diff --git a/.github/scripts/run-gigaspeech-pruned-transducer-stateless2-2022-05-12.sh b/.github/scripts/run-gigaspeech-pruned-transducer-stateless2-2022-05-12.sh index c8d9c6b77..b61a9d7b6 100755 --- a/.github/scripts/run-gigaspeech-pruned-transducer-stateless2-2022-05-12.sh +++ b/.github/scripts/run-gigaspeech-pruned-transducer-stateless2-2022-05-12.sh @@ -29,6 +29,9 @@ if [[ x"${GITHUB_EVENT_NAME}" == x"schedule" || x"${GITHUB_EVENT_LABEL_NAME}" == ls -lh data/fbank ls -lh pruned_transducer_stateless2/exp + ln -s data/fbank/cuts_DEV.jsonl.gz data/fbank/gigaspeech_cuts_DEV.jsonl.gz + ln -s data/fbank/cuts_TEST.jsonl.gz data/fbank/gigaspeech_cuts_TEST.jsonl.gz + log "Decoding dev and test" # use a small value for decoding with CPU diff --git a/.github/scripts/run-multi-zh_hans-zipformer.sh b/.github/scripts/run-multi-zh_hans-zipformer.sh new file mode 100755 index 000000000..2bc3137d8 --- /dev/null +++ b/.github/scripts/run-multi-zh_hans-zipformer.sh @@ -0,0 +1,51 @@ +#!/usr/bin/env bash + +set -e + +log() { + # This function is from espnet + local fname=${BASH_SOURCE[1]##*/} + echo -e "$(date '+%Y-%m-%d %H:%M:%S') (${fname}:${BASH_LINENO[0]}:${FUNCNAME[1]}) $*" +} + +cd egs/multi_zh-hans/ASR + +repo_url=https://huggingface.co/zrjin/icefall-asr-multi-zh-hans-zipformer-2023-9-2/ + +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/ +ls -lh $repo/test_wavs/*.wav + +pushd $repo/exp +ln -s epoch-20.pt epoch-99.pt +popd + +ls -lh $repo/exp/*.pt + + +./zipformer/pretrained.py \ + --checkpoint $repo/exp/epoch-99.pt \ + --tokens $repo/data/lang_bpe_2000/tokens.txt \ + --method greedy_search \ +$repo/test_wavs/DEV_T0000000000.wav \ +$repo/test_wavs/DEV_T0000000001.wav \ +$repo/test_wavs/DEV_T0000000002.wav + +for method in modified_beam_search fast_beam_search; do + log "$method" + + ./zipformer/pretrained.py \ + --method $method \ + --beam-size 4 \ + --checkpoint $repo/exp/epoch-99.pt \ + --tokens $repo/data/lang_bpe_2000/tokens.txt \ + $repo/test_wavs/DEV_T0000000000.wav \ + $repo/test_wavs/DEV_T0000000001.wav \ + $repo/test_wavs/DEV_T0000000002.wav +done diff --git a/.github/workflows/run-aishell-2022-06-20.yml b/.github/workflows/run-aishell-2022-06-20.yml index d14196f38..53fcb2c03 100644 --- a/.github/workflows/run-aishell-2022-06-20.yml +++ b/.github/workflows/run-aishell-2022-06-20.yml @@ -45,7 +45,7 @@ jobs: strategy: matrix: os: [ubuntu-latest] - python-version: [3.7, 3.8, 3.9] + python-version: [3.8] fail-fast: false diff --git a/.github/workflows/run-gigaspeech-2022-05-13.yml b/.github/workflows/run-gigaspeech-2022-05-13.yml index 0e47f7538..3121520c1 100644 --- a/.github/workflows/run-gigaspeech-2022-05-13.yml +++ b/.github/workflows/run-gigaspeech-2022-05-13.yml @@ -44,7 +44,7 @@ jobs: strategy: matrix: os: [ubuntu-latest] - python-version: [3.7, 3.8, 3.9] + python-version: [3.8] fail-fast: false diff --git a/.github/workflows/run-librispeech-2022-03-12.yml b/.github/workflows/run-librispeech-2022-03-12.yml index 3edbe43ec..f092e3c80 100644 --- a/.github/workflows/run-librispeech-2022-03-12.yml +++ b/.github/workflows/run-librispeech-2022-03-12.yml @@ -44,7 +44,7 @@ jobs: strategy: matrix: os: [ubuntu-latest] - python-version: [3.7, 3.8, 3.9] + python-version: [3.8] fail-fast: false diff --git a/.github/workflows/run-librispeech-2022-04-29.yml b/.github/workflows/run-librispeech-2022-04-29.yml index bb44a073b..f8f4d9977 100644 --- a/.github/workflows/run-librispeech-2022-04-29.yml +++ b/.github/workflows/run-librispeech-2022-04-29.yml @@ -44,7 +44,7 @@ jobs: strategy: matrix: os: [ubuntu-latest] - python-version: [3.7, 3.8, 3.9] + python-version: [3.8] fail-fast: false diff --git a/.github/workflows/run-librispeech-2022-05-13.yml b/.github/workflows/run-librispeech-2022-05-13.yml index e7b53b21c..dc20185da 100644 --- a/.github/workflows/run-librispeech-2022-05-13.yml +++ b/.github/workflows/run-librispeech-2022-05-13.yml @@ -44,7 +44,7 @@ jobs: strategy: matrix: os: [ubuntu-latest] - python-version: [3.7, 3.8, 3.9] + python-version: [3.8] fail-fast: false diff --git a/.github/workflows/run-librispeech-pruned-transducer-stateless3-2022-05-13.yml b/.github/workflows/run-librispeech-pruned-transducer-stateless3-2022-05-13.yml index bf73d4f18..3fb0920bc 100644 --- a/.github/workflows/run-librispeech-pruned-transducer-stateless3-2022-05-13.yml +++ b/.github/workflows/run-librispeech-pruned-transducer-stateless3-2022-05-13.yml @@ -44,7 +44,7 @@ jobs: strategy: matrix: os: [ubuntu-latest] - python-version: [3.7, 3.8, 3.9] + python-version: [3.8] fail-fast: false diff --git a/.github/workflows/run-librispeech-streaming-transducer-stateless2-2022-06-26.yml b/.github/workflows/run-librispeech-streaming-transducer-stateless2-2022-06-26.yml index 6ea308468..67a6f6fc4 100644 --- a/.github/workflows/run-librispeech-streaming-transducer-stateless2-2022-06-26.yml +++ b/.github/workflows/run-librispeech-streaming-transducer-stateless2-2022-06-26.yml @@ -44,7 +44,7 @@ jobs: strategy: matrix: os: [ubuntu-latest] - python-version: [3.7, 3.8, 3.9] + python-version: [3.8] fail-fast: false diff --git a/.github/workflows/run-librispeech-transducer-stateless2-2022-04-19.yml b/.github/workflows/run-librispeech-transducer-stateless2-2022-04-19.yml index 9fe2f0389..35ca08a31 100644 --- a/.github/workflows/run-librispeech-transducer-stateless2-2022-04-19.yml +++ b/.github/workflows/run-librispeech-transducer-stateless2-2022-04-19.yml @@ -44,7 +44,7 @@ jobs: strategy: matrix: os: [ubuntu-latest] - python-version: [3.7, 3.8, 3.9] + python-version: [3.8] fail-fast: false diff --git a/.github/workflows/run-multi-zh_hans-zipformer.yml b/.github/workflows/run-multi-zh_hans-zipformer.yml new file mode 100644 index 000000000..4ec81585f --- /dev/null +++ b/.github/workflows/run-multi-zh_hans-zipformer.yml @@ -0,0 +1,84 @@ +# Copyright 2023 Xiaomi Corp. (author: Zengrui Jin) + +# See ../../LICENSE for clarification regarding multiple authors +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +name: run-multi-zh_hans-zipformer + +on: + push: + branches: + - master + pull_request: + types: [labeled] + +concurrency: + group: run_multi-zh_hans_zipformer-${{ github.ref }} + cancel-in-progress: true + +jobs: + run_multi-zh_hans_zipformer: + if: github.event.label.name == 'onnx' || github.event.label.name == 'ready' || github.event_name == 'push' || github.event.label.name == 'multi-zh_hans' + runs-on: ${{ matrix.os }} + strategy: + matrix: + os: [ubuntu-latest] + python-version: [3.8] + + fail-fast: false + + steps: + - uses: actions/checkout@v2 + with: + fetch-depth: 0 + + - name: Setup Python ${{ matrix.python-version }} + uses: actions/setup-python@v2 + with: + python-version: ${{ matrix.python-version }} + cache: 'pip' + cache-dependency-path: '**/requirements-ci.txt' + + - name: Install Python dependencies + run: | + grep -v '^#' ./requirements-ci.txt | xargs -n 1 -L 1 pip install + pip uninstall -y protobuf + pip install --no-binary protobuf protobuf==3.20.* + + - name: Cache kaldifeat + id: my-cache + uses: actions/cache@v2 + with: + path: | + ~/tmp/kaldifeat + key: cache-tmp-${{ matrix.python-version }}-2023-05-22 + + - name: Install kaldifeat + if: steps.my-cache.outputs.cache-hit != 'true' + shell: bash + run: | + .github/scripts/install-kaldifeat.sh + + - name: Inference with pre-trained model + shell: bash + env: + GITHUB_EVENT_NAME: ${{ github.event_name }} + GITHUB_EVENT_LABEL_NAME: ${{ github.event.label.name }} + run: | + sudo apt-get -qq install git-lfs tree + export PYTHONPATH=$PWD:$PYTHONPATH + export PYTHONPATH=~/tmp/kaldifeat/kaldifeat/python:$PYTHONPATH + export PYTHONPATH=~/tmp/kaldifeat/build/lib:$PYTHONPATH + + .github/scripts/run-multi-zh_hans-zipformer.sh diff --git a/.github/workflows/run-pretrained-conformer-ctc.yml b/.github/workflows/run-pretrained-conformer-ctc.yml index bcd326b9d..6151a5a14 100644 --- a/.github/workflows/run-pretrained-conformer-ctc.yml +++ b/.github/workflows/run-pretrained-conformer-ctc.yml @@ -34,7 +34,7 @@ jobs: strategy: matrix: os: [ubuntu-latest] - python-version: [3.7, 3.8, 3.9] + python-version: [3.8] fail-fast: false diff --git a/.github/workflows/run-pretrained-transducer-stateless-librispeech-100h.yml b/.github/workflows/run-pretrained-transducer-stateless-librispeech-100h.yml index 1e5b25f5c..f8caee8e5 100644 --- a/.github/workflows/run-pretrained-transducer-stateless-librispeech-100h.yml +++ b/.github/workflows/run-pretrained-transducer-stateless-librispeech-100h.yml @@ -43,7 +43,7 @@ jobs: strategy: matrix: os: [ubuntu-latest] - python-version: [3.7, 3.8, 3.9] + python-version: [3.8] fail-fast: false diff --git a/.github/workflows/run-pretrained-transducer-stateless-librispeech-multi-datasets.yml b/.github/workflows/run-pretrained-transducer-stateless-librispeech-multi-datasets.yml index 9063c0ed6..7c3910eb8 100644 --- a/.github/workflows/run-pretrained-transducer-stateless-librispeech-multi-datasets.yml +++ b/.github/workflows/run-pretrained-transducer-stateless-librispeech-multi-datasets.yml @@ -43,7 +43,7 @@ jobs: strategy: matrix: os: [ubuntu-latest] - python-version: [3.7, 3.8, 3.9] + python-version: [3.8] fail-fast: false diff --git a/.github/workflows/run-pretrained-transducer-stateless-modified-2-aishell.yml b/.github/workflows/run-pretrained-transducer-stateless-modified-2-aishell.yml index 2d24528d3..ce6d6f92d 100644 --- a/.github/workflows/run-pretrained-transducer-stateless-modified-2-aishell.yml +++ b/.github/workflows/run-pretrained-transducer-stateless-modified-2-aishell.yml @@ -34,7 +34,7 @@ jobs: strategy: matrix: os: [ubuntu-latest] - python-version: [3.7, 3.8, 3.9] + python-version: [3.8] fail-fast: false diff --git a/.github/workflows/run-pretrained-transducer-stateless-modified-aishell.yml b/.github/workflows/run-pretrained-transducer-stateless-modified-aishell.yml index 761b26131..f0cebd94a 100644 --- a/.github/workflows/run-pretrained-transducer-stateless-modified-aishell.yml +++ b/.github/workflows/run-pretrained-transducer-stateless-modified-aishell.yml @@ -34,7 +34,7 @@ jobs: strategy: matrix: os: [ubuntu-latest] - python-version: [3.7, 3.8, 3.9] + python-version: [3.8] fail-fast: false diff --git a/.github/workflows/run-pretrained-transducer-stateless.yml b/.github/workflows/run-pretrained-transducer-stateless.yml index e46b9a849..1b69b97bf 100644 --- a/.github/workflows/run-pretrained-transducer-stateless.yml +++ b/.github/workflows/run-pretrained-transducer-stateless.yml @@ -43,7 +43,7 @@ jobs: strategy: matrix: os: [ubuntu-latest] - python-version: [3.7, 3.8, 3.9] + python-version: [3.8] fail-fast: false diff --git a/.github/workflows/run-pretrained-transducer.yml b/.github/workflows/run-pretrained-transducer.yml index 190e446bc..91d87f1c9 100644 --- a/.github/workflows/run-pretrained-transducer.yml +++ b/.github/workflows/run-pretrained-transducer.yml @@ -34,7 +34,7 @@ jobs: strategy: matrix: os: [ubuntu-latest] - python-version: [3.7, 3.8, 3.9] + python-version: [3.8] fail-fast: false diff --git a/README.md b/README.md index a876fb24e..523203aa4 100644 --- a/README.md +++ b/README.md @@ -338,7 +338,7 @@ We provide one model for this recipe: [Pruned stateless RNN-T: Conformer encoder #### Pruned stateless RNN-T: Conformer encoder + Embedding decoder + k2 pruned RNN-T loss -The best results for Chinese CER(%) and English WER(%) respectivly (zh: Chinese, en: English): +The best results for Chinese CER(%) and English WER(%) respectively (zh: Chinese, en: English): |decoding-method | dev | dev_zh | dev_en | test | test_zh | test_en | |--|--|--|--|--|--|--| |greedy_search| 7.30 | 6.48 | 19.19 |7.39| 6.66 | 19.13| diff --git a/egs/aidatatang_200zh/ASR/pruned_transducer_stateless2/asr_datamodule.py b/egs/aidatatang_200zh/ASR/pruned_transducer_stateless2/asr_datamodule.py index 167d5e15e..49a697bfd 100644 --- a/egs/aidatatang_200zh/ASR/pruned_transducer_stateless2/asr_datamodule.py +++ b/egs/aidatatang_200zh/ASR/pruned_transducer_stateless2/asr_datamodule.py @@ -37,7 +37,7 @@ from lhotse.dataset import ( DynamicBucketingSampler, K2SpeechRecognitionDataset, PrecomputedFeatures, - SingleCutSampler, + SimpleCutSampler, SpecAugment, ) from lhotse.dataset.input_strategies import OnTheFlyFeatures @@ -291,8 +291,8 @@ class Aidatatang_200zhAsrDataModule: drop_last=True, ) else: - logging.info("Using SingleCutSampler.") - train_sampler = SingleCutSampler( + logging.info("Using SimpleCutSampler.") + train_sampler = SimpleCutSampler( cuts_train, max_duration=self.args.max_duration, shuffle=self.args.shuffle, diff --git a/egs/aishell/ASR/tdnn_lstm_ctc/asr_datamodule.py b/egs/aishell/ASR/tdnn_lstm_ctc/asr_datamodule.py index efb32336a..180930747 100644 --- a/egs/aishell/ASR/tdnn_lstm_ctc/asr_datamodule.py +++ b/egs/aishell/ASR/tdnn_lstm_ctc/asr_datamodule.py @@ -30,7 +30,7 @@ from lhotse.dataset import ( DynamicBucketingSampler, K2SpeechRecognitionDataset, PrecomputedFeatures, - SingleCutSampler, + SimpleCutSampler, SpecAugment, ) from lhotse.dataset.input_strategies import OnTheFlyFeatures @@ -278,8 +278,8 @@ class AishellAsrDataModule: drop_last=self.args.drop_last, ) else: - logging.info("Using SingleCutSampler.") - train_sampler = SingleCutSampler( + logging.info("Using SimpleCutSampler.") + train_sampler = SimpleCutSampler( cuts_train, max_duration=self.args.max_duration, shuffle=self.args.shuffle, diff --git a/egs/aishell2/ASR/pruned_transducer_stateless5/asr_datamodule.py b/egs/aishell2/ASR/pruned_transducer_stateless5/asr_datamodule.py index 0f383a244..af37cc175 100644 --- a/egs/aishell2/ASR/pruned_transducer_stateless5/asr_datamodule.py +++ b/egs/aishell2/ASR/pruned_transducer_stateless5/asr_datamodule.py @@ -31,7 +31,7 @@ from lhotse.dataset import ( # noqa F401 for PrecomputedFeatures DynamicBucketingSampler, K2SpeechRecognitionDataset, PrecomputedFeatures, - SingleCutSampler, + SimpleCutSampler, SpecAugment, ) from lhotse.dataset.input_strategies import ( # noqa F401 For AudioSamples @@ -299,8 +299,8 @@ class AiShell2AsrDataModule: drop_last=self.args.drop_last, ) else: - logging.info("Using SingleCutSampler.") - train_sampler = SingleCutSampler( + logging.info("Using SimpleCutSampler.") + train_sampler = SimpleCutSampler( cuts_train, max_duration=self.args.max_duration, shuffle=self.args.shuffle, diff --git a/egs/aishell4/ASR/pruned_transducer_stateless5/asr_datamodule.py b/egs/aishell4/ASR/pruned_transducer_stateless5/asr_datamodule.py index d980a857f..da9da371e 100644 --- a/egs/aishell4/ASR/pruned_transducer_stateless5/asr_datamodule.py +++ b/egs/aishell4/ASR/pruned_transducer_stateless5/asr_datamodule.py @@ -30,7 +30,7 @@ from lhotse.dataset import ( # noqa F401 for PrecomputedFeatures DynamicBucketingSampler, K2SpeechRecognitionDataset, PrecomputedFeatures, - SingleCutSampler, + SimpleCutSampler, SpecAugment, ) from lhotse.dataset.input_strategies import ( # noqa F401 for AudioSamples @@ -310,8 +310,8 @@ class Aishell4AsrDataModule: drop_last=self.args.drop_last, ) else: - logging.info("Using SingleCutSampler.") - train_sampler = SingleCutSampler( + logging.info("Using SimpleCutSampler.") + train_sampler = SimpleCutSampler( cuts_train, max_duration=self.args.max_duration, shuffle=self.args.shuffle, diff --git a/egs/alimeeting/ASR/pruned_transducer_stateless2/asr_datamodule.py b/egs/alimeeting/ASR/pruned_transducer_stateless2/asr_datamodule.py index a9a4675a9..4799da19d 100644 --- a/egs/alimeeting/ASR/pruned_transducer_stateless2/asr_datamodule.py +++ b/egs/alimeeting/ASR/pruned_transducer_stateless2/asr_datamodule.py @@ -37,7 +37,7 @@ from lhotse.dataset import ( DynamicBucketingSampler, K2SpeechRecognitionDataset, PrecomputedFeatures, - SingleCutSampler, + SimpleCutSampler, SpecAugment, ) from lhotse.dataset.input_strategies import OnTheFlyFeatures @@ -292,8 +292,8 @@ class AlimeetingAsrDataModule: drop_last=True, ) else: - logging.info("Using SingleCutSampler.") - train_sampler = SingleCutSampler( + logging.info("Using SimpleCutSampler.") + train_sampler = SimpleCutSampler( cuts_train, max_duration=self.args.max_duration, shuffle=self.args.shuffle, diff --git a/egs/ami/SURT/dprnn_zipformer/asr_datamodule.py b/egs/ami/SURT/dprnn_zipformer/asr_datamodule.py index ec8106bc3..3dd786d33 100644 --- a/egs/ami/SURT/dprnn_zipformer/asr_datamodule.py +++ b/egs/ami/SURT/dprnn_zipformer/asr_datamodule.py @@ -257,7 +257,7 @@ class AmiAsrDataModule: drop_last=self.args.drop_last, ) else: - logging.info("Using SingleCutSampler.") + logging.info("Using SimpleCutSampler.") train_sampler = SimpleCutSampler( cuts_train, max_duration=self.args.max_duration, diff --git a/egs/commonvoice/ASR/pruned_transducer_stateless7/asr_datamodule.py b/egs/commonvoice/ASR/pruned_transducer_stateless7/asr_datamodule.py index 2c37244a4..73f2f1dce 100644 --- a/egs/commonvoice/ASR/pruned_transducer_stateless7/asr_datamodule.py +++ b/egs/commonvoice/ASR/pruned_transducer_stateless7/asr_datamodule.py @@ -30,7 +30,7 @@ from lhotse.dataset import ( # noqa F401 for PrecomputedFeatures DynamicBucketingSampler, K2SpeechRecognitionDataset, PrecomputedFeatures, - SingleCutSampler, + SimpleCutSampler, SpecAugment, ) from lhotse.dataset.input_strategies import ( # noqa F401 For AudioSamples @@ -311,8 +311,8 @@ class CommonVoiceAsrDataModule: drop_last=self.args.drop_last, ) else: - logging.info("Using SingleCutSampler.") - train_sampler = SingleCutSampler( + logging.info("Using SimpleCutSampler.") + train_sampler = SimpleCutSampler( cuts_train, max_duration=self.args.max_duration, shuffle=self.args.shuffle, diff --git a/egs/csj/ASR/local/utils/asr_datamodule.py b/egs/csj/ASR/local/utils/asr_datamodule.py index 619820a75..272486227 100644 --- a/egs/csj/ASR/local/utils/asr_datamodule.py +++ b/egs/csj/ASR/local/utils/asr_datamodule.py @@ -31,7 +31,7 @@ from lhotse.dataset import ( # noqa F401 for PrecomputedFeatures DynamicBucketingSampler, K2SpeechRecognitionDataset, PrecomputedFeatures, - SingleCutSampler, + SimpleCutSampler, SpecAugment, ) from lhotse.dataset.input_strategies import ( # noqa F401 For AudioSamples @@ -339,8 +339,8 @@ class CSJAsrDataModule: drop_last=self.args.drop_last, ) else: - logging.info("Using SingleCutSampler.") - train_sampler = SingleCutSampler( + logging.info("Using SimpleCutSampler.") + train_sampler = SimpleCutSampler( cuts_train, max_duration=self.args.max_duration, shuffle=self.args.shuffle, diff --git a/egs/gigaspeech/ASR/conformer_ctc/asr_datamodule.py b/egs/gigaspeech/ASR/conformer_ctc/asr_datamodule.py index 9437c935c..9d6e3c42a 100644 --- a/egs/gigaspeech/ASR/conformer_ctc/asr_datamodule.py +++ b/egs/gigaspeech/ASR/conformer_ctc/asr_datamodule.py @@ -27,7 +27,7 @@ from lhotse.dataset import ( DynamicBucketingSampler, K2SpeechRecognitionDataset, PrecomputedFeatures, - SingleCutSampler, + SimpleCutSampler, SpecAugment, ) from lhotse.dataset.input_strategies import OnTheFlyFeatures @@ -264,8 +264,8 @@ class GigaSpeechAsrDataModule: drop_last=True, ) else: - logging.info("Using SingleCutSampler.") - train_sampler = SingleCutSampler( + logging.info("Using SimpleCutSampler.") + train_sampler = SimpleCutSampler( cuts_train, max_duration=self.args.max_duration, shuffle=self.args.shuffle, diff --git a/egs/gigaspeech/ASR/pruned_transducer_stateless2/asr_datamodule.py b/egs/gigaspeech/ASR/pruned_transducer_stateless2/asr_datamodule.py index 4d5d2b8f9..29e72b408 100644 --- a/egs/gigaspeech/ASR/pruned_transducer_stateless2/asr_datamodule.py +++ b/egs/gigaspeech/ASR/pruned_transducer_stateless2/asr_datamodule.py @@ -30,7 +30,7 @@ from lhotse.dataset import ( DynamicBucketingSampler, K2SpeechRecognitionDataset, PrecomputedFeatures, - SingleCutSampler, + SimpleCutSampler, SpecAugment, ) from lhotse.dataset.input_strategies import OnTheFlyFeatures @@ -297,8 +297,8 @@ class GigaSpeechAsrDataModule: drop_last=True, ) else: - logging.info("Using SingleCutSampler.") - train_sampler = SingleCutSampler( + logging.info("Using SimpleCutSampler.") + train_sampler = SimpleCutSampler( cuts_train, max_duration=self.args.max_duration, shuffle=self.args.shuffle, diff --git a/egs/libricss/SURT/dprnn_zipformer/asr_datamodule.py b/egs/libricss/SURT/dprnn_zipformer/asr_datamodule.py index 51df91598..a72df89e0 100644 --- a/egs/libricss/SURT/dprnn_zipformer/asr_datamodule.py +++ b/egs/libricss/SURT/dprnn_zipformer/asr_datamodule.py @@ -259,7 +259,7 @@ class LibriCssAsrDataModule: drop_last=self.args.drop_last, ) else: - logging.info("Using SingleCutSampler.") + logging.info("Using SimpleCutSampler.") train_sampler = SimpleCutSampler( cuts_train, max_duration=self.args.max_duration, diff --git a/egs/libricss/SURT/dprnn_zipformer/scaling.py b/egs/libricss/SURT/dprnn_zipformer/scaling.py deleted file mode 120000 index 5f9be9fe0..000000000 --- a/egs/libricss/SURT/dprnn_zipformer/scaling.py +++ /dev/null @@ -1 +0,0 @@ -../../../librispeech/ASR/pruned_transducer_stateless7/scaling.py \ No newline at end of file diff --git a/egs/libricss/SURT/dprnn_zipformer/scaling.py b/egs/libricss/SURT/dprnn_zipformer/scaling.py new file mode 100644 index 000000000..4040a7b89 --- /dev/null +++ b/egs/libricss/SURT/dprnn_zipformer/scaling.py @@ -0,0 +1,1576 @@ +# Copyright 2022 Xiaomi Corp. (authors: Daniel Povey) +# +# 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 +import random +from typing import Optional, Tuple, Union + +import torch +import torch.backends.cudnn.rnn as rnn +import torch.nn as nn +from torch import _VF, Tensor + +from icefall.utils import is_jit_tracing + + +class ActivationBalancerFunction(torch.autograd.Function): + @staticmethod + def forward( + ctx, + x: Tensor, + scale_factor: Tensor, + sign_factor: Optional[Tensor], + channel_dim: int, + ) -> Tensor: + if channel_dim < 0: + channel_dim += x.ndim + ctx.channel_dim = channel_dim + xgt0 = x > 0 + if sign_factor is None: + ctx.save_for_backward(xgt0, scale_factor) + else: + ctx.save_for_backward(xgt0, scale_factor, sign_factor) + return x + + @staticmethod + def backward(ctx, x_grad: Tensor) -> Tuple[Tensor, None, None, None]: + if len(ctx.saved_tensors) == 3: + xgt0, scale_factor, sign_factor = ctx.saved_tensors + for _ in range(ctx.channel_dim, x_grad.ndim - 1): + scale_factor = scale_factor.unsqueeze(-1) + sign_factor = sign_factor.unsqueeze(-1) + factor = sign_factor + scale_factor * (xgt0.to(x_grad.dtype) - 0.5) + else: + xgt0, scale_factor = ctx.saved_tensors + for _ in range(ctx.channel_dim, x_grad.ndim - 1): + scale_factor = scale_factor.unsqueeze(-1) + factor = scale_factor * (xgt0.to(x_grad.dtype) - 0.5) + neg_delta_grad = x_grad.abs() * factor + return ( + x_grad - neg_delta_grad, + None, + None, + None, + ) + + +def _compute_scale_factor( + x: Tensor, + channel_dim: int, + min_abs: float, + max_abs: float, + gain_factor: float, + max_factor: float, +) -> Tensor: + if channel_dim < 0: + channel_dim += x.ndim + sum_dims = [d for d in range(x.ndim) if d != channel_dim] + x_abs_mean = torch.mean(x.abs(), dim=sum_dims).to(torch.float32) + + if min_abs == 0.0: + below_threshold = 0.0 + else: + # below_threshold is 0 if x_abs_mean > min_abs, can be at most max_factor if + # x_abs)_mean , min_abs. + below_threshold = ((min_abs - x_abs_mean) * (gain_factor / min_abs)).clamp( + min=0, max=max_factor + ) + + above_threshold = ((x_abs_mean - max_abs) * (gain_factor / max_abs)).clamp( + min=0, max=max_factor + ) + + return below_threshold - above_threshold + + +def _compute_sign_factor( + x: Tensor, + channel_dim: int, + min_positive: float, + max_positive: float, + gain_factor: float, + max_factor: float, +) -> Tensor: + if channel_dim < 0: + channel_dim += x.ndim + sum_dims = [d for d in range(x.ndim) if d != channel_dim] + proportion_positive = torch.mean((x > 0).to(torch.float32), dim=sum_dims) + if min_positive == 0.0: + factor1 = 0.0 + else: + # 0 if proportion_positive >= min_positive, else can be + # as large as max_factor. + factor1 = ( + (min_positive - proportion_positive) * (gain_factor / min_positive) + ).clamp_(min=0, max=max_factor) + + if max_positive == 1.0: + factor2 = 0.0 + else: + # 0 if self.proportion_positive <= max_positive, else can be + # as large as -max_factor. + factor2 = ( + (proportion_positive - max_positive) * (gain_factor / (1.0 - max_positive)) + ).clamp_(min=0, max=max_factor) + sign_factor = factor1 - factor2 + # require min_positive != 0 or max_positive != 1: + assert not isinstance(sign_factor, float) + return sign_factor + + +class ActivationScaleBalancerFunction(torch.autograd.Function): + """ + This object is used in class ActivationBalancer when the user specified + min_positive=0, max_positive=1, so there are no constraints on the signs + of the activations and only the absolute value has a constraint. + """ + + @staticmethod + def forward( + ctx, + x: Tensor, + sign_factor: Tensor, + scale_factor: Tensor, + channel_dim: int, + ) -> Tensor: + if channel_dim < 0: + channel_dim += x.ndim + ctx.channel_dim = channel_dim + xgt0 = x > 0 + ctx.save_for_backward(xgt0, sign_factor, scale_factor) + return x + + @staticmethod + def backward(ctx, x_grad: Tensor) -> Tuple[Tensor, None, None, None]: + xgt0, sign_factor, scale_factor = ctx.saved_tensors + for _ in range(ctx.channel_dim, x_grad.ndim - 1): + sign_factor = sign_factor.unsqueeze(-1) + scale_factor = scale_factor.unsqueeze(-1) + + factor = sign_factor + scale_factor * (xgt0.to(x_grad.dtype) - 0.5) + neg_delta_grad = x_grad.abs() * factor + return ( + x_grad - neg_delta_grad, + None, + None, + None, + ) + + +class RandomClampFunction(torch.autograd.Function): + @staticmethod + def forward( + ctx, + x: Tensor, + min: Optional[float], + max: Optional[float], + prob: float, + reflect: float, + ) -> Tensor: + x_clamped = torch.clamp(x, min=min, max=max) + mask = torch.rand_like(x) < prob + ans = torch.where(mask, x_clamped, x) + if x.requires_grad: + ctx.save_for_backward(ans == x) + ctx.reflect = reflect + if reflect != 0.0: + ans = ans * (1.0 + reflect) - (x * reflect) + return ans + + @staticmethod + def backward(ctx, ans_grad: Tensor) -> Tuple[Tensor, None, None, None, None]: + (is_same,) = ctx.saved_tensors + x_grad = ans_grad * is_same.to(ans_grad.dtype) + reflect = ctx.reflect + if reflect != 0.0: + x_grad = x_grad * (1.0 + reflect) - (ans_grad * reflect) + return x_grad, None, None, None, None + + +def random_clamp( + x: Tensor, + min: Optional[float] = None, + max: Optional[float] = None, + prob: float = 0.5, + reflect: float = 0.0, +): + return RandomClampFunction.apply(x, min, max, prob, reflect) + + +def random_cast_to_half(x: Tensor, min_abs: float = 5.0e-06) -> Tensor: + """ + A randomized way of casting a floating point value to half precision. + """ + if x.dtype == torch.float16: + return x + x_abs = x.abs() + is_too_small = x_abs < min_abs + # for elements where is_too_small is true, random_val will contain +-min_abs with + # probability (x.abs() / min_abs), and 0.0 otherwise. [so this preserves expectations, + # for those elements]. + random_val = min_abs * x.sign() * (torch.rand_like(x) * min_abs < x_abs) + return torch.where(is_too_small, random_val, x).to(torch.float16) + + +class RandomGradFunction(torch.autograd.Function): + """ + Does nothing in forward pass; in backward pass, gets rid of very small grads using + randomized approach that preserves expectations (intended to reduce roundoff). + """ + + @staticmethod + def forward(ctx, x: Tensor, min_abs: float) -> Tensor: + ctx.min_abs = min_abs + return x + + @staticmethod + def backward(ctx, ans_grad: Tensor) -> Tuple[Tensor, None]: + if ans_grad.dtype == torch.float16: + return ( + random_cast_to_half(ans_grad.to(torch.float32), min_abs=ctx.min_abs), + None, + ) + else: + return ans_grad, None + + +class RandomGrad(torch.nn.Module): + """ + Gets rid of very small gradients using an expectation-preserving method, intended to increase + accuracy of training when using amp (automatic mixed precision) + """ + + def __init__(self, min_abs: float = 5.0e-06): + super(RandomGrad, self).__init__() + self.min_abs = min_abs + + def forward(self, x: Tensor): + if torch.jit.is_scripting() or not self.training or torch.jit.is_tracing(): + return x + else: + return RandomGradFunction.apply(x, self.min_abs) + + +class SoftmaxFunction(torch.autograd.Function): + """ + Tries to handle half-precision derivatives in a randomized way that should + be more accurate for training than the default behavior. + """ + + @staticmethod + def forward(ctx, x: Tensor, dim: int): + ans = x.softmax(dim=dim) + # if x dtype is float16, x.softmax() returns a float32 because + # (presumably) that op does not support float16, and autocast + # is enabled. + if torch.is_autocast_enabled(): + ans = ans.to(torch.float16) + ctx.save_for_backward(ans) + ctx.x_dtype = x.dtype + ctx.dim = dim + return ans + + @staticmethod + def backward(ctx, ans_grad: Tensor): + (ans,) = ctx.saved_tensors + with torch.cuda.amp.autocast(enabled=False): + ans_grad = ans_grad.to(torch.float32) + ans = ans.to(torch.float32) + x_grad = ans_grad * ans + x_grad = x_grad - ans * x_grad.sum(dim=ctx.dim, keepdim=True) + return x_grad, None + + +def softmax(x: Tensor, dim: int): + if torch.jit.is_scripting() or torch.jit.is_tracing(): + return x.softmax(dim) + + return SoftmaxFunction.apply(x, dim) + + +class MaxEigLimiterFunction(torch.autograd.Function): + @staticmethod + def forward( + ctx, + x: Tensor, + coeffs: Tensor, + direction: Tensor, + channel_dim: int, + grad_scale: float, + ) -> Tensor: + ctx.channel_dim = channel_dim + ctx.grad_scale = grad_scale + ctx.save_for_backward(x.detach(), coeffs.detach(), direction.detach()) + return x + + @staticmethod + def backward(ctx, x_grad, *args): + with torch.enable_grad(): + (x_orig, coeffs, new_direction) = ctx.saved_tensors + x_orig.requires_grad = True + num_channels = x_orig.shape[ctx.channel_dim] + x = x_orig.transpose(ctx.channel_dim, -1).reshape(-1, num_channels) + new_direction.requires_grad = False + x = x - x.mean(dim=0) + x_var = (x**2).mean() + x_residual = x - coeffs * new_direction + x_residual_var = (x_residual**2).mean() + # `variance_proportion` is the proportion of the variance accounted for + # by the top eigen-direction. This is to be minimized. + variance_proportion = (x_var - x_residual_var) / (x_var + 1.0e-20) + variance_proportion.backward() + x_orig_grad = x_orig.grad + x_extra_grad = ( + x_orig.grad + * ctx.grad_scale + * x_grad.norm() + / (x_orig_grad.norm() + 1.0e-20) + ) + return x_grad + x_extra_grad.detach(), None, None, None, None + + +class GradientFilterFunction(torch.autograd.Function): + @staticmethod + def forward( + ctx, + x: Tensor, + batch_dim: int, # e.g., 1 + threshold: float, # e.g., 10.0 + *params: Tensor, # module parameters + ) -> Tuple[Tensor, ...]: + if x.requires_grad: + if batch_dim < 0: + batch_dim += x.ndim + ctx.batch_dim = batch_dim + ctx.threshold = threshold + return (x,) + params + + @staticmethod + def backward( + ctx, + x_grad: Tensor, + *param_grads: Tensor, + ) -> Tuple[Tensor, ...]: + eps = 1.0e-20 + dim = ctx.batch_dim + norm_dims = [d for d in range(x_grad.ndim) if d != dim] + norm_of_batch = (x_grad**2).mean(dim=norm_dims, keepdim=True).sqrt() + median_norm = norm_of_batch.median() + + cutoff = median_norm * ctx.threshold + inv_mask = (cutoff + norm_of_batch) / (cutoff + eps) + mask = 1.0 / (inv_mask + eps) + x_grad = x_grad * mask + + avg_mask = 1.0 / (inv_mask.mean() + eps) + param_grads = [avg_mask * g for g in param_grads] + + return (x_grad, None, None) + tuple(param_grads) + + +class GradientFilter(torch.nn.Module): + """This is used to filter out elements that have extremely large gradients + in batch and the module parameters with soft masks. + + Args: + batch_dim (int): + The batch dimension. + threshold (float): + For each element in batch, its gradient will be + filtered out if the gradient norm is larger than + `grad_norm_threshold * median`, where `median` is the median + value of gradient norms of all elememts in batch. + """ + + def __init__(self, batch_dim: int = 1, threshold: float = 10.0): + super(GradientFilter, self).__init__() + self.batch_dim = batch_dim + self.threshold = threshold + + def forward(self, x: Tensor, *params: Tensor) -> Tuple[Tensor, ...]: + if torch.jit.is_scripting() or is_jit_tracing(): + return (x,) + params + else: + return GradientFilterFunction.apply( + x, + self.batch_dim, + self.threshold, + *params, + ) + + +class BasicNorm(torch.nn.Module): + """ + This is intended to be a simpler, and hopefully cheaper, replacement for + LayerNorm. The observation this is based on, is that Transformer-type + networks, especially with pre-norm, sometimes seem to set one of the + feature dimensions to a large constant value (e.g. 50), which "defeats" + the LayerNorm because the output magnitude is then not strongly dependent + on the other (useful) features. Presumably the weight and bias of the + LayerNorm are required to allow it to do this. + + So the idea is to introduce this large constant value as an explicit + parameter, that takes the role of the "eps" in LayerNorm, so the network + doesn't have to do this trick. We make the "eps" learnable. + + Args: + num_channels: the number of channels, e.g. 512. + channel_dim: the axis/dimension corresponding to the channel, + interprted as an offset from the input's ndim if negative. + shis is NOT the num_channels; it should typically be one of + {-2, -1, 0, 1, 2, 3}. + eps: the initial "epsilon" that we add as ballast in: + scale = ((input_vec**2).mean() + epsilon)**-0.5 + Note: our epsilon is actually large, but we keep the name + to indicate the connection with conventional LayerNorm. + learn_eps: if true, we learn epsilon; if false, we keep it + at the initial value. + eps_min: float + eps_max: float + """ + + def __init__( + self, + num_channels: int, + channel_dim: int = -1, # CAUTION: see documentation. + eps: float = 0.25, + learn_eps: bool = True, + eps_min: float = -3.0, + eps_max: float = 3.0, + ) -> None: + super(BasicNorm, self).__init__() + self.num_channels = num_channels + self.channel_dim = channel_dim + if learn_eps: + self.eps = nn.Parameter(torch.tensor(eps).log().detach()) + else: + self.register_buffer("eps", torch.tensor(eps).log().detach()) + self.eps_min = eps_min + self.eps_max = eps_max + + def forward(self, x: Tensor) -> Tensor: + assert x.shape[self.channel_dim] == self.num_channels + eps = self.eps + if self.training and random.random() < 0.25: + # with probability 0.25, in training mode, clamp eps between the min + # and max; this will encourage it to learn parameters within the + # allowed range by making parameters that are outside the allowed + # range noisy. + + # gradients to allow the parameter to get back into the allowed + # region if it happens to exit it. + eps = eps.clamp(min=self.eps_min, max=self.eps_max) + scales = ( + torch.mean(x**2, dim=self.channel_dim, keepdim=True) + eps.exp() + ) ** -0.5 + return x * scales + + +class ScaledEmbedding(nn.Module): + r"""This is a modified version of nn.Embedding that introduces a learnable scale + on the parameters. Note: due to how we initialize it, it's best used with + schedulers like Noam that have a warmup period. + + It is a simple lookup table that stores embeddings of a fixed dictionary and size. + + This module is often used to store word embeddings and retrieve them using indices. + The input to the module is a list of indices, and the output is the corresponding + word embeddings. + + Args: + num_embeddings (int): size of the dictionary of embeddings + embedding_dim (int): the size of each embedding vector + padding_idx (int, optional): If given, pads the output with the embedding vector at :attr:`padding_idx` + (initialized to zeros) whenever it encounters the index. + scale_grad_by_freq (boolean, optional): If given, this will scale gradients by the inverse of frequency of + the words in the mini-batch. Default ``False``. + sparse (bool, optional): If ``True``, gradient w.r.t. :attr:`weight` matrix will be a sparse tensor. + See Notes for more details regarding sparse gradients. + + initial_speed (float, optional): This affects how fast the parameter will + learn near the start of training; you can set it to a value less than + one if you suspect that a module is contributing to instability near + the start of training. Note: regardless of the use of this option, + it's best to use schedulers like Noam that have a warm-up period. + Alternatively you can set it to more than 1 if you want it to + initially train faster. Must be greater than 0. + + + Attributes: + weight (Tensor): the learnable weights of the module of shape (num_embeddings, embedding_dim) + initialized from :math:`\mathcal{N}(0, 1)` + + Shape: + - Input: :math:`(*)`, LongTensor of arbitrary shape containing the indices to extract + - Output: :math:`(*, H)`, where `*` is the input shape and :math:`H=\text{embedding\_dim}` + + .. note:: + Keep in mind that only a limited number of optimizers support + sparse gradients: currently it's :class:`optim.SGD` (`CUDA` and `CPU`), + :class:`optim.SparseAdam` (`CUDA` and `CPU`) and :class:`optim.Adagrad` (`CPU`) + + .. note:: + With :attr:`padding_idx` set, the embedding vector at + :attr:`padding_idx` is initialized to all zeros. However, note that this + vector can be modified afterwards, e.g., using a customized + initialization method, and thus changing the vector used to pad the + output. The gradient for this vector from :class:`~torch.nn.Embedding` + is always zero. + + Examples:: + + >>> # an Embedding module containing 10 tensors of size 3 + >>> embedding = nn.Embedding(10, 3) + >>> # a batch of 2 samples of 4 indices each + >>> input = torch.LongTensor([[1,2,4,5],[4,3,2,9]]) + >>> embedding(input) + tensor([[[-0.0251, -1.6902, 0.7172], + [-0.6431, 0.0748, 0.6969], + [ 1.4970, 1.3448, -0.9685], + [-0.3677, -2.7265, -0.1685]], + + [[ 1.4970, 1.3448, -0.9685], + [ 0.4362, -0.4004, 0.9400], + [-0.6431, 0.0748, 0.6969], + [ 0.9124, -2.3616, 1.1151]]]) + + + >>> # example with padding_idx + >>> embedding = nn.Embedding(10, 3, padding_idx=0) + >>> input = torch.LongTensor([[0,2,0,5]]) + >>> embedding(input) + tensor([[[ 0.0000, 0.0000, 0.0000], + [ 0.1535, -2.0309, 0.9315], + [ 0.0000, 0.0000, 0.0000], + [-0.1655, 0.9897, 0.0635]]]) + + """ + __constants__ = [ + "num_embeddings", + "embedding_dim", + "padding_idx", + "scale_grad_by_freq", + "sparse", + ] + + num_embeddings: int + embedding_dim: int + padding_idx: int + scale_grad_by_freq: bool + weight: Tensor + sparse: bool + + def __init__( + self, + num_embeddings: int, + embedding_dim: int, + padding_idx: Optional[int] = None, + scale_grad_by_freq: bool = False, + sparse: bool = False, + initial_speed: float = 1.0, + ) -> None: + super(ScaledEmbedding, self).