transducer_stateless3/joiner.py @@ -0,0 +1 @@ +../pruned_transducer_stateless2/joiner.py \ No newline at end of file diff --git a/egs/librispeech/ASR/pruned_transducer_stateless3/librispeech.py b/egs/librispeech/ASR/pruned_transducer_stateless3/librispeech.py new file mode 100644 index 000000000..00b7c8334 --- /dev/null +++ b/egs/librispeech/ASR/pruned_transducer_stateless3/librispeech.py @@ -0,0 +1,74 @@ +# 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 LibriSpeech: + def __init__(self, manifest_dir: str): + """ + Args: + manifest_dir: + It is expected to contain the following files:: + + - cuts_dev-clean.json.gz + - cuts_dev-other.json.gz + - cuts_test-clean.json.gz + - cuts_test-other.json.gz + - cuts_train-clean-100.json.gz + - cuts_train-clean-360.json.gz + - cuts_train-other-500.json.gz + """ + self.manifest_dir = Path(manifest_dir) + + def train_clean_100_cuts(self) -> CutSet: + f = self.manifest_dir / "cuts_train-clean-100.json.gz" + logging.info(f"About to get train-clean-100 cuts from {f}") + return load_manifest(f) + + def train_clean_360_cuts(self) -> CutSet: + f = self.manifest_dir / "cuts_train-clean-360.json.gz" + logging.info(f"About to get train-clean-360 cuts from {f}") + return load_manifest(f) + + def train_other_500_cuts(self) -> CutSet: + f = self.manifest_dir / "cuts_train-other-500.json.gz" + logging.info(f"About to get train-other-500 cuts from {f}") + return load_manifest(f) + + def test_clean_cuts(self) -> CutSet: + f = self.manifest_dir / "cuts_test-clean.json.gz" + logging.info(f"About to get test-clean cuts from {f}") + return load_manifest(f) + + def test_other_cuts(self) -> CutSet: + f = self.manifest_dir / "cuts_test-other.json.gz" + logging.info(f"About to get test-other cuts from {f}") + return load_manifest(f) + + def dev_clean_cuts(self) -> CutSet: + f = self.manifest_dir / "cuts_dev-clean.json.gz" + logging.info(f"About to get dev-clean cuts from {f}") + return load_manifest(f) + + def dev_other_cuts(self) -> CutSet: + f = self.manifest_dir / "cuts_dev-other.json.gz" + logging.info(f"About to get dev-other cuts from {f}") + return load_manifest(f) diff --git a/egs/librispeech/ASR/pruned_transducer_stateless3/model.py b/egs/librispeech/ASR/pruned_transducer_stateless3/model.py new file mode 100644 index 000000000..5894361fc --- /dev/null +++ b/egs/librispeech/ASR/pruned_transducer_stateless3/model.py @@ -0,0 +1,235 @@ +# Copyright 2021 Xiaomi Corp. (authors: Fangjun Kuang, Wei Kang) +# +# See ../../../../LICENSE for clarification regarding multiple authors +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + + +from typing import Optional + +import k2 +import torch +import torch.nn as nn +from encoder_interface import EncoderInterface +from scaling import ScaledLinear + +from icefall.utils import add_sos + + +class Transducer(nn.Module): + """It implements https://arxiv.org/pdf/1211.3711.pdf + "Sequence Transduction with Recurrent Neural Networks" + """ + + def __init__( + self, + encoder: EncoderInterface, + decoder: nn.Module, + joiner: nn.Module, + encoder_dim: int, + decoder_dim: int, + joiner_dim: int, + vocab_size: int, + decoder_giga: Optional[nn.Module] = None, + joiner_giga: Optional[nn.Module] = None, + ): + """ + Args: + encoder: + It is the transcription network in the paper. Its accepts + two inputs: `x` of (N, T, encoder_dim) and `x_lens` of shape (N,). + It returns two tensors: `logits` of shape (N, T, encoder_dm) and + `logit_lens` of shape (N,). + decoder: + It is the prediction network in the paper. Its input shape + is (N, U) and its output shape is (N, U, decoder_dim). + It should contain one attribute: `blank_id`. + joiner: + It has two inputs with shapes: (N, T, encoder_dim) and + (N, U, decoder_dim). Its output shape is (N, T, U, vocab_size). + Note that its output contains + unnormalized probs, i.e., not processed by log-softmax. + encoder_dim: + Output dimension of the encoder network. + decoder_dim: + Output dimension of the decoder network. + joiner_dim: + Input dimension of the joiner network. + vocab_size: + Output dimension of the joiner network. + decoder_giga: + Optional. The decoder network for the GigaSpeech dataset. + joiner_giga: + Optional. The joiner network for the GigaSpeech dataset. + """ + super().