From f3ad32777a598de6169274198e418ee95fbf6ddc Mon Sep 17 00:00:00 2001 From: Zengwei Yao Date: Thu, 29 Sep 2022 11:15:43 +0800 Subject: [PATCH] Gradient filter for training lstm model (#564) * init files * add gradient filter module * refact getting median value * add cutoff for grad filter * delete comments * apply gradient filter in LSTM module, to filter both input and params * fix typing and refactor * filter with soft mask * rename lstm_transducer_stateless2 to lstm_transducer_stateless3 * fix typos, and update RESULTS.md * minor fix * fix return typing * fix typo --- .flake8 | 2 +- egs/librispeech/ASR/RESULTS.md | 134 +- .../lstm_transducer_stateless3/__init__.py | 1 + .../asr_datamodule.py | 1 + .../lstm_transducer_stateless3/beam_search.py | 1 + .../ASR/lstm_transducer_stateless3/decode.py | 818 ++++++++++++ .../ASR/lstm_transducer_stateless3/decoder.py | 1 + .../encoder_interface.py | 1 + .../ASR/lstm_transducer_stateless3/export.py | 388 ++++++ .../jit_pretrained.py | 322 +++++ .../ASR/lstm_transducer_stateless3/joiner.py | 1 + .../ASR/lstm_transducer_stateless3/lstm.py | 860 +++++++++++++ .../ASR/lstm_transducer_stateless3/model.py | 1 + .../ASR/lstm_transducer_stateless3/optim.py | 1 + .../lstm_transducer_stateless3/pretrained.py | 352 +++++ .../ASR/lstm_transducer_stateless3/scaling.py | 1 + .../scaling_converter.py | 1 + .../ASR/lstm_transducer_stateless3/stream.py | 1 + .../streaming_decode.py | 968 ++++++++++++++ .../lstm_transducer_stateless3/test_model.py | 92 ++ .../test_scaling_converter.py | 257 ++++ .../ASR/lstm_transducer_stateless3/train.py | 1138 +++++++++++++++++ .../pruned_transducer_stateless2/scaling.py | 134 +- 23 files changed, 5448 insertions(+), 28 deletions(-) create mode 120000 egs/librispeech/ASR/lstm_transducer_stateless3/__init__.py create mode 120000 egs/librispeech/ASR/lstm_transducer_stateless3/asr_datamodule.py create mode 120000 egs/librispeech/ASR/lstm_transducer_stateless3/beam_search.py create mode 100755 egs/librispeech/ASR/lstm_transducer_stateless3/decode.py create mode 120000 egs/librispeech/ASR/lstm_transducer_stateless3/decoder.py create mode 120000 egs/librispeech/ASR/lstm_transducer_stateless3/encoder_interface.py create mode 100755 egs/librispeech/ASR/lstm_transducer_stateless3/export.py create mode 100755 egs/librispeech/ASR/lstm_transducer_stateless3/jit_pretrained.py create mode 120000 egs/librispeech/ASR/lstm_transducer_stateless3/joiner.py create mode 100644 egs/librispeech/ASR/lstm_transducer_stateless3/lstm.py create mode 120000 egs/librispeech/ASR/lstm_transducer_stateless3/model.py create mode 120000 egs/librispeech/ASR/lstm_transducer_stateless3/optim.py create mode 100755 egs/librispeech/ASR/lstm_transducer_stateless3/pretrained.py create mode 120000 egs/librispeech/ASR/lstm_transducer_stateless3/scaling.py create mode 120000 egs/librispeech/ASR/lstm_transducer_stateless3/scaling_converter.py create mode 120000 egs/librispeech/ASR/lstm_transducer_stateless3/stream.py create mode 100755 egs/librispeech/ASR/lstm_transducer_stateless3/streaming_decode.py create mode 100755 egs/librispeech/ASR/lstm_transducer_stateless3/test_model.py create mode 100644 egs/librispeech/ASR/lstm_transducer_stateless3/test_scaling_converter.py create mode 100755 egs/librispeech/ASR/lstm_transducer_stateless3/train.py diff --git a/.flake8 b/.flake8 index 22cd63b3d..609fa2c03 100644 --- a/.flake8 +++ b/.flake8 @@ -9,7 +9,7 @@ per-file-ignores = egs/*/ASR/pruned_transducer_stateless*/*.py: E501, egs/*/ASR/*/optim.py: E501, egs/*/ASR/*/scaling.py: E501, - egs/librispeech/ASR/lstm_transducer_stateless/*.py: E501, E203 + egs/librispeech/ASR/lstm_transducer_stateless*/*.py: E501, E203 egs/librispeech/ASR/conv_emformer_transducer_stateless*/*.py: E501, E203 egs/librispeech/ASR/conformer_ctc2/*py: E501, egs/librispeech/ASR/RESULTS.md: E999, diff --git a/egs/librispeech/ASR/RESULTS.md b/egs/librispeech/ASR/RESULTS.md index 8a27b4b63..d5a67b619 100644 --- a/egs/librispeech/ASR/RESULTS.md +++ b/egs/librispeech/ASR/RESULTS.md @@ -1,12 +1,100 @@ ## Results -#### LibriSpeech BPE training results (Pruned Stateless LSTM RNN-T + multi-dataset) +### LibriSpeech BPE training results (Pruned Stateless LSTM RNN-T + gradient filter) -[lstm_transducer_stateless2](./lstm_transducer_stateless2) +#### [lstm_transducer_stateless3](./lstm_transducer_stateless3) + +It implements LSTM model with mechanisms in reworked model for streaming ASR. +Gradient filter is applied inside each lstm module to stabilize the training. + +See for more details. + +##### training on full librispeech + +This model contains 12 encoder layers (LSTM module + Feedforward module). The number of model parameters is 84689496. + +The WERs are: + +| | test-clean | test-other | comment | decoding mode | +|-------------------------------------|------------|------------|----------------------|----------------------| +| greedy search (max sym per frame 1) | 3.66 | 9.51 | --epoch 40 --avg 15 | simulated streaming | +| greedy search (max sym per frame 1) | 3.66 | 9.48 | --epoch 40 --avg 15 | streaming | +| fast beam search | 3.55 | 9.33 | --epoch 40 --avg 15 | simulated streaming | +| fast beam search | 3.57 | 9.25 | --epoch 40 --avg 15 | streaming | +| modified beam search | 3.55 | 9.28 | --epoch 40 --avg 15 | simulated streaming | +| modified beam search | 3.54 | 9.25 | --epoch 40 --avg 15 | streaming | + +Note: `simulated streaming` indicates feeding full utterance during decoding, while `streaming` indicates feeding certain number of frames at each time. + + +The training command is: + +```bash +./lstm_transducer_stateless3/train.py \ + --world-size 4 \ + --num-epochs 40 \ + --start-epoch 1 \ + --exp-dir lstm_transducer_stateless3/exp \ + --full-libri 1 \ + --max-duration 500 \ + --master-port 12325 \ + --num-encoder-layers 12 \ + --grad-norm-threshold 25.0 \ + --rnn-hidden-size 1024 +``` + +The tensorboard log can be found at + + +The simulated streaming decoding command using greedy search, fast beam search, and modified beam search is: +```bash +for decoding_method in greedy_search fast_beam_search modified_beam_search; do + ./lstm_transducer_stateless3/decode.py \ + --epoch 40 \ + --avg 15 \ + --exp-dir lstm_transducer_stateless3/exp \ + --max-duration 600 \ + --num-encoder-layers 12 \ + --rnn-hidden-size 1024 \ + --decoding-method $decoding_method \ + --use-averaged-model True \ + --beam 4 \ + --max-contexts 4 \ + --max-states 8 \ + --beam-size 4 +done +``` + +The streaming decoding command using greedy search, fast beam search, and modified beam search is: +```bash +for decoding_method in greedy_search fast_beam_search modified_beam_search; do + ./lstm_transducer_stateless3/streaming_decode.py \ + --epoch 40 \ + --avg 15 \ + --exp-dir lstm_transducer_stateless3/exp \ + --max-duration 600 \ + --num-encoder-layers 12 \ + --rnn-hidden-size 1024 \ + --decoding-method $decoding_method \ + --use-averaged-model True \ + --beam 4 \ + --max-contexts 4 \ + --max-states 8 \ + --beam-size 4 +done +``` + +Pretrained models, training logs, decoding logs, and decoding results +are available at + + + +### LibriSpeech BPE training results (Pruned Stateless LSTM RNN-T + multi-dataset) + +#### [lstm_transducer_stateless2](./lstm_transducer_stateless2) See for more details. - The WERs are: | | test-clean | test-other | comment | @@ -18,6 +106,7 @@ The WERs are: | modified_beam_search | 2.75 | 7.08 | --iter 472000 --avg 18 | | fast_beam_search | 2.77 | 7.29 | --iter 472000 --avg 18 | + The training command is: ```bash @@ -70,15 +159,16 @@ Pretrained models, training logs, decoding logs, and decoding results are available at -#### LibriSpeech BPE training results (Pruned Stateless LSTM RNN-T) -[lstm_transducer_stateless](./lstm_transducer_stateless) +### LibriSpeech BPE training results (Pruned Stateless LSTM RNN-T) + +#### [lstm_transducer_stateless](./lstm_transducer_stateless) It implements LSTM model with mechanisms in reworked model for streaming ASR. See for more details. -#### training on full librispeech +##### training on full librispeech This model contains 12 encoder layers (LSTM module + Feedforward module). The number of model parameters is 84689496. @@ -165,7 +255,7 @@ It is modified from [torchaudio](https://github.com/pytorch/audio). See for more details. -#### With lower latency setup, training on full librispeech +##### With lower latency setup, training on full librispeech In this model, the lengths of chunk and right context are 32 frames (i.e., 0.32s) and 8 frames (i.e., 0.08s), respectively. @@ -316,7 +406,7 @@ Pretrained models, training logs, decoding logs, and decoding results are available at -#### With higher latency setup, training on full librispeech +##### With higher latency setup, training on full librispeech In this model, the lengths of chunk and right context are 64 frames (i.e., 0.64s) and 16 frames (i.e., 0.16s), respectively. @@ -851,14 +941,14 @@ Pre-trained models, training and decoding logs, and decoding results are availab ### LibriSpeech BPE training results (Pruned Stateless Conv-Emformer RNN-T) -[conv_emformer_transducer_stateless](./conv_emformer_transducer_stateless) +#### [conv_emformer_transducer_stateless](./conv_emformer_transducer_stateless) It implements [Emformer](https://arxiv.org/abs/2010.10759) augmented with convolution module for streaming ASR. It is modified from [torchaudio](https://github.com/pytorch/audio). See for more details. -#### Training on full librispeech +##### Training on full librispeech In this model, the lengths of chunk and right context are 32 frames (i.e., 0.32s) and 8 frames (i.e., 0.08s), respectively. @@ -1011,7 +1101,7 @@ are available at ### LibriSpeech BPE training results (Pruned Stateless Emformer RNN-T) -[pruned_stateless_emformer_rnnt2](./pruned_stateless_emformer_rnnt2) +#### [pruned_stateless_emformer_rnnt2](./pruned_stateless_emformer_rnnt2) Use . @@ -1079,7 +1169,7 @@ results at: ### LibriSpeech BPE training results (Pruned Stateless Transducer 5) -[pruned_transducer_stateless5](./pruned_transducer_stateless5) +#### [pruned_transducer_stateless5](./pruned_transducer_stateless5) Same as `Pruned Stateless Transducer 2` but with more layers. @@ -1092,7 +1182,7 @@ The notations `large` and `medium` below are from the [Conformer](https://arxiv. paper, where the large model has about 118 M parameters and the medium model has 30.8 M parameters. -#### Large +##### Large Number of model parameters 118129516 (i.e, 118.13 M). @@ -1152,7 +1242,7 @@ results at: -#### Medium +##### Medium Number of model parameters 