__init__() + self.num_embeddings = num_embeddings + self.embedding_dim = embedding_dim + if padding_idx is not None: + if padding_idx > 0: + assert ( + padding_idx < self.num_embeddings + ), "Padding_idx must be within num_embeddings" + elif padding_idx < 0: + assert ( + padding_idx >= -self.num_embeddings + ), "Padding_idx must be within num_embeddings" + padding_idx = self.num_embeddings + padding_idx + self.padding_idx = padding_idx + self.scale_grad_by_freq = scale_grad_by_freq + + self.scale = nn.Parameter(torch.zeros(())) # see reset_parameters() + self.sparse = sparse + + self.weight = nn.Parameter(torch.Tensor(num_embeddings, embedding_dim)) + self.reset_parameters(initial_speed) + + def reset_parameters(self, initial_speed: float = 1.0) -> None: + std = 0.1 / initial_speed + nn.init.normal_(self.weight, std=std) + nn.init.constant_(self.scale, torch.tensor(1.0 / std).log()) + + if self.padding_idx is not None: + with torch.no_grad(): + self.weight[self.padding_idx].fill_(0) + + def forward(self, input: Tensor) -> Tensor: + F = torch.nn.functional + scale = self.scale.exp() + if input.numel() < self.num_embeddings: + return ( + F.embedding( + input, + self.weight, + self.padding_idx, + None, + 2.0, # None, 2.0 relate to normalization + self.scale_grad_by_freq, + self.sparse, + ) + * scale + ) + else: + return F.embedding( + input, + self.weight * scale, + self.padding_idx, + None, + 2.0, # None, 2.0 relates to normalization + self.scale_grad_by_freq, + self.sparse, + ) + + def extra_repr(self) -> str: + # s = "{num_embeddings}, {embedding_dim}, scale={scale}" + s = "{num_embeddings}, {embedding_dim}" + if self.padding_idx is not None: + s += ", padding_idx={padding_idx}" + if self.scale_grad_by_freq is not False: + s += ", scale_grad_by_freq={scale_grad_by_freq}" + if self.sparse is not False: + s += ", sparse=True" + return s.format(**self.__dict__) + + +def ScaledLinear(*args, initial_scale: float = 1.0, **kwargs) -> nn.Linear: + """ + Behaves like a constructor of a modified version of nn.Linear + that gives an easy way to set the default initial parameter scale. + + Args: + Accepts the standard args and kwargs that nn.Linear accepts + e.g. in_features, out_features, bias=False. + + initial_scale: you can override this if you want to increase + or decrease the initial magnitude of the module's output + (affects the initialization of weight_scale and bias_scale). + Another option, if you want to do something like this, is + to re-initialize the parameters. + """ + ans = nn.Linear(*args, **kwargs) + with torch.no_grad(): + ans.weight[:] *= initial_scale + if ans.bias is not None: + torch.nn.init.uniform_(ans.bias, -0.1 * initial_scale, 0.1 * initial_scale) + return ans + + +def ScaledConv1d(*args, initial_scale: float = 1.0, **kwargs) -> nn.Conv1d: + """ + Behaves like a constructor of a modified version of nn.Conv1d + that gives an easy way to set the default initial parameter scale. + + Args: + Accepts the standard args and kwargs that nn.Linear accepts + e.g. in_features, out_features, bias=False. + + initial_scale: you can override this if you want to increase + or decrease the initial magnitude of the module's output + (affects the initialization of weight_scale and bias_scale). + Another option, if you want to do something like this, is + to re-initialize the parameters. + """ + ans = nn.Conv1d(*args, **kwargs) + with torch.no_grad(): + ans.weight[:] *= initial_scale + if ans.bias is not None: + torch.nn.init.uniform_(ans.bias, -0.1 * initial_scale, 0.1 * initial_scale) + return ans + + +class ScaledLSTM(nn.LSTM): + # See docs for ScaledLinear. + # This class implements LSTM with scaling mechanism, using `torch._VF.lstm` + # Please refer to https://github.com/pytorch/pytorch/blob/master/torch/nn/modules/rnn.py + def __init__( + self, + *args, + initial_scale: float = 1.0, + initial_speed: float = 1.0, + grad_norm_threshold: float = 10.0, + **kwargs, + ): + super(ScaledLSTM, self).__init__(*args, **kwargs) + initial_scale = torch.tensor(initial_scale).log() + self._scales_names = [] + self._scales = [] + self.batch_dim = 0 if self.batch_first else 1 + self.num_directions = 1 + int(self.bidirectional) + for name in self._flat_weights_names: + scale_name = name + "_scale" + self._scales_names.append(scale_name) + param = nn.Parameter(initial_scale.clone().detach()) + setattr(self, scale_name, param) + self._scales.append(param) + + self.grad_filter = GradientFilter( + batch_dim=self.batch_dim, threshold=grad_norm_threshold + ) + + self._reset_parameters( + initial_speed + ) # Overrides the reset_parameters in base class + + def _reset_parameters(self, initial_speed: float): + std = 0.1 / initial_speed + a = (3**0.5) * std + scale = self.hidden_size**-0.5 + v = scale / std + for idx, name in enumerate(self._flat_weights_names): + if "weight" in name: + nn.init.uniform_(self._flat_weights[idx], -a, a) + with torch.no_grad(): + self._scales[idx] += torch.tensor(v).log() + elif "bias" in name: + nn.init.constant_(self._flat_weights[idx], 0.0) + + def _flatten_parameters(self, flat_weights) -> None: + """Resets parameter data pointer so that they can use faster code paths. + + Right now, this works only if the module is on the GPU and cuDNN is enabled. + Otherwise, it's a no-op. + + This function is modified from https://github.com/pytorch/pytorch/blob/master/torch/nn/modules/rnn.py # noqa + """ + # Short-circuits if _flat_weights is only partially instantiated + if len(flat_weights) != len(self._flat_weights_names): + return + + for w in flat_weights: + if not isinstance(w, Tensor): + return + # Short-circuits if any tensor in flat_weights is not acceptable to cuDNN + # or the tensors in flat_weights are of different dtypes + + first_fw = flat_weights[0] + dtype = first_fw.dtype + for fw in flat_weights: + if ( + not isinstance(fw.data, Tensor) + or not (fw.data.dtype == dtype) + or not fw.data.is_cuda + or not torch.backends.cudnn.is_acceptable(fw.data) + ): + return + + # If any parameters alias, we fall back to the slower, copying code path. This is + # a sufficient check, because overlapping parameter buffers that don't completely + # alias would break the assumptions of the uniqueness check in + # Module.named_parameters(). + unique_data_ptrs = set(p.data_ptr() for p in flat_weights) + if len(unique_data_ptrs) != len(flat_weights): + return + + with torch.cuda.device_of(first_fw): + + # Note: no_grad() is necessary since _cudnn_rnn_flatten_weight is + # an inplace operation on self._flat_weights + with torch.no_grad(): + if torch._use_cudnn_rnn_flatten_weight(): + num_weights = 4 if self.bias else 2 + if self.proj_size > 0: + num_weights += 1 + torch._cudnn_rnn_flatten_weight( + flat_weights, + num_weights, + self.input_size, + rnn.get_cudnn_mode(self.mode), + self.hidden_size, + self.proj_size, + self.num_layers, + self.batch_first, + bool(self.bidirectional), + ) + + def _get_flat_weights(self): + """Get scaled weights, and resets their data pointer.""" + flat_weights = [] + for idx in range(len(self._flat_weights_names)): + flat_weights.append(self._flat_weights[idx] * self._scales[idx].exp()) + self._flatten_parameters(flat_weights) + return flat_weights + + def forward(self, input: Tensor, hx: Optional[Tuple[Tensor, Tensor]] = None): + # This function is modified from https://github.com/pytorch/pytorch/blob/master/torch/nn/modules/rnn.py # noqa + # The change for calling `_VF.lstm()` is: + # self._flat_weights -> self._get_flat_weights() + if hx is None: + h_zeros = torch.zeros( + self.num_layers * self.num_directions, + input.size(self.batch_dim), + self.proj_size if self.proj_size > 0 else self.hidden_size, + dtype=input.dtype, + device=input.device, + ) + c_zeros = torch.zeros( + self.num_layers * self.num_directions, + input.size(self.batch_dim), + self.hidden_size, + dtype=input.dtype, + device=input.device, + ) + hx = (h_zeros, c_zeros) + + self.check_forward_args(input, hx, None) + + flat_weights = self._get_flat_weights() + input, *flat_weights = self.grad_filter(input, *flat_weights) + + result = _VF.lstm( + input, + hx, + flat_weights, + self.bias, + self.num_layers, + self.dropout, + self.training, + self.bidirectional, + self.batch_first, + ) + + output = result[0] + hidden = result[1:] + return output, hidden + + +class ActivationBalancer(torch.nn.Module): + """ + Modifies the backpropped derivatives of a function to try to encourage, for + each channel, that it is positive at least a proportion `threshold` of the + time. It does this by multiplying negative derivative values by up to + (1+max_factor), and positive derivative values by up to (1-max_factor), + interpolated from 1 at the threshold to those extremal values when none + of the inputs are positive. + + Args: + num_channels: the number of channels + channel_dim: the dimension/axis corresponding to the channel, e.g. + -1, 0, 1, 2; will be interpreted as an offset from x.ndim if negative. + min_positive: the minimum, per channel, of the proportion of the time + that (x > 0), below which we start to modify the derivatives. + max_positive: the maximum, per channel, of the proportion of the time + that (x > 0), above which we start to modify the derivatives. + max_factor: the maximum factor by which we modify the derivatives for + either the sign constraint or the magnitude constraint; + e.g. with max_factor=0.02, the the derivatives would be multiplied by + values in the range [0.98..1.02]. + sign_gain_factor: determines the 'gain' with which we increase the + change in gradient once the constraints on min_positive and max_positive + are violated. + scale_gain_factor: determines the 'gain' with which we increase the + change in gradient once the constraints on min_abs and max_abs + are violated. + min_abs: the minimum average-absolute-value difference from the mean + value per channel, which we allow, before we start to modify + the derivatives to prevent this. + max_abs: the maximum average-absolute-value difference from the mean + value per channel, which we allow, before we start to modify + the derivatives to prevent this. + min_prob: determines the minimum probability with which we modify the + gradients for the {min,max}_positive and {min,max}_abs constraints, + on each forward(). This is done randomly to prevent all layers + from doing it at the same time. Early in training we may use + higher probabilities than this; it will decay to this value. + """ + + def __init__( + self, + num_channels: int, + channel_dim: int, + min_positive: float = 0.05, + max_positive: float = 0.95, + max_factor: float = 0.04, + sign_gain_factor: float = 0.01, + scale_gain_factor: float = 0.02, + min_abs: float = 0.2, + max_abs: float = 100.0, + min_prob: float = 0.1, + ): + super(ActivationBalancer, self).__init__() + self.num_channels = num_channels + self.channel_dim = channel_dim + self.min_positive = min_positive + self.max_positive = max_positive + self.max_factor = max_factor + self.min_abs = min_abs + self.max_abs = max_abs + self.min_prob = min_prob + self.sign_gain_factor = sign_gain_factor + self.scale_gain_factor = scale_gain_factor + + # count measures how many times the forward() function has been called. + # We occasionally sync this to a tensor called `count`, that exists to + # make sure it is synced to disk when we load and save the model. + self.cpu_count = 0 + self.register_buffer("count", torch.tensor(0, dtype=torch.int64)) + + def forward(self, x: Tensor) -> Tensor: + if torch.jit.is_scripting() or not x.requires_grad or torch.jit.is_tracing(): + return _no_op(x) + + count = self.cpu_count + self.cpu_count += 1 + + if random.random() < 0.01: + # Occasionally sync self.cpu_count with self.count. + # count affects the decay of 'prob'. don't do this on every iter, + # because syncing with the GPU is slow. + self.cpu_count = max(self.cpu_count, self.count.item()) + self.count.fill_(self.cpu_count) + + # the prob of doing some work exponentially decreases from 0.5 till it hits + # a floor at min_prob (==0.1, by default) + prob = max(self.min_prob, 0.5 ** (1 + (count / 4000.0))) + + if random.random() < prob: + sign_gain_factor = 0.5 + if self.min_positive != 0.0 or self.max_positive != 1.0: + sign_factor = _compute_sign_factor( + x, + self.channel_dim, + self.min_positive, + self.max_positive, + gain_factor=self.sign_gain_factor / prob, + max_factor=self.max_factor, + ) + else: + sign_factor = None + + scale_factor = _compute_scale_factor( + x.detach(), + self.channel_dim, + min_abs=self.min_abs, + max_abs=self.max_abs, + gain_factor=self.scale_gain_factor / prob, + max_factor=self.max_factor, + ) + return ActivationBalancerFunction.apply( + x, + scale_factor, + sign_factor, + self.channel_dim, + ) + else: + return _no_op(x) + + +def penalize_abs_values_gt(x: Tensor, limit: float, penalty: float) -> Tensor: + """ + Returns x unmodified, but in backprop will put a penalty for the excess of + the absolute values of elements of x over the limit "limit". E.g. if + limit == 10.0, then if x has any values over 10 it will get a penalty. + + Caution: the value of this penalty will be affected by grad scaling used + in automatic mixed precision training. For this reasons we use this, + it shouldn't really matter, or may even be helpful; we just use this + to disallow really implausible values of scores to be given to softmax. + """ + x_sign = x.sign() + over_limit = (x.abs() - limit) > 0 + # The following is a memory efficient way to penalize the absolute values of + # x that's over the limit. (The memory efficiency comes when you think + # about which items torch needs to cache for the autograd, and which ones it + # can throw away). The numerical value of aux_loss as computed here will + # actually be larger than it should be, by limit * over_limit.sum(), but it + # has the same derivative as the real aux_loss which is penalty * (x.abs() - + # limit).relu(). + aux_loss = penalty * ((x_sign * over_limit).to(torch.int8) * x) + # note: we don't do sum() here on aux)_loss, but it's as if we had done + # sum() due to how with_loss() works. + x = with_loss(x, aux_loss) + # you must use x for something, or this will be ineffective. + return x + + +def _diag(x: Tensor): # like .diag(), but works for tensors with 3 dims. + if x.ndim == 2: + return x.diag() + else: + (batch, dim, dim) = x.shape + x = x.reshape(batch, dim * dim) + x = x[:, :: dim + 1] + assert x.shape == (batch, dim) + return x + + +def _whitening_metric(x: Tensor, num_groups: int): + """ + Computes the "whitening metric", a value which will be 1.0 if all the eigenvalues of + of the centered feature covariance are the same within each group's covariance matrix + and also between groups. + Args: + x: a Tensor of shape (*, num_channels) + num_groups: the number of groups of channels, a number >=1 that divides num_channels + Returns: + Returns a scalar Tensor that will be 1.0 if the data is "perfectly white" and + greater than 1.0 otherwise. + """ + assert x.dtype != torch.float16 + x = x.reshape(-1, x.shape[-1]) + (num_frames, num_channels) = x.shape + assert num_channels % num_groups == 0 + channels_per_group = num_channels // num_groups + x = x.reshape(num_frames, num_groups, channels_per_group).transpose(0, 1) + # x now has shape (num_groups, num_frames, channels_per_group) + # subtract the mean so we use the centered, not uncentered, covariance. + # My experience has been that when we "mess with the gradients" like this, + # it's better not do anything that tries to move the mean around, because + # that can easily cause instability. + x = x - x.mean(dim=1, keepdim=True) + # x_covar: (num_groups, channels_per_group, channels_per_group) + x_covar = torch.matmul(x.transpose(1, 2), x) + x_covar_mean_diag = _diag(x_covar).mean() + # the following expression is what we'd get if we took the matrix product + # of each covariance and measured the mean of its trace, i.e. + # the same as _diag(torch.matmul(x_covar, x_covar)).mean(). + x_covarsq_mean_diag = (x_covar**2).sum() / (num_groups * channels_per_group) + # this metric will be >= 1.0; the larger it is, the less 'white' the data was. + metric = x_covarsq_mean_diag / (x_covar_mean_diag**2 + 1.0e-20) + return metric + + +class WhiteningPenaltyFunction(torch.autograd.Function): + @staticmethod + def forward( + ctx, x: Tensor, num_groups: int, whitening_limit: float, grad_scale: float + ) -> Tensor: + ctx.save_for_backward(x) + ctx.num_groups = num_groups + ctx.whitening_limit = whitening_limit + ctx.grad_scale = grad_scale + return x + + @staticmethod + def backward(ctx, x_grad: Tensor): + (x_orig,) = ctx.saved_tensors + with torch.enable_grad(): + with torch.cuda.amp.autocast(enabled=False): + x_detached = x_orig.to(torch.float32).detach() + x_detached.requires_grad = True + + metric = _whitening_metric(x_detached, ctx.num_groups) + + if random.random() < 0.005 or __name__ == "__main__": + logging.info( + f"Whitening: num_groups={ctx.num_groups}, num_channels={x_orig.shape[-1]}, " + f"metric={metric.item():.2f} vs. limit={ctx.whitening_limit}" + ) + + (metric - ctx.whitening_limit).relu().backward() + penalty_grad = x_detached.grad + scale = ctx.grad_scale * ( + x_grad.to(torch.float32).norm() / (penalty_grad.norm() + 1.0e-20) + ) + penalty_grad = penalty_grad * scale + return x_grad + penalty_grad.to(x_grad.dtype), None, None, None + + +class Whiten(nn.Module): + def __init__( + self, + num_groups: int, + whitening_limit: float, + prob: Union[float, Tuple[float, float]], + grad_scale: float, + ): + """ + Args: + num_groups: the number of groups to divide the channel dim into before + whitening. We will attempt to make the feature covariance + within each group, after mean subtraction, as "white" as possible, + while having the same trace across all groups. + whitening_limit: a value greater than 1.0, that dictates how much + freedom we have to violate the constraints. 1.0 would mean perfectly + white, with exactly the same trace across groups; larger values + give more freedom. E.g. 2.0. + prob: the probability with which we apply the gradient modification + (also affects the grad scale). May be supplied as a float, + or as a pair (min_prob, max_prob) + + grad_scale: determines the scale on the gradient term from this object, + relative to the rest of the gradient on the attention weights. + E.g. 0.02 (you may want to use smaller values than this if prob is large) + """ + super(Whiten, self).__init__() + assert num_groups >= 1 + assert whitening_limit >= 1 + assert grad_scale >= 0 + self.num_groups = num_groups + self.whitening_limit = whitening_limit + if isinstance(prob, float): + assert 0 < prob <= 1 + self.prob = prob + else: + (self.min_prob, self.max_prob) = prob + assert 0 < self.min_prob < self.max_prob <= 1 + self.prob = self.max_prob + + self.grad_scale = grad_scale + + def forward(self, x: Tensor) -> Tensor: + """ + In the forward pass, this function just returns the input unmodified. + In the backward pass, it will modify the gradients to ensure that the + distribution in each group has close to (lambda times I) as the covariance + after mean subtraction, with the same lambda across groups. + For whitening_limit > 1, there will be more freedom to violate this + constraint. + + Args: + x: the input of shape (*, num_channels) + + Returns: + x, unmodified. You should make sure + you use the returned value, or the graph will be freed + and nothing will happen in backprop. + """ + if not x.requires_grad or random.random() > self.prob or self.grad_scale == 0: + return _no_op(x) + else: + if hasattr(self, "min_prob") and random.random() < 0.25: + # occasionally switch between min_prob and max_prob, based on whether + # we are above or below the threshold. + if ( + _whitening_metric(x.to(torch.float32), self.num_groups) + > self.whitening_limit + ): + # there would be a change to the grad. + self.prob = self.max_prob + else: + self.prob = self.min_prob + + return WhiteningPenaltyFunction.apply( + x, self.num_groups, self.whitening_limit, self.grad_scale + ) + + +class WithLoss(torch.autograd.Function): + @staticmethod + def forward(ctx, x: Tensor, y: Tensor): + ctx.y_shape = y.shape + return x + + @staticmethod + def backward(ctx, ans_grad: Tensor): + return ( + ans_grad, + torch.ones(ctx.y_shape, dtype=ans_grad.dtype, device=ans_grad.device), + ) + + +def with_loss(x, y): + if torch.jit.is_scripting() or torch.jit.is_tracing(): + return x + # returns x but adds y.sum() to the loss function. + return WithLoss.apply(x, y) + + +def _no_op(x: Tensor) -> Tensor: + if torch.jit.is_scripting() or torch.jit.is_tracing(): + return x + else: + # a no-op function that will have a node in the autograd graph, + # to avoid certain bugs relating to backward hooks + return x.chunk(1, dim=-1)[0] + + +class Identity(torch.nn.Module): + def __init__(self): + super(Identity, self).__init__() + + def forward(self, x): + return _no_op(x) + + +class MaxEig(torch.nn.Module): + """ + Modifies the backpropped derivatives of a function to try to discourage + that any given direction in activation space accounts for more than + a specified proportion of the covariance (e.g. 0.2). + + + Args: + num_channels: the number of channels + channel_dim: the dimension/axis corresponding to the channel, e.g. + -1, 0, 1, 2; will be interpreted as an offset from x.ndim if negative. + max_var_per_eig: the maximum proportion of the variance of the + features/channels, after mean subtraction, that can come from + any given eigenvalue. + min_prob: the minimum probability with which we apply this during any invocation + of forward(), assuming last time we applied the constraint it was + not active; supplied for speed. + scale: determines the scale with which we modify the gradients, relative + to the existing / unmodified gradients + """ + + def __init__( + self, + num_channels: int, + channel_dim: int, + max_var_per_eig: float = 0.2, + min_prob: float = 0.01, + scale: float = 0.01, + ): + super(MaxEig, self).__init__() + self.num_channels = num_channels + self.channel_dim = channel_dim + self.scale = scale + assert max_var_per_eig == 0.0 or max_var_per_eig > 1.0 / num_channels + self.max_var_per_eig = max_var_per_eig + + # we figure out the dominant direction using the power method: starting with + # a random vector, keep multiplying by the covariance and renormalizing. + with torch.no_grad(): + # arbitrary.. would use randn() but want to leave the rest of the model's + # random parameters unchanged for comparison + direction = torch.arange(num_channels).to(torch.float) + direction = direction / direction.norm() + self.register_buffer("max_eig_direction", direction) + + self.min_prob = min_prob + # cur_prob is the current probability we'll use to apply the ActivationBalancer. + # We'll regress this towards prob, each tiem we try to apply it and it is not + # active. + self.cur_prob = 1.0 + + def forward(self, x: Tensor) -> Tensor: + if ( + torch.jit.is_scripting() + or self.max_var_per_eig <= 0 + or random.random() > self.cur_prob + or torch.jit.is_tracing() + ): + return _no_op(x) + + with torch.cuda.amp.autocast(enabled=False): + eps = 1.0e-20 + orig_x = x + x = x.to(torch.float32) + with torch.no_grad(): + x = x.transpose(self.channel_dim, -1).reshape(-1, self.num_channels) + x = x - x.mean(dim=0) + new_direction, coeffs = self._find_direction_coeffs( + x, self.max_eig_direction + ) + x_var = (x**2).mean() + x_residual = x - coeffs * new_direction + x_residual_var = (x_residual**2).mean() + + # `variance_proportion` is the proportion of the variance accounted for + # by the top eigen-direction. + variance_proportion = (x_var - x_residual_var) / (x_var + 1.0e-20) + + # ensure new direction is nonzero even if x == 0, by including `direction`. + self._set_direction(0.1 * self.max_eig_direction + new_direction) + + if random.random() < 0.01 or __name__ == "__main__": + logging.info( + f"variance_proportion = {variance_proportion.item()}, shape={tuple(orig_x.shape)}, cur_prob={self.cur_prob}" + ) + + if variance_proportion >= self.max_var_per_eig: + # The constraint is active. Note, we should quite rarely + # reach here, only near the beginning of training if we are + # starting to diverge, should this constraint be active. + cur_prob = self.cur_prob + self.cur_prob = 1.0 # next time, do the update with probability 1.0. + return MaxEigLimiterFunction.apply( + orig_x, coeffs, new_direction, self.channel_dim, self.scale + ) + else: + # let self.cur_prob exponentially approach self.min_prob, as + # long as the constraint is inactive. + self.cur_prob = 0.75 * self.cur_prob + 0.25 * self.min_prob + return orig_x + + def _set_direction(self, direction: Tensor): + """ + Sets self.max_eig_direction to a normalized version of `direction` + """ + direction = direction.detach() + direction = direction / direction.norm() + direction_sum = direction.sum().item() + if direction_sum - direction_sum == 0: # no inf/nan + self.max_eig_direction[:] = direction + else: + logging.info( + f"Warning: sum of direction in MaxEig is {direction_sum}, " + "num_channels={self.num_channels}, channel_dim={self.channel_dim}" + ) + + def _find_direction_coeffs( + self, x: Tensor, prev_direction: Tensor + ) -> Tuple[Tensor, Tensor, Tensor]: + """ + Figure out (an approximation to) the proportion of the variance of a set of + feature vectors that can be attributed to the top eigen-direction. + Args: + x: a Tensor of shape (num_frames, num_channels), with num_frames > 1. + prev_direction: a Tensor of shape (num_channels,), that is our previous estimate + of the top eigen-direction, or a random direction if this is the first + iteration. Does not have to be normalized, but should be nonzero. + + Returns: (cur_direction, coeffs), where: + cur_direction: a Tensor of shape (num_channels,) that is the current + estimate of the top eigen-direction. + coeffs: a Tensor of shape (num_frames, 1) that minimizes, or + approximately minimizes, (x - coeffs * cur_direction).norm() + """ + (num_frames, num_channels) = x.shape + assert num_channels > 1 and num_frames > 1 + assert prev_direction.shape == (num_channels,) + # `coeffs` are the coefficients of `prev_direction` in x. + # actually represent the coeffs up to a constant positive factor. + coeffs = (x * prev_direction).sum(dim=1, keepdim=True) + 1.0e-10 + cur_direction = (x * coeffs).sum(dim=0) / ((coeffs**2).sum() + 1.0e-20) + return cur_direction, coeffs + + +class DoubleSwishFunction(torch.autograd.Function): + """ + double_swish(x) = x * torch.sigmoid(x-1) + This is a definition, originally motivated by its close numerical + similarity to swish(swish(x)), where swish(x) = x * sigmoid(x). + + Memory-efficient derivative computation: + double_swish(x) = x * s, where s(x) = torch.sigmoid(x-1) + double_swish'(x) = d/dx double_swish(x) = x * s'(x) + x' * s(x) = x * s'(x) + s(x). + Now, s'(x) = s(x) * (1-s(x)). + double_swish'(x) = x * s'(x) + s(x). + = x * s(x) * (1-s(x)) + s(x). + = double_swish(x) * (1-s(x)) + s(x) + ... so we just need to remember s(x) but not x itself. + """ + + @staticmethod + def forward(ctx, x: Tensor) -> Tensor: + requires_grad = x.requires_grad + x_dtype = x.dtype + if x.dtype == torch.float16: + x = x.to(torch.float32) + + s = torch.sigmoid(x - 1.0) + y = x * s + + if requires_grad: + deriv = y * (1 - s) + s + # notes on derivative of x * sigmoid(x - 1): + # https://www.wolframalpha.com/input?i=d%2Fdx+%28x+*+sigmoid%28x-1%29%29 + # min \simeq -0.043638. Take floor as -0.043637 so it's a lower bund + # max \simeq 1.1990. Take ceil to be 1.2 so it's an upper bound. + # the combination of "+ torch.rand_like(deriv)" and casting to torch.uint8 (which + # floors), should be expectation-preserving. + floor = -0.043637 + ceil = 1.2 + d_scaled = (deriv - floor) * (255.0 / (ceil - floor)) + torch.rand_like( + deriv + ) + if __name__ == "__main__": + # for self-testing only. + assert d_scaled.min() >= 0.0 + assert d_scaled.max() < 256.0 + d_int = d_scaled.to(torch.uint8) + ctx.save_for_backward(d_int) + if x.dtype == torch.float16 or torch.is_autocast_enabled(): + y = y.to(torch.float16) + return y + + @staticmethod + def backward(ctx, y_grad: Tensor) -> Tensor: + (d,) = ctx.saved_tensors + # the same constants as used in forward pass. + floor = -0.043637 + ceil = 1.2 + d = d * ((ceil - floor) / 255.0) + floor + return y_grad * d + + +class DoubleSwish(torch.nn.Module): + def forward(self, x: Tensor) -> Tensor: + """Return double-swish activation function which is an approximation to Swish(Swish(x)), + that we approximate closely with x * sigmoid(x-1). + """ + if torch.jit.is_scripting() or torch.jit.is_tracing(): + return x * torch.sigmoid(x - 1.0) + return DoubleSwishFunction.apply(x) + + +def _test_max_eig(): + for proportion in [0.1, 0.5, 10.0]: + logging.info(f"proportion = {proportion}") + x = torch.randn(100, 128) + direction = torch.randn(128) + coeffs = torch.randn(100, 1) + x += proportion * direction * coeffs + + x.requires_grad = True + + num_channels = 128 + m = MaxEig( + num_channels, 1, 0.5, scale=0.1 # channel_dim # max_var_per_eig + ) # grad_scale + + for _ in range(4): + y = m(x) + + y_grad = torch.randn_like(x) + y.backward(gradient=y_grad) + + if proportion < 0.2: + assert torch.allclose(x.grad, y_grad, atol=1.0e-02) + elif proportion > 1.0: + assert not torch.allclose(x.grad, y_grad) + + +def _test_whiten(): + for proportion in [0.1, 0.5, 10.0]: + logging.info(f"_test_whiten(): proportion = {proportion}") + x = torch.randn(100, 128) + direction = torch.randn(128) + coeffs = torch.randn(100, 1) + x += proportion * direction * coeffs + + x.requires_grad = True + + num_channels = 128 + m = Whiten( + 1, 5.0, prob=1.0, grad_scale=0.1 # num_groups # whitening_limit, + ) # grad_scale + + for _ in range(4): + y = m(x) + + y_grad = torch.randn_like(x) + y.backward(gradient=y_grad) + + if proportion < 0.2: + assert torch.allclose(x.grad, y_grad) + elif proportion > 1.0: + assert not torch.allclose(x.grad, y_grad) + + +def _test_activation_balancer_sign(): + probs = torch.arange(0, 1, 0.01) + N = 1000 + x = 1.0 * ((2.0 * (torch.rand(probs.numel(), N) < probs.unsqueeze(-1))) - 1.0) + x = x.detach() + x.requires_grad = True + m = ActivationBalancer( + probs.numel(), + channel_dim=0, + min_positive=0.05, + max_positive=0.95, + max_factor=0.2, + min_abs=0.0, + ) + + y_grad = torch.sign(torch.randn(probs.numel(), N)) + + y = m(x) + y.backward(gradient=y_grad) + print("_test_activation_balancer_sign: x = ", x) + print("_test_activation_balancer_sign: y grad = ", y_grad) + print("_test_activation_balancer_sign: x grad = ", x.grad) + + +def _test_activation_balancer_magnitude(): + magnitudes = torch.arange(0, 1, 0.01) + N = 1000 + x = torch.sign(torch.randn(magnitudes.numel(), N)) * magnitudes.unsqueeze(-1) + x = x.detach() + x.requires_grad = True + m = ActivationBalancer( + magnitudes.numel(), + channel_dim=0, + min_positive=0.0, + max_positive=1.0, + max_factor=0.2, + min_abs=0.2, + max_abs=0.8, + min_prob=1.0, + ) + + y_grad = torch.sign(torch.randn(magnitudes.numel(), N)) + + y = m(x) + y.backward(gradient=y_grad) + print("_test_activation_balancer_magnitude: x = ", x) + print("_test_activation_balancer_magnitude: y grad = ", y_grad) + print("_test_activation_balancer_magnitude: x grad = ", x.grad) + + +def _test_basic_norm(): + num_channels = 128 + m = BasicNorm(num_channels=num_channels, channel_dim=1) + + x = torch.randn(500, num_channels) + + y = m(x) + + assert y.shape == x.shape + x_rms = (x**2).mean().sqrt() + y_rms = (y**2).mean().sqrt() + print("x rms = ", x_rms) + print("y rms = ", y_rms) + assert y_rms < x_rms + assert y_rms > 0.5 * x_rms + + +def _test_double_swish_deriv(): + x = torch.randn(10, 12, dtype=torch.double) * 3.0 + x.requires_grad = True + m = DoubleSwish() + + tol = (1.2 - (-0.043637)) / 255.0 + torch.autograd.gradcheck(m, x, atol=tol) + + # for self-test. + x = torch.randn(1000, 1000, dtype=torch.double) * 3.0 + x.requires_grad = True + y = m(x) + + +def _test_softmax(): + a = torch.randn(2, 10, dtype=torch.float64) + b = a.clone() + a.requires_grad = True + b.requires_grad = True + a.softmax(dim=1)[:, 0].sum().backward() + print("a grad = ", a.grad) + softmax(b, dim=1)[:, 0].sum().backward() + print("b grad = ", b.grad) + assert torch.allclose(a.grad, b.grad) + + +if __name__ == "__main__": + logging.getLogger().setLevel(logging.INFO) + torch.set_num_threads(1) + torch.set_num_interop_threads(1) + _test_softmax() + _test_whiten() + _test_max_eig() + _test_activation_balancer_sign() + _test_activation_balancer_magnitude() + _test_basic_norm() + _test_double_swish_deriv() diff --git a/egs/libricss/SURT/prepare.sh b/egs/libricss/SURT/prepare.sh index 028240e44..3d2581d96 100755 --- a/egs/libricss/SURT/prepare.sh +++ b/egs/libricss/SURT/prepare.sh @@ -79,7 +79,7 @@ if [ $stage -le 0 ] && [ $stop_stage -ge 0 ]; then # ln -sfv /path/to/rirs_noises $dl_dir/ # if [ ! -d $dl_dir/rirs_noises ]; then - lhotse download rirs_noises $dl_dir + lhotse download rir-noise $dl_dir/rirs_noises fi fi @@ -89,6 +89,7 @@ if [ $stage -le 1 ] && [ $stop_stage -ge 1 ]; then # to $dl_dir/librispeech. We perform text normalization for the transcripts. # NOTE: Alignments are required for this recipe. mkdir -p data/manifests + lhotse prepare librispeech -p train-clean-100 -p train-clean-360 -p train-other-500 -p dev-clean \ -j 4 --alignments-dir $dl_dir/libri_alignments/LibriSpeech $dl_dir/librispeech data/manifests/ fi @@ -112,7 +113,7 @@ if [ $stage -le 3 ] && [ $stop_stage -ge 3 ]; then # We assume that you have downloaded the RIRS_NOISES corpus # to $dl_dir/rirs_noises - lhotse prepare rir-noise -p real_rir -p iso_noise $dl_dir/rirs_noises data/manifests + lhotse prepare rir-noise -p real_rir -p iso_noise $dl_dir/rirs_noises/RIRS_NOISES data/manifests fi if [ $stage -le 4 ] && [ $stop_stage -ge 4 ]; then diff --git a/egs/librispeech/ASR/pruned2_knowledge/asr_datamodule.py b/egs/librispeech/ASR/pruned2_knowledge/asr_datamodule.py index b839a4a4c..f8f558ce1 100644 --- a/egs/librispeech/ASR/pruned2_knowledge/asr_datamodule.py +++ b/egs/librispeech/ASR/pruned2_knowledge/asr_datamodule.py @@ -31,7 +31,7 @@ from lhotse.dataset import ( CutMix, K2SpeechRecognitionDataset, PrecomputedFeatures, - SingleCutSampler, + SimpleCutSampler, SpecAugment, ) from lhotse.dataset.input_strategies import OnTheFlyFeatures @@ -290,8 +290,8 @@ class LibriSpeechAsrDataModule: drop_last=True, ) else: - logging.info("Using SingleCutSampler.") - train_sampler = SingleCutSampler( + logging.info("Using SimpleCutSampler.") + train_sampler = SimpleCutSampler( cuts_train, max_duration=self.args.max_duration, shuffle=self.args.shuffle, diff --git a/egs/librispeech/ASR/pruned_transducer_stateless7/compute_ali.py b/egs/librispeech/ASR/pruned_transducer_stateless7/compute_ali.py index 8bcb56d62..27ef0a244 100755 --- a/egs/librispeech/ASR/pruned_transducer_stateless7/compute_ali.py +++ b/egs/librispeech/ASR/pruned_transducer_stateless7/compute_ali.py @@ -26,7 +26,7 @@ You can generate the checkpoint with the following command: ./pruned_transducer_stateless7/export.py \ --exp-dir ./pruned_transducer_stateless7/exp \ - --bpe-model data/lang_bpe_500/bpe.model \ + --tokens data/lang_bpe_500/tokens.txt \ --epoch 30 \ --avg 9 @@ -52,12 +52,12 @@ import torch import torch.nn as nn from alignment import batch_force_alignment from asr_datamodule import LibriSpeechAsrDataModule -from train import add_model_arguments, get_params, get_transducer_model - -from icefall.utils import AttributeDict, convert_timestamp, parse_timestamp from lhotse import CutSet from lhotse.serialization import SequentialJsonlWriter from lhotse.supervision import AlignmentItem +from train import add_model_arguments, get_params, get_transducer_model + +from icefall.utils import AttributeDict, convert_timestamp, parse_timestamp def get_parser(): diff --git a/egs/librispeech/ASR/pruned_transducer_stateless7/decoder.py b/egs/librispeech/ASR/pruned_transducer_stateless7/decoder.py index b085a1817..bfd019ff5 100644 --- a/egs/librispeech/ASR/pruned_transducer_stateless7/decoder.py +++ b/egs/librispeech/ASR/pruned_transducer_stateless7/decoder.py @@ -71,6 +71,10 @@ class Decoder(nn.Module): groups=decoder_dim // 4, # group size == 4 bias=False, ) + else: + # To avoid `RuntimeError: Module 'Decoder' has no attribute 'conv'` + # when inference with torch.jit.script and context_size == 1 + self.conv = nn.Identity() def forward(self, y: torch.Tensor, need_pad: bool = True) -> torch.Tensor: """ diff --git a/egs/librispeech/ASR/pruned_transducer_stateless7/gigaspeech.py b/egs/librispeech/ASR/pruned_transducer_stateless7/gigaspeech.py index 5c01d7190..75e153cb0 100644 --- a/egs/librispeech/ASR/pruned_transducer_stateless7/gigaspeech.py +++ b/egs/librispeech/ASR/pruned_transducer_stateless7/gigaspeech.py @@ -30,7 +30,7 @@ from lhotse.dataset import ( DynamicBucketingSampler, K2SpeechRecognitionDataset, PrecomputedFeatures, - SingleCutSampler, + SimpleCutSampler, SpecAugment, ) from lhotse.dataset.input_strategies import OnTheFlyFeatures @@ -297,8 +297,8 @@ class GigaSpeechAsrDataModule: drop_last=True, ) else: - logging.info("Using SingleCutSampler.") - train_sampler = SingleCutSampler( + logging.info("Using SimpleCutSampler.") + train_sampler = SimpleCutSampler( cuts_train, max_duration=self.args.max_duration, shuffle=self.args.shuffle, diff --git a/egs/librispeech/ASR/tdnn_lstm_ctc/asr_datamodule.py b/egs/librispeech/ASR/tdnn_lstm_ctc/asr_datamodule.py index c47964b07..20df469da 100644 --- a/egs/librispeech/ASR/tdnn_lstm_ctc/asr_datamodule.py +++ b/egs/librispeech/ASR/tdnn_lstm_ctc/asr_datamodule.py @@ -31,7 +31,7 @@ from lhotse.dataset import ( # noqa F401 for PrecomputedFeatures DynamicBucketingSampler, K2SpeechRecognitionDataset, PrecomputedFeatures, - SingleCutSampler, + SimpleCutSampler, SpecAugment, ) from lhotse.dataset.input_strategies import ( # noqa F401 For AudioSamples @@ -314,8 +314,8 @@ class LibriSpeechAsrDataModule: drop_last=self.args.drop_last, ) else: - logging.info("Using SingleCutSampler.") - train_sampler = SingleCutSampler( + logging.info("Using SimpleCutSampler.") + train_sampler = SimpleCutSampler( cuts_train, max_duration=self.args.max_duration, shuffle=self.args.shuffle, diff --git a/egs/librispeech/ASR/zipformer/zipformer.py b/egs/librispeech/ASR/zipformer/zipformer.py index b39af02b8..1a174b315 100644 --- a/egs/librispeech/ASR/zipformer/zipformer.py +++ b/egs/librispeech/ASR/zipformer/zipformer.py @@ -91,34 +91,34 @@ class Zipformer2(EncoderInterface): chunks. Must not be less than cnn_module_kernel (after factoring in rounding and downsampling); an error will be thrown if this is violated. """ + def __init__( - self, - output_downsampling_factor: int = 2, - downsampling_factor: Tuple[int] = (2, 4), - encoder_dim: Union[int, Tuple[int]] = 384, - num_encoder_layers: Union[int, Tuple[int]] = 4, - encoder_unmasked_dim: Union[int, Tuple[int]] = 256, - query_head_dim: Union[int, Tuple[int]] = 24, - pos_head_dim: Union[int, Tuple[int]] = 4, - value_head_dim: Union[int, Tuple[int]] = 12, - num_heads: Union[int, Tuple[int]] = 8, - feedforward_dim: Union[int, Tuple[int]] = 1536, - cnn_module_kernel: Union[int, Tuple[int]] = 31, - pos_dim: int = 192, - dropout: FloatLike = None, # see code below for default - warmup_batches: float = 4000.0, - causal: bool = False, - chunk_size: Tuple[int] = [-1], - left_context_frames: Tuple[int] = [-1], + self, + output_downsampling_factor: int = 2, + downsampling_factor: Tuple[int] = (2, 4), + encoder_dim: Union[int, Tuple[int]] = 384, + num_encoder_layers: Union[int, Tuple[int]] = 4, + encoder_unmasked_dim: Union[int, Tuple[int]] = 256, + query_head_dim: Union[int, Tuple[int]] = 24, + pos_head_dim: Union[int, Tuple[int]] = 4, + value_head_dim: Union[int, Tuple[int]] = 12, + num_heads: Union[int, Tuple[int]] = 8, + feedforward_dim: Union[int, Tuple[int]] = 1536, + cnn_module_kernel: Union[int, Tuple[int]] = 31, + pos_dim: int = 192, + dropout: FloatLike = None, # see code below for default + warmup_batches: float = 4000.0, + causal: bool = False, + chunk_size: Tuple[int] = [-1], + left_context_frames: Tuple[int] = [-1], ) -> None: super(Zipformer2, self).