__init__() + + assert isinstance(encoder, EncoderInterface), type(encoder) + assert hasattr(decoder, "blank_id") + + self.encoder = encoder + self.decoder = decoder + self.joiner = joiner + + self.decoder_giga = decoder_giga + self.joiner_giga = joiner_giga + + self.simple_am_proj = ScaledLinear( + encoder_dim, vocab_size, initial_speed=0.5 + ) + self.simple_lm_proj = ScaledLinear(decoder_dim, vocab_size) + + if decoder_giga is not None: + self.simple_am_proj_giga = ScaledLinear( + encoder_dim, vocab_size, initial_speed=0.5 + ) + self.simple_lm_proj_giga = ScaledLinear(decoder_dim, vocab_size) + + def forward( + self, + x: torch.Tensor, + x_lens: torch.Tensor, + y: k2.RaggedTensor, + libri: bool = True, + prune_range: int = 5, + am_scale: float = 0.0, + lm_scale: float = 0.0, + warmup: float = 1.0, + ) -> 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. + libri: + True to use the decoder and joiner for the LibriSpeech dataset. + False to use the decoder and joiner for the GigaSpeech dataset. + prune_range: + The prune range for rnnt loss, it means how many symbols(context) + we are considering for each frame to compute the loss. + am_scale: + The scale to smooth the loss with am (output of encoder network) + part + lm_scale: + The scale to smooth the loss with lm (output of predictor network) + part + warmup: + A value warmup >= 0 that determines which modules are active, values + warmup > 1 "are fully warmed up" and all modules will be active. + Returns: + Return the transducer loss. + + Note: + Regarding am_scale & lm_scale, it will make the loss-function one of + the form: + lm_scale * lm_probs + am_scale * am_probs + + (1-lm_scale-am_scale) * combined_probs + """ + assert x.ndim == 3, x.shape + assert x_lens.ndim == 1, x_lens.shape + assert y.num_axes == 2, y.num_axes + + assert x.size(0) == x_lens.size(0) == y.dim0 + + encoder_out, encoder_out_lens = self.encoder(x, x_lens, warmup=warmup) + assert torch.all(encoder_out_lens > 0) + + if libri: + decoder = self.decoder + simple_lm_proj = self.simple_lm_proj + simple_am_proj = self.simple_am_proj + joiner = self.joiner + else: + decoder = self.decoder_giga + simple_lm_proj = self.simple_lm_proj_giga + simple_am_proj = self.simple_am_proj_giga + joiner = self.joiner_giga + + # 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 = decoder.blank_id + sos_y = add_sos(y, sos_id=blank_id) + + # sos_y_padded: [B, S + 1], start with SOS. + sos_y_padded = sos_y.pad(mode="constant", padding_value=blank_id) + + # decoder_out: [B, S + 1, decoder_dim] + decoder_out = decoder(sos_y_padded) + + # Note: y does not start with SOS + # y_padded : [B, S] + y_padded = y.pad(mode="constant", padding_value=0) + + y_padded = y_padded.to(torch.int64) + boundary = torch.zeros( + (x.size(0), 4), dtype=torch.int64, device=x.device + ) + boundary[:, 2] = y_lens + boundary[:, 3] = encoder_out_lens + + lm = simple_lm_proj(decoder_out) + am = simple_am_proj(encoder_out) + + with torch.cuda.amp.autocast(enabled=False): + simple_loss, (px_grad, py_grad) = k2.rnnt_loss_smoothed( + lm=lm.float(), + am=am.float(), + symbols=y_padded, + termination_symbol=blank_id, + lm_only_scale=lm_scale, + am_only_scale=am_scale, + boundary=boundary, + reduction="sum", + return_grad=True, + ) + + # ranges : [B, T, prune_range] + ranges = k2.get_rnnt_prune_ranges( + px_grad=px_grad, + py_grad=py_grad, + boundary=boundary, + s_range=prune_range, + ) + + # am_pruned : [B, T, prune_range, encoder_dim] + # lm_pruned : [B, T, prune_range, decoder_dim] + am_pruned, lm_pruned = k2.do_rnnt_pruning( + am=joiner.encoder_proj(encoder_out), + lm=joiner.decoder_proj(decoder_out), + ranges=ranges, + ) + + # logits : [B, T, prune_range, vocab_size] + + # project_input=False since we applied the decoder's input projections + # prior to do_rnnt_pruning (this is an optimization for speed). + logits = joiner(am_pruned, lm_pruned, project_input=False) + + with torch.cuda.amp.autocast(enabled=False): + pruned_loss = k2.rnnt_loss_pruned( + logits=logits.float(), + symbols=y_padded, + ranges=ranges, + termination_symbol=blank_id, + boundary=boundary, + reduction="sum", + ) + + return (simple_loss, pruned_loss) diff --git a/egs/librispeech/ASR/pruned_transducer_stateless3/optim.py b/egs/librispeech/ASR/pruned_transducer_stateless3/optim.py new file mode 120000 index 000000000..e2deb4492 --- /dev/null +++ b/egs/librispeech/ASR/pruned_transducer_stateless3/optim.py @@ -0,0 +1 @@ +../pruned_transducer_stateless2/optim.py \ No newline at end of file diff --git a/egs/librispeech/ASR/pruned_transducer_stateless3/pretrained.py b/egs/librispeech/ASR/pruned_transducer_stateless3/pretrained.py new file mode 100755 index 000000000..d0fe5d24e --- /dev/null +++ b/egs/librispeech/ASR/pruned_transducer_stateless3/pretrained.py @@ -0,0 +1,306 @@ +#!/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. +""" +Usage: + +(1) greedy search +./pruned_transducer_stateless3/pretrained.py \ + --checkpoint ./pruned_transducer_stateless3/exp/pretrained.pt \ + --bpe-model ./data/lang_bpe_500/bpe.model \ + --method greedy_search \ + /path/to/foo.wav \ + /path/to/bar.wav \ + +(1) beam search +./pruned_transducer_stateless3/pretrained.py \ + --checkpoint ./pruned_transducer_stateless3/exp/pretrained.pt \ + --bpe-model ./data/lang_bpe_500/bpe.model \ + --method beam_search \ + --beam-size 4 \ + /path/to/foo.wav \ + /path/to/bar.wav \ + +You can also use `./pruned_transducer_stateless3/exp/epoch-xx.pt`. + +Note: ./pruned_transducer_stateless3/exp/pretrained.pt is generated by +./pruned_transducer_stateless3/export.py +""" + + +import argparse +import logging +import math +from typing import List + +import k2 +import kaldifeat +import sentencepiece