30896748 (i.e, 30.9 M). @@ -1212,7 +1302,7 @@ results at: -#### Baseline-2 +##### Baseline-2 It has 88.98 M parameters. Compared to the model in pruned_transducer_stateless2, its has more layers (24 v.s 12) but a narrower model (1536 feedforward dim and 384 encoder dim vs 2048 feed forward dim and 512 encoder dim). @@ -1273,13 +1363,13 @@ results at: ### LibriSpeech BPE training results (Pruned Stateless Transducer 4) -[pruned_transducer_stateless4](./pruned_transducer_stateless4) +#### [pruned_transducer_stateless4](./pruned_transducer_stateless4) This version saves averaged model during training, and decodes with averaged model. See for details about the idea of model averaging. -#### Training on full librispeech +##### Training on full librispeech See @@ -1355,7 +1445,7 @@ Pretrained models, training logs, decoding logs, and decoding results are available at -#### Training on train-clean-100 +##### Training on train-clean-100 See @@ -1392,7 +1482,7 @@ The tensorboard log can be found at ### LibriSpeech BPE training results (Pruned Stateless Transducer 3, 2022-04-29) -[pruned_transducer_stateless3](./pruned_transducer_stateless3) +#### [pruned_transducer_stateless3](./pruned_transducer_stateless3) Same as `Pruned Stateless Transducer 2` but using the XL subset from [GigaSpeech](https://github.com/SpeechColab/GigaSpeech) as extra training data. @@ -1606,10 +1696,10 @@ can be found at ### LibriSpeech BPE training results (Pruned Transducer 2) -[pruned_transducer_stateless2](./pruned_transducer_stateless2) +#### [pruned_transducer_stateless2](./pruned_transducer_stateless2) This is with a reworked version of the conformer encoder, with many changes. -#### Training on fulll librispeech +##### Training on full librispeech Using commit `34aad74a2c849542dd5f6359c9e6b527e8782fd6`. See @@ -1658,7 +1748,7 @@ can be found at -#### Training on train-clean-100: +##### Training on train-clean-100: Trained with 1 job: ``` diff --git a/egs/librispeech/ASR/lstm_transducer_stateless3/__init__.py b/egs/librispeech/ASR/lstm_transducer_stateless3/__init__.py new file mode 120000 index 000000000..b24e5e357 --- /dev/null +++ b/egs/librispeech/ASR/lstm_transducer_stateless3/__init__.py @@ -0,0 +1 @@ +../pruned_transducer_stateless2/__init__.py \ No newline at end of file diff --git a/egs/librispeech/ASR/lstm_transducer_stateless3/asr_datamodule.py b/egs/librispeech/ASR/lstm_transducer_stateless3/asr_datamodule.py new file mode 120000 index 000000000..a074d6085 --- /dev/null +++ b/egs/librispeech/ASR/lstm_transducer_stateless3/asr_datamodule.py @@ -0,0 +1 @@ +../pruned_transducer_stateless2/asr_datamodule.py \ No newline at end of file diff --git a/egs/librispeech/ASR/lstm_transducer_stateless3/beam_search.py b/egs/librispeech/ASR/lstm_transducer_stateless3/beam_search.py new file mode 120000 index 000000000..8554e44cc --- /dev/null +++ b/egs/librispeech/ASR/lstm_transducer_stateless3/beam_search.py @@ -0,0 +1 @@ +../pruned_transducer_stateless2/beam_search.py \ No newline at end of file diff --git a/egs/librispeech/ASR/lstm_transducer_stateless3/decode.py b/egs/librispeech/ASR/lstm_transducer_stateless3/decode.py new file mode 100755 index 000000000..9842bf00d --- /dev/null +++ b/egs/librispeech/ASR/lstm_transducer_stateless3/decode.py @@ -0,0 +1,818 @@ +#!/usr/bin/env python3 +# +# Copyright 2021-2022 Xiaomi Corporation (Author: Fangjun Kuang, +# Zengwei Yao) +# +# See ../../../../LICENSE for clarification regarding multiple authors +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +""" +Usage: +(1) greedy search +./lstm_transducer_stateless3/decode.py \ + --epoch 40 \ + --avg 20 \ + --exp-dir ./lstm_transducer_stateless3/exp \ + --max-duration 600 \ + --decoding-method greedy_search + +(2) beam search (not recommended) +./lstm_transducer_stateless2/decode.py \ + --epoch 40 \ + --avg 20 \ + --exp-dir ./lstm_transducer_stateless3/exp \ + --max-duration 600 \ + --decoding-method beam_search \ + --beam-size 4 + +(3) modified beam search +./lstm_transducer_stateless3/decode.py \ + --epoch 40 \ + --avg 20 \ + --exp-dir ./lstm_transducer_stateless3/exp \ + --max-duration 600 \ + --decoding-method modified_beam_search \ + --beam-size 4 + +(4) fast beam search (one best) +./lstm_transducer_stateless3/decode.py \ + --epoch 40 \ + --avg 20 \ + --exp-dir ./lstm_transducer_stateless3/exp \ + --max-duration 600 \ + --decoding-method fast_beam_search \ + --beam 20.0 \ + --max-contexts 8 \ + --max-states 64 + +(5) fast beam search (nbest) +./lstm_transducer_stateless3/decode.py \ + --epoch 40 \ + --avg 20 \ + --exp-dir ./pruned_transducer_stateless3/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) +./lstm_transducer_stateless3/decode.py \ + --epoch 40 \ + --avg 20 \ + --exp-dir ./lstm_transducer_stateless3/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) +./lstm_transducer_stateless3/decode.py \ + --epoch 40 \ + --avg 20 \ + --exp-dir ./lstm_transducer_stateless3/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 LibriSpeechAsrDataModule +from beam_search import ( + beam_search, + fast_beam_search_nbest, + fast_beam_search_nbest_LG, + fast_beam_search_nbest_oracle, + fast_beam_search_one_best, + greedy_search, + greedy_search_batch, + modified_beam_search, +) +from train import add_model_arguments, get_params, get_transducer_model + +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, + 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="lstm_transducer_stateless/exp", + help="The experiment dir", + ) + + parser.add_argument( + "--bpe-model", + type=str, + default="data/lang_bpe_500/bpe.model", + help="Path to the BPE model", + ) + + parser.add_argument( + "--lang-dir", + type=Path, + default="data/lang_bpe_500", + help="The lang dir containing word table and LG graph", + ) + + parser.add_argument( + "--decoding-method", + type=str, + default="greedy_search", + help="""Possible values are: + - greedy_search + - beam_search + - modified_beam_search + - 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 LG, Used + only when --decoding_method is fast_beam_search, fast_beam_search_nbest, + fast_beam_search_nbest_oracle, and fast_beam_search_nbest_LG. + Returns: + Return the decoding result. See above description for the format of + the returned dict. + """ + device = next(model.parameters()).device + feature = batch["inputs"] + assert feature.ndim == 3 + + feature = feature.to(device) + # at entry, feature is (N, T, C) + + supervisions = batch["supervisions"] + feature_lens = supervisions["num_frames"].to(device) + + # tail padding here to alleviate the tail deletion problem + num_tail_padded_frames = 35 + feature = torch.nn.functional.pad( + feature, + (0, 0, 0, num_tail_padded_frames), + mode="constant", + value=LOG_EPS, + ) + feature_lens += num_tail_padded_frames + + encoder_out, encoder_out_lens, _ = model.encoder( + x=feature, x_lens=feature_lens + ) + + hyps = [] + + if params.decoding_method == "fast_beam_search": + hyp_tokens = fast_beam_search_one_best( + model=model, + decoding_graph=decoding_graph, + encoder_out=encoder_out, + encoder_out_lens=encoder_out_lens, + beam=params.beam, + max_contexts=params.max_contexts, + max_states=params.max_states, + ) + for hyp in sp.decode(hyp_tokens): + hyps.append(hyp.split()) + elif params.decoding_method == "fast_beam_search_nbest_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[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 LG, Used + only when --decoding_method is fast_beam_search, fast_beam_search_nbest, + fast_beam_search_nbest_oracle, and fast_beam_search_nbest_LG. + Returns: + Return a dict, whose key may be "greedy_search" if greedy search + is used, or it may be "beam_7" if beam size of 7 is used. + Its value is a list of tuples. Each tuple contains two elements: + The first is the reference transcript, and the second is the + predicted result. + """ + num_cuts = 0 + + try: + num_batches = len(dl) + except TypeError: + num_batches = "?" + + if params.decoding_method == "greedy_search": + log_interval = 50 + else: + log_interval = 20 + + results = defaultdict(list) + for batch_idx, batch in enumerate(dl): + texts = batch["supervisions"]["text"] + 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): + ref_words = ref_text.split() + this_batch.append((cut_id, ref_words, hyp_words)) + + results[name].extend(this_batch) + + num_cuts += len(texts) + + if batch_idx % log_interval == 0: + batch_str = f"{batch_idx}/{num_batches}" + + logging.info( + f"batch {batch_str}, cuts processed until now is {num_cuts}" + ) + return results + + +def save_results( + params: AttributeDict, + test_set_name: str, + results_dict: Dict[str, List[Tuple[List[int], List[int]]]], +): + test_set_wers = dict() + for key, results in results_dict.items(): + recog_path = ( + params.res_dir / f"recogs-{test_set_name}-{key}-{params.suffix}.txt" + ) + results = sorted(results) + store_transcripts(filename=recog_path, texts=results) + logging.info(f"The transcripts are stored in {recog_path}") + + # The following prints out WERs, per-word error statistics and aligned + # ref/hyp pairs. + errs_filename = ( + params.res_dir / f"errs-{test_set_name}-{key}-{params.suffix}.txt" + ) + with open(errs_filename, "w") as f: + wer = write_error_stats( + f, f"{test_set_name}-{key}", results, enable_log=True + ) + test_set_wers[key] = wer + + logging.info("Wrote detailed error stats to {}".format(errs_filename)) + + test_set_wers = sorted(test_set_wers.items(), key=lambda x: x[1]) + errs_info = ( + params.res_dir + / f"wer-summary-{test_set_name}-{key}-{params.suffix}.txt" + ) + with open(errs_info, "w") as f: + print("settings\tWER", file=f) + for key, val in test_set_wers: + print("{}\t{}".format(key, val), file=f) + + s = "\nFor {}, WER of different settings are:\n".format(test_set_name) + note = "\tbest for {}".format(test_set_name) + for key, val in test_set_wers: + s += "{}\t{}{}\n".format(key, val, note) + note = "" + logging.info(s) + + +@torch.no_grad() +def main(): + parser = get_parser() + LibriSpeechAsrDataModule.add_arguments(parser) + args = parser.parse_args() + args.exp_dir = Path(args.exp_dir) + + params = get_params() + params.update(vars(args)) + + assert params.decoding_method in ( + "greedy_search", + "beam_search", + "fast_beam_search", + "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 "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_transducer_model(params) + + if not params.use_averaged_model: + if params.iter > 0: + filenames = find_checkpoints( + params.exp_dir, iteration=-params.iter + )[: params.avg] + if