__init__() if dropout is None: - dropout = ScheduledFloat((0.0, 0.3), - (20000.0, 0.1)) + dropout = ScheduledFloat((0.0, 0.3), (20000.0, 0.1)) def _to_tuple(x): - """ Converts a single int or a 1-tuple of an int to a tuple with the same length + """Converts a single int or a 1-tuple of an int to a tuple with the same length as downsampling_factor""" if isinstance(x, int): x = (x,) @@ -128,10 +128,12 @@ class Zipformer2(EncoderInterface): assert len(x) == len(downsampling_factor) and isinstance(x[0], int) return x - self.output_downsampling_factor = output_downsampling_factor # int - self.downsampling_factor = downsampling_factor # tuple - self.encoder_dim = encoder_dim = _to_tuple(encoder_dim) # tuple - self.encoder_unmasked_dim = encoder_unmasked_dim = _to_tuple(encoder_unmasked_dim) # tuple + self.output_downsampling_factor = output_downsampling_factor # int + self.downsampling_factor = downsampling_factor # tuple + self.encoder_dim = encoder_dim = _to_tuple(encoder_dim) # tuple + self.encoder_unmasked_dim = encoder_unmasked_dim = _to_tuple( + encoder_unmasked_dim + ) # tuple num_encoder_layers = _to_tuple(num_encoder_layers) self.num_encoder_layers = num_encoder_layers self.query_head_dim = query_head_dim = _to_tuple(query_head_dim) @@ -145,7 +147,7 @@ class Zipformer2(EncoderInterface): self.chunk_size = chunk_size self.left_context_frames = left_context_frames - for u,d in zip(encoder_unmasked_dim, encoder_dim): + for u, d in zip(encoder_unmasked_dim, encoder_dim): assert u <= d # each one will be Zipformer2Encoder or DownsampledZipformer2Encoder @@ -153,7 +155,6 @@ class Zipformer2(EncoderInterface): num_encoders = len(downsampling_factor) for i in range(num_encoders): - encoder_layer = Zipformer2EncoderLayer( embed_dim=encoder_dim[i], pos_dim=pos_dim, @@ -191,13 +192,11 @@ class Zipformer2(EncoderInterface): self.encoders = nn.ModuleList(encoders) - self.downsample_output = SimpleDownsample(max(encoder_dim), - downsample=output_downsampling_factor, - dropout=dropout) + self.downsample_output = SimpleDownsample( + max(encoder_dim), downsample=output_downsampling_factor, dropout=dropout + ) - def get_feature_masks( - self, - x: Tensor) -> Union[List[float], List[Tensor]]: + def get_feature_masks(self, x: Tensor) -> Union[List[float], List[Tensor]]: """ In eval mode, returns [1.0] * num_encoders; in training mode, returns a number of randomized feature masks, one per encoder. @@ -215,24 +214,30 @@ class Zipformer2(EncoderInterface): """ num_encoders = len(self.encoder_dim) if not self.training: - return [ 1.0 ] * num_encoders + return [1.0] * num_encoders (num_frames0, batch_size, _encoder_dims0) = x.shape - assert self.encoder_dim[0] == _encoder_dims0, (self.encoder_dim[0], _encoder_dims0) + assert self.encoder_dim[0] == _encoder_dims0, ( + self.encoder_dim[0], + _encoder_dims0, + ) feature_mask_dropout_prob = 0.125 # mask1 shape: (1, batch_size, 1) - mask1 = (torch.rand(1, batch_size, 1, - device=x.device) > - feature_mask_dropout_prob).to(x.dtype) + mask1 = ( + torch.rand(1, batch_size, 1, device=x.device) > feature_mask_dropout_prob + ).to(x.dtype) # mask2 has additional sequences masked, about twice the number. - mask2 = torch.logical_and(mask1, - (torch.rand(1, batch_size, 1, - device=x.device) > - feature_mask_dropout_prob).to(x.dtype)) + mask2 = torch.logical_and( + mask1, + ( + torch.rand(1, batch_size, 1, device=x.device) + > feature_mask_dropout_prob + ).to(x.dtype), + ) # dim: (1, batch_size, 2) mask = torch.cat((mask1, mask2), dim=-1) @@ -240,8 +245,9 @@ class Zipformer2(EncoderInterface): feature_masks = [] for i in range(num_encoders): channels = self.encoder_dim[i] - feature_mask = torch.ones(1, batch_size, channels, - dtype=x.dtype, device=x.device) + feature_mask = torch.ones( + 1, batch_size, channels, dtype=x.dtype, device=x.device + ) u1 = self.encoder_unmasked_dim[i] u2 = u1 + (channels - u1) // 2 @@ -281,7 +287,8 @@ class Zipformer2(EncoderInterface): return chunk_size, left_context_chunks def forward( - self, x: Tensor, + self, + x: Tensor, x_lens: Tensor, src_key_padding_mask: Optional[Tensor] = None, ) -> Tuple[Tensor, Tensor]: @@ -319,12 +326,17 @@ class Zipformer2(EncoderInterface): ds = self.downsampling_factor[i] x = convert_num_channels(x, self.encoder_dim[i]) - x = module(x, - chunk_size=chunk_size, - feature_mask=feature_masks[i], - src_key_padding_mask=(None if src_key_padding_mask is None - else src_key_padding_mask[...,::ds]), - attn_mask=attn_mask) + x = module( + x, + chunk_size=chunk_size, + feature_mask=feature_masks[i], + src_key_padding_mask=( + None + if src_key_padding_mask is None + else src_key_padding_mask[..., ::ds] + ), + attn_mask=attn_mask, + ) outputs.append(x) # if the last output has the largest dimension, x will be unchanged, @@ -345,9 +357,7 @@ class Zipformer2(EncoderInterface): return x, lengths def _get_attn_mask( - self, x: Tensor, - chunk_size: int, - left_context_chunks: int + self, x: Tensor, chunk_size: int, left_context_chunks: int ) -> Optional[Tensor]: """ Return None if chunk_size == -1, else return attention mask of shape @@ -362,9 +372,11 @@ class Zipformer2(EncoderInterface): assert all(chunk_size % d == 0 for d in self.downsampling_factor) if left_context_chunks >= 0: num_encoders = len(self.encoder_dim) - assert all (chunk_size * left_context_chunks >= - (self.cnn_module_kernel[i] // 2) * self.downsampling_factor[i] - for i in range(num_encoders)) + assert all( + chunk_size * left_context_chunks + >= (self.cnn_module_kernel[i] // 2) * self.downsampling_factor[i] + for i in range(num_encoders) + ) else: left_context_chunks = 1000000 @@ -382,8 +394,7 @@ class Zipformer2(EncoderInterface): src_c = c tgt_c = c.unsqueeze(-1) - attn_mask = torch.logical_or(src_c > tgt_c, - src_c < tgt_c - left_context_chunks) + attn_mask = torch.logical_or(src_c > tgt_c, src_c < tgt_c - left_context_chunks) if __name__ == "__main__": logging.info(f"attn_mask = {attn_mask}") return attn_mask @@ -392,7 +403,7 @@ class Zipformer2(EncoderInterface): num_encoders = len(self.encoder_dim) assert len(outputs) == num_encoders output_dim = max(self.encoder_dim) - output_pieces = [ outputs[-1] ] + output_pieces = [outputs[-1]] cur_dim = self.encoder_dim[-1] for i in range(num_encoders - 2, -1, -1): d = self.encoder_dim[i] @@ -489,21 +500,38 @@ class Zipformer2(EncoderInterface): nonlin_attn_head_dim = 3 * embed_dim // 4 conv_left_pad = self.cnn_module_kernel[i] // 2 for layer in range(num_layers): - cached_key = torch.zeros(downsample_left, batch_size, key_dim).to(device) - cached_nonlin_attn = torch.zeros(1, batch_size, downsample_left, nonlin_attn_head_dim).to(device) - cached_val1 = torch.zeros(downsample_left, batch_size, value_dim).to(device) - cached_val2 = torch.zeros(downsample_left, batch_size, value_dim).to(device) - cached_conv1 = torch.zeros(batch_size, embed_dim, conv_left_pad).to(device) - cached_conv2 = torch.zeros(batch_size, embed_dim, conv_left_pad).to(device) - states += [cached_key, cached_nonlin_attn, cached_val1, cached_val2, cached_conv1, cached_conv2] + cached_key = torch.zeros(downsample_left, batch_size, key_dim).to( + device + ) + cached_nonlin_attn = torch.zeros( + 1, batch_size, downsample_left, nonlin_attn_head_dim + ).to(device) + cached_val1 = torch.zeros(downsample_left, batch_size, value_dim).to( + device + ) + cached_val2 = torch.zeros(downsample_left, batch_size, value_dim).to( + device + ) + cached_conv1 = torch.zeros(batch_size, embed_dim, conv_left_pad).to( + device + ) + cached_conv2 = torch.zeros(batch_size, embed_dim, conv_left_pad).to( + device + ) + states += [ + cached_key, + cached_nonlin_attn, + cached_val1, + cached_val2, + cached_conv1, + cached_conv2, + ] return states def _whitening_schedule(x: float, ratio: float = 2.0) -> ScheduledFloat: - return ScheduledFloat((0.0, x), - (20000.0, ratio * x), - default=x) + return ScheduledFloat((0.0, x), (20000.0, ratio * x), default=x) def _balancer_schedule(min_prob: float): @@ -525,31 +553,45 @@ class Zipformer2EncoderLayer(nn.Module): >>> pos_emb = torch.rand(32, 19, 512) >>> out = encoder_layer(src, pos_emb) """ + def __init__( - self, - embed_dim: int, - pos_dim: int, - num_heads: int, - query_head_dim: int, - pos_head_dim: int, - value_head_dim: int, - feedforward_dim: int, - dropout: FloatLike = 0.1, - cnn_module_kernel: int = 31, - causal: bool = False, - attention_skip_rate: FloatLike = ScheduledFloat((0.0, 0.2), (4000.0, 0.05), (16000, 0.0), default=0), - conv_skip_rate: FloatLike = ScheduledFloat((0.0, 0.2), (4000.0, 0.05), (16000, 0.0), default=0), - const_attention_rate: FloatLike = ScheduledFloat((0.0, 0.25), (4000.0, 0.025), default=0), - ff2_skip_rate: FloatLike = ScheduledFloat((0.0, 0.1), (4000.0, 0.01), (50000.0, 0.0)), - ff3_skip_rate: FloatLike = ScheduledFloat((0.0, 0.1), (4000.0, 0.01), (50000.0, 0.0)), - bypass_skip_rate: FloatLike = ScheduledFloat((0.0, 0.5), (4000.0, 0.02), default=0), + self, + embed_dim: int, + pos_dim: int, + num_heads: int, + query_head_dim: int, + pos_head_dim: int, + value_head_dim: int, + feedforward_dim: int, + dropout: FloatLike = 0.1, + cnn_module_kernel: int = 31, + causal: bool = False, + attention_skip_rate: FloatLike = ScheduledFloat( + (0.0, 0.2), (4000.0, 0.05), (16000, 0.0), default=0 + ), + conv_skip_rate: FloatLike = ScheduledFloat( + (0.0, 0.2), (4000.0, 0.05), (16000, 0.0), default=0 + ), + const_attention_rate: FloatLike = ScheduledFloat( + (0.0, 0.25), (4000.0, 0.025), default=0 + ), + ff2_skip_rate: FloatLike = ScheduledFloat( + (0.0, 0.1), (4000.0, 0.01), (50000.0, 0.0) + ), + ff3_skip_rate: FloatLike = ScheduledFloat( + (0.0, 0.1), (4000.0, 0.01), (50000.0, 0.0) + ), + bypass_skip_rate: FloatLike = ScheduledFloat( + (0.0, 0.5), (4000.0, 0.02), default=0 + ), ) -> None: super(Zipformer2EncoderLayer, self).__init__() self.embed_dim = embed_dim # self.bypass implements layer skipping as well as bypass; see its default values. - self.bypass = BypassModule(embed_dim, skip_rate=bypass_skip_rate, - straight_through_rate=0) + self.bypass = BypassModule( + embed_dim, skip_rate=bypass_skip_rate, straight_through_rate=0 + ) # bypass_mid is bypass used in the middle of the layer. self.bypass_mid = BypassModule(embed_dim, straight_through_rate=0) @@ -567,39 +609,39 @@ class Zipformer2EncoderLayer(nn.Module): self.const_attention_rate = copy.deepcopy(const_attention_rate) self.self_attn_weights = RelPositionMultiheadAttentionWeights( - embed_dim, pos_dim=pos_dim, num_heads=num_heads, - query_head_dim=query_head_dim, pos_head_dim=pos_head_dim, + embed_dim, + pos_dim=pos_dim, + num_heads=num_heads, + query_head_dim=query_head_dim, + pos_head_dim=pos_head_dim, dropout=0.0, ) - self.self_attn1 = SelfAttention(embed_dim, num_heads, - value_head_dim) + self.self_attn1 = SelfAttention(embed_dim, num_heads, value_head_dim) - self.self_attn2 = SelfAttention(embed_dim, num_heads, - value_head_dim) + self.self_attn2 = SelfAttention(embed_dim, num_heads, value_head_dim) - self.feed_forward1 = FeedforwardModule(embed_dim, - (feedforward_dim * 3) // 4, - dropout) + self.feed_forward1 = FeedforwardModule( + embed_dim, (feedforward_dim * 3) // 4, dropout + ) - self.feed_forward2 = FeedforwardModule(embed_dim, - feedforward_dim, - dropout) + self.feed_forward2 = FeedforwardModule(embed_dim, feedforward_dim, dropout) - self.feed_forward3 = FeedforwardModule(embed_dim, - (feedforward_dim * 5) // 4, - dropout) + self.feed_forward3 = FeedforwardModule( + embed_dim, (feedforward_dim * 5) // 4, dropout + ) - self.nonlin_attention = NonlinAttention(embed_dim, - hidden_channels=3 * embed_dim // 4) + self.nonlin_attention = NonlinAttention( + embed_dim, hidden_channels=3 * embed_dim // 4 + ) - self.conv_module1 = ConvolutionModule(embed_dim, - cnn_module_kernel, - causal=causal) + self.conv_module1 = ConvolutionModule( + embed_dim, cnn_module_kernel, causal=causal + ) - self.conv_module2 = ConvolutionModule(embed_dim, - cnn_module_kernel, - causal=causal) + self.conv_module2 = ConvolutionModule( + embed_dim, cnn_module_kernel, causal=causal + ) # TODO: remove it self.bypass_scale = nn.Parameter(torch.full((embed_dim,), 0.5)) @@ -607,15 +649,20 @@ class Zipformer2EncoderLayer(nn.Module): self.norm = BiasNorm(embed_dim) self.balancer1 = Balancer( - embed_dim, channel_dim=-1, - min_positive=0.45, max_positive=0.55, - min_abs=0.2, max_abs=4.0, + embed_dim, + channel_dim=-1, + min_positive=0.45, + max_positive=0.55, + min_abs=0.2, + max_abs=4.0, ) # balancer for output of NonlinAttentionModule self.balancer_na = Balancer( - embed_dim, channel_dim=-1, - min_positive=0.3, max_positive=0.7, + embed_dim, + channel_dim=-1, + min_positive=0.3, + max_positive=0.7, min_abs=ScheduledFloat((0.0, 0.004), (4000.0, 0.02)), prob=0.05, # out of concern for memory usage ) @@ -624,34 +671,50 @@ class Zipformer2EncoderLayer(nn.Module): # small. give this a very small probability, even at the start of # training, it's to fix a rare problem and it's OK to fix it slowly. self.balancer_ff2 = Balancer( - embed_dim, channel_dim=-1, - min_positive=0.3, max_positive=0.7, + embed_dim, + channel_dim=-1, + min_positive=0.3, + max_positive=0.7, min_abs=ScheduledFloat((0.0, 0.0), (4000.0, 0.1), default=0.0), max_abs=2.0, prob=0.05, ) self.balancer_ff3 = Balancer( - embed_dim, channel_dim=-1, - min_positive=0.3, max_positive=0.7, + embed_dim, + channel_dim=-1, + min_positive=0.3, + max_positive=0.7, min_abs=ScheduledFloat((0.0, 0.0), (4000.0, 0.2), default=0.0), max_abs=4.0, prob=0.05, ) - self.whiten = Whiten(num_groups=1, - whitening_limit=_whitening_schedule(4.0, ratio=3.0), - prob=(0.025, 0.25), - grad_scale=0.01) - - self.balancer2 = Balancer( - embed_dim, channel_dim=-1, - min_positive=0.45, max_positive=0.55, - min_abs=0.1, max_abs=4.0, + self.whiten = Whiten( + num_groups=1, + whitening_limit=_whitening_schedule(4.0, ratio=3.0), + prob=(0.025, 0.25), + grad_scale=0.01, ) - def get_sequence_dropout_mask(self, x: Tensor, dropout_rate: float) -> Optional[Tensor]: - if dropout_rate == 0.0 or not self.training or torch.jit.is_scripting() or torch.jit.is_tracing(): + self.balancer2 = Balancer( + embed_dim, + channel_dim=-1, + min_positive=0.45, + max_positive=0.55, + min_abs=0.1, + max_abs=4.0, + ) + + def get_sequence_dropout_mask( + self, x: Tensor, dropout_rate: float + ) -> Optional[Tensor]: + if ( + dropout_rate == 0.0 + or not self.training + or torch.jit.is_scripting() + or torch.jit.is_tracing() + ): return None batch_size = x.shape[1] mask = (torch.rand(batch_size, 1, device=x.device) > dropout_rate).to(x.dtype) @@ -677,21 +740,21 @@ class Zipformer2EncoderLayer(nn.Module): src_key_padding_mask: Optional[Tensor] = None, ) -> Tensor: """ - Pass the input through the encoder layer. - Args: - src: the sequence to the encoder (required): shape (seq_len, batch_size, embedding_dim). - pos_emb: (1, 2*seq_len-1, pos_emb_dim) or (batch_size, 2*seq_len-1, pos_emb_dim) - chunk_size: the number of frames per chunk, of >= 0; if -1, no chunking. - feature_mask: something that broadcasts with src, that we'll multiply `src` - by at every layer: if a Tensor, likely of shape (seq_len, batch_size, embedding_dim) - attn_mask: the attention mask, of shape (batch_size, seq_len, seq_len) or (seq_len, seq_len), - interpreted as (batch_size, tgt_seq_len, src_seq_len) or (tgt_seq_len, src_seq_len). - True means masked position. May be None. - src_key_padding_mask: the mask for padding, of shape (batch_size, seq_len); True means - masked position. May be None. + Pass the input through the encoder layer. + Args: + src: the sequence to the encoder (required): shape (seq_len, batch_size, embedding_dim). + pos_emb: (1, 2*seq_len-1, pos_emb_dim) or (batch_size, 2*seq_len-1, pos_emb_dim) + chunk_size: the number of frames per chunk, of >= 0; if -1, no chunking. + feature_mask: something that broadcasts with src, that we'll multiply `src` + by at every layer: if a Tensor, likely of shape (seq_len, batch_size, embedding_dim) + attn_mask: the attention mask, of shape (batch_size, seq_len, seq_len) or (seq_len, seq_len), + interpreted as (batch_size, tgt_seq_len, src_seq_len) or (tgt_seq_len, src_seq_len). + True means masked position. May be None. + src_key_padding_mask: the mask for padding, of shape (batch_size, seq_len); True means + masked position. May be None. - Returns: - A tensor which has the same shape as src + Returns: + A tensor which has the same shape as src """ src_orig = src @@ -699,7 +762,9 @@ class Zipformer2EncoderLayer(nn.Module): if torch.jit.is_scripting() or torch.jit.is_tracing(): attention_skip_rate = 0.0 else: - attention_skip_rate = float(self.attention_skip_rate) if self.training else 0.0 + attention_skip_rate = ( + float(self.attention_skip_rate) if self.training else 0.0 + ) # attn_weights: (num_heads, batch_size, seq_len, seq_len) attn_weights = self.self_attn_weights( @@ -711,7 +776,9 @@ class Zipformer2EncoderLayer(nn.Module): src = src + self.feed_forward1(src) - self_attn_dropout_mask = self.get_sequence_dropout_mask(src, attention_skip_rate) + self_attn_dropout_mask = self.get_sequence_dropout_mask( + src, attention_skip_rate + ) selected_attn_weights = attn_weights[0:1] if torch.jit.is_scripting() or torch.jit.is_tracing(): @@ -722,53 +789,75 @@ class Zipformer2EncoderLayer(nn.Module): # averaging-over-time operation. # only need the mask, can just use the 1st one and expand later selected_attn_weights = selected_attn_weights[0:1] - selected_attn_weights = (selected_attn_weights > 0.0).to(selected_attn_weights.dtype) - selected_attn_weights = selected_attn_weights * (1.0 / selected_attn_weights.sum(dim=-1, keepdim=True)) + selected_attn_weights = (selected_attn_weights > 0.0).to( + selected_attn_weights.dtype + ) + selected_attn_weights = selected_attn_weights * ( + 1.0 / selected_attn_weights.sum(dim=-1, keepdim=True) + ) na = self.balancer_na(self.nonlin_attention(src, selected_attn_weights)) - src = src + (na if self_attn_dropout_mask is None else na * self_attn_dropout_mask) + src = src + ( + na if self_attn_dropout_mask is None else na * self_attn_dropout_mask + ) self_attn = self.self_attn1(src, attn_weights) - src = src + (self_attn if self_attn_dropout_mask is None else self_attn * self_attn_dropout_mask) + src = src + ( + self_attn + if self_attn_dropout_mask is None + else self_attn * self_attn_dropout_mask + ) if torch.jit.is_scripting() or torch.jit.is_tracing(): conv_skip_rate = 0.0 else: conv_skip_rate = float(self.conv_skip_rate) if self.training else 0.0 - src = src + self.sequence_dropout(self.conv_module1(src, chunk_size=chunk_size, - src_key_padding_mask=src_key_padding_mask), - conv_skip_rate) + src = src + self.sequence_dropout( + self.conv_module1( + src, chunk_size=chunk_size, src_key_padding_mask=src_key_padding_mask + ), + conv_skip_rate, + ) if torch.jit.is_scripting() or torch.jit.is_tracing(): ff2_skip_rate = 0.0 else: ff2_skip_rate = float(self.ff2_skip_rate) if self.training else 0.0 - src = src + self.sequence_dropout(self.balancer_ff2(self.feed_forward2(src)), - ff2_skip_rate) + src = src + self.sequence_dropout( + self.balancer_ff2(self.feed_forward2(src)), ff2_skip_rate + ) # bypass in the middle of the layer. src = self.bypass_mid(src_orig, src) self_attn = self.self_attn2(src, attn_weights) - src = src + (self_attn if self_attn_dropout_mask is None else self_attn * self_attn_dropout_mask) + src = src + ( + self_attn + if self_attn_dropout_mask is None + else self_attn * self_attn_dropout_mask + ) if torch.jit.is_scripting() or torch.jit.is_tracing(): conv_skip_rate = 0.0 else: conv_skip_rate = float(self.conv_skip_rate) if self.training else 0.0 - src = src + self.sequence_dropout(self.conv_module2(src, chunk_size=chunk_size, - src_key_padding_mask=src_key_padding_mask), - conv_skip_rate) + src = src + self.sequence_dropout( + self.conv_module2( + src, chunk_size=chunk_size, src_key_padding_mask=src_key_padding_mask + ), + conv_skip_rate, + ) if torch.jit.is_scripting() or torch.jit.is_tracing(): ff3_skip_rate = 0.0 else: ff3_skip_rate = float(self.ff3_skip_rate) if self.training else 0.0 - src = src + self.sequence_dropout(self.balancer_ff3(self.feed_forward3(src)), - ff3_skip_rate) + src = src + self.sequence_dropout( + self.balancer_ff3(self.feed_forward3(src)), ff3_skip_rate + ) src = self.balancer1(src) src = self.norm(src) @@ -912,20 +1001,22 @@ class Zipformer2Encoder(nn.Module): >>> src = torch.rand(10, 32, 512) >>> out = zipformer_encoder(src) """ + def __init__( - self, - encoder_layer: nn.Module, - num_layers: int, - pos_dim: int, - dropout: float, - warmup_begin: float, - warmup_end: float, - initial_layerdrop_rate: float = 0.5, - final_layerdrop_rate: float = 0.05, + self, + encoder_layer: nn.Module, + num_layers: int, + pos_dim: int, + dropout: float, + warmup_begin: float, + warmup_end: float, + initial_layerdrop_rate: float = 0.5, + final_layerdrop_rate: float = 0.05, ) -> None: super().__init__() - self.encoder_pos = CompactRelPositionalEncoding(pos_dim, dropout_rate=0.15, - length_factor=1.0) + self.encoder_pos = CompactRelPositionalEncoding( + pos_dim, dropout_rate=0.15, length_factor=1.0 + ) self.layers = nn.ModuleList( [copy.deepcopy(encoder_layer) for i in range(num_layers)] @@ -934,13 +1025,15 @@ class Zipformer2Encoder(nn.Module): assert 0 <= warmup_begin <= warmup_end - delta = (1. / num_layers) * (warmup_end - warmup_begin) + delta = (1.0 / num_layers) * (warmup_end - warmup_begin) cur_begin = warmup_begin # interpreted as a training batch index for i in range(num_layers): cur_end = cur_begin + delta - self.layers[i].bypass.skip_rate = ScheduledFloat((cur_begin, initial_layerdrop_rate), - (cur_end, final_layerdrop_rate), - default=0.0) + self.layers[i].bypass.skip_rate = ScheduledFloat( + (cur_begin, initial_layerdrop_rate), + (cur_end, final_layerdrop_rate), + default=0.0, + ) cur_begin = cur_end def forward( @@ -1014,8 +1107,13 @@ class Zipformer2Encoder(nn.Module): new_states = [] for i, mod in enumerate(self.layers): ( - cached_key, cached_nonlin_attn, cached_val1, cached_val2, cached_conv1, cached_conv2 - ) = states[i * 6: (i + 1) * 6] + cached_key, + cached_nonlin_attn, + cached_val1, + cached_val2, + cached_conv1, + cached_conv2, + ) = states[i * 6 : (i + 1) * 6] ( output, new_cached_key, @@ -1023,7 +1121,7 @@ class Zipformer2Encoder(nn.Module): new_cached_val1, new_cached_val2, new_cached_conv1, - new_cached_conv2 + new_cached_conv2, ) = mod.streaming_forward( output, pos_emb, @@ -1055,13 +1153,15 @@ class BypassModule(nn.Module): "straight-through", i.e. to not do the bypass operation much initially, in order to force all the modules to learn something. """ + def __init__( - self, - embed_dim: int, - skip_rate: FloatLike = 0.0, - straight_through_rate: FloatLike = 0.0, - scale_min: FloatLike = ScheduledFloat((0.0, 0.9), (20000.0, 0.2), default=0), - scale_max: FloatLike = 1.0): + self, + embed_dim: int, + skip_rate: FloatLike = 0.0, + straight_through_rate: FloatLike = 0.0, + scale_min: FloatLike = ScheduledFloat((0.0, 0.9), (20000.0, 0.2), default=0), + scale_max: FloatLike = 1.0, + ): super().__init__() self.bypass_scale = nn.Parameter(torch.full((embed_dim,), 0.5)) self.skip_rate = copy.deepcopy(skip_rate) @@ -1077,9 +1177,9 @@ class BypassModule(nn.Module): if torch.jit.is_scripting() or torch.jit.is_tracing() or not self.training: return self.bypass_scale else: - ans = limit_param_value(self.bypass_scale, - min=float(self.scale_min), - max=float(self.scale_max)) + ans = limit_param_value( + self.bypass_scale, min=float(self.scale_min), max=float(self.scale_max) + ) skip_rate = float(self.skip_rate) if skip_rate != 0.0: mask = torch.rand((batch_size, 1), device=ans.device) > skip_rate @@ -1088,13 +1188,14 @@ class BypassModule(nn.Module): # on which we have randomly chosen to do layer-skipping. straight_through_rate = float(self.straight_through_rate) if straight_through_rate != 0.0: - mask = torch.rand((batch_size, 1), device=ans.device) < straight_through_rate + mask = ( + torch.rand((batch_size, 1), device=ans.device) + < straight_through_rate + ) ans = torch.maximum(ans, mask.to(ans.dtype)) return ans - def forward(self, - src_orig: Tensor, - src: Tensor): + def forward(self, src_orig: Tensor, src: Tensor): """ Args: src_orig and src are both of shape (seq_len, batch_size, num_channels) Returns: something with the same shape as src and src_orig @@ -1109,15 +1210,13 @@ class DownsampledZipformer2Encoder(nn.Module): after convolutional downsampling, and then upsampled again at the output, and combined with the origin input, so that the output has the same shape as the input. """ - def __init__(self, - encoder: nn.Module, - dim: int, - downsample: int, - dropout: FloatLike): + + def __init__( + self, encoder: nn.Module, dim: int, downsample: int, dropout: FloatLike + ): super(DownsampledZipformer2Encoder, self).__init__() self.downsample_factor = downsample - self.downsample = SimpleDownsample(dim, - downsample, dropout) + self.downsample = SimpleDownsample(dim, downsample, dropout) self.num_layers = encoder.num_layers self.encoder = encoder self.upsample = SimpleUpsample(dim, downsample) @@ -1149,7 +1248,7 @@ class DownsampledZipformer2Encoder(nn.Module): src = self.downsample(src) ds = self.downsample_factor if attn_mask is not None: - attn_mask = attn_mask[::ds,::ds] + attn_mask = attn_mask[::ds, ::ds] src = self.encoder( src, @@ -1160,7 +1259,7 @@ class DownsampledZipformer2Encoder(nn.Module): ) src = self.upsample(src) # remove any extra frames that are not a multiple of downsample_factor - src = src[:src_orig.shape[0]] + src = src[: src_orig.shape[0]] return self.out_combiner(src_orig, src) @@ -1196,7 +1295,7 @@ class DownsampledZipformer2Encoder(nn.Module): ) src = self.upsample(src) # remove any extra frames that are not a multiple of downsample_factor - src = src[:src_orig.shape[0]] + src = src[: src_orig.shape[0]] return self.out_combiner(src_orig, src), new_states @@ -1205,10 +1304,8 @@ class SimpleDownsample(torch.nn.Module): """ Does downsampling with attention, by weighted sum, and a projection.. """ - def __init__(self, - channels: int, - downsample: int, - dropout: FloatLike): + + def __init__(self, channels: int, downsample: int, dropout: FloatLike): super(SimpleDownsample, self).__init__() self.bias = nn.Parameter(torch.zeros(downsample)) @@ -1218,8 +1315,7 @@ class SimpleDownsample(torch.nn.Module): self.downsample = downsample - def forward(self, - src: Tensor) -> Tensor: + def forward(self, src: Tensor) -> Tensor: """ x: (seq_len, batch_size, in_channels) Returns a tensor of shape @@ -1232,7 +1328,7 @@ class SimpleDownsample(torch.nn.Module): # Pad to an exact multiple of self.downsample # right-pad src, repeating the last element. pad = d_seq_len * ds - seq_len - src_extra = src[src.shape[0]-1:].expand(pad, src.shape[1], src.shape[2]) + src_extra = src[src.shape[0] - 1 :].expand(pad, src.shape[1], src.shape[2]) src = torch.cat((src, src_extra), dim=0) assert src.shape[0] == d_seq_len * ds @@ -1253,14 +1349,12 @@ class SimpleUpsample(torch.nn.Module): A very simple form of upsampling that mostly just repeats the input, but also adds a position-specific bias. """ - def __init__(self, - num_channels: int, - upsample: int): + + def __init__(self, num_channels: int, upsample: int): super(SimpleUpsample, self).__init__() self.upsample = upsample - def forward(self, - src: Tensor) -> Tensor: + def forward(self, src: Tensor) -> Tensor: """ x: (seq_len, batch_size, num_channels) Returns a tensor of shape @@ -1298,11 +1392,13 @@ class CompactRelPositionalEncoding(torch.nn.Module): length_factor: a heuristic scale (should be >= 1.0) which, if larger, gives less weight to small differences of offset near the origin. """ + def __init__( - self, embed_dim: int, - dropout_rate: FloatLike, - max_len: int = 1000, - length_factor: float = 1.0, + self, + embed_dim: int, + dropout_rate: FloatLike, + max_len: int = 1000, + length_factor: float = 1.0, ) -> None: """Construct a CompactRelPositionalEncoding object.""" super(CompactRelPositionalEncoding, self).__init__() @@ -1326,19 +1422,22 @@ class CompactRelPositionalEncoding(torch.nn.Module): return # if T == 4, x would contain [ -3, -2, 1, 0, 1, 2, 3 ] - x = torch.arange(-(T-1), T, - device=x.device).to(torch.float32).unsqueeze(1) + x = torch.arange(-(T - 1), T, device=x.device).to(torch.float32).unsqueeze(1) freqs = 1 + torch.arange(self.embed_dim // 2, device=x.device) # `compression_length` this is arbitrary/heuristic, if it is larger we have more resolution # for small time offsets but less resolution for large time offsets. - compression_length = (self.embed_dim ** 0.5) + compression_length = self.embed_dim**0.5 # x_compressed, like X, goes from -infinity to infinity as T goes from -infinity to infinity; # but it does so more slowly than T for large absolute values of T. # The formula is chosen so that d(x_compressed )/dx is 1 around x == 0, which # is important. - x_compressed = compression_length * x.sign() * ((x.abs() + compression_length).log() - math.log(compression_length)) + x_compressed = ( + compression_length + * x.sign() + * ((x.abs() + compression_length).log() - math.log(compression_length)) + ) # if self.length_factor == 1.0, then length_scale is chosen so that the # FFT can exactly separate points close to the origin (T == 0). So this @@ -1380,7 +1479,7 @@ class CompactRelPositionalEncoding(torch.nn.Module): - x_size_left + 1 : self.pe.size(0) // 2 # noqa E203 + x.size(0), - : + :, ] pos_emb = pos_emb.unsqueeze(0) return self.dropout(pos_emb) @@ -1407,15 +1506,14 @@ class RelPositionMultiheadAttentionWeights(nn.Module): """ def __init__( - self, - embed_dim: int, - pos_dim: int, - num_heads: int, - query_head_dim: int, - pos_head_dim: int, - dropout: float = 0.0, - pos_emb_skip_rate: FloatLike = ScheduledFloat((0.0, 0.5), - (4000.0, 0.0)) + self, + embed_dim: int, + pos_dim: int, + num_heads: int, + query_head_dim: int, + pos_head_dim: int, + dropout: float = 0.0, + pos_emb_skip_rate: FloatLike = ScheduledFloat((0.0, 0.5), (4000.0, 0.0)), ) -> None: super().