as spm +import torch +import torchaudio +from beam_search import ( + beam_search, + fast_beam_search_one_best, + greedy_search, + greedy_search_batch, + modified_beam_search, +) +from torch.nn.utils.rnn import pad_sequence +from train import get_params, get_transducer_model + + +def get_parser(): + parser = argparse.ArgumentParser( + formatter_class=argparse.ArgumentDefaultsHelpFormatter + ) + + parser.add_argument( + "--checkpoint", + type=str, + required=True, + help="Path to the checkpoint. " + "The checkpoint is assumed to be saved by " + "icefall.checkpoint.save_checkpoint().", + ) + + parser.add_argument( + "--bpe-model", + type=str, + help="""Path to bpe.model. + Used only when method is ctc-decoding. + """, + ) + + parser.add_argument( + "--method", + type=str, + default="greedy_search", + help="""Possible values are: + - greedy_search + - beam_search + - modified_beam_search + - fast_beam_search + """, + ) + + parser.add_argument( + "sound_files", + type=str, + nargs="+", + help="The input sound file(s) to transcribe. " + "Supported formats are those supported by torchaudio.load(). " + "For example, wav and flac are supported. " + "The sample rate has to be 16kHz.", + ) + + parser.add_argument( + "--sample-rate", + type=int, + default=16000, + help="The sample rate of the input sound file", + ) + + parser.add_argument( + "--beam-size", + type=int, + default=4, + help="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=1, + help="""Maximum number of symbols per frame. Used only when + --method is greedy_search. + """, + ) + + return parser + + +def read_sound_files( + filenames: List[str], expected_sample_rate: float +) -> List[torch.Tensor]: + """Read a list of sound files into a list 1-D float32 torch tensors. + Args: + filenames: + A list of sound filenames. + expected_sample_rate: + The expected sample rate of the sound files. + Returns: + Return a list of 1-D float32 torch tensors. + """ + ans = [] + for f in filenames: + wave, sample_rate = torchaudio.load(f) + assert sample_rate == expected_sample_rate, ( + f"expected sample rate: {expected_sample_rate}. " + f"Given: {sample_rate}" + ) + # We use only the first channel + ans.append(wave[0]) + return ans + + +@torch.no_grad() +def main(): + parser = get_parser() + args = parser.parse_args() + + params = get_params() + + params.update(vars(args)) + + sp = spm.SentencePieceProcessor() + sp.load(params.bpe_model) + + # is defined in local/train_bpe_model.py + params.blank_id = sp.piece_to_id("") + params.unk_id = sp.piece_to_id("") + params.vocab_size = sp.get_piece_size() + + logging.info(f"{params}") + + device = torch.device("cpu") + if torch.cuda.is_available(): + device = torch.device("cuda", 0) + + logging.info(f"device: {device}") + + logging.info("Creating model") + model = get_transducer_model(params) + + num_param = sum([p.numel() for p in model.parameters()]) + logging.info(f"Number of model parameters: {num_param}") + + checkpoint = torch.load(args.checkpoint, map_location="cpu") + model.load_state_dict(checkpoint["model"], strict=False) + model.to(device) + model.eval() + model.device = device + + logging.info("Constructing Fbank computer") + opts = kaldifeat.FbankOptions() + opts.device = device + opts.frame_opts.dither = 0 + opts.frame_opts.snip_edges = False + opts.frame_opts.samp_freq = params.sample_rate + opts.mel_opts.num_bins = params.feature_dim + + fbank = kaldifeat.Fbank(opts) + + logging.info(f"Reading sound files: {params.sound_files}") + waves = read_sound_files( + filenames=params.sound_files, expected_sample_rate=params.sample_rate + ) + waves = [w.to(device) for w in waves] + + logging.info("Decoding started") + features = fbank(waves) + feature_lengths = [f.size(0) for f in features] + + features = pad_sequence( + features, batch_first=True, padding_value=math.log(1e-10) + ) + + feature_lengths = torch.tensor(feature_lengths, device=device) + + encoder_out, encoder_out_lens = model.encoder( + x=features, x_lens=feature_lengths + ) + + num_waves = encoder_out.size(0) + hyps = [] + msg = f"Using {params.method}" + if params.method == "beam_search": + msg += f" with beam size {params.beam_size}" + logging.info(msg) + + if params.method == "fast_beam_search": + decoding_graph = k2.trivial_graph(params.vocab_size - 1, device=device) + hyp_tokens = fast_beam_search_one_best( + model=model, + decoding_graph=decoding_graph, + encoder_out=encoder_out, + encoder_out_lens=encoder_out_lens, + beam=8.0, + max_contexts=32, + max_states=8, + ) + for hyp in sp.decode(hyp_tokens): + hyps.append(hyp.split()) + elif params.method == "modified_beam_search": + hyp_tokens = modified_beam_search( + model=model, + encoder_out=encoder_out, + beam=params.beam_size, + ) + + for hyp in sp.decode(hyp_tokens): + hyps.append(hyp.split()) + elif params.method == "greedy_search" and params.max_sym_per_frame == 1: + hyp_tokens = greedy_search_batch( + model=model, + encoder_out=encoder_out, + ) + for hyp in