len(filenames) == 0: + raise ValueError( + f"No checkpoints found for" + f" --iter {params.iter}, --avg {params.avg}" + ) + elif len(filenames) < params.avg: + raise ValueError( + f"Not enough checkpoints ({len(filenames)}) found for" + f" --iter {params.iter}, --avg {params.avg}" + ) + logging.info(f"averaging {filenames}") + model.to(device) + model.load_state_dict(average_checkpoints(filenames, device=device)) + elif params.avg == 1: + load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model) + else: + start = params.epoch - params.avg + 1 + filenames = [] + for i in range(start, params.epoch + 1): + if i >= 1: + filenames.append(f"{params.exp_dir}/epoch-{i}.pt") + logging.info(f"averaging {filenames}") + model.to(device) + model.load_state_dict(average_checkpoints(filenames, device=device)) + else: + if params.iter > 0: + filenames = find_checkpoints( + params.exp_dir, iteration=-params.iter + )[: params.avg + 1] + if len(filenames) == 0: + raise ValueError( + f"No checkpoints found for" + f" --iter {params.iter}, --avg {params.avg}" + ) + elif len(filenames) < params.avg + 1: + raise ValueError( + f"Not enough checkpoints ({len(filenames)}) found for" + f" --iter {params.iter}, --avg {params.avg}" + ) + filename_start = filenames[-1] + filename_end = filenames[0] + logging.info( + "Calculating the averaged model over iteration checkpoints" + f" from {filename_start} (excluded) to {filename_end}" + ) + model.to(device) + model.load_state_dict( + average_checkpoints_with_averaged_model( + filename_start=filename_start, + filename_end=filename_end, + device=device, + ) + ) + else: + assert params.avg > 0, 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 + librispeech = LibriSpeechAsrDataModule(args) + + test_clean_cuts = librispeech.test_clean_cuts() + test_other_cuts = librispeech.test_other_cuts() + + test_clean_dl = librispeech.test_dataloaders(test_clean_cuts) + test_other_dl = librispeech.test_dataloaders(test_other_cuts) + + test_sets = ["test-clean", "test-other"] + test_dl = [test_clean_dl, test_other_dl] + + for test_set, test_dl in zip(test_sets, test_dl): + results_dict = decode_dataset( + dl=test_dl, + params=params, + model=model, + sp=sp, + word_table=word_table, + decoding_graph=decoding_graph, + ) + + save_results( + params=params, + test_set_name=test_set, + results_dict=results_dict, + ) + + logging.info("Done!") + + +if __name__ == "__main__": + main() diff --git a/egs/librispeech/ASR/lstm_transducer_stateless3/decoder.py b/egs/librispeech/ASR/lstm_transducer_stateless3/decoder.py new file mode 120000 index 000000000..0793c5709 --- /dev/null +++ b/egs/librispeech/ASR/lstm_transducer_stateless3/decoder.py @@ -0,0 +1 @@ +../pruned_transducer_stateless2/decoder.py \ No newline at end of file diff --git a/egs/librispeech/ASR/lstm_transducer_stateless3/encoder_interface.py b/egs/librispeech/ASR/lstm_transducer_stateless3/encoder_interface.py new file mode 120000 index 000000000..aa5d0217a --- /dev/null +++ b/egs/librispeech/ASR/lstm_transducer_stateless3/encoder_interface.py @@ -0,0 +1 @@ +../transducer_stateless/encoder_interface.py \ No newline at end of file diff --git a/egs/librispeech/ASR/lstm_transducer_stateless3/export.py b/egs/librispeech/ASR/lstm_transducer_stateless3/export.py new file mode 100755 index 000000000..212c7bad6 --- /dev/null +++ b/egs/librispeech/ASR/lstm_transducer_stateless3/export.py @@ -0,0 +1,388 @@ +#!/usr/bin/env python3 +# +# Copyright 2021-2022 Xiaomi Corporation (Author: Fangjun Kuang, Zengwei Yao) +# +# See ../../../../LICENSE for clarification regarding multiple authors +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +# This script converts several saved checkpoints +# to a single one using model averaging. +""" + +Usage: + +(1) Export to torchscript model using torch.jit.trace() + +./lstm_transducer_stateless3/export.py \ + --exp-dir ./lstm_transducer_stateless3/exp \ + --bpe-model data/lang_bpe_500/bpe.model \ + --epoch 40 \ + --avg 20 \ + --jit-trace 1 + +It will generate 3 files: `encoder_jit_trace.pt`, +`decoder_jit_trace.pt`, and `joiner_jit_trace.pt`. + +(2) Export `model.state_dict()` + +./lstm_transducer_stateless3/export.py \ + --exp-dir ./lstm_transducer_stateless3/exp \ + --bpe-model data/lang_bpe_500/bpe.model \ + --epoch 40 \ + --avg 20 + +It will generate a file `pretrained.pt` in the given `exp_dir`. You can later +load it by `icefall.checkpoint.load_checkpoint()`. + +To use the generated file with `lstm_transducer_stateless3/decode.py`, +you can do: + + cd /path/to/exp_dir + ln -s pretrained.pt epoch-9999.pt + + cd /path/to/egs/librispeech/ASR + ./lstm_transducer_stateless3/decode.py \ + --exp-dir ./lstm_transducer_stateless3/exp \ + --epoch 9999 \ + --avg 1 \ + --max-duration 600 \ + --decoding-method greedy_search \ + --bpe-model data/lang_bpe_500/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 + +https://huggingface.co/Zengwei/icefall-asr-librispeech-lstm-transducer-stateless-2022-08-18 + +with the following commands: + + sudo apt-get install git-lfs + git lfs install + git clone https://huggingface.co/Zengwei/icefall-asr-librispeech-lstm-transducer-stateless-2022-08-18 + # You will find the pre-trained model in icefall-asr-librispeech-lstm-transducer-stateless-2022-08-18/exp +""" + +import argparse +import logging +from pathlib import Path + +import sentencepiece as spm +import torch +import torch.nn as nn +from scaling_converter import convert_scaled_to_non_scaled +from train import add_model_arguments, get_params, get_transducer_model + +from icefall.checkpoint import ( + average_checkpoints, + average_checkpoints_with_averaged_model, + find_checkpoints, + load_checkpoint, +) +from icefall.utils import str2bool + + +def get_parser(): + parser = argparse.ArgumentParser( + formatter_class=argparse.ArgumentDefaultsHelpFormatter + ) + + parser.add_argument( + "--epoch", + type=int, + default=28, + help="""It specifies the checkpoint to use for averaging. + Note: Epoch counts from 0. + You can specify --avg to use more checkpoints for model averaging.""", + ) + + parser.add_argument( + "--iter", + type=int, + default=0, + help="""If positive, --epoch is ignored and it + will use the checkpoint exp_dir/checkpoint-iter.pt. + You can specify --avg to use more checkpoints for model averaging. + """, + ) + + parser.add_argument( + "--avg", + type=int, + default=15, + help="Number of checkpoints to average. Automatically select " + "consecutive checkpoints before the checkpoint specified by " + "'--epoch' and '--iter'", + ) + + parser.add_argument( + "--use-averaged-model", + type=str2bool, + default=True, + help="Whether to load averaged model. Currently it only supports " + "using --epoch. If True, it would decode with the averaged model " + "over the epoch range from `epoch-avg` (excluded) to `epoch`." + "Actually only the models with epoch number of `epoch-avg` and " + "`epoch` are loaded for averaging. ", + ) + + parser.add_argument( + "--exp-dir", + type=str, + default="pruned_transducer_stateless3/exp", + help="""It specifies the directory where all training related + files, e.g., checkpoints, log, etc, are saved + """, + ) + + parser.add_argument( + "--bpe-model", + type=str, + default="data/lang_bpe_500/bpe.model", + help="Path to the BPE model", + ) + + parser.add_argument( + "--jit-trace", + type=str2bool, + default=False, + help="""True to save a model after applying torch.jit.trace. + It will generate 3 files: + - encoder_jit_trace.pt + - decoder_jit_trace.pt + - joiner_jit_trace.pt + + Check ./jit_pretrained.py for how to use them. + """, + ) + + 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 + + +def export_encoder_model_jit_trace( + encoder_model: nn.Module, + encoder_filename: str, +) -> None: + """Export the given encoder model with torch.jit.trace() + + Note: The warmup argument is fixed to 1. + + Args: + encoder_model: + The input encoder model + encoder_filename: + The filename to save the exported model. + """ + x = torch.zeros(1, 100, 80, dtype=torch.float32) + x_lens = torch.tensor([100], dtype=torch.int64) + states = encoder_model.get_init_states() + + traced_model = torch.jit.trace(encoder_model, (x, x_lens, states)) + traced_model.save(encoder_filename) + logging.info(f"Saved to {encoder_filename}") + + +def export_decoder_model_jit_trace( + decoder_model: nn.Module, + decoder_filename: str, +) -> None: + """Export the given decoder model with torch.jit.trace() + + Note: The argument need_pad is fixed to False. + + Args: + decoder_model: + The input decoder model + decoder_filename: + The filename to save the exported model. + """ + y = torch.zeros(10, decoder_model.context_size, dtype=torch.int64) + need_pad = torch.tensor([False]) + + traced_model = torch.jit.trace(decoder_model, (y, need_pad)) + traced_model.save(decoder_filename) + logging.info(f"Saved to {decoder_filename}") + + +def export_joiner_model_jit_trace( + joiner_model: nn.Module, + joiner_filename: str, +) -> None: + """Export the given joiner model with torch.jit.trace() + + Note: The argument project_input is fixed to True. A user should not + project the encoder_out/decoder_out by himself/herself. The exported joiner + will do that for the user. + + Args: + joiner_model: + The input joiner model + joiner_filename: + The filename to save the exported model. + + """ + encoder_out_dim = joiner_model.encoder_proj.weight.shape[1] + decoder_out_dim = joiner_model.decoder_proj.weight.shape[1] + encoder_out = torch.rand(1, encoder_out_dim, dtype=torch.float32) + decoder_out = torch.rand(1, decoder_out_dim, dtype=torch.float32) + + traced_model = torch.jit.trace(joiner_model, (encoder_out, decoder_out)) + traced_model.save(joiner_filename) + logging.info(f"Saved to {joiner_filename}") + + +@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}") + + sp = spm.SentencePieceProcessor() + sp.load(params.bpe_model) + + # is defined in local/train_bpe_model.py + params.blank_id = sp.piece_to_id("") + params.vocab_size = sp.get_piece_size() + + logging.info(params) + + logging.info("About to create model") + model = get_transducer_model(params) + + if not params.use_averaged_model: + if params.iter > 0: + filenames = find_checkpoints( + params.exp_dir, iteration=-params.iter + )[: params.avg] + if len(filenames) == 0: + raise ValueError( + f"No checkpoints found for" + f" --iter {params.iter}, --avg {params.avg}" + ) + elif len(filenames) < params.avg: + raise ValueError( + f"Not enough checkpoints ({len(filenames)}) found for" + f" --iter {params.iter}, --avg {params.avg}" + ) + logging.info(f"averaging {filenames}") + model.to(device) + model.load_state_dict(average_checkpoints(filenames, device=device)) + elif params.avg == 1: + load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model) + else: + start = params.epoch - params.avg + 1 + filenames = [] + for i in range(start, params.epoch + 1): + if i >= 1: + filenames.append(f"{params.exp_dir}/epoch-{i}.pt") + logging.info(f"averaging {filenames}") + model.to(device) + model.load_state_dict(average_checkpoints(filenames, device=device)) + else: + if params.iter > 0: + filenames = find_checkpoints( + params.exp_dir, iteration=-params.iter + )[: params.avg + 1] + if len(filenames) == 0: + raise ValueError( + f"No checkpoints found for" + f" --iter {params.iter}, --avg {params.avg}" + ) + elif len(filenames) < params.avg + 1: + raise ValueError( + f"Not enough checkpoints ({len(filenames)}) found for" + f" --iter {params.iter}, --avg {params.avg}" + ) + filename_start = filenames[-1] + filename_end = filenames[0] + logging.info( + "Calculating the averaged model over iteration checkpoints" + f" from {filename_start} (excluded) to {filename_end}" + ) + model.to(device) + model.load_state_dict( + average_checkpoints_with_averaged_model( + filename_start=filename_start, + filename_end=filename_end, + device=device, + ) + ) + else: + assert params.avg > 0, params.avg + start = params.epoch - params.avg + assert start >= 1, start + filename_start = f"{params.exp_dir}/epoch-{start}.pt" + filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt" + logging.info( + f"Calculating the averaged model over epoch range from " + f"{start} (excluded) to {params.epoch}" + ) + model.to(device) + model.load_state_dict( + average_checkpoints_with_averaged_model( + filename_start=filename_start, + filename_end=filename_end, + device=device, + ) + ) + + model.to("cpu") + model.eval() + + if params.jit_trace is True: + convert_scaled_to_non_scaled(model, inplace=True) + logging.info("Using torch.jit.trace()") + encoder_filename = params.exp_dir / "encoder_jit_trace.pt" + export_encoder_model_jit_trace(model.encoder, encoder_filename) + + decoder_filename = params.exp_dir / "decoder_jit_trace.pt" + export_decoder_model_jit_trace(model.decoder, decoder_filename) + + joiner_filename = params.exp_dir / "joiner_jit_trace.pt" + export_joiner_model_jit_trace(model.joiner, joiner_filename) + else: + logging.info("Not using torchscript") + # Save it using a format so that it can be loaded + # by :func:`load_checkpoint` + filename = params.exp_dir / "pretrained.pt" + torch.save({"model": model.state_dict()}, str(filename)) + logging.info(f"Saved to {filename}") + + +if __name__ == "__main__": + formatter = ( + "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s" + ) + + logging.basicConfig(format=formatter, level=logging.INFO) + main() diff --git a/egs/librispeech/ASR/lstm_transducer_stateless3/jit_pretrained.py b/egs/librispeech/ASR/lstm_transducer_stateless3/jit_pretrained.py new file mode 100755 index 000000000..a3443cf0a --- /dev/null +++ b/egs/librispeech/ASR/lstm_transducer_stateless3/jit_pretrained.py @@ -0,0 +1,322 @@ +#!/usr/bin/env python3 +# Copyright 2022 Xiaomi Corp. (authors: Fangjun Kuang) +# +# See ../../../../LICENSE for clarification regarding multiple authors +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +""" +This script loads torchscript models, either exported by `torch.jit.trace()` +or by `torch.jit.script()`, and uses them to decode waves. +You can use the following command to get the exported models: + +./lstm_transducer_stateless3/export.py \ + --exp-dir ./lstm_transducer_stateless3/exp \ + --bpe-model data/lang_bpe_500/bpe.model \ + --epoch 40 \ + --avg 15 \ + --jit-trace 1 + +Usage of this script: + +./lstm_transducer_stateless3/jit_pretrained.py \ + --encoder-model-filename ./lstm_transducer_stateless3/exp/encoder_jit_trace.pt \ + --decoder-model-filename ./lstm_transducer_stateless3/exp/decoder_jit_trace.pt \ + --joiner-model-filename ./lstm_transducer_stateless3/exp/joiner_jit_trace.pt \ + --bpe-model ./data/lang_bpe_500/bpe.model \ + /path/to/foo.wav \ + /path/to/bar.wav +""" + +import argparse +import logging +import math +from typing import List + +import kaldifeat +import sentencepiece as spm +import torch +import torchaudio +from torch.nn.utils.rnn import pad_sequence + + +def get_parser(): + parser = argparse.ArgumentParser( + formatter_class=argparse.ArgumentDefaultsHelpFormatter + ) + + parser.add_argument( + "--encoder-model-filename", + type=str, + required=True, + help="Path to the encoder torchscript model. ", + ) + + parser.add_argument( + "--decoder-model-filename", + type=str, + required=True, + help="Path to the decoder torchscript model. ", + ) + + parser.add_argument( + "--joiner-model-filename", + type=str, + required=True, + help="Path to the joiner torchscript model. ", + ) + + parser.add_argument( + "--bpe-model", + type=str, + help="""Path to bpe.model.""", + ) + + 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( + "--context-size", + type=int, + default=2, + help="Context size of the decoder model", + ) + + return parser + + +def read_sound_files( + filenames: List[str], expected_sample_rate: float +) -> List[torch.Tensor]: + """Read a list of sound files into a list 1-D float32 torch tensors. + Args: + filenames: + A list of sound filenames. + expected_sample_rate: + The expected sample rate of the sound files. + Returns: + Return a list of 1-D float32 torch tensors. + """ + ans = [] + for f in filenames: + wave, sample_rate = torchaudio.load(f) + assert sample_rate == expected_sample_rate, ( + f"expected sample rate: {expected_sample_rate}. " + f"Given: {sample_rate}" + ) + # We use only the first channel + ans.append(wave[0]) + return ans + + +def greedy_search( + decoder: torch.jit.ScriptModule, + joiner: torch.jit.ScriptModule, + encoder_out: torch.Tensor, + encoder_out_lens: torch.Tensor, + context_size: int, +) -> List[List[int]]: + """Greedy search in batch mode. It hardcodes --max-sym-per-frame=1. + Args: + decoder: + The decoder model. + joiner: + The joiner model. + encoder_out: + A 3-D tensor of shape (N, T, C) + encoder_out_lens: + A 1-D tensor of shape (N,). + context_size: + The context size of the decoder model. + Returns: + Return the decoded results for each utterance. + """ + assert encoder_out.ndim == 3 + assert encoder_out.size(0) >= 1, encoder_out.size(0) + + packed_encoder_out = torch.nn.utils.rnn.pack_padded_sequence( + input=encoder_out, + lengths=encoder_out_lens.cpu(), + batch_first=True, + enforce_sorted=False, + ) + + device = encoder_out.device + blank_id = 0 # hard-code to 0 + + batch_size_list = packed_encoder_out.batch_sizes.tolist() + N = encoder_out.size(0) + + assert torch.all(encoder_out_lens > 0), encoder_out_lens + assert N == batch_size_list[0], (N, batch_size_list) + + hyps = [[blank_id] * context_size for _ in range(N)] + + decoder_input = torch.tensor( + hyps, + device=device, + dtype=torch.int64, + ) # (N, context_size) + + decoder_out = decoder( + decoder_input, + need_pad=torch.tensor([False]), + ).squeeze(1) + + offset = 0 + for batch_size in batch_size_list: + start = offset + end = offset + batch_size + current_encoder_out = packed_encoder_out.data[start:end] + current_encoder_out = current_encoder_out + # current_encoder_out's shape: (batch_size, encoder_out_dim) + offset = end + + decoder_out = decoder_out[:batch_size] + + logits = joiner( + current_encoder_out, + decoder_out, + ) + # logits'shape (batch_size, vocab_size) + + assert logits.ndim == 2, logits.shape + y = logits.argmax(dim=1).tolist() + emitted = False + for i, v in enumerate(y): + if v != blank_id: + hyps[i].append(v) + emitted = True + if emitted: + # update decoder output + decoder_input = [h[-context_size:] for h in hyps[:batch_size]] + decoder_input = torch.tensor( + decoder_input, + device=device, + dtype=torch.int64, + ) + decoder_out = decoder( + decoder_input, + need_pad=torch.tensor([False]), + ) + decoder_out = decoder_out.squeeze(1) + + sorted_ans = [h[context_size:] for h in hyps] + ans = [] + unsorted_indices = packed_encoder_out.unsorted_indices.tolist() + for i in range(N): + ans.append(sorted_ans[unsorted_indices[i]]) + + return ans + + +@torch.no_grad() +def main(): + parser = get_parser() + args = parser.parse_args() + logging.info(vars(args)) + + device = torch.device("cpu") + if torch.cuda.is_available(): + device = torch.device("cuda", 0) + + logging.info(f"device: {device}") + + encoder = torch.jit.load(args.encoder_model_filename) + decoder = torch.jit.load(args.decoder_model_filename) + joiner = torch.jit.load(args.joiner_model_filename) + + encoder.eval() + decoder.eval() + joiner.eval() + + encoder.to(device) + decoder.to(device) + joiner.to(device) + + sp = spm.SentencePieceProcessor() + sp.load(args.bpe_model) + + 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 = args.sample_rate + opts.mel_opts.num_bins = 80 + + fbank = kaldifeat.Fbank(opts) + + logging.info(f"Reading sound files: {args.sound_files}") + waves = read_sound_files( + filenames=args.sound_files, + expected_sample_rate=args.