__init__() self.embed_dim = embed_dim @@ -1434,13 +1532,16 @@ class RelPositionMultiheadAttentionWeights(nn.Module): # dividing it between the query and key. Note: this module is intended # to be used with the ScaledAdam optimizer; with most other optimizers, # it would be necessary to apply the scaling factor in the forward function. - self.in_proj = ScaledLinear(embed_dim, in_proj_dim, bias=True, - initial_scale=query_head_dim**-0.25) + self.in_proj = ScaledLinear( + embed_dim, in_proj_dim, bias=True, initial_scale=query_head_dim**-0.25 + ) - self.whiten_keys = Whiten(num_groups=num_heads, - whitening_limit=_whitening_schedule(3.0), - prob=(0.025, 0.25), - grad_scale=0.025) + self.whiten_keys = Whiten( + num_groups=num_heads, + whitening_limit=_whitening_schedule(3.0), + prob=(0.025, 0.25), + grad_scale=0.025, + ) # add a balancer for the keys that runs with very small probability, and # tries to enforce that all dimensions have mean around zero. The @@ -1450,19 +1551,20 @@ class RelPositionMultiheadAttentionWeights(nn.Module): # bias because the small numerical roundoff tends to have a non-random # sign. This module is intended to prevent that. Use a very small # probability; that should be suffixient to fix the problem. - self.balance_keys = Balancer(key_head_dim * num_heads, - channel_dim=-1, - min_positive=0.4, - max_positive=0.6, - min_abs=0.0, - max_abs=100.0, - prob=0.025) + self.balance_keys = Balancer( + key_head_dim * num_heads, + channel_dim=-1, + min_positive=0.4, + max_positive=0.6, + min_abs=0.0, + max_abs=100.0, + prob=0.025, + ) # linear transformation for positional encoding. - self.linear_pos = ScaledLinear(pos_dim, - num_heads * pos_head_dim, - bias=False, - initial_scale=0.05) + self.linear_pos = ScaledLinear( + pos_dim, num_heads * pos_head_dim, bias=False, initial_scale=0.05 + ) # the following are for diagnosics only, see --print-diagnostics option self.copy_pos_query = Identity() @@ -1498,10 +1600,10 @@ class RelPositionMultiheadAttentionWeights(nn.Module): query_dim = query_head_dim * num_heads # self-attention - q = x[...,0:query_dim] - k = x[...,query_dim:2*query_dim] + q = x[..., 0:query_dim] + k = x[..., query_dim : 2 * query_dim] # p is the position-encoding query - p = x[...,2*query_dim:] + p = x[..., 2 * query_dim :] assert p.shape[-1] == num_heads * pos_head_dim q = self.copy_query(q) # for diagnostics only, does nothing. @@ -1529,7 +1631,9 @@ class RelPositionMultiheadAttentionWeights(nn.Module): if use_pos_scores: pos_emb = self.linear_pos(pos_emb) seq_len2 = 2 * seq_len - 1 - pos_emb = pos_emb.reshape(-1, seq_len2, num_heads, pos_head_dim).permute(2, 0, 3, 1) + pos_emb = pos_emb.reshape(-1, seq_len2, num_heads, pos_head_dim).permute( + 2, 0, 3, 1 + ) # pos shape now: (head, {1 or batch_size}, pos_dim, seq_len2) # (head, batch, time1, pos_dim) x (head, 1, pos_dim, seq_len2) -> (head, batch, time1, seq_len2) @@ -1548,12 +1652,16 @@ class RelPositionMultiheadAttentionWeights(nn.Module): pos_scores = torch.gather(pos_scores, dim=1, index=indexes) pos_scores = pos_scores.reshape(num_heads, batch_size, time1, seq_len) else: - pos_scores = pos_scores.as_strided((num_heads, batch_size, seq_len, seq_len), - (pos_scores.stride(0), - pos_scores.stride(1), - pos_scores.stride(2)-pos_scores.stride(3), - pos_scores.stride(3)), - storage_offset=pos_scores.stride(3) * (seq_len - 1)) + pos_scores = pos_scores.as_strided( + (num_heads, batch_size, seq_len, seq_len), + ( + pos_scores.stride(0), + pos_scores.stride(1), + pos_scores.stride(2) - pos_scores.stride(3), + pos_scores.stride(3), + ), + storage_offset=pos_scores.stride(3) * (seq_len - 1), + ) attn_scores = attn_scores + pos_scores @@ -1572,10 +1680,9 @@ class RelPositionMultiheadAttentionWeights(nn.Module): # but we view this as a failsafe to avoid "implausible" parameter # values rather than a regularization method that should be active # under normal circumstances. - attn_scores = penalize_abs_values_gt(attn_scores, - limit=25.0, - penalty=1.0e-04, - name=self.name) + attn_scores = penalize_abs_values_gt( + attn_scores, limit=25.0, penalty=1.0e-04, name=self.name + ) assert attn_scores.shape == (num_heads, batch_size, seq_len, seq_len) @@ -1588,7 +1695,10 @@ class RelPositionMultiheadAttentionWeights(nn.Module): attn_scores = attn_scores.masked_fill(attn_mask, -1000) if key_padding_mask is not None: - assert key_padding_mask.shape == (batch_size, seq_len), key_padding_mask.shape + assert key_padding_mask.shape == ( + batch_size, + seq_len, + ), key_padding_mask.shape attn_scores = attn_scores.masked_fill( key_padding_mask.unsqueeze(1), -1000, @@ -1644,14 +1754,17 @@ class RelPositionMultiheadAttentionWeights(nn.Module): query_dim = query_head_dim * num_heads # self-attention - q = x[...,0:query_dim] - k = x[...,query_dim:2*query_dim] + q = x[..., 0:query_dim] + k = x[..., query_dim : 2 * query_dim] # p is the position-encoding query - p = x[...,2*query_dim:] + p = x[..., 2 * query_dim :] assert p.shape[-1] == num_heads * pos_head_dim # Pad cached left contexts - assert cached_key.shape[0] == left_context_len, (cached_key.shape[0], left_context_len) + assert cached_key.shape[0] == left_context_len, ( + cached_key.shape[0], + left_context_len, + ) k = torch.cat([cached_key, k], dim=0) # Update cached left contexts cached_key = k[-left_context_len:, ...] @@ -1672,13 +1785,15 @@ class RelPositionMultiheadAttentionWeights(nn.Module): pos_emb = self.linear_pos(pos_emb) seq_len2 = 2 * seq_len - 1 + left_context_len - pos_emb = pos_emb.reshape(-1, seq_len2, num_heads, pos_head_dim).permute(2, 0, 3, 1) + pos_emb = pos_emb.reshape(-1, seq_len2, num_heads, pos_head_dim).permute( + 2, 0, 3, 1 + ) # pos shape now: (head, {1 or batch_size}, pos_dim, seq_len2) # (head, batch, time1, pos_dim) x (head, 1, pos_dim, seq_len2) -> (head, batch, time1, seq_len2) # [where seq_len2 represents relative position.] pos_scores = torch.matmul(p, pos_emb) - + if torch.jit.is_tracing(): (num_heads, batch_size, time1, n) = pos_scores.shape rows = torch.arange(start=time1 - 1, end=-1, step=-1) @@ -1692,16 +1807,25 @@ class RelPositionMultiheadAttentionWeights(nn.Module): # to absolute position. I don't know whether I might have got the time-offsets backwards or # not, but let this code define which way round it is supposed to be. else: - pos_scores = pos_scores.as_strided((num_heads, batch_size, seq_len, k_len), - (pos_scores.stride(0), - pos_scores.stride(1), - pos_scores.stride(2)-pos_scores.stride(3), - pos_scores.stride(3)), - storage_offset=pos_scores.stride(3) * (seq_len - 1)) + pos_scores = pos_scores.as_strided( + (num_heads, batch_size, seq_len, k_len), + ( + pos_scores.stride(0), + pos_scores.stride(1), + pos_scores.stride(2) - pos_scores.stride(3), + pos_scores.stride(3), + ), + storage_offset=pos_scores.stride(3) * (seq_len - 1), + ) attn_scores = attn_scores + pos_scores - assert attn_scores.shape == (num_heads, batch_size, seq_len, k_len), attn_scores.shape + assert attn_scores.shape == ( + num_heads, + batch_size, + seq_len, + k_len, + ), attn_scores.shape if key_padding_mask is not None: assert key_padding_mask.shape == (batch_size, k_len), key_padding_mask.shape @@ -1714,18 +1838,21 @@ class RelPositionMultiheadAttentionWeights(nn.Module): return attn_weights, cached_key - def _print_attn_entropy( - self, - attn_weights: Tensor): + def _print_attn_entropy(self, attn_weights: Tensor): # attn_weights: (num_heads, batch_size, seq_len, seq_len) (num_heads, batch_size, seq_len, seq_len) = attn_weights.shape with torch.no_grad(): with torch.cuda.amp.autocast(enabled=False): attn_weights = attn_weights.to(torch.float32) - attn_weights_entropy = -((attn_weights + 1.0e-20).log() * attn_weights).sum( - dim=-1).mean(dim=(1,2)) - logging.info(f"name={self.name}, attn_weights_entropy = {attn_weights_entropy}") + attn_weights_entropy = ( + -((attn_weights + 1.0e-20).log() * attn_weights) + .sum(dim=-1) + .mean(dim=(1, 2)) + ) + logging.info( + f"name={self.name}, attn_weights_entropy = {attn_weights_entropy}" + ) class SelfAttention(nn.Module): @@ -1738,25 +1865,26 @@ class SelfAttention(nn.Module): num_heads: the number of attention heads value_head_dim: the value dimension per head """ + def __init__( - self, - embed_dim: int, - num_heads: int, - value_head_dim: int, + self, + embed_dim: int, + num_heads: int, + value_head_dim: int, ) -> None: super().__init__() - self.in_proj = nn.Linear(embed_dim, - num_heads * value_head_dim, - bias=True) + self.in_proj = nn.Linear(embed_dim, num_heads * value_head_dim, bias=True) - self.out_proj = ScaledLinear(num_heads * value_head_dim, - embed_dim, bias=True, - initial_scale=0.05) + self.out_proj = ScaledLinear( + num_heads * value_head_dim, embed_dim, bias=True, initial_scale=0.05 + ) - self.whiten = Whiten(num_groups=1, - whitening_limit=_whitening_schedule(7.5, ratio=3.0), - prob=(0.025, 0.25), - grad_scale=0.01) + self.whiten = Whiten( + num_groups=1, + whitening_limit=_whitening_schedule(7.5, ratio=3.0), + prob=(0.025, 0.25), + grad_scale=0.01, + ) def forward( self, @@ -1785,8 +1913,11 @@ class SelfAttention(nn.Module): x = torch.matmul(attn_weights, x) # v: (num_heads, batch_size, seq_len, value_head_dim) - x = x.permute(2, 1, 0, 3).contiguous().view( - seq_len, batch_size, num_heads * value_head_dim) + x = ( + x.permute(2, 1, 0, 3) + .contiguous() + .view(seq_len, batch_size, num_heads * value_head_dim) + ) # returned value is of shape (seq_len, batch_size, embed_dim), like the input. x = self.out_proj(x) @@ -1823,7 +1954,10 @@ class SelfAttention(nn.Module): x = self.in_proj(x) # (seq_len, batch_size, num_heads * value_head_dim) # Pad cached left contexts - assert cached_val.shape[0] == left_context_len, (cached_val.shape[0], left_context_len) + assert cached_val.shape[0] == left_context_len, ( + cached_val.shape[0], + left_context_len, + ) x = torch.cat([cached_val, x], dim=0) # Update cached left contexts cached_val = x[-left_context_len:, ...] @@ -1836,8 +1970,11 @@ class SelfAttention(nn.Module): x = torch.matmul(attn_weights, x) # v: (num_heads, batch_size, seq_len, value_head_dim) - x = x.permute(2, 1, 0, 3).contiguous().view( - seq_len, batch_size, num_heads * value_head_dim) + x = ( + x.permute(2, 1, 0, 3) + .contiguous() + .view(seq_len, batch_size, num_heads * value_head_dim) + ) # returned value is of shape (seq_len, batch_size, embed_dim), like the input. x = self.out_proj(x) @@ -1846,33 +1983,38 @@ class SelfAttention(nn.Module): class FeedforwardModule(nn.Module): - """Feedforward module in Zipformer2 model. - """ - def __init__(self, - embed_dim: int, - feedforward_dim: int, - dropout: FloatLike): + """Feedforward module in Zipformer2 model.""" + + def __init__(self, embed_dim: int, feedforward_dim: int, dropout: FloatLike): super(FeedforwardModule, self).__init__() self.in_proj = nn.Linear(embed_dim, feedforward_dim) - self.hidden_balancer = Balancer(feedforward_dim, - channel_dim=-1, - min_positive=0.3, - max_positive=1.0, - min_abs=0.75, - max_abs=5.0) + self.hidden_balancer = Balancer( + feedforward_dim, + channel_dim=-1, + min_positive=0.3, + max_positive=1.0, + min_abs=0.75, + max_abs=5.0, + ) # shared_dim=0 means we share the dropout mask along the time axis - self.out_proj = ActivationDropoutAndLinear(feedforward_dim, embed_dim, - activation='SwooshL', - dropout_p=dropout, - dropout_shared_dim=0, bias=True, - initial_scale=0.1) + self.out_proj = ActivationDropoutAndLinear( + feedforward_dim, + embed_dim, + activation="SwooshL", + dropout_p=dropout, + dropout_shared_dim=0, + bias=True, + initial_scale=0.1, + ) - self.out_whiten = Whiten(num_groups=1, - whitening_limit=_whitening_schedule(7.5), - prob=(0.025, 0.25), - grad_scale=0.01) + self.out_whiten = Whiten( + num_groups=1, + whitening_limit=_whitening_schedule(7.5), + prob=(0.025, 0.25), + grad_scale=0.01, + ) def forward(self, x: Tensor): x = self.in_proj(x) @@ -1893,9 +2035,9 @@ class NonlinAttention(nn.Module): """ def __init__( - self, - channels: int, - hidden_channels: int, + self, + channels: int, + hidden_channels: int, ) -> None: super().__init__() @@ -1908,7 +2050,8 @@ class NonlinAttention(nn.Module): # starting from about 3, and poorly-trained instances of the module have smaller abs values # before the sigmoid. self.balancer = Balancer( - hidden_channels, channel_dim=-1, + hidden_channels, + channel_dim=-1, min_positive=ScheduledFloat((0.0, 0.25), (20000.0, 0.05)), max_positive=ScheduledFloat((0.0, 0.75), (20000.0, 0.95)), min_abs=0.5, @@ -1920,19 +2063,23 @@ class NonlinAttention(nn.Module): self.identity2 = Identity() # for diagnostics. self.identity3 = Identity() # for diagnostics. - self.out_proj = ScaledLinear(hidden_channels, channels, - bias=True, - initial_scale=0.05) + self.out_proj = ScaledLinear( + hidden_channels, channels, bias=True, initial_scale=0.05 + ) - self.whiten1 = Whiten(num_groups=1, - whitening_limit=_whitening_schedule(5.0), - prob=(0.025, 0.25), - grad_scale=0.01) + self.whiten1 = Whiten( + num_groups=1, + whitening_limit=_whitening_schedule(5.0), + prob=(0.025, 0.25), + grad_scale=0.01, + ) - self.whiten2 = Whiten(num_groups=1, - whitening_limit=_whitening_schedule(5.0, ratio=3.0), - prob=(0.025, 0.25), - grad_scale=0.01) + self.whiten2 = Whiten( + num_groups=1, + whitening_limit=_whitening_schedule(5.0, ratio=3.0), + prob=(0.025, 0.25), + grad_scale=0.01, + ) def forward( self, @@ -1940,11 +2087,11 @@ class NonlinAttention(nn.Module): attn_weights: Tensor, ) -> Tensor: """. - Args: - x: a Tensor of shape (seq_len, batch_size, num_channels) -attn_weights: a Tensor of shape (num_heads, batch_size, seq_len, seq_len) - Returns: - a Tensor with the same shape as x + Args: + x: a Tensor of shape (seq_len, batch_size, num_channels) + attn_weights: a Tensor of shape (num_heads, batch_size, seq_len, seq_len) + Returns: + a Tensor with the same shape as x """ x = self.in_proj(x) @@ -2014,13 +2161,21 @@ attn_weights: a Tensor of shape (num_heads, batch_size, seq_len, seq_len) (seq_len, batch_size, embed_dim) = x.shape num_heads = attn_weights.shape[0] - assert attn_weights.shape == (num_heads, batch_size, seq_len, left_context_len + seq_len) + assert attn_weights.shape == ( + num_heads, + batch_size, + seq_len, + left_context_len + seq_len, + ) x = x.reshape(seq_len, batch_size, num_heads, -1).permute(2, 1, 0, 3) # now x: (num_heads, batch_size, seq_len, head_dim) # Pad cached tensor - assert cached_x.shape[2] == left_context_len, (cached_x.shape[2], left_context_len) + assert cached_x.shape[2] == left_context_len, ( + cached_x.shape[2], + left_context_len, + ) x_pad = torch.cat([cached_x, x], dim=2) # Update cached tensor cached_x = x_pad[:, :, -left_context_len:, :] @@ -2045,8 +2200,12 @@ class ConvolutionModule(nn.Module): bias (bool): Whether to use bias in conv layers (default=True). """ + def __init__( - self, channels: int, kernel_size: int, causal: bool, + self, + channels: int, + kernel_size: int, + causal: bool, ) -> None: """Construct a ConvolutionModule object.""" super(ConvolutionModule, self).__init__() @@ -2057,7 +2216,8 @@ class ConvolutionModule(nn.Module): self.causal = causal self.in_proj = nn.Linear( - channels, 2 * bottleneck_dim, + channels, + 2 * bottleneck_dim, ) # the gradients on in_proj are a little noisy, likely to do with the # sigmoid in glu. @@ -2076,7 +2236,8 @@ class ConvolutionModule(nn.Module): # it will be in a better position to start learning something, i.e. to latch onto # the correct range. self.balancer1 = Balancer( - bottleneck_dim, channel_dim=-1, + bottleneck_dim, + channel_dim=-1, min_positive=ScheduledFloat((0.0, 0.05), (8000.0, 0.025)), max_positive=1.0, min_abs=1.5, @@ -2091,31 +2252,40 @@ class ConvolutionModule(nn.Module): assert kernel_size % 2 == 1 - self.depthwise_conv = ChunkCausalDepthwiseConv1d( - channels=bottleneck_dim, - kernel_size=kernel_size) if causal else nn.Conv1d( - in_channels=bottleneck_dim, - out_channels=bottleneck_dim, - groups=bottleneck_dim, - kernel_size=kernel_size, - padding=kernel_size // 2) + self.depthwise_conv = ( + ChunkCausalDepthwiseConv1d(channels=bottleneck_dim, kernel_size=kernel_size) + if causal + else nn.Conv1d( + in_channels=bottleneck_dim, + out_channels=bottleneck_dim, + groups=bottleneck_dim, + kernel_size=kernel_size, + padding=kernel_size // 2, + ) + ) self.balancer2 = Balancer( - bottleneck_dim, channel_dim=1, + bottleneck_dim, + channel_dim=1, min_positive=ScheduledFloat((0.0, 0.1), (8000.0, 0.05)), max_positive=1.0, min_abs=ScheduledFloat((0.0, 0.2), (20000.0, 0.5)), max_abs=10.0, ) - self.whiten = Whiten(num_groups=1, - whitening_limit=_whitening_schedule(7.5), - prob=(0.025, 0.25), - grad_scale=0.01) + self.whiten = Whiten( + num_groups=1, + whitening_limit=_whitening_schedule(7.5), + prob=(0.025, 0.25), + grad_scale=0.01, + ) self.out_proj = ActivationDropoutAndLinear( - bottleneck_dim, channels, activation='SwooshR', - dropout_p=0.0, initial_scale=0.05, + bottleneck_dim, + channels, + activation="SwooshR", + dropout_p=0.0, + initial_scale=0.05, ) def forward( @@ -2153,9 +2323,15 @@ class ConvolutionModule(nn.Module): if src_key_padding_mask is not None: x = x.masked_fill(src_key_padding_mask.unsqueeze(1).expand_as(x), 0.0) - if not torch.jit.is_scripting() and not torch.jit.is_tracing() and chunk_size >= 0: + if ( + not torch.jit.is_scripting() + and not torch.jit.is_tracing() + and chunk_size >= 0 + ): # Not support exporting a model for simulated streaming decoding - assert self.causal, "Must initialize model with causal=True if you use chunk_size" + assert ( + self.causal + ), "Must initialize model with causal=True if you use chunk_size" x = self.depthwise_conv(x, chunk_size=chunk_size) else: x = self.depthwise_conv(x) @@ -2225,10 +2401,12 @@ def _test_zipformer_main(causal: bool = False): # Just make sure the forward pass runs. c = Zipformer2( - encoder_dim=(64, 96), encoder_unmasked_dim=(48, 64), num_heads=(4, 4), + encoder_dim=(64, 96), + encoder_unmasked_dim=(48, 64), + num_heads=(4, 4), causal=causal, chunk_size=(4,) if causal else (-1,), - left_context_frames=(64,) + left_context_frames=(64,), ) batch_size = 5 seq_len = 20 diff --git a/egs/mgb2/ASR/conformer_ctc/asr_datamodule.py b/egs/mgb2/ASR/conformer_ctc/asr_datamodule.py index 8242e986d..442ff85c2 100644 --- a/egs/mgb2/ASR/conformer_ctc/asr_datamodule.py +++ b/egs/mgb2/ASR/conformer_ctc/asr_datamodule.py @@ -17,7 +17,7 @@ from lhotse.dataset import ( DynamicBucketingSampler, K2SpeechRecognitionDataset, PrecomputedFeatures, - SingleCutSampler, + SimpleCutSampler, SpecAugment, ) from lhotse.dataset.input_strategies import OnTheFlyFeatures @@ -270,8 +270,8 @@ class MGB2AsrDataModule: drop_last=self.args.drop_last, ) else: - logging.info("Using SingleCutSampler.") - train_sampler = SingleCutSampler( + logging.info("Using SimpleCutSampler.") + train_sampler = SimpleCutSampler( cuts_train, max_duration=self.args.max_duration, shuffle=self.args.shuffle, diff --git a/egs/multi_zh-hans/ASR/README.md b/egs/multi_zh-hans/ASR/README.md new file mode 100644 index 000000000..537816a5d --- /dev/null +++ b/egs/multi_zh-hans/ASR/README.md @@ -0,0 +1,39 @@ + +# Introduction + +This recipe includes scripts for training Zipformer model using multiple Chinese datasets. + +# Included Training Sets +1. THCHS-30 +2. AiShell-{1,2,4} +3. ST-CMDS +4. Primewords +5. MagicData +6. Aidatatang_200zh +7. AliMeeting +8. WeNetSpeech +9. KeSpeech-ASR + +|Datset| Number of hours| URL| +|---|---:|---| +|**TOTAL**|14,106|---| +|THCHS-30|35|https://www.openslr.org/18/| +|AiShell-1|170|https://www.openslr.org/33/| +|AiShell-2|1,000|http://www.aishelltech.com/aishell_2| +|AiShell-4|120|https://www.openslr.org/111/| +|ST-CMDS|110|https://www.openslr.org/38/| +|Primewords|99|https://www.openslr.org/47/| +|aidatatang_200zh|200|https://www.openslr.org/62/| +|MagicData|755|https://www.openslr.org/68/| +|AliMeeting|100|https://openslr.org/119/| +|WeNetSpeech|10,000|https://github.com/wenet-e2e/WenetSpeech| +|KeSpeech|1,542|https://github.com/KeSpeech/KeSpeech| + + +# Included Test Sets +1. Aishell-{1,2,4} +2. Aidatatang_200zh +3. AliMeeting +4. MagicData +5. KeSpeech-ASR +6. WeNetSpeech \ No newline at end of file diff --git a/egs/multi_zh-hans/ASR/RESULTS.md b/egs/multi_zh-hans/ASR/RESULTS.md new file mode 100644 index 000000000..31fbd9700 --- /dev/null +++ b/egs/multi_zh-hans/ASR/RESULTS.md @@ -0,0 +1,38 @@ +## Results + +### Multi Chinese datasets char-based training results (Non-streaming) on zipformer model + +This is the [pull request #1238](https://github.com/k2-fsa/icefall/pull/1238) in icefall. + +#### Non-streaming + +Best results (num of params : ~69M): + +The training command: + +``` +./zipformer/train.py \ + --world-size 4 \ + --num-epochs 20 \ + --use-fp16 1 \ + --max-duration 600 \ + --num-workers 8 +``` + +The decoding command: + +``` +./zipformer/decode.py \ + --epoch 20 \ + --avg 1 +``` + +Character Error Rates (CERs) listed below are produced by the checkpoint of the 20th epoch using greedy search and BPE model ( # tokens is 2000, byte fallback enabled). + +| Datasets | aidatatang _200zh | aidatatang _200zh | alimeeting | alimeeting | aishell-1 | aishell-1 | aishell-2 | aishell-2 | aishell-4 | magicdata | magicdata | kespeech-asr | kespeech-asr | kespeech-asr | WenetSpeech | WenetSpeech | WenetSpeech | +|--------------------------------|------------------------------|-------------|-------------------|--------------|----------------|-------------|------------------|-------------|------------------|------------------|-------------|-----------------------|-----------------------|-------------|--------------------|-------------------------|---------------------| +| Zipformer CER (%) | dev | test | eval | test | dev | test | dev | test | test | dev | test | dev phase1 | dev phase2 | test | dev | test meeting | test net | +| | 3.2 | 3.67 | 23.15 | 24.78 | 2.91 | 3.04 | 3.59 | 4.03 | 15.68 | 3.68 | 3.12 | 6.69 | 3.19 | 8.01 | 9.32 | 7.05 | 8.78 | + + +The pre-trained model is available here : https://huggingface.co/zrjin/icefall-asr-multi-zh-hans-zipformer-2023-9-2 diff --git a/egs/multi_zh-hans/ASR/local/bpe_model_to_tokens.py b/egs/multi_zh-hans/ASR/local/bpe_model_to_tokens.py new file mode 100755 index 000000000..d078e5b98 --- /dev/null +++ b/egs/multi_zh-hans/ASR/local/bpe_model_to_tokens.py @@ -0,0 +1,37 @@ +#!/usr/bin/env python3 + +""" +This script takes `bpe.model` as input and generates a file `tokens.txt` +from it. + +Usage: +./bpe_model_to_tokens.py /path/to/input/bpe.model > tokens.txt +""" +import argparse + +import sentencepiece as spm + + +def get_args(): + parser = argparse.ArgumentParser() + parser.add_argument( + "bpe_model", + type=str, + help="Path to the input bpe.model", + ) + + return parser.parse_args() + + +def main(): + args = get_args() + + sp = spm.SentencePieceProcessor() + sp.load(args.bpe_model) + + for i in range(sp.vocab_size()): + print(sp.id_to_piece(i), i) + + +if __name__ == "__main__": + main() diff --git a/egs/multi_zh-hans/ASR/local/compile_lg.py b/egs/multi_zh-hans/ASR/local/compile_lg.py new file mode 120000 index 000000000..462d6d3fb --- /dev/null +++ b/egs/multi_zh-hans/ASR/local/compile_lg.py @@ -0,0 +1 @@ +../../../librispeech/ASR/local/compile_lg.py \ No newline at end of file diff --git a/egs/multi_zh-hans/ASR/local/compute_fbank_kespeech_dev_test.py b/egs/multi_zh-hans/ASR/local/compute_fbank_kespeech_dev_test.py new file mode 100755 index 000000000..2581ee42f --- /dev/null +++ b/egs/multi_zh-hans/ASR/local/compute_fbank_kespeech_dev_test.py @@ -0,0 +1,93 @@ +#!/usr/bin/env python3 +# Copyright 2021 Johns Hopkins University (Piotr Żelasko) +# Copyright 2021 Xiaomi Corp. (Fangjun Kuang) +# Copyright 2023 Xiaomi Corp. (Zengrui Jin) +# +# See ../../../../LICENSE for clarification regarding multiple authors +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import logging +from pathlib import Path + +import torch +from lhotse import CutSet, KaldifeatFbank, KaldifeatFbankConfig, LilcomChunkyWriter + +# 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_kespeech_dev_test(): + in_out_dir = Path("data/fbank/kespeech") + # number of workers in dataloader + num_workers = 42 + + # number of seconds in a batch + batch_duration = 600 + + subsets = ( + "dev_phase1", + "dev_phase2", + "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"kespeech-asr_cuts_{partition}.jsonl.gz" + if cuts_path.is_file(): + logging.info(f"{cuts_path} exists - skipping") + continue + + raw_cuts_path = in_out_dir / f"kespeech-asr_cuts_{partition}_raw.jsonl.gz" + + logging.info(f"Loading {raw_cuts_path}") + cut_set = CutSet.from_file(raw_cuts_path) + + logging.info("Splitting cuts into smaller chunks") + cut_set = cut_set.trim_to_supervisions( + keep_overlapping=False, min_duration=None + ) + + 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, + storage_type=LilcomChunkyWriter, + overwrite=True, + ) + + logging.info(f"Saving to {cuts_path}") + cut_set.to_file(cuts_path) + + +def main(): + formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s" + logging.basicConfig(format=formatter, level=logging.INFO) + + compute_fbank_kespeech_dev_test() + + +if __name__ == "__main__": + main() diff --git a/egs/multi_zh-hans/ASR/local/compute_fbank_kespeech_splits.py b/egs/multi_zh-hans/ASR/local/compute_fbank_kespeech_splits.py new file mode 100755 index 000000000..8bfbc7b50 --- /dev/null +++ b/egs/multi_zh-hans/ASR/local/compute_fbank_kespeech_splits.py @@ -0,0 +1,180 @@ +#!/usr/bin/env python3 +# Copyright 2021 Johns Hopkins University (Piotr Żelasko) +# Copyright 2021 Xiaomi Corp. (Fangjun Kuang) +# Copyright 2023 Xiaomi Corp. (Zengrui Jin) +# +# See ../../../../LICENSE for clarification regarding multiple authors +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import argparse +import logging +from datetime import datetime +from pathlib import Path + +import torch +from lhotse import ( + CutSet, + KaldifeatFbank, + KaldifeatFbankConfig, + LilcomChunkyWriter, + set_audio_duration_mismatch_tolerance, + set_caching_enabled, +) + +# 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( + "--training-subset", + type=str, + default="train_phase1", + choices=["train_phase1", "train_phase2"], + help="The training subset for computing fbank feature.", + ) + + 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 given 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_kespeech_splits(args): + subset = args.training_subset + subset = str(subset) + num_splits = args.num_splits + output_dir = f"data/fbank/kespeech/{subset}_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}") + + set_audio_duration_mismatch_tolerance(0.01) # 10ms tolerance + set_caching_enabled(False) + for i in range(start, stop): + idx = f"{i + 1}".zfill(num_digits) + logging.info(f"Processing {idx}/{num_splits}") + + cuts_path = output_dir / f"kespeech-asr_cuts_{subset}.{idx}.jsonl.gz" + if cuts_path.is_file(): + logging.info(f"{cuts_path} exists - skipping") + continue + + raw_cuts_path = output_dir / f"kespeech-asr_cuts_{subset}_raw.{idx}.jsonl.gz" + + logging.info(f"Loading {raw_cuts_path}") + cut_set = CutSet.from_file(raw_cuts_path) + + logging.info("Splitting cuts into smaller chunks.") + cut_set = cut_set.trim_to_supervisions( + keep_overlapping=False, min_duration=None + ) + + logging.info("Computing features") + cut_set = cut_set.compute_and_store_features_batch( + extractor=extractor, + storage_path=f"{output_dir}/feats_{subset}_{idx}", + num_workers=args.num_workers, + batch_duration=args.batch_duration, + storage_type=LilcomChunkyWriter, + overwrite=True, + ) + + logging.info(f"Saving to {cuts_path}") + cut_set.to_file(cuts_path) + + +def main(): + now = datetime.now() + date_time = now.strftime("%Y-%m-%d-%H-%M-%S") + + log_filename = "log-compute_fbank_kespeech_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_kespeech_splits(args) + + +if __name__ == "__main__": + main() diff --git a/egs/multi_zh-hans/ASR/local/compute_fbank_magicdata.py b/egs/multi_zh-hans/ASR/local/compute_fbank_magicdata.py new file mode 100755 index 000000000..5649d3815 --- /dev/null +++ b/egs/multi_zh-hans/ASR/local/compute_fbank_magicdata.py @@ -0,0 +1,122 @@ +#!/usr/bin/env python3 +# Copyright 2023 Xiaomi Corp. (authors: Fangjun Kuang +# Zengrui Jin) +# +# See ../../../../LICENSE for clarification regarding multiple authors +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + + +""" +This file computes fbank features of the MagicData dataset. +It looks for manifests in the directory data/manifests/magicdata. + +The generated fbank features are saved in data/fbank. +""" + +import argparse +import logging +import os +from pathlib import Path + +import torch +from lhotse import CutSet, Fbank, FbankConfig, LilcomChunkyWriter +from lhotse.recipes.utils import read_manifests_if_cached + +from icefall.utils import get_executor + +# 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_magicdata(num_mel_bins: int = 80, speed_perturb: bool = False): + src_dir = Path("data/manifests/magicdata") + output_dir = Path("data/fbank") + num_jobs = min(30, os.cpu_count()) + + dataset_parts = ("train", "test", "dev") + prefix = "magicdata" + suffix = "jsonl.gz" + manifests = read_manifests_if_cached( + dataset_parts=dataset_parts, + output_dir=src_dir, + prefix=prefix, + suffix=suffix, + ) + assert manifests is not None + + assert len(manifests) == len(dataset_parts), ( + len(manifests), + len(dataset_parts), + list(manifests.keys()), + dataset_parts, + ) + + extractor = Fbank(FbankConfig(num_mel_bins=num_mel_bins)) + + with get_executor() as ex: # Initialize the executor only once. + for partition, m in manifests.items(): + if (output_dir / f"{prefix}_cuts_{partition}.{suffix}").is_file(): + logging.info(f"{partition} already exists - skipping.") + continue + logging.info(f"Processing {partition}") + cut_set = CutSet.from_manifests( + recordings=m["recordings"], + supervisions=m["supervisions"], + ) + if "train" in partition and speed_perturb: + cut_set = ( + cut_set + cut_set.perturb_speed(0.9) + cut_set.perturb_speed(1.1) + ) + cut_set = cut_set.compute_and_store_features( + extractor=extractor, + storage_path=f"{output_dir}/{prefix}_feats_{partition}", + # when an executor is specified, make more partitions + num_jobs=num_jobs if ex is None else 80, + executor=ex, + storage_type=LilcomChunkyWriter, + ) + cut_set.to_file(output_dir / f"{prefix}_cuts_{partition}.{suffix}") + + +def get_args(): + parser = argparse.ArgumentParser() + parser.add_argument( + "--num-mel-bins", + type=int, + default=80, + help="""The number of mel bins for Fbank""", + ) + parser.add_argument( + "--speed-perturb", + type=bool, + default=False, + help="Enable 0.9 and 1.1 speed perturbation for data augmentation. Default: False.", + ) + + return parser.parse_args() + + +if __name__ == "__main__": + formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s" + + logging.basicConfig(format=formatter, level=logging.INFO) + + args = get_args() + compute_fbank_magicdata( + num_mel_bins=args.num_mel_bins, speed_perturb=args.speed_perturb + ) diff --git a/egs/multi_zh-hans/ASR/local/compute_fbank_primewords.py b/egs/multi_zh-hans/ASR/local/compute_fbank_primewords.py new file mode 100755 index 000000000..303a16580 --- /dev/null +++ b/egs/multi_zh-hans/ASR/local/compute_fbank_primewords.py @@ -0,0 +1,122 @@ +#!/usr/bin/env python3 +# Copyright 2023 Xiaomi Corp. (authors: Fangjun Kuang +# Zengrui Jin) +# +# See ../../../../LICENSE for clarification regarding multiple authors +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + + +""" +This file computes fbank features of the Primewords dataset. +It looks for manifests in the directory data/manifests/primewords. + +The generated fbank features are saved in data/fbank. +""" + +import argparse +import logging +import os +from pathlib import Path + +import torch +from lhotse import CutSet, Fbank, FbankConfig, LilcomChunkyWriter +from lhotse.recipes.utils import read_manifests_if_cached + +from icefall.utils import get_executor + +# 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_primewords(num_mel_bins: int = 80, speed_perturb: bool = False): + src_dir = Path("data/manifests/primewords") + output_dir = Path("data/fbank") + num_jobs = min(15, os.cpu_count()) + + dataset_parts = ("train",) + prefix = "primewords" + suffix = "jsonl.gz" + manifests = read_manifests_if_cached( + dataset_parts=dataset_parts, + output_dir=src_dir, + prefix=prefix, + suffix=suffix, + ) + assert manifests is not None + + assert len(manifests) == len(dataset_parts), ( + len(manifests), + len(dataset_parts), + list(manifests.keys()), + dataset_parts, + ) + + extractor = Fbank(FbankConfig(num_mel_bins=num_mel_bins)) + + with get_executor() as ex: # Initialize the executor only once. + for partition, m in manifests.items(): + if (output_dir / f"{prefix}_cuts_{partition}.{suffix}").is_file(): + logging.info(f"{partition} already exists - skipping.") + continue + logging.info(f"Processing {partition}") + cut_set = CutSet.from_manifests( + recordings=m["recordings"], + supervisions=m["supervisions"], + ) + if "train" in partition and speed_perturb: + cut_set = ( + cut_set + cut_set.perturb_speed(0.9) + cut_set.perturb_speed(1.1) + ) + cut_set = cut_set.compute_and_store_features( + extractor=extractor, + storage_path=f"{output_dir}/{prefix}_feats_{partition}", + # when an executor is specified, make more partitions + num_jobs=num_jobs if ex is None else 80, + executor=ex, + storage_type=LilcomChunkyWriter, + ) + cut_set.to_file(output_dir / f"{prefix}_cuts_{partition}.{suffix}") + + +def get_args(): + parser = argparse.ArgumentParser() + parser.add_argument( + "--num-mel-bins", + type=int, + default=80, + help="""The number of mel bins for Fbank""", + ) + parser.add_argument( + "--speed-perturb", + type=bool, + default=False, + help="Enable 0.9 and 1.1 speed perturbation for data augmentation. Default: False.", + ) + + return parser.parse_args() + + +if __name__ == "__main__": + formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s" + + logging.basicConfig(format=formatter, level=logging.INFO) + + args = get_args() + compute_fbank_primewords( + num_mel_bins=args.num_mel_bins, speed_perturb=args.speed_perturb + ) diff --git a/egs/multi_zh-hans/ASR/local/compute_fbank_stcmds.py b/egs/multi_zh-hans/ASR/local/compute_fbank_stcmds.py new file mode 100755 index 000000000..730806954 --- /dev/null +++ b/egs/multi_zh-hans/ASR/local/compute_fbank_stcmds.py @@ -0,0 +1,121 @@ +#!/usr/bin/env python3 +# Copyright 2023 Xiaomi Corp. (authors: Fangjun Kuang +# Zengrui Jin) +# +# See ../../../../LICENSE for clarification regarding multiple authors +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + + +""" +This file computes fbank features of the ST-CMDS dataset. +It looks for manifests in the directory data/manifests/stcmds. + +The generated fbank features are saved in data/fbank. +""" + +import argparse +import logging +import os +from pathlib import Path + +import torch +from lhotse import CutSet, Fbank, FbankConfig, LilcomChunkyWriter +from lhotse.recipes.utils import read_manifests_if_cached + +from icefall.utils import get_executor + +# 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_stcmds(num_mel_bins: int = 80, speed_perturb: bool = False): + src_dir = Path("data/manifests/stcmds") + output_dir = Path("data/fbank") + num_jobs = min(15, os.cpu_count()) + + dataset_parts = ("train",) + prefix = "stcmds" + suffix = "jsonl.gz" + manifests = read_manifests_if_cached( + dataset_parts=dataset_parts, + output_dir=src_dir, + prefix=prefix, + suffix=suffix, + ) + assert manifests is not None + + assert len(manifests) == len(dataset_parts), ( + len(manifests), + len(dataset_parts), + list(manifests.keys()), + dataset_parts, + ) + + extractor = Fbank(FbankConfig(num_mel_bins=num_mel_bins)) + + with get_executor() as ex: # Initialize the executor only once. + for partition, m in manifests.items(): + if (output_dir / f"{prefix}_cuts_{partition}.{suffix}").is_file(): + logging.info(f"{partition} already exists - skipping.") + continue + logging.info(f"Processing {partition}") + cut_set = CutSet.from_manifests( + recordings=m["recordings"], + supervisions=m["supervisions"], + ) + if "train" in partition and speed_perturb: + cut_set = ( + cut_set + cut_set.perturb_speed(0.9) + cut_set.perturb_speed(1.1) + ) + cut_set = cut_set.compute_and_store_features( + extractor=extractor, + storage_path=f"{output_dir}/{prefix}_feats_{partition}", + # when an executor is specified, make more partitions + num_jobs=num_jobs if ex is None else 80, + executor=ex, + storage_type=LilcomChunkyWriter, + ) + cut_set.to_file(output_dir / f"{prefix}_cuts_{partition}.{suffix}") + + +def get_args(): + parser = argparse.ArgumentParser() + parser.add_argument( + "--num-mel-bins", + type=int, + default=80, + help="""The number of mel bins for Fbank""", + ) + parser.add_argument( + "--speed-perturb", + type=bool, + default=False, + help="Enable 0.9 and 1.1 speed perturbation for data augmentation. Default: False.", + ) + return parser.parse_args() + + +if __name__ == "__main__": + formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s" + + logging.basicConfig(format=formatter, level=logging.INFO) + + args = get_args() + compute_fbank_stcmds( + num_mel_bins=args.num_mel_bins, speed_perturb=args.speed_perturb + ) diff --git a/egs/multi_zh-hans/ASR/local/compute_fbank_thchs30.py b/egs/multi_zh-hans/ASR/local/compute_fbank_thchs30.py new file mode 100755 index 000000000..58bb8002a --- /dev/null +++ b/egs/multi_zh-hans/ASR/local/compute_fbank_thchs30.py @@ -0,0 +1,127 @@ +#!/usr/bin/env python3 +# Copyright 2023 Xiaomi Corp. (authors: Fangjun Kuang +# Zengrui Jin) +# +# See ../../../../LICENSE for clarification regarding multiple authors +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + + +""" +This file computes fbank features of the THCHS-30 dataset. +It looks for manifests in the directory data/manifests/thchs30. + +The generated fbank features are saved in data/fbank. +""" + +import argparse +import logging +import os +from pathlib import Path + +import torch +from lhotse import CutSet, Fbank, FbankConfig, LilcomChunkyWriter +from lhotse.recipes.utils import read_manifests_if_cached + +from icefall.utils import get_executor + +# 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_thchs30(num_mel_bins: int = 80, speed_perturb: bool = False): + src_dir = Path("data/manifests/thchs30") + output_dir = Path("data/fbank") + num_jobs = min(15, os.cpu_count()) + + dataset_parts = ( + "train", + "dev", + "test", + ) + prefix = "thchs_30" + suffix = "jsonl.gz" + manifests = read_manifests_if_cached( + dataset_parts=dataset_parts, + output_dir=src_dir, + prefix=prefix, + suffix=suffix, + ) + assert manifests is not None + + assert len(manifests) == len(dataset_parts), ( + len(manifests), + len(dataset_parts), + list(manifests.keys()), + dataset_parts, + ) + + extractor = Fbank(FbankConfig(num_mel_bins=num_mel_bins)) + + with get_executor() as ex: # Initialize the executor only once. + for partition, m in manifests.items(): + if (output_dir / f"{prefix}_cuts_{partition}.{suffix}").is_file(): + logging.info(f"{partition} already exists - skipping.") + continue + logging.info(f"Processing {partition}") + cut_set = CutSet.from_manifests( + recordings=m["recordings"], + supervisions=m["supervisions"], + ) + if "train" in partition: + cut_set = ( + (cut_set + cut_set.perturb_speed(0.9) + cut_set.perturb_speed(1.1)) + if speed_perturb + else cut_set + ) + cut_set = cut_set.compute_and_store_features( + extractor=extractor, + storage_path=f"{output_dir}/{prefix}_feats_{partition}", + # when an executor is specified, make more partitions + num_jobs=num_jobs if ex is None else 80, + executor=ex, + storage_type=LilcomChunkyWriter, + ) + cut_set.to_file(output_dir / f"{prefix}_cuts_{partition}.