sp.decode(hyp_tokens): + hyps.append(hyp.split()) + else: + for i in range(num_waves): + # fmt: off + encoder_out_i = encoder_out[i:i+1, :encoder_out_lens[i]] + # fmt: on + if params.method == "greedy_search": + hyp = greedy_search( + model=model, + encoder_out=encoder_out_i, + max_sym_per_frame=params.max_sym_per_frame, + ) + elif params.method == "beam_search": + hyp = beam_search( + model=model, + encoder_out=encoder_out_i, + beam=params.beam_size, + ) + else: + raise ValueError(f"Unsupported method: {params.method}") + + hyps.append(sp.decode(hyp).split()) + + s = "\n" + for filename, hyp in zip(params.sound_files, hyps): + words = " ".join(hyp) + s += f"{filename}:\n{words}\n\n" + logging.info(s) + + logging.info("Decoding Done") + + +if __name__ == "__main__": + formatter = ( + "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s" + ) + + logging.basicConfig(format=formatter, level=logging.INFO) + main() diff --git a/egs/librispeech/ASR/pruned_transducer_stateless3/scaling.py b/egs/librispeech/ASR/pruned_transducer_stateless3/scaling.py new file mode 120000 index 000000000..09d802cc4 --- /dev/null +++ b/egs/librispeech/ASR/pruned_transducer_stateless3/scaling.py @@ -0,0 +1 @@ +../pruned_transducer_stateless2/scaling.py \ No newline at end of file diff --git a/egs/librispeech/ASR/pruned_transducer_stateless3/train.py b/egs/librispeech/ASR/pruned_transducer_stateless3/train.py new file mode 100755 index 000000000..4966ea57f --- /dev/null +++ b/egs/librispeech/ASR/pruned_transducer_stateless3/train.py @@ -0,0 +1,1125 @@ +#!/usr/bin/env python3 +# Copyright 2021 Xiaomi Corp. (authors: Fangjun Kuang, +# Wei Kang +# Mingshuang Luo) +# +# See ../../../../LICENSE for clarification regarding multiple authors +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +""" +Usage: + +export CUDA_VISIBLE_DEVICES="0,1,2,3" + +cd egs/librispeech/ASR/ +./prepare.sh +./prepare_giga_speech.sh + +./pruned_transducer_stateless3/train.py \ + --world-size 4 \ + --num-epochs 30 \ + --start-epoch 0 \ + --exp-dir pruned_transducer_stateless3/exp \ + --full-libri 1 \ + --max-duration 300 + +# For mix precision training: + +./pruned_transducer_stateless3/train.py \ + --world-size 4 \ + --num-epochs 30 \ + --start-epoch 0 \ + --use_fp16 1 \ + --exp-dir pruned_transducer_stateless3/exp \ + --full-libri 1 \ + --max-duration 550 + +""" + + +import argparse +import logging +import random +import warnings +from pathlib import Path +from shutil import copyfile +from typing import Any, Dict, Optional, Tuple, Union + +import k2 +import optim +import sentencepiece as spm +import torch +import torch.multiprocessing as mp +import torch.nn as nn +from asr_datamodule import AsrDataModule +from conformer import Conformer +from decoder import Decoder +from gigaspeech import GigaSpeech +from joiner import Joiner +from lhotse import CutSet, load_manifest +from lhotse.cut import Cut +from lhotse.dataset.sampling.base import CutSampler +from lhotse.utils import fix_random_seed +from librispeech import LibriSpeech +from model import Transducer +from optim import Eden, Eve +from torch import Tensor +from torch.cuda.amp import GradScaler +from torch.nn.parallel import DistributedDataParallel as DDP +from torch.utils.tensorboard import SummaryWriter + +from icefall import diagnostics +from icefall.checkpoint import load_checkpoint, remove_checkpoints +from icefall.checkpoint import save_checkpoint as save_checkpoint_impl +from icefall.checkpoint import save_checkpoint_with_global_batch_idx +from icefall.dist import cleanup_dist, setup_dist +from icefall.env import get_env_info +from icefall.utils import AttributeDict, MetricsTracker, setup_logger, str2bool + +LRSchedulerType = Union[ + torch.optim.lr_scheduler._LRScheduler, optim.LRScheduler +] + + +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( + "--full-libri", + type=str2bool, + default=True, + help="When enabled, use 960h LibriSpeech. " + "Otherwise, use 100h subset.", + ) + + 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_stateless3/exp/epoch-{start_epoch-1}.pt + """, + ) + + parser.add_argument( + "--start-batch", + type=int, + default=0, + help="""If positive, --start-epoch is ignored and + it loads the checkpoint from exp-dir/checkpoint-{start_batch}.pt + """, + ) + + parser.add_argument( + "--exp-dir", + type=str, + default="pruned_transducer_stateless3/exp", + help="""The experiment dir. + It specifies the directory where all training related + files, e.g., checkpoints, log, etc, are saved + """, + ) + + parser.add_argument( + "--bpe-model", + type=str, + default="data/lang_bpe_500/bpe.model", + help="Path to the BPE model", + ) + + parser.add_argument( + "--initial-lr", + type=float, + default=0.003, + help="The initial learning rate. This value should not need " + "to be changed.", + ) + + parser.add_argument( + "--lr-batches", + type=float, + default=5000, + help="""Number of steps that affects how rapidly the learning rate decreases. + We suggest not to change this.""", + ) + + parser.add_argument( + "--lr-epochs", + type=float, + default=4, + help="""Number of epochs that affects how rapidly the learning rate decreases. + """, + ) + + parser.add_argument( + "--context-size", + type=int, + default=2, + help="The context size in the decoder. 