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) + + states = encoder.get_init_states(batch_size=features.size(0), device=device) + + encoder_out, encoder_out_lens, _ = encoder( + x=features, + x_lens=feature_lengths, + states=states, + ) + + hyps = greedy_search( + decoder=decoder, + joiner=joiner, + encoder_out=encoder_out, + encoder_out_lens=encoder_out_lens, + context_size=args.context_size, + ) + s = "\n" + for filename, hyp in zip(args.sound_files, hyps): + words = sp.decode(hyp) + s += f"{filename}:\n{words}\n\n" + logging.info(s) + + logging.info("Decoding Done") + + +if __name__ == "__main__": + formatter = ( + "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s" + ) + + logging.basicConfig(format=formatter, level=logging.INFO) + main() diff --git a/egs/librispeech/ASR/lstm_transducer_stateless3/joiner.py b/egs/librispeech/ASR/lstm_transducer_stateless3/joiner.py new file mode 120000 index 000000000..815fd4bb6 --- /dev/null +++ b/egs/librispeech/ASR/lstm_transducer_stateless3/joiner.py @@ -0,0 +1 @@ +../pruned_transducer_stateless2/joiner.py \ No newline at end of file diff --git a/egs/librispeech/ASR/lstm_transducer_stateless3/lstm.py b/egs/librispeech/ASR/lstm_transducer_stateless3/lstm.py new file mode 100644 index 000000000..90bc351f4 --- /dev/null +++ b/egs/librispeech/ASR/lstm_transducer_stateless3/lstm.py @@ -0,0 +1,860 @@ +# Copyright 2022 Xiaomi Corp. (authors: 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. + +import copy +import math +from typing import List, Optional, Tuple + +import torch +from encoder_interface import EncoderInterface +from scaling import ( + ActivationBalancer, + BasicNorm, + DoubleSwish, + ScaledConv2d, + ScaledLinear, + ScaledLSTM, +) +from torch import nn + +LOG_EPSILON = math.log(1e-10) + + +def unstack_states( + states: Tuple[torch.Tensor, torch.Tensor] +) -> List[Tuple[torch.Tensor, torch.Tensor]]: + """ + Unstack the lstm states corresponding to a batch of utterances into a list + of states, where the i-th entry is the state from the i-th utterance. + + Args: + states: + A tuple of 2 elements. + ``states[0]`` is the lstm hidden states, of a batch of utterance. + ``states[1]`` is the lstm cell states, of a batch of utterances. + + Returns: + A list of states. + ``states[i]`` is a tuple of 2 elememts of i-th utterance. + ``states[i][0]`` is the lstm hidden states of i-th utterance. + ``states[i][1]`` is the lstm cell states of i-th utterance. + """ + hidden_states, cell_states = states + + list_hidden_states = hidden_states.unbind(dim=1) + list_cell_states = cell_states.unbind(dim=1) + + ans = [ + (h.unsqueeze(1), c.unsqueeze(1)) + for (h, c) in zip(list_hidden_states, list_cell_states) + ] + return ans + + +def stack_states( + states_list: List[Tuple[torch.Tensor, torch.Tensor]] +) -> Tuple[torch.Tensor, torch.Tensor]: + """ + Stack list of lstm states corresponding to separate utterances into a single + lstm state so that it can be used as an input for lstm when those utterances + are formed into a batch. + + Args: + state_list: + Each element in state_list corresponds to the lstm state for a single + utterance. + ``states[i]`` is a tuple of 2 elememts of i-th utterance. + ``states[i][0]`` is the lstm hidden states of i-th utterance. + ``states[i][1]`` is the lstm cell states of i-th utterance. + + + Returns: + A new state corresponding to a batch of utterances. + It is a tuple of 2 elements. + ``states[0]`` is the lstm hidden states, of a batch of utterance. + ``states[1]`` is the lstm cell states, of a batch of utterances. + """ + hidden_states = torch.cat([s[0] for s in states_list], dim=1) + cell_states = torch.cat([s[1] for s in states_list], dim=1) + ans = (hidden_states, cell_states) + return ans + + +class RNN(EncoderInterface): + """ + Args: + num_features (int): + Number of input features. + subsampling_factor (int): + Subsampling factor of encoder (convolution layers before lstm layers) (default=4). # noqa + d_model (int): + Output dimension (default=512). + dim_feedforward (int): + Feedforward dimension (default=2048). + rnn_hidden_size (int): + Hidden dimension for lstm layers (default=1024). + grad_norm_threshold: + For each sequence 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. + num_encoder_layers (int): + Number of encoder layers (default=12). + dropout (float): + Dropout rate (default=0.1). + layer_dropout (float): + Dropout value for model-level warmup (default=0.075). + aux_layer_period (int): + Period of auxiliary layers used for random combiner during training. + If set to 0, will not use the random combiner (Default). + You can set a positive integer to use the random combiner, e.g., 3. + """ + + def __init__( + self, + num_features: int, + subsampling_factor: int = 4, + d_model: int = 512, + dim_feedforward: int = 2048, + rnn_hidden_size: int = 1024, + grad_norm_threshold: float = 10.0, + num_encoder_layers: int = 12, + dropout: float = 0.1, + layer_dropout: float = 0.075, + aux_layer_period: int = 0, + ) -> None: + super(RNN, self).__init__() + + self.num_features = num_features + self.subsampling_factor = subsampling_factor + if subsampling_factor != 4: + raise NotImplementedError("Support only 'subsampling_factor=4'.") + + # self.encoder_embed converts the input of shape (N, T, num_features) + # to the shape (N, T//subsampling_factor, d_model). + # That is, it does two things simultaneously: + # (1) subsampling: T -> T//subsampling_factor + # (2) embedding: num_features -> d_model + self.encoder_embed = Conv2dSubsampling(num_features, d_model) + + self.num_encoder_layers = num_encoder_layers + self.d_model = d_model + self.rnn_hidden_size = rnn_hidden_size + + encoder_layer = RNNEncoderLayer( + d_model=d_model, + dim_feedforward=dim_feedforward, + rnn_hidden_size=rnn_hidden_size, + grad_norm_threshold=grad_norm_threshold, + dropout=dropout, + layer_dropout=layer_dropout, + ) + self.encoder = RNNEncoder( + encoder_layer, + num_encoder_layers, + aux_layers=list( + range( + num_encoder_layers // 3, + num_encoder_layers - 1, + aux_layer_period, + ) + ) + if aux_layer_period > 0 + else None, + ) + + def forward( + self, + x: torch.Tensor, + x_lens: torch.Tensor, + states: Optional[Tuple[torch.Tensor, torch.Tensor]] = None, + warmup: float = 1.0, + ) -> Tuple[torch.Tensor, torch.Tensor, Tuple[torch.Tensor, torch.Tensor]]: + """ + Args: + x: + The input tensor. Its shape is (N, T, C), where N is the batch size, + T is the sequence length, C is the feature dimension. + x_lens: + A tensor of shape (N,), containing the number of frames in `x` + before padding. + states: + A tuple of 2 tensors (optional). It is for streaming inference. + states[0] is the hidden states of all layers, + with shape of (num_layers, N, d_model); + states[1] is the cell states of all layers, + with shape of (num_layers, N, rnn_hidden_size). + warmup: + A floating point value that gradually increases from 0 throughout + training; when it is >= 1.0 we are "fully warmed up". It is used + to turn modules on sequentially. + + Returns: + A tuple of 3 tensors: + - embeddings: its shape is (N, T', d_model), where T' is the output + sequence lengths. + - lengths: a tensor of shape (batch_size,) containing the number of + frames in `embeddings` before padding. + - updated states, whose shape is the same as the input states. + """ + x = self.encoder_embed(x) + x = x.permute(1, 0, 2) # (N, T, C) -> (T, N, C) + + # lengths = ((x_lens - 3) // 2 - 1) // 2 # issue an warning + # + # Note: rounding_mode in torch.div() is available only in torch >= 1.8.0 + lengths = (((x_lens - 3) >> 1) - 1) >> 1 + if not torch.jit.is_tracing(): + assert x.size(0) == lengths.max().item() + + if states is None: + x = self.encoder(x, warmup=warmup)[0] + # torch.jit.trace requires returned types to be the same as annotated # noqa + new_states = (torch.empty(0), torch.empty(0)) + else: + assert not self.training + assert len(states) == 2 + if not torch.jit.is_tracing(): + # for hidden state + assert states[0].shape == ( + self.num_encoder_layers, + x.size(1), + self.d_model, + ) + # for cell state + assert states[1].shape == ( + self.num_encoder_layers, + x.size(1), + self.rnn_hidden_size, + ) + x, new_states = self.encoder(x, states) + + x = x.permute(1, 0, 2) # (T, N, C) -> (N, T, C) + return x, lengths, new_states + + @torch.jit.export + def get_init_states( + self, batch_size: int = 1, device: torch.device = torch.device("cpu") + ) -> Tuple[torch.Tensor, torch.Tensor]: + """Get model initial states.""" + # for rnn hidden states + hidden_states = torch.zeros( + (self.num_encoder_layers, batch_size, self.d_model), device=device + ) + cell_states = torch.zeros( + (self.num_encoder_layers, batch_size, self.rnn_hidden_size), + device=device, + ) + return (hidden_states, cell_states) + + +class RNNEncoderLayer(nn.Module): + """ + RNNEncoderLayer is made up of lstm and feedforward networks. + For stable training, in each lstm module, gradient filter + is applied to filter out extremely large elements in batch gradients + and also the module parameters with soft masks. + + Args: + d_model: + The number of expected features in the input (required). + dim_feedforward: + The dimension of feedforward network model (default=2048). + rnn_hidden_size: + The hidden dimension of rnn layer. + grad_norm_threshold: + For each sequence 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. + dropout: + The dropout value (default=0.1). + layer_dropout: + The dropout value for model-level warmup (default=0.075). + """ + + def __init__( + self, + d_model: int, + dim_feedforward: int, + rnn_hidden_size: int, + grad_norm_threshold: float = 10.0, + dropout: float = 0.1, + layer_dropout: float = 0.075, + ) -> None: + super(RNNEncoderLayer, self).__init__() + self.layer_dropout = layer_dropout + self.d_model = d_model + self.rnn_hidden_size = rnn_hidden_size + + assert rnn_hidden_size >= d_model, (rnn_hidden_size, d_model) + + self.lstm = ScaledLSTM( + input_size=d_model, + hidden_size=rnn_hidden_size, + proj_size=d_model if rnn_hidden_size > d_model else 0, + num_layers=1, + dropout=0.0, + grad_norm_threshold=grad_norm_threshold, + ) + self.feed_forward = nn.Sequential( + ScaledLinear(d_model, dim_feedforward), + ActivationBalancer(channel_dim=-1), + DoubleSwish(), + nn.Dropout(dropout), + ScaledLinear(dim_feedforward, d_model, initial_scale=0.25), + ) + self.norm_final = BasicNorm(d_model) + + # try to ensure the output is close to zero-mean (or at least, zero-median). # noqa + self.balancer = ActivationBalancer( + channel_dim=-1, min_positive=0.45, max_positive=0.55, max_abs=6.0 + ) + self.dropout = nn.Dropout(dropout) + + def forward( + self, + src: torch.Tensor, + states: Optional[Tuple[torch.Tensor, torch.Tensor]] = None, + warmup: float = 1.0, + ) -> Tuple[torch.Tensor, Tuple[torch.Tensor, torch.Tensor]]: + """ + Pass the input through the encoder layer. + + Args: + src: + The sequence to the encoder layer (required). + Its shape is (S, N, E), where S is the sequence length, + N is the batch size, and E is the feature number. + states: + A tuple of 2 tensors (optional). It is for streaming inference. + states[0] is the hidden states of all layers, + with shape of (1, N, d_model); + states[1] is the cell states of all layers, + with shape of (1, N, rnn_hidden_size). + warmup: + It controls selective bypass of of layers; if < 1.0, we will + bypass layers more frequently. + """ + src_orig = src + + warmup_scale = min(0.1 + warmup, 1.0) + # alpha = 1.0 means fully use this encoder layer, 0.0 would mean + # completely bypass it. + if self.training: + alpha = ( + warmup_scale + if torch.rand(()).item() <= (1.0 - self.layer_dropout) + else 0.1 + ) + else: + alpha = 1.0 + + # lstm module + if states is None: + src_lstm = self.lstm(src)[0] + # torch.jit.trace requires returned types be the same as annotated + new_states = (torch.empty(0), torch.empty(0)) + else: + assert not self.training + assert len(states) == 2 + if not torch.jit.is_tracing(): + # for hidden state + assert states[0].shape == (1, src.size(1), self.d_model) + # for cell state + assert states[1].shape == (1, src.size(1), self.rnn_hidden_size) + src_lstm, new_states = self.lstm(src, states) + src = src + self.dropout(src_lstm) + + # feed forward module + src = src + self.dropout(self.feed_forward(src)) + + src = self.norm_final(self.balancer(src)) + + if alpha != 1.0: + src = alpha * src + (1 - alpha) * src_orig + + return src, new_states + + +class RNNEncoder(nn.Module): + """ + RNNEncoder is a stack of N encoder layers. + + Args: + encoder_layer: + An instance of the RNNEncoderLayer() class (required). + num_layers: + The number of sub-encoder-layers in the encoder (required). + """ + + def __init__( + self, + encoder_layer: nn.Module, + num_layers: int, + aux_layers: Optional[List[int]] = None, + ) -> None: + super(RNNEncoder, self).