{suffix}") + + +def get_args(): + parser = argparse.ArgumentParser() + parser.add_argument( + "--num-mel-bins", + type=int, + default=80, + help="""The number of mel bins for Fbank""", + ) + parser.add_argument( + "--speed-perturb", + type=bool, + default=False, + help="Enable 0.9 and 1.1 speed perturbation for data augmentation. Default: False.", + ) + return parser.parse_args() + + +if __name__ == "__main__": + formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s" + + logging.basicConfig(format=formatter, level=logging.INFO) + + args = get_args() + compute_fbank_thchs30( + num_mel_bins=args.num_mel_bins, speed_perturb=args.speed_perturb + ) diff --git a/egs/multi_zh-hans/ASR/local/prepare_char.py b/egs/multi_zh-hans/ASR/local/prepare_char.py new file mode 120000 index 000000000..be7da61af --- /dev/null +++ b/egs/multi_zh-hans/ASR/local/prepare_char.py @@ -0,0 +1 @@ +../../../wenetspeech/ASR/local/prepare_char.py \ No newline at end of file diff --git a/egs/multi_zh-hans/ASR/local/prepare_for_bpe_model.py b/egs/multi_zh-hans/ASR/local/prepare_for_bpe_model.py new file mode 100755 index 000000000..020800c15 --- /dev/null +++ b/egs/multi_zh-hans/ASR/local/prepare_for_bpe_model.py @@ -0,0 +1,65 @@ +#!/usr/bin/env python3 +# Copyright 2023 Xiaomi Corp. (authors: Zengrui Jin) +# +# See ../../../../LICENSE for clarification regarding multiple authors +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +# This script tokenizes the training transcript by CJK characters +# and saves the result to transcript_chars.txt, which is used +# to train the BPE model later. + +import argparse +from pathlib import Path + +from tqdm.auto import tqdm + +from icefall.utils import tokenize_by_CJK_char + + +def get_args(): + parser = argparse.ArgumentParser() + parser.add_argument( + "--lang-dir", + type=str, + help="""Output directory. + The generated transcript_chars.txt is saved to this directory. + """, + ) + + parser.add_argument( + "--text", + type=str, + help="WenetSpeech training transcript.", + ) + + return parser.parse_args() + + +def main(): + args = get_args() + lang_dir = Path(args.lang_dir) + text = Path(args.text) + + assert lang_dir.exists() and text.exists(), f"{lang_dir} or {text} does not exist!" + + transcript_path = lang_dir / "transcript_chars.txt" + + with open(text, "r", encoding="utf-8") as fin: + with open(transcript_path, "w+", encoding="utf-8") as fout: + for line in fin: + fout.write(tokenize_by_CJK_char(line) + "\n") + + +if __name__ == "__main__": + main() diff --git a/egs/multi_zh-hans/ASR/local/prepare_lang.py b/egs/multi_zh-hans/ASR/local/prepare_lang.py new file mode 120000 index 000000000..747f2ab39 --- /dev/null +++ b/egs/multi_zh-hans/ASR/local/prepare_lang.py @@ -0,0 +1 @@ +../../../librispeech/ASR/local/prepare_lang.py \ No newline at end of file diff --git a/egs/multi_zh-hans/ASR/local/prepare_lang_bpe.py b/egs/multi_zh-hans/ASR/local/prepare_lang_bpe.py new file mode 120000 index 000000000..36b40e7fc --- /dev/null +++ b/egs/multi_zh-hans/ASR/local/prepare_lang_bpe.py @@ -0,0 +1 @@ +../../../librispeech/ASR/local/prepare_lang_bpe.py \ No newline at end of file diff --git a/egs/multi_zh-hans/ASR/local/preprocess_kespeech.py b/egs/multi_zh-hans/ASR/local/preprocess_kespeech.py new file mode 100755 index 000000000..20274263f --- /dev/null +++ b/egs/multi_zh-hans/ASR/local/preprocess_kespeech.py @@ -0,0 +1,151 @@ +#!/usr/bin/env python3 +# Copyright 2021 Johns Hopkins University (Piotr Żelasko) +# Copyright 2021 Xiaomi Corp. (Fangjun Kuang) +# Copyright 2023 Xiaomi Corp. (Zengrui Jin) +# +# See ../../../../LICENSE for clarification regarding multiple authors +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import argparse +import logging +import re +from pathlib import Path + +from lhotse import CutSet, SupervisionSegment +from lhotse.recipes.utils import read_manifests_if_cached + +from icefall import setup_logger + +# Similar text filtering and normalization procedure as in: +# https://github.com/SpeechColab/WenetSpeech/blob/main/toolkits/kaldi/wenetspeech_data_prep.sh + + +def normalize_text( + utt: str, + punct_pattern=re.compile(r"<(PERIOD|QUESTIONMARK|EXCLAMATIONPOINT)>"), + whitespace_pattern=re.compile(r"\s\s+"), +) -> str: + return whitespace_pattern.sub(" ", punct_pattern.sub("", utt)) + + +def has_no_oov( + sup: SupervisionSegment, + oov_pattern=re.compile(r"<(SIL|MUSIC|NOISE|OTHER|SPOKEN_NOISE)>"), +) -> bool: + return oov_pattern.search(sup.text) is None + + +def preprocess_kespeech(speed_perturb: bool = False): + src_dir = Path("data/manifests/kespeech") + output_dir = Path("data/fbank/kespeech") + output_dir.mkdir(exist_ok=True) + + # Note: By default, we preprocess all sub-parts. + # You can delete those that you don't need. + # For instance, if you don't want to use the test subpart, just remove + # the line below containing "test" + dataset_parts = ( + "dev_phase1", + "dev_phase2", + "test", + "train_phase1", + "train_phase2", + ) + + logging.info("Loading manifest (may take 10 minutes)") + manifests = read_manifests_if_cached( + dataset_parts=dataset_parts, + output_dir=src_dir, + suffix="jsonl.gz", + prefix="kespeech-asr", + ) + assert manifests is not None + + assert len(manifests) == len(dataset_parts), ( + len(manifests), + len(dataset_parts), + list(manifests.keys()), + dataset_parts, + ) + logging_threshold = 50 + logging_count = 0 + + for partition, m in manifests.items(): + logging.info(f"Processing {partition}") + raw_cuts_path = output_dir / f"kespeech-asr_cuts_{partition}_raw.jsonl.gz" + if raw_cuts_path.is_file(): + logging.info(f"{partition} already exists - skipping") + continue + + # Note this step makes the recipe different than LibriSpeech: + # We must filter out some utterances and remove punctuation + # to be consistent with Kaldi. + logging.info("Filtering OOV utterances from supervisions") + m["supervisions"] = m["supervisions"].filter(has_no_oov) + logging.info(f"Normalizing text in {partition}") + for sup in m["supervisions"]: + orig_text = sup.text + sup.text = normalize_text(sup.text) + if logging_count < logging_threshold and len(orig_text) != len(sup.text): + logging_count += 1 + logging.info( + f"\nOriginal text vs normalized text:\n{orig_text}\n{sup.text}" + ) + + # Create long-recording cut manifests. + logging.info(f"Processing {partition}") + cut_set = CutSet.from_manifests( + recordings=m["recordings"], + supervisions=m["supervisions"], + ) + # Run data augmentation that needs to be done in the + # time domain. + if partition not in [ + "dev_phase1", + "dev_phase2", + "test", + ]: + if speed_perturb: + 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(f"Saving to {raw_cuts_path}") + cut_set.to_file(raw_cuts_path) + + +def get_args(): + parser = argparse.ArgumentParser() + parser.add_argument( + "--speed-perturb", + type=bool, + default=False, + help="Enable 0.9 and 1.1 speed perturbation for data augmentation. Default: False.", + ) + return parser.parse_args() + + +def main(): + setup_logger(log_filename="./log-preprocess-kespeech") + + args = get_args() + preprocess_kespeech(speed_perturb=args.speed_perturb) + logging.info("Done") + + +if __name__ == "__main__": + main() diff --git a/egs/multi_zh-hans/ASR/local/text2token.py b/egs/multi_zh-hans/ASR/local/text2token.py new file mode 120000 index 000000000..ce5cfd537 --- /dev/null +++ b/egs/multi_zh-hans/ASR/local/text2token.py @@ -0,0 +1 @@ +../../../wenetspeech/ASR/local/text2token.py \ No newline at end of file diff --git a/egs/multi_zh-hans/ASR/local/train_bpe_model.py b/egs/multi_zh-hans/ASR/local/train_bpe_model.py new file mode 100755 index 000000000..976ea0ba8 --- /dev/null +++ b/egs/multi_zh-hans/ASR/local/train_bpe_model.py @@ -0,0 +1,109 @@ +#!/usr/bin/env python3 +# Copyright 2021 Xiaomi Corp. (authors: Fangjun Kuang) +# Copyright 2023 Xiaomi Corp. (authors: Zengrui Jin) +# +# See ../../../../LICENSE for clarification regarding multiple authors +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + + +# You can install sentencepiece via: +# +# pip install sentencepiece +# +# Due to an issue reported in +# https://github.com/google/sentencepiece/pull/642#issuecomment-857972030 +# +# Please install a version >=0.1.96 + +import argparse +import shutil +from pathlib import Path + +import sentencepiece as spm + + +def get_args(): + parser = argparse.ArgumentParser() + parser.add_argument( + "--lang-dir", + type=str, + help="""Input and output directory. + The generated bpe.model is saved to this directory. + """, + ) + + parser.add_argument( + "--transcript", + type=str, + help="Training transcript.", + ) + + parser.add_argument( + "--vocab-size", + type=int, + help="Vocabulary size for BPE training", + ) + + parser.add_argument( + "--byte-fallback", + type=bool, + default=True, + help="Enable byte fallback for BPE model.", + ) + + return parser.parse_args() + + +def main(): + args = get_args() + vocab_size = args.vocab_size + lang_dir = Path(args.lang_dir) + + model_type = "unigram" + + model_prefix = f"{lang_dir}/{model_type}_{vocab_size}" + train_text = args.transcript + character_coverage = 0.98 + input_sentence_size = 100000000 + + user_defined_symbols = ["", ""] + unk_id = len(user_defined_symbols) + # Note: unk_id is fixed to 2. + # If you change it, you should also change other + # places that are using it. + + model_file = Path(model_prefix + ".model") + if not model_file.is_file(): + spm.SentencePieceTrainer.train( + input=train_text, + vocab_size=vocab_size, + model_type=model_type, + model_prefix=model_prefix, + input_sentence_size=input_sentence_size, + character_coverage=character_coverage, + user_defined_symbols=user_defined_symbols, + unk_id=unk_id, + bos_id=-1, + eos_id=-1, + byte_fallback=args.byte_fallback, + ) + else: + print(f"{model_file} exists - skipping") + return + + shutil.copyfile(model_file, f"{lang_dir}/bpe.model") + + +if __name__ == "__main__": + main() diff --git a/egs/multi_zh-hans/ASR/local/validate_bpe_lexicon.py b/egs/multi_zh-hans/ASR/local/validate_bpe_lexicon.py new file mode 120000 index 000000000..721bb48e7 --- /dev/null +++ b/egs/multi_zh-hans/ASR/local/validate_bpe_lexicon.py @@ -0,0 +1 @@ +../../../librispeech/ASR/local/validate_bpe_lexicon.py \ No newline at end of file diff --git a/egs/multi_zh-hans/ASR/prepare.sh b/egs/multi_zh-hans/ASR/prepare.sh new file mode 100755 index 000000000..c09b9c1de --- /dev/null +++ b/egs/multi_zh-hans/ASR/prepare.sh @@ -0,0 +1,373 @@ +#!/usr/bin/env bash + +# fix segmentation fault reported in https://github.com/k2-fsa/icefall/issues/674 +export PROTOCOL_BUFFERS_PYTHON_IMPLEMENTATION=python + +set -eou pipefail + +stage=-1 +stop_stage=100 +num_splits=100 + +dl_dir=$PWD/download + +. shared/parse_options.sh || exit 1 + +vocab_sizes=( + 2000 +) + +# All files generated by this script are saved in "data". +# You can safely remove "data" and rerun this script to regenerate it. +mkdir -p data + +log() { + # This function is from espnet + local fname=${BASH_SOURCE[1]##*/} + echo -e "$(date '+%Y-%m-%d %H:%M:%S') (${fname}:${BASH_LINENO[0]}:${FUNCNAME[1]}) $*" +} + +log "dl_dir: $dl_dir" + +log "Dataset: musan" +if [ $stage -le 1 ] && [ $stop_stage -ge 1 ]; then + log "Stage 1: Soft link fbank of musan" + mkdir -p data/fbank + if [ -e ../../librispeech/ASR/data/fbank/.musan.done ]; then + cd data/fbank + ln -svf $(realpath ../../../../librispeech/ASR/data/fbank/musan_feats) . + ln -svf $(realpath ../../../../librispeech/ASR/data/fbank/musan_cuts.jsonl.gz) . + cd ../.. + else + log "Abort! Please run ../../librispeech/ASR/prepare.sh --stage 4 --stop-stage 4" + exit 1 + fi +fi + +log "Dataset: THCHS-30" +if [ $stage -le 2 ] && [ $stop_stage -ge 2 ]; then + log "Stage 2: Prepare THCHS-30" + if [ ! -d $dl_dir/thchs30 ]; then + log "Downloading THCHS-30" + lhotse download thchs-30 $dl_dir/thchs30 + fi + + if [ ! -f data/manifests/.thchs30.done ]; then + mkdir -p data/manifests + lhotse prepare thchs-30 $dl_dir/thchs30 data/manifests/thchs30 + touch data/manifests/.thchs30.done + fi + + if [ ! -f data/fbank/.thchs30.done ]; then + mkdir -p data/fbank + ./local/compute_fbank_thchs30.py + touch data/fbank/.thchs30.done + fi +fi + +log "Dataset: AISHELL-1" +if [ $stage -le 3 ] && [ $stop_stage -ge 3 ]; then + log "Stage 3: Prepare AISHELL-1" + if [ -e ../../aishell/ASR/data/fbank/.aishell.done ]; then + cd data/fbank + ln -svf $(realpath ../../../../aishell/ASR/data/fbank/aishell_feats_train) . + ln -svf $(realpath ../../../../aishell/ASR/data/fbank/aishell_feats_dev) . + ln -svf $(realpath ../../../../aishell/ASR/data/fbank/aishell_feats_test) . + ln -svf $(realpath ../../../../aishell/ASR/data/fbank/aishell_cuts_train.jsonl.gz) . + ln -svf $(realpath ../../../../aishell/ASR/data/fbank/aishell_cuts_dev.jsonl.gz) . + ln -svf $(realpath ../../../../aishell/ASR/data/fbank/aishell_cuts_test.jsonl.gz) . + cd ../.. + else + log "Abort! Please run ../../aishell/ASR/prepare.sh --stage 3 --stop-stage 3" + exit 1 + fi +fi + +log "Dataset: AISHELL-2" +if [ $stage -le 4 ] && [ $stop_stage -ge 4 ]; then + log "Stage 4: Prepare AISHELL-2" + if [ -e ../../aishell/ASR/data/fbank/.aishell2.done ]; then + cd data/fbank + ln -svf $(realpath ../../../../aishell2/ASR/data/fbank/aishell2_feats_train) . + ln -svf $(realpath ../../../../aishell2/ASR/data/fbank/aishell2_feats_dev) . + ln -svf $(realpath ../../../../aishell2/ASR/data/fbank/aishell2_feats_test) . + ln -svf $(realpath ../../../../aishell2/ASR/data/fbank/aishell2_cuts_train.jsonl.gz) . + ln -svf $(realpath ../../../../aishell2/ASR/data/fbank/aishell2_cuts_dev.jsonl.gz) . + ln -svf $(realpath ../../../../aishell2/ASR/data/fbank/aishell2_cuts_test.jsonl.gz) . + cd ../.. + else + log "Abort! Please run ../../aishell2/ASR/prepare.sh --stage 3 --stop-stage 3" + exit 1 + fi +fi + +log "Dataset: AISHELL-4" +if [ $stage -le 5 ] && [ $stop_stage -ge 5 ]; then + log "Stage 5: Prepare AISHELL-4" + if [ -e ../../aishell/ASR/data/fbank/.aishell4.done ]; then + cd data/fbank + ln -svf $(realpath ../../../../aishell4/ASR/data/fbank/aishell4_feats_train) . + ln -svf $(realpath ../../../../aishell4/ASR/data/fbank/aishell4_feats_dev) . + ln -svf $(realpath ../../../../aishell4/ASR/data/fbank/aishell4_feats_test) . + ln -svf $(realpath ../../../../aishell4/ASR/data/fbank/aishell4_cuts_train_L.jsonl.gz) . + ln -svf $(realpath ../../../../aishell4/ASR/data/fbank/aishell4_cuts_train_M.jsonl.gz) . + ln -svf $(realpath ../../../../aishell4/ASR/data/fbank/aishell4_cuts_train_S.jsonl.gz) . + ln -svf $(realpath ../../../../aishell4/ASR/data/fbank/aishell4_cuts_dev.jsonl.gz) . + ln -svf $(realpath ../../../../aishell4/ASR/data/fbank/aishell4_cuts_test.jsonl.gz) . + cd ../.. + else + log "Abort! Please run ../../aishell4/ASR/prepare.sh --stage 3 --stop-stage 3" + exit 1 + fi +fi + +log "Dataset: ST-CMDS" +if [ $stage -le 6 ] && [ $stop_stage -ge 6 ]; then + log "Stage 6: Prepare ST-CMDS" + if [ ! -f $dl_dir/stcmds/ST-CMDS-20170001_1-OS.tar.gz ]; then + log "Downloading ST-CMDS" + lhotse download stcmds $dl_dir/stcmds + fi + + if [ ! -f data/manifests/.stcmds.done ]; then + mkdir -p data/manifests + lhotse prepare stcmds $dl_dir/stcmds data/manifests/stcmds + touch data/manifests/.stcmds.done + fi + + if [ ! -f data/fbank/.stcmds.done ]; then + mkdir -p data/fbank + ./local/compute_fbank_stcmds.py + touch data/fbank/.stcmds.done + fi +fi + + +log "Dataset: Primewords" +if [ $stage -le 7 ] && [ $stop_stage -ge 7 ]; then + log "Stage 7: Prepare Primewords" + if [ ! -f $dl_dir/primewords/primewords_md_2018_set1.tar.gz ]; then + log "Downloading Primewords" + lhotse download primewords $dl_dir/primewords + fi + + if [ ! -f data/manifests/.stcmds.done ]; then + mkdir -p data/manifests + lhotse prepare stcmds $dl_dir/primewords data/manifests/primewords + touch data/manifests/.primewords.done + fi + + if [ ! -f data/fbank/.primewords.done ]; then + mkdir -p data/fbank + ./local/compute_fbank_primewords.py + touch data/fbank/.primewords.done + fi +fi + +log "Dataset: MagicData" +if [ $stage -le 8 ] && [ $stop_stage -ge 8 ]; then + log "Stage 8: Prepare MagicData" + if [ ! -f $dl_dir/magicdata/train_set.tar.gz ]; then + log "Downloading MagicData" + lhotse download magicdata $dl_dir/magicdata + fi + + if [ ! -f data/manifests/.magicdata.done ]; then + mkdir -p data/manifests + lhotse prepare magicdata $dl_dir/magicdata data/manifests/magicdata + touch data/manifests/.magicdata.done + fi + + if [ ! -f data/fbank/.magicdata.done ]; then + mkdir -p data/fbank + ./local/compute_fbank_magicdata.py + touch data/fbank/.magicdata.done + fi +fi + +log "Dataset: aidatatang_200zh" +if [ $stage -le 9 ] && [ $stop_stage -ge 9 ]; then + log "Stage 9: Prepare aidatatang_200zh" + if [ -e ../../aidatatang_200zh/ASR/data/fbank/.aidatatang_200zh.done ]; then + cd data/fbank + ln -svf $(realpath ../../../../aidatatang_200zh/ASR/data/fbank/aidatatang_feats_train) . + ln -svf $(realpath ../../../../aidatatang_200zh/ASR/data/fbank/aidatatang_feats_dev) . + ln -svf $(realpath ../../../../aidatatang_200zh/ASR/data/fbank/aidatatang_feats_test) . + ln -svf $(realpath ../../../../aidatatang_200zh/ASR/data/fbank/aidatatang_cuts_train.jsonl.gz) . + ln -svf $(realpath ../../../../aidatatang_200zh/ASR/data/fbank/aidatatang_cuts_dev.jsonl.gz) . + ln -svf $(realpath ../../../../aidatatang_200zh/ASR/data/fbank/aidatatang_cuts_test.jsonl.gz) . + cd ../.. + else + log "Abort! Please run ../../aidatatang_200zh/ASR/prepare.sh --stage 4 --stop-stage 4" + exit 1 + fi +fi + +log "Dataset: Ali-Meeting" +if [ $stage -le 10 ] && [ $stop_stage -ge 10 ]; then + log "Stage 10: Prepare Ali-Meeting" + if [ -e ../../alimeeting/ASR/data/fbank/.fbank.done ]; then + cd data/fbank + ln -svf $(realpath ../../../../alimeeting/ASR/data/fbank/alimeeting-far_feats_train) . + ln -svf $(realpath ../../../../alimeeting/ASR/data/fbank/alimeeting-far_feats_eval) . + ln -svf $(realpath ../../../../alimeeting/ASR/data/fbank/alimeeting-far_feats_test) . + ln -svf $(realpath ../../../../alimeeting/ASR/data/fbank/alimeeting-far_cuts_train.jsonl.gz) . + ln -svf $(realpath ../../../../alimeeting/ASR/data/fbank/alimeeting-far_cuts_eval.jsonl.gz) . + ln -svf $(realpath ../../../../alimeeting/ASR/data/fbank/alimeeting-far_cuts_test.jsonl.gz) . + cd ../.. + else + log "Abort! Please run ../../alimeeting/ASR/prepare.sh --stage 5 --stop-stage 5" + exit 1 + fi +fi + +log "Dataset: WenetSpeech" +if [ $stage -le 11 ] && [ $stop_stage -ge 11 ]; then + log "Stage 11: Prepare WenetSpeech" + if [ -e ../../wenetspeech/ASR/data/fbank/.preprocess_complete ]; then + cd data/fbank + ln -svf $(realpath ../../../../wenetspeech/ASR/data/fbank/cuts_DEV.jsonl.gz) . + ln -svf $(realpath ../../../../wenetspeech/ASR/data/fbank/cuts_L.jsonl.gz) . + ln -svf $(realpath ../../../../wenetspeech/ASR/data/fbank/cuts_TEST_MEETING.jsonl.gz) . + ln -svf $(realpath ../../../../wenetspeech/ASR/data/fbank/cuts_TEST_NET.jsonl.gz) . + + ln -svf $(realpath ../../../../wenetspeech/ASR/data/fbank/L_split_1000) . + ln -svf $(realpath ../../../../wenetspeech/ASR/data/fbank/*.lca) . + ln -svf $(realpath ../../../../wenetspeech/ASR/data/fbank/) ./wenetspeech + cd ../.. + else + log "Abort! Please run ../../wenetspeech/ASR/prepare.sh" + exit 1 + fi + + if [ -d ../../wenetspeech/ASR/data/lang_char/ ]; then + cd data + cp -r ../../../../wenetspeech/ASR/data/lang_char . + cd .. + else + log "Abort! Please run ../../wenetspeech/ASR/prepare.sh" + exit 1 + fi +fi + +log "Dataset: KeSpeech" +if [ $stage -le 12 ] && [ $stop_stage -ge 12 ]; then + log "Stage 12: Prepare KeSpeech" + if [ ! -d $dl_dir/KeSpeech ]; then + log "Abort! Please download KeSpeech first." + log "KeSpeech download link: https://github.com/KeSpeech/KeSpeech" + exit 1 + fi + + if [ ! -f data/manifests/.kespeech.done ]; then + mkdir -p data/manifests + lhotse prepare kespeech -j 16 $dl_dir/KeSpeech data/manifests/kespeech + touch data/manifests/.kespeech.done + fi + + if [ ! -f data/fbank/.kespeech.done ]; then + mkdir -p data/fbank + + log "Preprocess KeSpeech manifest" + if [ ! -f data/fbank/.kespeech_preprocess_complete ]; then + python3 ./local/preprocess_kespeech.py + touch data/fbank/.kespeech_preprocess_complete + fi + + if [ -f data/fbank/.kespeech.train_phase1.split.${num_splits}.done ]; then + log "Spliting KeSpeech train_phase1" + lhotse split ${num_splits} \ + data/fbank/kespeech/kespeech-asr_cuts_train_phase1_raw.jsonl.gz \ + data/fbank/kespeech/train_phase1_split_${num_splits} + touch data/fbank/.kespeech.train_phase1.split.${num_splits}.done + fi + + if [ -f data/fbank/.kespeech.train_phase2.split.${num_splits}.done ]; then + log "Spliting KeSpeech train_phase2" + lhotse split ${num_splits} \ + data/fbank/kespeech/kespeech-asr_cuts_train_phase2_raw.jsonl.gz \ + data/fbank/kespeech/train_phase2_split_${num_splits} + touch data/fbank/.kespeech.train_phase2.split.${num_splits}.done + fi + + log "Compute KeSpeech fbank for train_phase1" + ./local/compute_fbank_kespeech_splits.py --num-splits ${num_splits} --training-subset train_phase1 + + log "Compute KeSpeech fbank for train_phase2" + ./local/compute_fbank_kespeech_splits.py --num-splits ${num_splits} --training-subset train_phase2 + + log "Compute KeSpeech fbank for test/dev" + ./local/compute_fbank_kespeech_dev_test.py + + touch data/fbank/.kespeech.done + fi +fi + +if [ $stage -le 13 ] && [ $stop_stage -ge 13 ]; then + log "Stage 13: BPE model training (note that we use transcripts of wenetspeech only for BPE training)" + ./local/prepare_for_bpe_model.py --lang-dir ./data/lang_char --text ./data/lang_char/text + + for vocab_size in ${vocab_sizes[@]}; do + lang_dir=data/lang_bpe_${vocab_size} + + mkdir -p $lang_dir + if [ ! -f $lang_dir/bpe.model ]; then + ./local/train_bpe_model.py \ + --lang-dir $lang_dir \ + --transcript ./data/lang_char/transcript_chars.txt \ + --vocab-size $vocab_size + + ./local/bpe_model_to_tokens.py $lang_dir/bpe.model > $lang_dir/tokens.txt + fi + + if [ ! -f $lang_dir/L_disambig.pt ]; then + cp data/lang_char/words.txt $lang_dir + + ./local/prepare_lang_bpe.py --lang-dir $lang_dir + log "Validating $lang_dir/lexicon.txt" + ./local/validate_bpe_lexicon.py \ + --lexicon $lang_dir/lexicon.txt \ + --bpe-model $lang_dir/bpe.model + fi + + if [ ! -f $lang_dir/L.fst ]; then + log "Converting L.pt to L.fst" + ./shared/convert-k2-to-openfst.py \ + --olabels aux_labels \ + $lang_dir/L.pt \ + $lang_dir/L.fst + fi + + if [ ! -f $lang_dir/L_disambig.fst ]; then + log "Converting L_disambig.pt to L_disambig.fst" + ./shared/convert-k2-to-openfst.py \ + --olabels aux_labels \ + $lang_dir/L_disambig.pt \ + $lang_dir/L_disambig.fst + fi + done +fi + +if [ $stage -le 14 ] && [ $stop_stage -ge 14 ]; then + log "Stage 14: Prepare G (note that we use ngram lm of wenetspeech only for G preparation)" + + if [ -d ../../wenetspeech/ASR/data/lang_char/ ]; then + cd data + ln -s ../../../../wenetspeech/ASR/data/lm . + cd .. + else + log "Abort! Please run ../../wenetspeech/ASR/prepare.sh" + exit 1 + fi +fi + +if [ $stage -le 15 ] && [ $stop_stage -ge 15 ]; then + log "Stage 15: Compile LG" + for vocab_size in ${vocab_sizes[@]}; do + lang_dir=data/lang_bpe_${vocab_size} + + python ./local/compile_lg.py --lang-dir $lang_dir + done +fi + + diff --git a/egs/multi_zh-hans/ASR/shared b/egs/multi_zh-hans/ASR/shared new file mode 120000 index 000000000..4cbd91a7e --- /dev/null +++ b/egs/multi_zh-hans/ASR/shared @@ -0,0 +1 @@ +../../../icefall/shared \ No newline at end of file diff --git a/egs/multi_zh-hans/ASR/zipformer/asr_datamodule.py b/egs/multi_zh-hans/ASR/zipformer/asr_datamodule.py new file mode 100644 index 000000000..3d58ebf3a --- /dev/null +++ b/egs/multi_zh-hans/ASR/zipformer/asr_datamodule.py @@ -0,0 +1,388 @@ +# Copyright 2021 Piotr Żelasko +# Copyright 2022 Xiaomi Corporation (Author: Mingshuang Luo) +# +# See ../../../../LICENSE for clarification regarding multiple authors +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + + +import argparse +import inspect +import logging +from functools import lru_cache +from pathlib import Path +from typing import Any, Dict, Optional + +import torch +from lhotse import CutSet, Fbank, FbankConfig, load_manifest, load_manifest_lazy +from lhotse.dataset import ( # noqa F401 for PrecomputedFeatures + CutConcatenate, + CutMix, + DynamicBucketingSampler, + K2SpeechRecognitionDataset, + PrecomputedFeatures, + SimpleCutSampler, + SpecAugment, +) +from lhotse.dataset.input_strategies import ( # noqa F401 For AudioSamples + AudioSamples, + OnTheFlyFeatures, +) +from lhotse.utils import fix_random_seed +from torch.utils.data import DataLoader + +from icefall.utils import str2bool + + +class _SeedWorkers: + def __init__(self, seed: int): + self.seed = seed + + def __call__(self, worker_id: int): + fix_random_seed(self.seed + worker_id) + + +class AsrDataModule: + """ + DataModule for k2 ASR experiments. + It assumes there is always one train and valid dataloader, + but there can be multiple test dataloaders (e.g. LibriSpeech test-clean + and test-other). + + It contains all the common data pipeline modules used in ASR + experiments, e.g.: + - dynamic batch size, + - bucketing samplers, + - cut concatenation, + - augmentation, + - on-the-fly feature extraction + + This class should be derived for specific corpora used in ASR tasks. + """ + + def __init__(self, args: argparse.Namespace): + self.args = args + + @classmethod + def add_arguments(cls, parser: argparse.ArgumentParser): + group = parser.add_argument_group( + title="ASR data related options", + description="These options are used for the preparation of " + "PyTorch DataLoaders from Lhotse CutSet's -- they control the " + "effective batch sizes, sampling strategies, applied data " + "augmentations, etc.", + ) + group.add_argument( + "--manifest-dir", + type=Path, + default=Path("data/fbank"), + help="Path to directory with train/valid/test cuts.", + ) + group.add_argument( + "--max-duration", + type=int, + default=300.0, + help="Maximum pooled recordings duration (seconds) in a " + "single batch. You can reduce it if it causes CUDA OOM.", + ) + group.add_argument( + "--bucketing-sampler", + type=str2bool, + default=True, + help="When enabled, the batches will come from buckets of " + "similar duration (saves padding frames).", + ) + group.add_argument( + "--num-buckets", + type=int, + default=30, + help="The number of buckets for the DynamicBucketingSampler" + "(you might want to increase it for larger datasets).", + ) + group.add_argument( + "--concatenate-cuts", + type=str2bool, + default=False, + help="When enabled, utterances (cuts) will be concatenated " + "to minimize the amount of padding.", + ) + group.add_argument( + "--duration-factor", + type=float, + default=1.0, + help="Determines the maximum duration of a concatenated cut " + "relative to the duration of the longest cut in a batch.", + ) + group.add_argument( + "--gap", + type=float, + default=1.0, + help="The amount of padding (in seconds) inserted between " + "concatenated cuts. This padding is filled with noise when " + "noise augmentation is used.", + ) + group.add_argument( + "--on-the-fly-feats", + type=str2bool, + default=False, + help="When enabled, use on-the-fly cut mixing and feature " + "extraction. Will drop existing precomputed feature manifests " + "if available.", + ) + group.add_argument( + "--shuffle", + type=str2bool, + default=True, + help="When enabled (=default), the examples will be " + "shuffled for each epoch.", + ) + group.add_argument( + "--drop-last", + type=str2bool, + default=True, + help="Whether to drop last batch. Used by sampler.", + ) + group.add_argument( + "--return-cuts", + type=str2bool, + default=True, + help="When enabled, each batch will have the " + "field: batch['supervisions']['cut'] with the cuts that " + "were used to construct it.", + ) + + group.add_argument( + "--num-workers", + type=int, + default=2, + help="The number of training dataloader workers that " + "collect the batches.", + ) + + group.add_argument( + "--enable-spec-aug", + type=str2bool, + default=True, + help="When enabled, use SpecAugment for training dataset.", + ) + + group.add_argument( + "--spec-aug-time-warp-factor", + type=int, + default=80, + help="Used only when --enable-spec-aug is True. " + "It specifies the factor for time warping in SpecAugment. " + "Larger values mean more warping. " + "A value less than 1 means to disable time warp.", + ) + + group.add_argument( + "--enable-musan", + type=str2bool, + default=True, + help="When enabled, select noise from MUSAN and mix it" + "with training dataset. ", + ) + + group.add_argument( + "--input-strategy", + type=str, + default="PrecomputedFeatures", + help="AudioSamples or PrecomputedFeatures", + ) + + def train_dataloaders( + self, + cuts_train: CutSet, + sampler_state_dict: Optional[Dict[str, Any]] = None, + ) -> DataLoader: + """ + Args: + cuts_train: + CutSet for training. + sampler_state_dict: + The state dict for the training sampler. + """ + transforms = [] + if self.args.enable_musan: + logging.info("Enable MUSAN") + logging.info("About to get Musan cuts") + cuts_musan = load_manifest(self.args.manifest_dir / "musan_cuts.jsonl.gz") + transforms.append( + CutMix(cuts=cuts_musan, prob=0.5, snr=(10, 20), preserve_id=True) + ) + else: + logging.info("Disable MUSAN") + + if self.args.concatenate_cuts: + logging.info( + f"Using cut concatenation with duration factor " + f"{self.args.duration_factor} and gap {self.args.gap}." + ) + # Cut concatenation should be the first transform in the list, + # so that if we e.g. mix noise in, it will fill the gaps between + # different utterances. + transforms = [ + CutConcatenate( + duration_factor=self.args.duration_factor, gap=self.args.gap + ) + ] + transforms + + input_transforms = [] + if self.args.enable_spec_aug: + logging.info("Enable SpecAugment") + logging.info(f"Time warp factor: {self.args.spec_aug_time_warp_factor}") + # Set the value of num_frame_masks according to Lhotse's version. + # In different Lhotse's versions, the default of num_frame_masks is + # different. + num_frame_masks = 10 + num_frame_masks_parameter = inspect.signature( + SpecAugment.__init__ + ).parameters["num_frame_masks"] + if num_frame_masks_parameter.default == 1: + num_frame_masks = 2 + logging.info(f"Num frame mask: {num_frame_masks}") + input_transforms.append( + SpecAugment( + time_warp_factor=self.args.spec_aug_time_warp_factor, + num_frame_masks=num_frame_masks, + features_mask_size=27, + num_feature_masks=2, + frames_mask_size=100, + ) + ) + else: + logging.info("Disable SpecAugment") + + logging.info("About to create train dataset") + train = K2SpeechRecognitionDataset( + input_strategy=eval(self.args.input_strategy)(), + cut_transforms=transforms, + input_transforms=input_transforms, + return_cuts=self.args.return_cuts, + ) + + if self.args.on_the_fly_feats: + # NOTE: the PerturbSpeed transform should be added only if we + # remove it from data prep stage. + # Add on-the-fly speed perturbation; since originally it would + # have increased epoch size by 3, we will apply prob 2/3 and use + # 3x more epochs. + # Speed perturbation probably should come first before + # concatenation, but in principle the transforms order doesn't have + # to be strict (e.g. could be randomized) + # transforms = [PerturbSpeed(factors=[0.9, 1.1], p=2/3)] + transforms # noqa + # Drop feats to be on the safe side. + train = K2SpeechRecognitionDataset( + cut_transforms=transforms, + input_strategy=OnTheFlyFeatures(Fbank(FbankConfig(num_mel_bins=80))), + input_transforms=input_transforms, + return_cuts=self.args.return_cuts, + ) + + if self.args.bucketing_sampler: + logging.info("Using DynamicBucketingSampler.") + train_sampler = DynamicBucketingSampler( + cuts_train, + max_duration=self.args.max_duration, + shuffle=self.args.shuffle, + num_buckets=self.args.num_buckets, + drop_last=self.args.drop_last, + ) + else: + logging.info("Using SimpleCutSampler.") + train_sampler = SimpleCutSampler( + cuts_train, + max_duration=self.args.max_duration, + shuffle=self.args.shuffle, + ) + logging.info("About to create train dataloader") + + if sampler_state_dict is not None: + logging.info("Loading sampler state dict") + train_sampler.load_state_dict(sampler_state_dict) + + # 'seed' is derived from the current random state, which will have + # previously been set in the main process. + seed = torch.randint(0, 100000, ()).item() + worker_init_fn = _SeedWorkers(seed) + + train_dl = DataLoader( + train, + sampler=train_sampler, + batch_size=None, + num_workers=self.args.num_workers, + persistent_workers=True, + worker_init_fn=worker_init_fn, + ) + + return train_dl + + def valid_dataloaders(self, cuts_valid: CutSet) -> DataLoader: + transforms = [] + if self.args.concatenate_cuts: + transforms = [ + CutConcatenate( + duration_factor=self.args.duration_factor, gap=self.args.gap + ) + ] + transforms + + logging.info("About to create dev dataset") + if self.args.on_the_fly_feats: + validate = K2SpeechRecognitionDataset( + cut_transforms=transforms, + input_strategy=OnTheFlyFeatures(Fbank(FbankConfig(num_mel_bins=80))), + return_cuts=self.args.return_cuts, + ) + else: + validate = K2SpeechRecognitionDataset( + cut_transforms=transforms, + return_cuts=self.args.return_cuts, + ) + valid_sampler = DynamicBucketingSampler( + cuts_valid, + max_duration=self.args.max_duration, + shuffle=False, + ) + logging.info("About to create dev dataloader") + valid_dl = DataLoader( + validate, + sampler=valid_sampler, + batch_size=None, + num_workers=2, + persistent_workers=False, + ) + + return valid_dl + + def test_dataloaders(self, cuts: CutSet) -> DataLoader: + logging.debug("About to create test dataset") + test = K2SpeechRecognitionDataset( + input_strategy=OnTheFlyFeatures(Fbank(FbankConfig(num_mel_bins=80))) + if self.args.on_the_fly_feats + else eval(self.args.input_strategy)(), + return_cuts=self.args.return_cuts, + ) + sampler = DynamicBucketingSampler( + cuts, + max_duration=self.args.max_duration, + shuffle=False, + ) + logging.debug("About to create test dataloader") + test_dl = DataLoader( + test, + batch_size=None, + sampler=sampler, + num_workers=self.args.num_workers, + ) + return test_dl diff --git a/egs/multi_zh-hans/ASR/zipformer/beam_search.py b/egs/multi_zh-hans/ASR/zipformer/beam_search.py new file mode 120000 index 000000000..8e2c0a65c --- /dev/null +++ b/egs/multi_zh-hans/ASR/zipformer/beam_search.py @@ -0,0 +1 @@ +../../../librispeech/ASR/zipformer/beam_search.py \ No newline at end of file diff --git a/egs/multi_zh-hans/ASR/zipformer/decode.py b/egs/multi_zh-hans/ASR/zipformer/decode.py new file mode 100755 index 000000000..f501c3c30 --- /dev/null +++ b/egs/multi_zh-hans/ASR/zipformer/decode.py @@ -0,0 +1,828 @@ +#!