1 means bigram; " + "2 means tri-gram", + ) + + parser.add_argument( + "--prune-range", + type=int, + default=5, + help="The prune range for rnnt loss, it means how many symbols(context)" + "we are using to compute the loss", + ) + + parser.add_argument( + "--lm-scale", + type=float, + default=0.25, + help="The scale to smooth the loss with lm " + "(output of prediction network) part.", + ) + + parser.add_argument( + "--am-scale", + type=float, + default=0.0, + help="The scale to smooth the loss with am (output of encoder network)" + "part.", + ) + + parser.add_argument( + "--simple-loss-scale", + type=float, + default=0.5, + help="To get pruning ranges, we will calculate a simple version" + "loss(joiner is just addition), this simple loss also uses for" + "training (as a regularization item). We will scale the simple loss" + "with this parameter before adding to the final loss.", + ) + + parser.add_argument( + "--seed", + type=int, + default=42, + help="The seed for random generators intended for reproducibility", + ) + + parser.add_argument( + "--print-diagnostics", + type=str2bool, + default=False, + help="Accumulate stats on activations, print them and exit.", + ) + + parser.add_argument( + "--save-every-n", + type=int, + default=8000, + help="""Save checkpoint after processing this number of batches" + periodically. We save checkpoint to exp-dir/ whenever + params.batch_idx_train % save_every_n == 0. The checkpoint filename + has the form: f'exp-dir/checkpoint-{params.batch_idx_train}.pt' + Note: It also saves checkpoint to `exp-dir/epoch-xxx.pt` at the + end of each epoch where `xxx` is the epoch number counting from 0. + """, + ) + + parser.add_argument( + "--keep-last-k", + type=int, + default=20, + help="""Only keep this number of checkpoints on disk. + For instance, if it is 3, there are only 3 checkpoints + in the exp-dir with filenames `checkpoint-xxx.pt`. + It does not affect checkpoints with name `epoch-xxx.pt`. + """, + ) + + parser.add_argument( + "--use-fp16", + type=str2bool, + default=False, + help="Whether to use half precision training.", + ) + + parser.add_argument( + "--giga-prob", + type=float, + default=0.5, + help="The probability to select a batch from the GigaSpeech 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. + + - encoder_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": 3000, # For the 100h subset, use 800 + # parameters for conformer + "feature_dim": 80, + "subsampling_factor": 4, + "encoder_dim": 512, + "nhead": 8, + "dim_feedforward": 2048, + "num_encoder_layers": 12, + # parameters for decoder + "decoder_dim": 512, + # parameters for joiner + "joiner_dim": 512, + # parameters for Noam + "model_warm_step": 3000, # arg given to model, not for lrate + "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, + subsampling_factor=params.subsampling_factor, + d_model=params.encoder_dim, + nhead=params.nhead, + dim_feedforward=params.dim_feedforward, + num_encoder_layers=params.num_encoder_layers, + ) + return encoder + + +def get_decoder_model(params: AttributeDict) -> nn.Module: + decoder = Decoder( + vocab_size=params.vocab_size, + decoder_dim=params.decoder_dim, + blank_id=params.blank_id, + context_size=params.context_size, + ) + return decoder + + +def get_joiner_model(params: AttributeDict) -> nn.Module: + joiner = Joiner( + encoder_dim=params.encoder_dim, + decoder_dim=params.decoder_dim, + joiner_dim=params.joiner_dim, + vocab_size=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_giga = get_decoder_model(params) + joiner_giga = get_joiner_model(params) + + model = Transducer( + encoder=encoder, + decoder=decoder, + joiner=joiner, + decoder_giga=decoder_giga, + joiner_giga=joiner_giga, + encoder_dim=params.encoder_dim, + decoder_dim=params.decoder_dim, + joiner_dim=params.joiner_dim, + vocab_size=params.vocab_size, + ) + return model + + +def load_checkpoint_if_available( + params: AttributeDict, + model: nn.Module, + optimizer: Optional[torch.optim.Optimizer] = None, + scheduler: Optional[LRSchedulerType] = None, +) -> Optional[Dict[str, Any]]: + """Load checkpoint from file. + + If params.start_batch is positive, it will load the checkpoint from + `params.exp_dir/checkpoint-{params.start_batch}.pt`. Otherwise, if + params.start_epoch is positive, it will load the checkpoint from + `params.start_epoch - 1`. + + Apart from loading state dict for `model` and `optimizer` it also updates + `best_train_epoch`, `best_train_loss`, `best_valid_epoch`, + and `best_valid_loss` in `params`. + + Args: + params: + The return value of :func:`get_params`. + model: + The training model. + optimizer: + The optimizer that we are using. + scheduler: + The scheduler that we are using. + Returns: + Return a dict containing previously saved training info. + """ + if params.start_batch > 0: + filename = params.exp_dir / f"checkpoint-{params.start_batch}.pt" + elif params.start_epoch > 0: + filename = params.exp_dir / f"epoch-{params.start_epoch-1}.pt" + else: + return None + + assert filename.is_file(), f"{filename} does not exist!" + + saved_params = load_checkpoint( + filename, + model=model, + optimizer=optimizer, + scheduler=scheduler, + ) + + keys = [ + "best_train_epoch", + "best_valid_epoch", + "batch_idx_train", + "best_train_loss", + "best_valid_loss", + ] + for k in keys: + params[k] = saved_params[k] + + if params.start_batch > 0: + if "cur_epoch" in saved_params: + params["start_epoch"] = saved_params["cur_epoch"] + + return saved_params + + +def save_checkpoint( + params: AttributeDict, + model: nn.Module, + optimizer: Optional[torch.optim.Optimizer] = None, + scheduler: Optional[LRSchedulerType] = None, + sampler: Optional[CutSampler] = None, + scaler: Optional[GradScaler] = None, + rank: int = 0, +) -> None: + """Save model, optimizer, scheduler and training stats to file. + + Args: + params: + It is returned by :func:`get_params`. + model: + The training model. + optimizer: + The optimizer used in the training. + sampler: + The sampler for the training dataset. + scaler: + The scaler used for mix precision training. + """ + if rank != 0: + return + filename = params.exp_dir / f"epoch-{params.cur_epoch}.pt" + save_checkpoint_impl( + filename=filename, + model=model, + params=params, + optimizer=optimizer, + scheduler=scheduler, + sampler=sampler, + scaler=scaler, + rank=rank, + ) + + if params.best_train_epoch == params.cur_epoch: + best_train_filename = params.exp_dir / "best-train-loss.pt" + copyfile(src=filename, dst=best_train_filename) + + if params.best_valid_epoch == params.cur_epoch: + best_valid_filename = params.exp_dir / "best-valid-loss.pt" + copyfile(src=filename, dst=best_valid_filename) + + +def is_libri(c: Cut) -> bool: + """Return True if this cut is from the LibriSpeech dataset. + + Note: + During data preparation, we set the custom field in + the supervision segment of GigaSpeech to dict(origin='giga') + See ../local/preprocess_gigaspeech.py. + """ + return c.supervisions[0].custom is None + + +def compute_loss( + params: AttributeDict, + model: nn.Module, + sp: spm.SentencePieceProcessor, + batch: dict, + is_training: bool, + warmup: float = 1.0, +) -> 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. + warmup: a floating point value which increases throughout training; + values >= 1.0 are fully warmed up and have all modules present. + """ + 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) + + libri = is_libri(supervisions["cut"][0]) + + texts = batch["supervisions"]["text"] + y = sp.encode(texts, out_type=int) + y = k2.RaggedTensor(y).to(device) + + with torch.set_grad_enabled(is_training): + simple_loss, pruned_loss = model( + x=feature, + x_lens=feature_lens, + y=y, + libri=libri, + prune_range=params.prune_range, + am_scale=params.am_scale, + lm_scale=params.lm_scale, + warmup=warmup, + ) + # after the main warmup step, we keep pruned_loss_scale small + # for the same amount of time (model_warm_step), to avoid + # overwhelming the simple_loss and causing it to diverge, + # in case it had not fully learned the alignment yet. + pruned_loss_scale = ( + 0.0 + if warmup < 1.0 + else (0.1 if warmup > 1.0 and warmup < 2.0 else 1.0) + ) + loss = ( + params.simple_loss_scale * simple_loss + + pruned_loss_scale * pruned_loss + ) + + assert loss.requires_grad == is_training + + info = MetricsTracker() + with warnings.catch_warnings(): + warnings.simplefilter("ignore") + info["frames"] = ( + (feature_lens // params.subsampling_factor).sum().item() + ) + + # Note: We use reduction=sum while computing the loss. + info["loss"] = loss.detach().cpu().item() + info["simple_loss"] = simple_loss.detach().cpu().item() + info["pruned_loss"] = pruned_loss.detach().cpu().item() + + return loss, info + + +def compute_validation_loss( + params: AttributeDict, + model: nn.Module, + sp: spm.SentencePieceProcessor, + valid_dl: torch.utils.data.DataLoader, + world_size: int = 1, +) -> MetricsTracker: + """Run the validation process.""" + model.eval() + + tot_loss = MetricsTracker() + + for batch_idx, batch in enumerate(valid_dl): + loss, loss_info = compute_loss( + params=params, + model=model, + sp=sp, + batch=batch, + is_training=False, + ) + assert loss.requires_grad is False + tot_loss = tot_loss + loss_info + + if world_size > 1: + tot_loss.reduce(loss.device) + + loss_value = tot_loss["loss"] / tot_loss["frames"] + if loss_value < params.best_valid_loss: + params.best_valid_epoch = params.cur_epoch + params.best_valid_loss = loss_value + + return tot_loss + + +def train_one_epoch( + params: AttributeDict, + model: nn.Module, + optimizer: torch.optim.Optimizer, + scheduler: LRSchedulerType, + sp: spm.SentencePieceProcessor, + train_dl: torch.utils.data.DataLoader, + giga_train_dl: torch.utils.data.DataLoader, + valid_dl: torch.utils.data.DataLoader, + rng: random.Random, + scaler: GradScaler, + tb_writer: Optional[SummaryWriter] = None, + world_size: int = 1, + rank: int = 0, +) -> None: + """Train the model for one epoch. + + The training loss from the mean of all frames is saved in + `params.train_loss`. It runs the validation process every + `params.valid_interval` batches. + + Args: + params: + It is returned by :func:`get_params`. + model: + The model for training. + optimizer: + The optimizer we are using. + scheduler: + The learning rate scheduler, we call step() every step. + train_dl: + Dataloader for the training dataset. + giga_train_dl: + Dataloader for the GigaSpeech training dataset. + valid_dl: + Dataloader for the validation dataset. + rng: + For selecting which dataset to use. + scaler: + The scaler used for mix precision training. + tb_writer: + Writer to write log messages to tensorboard. + world_size: + Number of nodes in DDP training. If it is 1, DDP is disabled. + rank: + The rank of the node in DDP training. If no DDP is used, it should + be set to 0. + """ + model.train() + + libri_tot_loss = MetricsTracker() + giga_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.giga_prob, params.giga_prob] + + iter_libri = iter(train_dl) + iter_giga = iter(giga_train_dl) + + batch_idx = 0 + + while True: + idx = rng.choices((0, 1), weights=dl_weights, k=1)[0] + dl = iter_libri if idx == 0 else iter_giga + + try: + batch = next(dl) + except StopIteration: + name = "libri" if idx == 0 else "giga" + logging.info(f"{name} reaches end of dataloader") + break + + batch_idx += 1 + + params.batch_idx_train += 1 + batch_size = len(batch["supervisions"]["text"]) + + libri = is_libri(batch["supervisions"]["cut"][0]) + + with torch.cuda.amp.autocast(enabled=params.use_fp16): + loss, loss_info = compute_loss( + params=params, + model=model, + sp=sp, + batch=batch, + is_training=True, + warmup=(params.batch_idx_train / params.model_warm_step), + ) + # summary stats + tot_loss = (tot_loss * (1 - 1 / params.reset_interval)) + loss_info + + if libri: + libri_tot_loss = ( + libri_tot_loss * (1 - 1 / params.reset_interval) + ) + loss_info + prefix = "libri" # for logging only + else: + giga_tot_loss = ( + giga_tot_loss * (1 - 1 / params.reset_interval) + ) + loss_info + prefix = "giga" + + # NOTE: We use reduction==sum and loss is computed over utterances + # in the batch and there is no normalization to it so far. + scaler.scale(loss).backward() + scheduler.step_batch(params.batch_idx_train) + scaler.step(optimizer) + scaler.update() + optimizer.zero_grad() + + if params.print_diagnostics and batch_idx == 5: + return + + if ( + params.batch_idx_train > 0 + and params.batch_idx_train % params.save_every_n == 0 + ): + save_checkpoint_with_global_batch_idx( + out_dir=params.exp_dir, + global_batch_idx=params.batch_idx_train, + model=model, + params=params, + optimizer=optimizer, + scheduler=scheduler, + sampler=train_dl.sampler, + scaler=scaler, + rank=rank, + ) + remove_checkpoints( + out_dir=params.exp_dir, + topk=params.keep_last_k, + rank=rank, + ) + + if batch_idx % params.log_interval == 0: + cur_lr = scheduler.get_last_lr()[0] + logging.info( + f"Epoch {params.cur_epoch}, " + f"batch {batch_idx}, {prefix}_loss[{loss_info}], " + f"tot_loss[{tot_loss}], " + f"libri_tot_loss[{libri_tot_loss}], " + f"giga_tot_loss[{giga_tot_loss}], " + f"batch size: {batch_size}" + f"lr: {cur_lr:.2e}" + ) + + if tb_writer is not None: + tb_writer.add_scalar( + "train/learning_rate", cur_lr, params.batch_idx_train + ) + + loss_info.write_summary( + tb_writer, + f"train/current_{prefix}_", + params.batch_idx_train, + ) + tot_loss.write_summary( + tb_writer, "train/tot_", params.batch_idx_train + ) + libri_tot_loss.write_summary( + tb_writer, "train/libri_tot_", params.batch_idx_train + ) + giga_tot_loss.write_summary( + tb_writer, "train/giga_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, + sp=sp, + 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 20 seconds + # + # Caution: There is a reason to select 20.0 here. Please see + # ../local/display_manifest_statistics.py + # + # You should use ../local/display_manifest_statistics.py to get + # an utterance duration distribution for your dataset to select + # the threshold + return 1.0 <= c.duration <= 20.0 + + cuts = cuts.filter(remove_short_and_long_utt) + + 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)) + if params.full_libri is False: + params.valid_interval = 1600 + + fix_random_seed(params.seed) + rng = random.Random(params.seed) + if world_size > 1: + setup_dist(rank, world_size, params.master_port) + + setup_logger(f"{params.exp_dir}/log/log-train") + logging.info("Training started") + + if args.tensorboard and rank == 0: + tb_writer = SummaryWriter(log_dir=f"{params.exp_dir}/tensorboard") + else: + tb_writer = None + + device = torch.device("cpu") + if torch.cuda.is_available(): + device = torch.device("cuda", rank) + logging.info(f"Device: {device}") + + sp = spm.SentencePieceProcessor() + sp.load(params.bpe_model) + + # is defined in local/train_bpe_model.py + params.blank_id = sp.piece_to_id("") + params.vocab_size = sp.get_piece_size() + + 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 = Eve(model.parameters(), lr=params.initial_lr) + + scheduler = Eden(optimizer, params.lr_batches, params.lr_epochs) + + if checkpoints and "optimizer" in checkpoints: + logging.info("Loading optimizer state dict") + optimizer.load_state_dict(checkpoints["optimizer"]) + + if ( + checkpoints + and "scheduler" in checkpoints + and checkpoints["scheduler"] is not None + ): + logging.info("Loading scheduler state dict") + scheduler.load_state_dict(checkpoints["scheduler"]) + + if params.print_diagnostics: + opts = diagnostics.TensorDiagnosticOptions( + 2 ** 22 + ) # allow 4 megabytes per sub-module + diagnostic = diagnostics.attach_diagnostics(model, opts) + + librispeech = LibriSpeech(manifest_dir=args.manifest_dir) + + train_cuts = librispeech.train_clean_100_cuts() + if params.full_libri: + train_cuts += librispeech.train_clean_360_cuts() + train_cuts += librispeech.train_other_500_cuts() + + train_cuts = filter_short_and_long_utterances(train_cuts) + + gigaspeech = GigaSpeech(manifest_dir=args.manifest_dir) + # XL 10k hours + # L 2.5k hours + # M 1k hours + # S 250 hours + # XS 10 hours + # DEV 12 hours + # Test 40 hours + if params.full_libri: + logging.info("Using the XL subset of GigaSpeech (10k hours)") + train_giga_cuts = gigaspeech.train_XL_cuts() + else: + logging.info("Using the S subset of GigaSpeech (250 hours)") + train_giga_cuts = gigaspeech.train_S_cuts() + + train_giga_cuts = filter_short_and_long_utterances(train_giga_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, + ) + + giga_train_dl = asr_datamodule.train_dataloaders( + train_giga_cuts, + dynamic_bucketing=True, + on_the_fly_feats=False, + cuts_musan=cuts_musan, + ) + + valid_cuts = librispeech.dev_clean_cuts() + valid_cuts += librispeech.dev_other_cuts() + valid_dl = asr_datamodule.valid_dataloaders(valid_cuts) + + # It's time consuming to include `giga_train_dl` here + # for dl in [train_dl, giga_train_dl]: + for dl in [train_dl]: + scan_pessimistic_batches_for_oom( + model=model, + train_dl=dl, + optimizer=optimizer, + sp=sp, + params=params, + ) + + scaler = GradScaler(enabled=params.use_fp16) + if checkpoints and "grad_scaler" in checkpoints: + logging.info("Loading grad scaler state dict") + scaler.load_state_dict(checkpoints["grad_scaler"]) + + for epoch in range(params.start_epoch, params.num_epochs): + scheduler.step_epoch(epoch) + fix_random_seed(params.seed + epoch) + train_dl.sampler.set_epoch(epoch) + + if tb_writer is not None: + tb_writer.add_scalar("train/epoch", epoch, params.batch_idx_train) + + params.cur_epoch = epoch + + train_one_epoch( + params=params, + model=model, + optimizer=optimizer, + scheduler=scheduler, + sp=sp, + train_dl=train_dl, + giga_train_dl=giga_train_dl, + valid_dl=valid_dl, + rng=rng, + scaler=scaler, + tb_writer=tb_writer, + world_size=world_size, + rank=rank, + ) + + if params.print_diagnostics: + diagnostic.print_diagnostics() + break + + save_checkpoint( + params=params, + model=model, + optimizer=optimizer, + scheduler=scheduler, + sampler=train_dl.sampler, + scaler=scaler, + rank=rank, + ) + + logging.info("Done!") + + if world_size > 1: + torch.distributed.barrier() + cleanup_dist() + + +def scan_pessimistic_batches_for_oom( + model: nn.Module, + train_dl: torch.utils.data.DataLoader, + optimizer: torch.optim.Optimizer, + sp: spm.SentencePieceProcessor, + params: AttributeDict, +): + from lhotse.dataset import find_pessimistic_batches + + logging.info( + "Sanity check -- see if any of the batches in epoch 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: + # warmup = 0.0 is so that the derivs for the pruned loss stay zero + # (i.e. are not remembered by the decaying-average in adam), because + # we want to avoid these params being subject to shrinkage in adam. + with torch.cuda.amp.autocast(enabled=params.use_fp16): + loss, _ = compute_loss( + params=params, + model=model, + sp=sp, + batch=batch, + is_training=True, + warmup=0.0, + ) + loss.backward() + optimizer.step() + optimizer.zero_grad() + 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) + + assert 0 <= args.giga_prob < 1, args.giga_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/requirements-ci.txt b/requirements-ci.txt index 7fb4b1665..4f507285b 100644 --- a/requirements-ci.txt +++ b/requirements-ci.txt @@ -11,7 +11,7 @@ graphviz==0.19.1 -f https://download.pytorch.org/whl/cpu/torch_stable.html torch==1.10.0+cpu -f https://download.pytorch.org/whl/cpu/torch_stable.html torchaudio==0.10.0+cpu --f https://k2-fsa.org/nightly/ k2==1.14.dev20220316+cpu.torch1.10.0 +-f https://k2-fsa.org/nightly/ k2==1.15.1.dev20220426+cpu.torch1.10.0 git+https://github.com/lhotse-speech/lhotse kaldilm==1.11