__init__() + self.layers = nn.ModuleList( + [copy.deepcopy(encoder_layer) for i in range(num_layers)] + ) + self.num_layers = num_layers + self.d_model = encoder_layer.d_model + self.rnn_hidden_size = encoder_layer.rnn_hidden_size + + self.aux_layers: List[int] = [] + self.combiner: Optional[nn.Module] = None + if aux_layers is not None: + assert len(set(aux_layers)) == len(aux_layers) + assert num_layers - 1 not in aux_layers + self.aux_layers = aux_layers + [num_layers - 1] + self.combiner = RandomCombine( + num_inputs=len(self.aux_layers), + final_weight=0.5, + pure_prob=0.333, + stddev=2.0, + ) + + def forward( + self, + src: torch.Tensor, + states: Optional[Tuple[torch.Tensor, torch.Tensor]] = None, + warmup: float = 1.0, + ) -> Tuple[torch.Tensor, Tuple[torch.Tensor, torch.Tensor]]: + """ + Pass the input through the encoder layer in turn. + + Args: + src: + The sequence to the encoder layer (required). + Its shape is (S, N, E), where S is the sequence length, + N is the batch size, and E is the feature number. + states: + A tuple of 2 tensors (optional). It is for streaming inference. + states[0] is the hidden states of all layers, + with shape of (num_layers, N, d_model); + states[1] is the cell states of all layers, + with shape of (num_layers, N, rnn_hidden_size). + warmup: + It controls selective bypass of of layers; if < 1.0, we will + bypass layers more frequently. + """ + if states is not None: + assert not self.training + assert len(states) == 2 + if not torch.jit.is_tracing(): + # for hidden state + assert states[0].shape == ( + self.num_layers, + src.size(1), + self.d_model, + ) + # for cell state + assert states[1].shape == ( + self.num_layers, + src.size(1), + self.rnn_hidden_size, + ) + + output = src + + outputs = [] + + new_hidden_states = [] + new_cell_states = [] + + for i, mod in enumerate(self.layers): + if states is None: + output = mod(output, warmup=warmup)[0] + else: + layer_state = ( + states[0][i : i + 1, :, :], # h: (1, N, d_model) + states[1][i : i + 1, :, :], # c: (1, N, rnn_hidden_size) + ) + output, (h, c) = mod(output, layer_state) + new_hidden_states.append(h) + new_cell_states.append(c) + + if self.combiner is not None and i in self.aux_layers: + outputs.append(output) + + if self.combiner is not None: + output = self.combiner(outputs) + + if states is None: + new_states = (torch.empty(0), torch.empty(0)) + else: + new_states = ( + torch.cat(new_hidden_states, dim=0), + torch.cat(new_cell_states, dim=0), + ) + + return output, new_states + + +class Conv2dSubsampling(nn.Module): + """Convolutional 2D subsampling (to 1/4 length). + + Convert an input of shape (N, T, idim) to an output + with shape (N, T', odim), where + T' = ((T-3)//2-1)//2, which approximates T' == T//4 + + It is based on + https://github.com/espnet/espnet/blob/master/espnet/nets/pytorch_backend/transformer/subsampling.py # noqa + """ + + def __init__( + self, + in_channels: int, + out_channels: int, + layer1_channels: int = 8, + layer2_channels: int = 32, + layer3_channels: int = 128, + ) -> None: + """ + Args: + in_channels: + Number of channels in. The input shape is (N, T, in_channels). + Caution: It requires: T >= 9, in_channels >= 9. + out_channels + Output dim. The output shape is (N, ((T-3)//2-1)//2, out_channels) + layer1_channels: + Number of channels in layer1 + layer1_channels: + Number of channels in layer2 + """ + assert in_channels >= 9 + super().__init__() + + self.conv = nn.Sequential( + ScaledConv2d( + in_channels=1, + out_channels=layer1_channels, + kernel_size=3, + padding=0, + ), + ActivationBalancer(channel_dim=1), + DoubleSwish(), + ScaledConv2d( + in_channels=layer1_channels, + out_channels=layer2_channels, + kernel_size=3, + stride=2, + ), + ActivationBalancer(channel_dim=1), + DoubleSwish(), + ScaledConv2d( + in_channels=layer2_channels, + out_channels=layer3_channels, + kernel_size=3, + stride=2, + ), + ActivationBalancer(channel_dim=1), + DoubleSwish(), + ) + self.out = ScaledLinear( + layer3_channels * (((in_channels - 3) // 2 - 1) // 2), out_channels + ) + # set learn_eps=False because out_norm is preceded by `out`, and `out` + # itself has learned scale, so the extra degree of freedom is not + # needed. + self.out_norm = BasicNorm(out_channels, learn_eps=False) + # constrain median of output to be close to zero. + self.out_balancer = ActivationBalancer( + channel_dim=-1, min_positive=0.45, max_positive=0.55 + ) + + def forward(self, x: torch.Tensor) -> torch.Tensor: + """Subsample x. + + Args: + x: + Its shape is (N, T, idim). + + Returns: + Return a tensor of shape (N, ((T-3)//2-1)//2, odim) + """ + # On entry, x is (N, T, idim) + x = x.unsqueeze(1) # (N, T, idim) -> (N, 1, T, idim) i.e., (N, C, H, W) + x = self.conv(x) + # Now x is of shape (N, odim, ((T-3)//2-1)//2, ((idim-3)//2-1)//2) + b, c, t, f = x.size() + x = self.out(x.transpose(1, 2).contiguous().view(b, t, c * f)) + # Now x is of shape (N, ((T-3)//2-1))//2, odim) + x = self.out_norm(x) + x = self.out_balancer(x) + return x + + +class RandomCombine(nn.Module): + """ + This module combines a list of Tensors, all with the same shape, to + produce a single output of that same shape which, in training time, + is a random combination of all the inputs; but which in test time + will be just the last input. + + The idea is that the list of Tensors will be a list of outputs of multiple + conformer layers. This has a similar effect as iterated loss. (See: + DEJA-VU: DOUBLE FEATURE PRESENTATION AND ITERATED LOSS IN DEEP TRANSFORMER + NETWORKS). + """ + + def __init__( + self, + num_inputs: int, + final_weight: float = 0.5, + pure_prob: float = 0.5, + stddev: float = 2.0, + ) -> None: + """ + Args: + num_inputs: + The number of tensor inputs, which equals the number of layers' + outputs that are fed into this module. E.g. in an 18-layer neural + net if we output layers 16, 12, 18, num_inputs would be 3. + final_weight: + The amount of weight or probability we assign to the + final layer when randomly choosing layers or when choosing + continuous layer weights. + pure_prob: + The probability, on each frame, with which we choose + only a single layer to output (rather than an interpolation) + stddev: + A standard deviation that we add to log-probs for computing + randomized weights. + + The method of choosing which layers, or combinations of layers, to use, + is conceptually as follows:: + + With probability `pure_prob`:: + With probability `final_weight`: choose final layer, + Else: choose random non-final layer. + Else:: + Choose initial log-weights that correspond to assigning + weight `final_weight` to the final layer and equal + weights to other layers; then add Gaussian noise + with variance `stddev` to these log-weights, and normalize + to weights (note: the average weight assigned to the + final layer here will not be `final_weight` if stddev>0). + """ + super().__init__() + assert 0 <= pure_prob <= 1, pure_prob + assert 0 < final_weight < 1, final_weight + assert num_inputs >= 1 + + self.num_inputs = num_inputs + self.final_weight = final_weight + self.pure_prob = pure_prob + self.stddev = stddev + + self.final_log_weight = ( + torch.tensor( + (final_weight / (1 - final_weight)) * (self.num_inputs - 1) + ) + .log() + .item() + ) + + def forward(self, inputs: List[torch.Tensor]) -> torch.Tensor: + """Forward function. + Args: + inputs: + A list of Tensor, e.g. from various layers of a transformer. + All must be the same shape, of (*, num_channels) + Returns: + A Tensor of shape (*, num_channels). In test mode + this is just the final input. + """ + num_inputs = self.num_inputs + assert len(inputs) == num_inputs + if not self.training or torch.jit.is_scripting(): + return inputs[-1] + + # Shape of weights: (*, num_inputs) + num_channels = inputs[0].shape[-1] + num_frames = inputs[0].numel() // num_channels + + ndim = inputs[0].ndim + # stacked_inputs: (num_frames, num_channels, num_inputs) + stacked_inputs = torch.stack(inputs, dim=ndim).reshape( + (num_frames, num_channels, num_inputs) + ) + + # weights: (num_frames, num_inputs) + weights = self._get_random_weights( + inputs[0].dtype, inputs[0].device, num_frames + ) + + weights = weights.reshape(num_frames, num_inputs, 1) + # ans: (num_frames, num_channels, 1) + ans = torch.matmul(stacked_inputs, weights) + # ans: (*, num_channels) + + ans = ans.reshape(inputs[0].shape[:-1] + (num_channels,)) + + # The following if causes errors for torch script in torch 1.6.0 + # if __name__ == "__main__": + # # for testing only... + # print("Weights = ", weights.reshape(num_frames, num_inputs)) + return ans + + def _get_random_weights( + self, dtype: torch.dtype, device: torch.device, num_frames: int + ) -> torch.Tensor: + """Return a tensor of random weights, of shape + `(num_frames, self.num_inputs)`, + Args: + dtype: + The data-type desired for the answer, e.g. float, double. + device: + The device needed for the answer. + num_frames: + The number of sets of weights desired + Returns: + A tensor of shape (num_frames, self.num_inputs), such that + `ans.sum(dim=1)` is all ones. + """ + pure_prob = self.pure_prob + if pure_prob == 0.0: + return self._get_random_mixed_weights(dtype, device, num_frames) + elif pure_prob == 1.0: + return self._get_random_pure_weights(dtype, device, num_frames) + else: + p = self._get_random_pure_weights(dtype, device, num_frames) + m = self._get_random_mixed_weights(dtype, device, num_frames) + return torch.where( + torch.rand(num_frames, 1, device=device) < self.pure_prob, p, m + ) + + def _get_random_pure_weights( + self, dtype: torch.dtype, device: torch.device, num_frames: int + ): + """Return a tensor of random one-hot weights, of shape + `(num_frames, self.num_inputs)`, + Args: + dtype: + The data-type desired for the answer, e.g. float, double. + device: + The device needed for the answer. + num_frames: + The number of sets of weights desired. + Returns: + A one-hot tensor of shape `(num_frames, self.num_inputs)`, with + exactly one weight equal to 1.0 on each frame. + """ + final_prob = self.final_weight + + # final contains self.num_inputs - 1 in all elements + final = torch.full((num_frames,), self.num_inputs - 1, device=device) + # nonfinal contains random integers in [0..num_inputs - 2], these are for non-final weights. # noqa + nonfinal = torch.randint( + self.num_inputs - 1, (num_frames,), device=device + ) + + indexes = torch.where( + torch.rand(num_frames, device=device) < final_prob, final, nonfinal + ) + ans = torch.nn.functional.one_hot( + indexes, num_classes=self.num_inputs + ).to(dtype=dtype) + return ans + + def _get_random_mixed_weights( + self, dtype: torch.dtype, device: torch.device, num_frames: int + ): + """Return a tensor of random one-hot weights, of shape + `(num_frames, self.num_inputs)`, + Args: + dtype: + The data-type desired for the answer, e.g. float, double. + device: + The device needed for the answer. + num_frames: + The number of sets of weights desired. + Returns: + A tensor of shape (num_frames, self.num_inputs), which elements + in [0..1] that sum to one over the second axis, i.e. + `ans.sum(dim=1)` is all ones. + """ + logprobs = ( + torch.randn(num_frames, self.num_inputs, dtype=dtype, device=device) + * self.stddev + ) + logprobs[:, -1] += self.final_log_weight + return logprobs.softmax(dim=1) + + +def _test_random_combine(final_weight: float, pure_prob: float, stddev: float): + print( + f"_test_random_combine: final_weight={final_weight}, pure_prob={pure_prob}, stddev={stddev}" # noqa + ) + num_inputs = 3 + num_channels = 50 + m = RandomCombine( + num_inputs=num_inputs, + final_weight=final_weight, + pure_prob=pure_prob, + stddev=stddev, + ) + + x = [torch.ones(3, 4, num_channels) for _ in range(num_inputs)] + + y = m(x) + assert y.shape == x[0].shape + assert torch.allclose(y, x[0]) # .. since actually all ones. + + +def _test_random_combine_main(): + _test_random_combine(0.999, 0, 0.0) + _test_random_combine(0.5, 0, 0.0) + _test_random_combine(0.999, 0, 0.0) + _test_random_combine(0.5, 0, 0.3) + _test_random_combine(0.5, 1, 0.3) + _test_random_combine(0.5, 0.5, 0.3) + + feature_dim = 50 + c = RNN(num_features=feature_dim, d_model=128) + batch_size = 5 + seq_len = 20 + # Just make sure the forward pass runs. + f = c( + torch.randn(batch_size, seq_len, feature_dim), + torch.full((batch_size,), seq_len, dtype=torch.int64), + ) + f # to remove flake8 warnings + + +if __name__ == "__main__": + feature_dim = 80 + m = RNN( + num_features=feature_dim, + d_model=512, + rnn_hidden_size=1024, + dim_feedforward=2048, + num_encoder_layers=12, + ) + batch_size = 5 + seq_len = 20 + # Just make sure the forward pass runs. + f = m( + torch.randn(batch_size, seq_len, feature_dim), + torch.full((batch_size,), seq_len, dtype=torch.int64), + warmup=0.5, + ) + num_param = sum([p.numel() for p in m.parameters()]) + print(f"Number of model parameters: {num_param}") + + _test_random_combine_main() diff --git a/egs/librispeech/ASR/lstm_transducer_stateless3/model.py b/egs/librispeech/ASR/lstm_transducer_stateless3/model.py new file mode 120000 index 000000000..1bf04f3a4 --- /dev/null +++ b/egs/librispeech/ASR/lstm_transducer_stateless3/model.py @@ -0,0 +1 @@ +../lstm_transducer_stateless/model.py \ No newline at end of file diff --git a/egs/librispeech/ASR/lstm_transducer_stateless3/optim.py b/egs/librispeech/ASR/lstm_transducer_stateless3/optim.py new file mode 120000 index 000000000..e2deb4492 --- /dev/null +++ b/egs/librispeech/ASR/lstm_transducer_stateless3/optim.py @@ -0,0 +1 @@ +../pruned_transducer_stateless2/optim.py \ No newline at end of file diff --git a/egs/librispeech/ASR/lstm_transducer_stateless3/pretrained.py b/egs/librispeech/ASR/lstm_transducer_stateless3/pretrained.py new file mode 100755 index 000000000..0e48fef04 --- /dev/null +++ b/egs/librispeech/ASR/lstm_transducer_stateless3/pretrained.py @@ -0,0 +1,352 @@ +#!/usr/bin/env python3 +# Copyright 2021 Xiaomi Corp. (authors: Fangjun Kuang) +# +# See ../../../../LICENSE for clarification regarding multiple authors +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +""" +Usage: + +(1) greedy search +./lstm_transducer_stateless3/pretrained.py \ + --checkpoint ./lstm_transducer_stateless3/exp/pretrained.pt \ + --bpe-model ./data/lang_bpe_500/bpe.model \ + --method greedy_search \ + /path/to/foo.wav \ + /path/to/bar.wav + +(2) beam search +./lstm_transducer_stateless3/pretrained.py \ + --checkpoint ./lstm_transducer_stateless3/exp/pretrained.pt \ + --bpe-model ./data/lang_bpe_500/bpe.model \ + --method beam_search \ + --beam-size 4 \ + /path/to/foo.wav \ + /path/to/bar.wav + +(3) modified beam search +./lstm_transducer_stateless3/pretrained.py \ + --checkpoint ./lstm_transducer_stateless3/exp/pretrained.pt \ + --bpe-model ./data/lang_bpe_500/bpe.model \ + --method modified_beam_search \ + --beam-size 4 \ + /path/to/foo.wav \ + /path/to/bar.wav + +(4) fast beam search +./lstm_transducer_stateless3/pretrained.py \ + --checkpoint ./lstm_transducer_stateless3/exp/pretrained.pt \ + --bpe-model ./data/lang_bpe_500/bpe.model \ + --method fast_beam_search \ + --beam-size 4 \ + /path/to/foo.wav \ + /path/to/bar.wav + +You can also use `./lstm_transducer_stateless3/exp/epoch-xx.pt`. + +Note: ./lstm_transducer_stateless3/exp/pretrained.pt is generated by +./lstm_transducer_stateless3/export.py +""" + + +import argparse +import logging +import math +from typing import List + +import k2 +import kaldifeat +import sentencepiece as spm +import torch +import torchaudio +from beam_search import ( + beam_search, + fast_beam_search_one_best, + greedy_search, + greedy_search_batch, + modified_beam_search, +) +from torch.nn.utils.rnn import pad_sequence +from train import add_model_arguments, get_params, get_transducer_model + + +def get_parser(): + parser = argparse.ArgumentParser( + formatter_class=argparse.ArgumentDefaultsHelpFormatter + ) + + parser.add_argument( + "--checkpoint", + type=str, + required=True, + help="Path to the checkpoint. " + "The checkpoint is assumed to be saved by " + "icefall.checkpoint.save_checkpoint().", + ) + + parser.add_argument( + "--bpe-model", + type=str, + help="""Path to bpe.model.""", + ) + + parser.add_argument( + "--method", + type=str, + default="greedy_search", + help="""Possible values are: + - greedy_search + - beam_search + - modified_beam_search + - fast_beam_search + """, + ) + + parser.add_argument( + "sound_files", + type=str, + nargs="+", + help="The input sound file(s) to transcribe. " + "Supported formats are those supported by torchaudio.load(). " + "For example, wav and flac are supported. " + "The sample rate has to be 16kHz.", + ) + + parser.add_argument( + "--sample-rate", + type=int, + default=16000, + help="The sample rate of the input sound file", + ) + + parser.add_argument( + "--beam-size", + type=int, + default=4, + help="""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}. " + f"Given: {sample_rate}" + ) + # We use only the first channel + ans.append(wave[0]) + return ans + + +@torch.no_grad() +def main(): + parser = get_parser() + args = parser.parse_args() + + params = get_params() + + params.update(vars(args)) + + sp = spm.SentencePieceProcessor() + sp.load(params.bpe_model) + + # is defined in local/train_bpe_model.py + params.blank_id = sp.piece_to_id("") + params.unk_id = sp.piece_to_id("") + params.vocab_size = sp.get_piece_size() + + logging.info(f"{params}") + + device = torch.device("cpu") + if torch.cuda.is_available(): + device = torch.device("cuda", 0) + + logging.info(f"device: {device}") + + logging.info("Creating model") + model = get_transducer_model(params) + + num_param = sum([p.numel() for p in model.parameters()]) + logging.info(f"Number of model parameters: {num_param}") + + checkpoint = torch.load(args.checkpoint, map_location="cpu") + model.load_state_dict(checkpoint["model"], strict=False) + model.to(device) + model.eval() + model.device = device + + logging.info("Constructing Fbank computer") + opts = kaldifeat.FbankOptions() + opts.device = device + opts.frame_opts.dither = 0 + opts.frame_opts.snip_edges = False + opts.frame_opts.samp_freq = params.sample_rate + opts.mel_opts.num_bins = params.feature_dim + + fbank = kaldifeat.Fbank(opts) + + logging.info(f"Reading sound files: {params.sound_files}") + waves = read_sound_files( + filenames=params.sound_files, expected_sample_rate=params.sample_rate + ) + waves = [w.to(device) for w in waves] + + logging.info("Decoding started") + features = fbank(waves) + feature_lengths = [f.size(0) for f in features] + + features = pad_sequence( + features, batch_first=True, padding_value=math.log(1e-10) + ) + + feature_lengths = torch.tensor(feature_lengths, device=device) + + encoder_out, encoder_out_lens, _ = model.encoder( + x=features, x_lens=feature_lengths + ) + + num_waves = encoder_out.size(0) + hyps = [] + msg = f"Using {params.method}" + if params.method == "beam_search": + msg += f" with beam size {params.beam_size}" + logging.info(msg) + + if params.method == "fast_beam_search": + decoding_graph = k2.trivial_graph(params.vocab_size - 1, device=device) + hyp_tokens = fast_beam_search_one_best( + model=model, + decoding_graph=decoding_graph, + encoder_out=encoder_out, + encoder_out_lens=encoder_out_lens, + beam=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.