/usr/bin/env python3 +# +# Copyright 2021-2023 Xiaomi Corporation (Author: Fangjun Kuang, +# Zengwei Yao) +# +# See ../../../../LICENSE for clarification regarding multiple authors +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +""" +Usage: +(1) greedy search +./zipformer/decode.py \ + --epoch 28 \ + --avg 15 \ + --exp-dir ./zipformer/exp \ + --max-duration 600 \ + --decoding-method greedy_search + +(2) beam search (not recommended) +./zipformer/decode.py \ + --epoch 28 \ + --avg 15 \ + --exp-dir ./zipformer/exp \ + --max-duration 600 \ + --decoding-method beam_search \ + --beam-size 4 + +(3) modified beam search +./zipformer/decode.py \ + --epoch 28 \ + --avg 15 \ + --exp-dir ./zipformer/exp \ + --max-duration 600 \ + --decoding-method modified_beam_search \ + --beam-size 4 + +(4) fast beam search (one best) +./zipformer/decode.py \ + --epoch 28 \ + --avg 15 \ + --exp-dir ./zipformer/exp \ + --max-duration 600 \ + --decoding-method fast_beam_search \ + --beam 20.0 \ + --max-contexts 8 \ + --max-states 64 + +(5) fast beam search (nbest) +./zipformer/decode.py \ + --epoch 28 \ + --avg 15 \ + --exp-dir ./zipformer/exp \ + --max-duration 600 \ + --decoding-method fast_beam_search_nbest \ + --beam 20.0 \ + --max-contexts 8 \ + --max-states 64 \ + --num-paths 200 \ + --nbest-scale 0.5 + +(6) fast beam search (nbest oracle WER) +./zipformer/decode.py \ + --epoch 28 \ + --avg 15 \ + --exp-dir ./zipformer/exp \ + --max-duration 600 \ + --decoding-method fast_beam_search_nbest_oracle \ + --beam 20.0 \ + --max-contexts 8 \ + --max-states 64 \ + --num-paths 200 \ + --nbest-scale 0.5 + +(7) fast beam search (with LG) +./zipformer/decode.py \ + --epoch 28 \ + --avg 15 \ + --exp-dir ./zipformer/exp \ + --max-duration 600 \ + --decoding-method fast_beam_search_nbest_LG \ + --beam 20.0 \ + --max-contexts 8 \ + --max-states 64 +""" + + +import argparse +import logging +import math +from collections import defaultdict +from pathlib import Path +from typing import Dict, List, Optional, Tuple + +import k2 +import sentencepiece as spm +import torch +import torch.nn as nn +from asr_datamodule import AsrDataModule +from beam_search import ( + beam_search, + fast_beam_search_nbest, + fast_beam_search_nbest_LG, + fast_beam_search_nbest_oracle, + fast_beam_search_one_best, + greedy_search, + greedy_search_batch, + modified_beam_search, +) +from lhotse.cut import Cut +from multi_dataset import MultiDataset +from train import add_model_arguments, get_model, get_params + +from icefall.checkpoint import ( + average_checkpoints, + average_checkpoints_with_averaged_model, + find_checkpoints, + load_checkpoint, +) +from icefall.lexicon import Lexicon +from icefall.utils import ( + AttributeDict, + make_pad_mask, + setup_logger, + store_transcripts, + str2bool, + write_error_stats, +) + +LOG_EPS = math.log(1e-10) + + +def get_parser(): + parser = argparse.ArgumentParser( + formatter_class=argparse.ArgumentDefaultsHelpFormatter + ) + + parser.add_argument( + "--epoch", + type=int, + default=30, + help="""It specifies the checkpoint to use for decoding. + Note: Epoch counts from 1. + You can specify --avg to use more checkpoints for model averaging.""", + ) + + parser.add_argument( + "--iter", + type=int, + default=0, + help="""If positive, --epoch is ignored and it + will use the checkpoint exp_dir/checkpoint-iter.pt. + You can specify --avg to use more checkpoints for model averaging. + """, + ) + + parser.add_argument( + "--avg", + type=int, + default=15, + help="Number of checkpoints to average. Automatically select " + "consecutive checkpoints before the checkpoint specified by " + "'--epoch' and '--iter'", + ) + + parser.add_argument( + "--use-averaged-model", + type=str2bool, + default=True, + help="Whether to load averaged model. Currently it only supports " + "using --epoch. If True, it would decode with the averaged model " + "over the epoch range from `epoch-avg` (excluded) to `epoch`." + "Actually only the models with epoch number of `epoch-avg` and " + "`epoch` are loaded for averaging. ", + ) + + parser.add_argument( + "--exp-dir", + type=str, + default="zipformer/exp", + help="The experiment dir", + ) + + parser.add_argument( + "--bpe-model", + type=str, + default="data/lang_bpe_2000/bpe.model", + help="Path to the BPE model", + ) + + parser.add_argument( + "--lang-dir", + type=Path, + default="data/lang_bpe_2000", + help="The lang dir containing word table and LG graph", + ) + + parser.add_argument( + "--decoding-method", + type=str, + default="greedy_search", + help="""Possible values are: + - greedy_search + - beam_search + - modified_beam_search + - fast_beam_search + - fast_beam_search_nbest + - fast_beam_search_nbest_oracle + - fast_beam_search_nbest_LG + If you use fast_beam_search_nbest_LG, you have to specify + `--lang-dir`, which should contain `LG.pt`. + """, + ) + + parser.add_argument( + "--beam-size", + type=int, + default=4, + help="""An integer indicating how many candidates we will keep for each + frame. Used only when --decoding-method is beam_search or + modified_beam_search.""", + ) + + parser.add_argument( + "--beam", + type=float, + default=20.0, + help="""A floating point value to calculate the cutoff score during beam + search (i.e., `cutoff = max-score - beam`), which is the same as the + `beam` in Kaldi. + Used only when --decoding-method is fast_beam_search, + fast_beam_search_nbest, fast_beam_search_nbest_LG, + and fast_beam_search_nbest_oracle + """, + ) + + parser.add_argument( + "--ngram-lm-scale", + type=float, + default=0.01, + help=""" + Used only when --decoding_method is fast_beam_search_nbest_LG. + It specifies the scale for n-gram LM scores. + """, + ) + + parser.add_argument( + "--max-contexts", + type=int, + default=8, + help="""Used only when --decoding-method is + fast_beam_search, fast_beam_search_nbest, fast_beam_search_nbest_LG, + and fast_beam_search_nbest_oracle""", + ) + + parser.add_argument( + "--max-states", + type=int, + default=64, + help="""Used only when --decoding-method is + fast_beam_search, fast_beam_search_nbest, fast_beam_search_nbest_LG, + and fast_beam_search_nbest_oracle""", + ) + + parser.add_argument( + "--context-size", + type=int, + default=2, + help="The context size in the decoder. 1 means bigram; " "2 means tri-gram", + ) + parser.add_argument( + "--max-sym-per-frame", + type=int, + default=1, + help="""Maximum number of symbols per frame. + Used only when --decoding_method is greedy_search""", + ) + + parser.add_argument( + "--num-paths", + type=int, + default=200, + help="""Number of paths for nbest decoding. + Used only when the decoding method is fast_beam_search_nbest, + fast_beam_search_nbest_LG, and fast_beam_search_nbest_oracle""", + ) + + parser.add_argument( + "--nbest-scale", + type=float, + default=0.5, + help="""Scale applied to lattice scores when computing nbest paths. + Used only when the decoding method is fast_beam_search_nbest, + fast_beam_search_nbest_LG, and fast_beam_search_nbest_oracle""", + ) + + add_model_arguments(parser) + + return parser + + +def decode_one_batch( + params: AttributeDict, + model: nn.Module, + sp: spm.SentencePieceProcessor, + batch: dict, + word_table: Optional[k2.SymbolTable] = None, + decoding_graph: Optional[k2.Fsa] = None, +) -> Dict[str, List[List[str]]]: + """Decode one batch and return the result in a dict. The dict has the + following format: + + - key: It indicates the setting used for decoding. For example, + if greedy_search is used, it would be "greedy_search" + If beam search with a beam size of 7 is used, it would be + "beam_7" + - value: It contains the decoding result. `len(value)` equals to + batch size. `value[i]` is the decoding result for the i-th + utterance in the given batch. + Args: + params: + It's the return value of :func:`get_params`. + model: + The neural model. + sp: + The BPE model. + batch: + It is the return value from iterating + `lhotse.dataset.K2SpeechRecognitionDataset`. See its documentation + for the format of the `batch`. + word_table: + The word symbol table. + decoding_graph: + The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used + only when --decoding_method is fast_beam_search, fast_beam_search_nbest, + fast_beam_search_nbest_oracle, and fast_beam_search_nbest_LG. + Returns: + Return the decoding result. See above description for the format of + the returned dict. + """ + device = next(model.parameters()).device + feature = batch["inputs"] + assert feature.ndim == 3 + + feature = feature.to(device) + # at entry, feature is (N, T, C) + + supervisions = batch["supervisions"] + feature_lens = supervisions["num_frames"].to(device) + + if params.causal: + # this seems to cause insertions at the end of the utterance if used with zipformer. + pad_len = 30 + feature_lens += pad_len + feature = torch.nn.functional.pad( + feature, + pad=(0, 0, 0, pad_len), + value=LOG_EPS, + ) + + encoder_out, encoder_out_lens = model.forward_encoder(feature, feature_lens) + + hyps = [] + + if params.decoding_method == "fast_beam_search": + hyp_tokens = fast_beam_search_one_best( + model=model, + decoding_graph=decoding_graph, + encoder_out=encoder_out, + encoder_out_lens=encoder_out_lens, + beam=params.beam, + max_contexts=params.max_contexts, + max_states=params.max_states, + ) + for hyp in sp.decode(hyp_tokens): + hyps.append(hyp.split()) + elif params.decoding_method == "fast_beam_search_nbest_LG": + hyp_tokens = fast_beam_search_nbest_LG( + model=model, + decoding_graph=decoding_graph, + encoder_out=encoder_out, + encoder_out_lens=encoder_out_lens, + beam=params.beam, + max_contexts=params.max_contexts, + max_states=params.max_states, + num_paths=params.num_paths, + nbest_scale=params.nbest_scale, + ) + for hyp in hyp_tokens: + hyps.append([word_table[i] for i in hyp]) + elif params.decoding_method == "fast_beam_search_nbest": + hyp_tokens = fast_beam_search_nbest( + model=model, + decoding_graph=decoding_graph, + encoder_out=encoder_out, + encoder_out_lens=encoder_out_lens, + beam=params.beam, + max_contexts=params.max_contexts, + max_states=params.max_states, + num_paths=params.num_paths, + nbest_scale=params.nbest_scale, + ) + for hyp in sp.decode(hyp_tokens): + hyps.append(hyp.split()) + elif params.decoding_method == "fast_beam_search_nbest_oracle": + hyp_tokens = fast_beam_search_nbest_oracle( + model=model, + decoding_graph=decoding_graph, + encoder_out=encoder_out, + encoder_out_lens=encoder_out_lens, + beam=params.beam, + max_contexts=params.max_contexts, + max_states=params.max_states, + num_paths=params.num_paths, + ref_texts=sp.encode(supervisions["text"]), + nbest_scale=params.nbest_scale, + ) + for hyp in sp.decode(hyp_tokens): + hyps.append(hyp.split()) + elif params.decoding_method == "greedy_search" and params.max_sym_per_frame == 1: + hyp_tokens = greedy_search_batch( + model=model, + encoder_out=encoder_out, + encoder_out_lens=encoder_out_lens, + ) + for hyp in sp.decode(hyp_tokens): + hyps.append(hyp.split()) + elif params.decoding_method == "modified_beam_search": + hyp_tokens = modified_beam_search( + model=model, + encoder_out=encoder_out, + encoder_out_lens=encoder_out_lens, + beam=params.beam_size, + ) + for hyp in sp.decode(hyp_tokens): + hyps.append(hyp.split()) + else: + batch_size = encoder_out.size(0) + + for i in range(batch_size): + # fmt: off + encoder_out_i = encoder_out[i:i+1, :encoder_out_lens[i]] + # fmt: on + if params.decoding_method == "greedy_search": + hyp = greedy_search( + model=model, + encoder_out=encoder_out_i, + max_sym_per_frame=params.max_sym_per_frame, + ) + elif params.decoding_method == "beam_search": + hyp = beam_search( + model=model, + encoder_out=encoder_out_i, + beam=params.beam_size, + ) + else: + raise ValueError( + f"Unsupported decoding method: {params.decoding_method}" + ) + hyps.append(sp.decode(hyp).split()) + + if params.decoding_method == "greedy_search": + return {"greedy_search": hyps} + elif "fast_beam_search" in params.decoding_method: + key = f"beam_{params.beam}_" + key += f"max_contexts_{params.max_contexts}_" + key += f"max_states_{params.max_states}" + if "nbest" in params.decoding_method: + key += f"_num_paths_{params.num_paths}_" + key += f"nbest_scale_{params.nbest_scale}" + if "LG" in params.decoding_method: + key += f"_ngram_lm_scale_{params.ngram_lm_scale}" + + return {key: hyps} + else: + return {f"beam_size_{params.beam_size}": hyps} + + +def decode_dataset( + dl: torch.utils.data.DataLoader, + params: AttributeDict, + model: nn.Module, + sp: spm.SentencePieceProcessor, + word_table: Optional[k2.SymbolTable] = None, + decoding_graph: Optional[k2.Fsa] = None, +) -> Dict[str, List[Tuple[str, List[str], List[str]]]]: + """Decode dataset. + + Args: + dl: + PyTorch's dataloader containing the dataset to decode. + params: + It is returned by :func:`get_params`. + model: + The neural model. + sp: + The BPE model. + word_table: + The word symbol table. + decoding_graph: + The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used + only when --decoding_method is fast_beam_search, fast_beam_search_nbest, + fast_beam_search_nbest_oracle, and fast_beam_search_nbest_LG. + Returns: + Return a dict, whose key may be "greedy_search" if greedy search + is used, or it may be "beam_7" if beam size of 7 is used. + Its value is a list of tuples. Each tuple contains two elements: + The first is the reference transcript, and the second is the + predicted result. + """ + num_cuts = 0 + + try: + num_batches = len(dl) + except TypeError: + num_batches = "?" + + if params.decoding_method == "greedy_search": + log_interval = 50 + else: + log_interval = 20 + + results = defaultdict(list) + for batch_idx, batch in enumerate(dl): + texts = batch["supervisions"]["text"] + texts = [list(str(text).replace(" ", "")) for text in texts] + cut_ids = [cut.id for cut in batch["supervisions"]["cut"]] + + hyps_dict = decode_one_batch( + params=params, + model=model, + sp=sp, + decoding_graph=decoding_graph, + word_table=word_table, + batch=batch, + ) + + for name, hyps in hyps_dict.items(): + this_batch = [] + assert len(hyps) == len(texts) + for cut_id, hyp_words, ref_text in zip(cut_ids, hyps, texts): + hyp_text = "".join(hyp_words) + this_batch.append((cut_id, ref_text, hyp_text)) + + results[name].extend(this_batch) + + num_cuts += len(texts) + + if batch_idx % log_interval == 0: + batch_str = f"{batch_idx}/{num_batches}" + + logging.info(f"batch {batch_str}, cuts processed until now is {num_cuts}") + return results + + +def save_results( + params: AttributeDict, + test_set_name: str, + results_dict: Dict[str, List[Tuple[str, List[str], List[str]]]], +): + test_set_wers = dict() + for key, results in results_dict.items(): + recog_path = ( + params.res_dir / f"recogs-{test_set_name}-{key}-{params.suffix}.txt" + ) + results = sorted(results) + store_transcripts(filename=recog_path, texts=results) + logging.info(f"The transcripts are stored in {recog_path}") + + # The following prints out WERs, per-word error statistics and aligned + # ref/hyp pairs. + errs_filename = ( + params.res_dir / f"errs-{test_set_name}-{key}-{params.suffix}.txt" + ) + with open(errs_filename, "w") as f: + wer = write_error_stats( + f, f"{test_set_name}-{key}", results, enable_log=True + ) + test_set_wers[key] = wer + + logging.info("Wrote detailed error stats to {}".format(errs_filename)) + + test_set_wers = sorted(test_set_wers.items(), key=lambda x: x[1]) + errs_info = ( + params.res_dir / f"wer-summary-{test_set_name}-{key}-{params.suffix}.txt" + ) + with open(errs_info, "w") as f: + print("settings\tWER", file=f) + for key, val in test_set_wers: + print("{}\t{}".format(key, val), file=f) + + s = "\nFor {}, WER of different settings are:\n".format(test_set_name) + note = "\tbest for {}".format(test_set_name) + for key, val in test_set_wers: + s += "{}\t{}{}\n".format(key, val, note) + note = "" + logging.info(s) + + +@torch.no_grad() +def main(): + parser = get_parser() + AsrDataModule.add_arguments(parser) + args = parser.parse_args() + args.exp_dir = Path(args.exp_dir) + + params = get_params() + params.update(vars(args)) + + assert params.decoding_method in ( + "greedy_search", + "beam_search", + "fast_beam_search", + "fast_beam_search_nbest", + "fast_beam_search_nbest_LG", + "fast_beam_search_nbest_oracle", + "modified_beam_search", + ) + params.res_dir = params.exp_dir / params.decoding_method + + if params.iter > 0: + params.suffix = f"iter-{params.iter}-avg-{params.avg}" + else: + params.suffix = f"epoch-{params.epoch}-avg-{params.avg}" + + if params.causal: + assert ( + "," not in params.chunk_size + ), "chunk_size should be one value in decoding." + assert ( + "," not in params.left_context_frames + ), "left_context_frames should be one value in decoding." + params.suffix += f"-chunk-{params.chunk_size}" + params.suffix += f"-left-context-{params.left_context_frames}" + + if "fast_beam_search" in params.decoding_method: + params.suffix += f"-beam-{params.beam}" + params.suffix += f"-max-contexts-{params.max_contexts}" + params.suffix += f"-max-states-{params.max_states}" + if "nbest" in params.decoding_method: + params.suffix += f"-nbest-scale-{params.nbest_scale}" + params.suffix += f"-num-paths-{params.num_paths}" + if "LG" in params.decoding_method: + params.suffix += f"-ngram-lm-scale-{params.ngram_lm_scale}" + elif "beam_search" in params.decoding_method: + params.suffix += f"-{params.decoding_method}-beam-size-{params.beam_size}" + else: + params.suffix += f"-context-{params.context_size}" + params.suffix += f"-max-sym-per-frame-{params.max_sym_per_frame}" + + if params.use_averaged_model: + params.suffix += "-use-averaged-model" + + setup_logger(f"{params.res_dir}/log-decode-{params.suffix}") + logging.info("Decoding started") + + device = torch.device("cpu") + if torch.cuda.is_available(): + device = torch.device("cuda", 0) + + logging.info(f"Device: {device}") + + sp = spm.SentencePieceProcessor() + sp.load(params.bpe_model) + + # and are defined in local/train_bpe_model.py + params.blank_id = sp.piece_to_id("") + params.unk_id = sp.piece_to_id("") + params.vocab_size = sp.get_piece_size() + + logging.info(params) + + logging.info("About to create model") + model = get_model(params) + + if not params.use_averaged_model: + if params.iter > 0: + filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[ + : params.avg + ] + if len(filenames) == 0: + raise ValueError( + f"No checkpoints found for" + f" --iter {params.iter}, --avg {params.avg}" + ) + elif len(filenames) < params.avg: + raise ValueError( + f"Not enough checkpoints ({len(filenames)}) found for" + f" --iter {params.iter}, --avg {params.avg}" + ) + logging.info(f"averaging {filenames}") + model.to(device) + model.load_state_dict(average_checkpoints(filenames, device=device)) + elif params.avg == 1: + load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model) + else: + start = params.epoch - params.avg + 1 + filenames = [] + for i in range(start, params.epoch + 1): + if i >= 1: + filenames.append(f"{params.exp_dir}/epoch-{i}.pt") + logging.info(f"averaging {filenames}") + model.to(device) + model.load_state_dict(average_checkpoints(filenames, device=device)) + else: + if params.iter > 0: + filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[ + : params.avg + 1 + ] + if len(filenames) == 0: + raise ValueError( + f"No checkpoints found for" + f" --iter {params.iter}, --avg {params.avg}" + ) + elif len(filenames) < params.avg + 1: + raise ValueError( + f"Not enough checkpoints ({len(filenames)}) found for" + f" --iter {params.iter}, --avg {params.avg}" + ) + filename_start = filenames[-1] + filename_end = filenames[0] + logging.info( + "Calculating the averaged model over iteration checkpoints" + f" from {filename_start} (excluded) to {filename_end}" + ) + model.to(device) + model.load_state_dict( + average_checkpoints_with_averaged_model( + filename_start=filename_start, + filename_end=filename_end, + device=device, + ) + ) + else: + assert params.avg > 0, params.avg + start = params.epoch - params.avg + assert start >= 1, start + filename_start = f"{params.exp_dir}/epoch-{start}.pt" + filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt" + logging.info( + f"Calculating the averaged model over epoch range from " + f"{start} (excluded) to {params.epoch}" + ) + model.to(device) + model.load_state_dict( + average_checkpoints_with_averaged_model( + filename_start=filename_start, + filename_end=filename_end, + device=device, + ) + ) + + model.to(device) + model.eval() + + if "fast_beam_search" in params.decoding_method: + if params.decoding_method == "fast_beam_search_nbest_LG": + lexicon = Lexicon(params.lang_dir) + word_table = lexicon.word_table + lg_filename = params.lang_dir / "LG.pt" + logging.info(f"Loading {lg_filename}") + decoding_graph = k2.Fsa.from_dict( + torch.load(lg_filename, map_location=device) + ) + decoding_graph.scores *= params.ngram_lm_scale + else: + word_table = None + decoding_graph = k2.trivial_graph(params.vocab_size - 1, device=device) + else: + decoding_graph = None + word_table = None + + num_param = sum([p.numel() for p in model.parameters()]) + logging.info(f"Number of model parameters: {num_param}") + + # we need cut ids to display recognition results. + args.return_cuts = True + data_module = AsrDataModule(args) + multi_dataset = MultiDataset(args.manifest_dir) + + def remove_short_utt(c: Cut): + T = ((c.num_frames - 7) // 2 + 1) // 2 + if T <= 0: + logging.warning( + f"Excluding cut with ID: {c.id} from decoding, num_frames: {c.num_frames}" + ) + return T > 0 + + test_sets_cuts = multi_dataset.test_cuts() + + test_sets = test_sets_cuts.keys() + test_dl = [ + data_module.test_dataloaders(test_sets_cuts[cuts_name].filter(remove_short_utt)) + for cuts_name in test_sets + ] + + for test_set, test_dl in zip(test_sets, test_dl): + logging.info(f"Start decoding test set: {test_set}") + + results_dict = decode_dataset( + dl=test_dl, + params=params, + model=model, + sp=sp, + word_table=word_table, + decoding_graph=decoding_graph, + ) + + save_results( + params=params, + test_set_name=test_set, + results_dict=results_dict, + ) + + logging.info("Done!") + + +if __name__ == "__main__": + main() diff --git a/egs/multi_zh-hans/ASR/zipformer/decoder.py b/egs/multi_zh-hans/ASR/zipformer/decoder.py new file mode 120000 index 000000000..5a8018680 --- /dev/null +++ b/egs/multi_zh-hans/ASR/zipformer/decoder.py @@ -0,0 +1 @@ +../../../librispeech/ASR/zipformer/decoder.py \ No newline at end of file diff --git a/egs/multi_zh-hans/ASR/zipformer/encoder_interface.py b/egs/multi_zh-hans/ASR/zipformer/encoder_interface.py new file mode 120000 index 000000000..c2eaca671 --- /dev/null +++ b/egs/multi_zh-hans/ASR/zipformer/encoder_interface.py @@ -0,0 +1 @@ +../../../librispeech/ASR/zipformer/encoder_interface.py \ No newline at end of file diff --git a/egs/multi_zh-hans/ASR/zipformer/export-onnx-streaming.py b/egs/multi_zh-hans/ASR/zipformer/export-onnx-streaming.py new file mode 120000 index 000000000..2962eb784 --- /dev/null +++ b/egs/multi_zh-hans/ASR/zipformer/export-onnx-streaming.py @@ -0,0 +1 @@ +../../../librispeech/ASR/zipformer/export-onnx-streaming.py \ No newline at end of file diff --git a/egs/multi_zh-hans/ASR/zipformer/export-onnx.py b/egs/multi_zh-hans/ASR/zipformer/export-onnx.py new file mode 120000 index 000000000..70a15683c --- /dev/null +++ b/egs/multi_zh-hans/ASR/zipformer/export-onnx.py @@ -0,0 +1 @@ +../../../librispeech/ASR/zipformer/export-onnx.py \ No newline at end of file diff --git a/egs/multi_zh-hans/ASR/zipformer/export.py b/egs/multi_zh-hans/ASR/zipformer/export.py new file mode 100755 index 000000000..723288191 --- /dev/null +++ b/egs/multi_zh-hans/ASR/zipformer/export.py @@ -0,0 +1,541 @@ +#!/usr/bin/env python3 +# +# Copyright 2021-2023 Xiaomi Corporation (Author: Fangjun Kuang, +# Zengwei Yao, +# Wei Kang) +# +# See ../../../../LICENSE for clarification regarding multiple authors +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +# This script converts several saved checkpoints +# to a single one using model averaging. +""" + +Usage: + +Note: This is a example for librispeech dataset, if you are using different +dataset, you should change the argument values according to your dataset. + +(1) Export to torchscript model using torch.jit.script() + +- For non-streaming model: + +./zipformer/export.py \ + --exp-dir ./zipformer/exp \ + --tokens data/lang_bpe_2000/tokens.txt \ + --epoch 20 \ + --avg 1 \ + --jit 1 + +It will generate a file `jit_script.pt` in the given `exp_dir`. You can later +load it by `torch.jit.load("jit_script.pt")`. + +Check ./jit_pretrained.py for its usage. + +Check https://github.com/k2-fsa/sherpa +for how to use the exported models outside of icefall. + +- For streaming model: + +./zipformer/export.py \ + --exp-dir ./zipformer/exp \ + --causal 1 \ + --chunk-size 16 \ + --left-context-frames 128 \ + --tokens data/lang_bpe_2000/tokens.txt \ + --epoch 20 \ + --avg 1 \ + --jit 1 + +It will generate a file `jit_script_chunk_16_left_128.pt` in the given `exp_dir`. +You can later load it by `torch.jit.load("jit_script_chunk_16_left_128.pt")`. + +Check ./jit_pretrained_streaming.py for its usage. + +Check https://github.com/k2-fsa/sherpa +for how to use the exported models outside of icefall. + +(2) Export `model.state_dict()` + +- For non-streaming model: + +./zipformer/export.py \ + --exp-dir ./zipformer/exp \ + --tokens data/lang_bpe_2000/tokens.txt \ + --epoch 20 \ + --avg 1 + +- For streaming model: + +./zipformer/export.py \ + --exp-dir ./zipformer/exp \ + --causal 1 \ + --tokens data/lang_bpe_2000/tokens.txt \ + --epoch 20 \ + --avg 1 + +It will generate a file `pretrained.pt` in the given `exp_dir`. You can later +load it by `icefall.checkpoint.load_checkpoint()`. + +- For non-streaming model: + +To use the generated file with `zipformer/decode.py`, +you can do: + + cd /path/to/exp_dir + ln -s pretrained.pt epoch-9999.pt + + cd /path/to/egs/librispeech/ASR + ./zipformer/decode.py \ + --exp-dir ./zipformer/exp \ + --epoch 9999 \ + --avg 1 \ + --max-duration 600 \ + --decoding-method greedy_search \ + --bpe-model data/lang_bpe_2000/bpe.model + +- For streaming model: + +To use the generated file with `zipformer/decode.py` and `zipformer/streaming_decode.py`, you can do: + + cd /path/to/exp_dir + ln -s pretrained.pt epoch-9999.pt + + cd /path/to/egs/librispeech/ASR + + # simulated streaming decoding + ./zipformer/decode.py \ + --exp-dir ./zipformer/exp \ + --epoch 9999 \ + --avg 1 \ + --max-duration 600 \ + --causal 1 \ + --chunk-size 16 \ + --left-context-frames 128 \ + --decoding-method greedy_search \ + --bpe-model data/lang_bpe_2000/bpe.model + + # chunk-wise streaming decoding + ./zipformer/streaming_decode.py \ + --exp-dir ./zipformer/exp \ + --epoch 9999 \ + --avg 1 \ + --max-duration 600 \ + --causal 1 \ + --chunk-size 16 \ + --left-context-frames 128 \ + --decoding-method greedy_search \ + --bpe-model data/lang_bpe_2000/bpe.model + +Check ./pretrained.py for its usage. + +Note: If you don't want to train a model from scratch, we have +provided one for you. You can get it at + +- non-streaming model: +https://huggingface.co/zrjin/icefall-asr-multi-zh-hans-zipformer-2023-9-2/ + +with the following commands: + + sudo apt-get install git-lfs + git lfs install + git clone https://huggingface.co/zrjin/icefall-asr-multi-zh-hans-zipformer-2023-9-2/ + # You will find the pre-trained models in exp dir +""" + +import argparse +import logging +import re +from pathlib import Path +from typing import List, Tuple + +import k2 +import torch +from scaling_converter import convert_scaled_to_non_scaled +from torch import Tensor, nn +from train import add_model_arguments, get_model, get_params + +from icefall.checkpoint import ( + average_checkpoints, + average_checkpoints_with_averaged_model, + find_checkpoints, + load_checkpoint, +) +from icefall.utils import make_pad_mask, str2bool + + +def num_tokens( + token_table: k2.SymbolTable, disambig_pattern: str = re.compile(r"^#\d+$") +) -> int: + """Return the number of tokens excluding those from + disambiguation symbols. + + Caution: + 0 is not a token ID so it is excluded from the return value. + """ + symbols = token_table.symbols + ans = [] + for s in symbols: + if not disambig_pattern.match(s): + ans.append(token_table[s]) + num_tokens = len(ans) + if 0 in ans: + num_tokens -= 1 + return num_tokens + + +def get_parser(): + parser = argparse.ArgumentParser( + formatter_class=argparse.ArgumentDefaultsHelpFormatter + ) + + parser.add_argument( + "--epoch", + type=int, + default=20, + help="""It specifies the checkpoint to use for decoding. + Note: Epoch counts from 1. + You can specify --avg to use more checkpoints for model averaging.""", + ) + + parser.add_argument( + "--iter", + type=int, + default=0, + help="""If positive, --epoch is ignored and it + will use the checkpoint exp_dir/checkpoint-iter.pt. + You can specify --avg to use more checkpoints for model averaging. + """, + ) + + parser.add_argument( + "--avg", + type=int, + default=1, + help="Number of checkpoints to average. Automatically select " + "consecutive checkpoints before the checkpoint specified by " + "'--epoch' and '--iter'", + ) + + parser.add_argument( + "--use-averaged-model", + type=str2bool, + default=True, + help="Whether to load averaged model. Currently it only supports " + "using --epoch. If True, it would decode with the averaged model " + "over the epoch range from `epoch-avg` (excluded) to `epoch`." + "Actually only the models with epoch number of `epoch-avg` and " + "`epoch` are loaded for averaging. ", + ) + + parser.add_argument( + "--exp-dir", + type=str, + default="zipformer/exp", + help="""It specifies the directory where all training related + files, e.g., checkpoints, log, etc, are saved + """, + ) + + parser.add_argument( + "--tokens", + type=str, + default="data/lang_bpe_2000/tokens.txt", + help="Path to the tokens.txt", + ) + + parser.add_argument( + "--jit", + type=str2bool, + default=False, + help="""True to save a model after applying torch.jit.script. + It will generate a file named jit_script.pt. + Check ./jit_pretrained.py for how to use it. + """, + ) + + parser.add_argument( + "--context-size", + type=int, + default=2, + help="The context size in the decoder. 1 means bigram; 2 means tri-gram", + ) + + add_model_arguments(parser) + + return parser + + +class EncoderModel(nn.Module): + """A wrapper for encoder and encoder_embed""" + + def __init__(self, encoder: nn.Module, encoder_embed: nn.Module) -> None: + super().__init__() + self.encoder = encoder + self.encoder_embed = encoder_embed + + def forward( + self, features: Tensor, feature_lengths: Tensor + ) -> Tuple[Tensor, Tensor]: + """ + Args: + features: (N, T, C) + feature_lengths: (N,) + """ + x, x_lens = self.encoder_embed(features, feature_lengths) + + src_key_padding_mask = make_pad_mask(x_lens) + x = x.permute(1, 0, 2) # (N, T, C) -> (T, N, C) + + encoder_out, encoder_out_lens = self.encoder(x, x_lens, src_key_padding_mask) + encoder_out = encoder_out.permute(1, 0, 2) # (T, N, C) ->(N, T, C) + + return encoder_out, encoder_out_lens + + +class StreamingEncoderModel(nn.Module): + """A wrapper for encoder and encoder_embed""" + + def __init__(self, encoder: nn.Module, encoder_embed: nn.Module) -> None: + super().__init__() + assert len(encoder.chunk_size) == 1, encoder.chunk_size + assert len(encoder.left_context_frames) == 1, encoder.left_context_frames + self.chunk_size = encoder.chunk_size[0] + self.left_context_len = encoder.left_context_frames[0] + + # The encoder_embed subsample features (T - 7) // 2 + # The ConvNeXt module needs (7 - 1) // 2 = 3 frames of right padding after subsampling + self.pad_length = 7 + 2 * 3 + + self.encoder = encoder + self.encoder_embed = encoder_embed + + def forward( + self, features: Tensor, feature_lengths: Tensor, states: List[Tensor] + ) -> Tuple[Tensor, Tensor, List[Tensor]]: + """Streaming forward for encoder_embed and encoder. + + Args: + features: (N, T, C) + feature_lengths: (N,) + states: a list of Tensors + + Returns encoder outputs, output lengths, and updated states. + """ + chunk_size = self.chunk_size + left_context_len = self.left_context_len + + cached_embed_left_pad = states[-2] + x, x_lens, new_cached_embed_left_pad = self.encoder_embed.streaming_forward( + x=features, + x_lens=feature_lengths, + cached_left_pad=cached_embed_left_pad, + ) + assert x.size(1) == chunk_size, (x.size(1), chunk_size) + + src_key_padding_mask = make_pad_mask(x_lens) + + # processed_mask is used to mask out initial states + processed_mask = torch.arange(left_context_len, device=x.device).expand( + x.size(0), left_context_len + ) + processed_lens = states[-1] # (batch,) + # (batch, left_context_size) + processed_mask = (processed_lens.unsqueeze(1) <= processed_mask).flip(1) + # Update processed lengths + new_processed_lens = processed_lens + x_lens + + # (batch, left_context_size + chunk_size) + src_key_padding_mask = torch.cat([processed_mask, src_key_padding_mask], dim=1) + + x = x.permute(1, 0, 2) # (N, T, C) -> (T, N, C) + encoder_states = states[:-2] + + ( + encoder_out, + encoder_out_lens, + new_encoder_states, + ) = self.encoder.streaming_forward( + x=x, + x_lens=x_lens, + states=encoder_states, + src_key_padding_mask=src_key_padding_mask, + ) + encoder_out = encoder_out.permute(1, 0, 2) # (T, N, C) ->(N, T, C) + + new_states = new_encoder_states + [ + new_cached_embed_left_pad, + new_processed_lens, + ] + return encoder_out, encoder_out_lens, new_states + + @torch.jit.export + def get_init_states( + self, + batch_size: int = 1, + device: torch.device = torch.device("cpu"), + ) -> List[torch.Tensor]: + """ + Returns a list of cached tensors of all encoder layers. For layer-i, states[i*6:(i+1)*6] + is (cached_key, cached_nonlin_attn, cached_val1, cached_val2, cached_conv1, cached_conv2). + states[-2] is the cached left padding for ConvNeXt module, + of shape (batch_size, num_channels, left_pad, num_freqs) + states[-1] is processed_lens of shape (batch,), which records the number + of processed frames (at 50hz frame rate, after encoder_embed) for each sample in batch. + """ + states = self.encoder.get_init_states(batch_size, device) + + embed_states = self.encoder_embed.get_init_states(batch_size, device) + states.append(embed_states) + + processed_lens = torch.zeros(batch_size, dtype=torch.int32, device=device) + states.append(processed_lens) + + return states + + +@torch.no_grad() +def main(): + args = get_parser().parse_args() + args.exp_dir = Path(args.exp_dir) + + params = get_params() + params.update(vars(args)) + + device = torch.device("cpu") + # if torch.cuda.is_available(): + # device = torch.device("cuda", 0) + + logging.info(f"device: {device}") + + token_table = k2.SymbolTable.from_file(params.tokens) + params.blank_id = token_table[""] + params.vocab_size = num_tokens(token_table) + 1 + + logging.info(params) + + logging.info("About to create model") + model = get_model(params) + + if not params.use_averaged_model: + if params.iter > 0: + filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[ + : params.avg + ] + if len(filenames) == 0: + raise ValueError( + f"No checkpoints found for" + f" --iter {params.iter}, --avg {params.avg}" + ) + elif len(filenames) < params.avg: + raise ValueError( + f"Not enough checkpoints ({len(filenames)}) found for" + f" --iter {params.iter}, --avg {params.avg}" + ) + logging.info(f"averaging {filenames}") + model.load_state_dict(average_checkpoints(filenames, device=device)) + elif params.avg == 1: + load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model) + else: + start = params.epoch - params.avg + 1 + filenames = [] + for i in range(start, params.epoch + 1): + if i >= 1: + filenames.append(f"{params.exp_dir}/epoch-{i}.pt") + logging.info(f"averaging {filenames}") + model.load_state_dict(average_checkpoints(filenames, device=device)) + else: + if params.iter > 0: + filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[ + : params.avg + 1 + ] + if len(filenames) == 0: + raise ValueError( + f"No checkpoints found for" + f" --iter {params.iter}, --avg {params.avg}" + ) + elif len(filenames) < params.avg + 1: + raise ValueError( + f"Not enough checkpoints ({len(filenames)}) found for" + f" --iter {params.iter}, --avg {params.avg}" + ) + filename_start = filenames[-1] + filename_end = filenames[0] + logging.info( + "Calculating the averaged model over iteration checkpoints" + f" from {filename_start} (excluded) to {filename_end}" + ) + model.load_state_dict( + average_checkpoints_with_averaged_model( + filename_start=filename_start, + filename_end=filename_end, + device=device, + ) + ) + else: + assert params.avg > 0, params.avg + start = params.epoch - params.avg + assert start >= 1, start + filename_start = f"{params.exp_dir}/epoch-{start}.pt" + filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt" + logging.info( + f"Calculating the averaged model over epoch range from " + f"{start} (excluded) to {params.epoch}" + ) + model.load_state_dict( + average_checkpoints_with_averaged_model( + filename_start=filename_start, + filename_end=filename_end, + device=device, + ) + ) + + model.eval() + + if params.jit is True: + convert_scaled_to_non_scaled(model, inplace=True) + # We won't use the forward() method of the model in C++, so just ignore + # it here. + # Otherwise, one of its arguments is a ragged tensor and is not + # torch scriptabe. + model.__class__.forward = torch.jit.ignore(model.__class__.forward) + + # Wrap encoder and encoder_embed as a module + if params.causal: + model.encoder = StreamingEncoderModel(model.encoder, model.encoder_embed) + chunk_size = model.encoder.chunk_size + left_context_len = model.encoder.left_context_len + filename = f"jit_script_chunk_{chunk_size}_left_{left_context_len}.pt" + else: + model.encoder = EncoderModel(model.encoder, model.encoder_embed) + filename = "jit_script.pt" + + logging.info("Using torch.jit.script") + model = torch.jit.script(model) + model.save(str(params.exp_dir / filename)) + logging.info(f"Saved to {filename}") + else: + logging.info("Not using torchscript. Export model.state_dict()") + # Save it using a format so that it can be loaded + # by :func:`load_checkpoint` + filename = params.exp_dir / "pretrained.pt" + torch.save({"model": model.state_dict()}, str(filename)) + logging.info(f"Saved to {filename}") + + +if __name__ == "__main__": + formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s" + + logging.basicConfig(format=formatter, level=logging.INFO) + main() diff --git a/egs/multi_zh-hans/ASR/zipformer/generate_averaged_model.py b/egs/multi_zh-hans/ASR/zipformer/generate_averaged_model.py new file mode 100755 index 000000000..68111fad7 --- /dev/null +++ b/egs/multi_zh-hans/ASR/zipformer/generate_averaged_model.py @@ -0,0 +1,193 @@ +#!/usr/bin/env python3 +# +# Copyright 2021-2022 Xiaomi Corporation (Author: Yifan Yang) +# +# See ../../../../LICENSE for clarification regarding multiple authors +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +""" +Usage: +(1) use the checkpoint exp_dir/epoch-xxx.pt +./zipformer/generate_averaged_model.py \ + --epoch 28 \ + --avg 15 \ + --exp-dir ./zipformer/exp + +It will generate a file `epoch-28-avg-15.pt` in the given `exp_dir`. +You can later load it by `torch.load("epoch-28-avg-15.pt")`. + +(2) use the checkpoint exp_dir/checkpoint-iter.pt +./zipformer/generate_averaged_model.py \ + --iter 22000 \ + --avg 5 \ + --exp-dir ./zipformer/exp + +It will generate a file `iter-22000-avg-5.pt` in the given `exp_dir`. +You can later load it by `torch.load("iter-22000-avg-5.pt")`. +""" + + +import argparse +from pathlib import Path + +import k2 +import torch +from train import add_model_arguments, get_model, get_params + +from icefall.checkpoint import average_checkpoints_with_averaged_model, find_checkpoints + + +def get_parser(): + parser = argparse.ArgumentParser( + formatter_class=argparse.ArgumentDefaultsHelpFormatter + ) + + parser.add_argument( + "--epoch", + type=int, + default=30, + help="""It specifies the checkpoint to use for decoding. + Note: Epoch counts from 1. + You can specify --avg to use more checkpoints for model averaging.""", + ) + + parser.add_argument( + "--iter", + type=int, + default=0, + help="""If positive, --epoch is ignored and it + will use the checkpoint exp_dir/checkpoint-iter.pt. + You can specify --avg to use more checkpoints for model averaging. + """, + ) + + parser.add_argument( + "--avg", + type=int, + default=9, + help="Number of checkpoints to average. Automatically select " + "consecutive checkpoints before the checkpoint specified by " + "'--epoch' and '--iter'", + ) + + parser.add_argument( + "--exp-dir", + type=str, + default="zipformer/exp", + help="The experiment dir", + ) + + parser.add_argument( + "--tokens", + type=str, + default="data/lang_bpe_500/tokens.txt", + help="Path to the tokens.txt", + ) + + parser.add_argument( + "--context-size", + type=int, + default=2, + help="The context size in the decoder. 1 means bigram; 2 means tri-gram", + ) + + add_model_arguments(parser) + + return parser + + +@torch.no_grad() +def main(): + parser = get_parser() + args = parser.parse_args() + args.exp_dir = Path(args.exp_dir) + + params = get_params() + params.update(vars(args)) + + if params.iter > 0: + params.suffix = f"iter-{params.iter}-avg-{params.avg}" + else: + params.suffix = f"epoch-{params.epoch}-avg-{params.avg}" + + print("Script started") + + device = torch.device("cpu") + print(f"Device: {device}") + + symbol_table = k2.SymbolTable.from_file(params.tokens) + params.blank_id = symbol_table[""] + params.unk_id = symbol_table[""] + params.vocab_size = len(symbol_table) + + print("About to create model") + model = get_model(params) + + if params.iter > 0: + filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[ + : params.avg + 1 + ] + if len(filenames) == 0: + raise ValueError( + f"No checkpoints found for --iter {params.iter}, --avg {params.avg}" + ) + elif len(filenames) < params.avg + 1: + raise ValueError( + f"Not enough checkpoints ({len(filenames)}) found for" + f" --iter {params.iter}, --avg {params.avg}" + ) + filename_start = filenames[-1] + filename_end = filenames[0] + print( + "Calculating the averaged model over iteration checkpoints" + f" from {filename_start} (excluded) to {filename_end}" + ) + model.to(device) + model.load_state_dict( + average_checkpoints_with_averaged_model( + filename_start=filename_start, + filename_end=filename_end, + device=device, + ) + ) + filename = params.exp_dir / f"iter-{params.iter}-avg-{params.avg}.pt" + torch.save({"model": model.state_dict()}, filename) + else: + assert params.avg > 0, params.avg + start = params.epoch - params.avg + assert start >= 1, start + filename_start = f"{params.exp_dir}/epoch-{start}.pt" + filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt" + print( + f"Calculating the averaged model over epoch range from " + f"{start} (excluded) to {params.epoch}" + ) + model.to(device) + model.load_state_dict( + average_checkpoints_with_averaged_model( + filename_start=filename_start, + filename_end=filename_end, + device=device, + ) + ) + filename = params.exp_dir / f"epoch-{params.epoch}-avg-{params.avg}.pt" + torch.save({"model": model.state_dict()}, filename) + + num_param = sum([p.numel() for p in model.parameters()]) + print(f"Number of model parameters: {num_param}") + + print("Done!") + + +if __name__ == "__main__": + main() diff --git a/egs/multi_zh-hans/ASR/zipformer/jit_pretrained.py b/egs/multi_zh-hans/ASR/zipformer/jit_pretrained.py new file mode 120000 index 000000000..25108391f --- /dev/null +++ b/egs/multi_zh-hans/ASR/zipformer/jit_pretrained.py @@ -0,0 +1 @@ +../../../librispeech/ASR/zipformer/jit_pretrained.py \ No newline at end of file diff --git a/egs/multi_zh-hans/ASR/zipformer/jit_pretrained_ctc.py b/egs/multi_zh-hans/ASR/zipformer/jit_pretrained_ctc.py new file mode 120000 index 000000000..9a8da5844 --- /dev/null +++ b/egs/multi_zh-hans/ASR/zipformer/jit_pretrained_ctc.py @@ -0,0 +1 @@ +../../../librispeech/ASR/zipformer/jit_pretrained_ctc.py \ No newline at end of file diff --git a/egs/multi_zh-hans/ASR/zipformer/jit_pretrained_streaming.py b/egs/multi_zh-hans/ASR/zipformer/jit_pretrained_streaming.py new file mode 120000 index 000000000..1962351e9 --- /dev/null +++ b/egs/multi_zh-hans/ASR/zipformer/jit_pretrained_streaming.py @@ -0,0 +1 @@ +../../../librispeech/ASR/zipformer/jit_pretrained_streaming.py \ No newline at end of file diff --git a/egs/multi_zh-hans/ASR/zipformer/joiner.py b/egs/multi_zh-hans/ASR/zipformer/joiner.py new file mode 120000 index 000000000..5b8a36332 --- /dev/null +++ b/egs/multi_zh-hans/ASR/zipformer/joiner.py @@ -0,0 +1 @@ +../../../librispeech/ASR/zipformer/joiner.py \ No newline at end of file diff --git a/egs/multi_zh-hans/ASR/zipformer/model.py b/egs/multi_zh-hans/ASR/zipformer/model.py new file mode 120000 index 000000000..cd7e07d72 --- /dev/null +++ b/egs/multi_zh-hans/ASR/zipformer/model.py @@ -0,0 +1 @@ +../../../librispeech/ASR/zipformer/model.py \ No newline at end of file diff --git a/egs/multi_zh-hans/ASR/zipformer/multi_dataset.py b/egs/multi_zh-hans/ASR/zipformer/multi_dataset.py new file mode 100644 index 000000000..b1920e62e --- /dev/null +++ b/egs/multi_zh-hans/ASR/zipformer/multi_dataset.py @@ -0,0 +1,316 @@ +# Copyright 2023 Xiaomi Corp. (authors: Zengrui Jin) +# +# See ../../../../LICENSE for clarification regarding multiple authors +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + + +import glob +import logging +import re +from pathlib import Path +from typing import Dict, List + +import lhotse +from lhotse import CutSet, load_manifest_lazy + + +class MultiDataset: + def __init__(self, fbank_dir: str): + """ + Args: + manifest_dir: + It is expected to contain the following files: + - aidatatang_cuts_train.jsonl.gz + - aishell_cuts_train.jsonl.gz + - aishell2_cuts_train.jsonl.gz + - aishell4_cuts_train_L.jsonl.gz + - aishell4_cuts_train_M.jsonl.gz + - aishell4_cuts_train_S.jsonl.gz + - alimeeting-far_cuts_train.jsonl.gz + - magicdata_cuts_train.jsonl.gz + - primewords_cuts_train.jsonl.gz + - stcmds_cuts_train.jsonl.gz + - thchs_30_cuts_train.jsonl.gz + - kespeech/kespeech-asr_cuts_train_phase1.jsonl.gz + - kespeech/kespeech-asr_cuts_train_phase2.jsonl.gz + - wenetspeech/cuts_L.jsonl.gz + """ + self.fbank_dir = Path(fbank_dir) + + def train_cuts(self) -> CutSet: + logging.info("About to get multidataset train cuts") + + # THCHS-30 + logging.info("Loading THCHS-30 in lazy mode") + thchs_30_cuts = load_manifest_lazy( + self.fbank_dir / "thchs_30_cuts_train.jsonl.gz" + ) + + # AISHELL-1 + logging.info("Loading Aishell-1 in lazy mode") + aishell_cuts = load_manifest_lazy( + self.fbank_dir / "aishell_cuts_train.jsonl.gz" + ) + + # AISHELL-2 + logging.info("Loading Aishell-2 in lazy mode") + aishell_2_cuts = load_manifest_lazy( + self.fbank_dir / "aishell2_cuts_train.jsonl.gz" + ) + + # AISHELL-4 + logging.info("Loading Aishell-4 in lazy mode") + aishell_4_L_cuts = load_manifest_lazy( + self.fbank_dir / "aishell4_cuts_train_L.jsonl.gz" + ) + aishell_4_M_cuts = load_manifest_lazy( + self.fbank_dir / "aishell4_cuts_train_M.jsonl.gz" + ) + aishell_4_S_cuts = load_manifest_lazy( + self.fbank_dir / "aishell4_cuts_train_S.jsonl.gz" + ) + + # ST-CMDS + logging.info("Loading ST-CMDS in lazy mode") + stcmds_cuts = load_manifest_lazy(self.fbank_dir / "stcmds_cuts_train.jsonl.gz") + + # Primewords + logging.info("Loading Primewords in lazy mode") + primewords_cuts = load_manifest_lazy( + self.fbank_dir / "primewords_cuts_train.jsonl.gz" + ) + + # MagicData + logging.info("Loading MagicData in lazy mode") + magicdata_cuts = load_manifest_lazy( + self.fbank_dir / "magicdata_cuts_train.jsonl.gz" + ) + + # Aidatatang_200zh + logging.info("Loading Aidatatang_200zh in lazy mode") + aidatatang_200zh_cuts = load_manifest_lazy( + self.fbank_dir / "aidatatang_cuts_train.jsonl.gz" + ) + + # Ali-Meeting + logging.info("Loading Ali-Meeting in lazy mode") + alimeeting_cuts = load_manifest_lazy( + self.fbank_dir / "alimeeting-far_cuts_train.jsonl.gz" + ) + + # WeNetSpeech + logging.info("Loading WeNetSpeech in lazy mode") + wenetspeech_L_cuts = load_manifest_lazy( + self.fbank_dir / "wenetspeech" / "cuts_L.jsonl.gz" + ) + + # KeSpeech + logging.info("Loading KeSpeech in lazy mode") + kespeech_1_cuts = load_manifest_lazy( + self.fbank_dir / "kespeech" / "kespeech-asr_cuts_train_phase1.jsonl.gz" + ) + kespeech_2_cuts = load_manifest_lazy( + self.fbank_dir / "kespeech" / "kespeech-asr_cuts_train_phase2.jsonl.gz" + ) + + return CutSet.mux( + thchs_30_cuts, + aishell_cuts, + aishell_2_cuts, + aishell_4_L_cuts, + aishell_4_M_cuts, + aishell_4_S_cuts, + stcmds_cuts, + primewords_cuts, + magicdata_cuts, + aidatatang_200zh_cuts, + alimeeting_cuts, + wenetspeech_L_cuts, + kespeech_1_cuts, + kespeech_2_cuts, + weights=[ + len(thchs_30_cuts), + len(aishell_cuts), + len(aishell_2_cuts), + len(aishell_4_L_cuts), + len(aishell_4_M_cuts), + len(aishell_4_S_cuts), + len(stcmds_cuts), + len(primewords_cuts), + len(magicdata_cuts), + len(aidatatang_200zh_cuts), + len(alimeeting_cuts), + len(wenetspeech_L_cuts), + len(kespeech_1_cuts), + len(kespeech_2_cuts), + ], + ) + + def dev_cuts(self) -> CutSet: + logging.info("About to get multidataset dev cuts") + + # Aidatatang_200zh + logging.info("Loading Aidatatang_200zh DEV set in lazy mode") + aidatatang_dev_cuts = load_manifest_lazy( + self.fbank_dir / "aidatatang_cuts_dev.jsonl.gz" + ) + + # AISHELL + logging.info("Loading Aishell DEV set in lazy mode") + aishell_dev_cuts = load_manifest_lazy( + self.fbank_dir / "aishell_cuts_dev.jsonl.gz" + ) + + # AISHELL-2 + logging.info("Loading Aishell-2 DEV set in lazy mode") + aishell2_dev_cuts = load_manifest_lazy( + self.fbank_dir / "aishell2_cuts_dev.jsonl.gz" + ) + + # Ali-Meeting + logging.info("Loading Ali-Meeting DEV set in lazy mode") + alimeeting_dev_cuts = load_manifest_lazy( + self.fbank_dir / "alimeeting-far_cuts_eval.jsonl.gz" + ) + + # MagicData + logging.info("Loading MagicData DEV set in lazy mode") + magicdata_dev_cuts = load_manifest_lazy( + self.fbank_dir / "magicdata_cuts_dev.jsonl.gz" + ) + + # KeSpeech + logging.info("Loading KeSpeech DEV set in lazy mode") + kespeech_dev_phase1_cuts = load_manifest_lazy( + self.fbank_dir / "kespeech" / "kespeech-asr_cuts_dev_phase1.jsonl.gz" + ) + kespeech_dev_phase2_cuts = load_manifest_lazy( + self.fbank_dir / "kespeech" / "kespeech-asr_cuts_dev_phase2.jsonl.gz" + ) + + # WeNetSpeech + logging.info("Loading WeNetSpeech DEV set in lazy mode") + wenetspeech_dev_cuts = load_manifest_lazy( + self.fbank_dir / "wenetspeech" / "cuts_DEV.jsonl.gz" + ) + + return wenetspeech_dev_cuts + # return [ + # aidatatang_dev_cuts, + # aishell_dev_cuts, + # aishell2_dev_cuts, + # alimeeting_dev_cuts, + # magicdata_dev_cuts, + # kespeech_dev_phase1_cuts, + # kespeech_dev_phase2_cuts, + # wenetspeech_dev_cuts, + # ] + + def test_cuts(self) -> Dict[str, CutSet]: + logging.info("About to get multidataset test cuts") + + # Aidatatang_200zh + logging.info("Loading Aidatatang_200zh set in lazy mode") + aidatatang_test_cuts = load_manifest_lazy( + self.fbank_dir / "aidatatang_cuts_test.jsonl.gz" + ) + aidatatang_dev_cuts = load_manifest_lazy( + self.fbank_dir / "aidatatang_cuts_dev.jsonl.gz" + ) + + # AISHELL + logging.info("Loading Aishell set in lazy mode") + aishell_test_cuts = load_manifest_lazy( + self.fbank_dir / "aishell_cuts_test.jsonl.gz" + ) + aishell_dev_cuts = load_manifest_lazy( + self.fbank_dir / "aishell_cuts_dev.jsonl.gz" + ) + + # AISHELL-2 + logging.info("Loading Aishell-2 set in lazy mode") + aishell2_test_cuts = load_manifest_lazy( + self.fbank_dir / "aishell2_cuts_test.jsonl.gz" + ) + aishell2_dev_cuts = load_manifest_lazy( + self.fbank_dir / "aishell2_cuts_dev.jsonl.gz" + ) + + # AISHELL-4 + logging.info("Loading Aishell-4 TEST set in lazy mode") + aishell4_test_cuts = load_manifest_lazy( + self.fbank_dir / "aishell4_cuts_test.jsonl.gz" + ) + + # Ali-Meeting + logging.info("Loading Ali-Meeting set in lazy mode") + alimeeting_test_cuts = load_manifest_lazy( + self.fbank_dir / "alimeeting-far_cuts_test.jsonl.gz" + ) + alimeeting_eval_cuts = load_manifest_lazy( + self.fbank_dir / "alimeeting-far_cuts_eval.jsonl.gz" + ) + + # MagicData + logging.info("Loading MagicData set in lazy mode") + magicdata_test_cuts = load_manifest_lazy( + self.fbank_dir / "magicdata_cuts_test.jsonl.gz" + ) + magicdata_dev_cuts = load_manifest_lazy( + self.fbank_dir / "magicdata_cuts_dev.jsonl.gz" + ) + + # KeSpeech + logging.info("Loading KeSpeech set in lazy mode") + kespeech_test_cuts = load_manifest_lazy( + self.fbank_dir / "kespeech" / "kespeech-asr_cuts_test.jsonl.gz" + ) + kespeech_dev_phase1_cuts = load_manifest_lazy( + self.fbank_dir / "kespeech" / "kespeech-asr_cuts_dev_phase1.jsonl.gz" + ) + kespeech_dev_phase2_cuts = load_manifest_lazy( + self.fbank_dir / "kespeech" / "kespeech-asr_cuts_dev_phase2.jsonl.gz" + ) + + # WeNetSpeech + logging.info("Loading WeNetSpeech set in lazy mode") + wenetspeech_test_meeting_cuts = load_manifest_lazy( + self.fbank_dir / "wenetspeech" / "cuts_TEST_MEETING.jsonl.gz" + ) + wenetspeech_test_net_cuts = load_manifest_lazy( + self.fbank_dir / "wenetspeech" / "cuts_TEST_NET.jsonl.gz" + ) + wenetspeech_dev_cuts = load_manifest_lazy( + self.fbank_dir / "wenetspeech" / "cuts_DEV.jsonl.gz" + ) + + return { + "aidatatang_test": aidatatang_test_cuts, + "aidatatang_dev": aidatatang_dev_cuts, + "alimeeting_test": alimeeting_test_cuts, + "alimeeting_eval": alimeeting_eval_cuts, + "aishell_test": aishell_test_cuts, + "aishell_dev": aishell_dev_cuts, + "aishell-2_test": aishell2_test_cuts, + "aishell-2_dev": aishell2_dev_cuts, + "aishell-4": aishell4_test_cuts, + "magicdata_test": magicdata_test_cuts, + "magicdata_dev": magicdata_dev_cuts, + "kespeech-asr_test": kespeech_test_cuts, + "kespeech-asr_dev_phase1": kespeech_dev_phase1_cuts, + "kespeech-asr_dev_phase2": kespeech_dev_phase2_cuts, + "wenetspeech-meeting_test": wenetspeech_test_meeting_cuts, + "wenetspeech-net_test": wenetspeech_test_net_cuts, + "wenetspeech_dev": wenetspeech_dev_cuts, + } diff --git a/egs/multi_zh-hans/ASR/zipformer/onnx_check.py b/egs/multi_zh-hans/ASR/zipformer/onnx_check.py new file mode 120000 index 000000000..f3dd42004 --- /dev/null +++ b/egs/multi_zh-hans/ASR/zipformer/onnx_check.py @@ -0,0 +1 @@ +../../../librispeech/ASR/zipformer/onnx_check.py \ No newline at end of file diff --git a/egs/multi_zh-hans/ASR/zipformer/onnx_decode.py b/egs/multi_zh-hans/ASR/zipformer/onnx_decode.py new file mode 120000 index 000000000..0573b88c5 --- /dev/null +++ b/egs/multi_zh-hans/ASR/zipformer/onnx_decode.py @@ -0,0 +1 @@ +../../../librispeech/ASR/zipformer/onnx_decode.py \ No newline at end of file diff --git a/egs/multi_zh-hans/ASR/zipformer/onnx_pretrained-streaming.py b/egs/multi_zh-hans/ASR/zipformer/onnx_pretrained-streaming.py new file mode 120000 index 000000000..cfea104c2 --- /dev/null +++ b/egs/multi_zh-hans/ASR/zipformer/onnx_pretrained-streaming.py @@ -0,0 +1 @@ +../../../librispeech/ASR/zipformer/onnx_pretrained-streaming.py \ No newline at end of file diff --git a/egs/multi_zh-hans/ASR/zipformer/onnx_pretrained.py b/egs/multi_zh-hans/ASR/zipformer/onnx_pretrained.py new file mode 120000 index 000000000..8f32f4ee7 --- /dev/null +++ b/egs/multi_zh-hans/ASR/zipformer/onnx_pretrained.py @@ -0,0 +1 @@ +../../../librispeech/ASR/zipformer/onnx_pretrained.py \ No newline at end of file diff --git a/egs/multi_zh-hans/ASR/zipformer/optim.py b/egs/multi_zh-hans/ASR/zipformer/optim.py new file mode 120000 index 000000000..5eaa3cffd --- /dev/null +++ b/egs/multi_zh-hans/ASR/zipformer/optim.py @@ -0,0 +1 @@ +../../../librispeech/ASR/zipformer/optim.py \ No newline at end of file diff --git a/egs/multi_zh-hans/ASR/zipformer/pretrained.py b/egs/multi_zh-hans/ASR/zipformer/pretrained.py new file mode 100755 index 000000000..69ff382da --- /dev/null +++ b/egs/multi_zh-hans/ASR/zipformer/pretrained.py @@ -0,0 +1,381 @@ +#!/usr/bin/env python3 +# Copyright 2021-2023 Xiaomi Corp. (authors: Fangjun Kuang, Zengwei Yao) +# +# See ../../../../LICENSE for clarification regarding multiple authors +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +""" +This script loads a checkpoint and uses it to decode waves. +You can generate the checkpoint with the following command: + +Note: This is a example for librispeech dataset, if you are using different +dataset, you should change the argument values according to your dataset. + +- For non-streaming model: + +./zipformer/export.py \ + --exp-dir ./zipformer/exp \ + --tokens data/lang_bpe_2000/tokens.txt \ + --epoch 23 \ + --avg 1 + +- For streaming model: + +./zipformer/export.py \ + --exp-dir ./zipformer/exp \ + --causal 1 \ + --tokens data/lang_bpe_2000/tokens.txt \ + --epoch 23 \ + --avg 1 + +Usage of this script: + +- For non-streaming model: + +(1) greedy search +./zipformer/pretrained.py \ + --checkpoint ./zipformer/exp/pretrained.pt \ + --tokens data/lang_bpe_2000/tokens.txt \ + --method greedy_search \ + /path/to/foo.wav \ + /path/to/bar.wav + +(2) modified beam search +./zipformer/pretrained.py \ + --checkpoint ./zipformer/exp/pretrained.pt \ + --tokens ./data/lang_bpe_2000/tokens.txt \ + --method modified_beam_search \ + /path/to/foo.wav \ + /path/to/bar.wav + +(3) fast beam search +./zipformer/pretrained.py \ + --checkpoint ./zipformer/exp/pretrained.pt \ + --tokens ./data/lang_bpe_2000/tokens.txt \ + --method fast_beam_search \ + /path/to/foo.wav \ + /path/to/bar.wav + +- For streaming model: + +(1) greedy search +./zipformer/pretrained.py \ + --checkpoint ./zipformer/exp/pretrained.pt \ + --causal 1 \ + --chunk-size 16 \ + --left-context-frames 128 \ + --tokens ./data/lang_bpe_2000/tokens.txt \ + --method greedy_search \ + /path/to/foo.wav \ + /path/to/bar.wav + +(2) modified beam search +./zipformer/pretrained.py \ + --checkpoint ./zipformer/exp/pretrained.pt \ + --causal 1 \ + --chunk-size 16 \ + --left-context-frames 128 \ + --tokens ./data/lang_bpe_2000/tokens.txt \ + --method modified_beam_search \ + /path/to/foo.wav \ + /path/to/bar.wav + +(3) fast beam search +./zipformer/pretrained.py \ + --checkpoint ./zipformer/exp/pretrained.pt \ + --causal 1 \ + --chunk-size 16 \ + --left-context-frames 128 \ + --tokens ./data/lang_bpe_2000/tokens.txt \ + --method fast_beam_search \ + /path/to/foo.wav \ + /path/to/bar.wav + + +You can also use `./zipformer/exp/epoch-xx.pt`. + +Note: ./zipformer/exp/pretrained.pt is generated by ./zipformer/export.py +""" + + +import argparse +import logging +import math +from typing import List + +import k2 +import kaldifeat +import torch +import torchaudio +from beam_search import ( + fast_beam_search_one_best, + greedy_search_batch, + modified_beam_search, +) +from export import num_tokens +from torch.nn.utils.rnn import pad_sequence +from train import add_model_arguments, get_model, get_params + +from icefall.utils import make_pad_mask + + +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( + "--tokens", + type=str, + help="""Path to tokens.txt.""", + ) + + parser.add_argument( + "--method", + type=str, + default="greedy_search", + help="""Possible values are: + - greedy_search + - modified_beam_search + - fast_beam_search + """, + ) + + parser.add_argument( + "sound_files", + type=str, + nargs="+", + help="The input sound file(s) to transcribe. " + "Supported formats are those supported by torchaudio.load(). " + "For example, wav and flac are supported. " + "The sample rate has to be 16kHz.", + ) + + parser.add_argument( + "--sample-rate", + type=int, + default=16000, + help="The sample rate of the input sound file", + ) + + parser.add_argument( + "--beam-size", + type=int, + default=4, + help="""An integer indicating how many candidates we will keep for each + frame. Used only when --method is beam_search or + modified_beam_search.""", + ) + + parser.add_argument( + "--beam", + type=float, + default=4, + help="""A floating point value to calculate the cutoff score during beam + search (i.e., `cutoff = max-score - beam`), which is the same as the + `beam` in Kaldi. + Used only when --method is fast_beam_search""", + ) + + parser.add_argument( + "--max-contexts", + type=int, + default=4, + help="""Used only when --method is fast_beam_search""", + ) + + parser.add_argument( + "--max-states", + type=int, + default=8, + help="""Used only when --method is fast_beam_search""", + ) + + parser.add_argument( + "--context-size", + type=int, + default=2, + help="The context size in the decoder. 1 means bigram; 2 means tri-gram", + ) + + parser.add_argument( + "--max-sym-per-frame", + type=int, + default=1, + help="""Maximum number of symbols per frame. Used only when + --method is greedy_search. + """, + ) + + add_model_arguments(parser) + + return parser + + +def read_sound_files( + filenames: List[str], expected_sample_rate: float +) -> List[torch.Tensor]: + """Read a list of sound files into a list 1-D float32 torch tensors. + Args: + filenames: + A list of sound filenames. + expected_sample_rate: + The expected sample rate of the sound files. + Returns: + Return a list of 1-D float32 torch tensors. + """ + ans = [] + for f in filenames: + wave, sample_rate = torchaudio.load(f) + assert ( + sample_rate == expected_sample_rate + ), f"expected sample rate: {expected_sample_rate}. Given: {sample_rate}" + # We use only the first channel + ans.append(wave[0].contiguous()) + return ans + + +@torch.no_grad() +def main(): + parser = get_parser() + args = parser.parse_args() + + params = get_params() + + params.update(vars(args)) + + token_table = k2.SymbolTable.from_file(params.tokens) + + params.blank_id = token_table[""] + params.unk_id = token_table[""] + params.vocab_size = num_tokens(token_table) + 1 + + logging.info(f"{params}") + + device = torch.device("cpu") + if torch.cuda.is_available(): + device = torch.device("cuda", 0) + + logging.info(f"device: {device}") + + if params.causal: + assert ( + "," not in params.chunk_size + ), "chunk_size should be one value in decoding." + assert ( + "," not in params.left_context_frames + ), "left_context_frames should be one value in decoding." + + logging.info("Creating model") + model = get_model(params) + + num_param = sum([p.numel() for p in model.parameters()]) + logging.info(f"Number of model parameters: {num_param}") + + checkpoint = torch.load(args.checkpoint, map_location="cpu") + model.load_state_dict(checkpoint["model"], strict=False) + model.to(device) + model.eval() + + logging.info("Constructing Fbank computer") + opts = kaldifeat.FbankOptions() + opts.device = device + opts.frame_opts.dither = 0 + opts.frame_opts.snip_edges = False + opts.frame_opts.samp_freq = params.sample_rate + opts.mel_opts.num_bins = params.feature_dim + + fbank = kaldifeat.Fbank(opts) + + logging.info(f"Reading sound files: {params.sound_files}") + waves = read_sound_files( + filenames=params.sound_files, expected_sample_rate=params.sample_rate + ) + waves = [w.to(device) for w in waves] + + logging.info("Decoding started") + features = fbank(waves) + feature_lengths = [f.size(0) for f in features] + + features = pad_sequence(features, batch_first=True, padding_value=math.log(1e-10)) + feature_lengths = torch.tensor(feature_lengths, device=device) + + # model forward + encoder_out, encoder_out_lens = model.forward_encoder(features, feature_lengths) + + hyps = [] + msg = f"Using {params.method}" + logging.info(msg) + + def token_ids_to_words(token_ids: List[int]) -> str: + text = "" + for i in token_ids: + text += token_table[i] + return text.replace("▁", " ").strip() + + if params.method == "fast_beam_search": + decoding_graph = k2.trivial_graph(params.vocab_size - 1, device=device) + hyp_tokens = fast_beam_search_one_best( + model=model, + decoding_graph=decoding_graph, + encoder_out=encoder_out, + encoder_out_lens=encoder_out_lens, + beam=params.beam, + max_contexts=params.max_contexts, + max_states=params.max_states, + ) + for hyp in hyp_tokens: + hyps.append(token_ids_to_words(hyp)) + elif params.method == "modified_beam_search": + hyp_tokens = modified_beam_search( + model=model, + encoder_out=encoder_out, + encoder_out_lens=encoder_out_lens, + beam=params.beam_size, + ) + + for hyp in hyp_tokens: + hyps.append(token_ids_to_words(hyp)) + elif params.method == "greedy_search" and params.max_sym_per_frame == 1: + hyp_tokens = greedy_search_batch( + model=model, + encoder_out=encoder_out, + encoder_out_lens=encoder_out_lens, + ) + for hyp in hyp_tokens: + hyps.append(token_ids_to_words(hyp)) + else: + raise ValueError(f"Unsupported method: {params.method}") + + s = "\n" + for filename, hyp in zip(params.sound_files, hyps): + s += f"{filename}:\n{hyp}\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/multi_zh-hans/ASR/zipformer/scaling.py b/egs/multi_zh-hans/ASR/zipformer/scaling.py new file mode 120000 index 000000000..6f398f431 --- /dev/null +++ b/egs/multi_zh-hans/ASR/zipformer/scaling.py @@ -0,0 +1 @@ +../../../librispeech/ASR/zipformer/scaling.py \ No newline at end of file diff --git a/egs/multi_zh-hans/ASR/zipformer/scaling_converter.py b/egs/multi_zh-hans/ASR/zipformer/scaling_converter.py new file mode 120000 index 000000000..b0ecee05e --- /dev/null +++ b/egs/multi_zh-hans/ASR/zipformer/scaling_converter.py @@ -0,0 +1 @@ +../../../librispeech/ASR/zipformer/scaling_converter.py \ No newline at end of file diff --git a/egs/multi_zh-hans/ASR/zipformer/streaming_beam_search.py b/egs/multi_zh-hans/ASR/zipformer/streaming_beam_search.py new file mode 120000 index 000000000..b1ed54557 --- /dev/null +++ b/egs/multi_zh-hans/ASR/zipformer/streaming_beam_search.py @@ -0,0 +1 @@ +../../../librispeech/ASR/zipformer/streaming_beam_search.py \ No newline at end of file diff --git a/egs/multi_zh-hans/ASR/zipformer/streaming_decode.py b/egs/multi_zh-hans/ASR/zipformer/streaming_decode.py new file mode 120000 index 000000000..13fd02a78 --- /dev/null +++ b/egs/multi_zh-hans/ASR/zipformer/streaming_decode.py @@ -0,0 +1 @@ +../../../librispeech/ASR/zipformer/streaming_decode.py \ No newline at end of file diff --git a/egs/multi_zh-hans/ASR/zipformer/subsampling.py b/egs/multi_zh-hans/ASR/zipformer/subsampling.py new file mode 120000 index 000000000..01ae9002c --- /dev/null +++ b/egs/multi_zh-hans/ASR/zipformer/subsampling.py @@ -0,0 +1 @@ +../../../librispeech/ASR/zipformer/subsampling.py \ No newline at end of file diff --git a/egs/multi_zh-hans/ASR/zipformer/train.py b/egs/multi_zh-hans/ASR/zipformer/train.py new file mode 100755 index 000000000..4f2d728be --- /dev/null +++ b/egs/multi_zh-hans/ASR/zipformer/train.py @@ -0,0 +1,1385 @@ +#!/usr/bin/env python3 +# Copyright 2021-2023 Xiaomi Corp. (authors: Fangjun Kuang, +# Wei Kang, +# Mingshuang Luo, +# Zengwei Yao, +# Daniel Povey) +# +# 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" + +# For non-streaming model training: +./zipformer/train.py \ + --world-size 4 \ + --num-epochs 30 \ + --start-epoch 1 \ + --use-fp16 1 \ + --exp-dir zipformer/exp \ + --max-duration 1000 + +# For streaming model training: +./zipformer/train.py \ + --world-size 4 \ + --num-epochs 30 \ + --start-epoch 1 \ + --use-fp16 1 \ + --exp-dir zipformer/exp \ + --causal 1 \ + --max-duration 1000 + +It supports training with: + - transducer loss (default), with `--use-transducer True --use-ctc False` + - ctc loss (not recommended), with `--use-transducer False --use-ctc True` + - transducer loss & ctc loss, with `--use-transducer True --use-ctc True` +""" + + +import argparse +import copy +import logging +import warnings +from pathlib import Path +from shutil import copyfile +from typing import Any, Dict, Optional, Tuple, Union + +import k2 +import optim +import sentencepiece as spm +import torch +import torch.multiprocessing as mp +import torch.nn as nn +from asr_datamodule import AsrDataModule +from decoder import Decoder +from joiner import Joiner +from lhotse.cut import Cut +from lhotse.dataset.sampling.base import CutSampler +from lhotse.utils import fix_random_seed +from model import AsrModel +from multi_dataset import MultiDataset +from optim import Eden, ScaledAdam +from scaling import ScheduledFloat +from subsampling import Conv2dSubsampling +from torch import Tensor +from torch.cuda.amp import GradScaler +from torch.nn.parallel import DistributedDataParallel as DDP +from torch.utils.tensorboard import SummaryWriter +from zipformer import Zipformer2 + +from icefall import diagnostics +from icefall.checkpoint import load_checkpoint, remove_checkpoints +from icefall.checkpoint import save_checkpoint as save_checkpoint_impl +from icefall.checkpoint import ( + save_checkpoint_with_global_batch_idx, + update_averaged_model, +) +from icefall.dist import cleanup_dist, setup_dist +from icefall.env import get_env_info +from icefall.hooks import register_inf_check_hooks +from icefall.utils import ( + AttributeDict, + MetricsTracker, + get_parameter_groups_with_lrs, + setup_logger, + str2bool, +) + +LRSchedulerType = Union[torch.optim.lr_scheduler._LRScheduler, optim.LRScheduler] + + +def get_adjusted_batch_count(params: AttributeDict) -> float: + # returns the number of batches we would have used so far if we had used the reference + # duration. This is for purposes of set_batch_count(). + return ( + params.batch_idx_train + * (params.max_duration * params.world_size) + / params.ref_duration + ) + + +def set_batch_count(model: Union[nn.Module, DDP], batch_count: float) -> None: + if isinstance(model, DDP): + # get underlying nn.Module + model = model.module + for name, module in model.named_modules(): + if hasattr(module, "batch_count"): + module.batch_count = batch_count + if hasattr(module, "name"): + module.name = name + + +def add_model_arguments(parser: argparse.ArgumentParser): + parser.add_argument( + "--num-encoder-layers", + type=str, + default="2,2,3,4,3,2", + help="Number of zipformer encoder layers per stack, comma separated.", + ) + + parser.add_argument( + "--downsampling-factor", + type=str, + default="1,2,4,8,4,2", + help="Downsampling factor for each stack of encoder layers.", + ) + + parser.add_argument( + "--feedforward-dim", + type=str, + default="512,768,1024,1536,1024,768", + help="Feedforward dimension of the zipformer encoder layers, per stack, comma separated.", + ) + + parser.add_argument( + "--num-heads", + type=str, + default="4,4,4,8,4,4", + help="Number of attention heads in the zipformer encoder layers: a single int or comma-separated list.", + ) + + parser.add_argument( + "--encoder-dim", + type=str, + default="192,256,384,512,384,256", + help="Embedding dimension in encoder stacks: a single int or comma-separated list.", + ) + + parser.add_argument( + "--query-head-dim", + type=str, + default="32", + help="Query/key dimension per head in encoder stacks: a single int or comma-separated list.", + ) + + parser.add_argument( + "--value-head-dim", + type=str, + default="12", + help="Value dimension per head in encoder stacks: a single int or comma-separated list.", + ) + + parser.add_argument( + "--pos-head-dim", + type=str, + default="4", + help="Positional-encoding dimension per head in encoder stacks: a single int or comma-separated list.", + ) + + parser.add_argument( + "--pos-dim", + type=int, + default="48", + help="Positional-encoding embedding dimension", + ) + + parser.add_argument( + "--encoder-unmasked-dim", + type=str, + default="192,192,256,256,256,192", + help="Unmasked dimensions in the encoders, relates to augmentation during training. " + "A single int or comma-separated list. Must be <= each corresponding encoder_dim.", + ) + + parser.add_argument( + "--cnn-module-kernel", + type=str, + default="31,31,15,15,15,31", + help="Sizes of convolutional kernels in convolution modules in each encoder stack: " + "a single int or comma-separated list.", + ) + + parser.add_argument( + "--decoder-dim", + type=int, + default=512, + help="Embedding dimension in the decoder model.", + ) + + parser.add_argument( + "--joiner-dim", + type=int, + default=512, + help="""Dimension used in the joiner model. + Outputs from the encoder and decoder model are projected + to this dimension before adding. + """, + ) + + parser.add_argument( + "--causal", + type=str2bool, + default=False, + help="If True, use causal version of model.", + ) + + parser.add_argument( + "--chunk-size", + type=str, + default="16,32,64,-1", + help="Chunk sizes (at 50Hz frame rate) will be chosen randomly from this list during training. " + " Must be just -1 if --causal=False", + ) + + parser.add_argument( + "--left-context-frames", + type=str, + default="64,128,256,-1", + help="Maximum left-contexts for causal training, measured in frames which will " + "be converted to a number of chunks. If splitting into chunks, " + "chunk left-context frames will be chosen randomly from this list; else not relevant.", + ) + + parser.add_argument( + "--use-transducer", + type=str2bool, + default=True, + help="If True, use Transducer head.", + ) + + parser.add_argument( + "--use-ctc", + type=str2bool, + default=False, + help="If True, use CTC head.", + ) + + +def get_parser(): + parser = argparse.ArgumentParser( + formatter_class=argparse.ArgumentDefaultsHelpFormatter + ) + + parser.add_argument( + "--world-size", + type=int, + default=1, + help="Number of GPUs for DDP training.", + ) + + parser.add_argument( + "--master-port", + type=int, + default=12354, + help="Master port to use for DDP training.", + ) + + parser.add_argument( + "--tensorboard", + type=str2bool, + default=True, + help="Should various information be logged in tensorboard.", + ) + + parser.add_argument( + "--num-epochs", + type=int, + default=30, + help="Number of epochs to train.", + ) + + parser.add_argument( + "--start-epoch", + type=int, + default=1, + help="""Resume training from this epoch. It should be positive. + If larger than 1, it will load checkpoint from + exp-dir/epoch-{start_epoch-1}.pt + """, + ) + + parser.add_argument( + "--start-batch", + type=int, + default=0, + help="""If positive, --start-epoch is ignored and + it loads the checkpoint from exp-dir/checkpoint-{start_batch}.pt + """, + ) + + parser.add_argument( + "--exp-dir", + type=str, + default="zipformer/exp", + help="""The experiment dir. + It specifies the directory where all training related + files, e.g., checkpoints, log, etc, are saved + """, + ) + + parser.add_argument( + "--bpe-model", + type=str, + default="data/lang_bpe_2000/bpe.model", + help="Path to the BPE model", + ) + + parser.add_argument( + "--base-lr", type=float, default=0.045, help="The base learning rate." + ) + + parser.add_argument( + "--lr-batches", + type=float, + default=7500, + help="""Number of steps that affects how rapidly the learning rate + decreases. We suggest not to change this.""", + ) + + parser.add_argument( + "--lr-epochs", + type=float, + default=3.5, + help="""Number of epochs that affects how rapidly the learning rate decreases. + """, + ) + + parser.add_argument( + "--ref-duration", + type=float, + default=600, + help="Reference batch duration for purposes of adjusting batch counts for setting various " + "schedules inside the model", + ) + + parser.add_argument( + "--context-size", + type=int, + default=2, + help="The context size in the decoder. 1 means bigram; " "2 means tri-gram", + ) + + parser.add_argument( + "--prune-range", + type=int, + default=5, + help="The prune range for rnnt loss, it means how many symbols(context)" + "we are using to compute the loss", + ) + + parser.add_argument( + "--lm-scale", + type=float, + default=0.25, + help="The scale to smooth the loss with lm " + "(output of prediction network) part.", + ) + + parser.add_argument( + "--am-scale", + type=float, + default=0.0, + help="The scale to smooth the loss with am (output of encoder network)" "part.", + ) + + parser.add_argument( + "--simple-loss-scale", + type=float, + default=0.5, + help="To get pruning ranges, we will calculate a simple version" + "loss(joiner is just addition), this simple loss also uses for" + "training (as a regularization item). We will scale the simple loss" + "with this parameter before adding to the final loss.", + ) + + parser.add_argument( + "--ctc-loss-scale", + type=float, + default=0.2, + help="Scale for CTC loss.", + ) + + parser.add_argument( + "--seed", + type=int, + default=42, + help="The seed for random generators intended for reproducibility", + ) + + parser.add_argument( + "--print-diagnostics", + type=str2bool, + default=False, + help="Accumulate stats on activations, print them and exit.", + ) + + parser.add_argument( + "--inf-check", + type=str2bool, + default=False, + help="Add hooks to check for infinite module outputs and gradients.", + ) + + parser.add_argument( + "--save-every-n", + type=int, + default=4000, + help="""Save checkpoint after processing this number of batches" + periodically. We save checkpoint to exp-dir/ whenever + params.batch_idx_train % save_every_n == 0. The checkpoint filename + has the form: f'exp-dir/checkpoint-{params.batch_idx_train}.pt' + Note: It also saves checkpoint to `exp-dir/epoch-xxx.pt` at the + end of each epoch where `xxx` is the epoch number counting from 1. + """, + ) + + parser.add_argument( + "--keep-last-k", + type=int, + default=30, + help="""Only keep this number of checkpoints on disk. + For instance, if it is 3, there are only 3 checkpoints + in the exp-dir with filenames `checkpoint-xxx.pt`. + It does not affect checkpoints with name `epoch-xxx.pt`. + """, + ) + + parser.add_argument( + "--average-period", + type=int, + default=200, + help="""Update the averaged model, namely `model_avg`, after processing + this number of batches. `model_avg` is a separate version of model, + in which each floating-point parameter is the average of all the + parameters from the start of training. Each time we take the average, + we do: `model_avg = model * (average_period / batch_idx_train) + + model_avg * ((batch_idx_train - average_period) / batch_idx_train)`. + """, + ) + + parser.add_argument( + "--use-fp16", + type=str2bool, + default=False, + help="Whether to use half precision training.", + ) + + add_model_arguments(parser) + + return parser + + +def get_params() -> AttributeDict: + """Return a dict containing training parameters. + + All training related parameters that are not passed from the commandline + are saved in the variable `params`. + + Commandline options are merged into `params` after they are parsed, so + you can also access them via `params`. + + Explanation of options saved in `params`: + + - best_train_loss: Best training loss so far. It is used to select + the model that has the lowest training loss. It is + updated during the training. + + - best_valid_loss: Best validation loss so far. It is used to select + the model that has the lowest validation loss. It is + updated during the training. + + - best_train_epoch: It is the epoch that has the best training loss. + + - best_valid_epoch: It is the epoch that has the best validation loss. + + - batch_idx_train: Used to writing statistics to tensorboard. It + contains number of batches trained so far across + epochs. + + - log_interval: Print training loss if batch_idx % log_interval` is 0 + + - reset_interval: Reset statistics if batch_idx % reset_interval is 0 + + - valid_interval: Run validation if batch_idx % valid_interval is 0 + + - feature_dim: The model input dim. It has to match the one used + in computing features. + + - subsampling_factor: The subsampling factor for the model. + + - encoder_dim: Hidden dim for multi-head attention model. + + - num_decoder_layers: Number of decoder layer of transformer decoder. + + - warm_step: The warmup period that dictates the decay of the + scale on "simple" (un-pruned) loss. + """ + params = AttributeDict( + { + "best_train_loss": float("inf"), + "best_valid_loss": float("inf"), + "best_train_epoch": -1, + "best_valid_epoch": -1, + "batch_idx_train": 0, + "log_interval": 50, + "reset_interval": 200, + "valid_interval": 3000, # For the 100h subset, use 800 + # parameters for zipformer + "feature_dim": 80, + "subsampling_factor": 4, # not passed in, this is fixed. + "warm_step": 2000, + "env_info": get_env_info(), + } + ) + + return params + + +def _to_int_tuple(s: str): + return tuple(map(int, s.split(","))) + + +def get_encoder_embed(params: AttributeDict) -> nn.Module: + # encoder_embed converts the input of shape (N, T, num_features) + # to the shape (N, (T - 7) // 2, encoder_dims). + # That is, it does two things simultaneously: + # (1) subsampling: T -> (T - 7) // 2 + # (2) embedding: num_features -> encoder_dims + # In the normal configuration, we will downsample once more at the end + # by a factor of 2, and most of the encoder stacks will run at a lower + # sampling rate. + encoder_embed = Conv2dSubsampling( + in_channels=params.feature_dim, + out_channels=_to_int_tuple(params.encoder_dim)[0], + dropout=ScheduledFloat((0.0, 0.3), (20000.0, 0.1)), + ) + return encoder_embed + + +def get_encoder_model(params: AttributeDict) -> nn.Module: + encoder = Zipformer2( + output_downsampling_factor=2, + downsampling_factor=_to_int_tuple(params.downsampling_factor), + num_encoder_layers=_to_int_tuple(params.num_encoder_layers), + encoder_dim=_to_int_tuple(params.encoder_dim), + encoder_unmasked_dim=_to_int_tuple(params.encoder_unmasked_dim), + query_head_dim=_to_int_tuple(params.query_head_dim), + pos_head_dim=_to_int_tuple(params.pos_head_dim), + value_head_dim=_to_int_tuple(params.value_head_dim), + pos_dim=params.pos_dim, + num_heads=_to_int_tuple(params.num_heads), + feedforward_dim=_to_int_tuple(params.feedforward_dim), + cnn_module_kernel=_to_int_tuple(params.cnn_module_kernel), + dropout=ScheduledFloat((0.0, 0.3), (20000.0, 0.1)), + warmup_batches=4000.0, + causal=params.causal, + chunk_size=_to_int_tuple(params.chunk_size), + left_context_frames=_to_int_tuple(params.left_context_frames), + ) + return encoder + + +def get_decoder_model(params: AttributeDict) -> nn.Module: + decoder = Decoder( + vocab_size=params.vocab_size, + decoder_dim=params.decoder_dim, + blank_id=params.blank_id, + context_size=params.context_size, + ) + return decoder + + +def get_joiner_model(params: AttributeDict) -> nn.Module: + joiner = Joiner( + encoder_dim=max(_to_int_tuple(params.encoder_dim)), + decoder_dim=params.decoder_dim, + joiner_dim=params.joiner_dim, + vocab_size=params.vocab_size, + ) + return joiner + + +def get_model(params: AttributeDict) -> nn.Module: + assert params.use_transducer or params.use_ctc, ( + f"At least one of them should be True, " + f"but got params.use_transducer={params.use_transducer}, " + f"params.use_ctc={params.use_ctc}" + ) + + encoder_embed = get_encoder_embed(params) + encoder = get_encoder_model(params) + + if params.use_transducer: + decoder = get_decoder_model(params) + joiner = get_joiner_model(params) + else: + decoder = None + joiner = None + + model = AsrModel( + encoder_embed=encoder_embed, + encoder=encoder, + decoder=decoder, + joiner=joiner, + encoder_dim=max(_to_int_tuple(params.encoder_dim)), + decoder_dim=params.decoder_dim, + vocab_size=params.vocab_size, + use_transducer=params.use_transducer, + use_ctc=params.use_ctc, + ) + return model + + +def load_checkpoint_if_available( + params: AttributeDict, + model: nn.Module, + model_avg: nn.Module = None, + optimizer: Optional[torch.optim.Optimizer] = None, + scheduler: Optional[LRSchedulerType] = None, +) -> Optional[Dict[str, Any]]: + """Load checkpoint from file. + + If params.start_batch is positive, it will load the checkpoint from + `params.exp_dir/checkpoint-{params.start_batch}.pt`. Otherwise, if + params.start_epoch is larger than 1, it will load the checkpoint from + `params.start_epoch - 1`. + + Apart from loading state dict for `model` and `optimizer` it also updates + `best_train_epoch`, `best_train_loss`, `best_valid_epoch`, + and `best_valid_loss` in `params`. + + Args: + params: + The return value of :func:`get_params`. + model: + The training model. + model_avg: + The stored model averaged from the start of training. + optimizer: + The optimizer that we are using. + scheduler: + The scheduler that we are using. + Returns: + Return a dict containing previously saved training info. + """ + if params.start_batch > 0: + filename = params.exp_dir / f"checkpoint-{params.start_batch}.pt" + elif params.start_epoch > 1: + filename = params.exp_dir / f"epoch-{params.start_epoch-1}.pt" + else: + return None + + assert filename.is_file(), f"{filename} does not exist!" + + saved_params = load_checkpoint( + filename, + model=model, + model_avg=model_avg, + optimizer=optimizer, + scheduler=scheduler, + ) + + keys = [ + "best_train_epoch", + "best_valid_epoch", + "batch_idx_train", + "best_train_loss", + "best_valid_loss", + ] + for k in keys: + params[k] = saved_params[k] + + if params.start_batch > 0: + if "cur_epoch" in saved_params: + params["start_epoch"] = saved_params["cur_epoch"] + + return saved_params + + +def save_checkpoint( + params: AttributeDict, + model: Union[nn.Module, DDP], + model_avg: Optional[nn.Module] = None, + optimizer: Optional[torch.optim.Optimizer] = None, + scheduler: Optional[LRSchedulerType] = None, + sampler: Optional[CutSampler] = None, + scaler: Optional[GradScaler] = None, + rank: int = 0, +) -> None: + """Save model, optimizer, scheduler and training stats to file. + + Args: + params: + It is returned by :func:`get_params`. + model: + The training model. + model_avg: + The stored model averaged from the start of training. + optimizer: + The optimizer used in the training. + sampler: + The sampler for the training dataset. + scaler: + The scaler used for mix precision training. + """ + if rank != 0: + return + filename = params.exp_dir / f"epoch-{params.cur_epoch}.pt" + save_checkpoint_impl( + filename=filename, + model=model, + model_avg=model_avg, + params=params, + optimizer=optimizer, + scheduler=scheduler, + sampler=sampler, + scaler=scaler, + rank=rank, + ) + + if params.best_train_epoch == params.cur_epoch: + best_train_filename = params.exp_dir / "best-train-loss.pt" + copyfile(src=filename, dst=best_train_filename) + + if params.best_valid_epoch == params.cur_epoch: + best_valid_filename = params.exp_dir / "best-valid-loss.pt" + copyfile(src=filename, dst=best_valid_filename) + + +def compute_loss( + params: AttributeDict, + model: Union[nn.Module, DDP], + sp: spm.SentencePieceProcessor, + batch: dict, + is_training: bool, +) -> Tuple[Tensor, MetricsTracker]: + """ + Compute loss given the model and its inputs. + + Args: + params: + Parameters for training. See :func:`get_params`. + model: + The model for training. It is an instance of Zipformer in our case. + batch: + A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()` + for the content in it. + is_training: + True for training. False for validation. When it is True, this + function enables autograd during computation; when it is False, it + disables autograd. + warmup: a floating point value which increases throughout training; + values >= 1.0 are fully warmed up and have all modules present. + """ + device = model.device if isinstance(model, DDP) else next(model.parameters()).device + feature = batch["inputs"] + # at entry, feature is (N, T, C) + assert feature.ndim == 3 + feature = feature.to(device) + + supervisions = batch["supervisions"] + feature_lens = supervisions["num_frames"].to(device) + + batch_idx_train = params.batch_idx_train + warm_step = params.warm_step + + texts = batch["supervisions"]["text"] + y = sp.encode(texts, out_type=int) + y = k2.RaggedTensor(y) + + with torch.set_grad_enabled(is_training): + simple_loss, pruned_loss, ctc_loss = model( + x=feature, + x_lens=feature_lens, + y=y, + prune_range=params.prune_range, + am_scale=params.am_scale, + lm_scale=params.lm_scale, + ) + + loss = 0.0 + + if params.use_transducer: + s = params.simple_loss_scale + # take down the scale on the simple loss from 1.0 at the start + # to params.simple_loss scale by warm_step. + simple_loss_scale = ( + s + if batch_idx_train >= warm_step + else 1.0 - (batch_idx_train / warm_step) * (1.0 - s) + ) + pruned_loss_scale = ( + 1.0 + if batch_idx_train >= warm_step + else 0.1 + 0.9 * (batch_idx_train / warm_step) + ) + loss += simple_loss_scale * simple_loss + pruned_loss_scale * pruned_loss + + if params.use_ctc: + loss += params.ctc_loss_scale * ctc_loss + + assert loss.requires_grad == is_training + + info = MetricsTracker() + with warnings.catch_warnings(): + warnings.simplefilter("ignore") + info["frames"] = (feature_lens // params.subsampling_factor).sum().item() + + # Note: We use reduction=sum while computing the loss. + info["loss"] = loss.detach().cpu().item() + if params.use_transducer: + info["simple_loss"] = simple_loss.detach().cpu().item() + info["pruned_loss"] = pruned_loss.detach().cpu().item() + if params.use_ctc: + info["ctc_loss"] = ctc_loss.detach().cpu().item() + + return loss, info + + +def compute_validation_loss( + params: AttributeDict, + model: Union[nn.Module, DDP], + sp: spm.SentencePieceProcessor, + valid_dl: torch.utils.data.DataLoader, + world_size: int = 1, +) -> MetricsTracker: + """Run the validation process.""" + model.eval() + + tot_loss = MetricsTracker() + + for batch_idx, batch in enumerate(valid_dl): + loss, loss_info = compute_loss( + params=params, + model=model, + sp=sp, + batch=batch, + is_training=False, + ) + assert loss.requires_grad is False + tot_loss = tot_loss + loss_info + + if world_size > 1: + tot_loss.reduce(loss.device) + + loss_value = tot_loss["loss"] / tot_loss["frames"] + if loss_value < params.best_valid_loss: + params.best_valid_epoch = params.cur_epoch + params.best_valid_loss = loss_value + + return tot_loss + + +def train_one_epoch( + params: AttributeDict, + model: Union[nn.Module, DDP], + optimizer: torch.optim.Optimizer, + scheduler: LRSchedulerType, + sp: spm.SentencePieceProcessor, + train_dl: torch.utils.data.DataLoader, + valid_dl: torch.utils.data.DataLoader, + scaler: GradScaler, + model_avg: Optional[nn.Module] = None, + tb_writer: Optional[SummaryWriter] = None, + world_size: int = 1, + rank: int = 0, +) -> None: + """Train the model for one epoch. + + The training loss from the mean of all frames is saved in + `params.train_loss`. It runs the validation process every + `params.valid_interval` batches. + + Args: + params: + It is returned by :func:`get_params`. + model: + The model for training. + optimizer: + The optimizer we are using. + scheduler: + The learning rate scheduler, we call step() every step. + train_dl: + Dataloader for the training dataset. + valid_dl: + Dataloader for the validation dataset. + scaler: + The scaler used for mix precision training. + model_avg: + The stored model averaged from the start of training. + tb_writer: + Writer to write log messages to tensorboard. + world_size: + Number of nodes in DDP training. If it is 1, DDP is disabled. + rank: + The rank of the node in DDP training. If no DDP is used, it should + be set to 0. + """ + model.train() + + tot_loss = MetricsTracker() + + saved_bad_model = False + + def save_bad_model(suffix: str = ""): + save_checkpoint_impl( + filename=params.exp_dir / f"bad-model{suffix}-{rank}.pt", + model=model, + model_avg=model_avg, + params=params, + optimizer=optimizer, + scheduler=scheduler, + sampler=train_dl.sampler, + scaler=scaler, + rank=0, + ) + + for batch_idx, batch in enumerate(train_dl): + if batch_idx % 10 == 0: + set_batch_count(model, get_adjusted_batch_count(params)) + + params.batch_idx_train += 1 + batch_size = len(batch["supervisions"]["text"]) + + try: + with torch.cuda.amp.autocast(enabled=params.use_fp16): + loss, loss_info = compute_loss( + params=params, + model=model, + sp=sp, + batch=batch, + is_training=True, + ) + # summary stats + tot_loss = (tot_loss * (1 - 1 / params.reset_interval)) + loss_info + + # NOTE: We use reduction==sum and loss is computed over utterances + # in the batch and there is no normalization to it so far. + scaler.scale(loss).backward() + scheduler.step_batch(params.batch_idx_train) + + scaler.step(optimizer) + scaler.update() + optimizer.zero_grad() + except: # noqa + save_bad_model() + display_and_save_batch(batch, params=params, sp=sp) + raise + + if params.print_diagnostics and batch_idx == 5: + return + + if ( + rank == 0 + and params.batch_idx_train > 0 + and params.batch_idx_train % params.average_period == 0 + ): + update_averaged_model( + params=params, + model_cur=model, + model_avg=model_avg, + ) + + if ( + params.batch_idx_train > 0 + and params.batch_idx_train % params.save_every_n == 0 + ): + save_checkpoint_with_global_batch_idx( + out_dir=params.exp_dir, + global_batch_idx=params.batch_idx_train, + model=model, + model_avg=model_avg, + params=params, + optimizer=optimizer, + scheduler=scheduler, + sampler=train_dl.sampler, + scaler=scaler, + rank=rank, + ) + remove_checkpoints( + out_dir=params.exp_dir, + topk=params.keep_last_k, + rank=rank, + ) + + if batch_idx % 100 == 0 and params.use_fp16: + # If the grad scale was less than 1, try increasing it. The _growth_interval + # of the grad scaler is configurable, but we can't configure it to have different + # behavior depending on the current grad scale. + cur_grad_scale = scaler._scale.item() + + if cur_grad_scale < 8.0 or (cur_grad_scale < 32.0 and batch_idx % 400 == 0): + scaler.update(cur_grad_scale * 2.0) + if cur_grad_scale < 0.01: + if not saved_bad_model: + save_bad_model(suffix="-first-warning") + saved_bad_model = True + logging.warning(f"Grad scale is small: {cur_grad_scale}") + if cur_grad_scale < 1.0e-05: + save_bad_model() + raise RuntimeError( + f"grad_scale is too small, exiting: {cur_grad_scale}" + ) + + if batch_idx % params.log_interval == 0: + cur_lr = max(scheduler.get_last_lr()) + cur_grad_scale = scaler._scale.item() if params.use_fp16 else 1.0 + + logging.info( + f"Epoch {params.cur_epoch}, " + f"batch {batch_idx}, loss[{loss_info}], " + f"tot_loss[{tot_loss}], batch size: {batch_size}, " + f"lr: {cur_lr:.2e}, " + + (f"grad_scale: {scaler._scale.item()}" if params.use_fp16 else "") + ) + + if tb_writer is not None: + tb_writer.add_scalar( + "train/learning_rate", cur_lr, params.batch_idx_train + ) + + loss_info.write_summary( + tb_writer, "train/current_", params.batch_idx_train + ) + tot_loss.write_summary(tb_writer, "train/tot_", params.batch_idx_train) + if params.use_fp16: + tb_writer.add_scalar( + "train/grad_scale", cur_grad_scale, params.batch_idx_train + ) + + if batch_idx % params.valid_interval == 0 and not params.print_diagnostics: + logging.info("Computing validation loss") + valid_info = compute_validation_loss( + params=params, + model=model, + sp=sp, + valid_dl=valid_dl, + world_size=world_size, + ) + model.train() + logging.info(f"Epoch {params.cur_epoch}, validation: {valid_info}") + logging.info( + f"Maximum memory allocated so far is {torch.cuda.max_memory_allocated()//1000000}MB" + ) + if tb_writer is not None: + valid_info.write_summary( + tb_writer, "train/valid_", params.batch_idx_train + ) + + loss_value = tot_loss["loss"] / tot_loss["frames"] + params.train_loss = loss_value + if params.train_loss < params.best_train_loss: + params.best_train_epoch = params.cur_epoch + params.best_train_loss = params.train_loss + + +def run(rank, world_size, args): + """ + Args: + rank: + It is a value between 0 and `world_size-1`, which is + passed automatically by `mp.spawn()` in :func:`main`. + The node with rank 0 is responsible for saving checkpoint. + world_size: + Number of GPUs for DDP training. + args: + The return value of get_parser().parse_args() + """ + params = get_params() + params.update(vars(args)) + + fix_random_seed(params.seed) + if world_size > 1: + setup_dist(rank, world_size, params.master_port) + + setup_logger(f"{params.exp_dir}/log/log-train") + logging.info("Training started") + + if args.tensorboard and rank == 0: + tb_writer = SummaryWriter(log_dir=f"{params.exp_dir}/tensorboard") + else: + tb_writer = None + + device = torch.device("cpu") + if torch.cuda.is_available(): + device = torch.device("cuda", rank) + logging.info(f"Device: {device}") + + sp = spm.SentencePieceProcessor() + sp.load(params.bpe_model) + + # is defined in local/train_bpe_model.py + params.blank_id = sp.piece_to_id("") + params.vocab_size = sp.get_piece_size() + + if not params.use_transducer: + params.ctc_loss_scale = 1.0 + + logging.info(params) + + logging.info("About to create model") + model = get_model(params) + + num_param = sum([p.numel() for p in model.parameters()]) + logging.info(f"Number of model parameters: {num_param}") + + assert params.save_every_n >= params.average_period + model_avg: Optional[nn.Module] = None + if rank == 0: + # model_avg is only used with rank 0 + model_avg = copy.deepcopy(model).to(torch.float64) + + assert params.start_epoch > 0, params.start_epoch + checkpoints = load_checkpoint_if_available( + params=params, model=model, model_avg=model_avg + ) + + model.to(device) + if world_size > 1: + logging.info("Using DDP") + model = DDP(model, device_ids=[rank], find_unused_parameters=True) + + optimizer = ScaledAdam( + get_parameter_groups_with_lrs(model, lr=params.base_lr, include_names=True), + lr=params.base_lr, # should have no effect + clipping_scale=2.0, + ) + + scheduler = Eden(optimizer, params.lr_batches, params.lr_epochs) + + if checkpoints and "optimizer" in checkpoints: + logging.info("Loading optimizer state dict") + optimizer.load_state_dict(checkpoints["optimizer"]) + + if ( + checkpoints + and "scheduler" in checkpoints + and checkpoints["scheduler"] is not None + ): + logging.info("Loading scheduler state dict") + scheduler.load_state_dict(checkpoints["scheduler"]) + + if params.print_diagnostics: + opts = diagnostics.TensorDiagnosticOptions( + 2**22 + ) # allow 4 megabytes per sub-module + diagnostic = diagnostics.attach_diagnostics(model, opts) + + if params.inf_check: + register_inf_check_hooks(model) + + data_module = AsrDataModule(args) + multi_dataset = MultiDataset(args.manifest_dir) + + train_cuts = multi_dataset.train_cuts() + + def remove_short_and_long_utt(c: Cut): + # Keep only utterances with duration between 1 second and 20 seconds + # + # Caution: There is a reason to select 20.0 here. Please see + # ../local/display_manifest_statistics.py + # + # You should use ../local/display_manifest_statistics.py to get + # an utterance duration distribution for your dataset to select + # the threshold + if c.duration < 1.0 or c.duration > 20.0: + # logging.warning( + # f"Exclude cut with ID {c.id} from training. Duration: {c.duration}" + # ) + return False + + # In pruned RNN-T, we require that T >= S + # where T is the number of feature frames after subsampling + # and S is the number of tokens in the utterance + + # In ./zipformer.py, the conv module uses the following expression + # for subsampling + T = ((c.num_frames - 7) // 2 + 1) // 2 + tokens = sp.encode(c.supervisions[0].text, out_type=str) + + if T < len(tokens): + logging.warning( + f"Exclude cut with ID {c.id} from training. " + f"Number of frames (before subsampling): {c.num_frames}. " + f"Number of frames (after subsampling): {T}. " + f"Text: {c.supervisions[0].text}. " + f"Tokens: {tokens}. " + f"Number of tokens: {len(tokens)}" + ) + return False + + return True + + train_cuts = train_cuts.filter(remove_short_and_long_utt) + + if params.start_batch > 0 and checkpoints and "sampler" in checkpoints: + # We only load the sampler's state dict when it loads a checkpoint + # saved in the middle of an epoch + sampler_state_dict = checkpoints["sampler"] + else: + sampler_state_dict = None + + train_dl = data_module.train_dataloaders( + train_cuts, sampler_state_dict=sampler_state_dict + ) + + valid_cuts = multi_dataset.dev_cuts() + valid_dl = data_module.valid_dataloaders(valid_cuts) + + if not params.print_diagnostics: + scan_pessimistic_batches_for_oom( + model=model, + train_dl=train_dl, + optimizer=optimizer, + sp=sp, + params=params, + ) + + scaler = GradScaler(enabled=params.use_fp16, init_scale=1.0) + if checkpoints and "grad_scaler" in checkpoints: + logging.info("Loading grad scaler state dict") + scaler.load_state_dict(checkpoints["grad_scaler"]) + + for epoch in range(params.start_epoch, params.num_epochs + 1): + scheduler.step_epoch(epoch - 1) + fix_random_seed(params.seed + epoch - 1) + train_dl.sampler.set_epoch(epoch - 1) + + if tb_writer is not None: + tb_writer.add_scalar("train/epoch", epoch, params.batch_idx_train) + + params.cur_epoch = epoch + + train_one_epoch( + params=params, + model=model, + model_avg=model_avg, + optimizer=optimizer, + scheduler=scheduler, + sp=sp, + train_dl=train_dl, + valid_dl=valid_dl, + scaler=scaler, + tb_writer=tb_writer, + world_size=world_size, + rank=rank, + ) + + if params.print_diagnostics: + diagnostic.print_diagnostics() + break + + save_checkpoint( + params=params, + model=model, + model_avg=model_avg, + optimizer=optimizer, + scheduler=scheduler, + sampler=train_dl.sampler, + scaler=scaler, + rank=rank, + ) + + logging.info("Done!") + + if world_size > 1: + torch.distributed.barrier() + cleanup_dist() + + +def display_and_save_batch( + batch: dict, + params: AttributeDict, + sp: spm.SentencePieceProcessor, +) -> None: + """Display the batch statistics and save the batch into disk. + + Args: + batch: + A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()` + for the content in it. + params: + Parameters for training. See :func:`get_params`. + sp: + The BPE model. + """ + from lhotse.utils import uuid4 + + filename = f"{params.exp_dir}/batch-{uuid4()}.pt" + logging.info(f"Saving batch to {filename}") + torch.save(batch, filename) + + supervisions = batch["supervisions"] + features = batch["inputs"] + + logging.info(f"features shape: {features.shape}") + + y = sp.encode(supervisions["text"], out_type=int) + num_tokens = sum(len(i) for i in y) + logging.info(f"num tokens: {num_tokens}") + + +def scan_pessimistic_batches_for_oom( + model: Union[nn.Module, DDP], + train_dl: torch.utils.data.DataLoader, + optimizer: torch.optim.Optimizer, + sp: spm.SentencePieceProcessor, + params: AttributeDict, +): + from lhotse.dataset import find_pessimistic_batches + + logging.info( + "Sanity check -- see if any of the batches in epoch 1 would cause OOM." + ) + batches, crit_values = find_pessimistic_batches(train_dl.sampler) + for criterion, cuts in batches.items(): + batch = train_dl.dataset[cuts] + try: + with torch.cuda.amp.autocast(enabled=params.use_fp16): + loss, _ = compute_loss( + params=params, + model=model, + sp=sp, + batch=batch, + is_training=True, + ) + loss.backward() + optimizer.zero_grad() + except Exception as e: + if "CUDA out of memory" in str(e): + logging.error( + "Your GPU ran out of memory with the current " + "max_duration setting. We recommend decreasing " + "max_duration and trying again.\n" + f"Failing criterion: {criterion} " + f"(={crit_values[criterion]}) ..." + ) + display_and_save_batch(batch, params=params, sp=sp) + raise + logging.info( + f"Maximum memory allocated so far is {torch.cuda.max_memory_allocated()//1000000}MB" + ) + + +def main(): + parser = get_parser() + AsrDataModule.add_arguments(parser) + args = parser.parse_args() + args.exp_dir = Path(args.exp_dir) + + world_size = args.world_size + assert world_size >= 1 + if world_size > 1: + mp.spawn(run, args=(world_size, args), nprocs=world_size, join=True) + else: + run(rank=0, world_size=1, args=args) + + +torch.set_num_threads(1) +torch.set_num_interop_threads(1) + +if __name__ == "__main__": + main() diff --git a/egs/multi_zh-hans/ASR/zipformer/zipformer.py b/egs/multi_zh-hans/ASR/zipformer/zipformer.py new file mode 120000 index 000000000..23011dda7 --- /dev/null +++ b/egs/multi_zh-hans/ASR/zipformer/zipformer.py @@ -0,0 +1 @@ +../../../librispeech/ASR/zipformer/zipformer.py \ No newline at end of file diff --git a/egs/tal_csasr/ASR/pruned_transducer_stateless5/asr_datamodule.py b/egs/tal_csasr/ASR/pruned_transducer_stateless5/asr_datamodule.py index 2240c1c1d..39beffdcf 100644 --- a/egs/tal_csasr/ASR/pruned_transducer_stateless5/asr_datamodule.py +++ b/egs/tal_csasr/ASR/pruned_transducer_stateless5/asr_datamodule.py @@ -30,7 +30,7 @@ from lhotse.dataset import ( # noqa F401 for PrecomputedFeatures DynamicBucketingSampler, K2SpeechRecognitionDataset, PrecomputedFeatures, - SingleCutSampler, + SimpleCutSampler, SpecAugment, ) from lhotse.dataset.input_strategies import ( # noqa F401 for AudioSamples @@ -311,8 +311,8 @@ class TAL_CSASRAsrDataModule: drop_last=self.args.drop_last, ) else: - logging.info("Using SingleCutSampler.") - train_sampler = SingleCutSampler( + logging.info("Using SimpleCutSampler.") + train_sampler = SimpleCutSampler( cuts_train, max_duration=self.args.max_duration, shuffle=self.args.shuffle, diff --git a/egs/tal_csasr/ASR/pruned_transducer_stateless5/decode.py b/egs/tal_csasr/ASR/pruned_transducer_stateless5/decode.py index 3bfb832fb..3485d4005 100755 --- a/egs/tal_csasr/ASR/pruned_transducer_stateless5/decode.py +++ b/egs/tal_csasr/ASR/pruned_transducer_stateless5/decode.py @@ -724,12 +724,12 @@ def main(): ) save_results( params=params, - test_set_name=test_set, + test_set_name=test_set + "-zh", results_dict=zh_results_dict, ) save_results( params=params, - test_set_name=test_set, + test_set_name=test_set + "-en", results_dict=en_results_dict, ) diff --git a/egs/tedlium3/ASR/transducer_stateless/asr_datamodule.py b/egs/tedlium3/ASR/transducer_stateless/asr_datamodule.py index c647392f0..28d0d3826 100644 --- a/egs/tedlium3/ASR/transducer_stateless/asr_datamodule.py +++ b/egs/tedlium3/ASR/transducer_stateless/asr_datamodule.py @@ -28,7 +28,7 @@ from lhotse.dataset import ( CutMix, DynamicBucketingSampler, K2SpeechRecognitionDataset, - SingleCutSampler, + SimpleCutSampler, SpecAugment, ) from lhotse.dataset.input_strategies import OnTheFlyFeatures @@ -259,8 +259,8 @@ class TedLiumAsrDataModule: drop_last=True, ) else: - logging.info("Using SingleCutSampler.") - train_sampler = SingleCutSampler( + logging.info("Using SimpleCutSampler.") + train_sampler = SimpleCutSampler( cuts_train, max_duration=self.args.max_duration, shuffle=self.args.shuffle, @@ -282,7 +282,6 @@ class TedLiumAsrDataModule: return train_dl def valid_dataloaders(self, cuts_valid: CutSet) -> DataLoader: - transforms = [] if self.args.concatenate_cuts: transforms = [ @@ -322,7 +321,6 @@ class TedLiumAsrDataModule: return valid_dl def test_dataloaders(self, cuts_test: CutSet) -> DataLoader: - logging.debug("About to create test dataset") if self.args.on_the_fly_feats: test = K2SpeechRecognitionDataset( diff --git a/egs/timit/ASR/tdnn_lstm_ctc/asr_datamodule.py b/egs/timit/ASR/tdnn_lstm_ctc/asr_datamodule.py index 51ca4cc6e..7c299d601 100644 --- a/egs/timit/ASR/tdnn_lstm_ctc/asr_datamodule.py +++ b/egs/timit/ASR/tdnn_lstm_ctc/asr_datamodule.py @@ -30,7 +30,7 @@ from lhotse.dataset import ( DynamicBucketingSampler, K2SpeechRecognitionDataset, PrecomputedFeatures, - SingleCutSampler, + SimpleCutSampler, SpecAugment, ) from lhotse.dataset.input_strategies import OnTheFlyFeatures @@ -225,8 +225,8 @@ class TimitAsrDataModule(DataModule): drop_last=True, ) else: - logging.info("Using SingleCutSampler.") - train_sampler = SingleCutSampler( + logging.info("Using SimpleCutSampler.") + train_sampler = SimpleCutSampler( cuts_train, max_duration=self.args.max_duration, shuffle=self.args.shuffle, @@ -267,7 +267,7 @@ class TimitAsrDataModule(DataModule): cut_transforms=transforms, return_cuts=self.args.return_cuts, ) - valid_sampler = SingleCutSampler( + valid_sampler = SimpleCutSampler( cuts_valid, max_duration=self.args.max_duration, shuffle=False, @@ -298,7 +298,7 @@ class TimitAsrDataModule(DataModule): else PrecomputedFeatures(), return_cuts=self.args.return_cuts, ) - sampler = SingleCutSampler(cuts_test, max_duration=self.args.max_duration) + sampler = SimpleCutSampler(cuts_test, max_duration=self.args.max_duration) logging.debug("About to create test dataloader") test_dl = DataLoader(test, batch_size=None, sampler=sampler, num_workers=1) test_loaders.append(test_dl) diff --git a/egs/wenetspeech/ASR/local/compute_fbank_wenetspeech_dev_test.py b/egs/wenetspeech/ASR/local/compute_fbank_wenetspeech_dev_test.py index 20d7341db..1af08fee2 100755 --- a/egs/wenetspeech/ASR/local/compute_fbank_wenetspeech_dev_test.py +++ b/egs/wenetspeech/ASR/local/compute_fbank_wenetspeech_dev_test.py @@ -28,6 +28,7 @@ from lhotse import CutSet, KaldifeatFbank, KaldifeatFbankConfig, LilcomChunkyWri # even when we are not invoking the main (e.g. when spawning subprocesses). torch.set_num_threads(1) torch.set_num_interop_threads(1) +torch.multiprocessing.set_sharing_strategy("file_system") def compute_fbank_wenetspeech_dev_test(): diff --git a/egs/wenetspeech/ASR/local/compute_fbank_wenetspeech_splits.py b/egs/wenetspeech/ASR/local/compute_fbank_wenetspeech_splits.py index 1b257fb70..99d39bbdc 100755 --- a/egs/wenetspeech/ASR/local/compute_fbank_wenetspeech_splits.py +++ b/egs/wenetspeech/ASR/local/compute_fbank_wenetspeech_splits.py @@ -37,6 +37,7 @@ from lhotse import ( # even when we are not invoking the main (e.g. when spawning subprocesses). torch.set_num_threads(1) torch.set_num_interop_threads(1) +torch.multiprocessing.set_sharing_strategy("file_system") def get_parser(): diff --git a/egs/wenetspeech/ASR/pruned_transducer_stateless2/asr_datamodule.py b/egs/wenetspeech/ASR/pruned_transducer_stateless2/asr_datamodule.py index 746b212ff..c5967f10a 100644 --- a/egs/wenetspeech/ASR/pruned_transducer_stateless2/asr_datamodule.py +++ b/egs/wenetspeech/ASR/pruned_transducer_stateless2/asr_datamodule.py @@ -37,7 +37,7 @@ from lhotse.dataset import ( DynamicBucketingSampler, K2SpeechRecognitionDataset, PrecomputedFeatures, - SingleCutSampler, + SimpleCutSampler, SpecAugment, ) from lhotse.dataset.input_strategies import OnTheFlyFeatures @@ -296,8 +296,8 @@ class WenetSpeechAsrDataModule: drop_last=True, ) else: - logging.info("Using SingleCutSampler.") - train_sampler = SingleCutSampler( + logging.info("Using SimpleCutSampler.") + train_sampler = SimpleCutSampler( cuts_train, max_duration=self.args.max_duration, shuffle=self.args.shuffle, diff --git a/egs/xbmu_amdo31/ASR/pruned_transducer_stateless5/asr_datamodule.py b/egs/xbmu_amdo31/ASR/pruned_transducer_stateless5/asr_datamodule.py index 55d5f4636..6362ab7cd 100644 --- a/egs/xbmu_amdo31/ASR/pruned_transducer_stateless5/asr_datamodule.py +++ b/egs/xbmu_amdo31/ASR/pruned_transducer_stateless5/asr_datamodule.py @@ -32,7 +32,7 @@ from lhotse.dataset import ( DynamicBucketingSampler, K2SpeechRecognitionDataset, PrecomputedFeatures, - SingleCutSampler, + SimpleCutSampler, SpecAugment, ) from lhotse.dataset.input_strategies import AudioSamples # noqa F401 For AudioSamples @@ -299,8 +299,8 @@ class Xbmu_AmdoAsrDataModule: drop_last=self.args.drop_last, ) else: - logging.info("Using SingleCutSampler.") - train_sampler = SingleCutSampler( + logging.info("Using SimpleCutSampler.") + train_sampler = SimpleCutSampler( cuts_train, max_duration=self.args.max_duration, shuffle=self.args.shuffle, diff --git a/egs/yesno/ASR/tdnn/asr_datamodule.py b/egs/yesno/ASR/tdnn/asr_datamodule.py index ada8c1a6c..dc66b217d 100644 --- a/egs/yesno/ASR/tdnn/asr_datamodule.py +++ b/egs/yesno/ASR/tdnn/asr_datamodule.py @@ -26,7 +26,7 @@ from lhotse.dataset import ( DynamicBucketingSampler, K2SpeechRecognitionDataset, PrecomputedFeatures, - SingleCutSampler, + SimpleCutSampler, ) from lhotse.dataset.input_strategies import OnTheFlyFeatures from torch.utils.data import DataLoader @@ -196,8 +196,8 @@ class YesNoAsrDataModule(DataModule): drop_last=True, ) else: - logging.info("Using SingleCutSampler.") - train_sampler = SingleCutSampler( + logging.info("Using SimpleCutSampler.") + train_sampler = SimpleCutSampler( cuts_train, max_duration=self.args.max_duration, shuffle=self.args.shuffle, diff --git a/icefall/utils.py b/icefall/utils.py index b01cd2770..947d79438 100644 --- a/icefall/utils.py +++ b/icefall/utils.py @@ -493,6 +493,7 @@ def write_error_stats( test_set_name: str, results: List[Tuple[str, str]], enable_log: bool = True, + sclite_mode: bool = False, ) -> float: """Write statistics based on predicted results and reference transcripts. @@ -538,7 +539,7 @@ def write_error_stats( num_corr = 0 ERR = "*" for cut_id, ref, hyp in results: - ali = kaldialign.align(ref, hyp, ERR) + ali = kaldialign.align(ref, hyp, ERR, sclite_mode=sclite_mode) for ref_word, hyp_word in ali: if ref_word == ERR: ins[hyp_word] += 1 diff --git a/requirements-ci.txt b/requirements-ci.txt index 3c2eb5f65..2433e190b 100644 --- a/requirements-ci.txt +++ b/requirements-ci.txt @@ -10,12 +10,13 @@ graphviz==0.19.1 -f https://download.pytorch.org/whl/cpu/torch_stable.html torch==1.13.1+cpu -f https://download.pytorch.org/whl/cpu/torch_stable.html torchaudio==0.13.1+cpu +six -f https://k2-fsa.org/nightly/ k2==1.23.4.dev20230319+cpu.torch1.13.1 git+https://github.com/lhotse-speech/lhotse kaldilm==1.11 -kaldialign==0.2 +kaldialign==0.7.1 sentencepiece==0.1.96 tensorboard==2.8.0 typeguard==2.13.3 diff --git a/test/test_ali.py b/test/test_ali.py index b107a6d80..d607e40aa 100755 --- a/test/test_ali.py +++ b/test/test_ali.py @@ -26,7 +26,7 @@ from pathlib import Path from lhotse import CutSet, load_manifest -from lhotse.dataset import K2SpeechRecognitionDataset, SingleCutSampler +from lhotse.dataset import K2SpeechRecognitionDataset, SimpleCutSampler from lhotse.dataset.collation import collate_custom_field from torch.utils.data import DataLoader @@ -44,7 +44,7 @@ def get_dataloader(): cuts = load_manifest(cuts_json) print(cuts[0]) cuts = cuts.with_features_path_prefix(egs_dir) - sampler = SingleCutSampler( + sampler = SimpleCutSampler( cuts, max_duration=10, shuffle=False,