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()) + 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 sp.decode(hyp_tokens): + hyps.append(hyp.split()) + else: + for i in range(num_waves): + # fmt: off + encoder_out_i = encoder_out[i:i+1, :encoder_out_lens[i]] + # fmt: on + if params.method == "greedy_search": + hyp = greedy_search( + model=model, + encoder_out=encoder_out_i, + max_sym_per_frame=params.max_sym_per_frame, + ) + elif params.method == "beam_search": + hyp = beam_search( + model=model, + encoder_out=encoder_out_i, + beam=params.beam_size, + ) + else: + raise ValueError(f"Unsupported method: {params.method}") + + hyps.append(sp.decode(hyp).split()) + + s = "\n" + for filename, hyp in zip(params.sound_files, hyps): + words = " ".join(hyp) + s += f"{filename}:\n{words}\n\n" + logging.info(s) + + logging.info("Decoding Done") + + +if __name__ == "__main__": + formatter = ( + "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s" + ) + + logging.basicConfig(format=formatter, level=logging.INFO) + main() diff --git a/egs/librispeech/ASR/lstm_transducer_stateless3/scaling.py b/egs/librispeech/ASR/lstm_transducer_stateless3/scaling.py new file mode 120000 index 000000000..09d802cc4 --- /dev/null +++ b/egs/librispeech/ASR/lstm_transducer_stateless3/scaling.py @@ -0,0 +1 @@ +../pruned_transducer_stateless2/scaling.py \ No newline at end of file diff --git a/egs/librispeech/ASR/lstm_transducer_stateless3/scaling_converter.py b/egs/librispeech/ASR/lstm_transducer_stateless3/scaling_converter.py new file mode 120000 index 000000000..3b667058d --- /dev/null +++ b/egs/librispeech/ASR/lstm_transducer_stateless3/scaling_converter.py @@ -0,0 +1 @@ +../pruned_transducer_stateless3/scaling_converter.py \ No newline at end of file diff --git a/egs/librispeech/ASR/lstm_transducer_stateless3/stream.py b/egs/librispeech/ASR/lstm_transducer_stateless3/stream.py new file mode 120000 index 000000000..71ea6dff1 --- /dev/null +++ b/egs/librispeech/ASR/lstm_transducer_stateless3/stream.py @@ -0,0 +1 @@ +../lstm_transducer_stateless/stream.py \ No newline at end of file diff --git a/egs/librispeech/ASR/lstm_transducer_stateless3/streaming_decode.py b/egs/librispeech/ASR/lstm_transducer_stateless3/streaming_decode.py new file mode 100755 index 000000000..cfa918ed5 --- /dev/null +++ b/egs/librispeech/ASR/lstm_transducer_stateless3/streaming_decode.py @@ -0,0 +1,968 @@ +#!/usr/bin/env python3 +# +# Copyright 2021-2022 Xiaomi Corporation (Author: Fangjun Kuang, +# Zengwei Yao) +# +# See ../../../../LICENSE for clarification regarding multiple authors +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +""" +Usage: +(1) greedy search +./lstm_transducer_stateless3/streaming_decode.py \ + --epoch 40 \ + --avg 20 \ + --exp-dir lstm_transducer_stateless3/exp \ + --num-decode-streams 2000 \ + --num-encoder-layers 12 \ + --rnn-hidden-size 1024 \ + --decoding-method greedy_search \ + --use-averaged-model True + +(2) modified beam search +./lstm_transducer_stateless3/streaming_decode.py \ + --epoch 40 \ + --avg 20 \ + --exp-dir lstm_transducer_stateless3/exp \ + --num-decode-streams 2000 \ + --num-encoder-layers 12 \ + --rnn-hidden-size 1024 \ + --decoding-method modified_beam_search \ + --use-averaged-model True \ + --beam-size 4 + +(3) fast beam search +./lstm_transducer_stateless3/streaming_decode.py \ + --epoch 40 \ + --avg 20 \ + --exp-dir lstm_transducer_stateless3/exp \ + --num-decode-streams 2000 \ + --num-encoder-layers 12 \ + --rnn-hidden-size 1024 \ + --decoding-method fast_beam_search \ + --use-averaged-model True \ + --beam 4 \ + --max-contexts 4 \ + --max-states 8 +""" +import argparse +import logging +import warnings +from pathlib import Path +from typing import Dict, List, Optional, Tuple + +import k2 +import numpy as np +import sentencepiece as spm +import torch +import torch.nn as nn +from asr_datamodule import LibriSpeechAsrDataModule +from beam_search import Hypothesis, HypothesisList, get_hyps_shape +from kaldifeat import Fbank, FbankOptions +from lhotse import CutSet +from lstm import LOG_EPSILON, stack_states, unstack_states +from stream import Stream +from torch.nn.utils.rnn import pad_sequence +from train import add_model_arguments, get_params, get_transducer_model + +from icefall.checkpoint import ( + average_checkpoints, + average_checkpoints_with_averaged_model, + find_checkpoints, + load_checkpoint, +) +from icefall.decode import one_best_decoding +from icefall.utils import ( + AttributeDict, + get_texts, + setup_logger, + store_transcripts, + str2bool, + write_error_stats, +) + + +def get_parser(): + parser = argparse.ArgumentParser( + formatter_class=argparse.ArgumentDefaultsHelpFormatter + ) + + parser.add_argument( + "--epoch", + type=int, + default=40, + help="It specifies the checkpoint to use for decoding." + "Note: Epoch counts from 0.", + ) + + 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=20, + help="Number of checkpoints to average. Automatically select " + "consecutive checkpoints before the checkpoint specified by " + "'--epoch'. ", + ) + + parser.add_argument( + "--use-averaged-model", + type=str2bool, + default=False, + help="Whether to load averaged model. Currently it only supports " + "using --epoch. If True, it would decode with the averaged model " + "over the epoch range from `epoch-avg` (excluded) to `epoch`." + "Actually only the models with epoch number of `epoch-avg` and " + "`epoch` are loaded for averaging. ", + ) + + parser.add_argument( + "--exp-dir", + type=str, + default="lstm_transducer_stateless3/exp", + help="The experiment dir", + ) + + parser.add_argument( + "--bpe-model", + type=str, + default="data/lang_bpe_500/bpe.model", + help="Path to the BPE model", + ) + + parser.add_argument( + "--decoding-method", + type=str, + default="greedy_search", + help="""Possible values are: + - greedy_search + - modified_beam_search + - fast_beam_search + """, + ) + + parser.add_argument( + "--beam-size", + type=int, + default=4, + help="""An interger 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""", + ) + + parser.add_argument( + "--max-contexts", + type=int, + default=8, + help="""Used only when --decoding-method is + fast_beam_search""", + ) + + parser.add_argument( + "--max-states", + type=int, + default=64, + help="""Used only when --decoding-method is + fast_beam_search""", + ) + + parser.add_argument( + "--context-size", + type=int, + default=2, + help="The context size in the decoder. 1 means bigram; " + "2 means tri-gram", + ) + parser.add_argument( + "--max-sym-per-frame", + type=int, + default=1, + help="""Maximum number of symbols per frame. + Used only when --decoding_method is greedy_search""", + ) + + parser.add_argument( + "--sampling-rate", + type=float, + default=16000, + help="Sample rate of the audio", + ) + + parser.add_argument( + "--num-decode-streams", + type=int, + default=2000, + help="The number of streams that can be decoded in parallel", + ) + + add_model_arguments(parser) + + return parser + + +def greedy_search( + model: nn.Module, + encoder_out: torch.Tensor, + streams: List[Stream], +) -> None: + """Greedy search in batch mode. It hardcodes --max-sym-per-frame=1. + + Args: + model: + The transducer model. + encoder_out: + Output from the encoder. Its shape is (N, T, C), where N >= 1. + streams: + A list of Stream objects. + """ + assert len(streams) == encoder_out.size(0) + assert encoder_out.ndim == 3 + + blank_id = model.decoder.blank_id + context_size = model.decoder.context_size + device = next(model.parameters()).device + T = encoder_out.size(1) + + encoder_out = model.joiner.encoder_proj(encoder_out) + + decoder_input = torch.tensor( + [stream.hyp[-context_size:] for stream in streams], + device=device, + dtype=torch.int64, + ) + # decoder_out is of shape (batch_size, 1, decoder_out_dim) + decoder_out = model.decoder(decoder_input, need_pad=False) + decoder_out = model.joiner.decoder_proj(decoder_out) + + for t in range(T): + # current_encoder_out's shape: (batch_size, 1, encoder_out_dim) + current_encoder_out = encoder_out[:, t : t + 1, :] # noqa + + logits = model.joiner( + current_encoder_out.unsqueeze(2), + decoder_out.unsqueeze(1), + project_input=False, + ) + # logits'shape (batch_size, vocab_size) + logits = logits.squeeze(1).squeeze(1) + + assert logits.ndim == 2, logits.shape + y = logits.argmax(dim=1).tolist() + emitted = False + for i, v in enumerate(y): + if v != blank_id: + streams[i].hyp.append(v) + emitted = True + if emitted: + # update decoder output + decoder_input = torch.tensor( + [stream.hyp[-context_size:] for stream in streams], + device=device, + dtype=torch.int64, + ) + decoder_out = model.decoder( + decoder_input, + need_pad=False, + ) + decoder_out = model.joiner.decoder_proj(decoder_out) + + +def modified_beam_search( + model: nn.Module, + encoder_out: torch.Tensor, + streams: List[Stream], + beam: int = 4, +): + """Beam search in batch mode with --max-sym-per-frame=1 being hardcoded. + + Args: + model: + The RNN-T model. + encoder_out: + A 3-D tensor of shape (N, T, encoder_out_dim) containing the output of + the encoder model. + streams: + A list of stream objects. + beam: + Number of active paths during the beam search. + """ + assert encoder_out.ndim == 3, encoder_out.shape + assert len(streams) == encoder_out.size(0) + + blank_id = model.decoder.blank_id + context_size = model.decoder.context_size + device = next(model.parameters()).device + batch_size = len(streams) + T = encoder_out.size(1) + + B = [stream.hyps for stream in streams] + + encoder_out = model.joiner.encoder_proj(encoder_out) + + for t in range(T): + current_encoder_out = encoder_out[:, t].unsqueeze(1).unsqueeze(1) + # current_encoder_out's shape: (batch_size, 1, 1, encoder_out_dim) + + hyps_shape = get_hyps_shape(B).to(device) + + A = [list(b) for b in B] + B = [HypothesisList() for _ in range(batch_size)] + + ys_log_probs = torch.stack( + [hyp.log_prob.reshape(1) for hyps in A for hyp in hyps], dim=0 + ) # (num_hyps, 1) + + decoder_input = torch.tensor( + [hyp.ys[-context_size:] for hyps in A for hyp in hyps], + device=device, + dtype=torch.int64, + ) # (num_hyps, context_size) + + decoder_out = model.decoder(decoder_input, need_pad=False).unsqueeze(1) + decoder_out = model.joiner.decoder_proj(decoder_out) + # decoder_out is of shape (num_hyps, 1, 1, decoder_output_dim) + + # Note: For torch 1.7.1 and below, it requires a torch.int64 tensor + # as index, so we use `to(torch.int64)` below. + current_encoder_out = torch.index_select( + current_encoder_out, + dim=0, + index=hyps_shape.row_ids(1).to(torch.int64), + ) # (num_hyps, encoder_out_dim) + + logits = model.joiner( + current_encoder_out, decoder_out, project_input=False + ) + # logits is of shape (num_hyps, 1, 1, vocab_size) + + logits = logits.squeeze(1).squeeze(1) + + log_probs = logits.log_softmax(dim=-1) # (num_hyps, vocab_size) + + log_probs.add_(ys_log_probs) + + vocab_size = log_probs.size(-1) + + log_probs = log_probs.reshape(-1) + + row_splits = hyps_shape.row_splits(1) * vocab_size + log_probs_shape = k2.ragged.create_ragged_shape2( + row_splits=row_splits, cached_tot_size=log_probs.numel() + ) + ragged_log_probs = k2.RaggedTensor( + shape=log_probs_shape, value=log_probs + ) + + for i in range(batch_size): + topk_log_probs, topk_indexes = ragged_log_probs[i].topk(beam) + + with warnings.catch_warnings(): + warnings.simplefilter("ignore") + topk_hyp_indexes = (topk_indexes // vocab_size).tolist() + topk_token_indexes = (topk_indexes % vocab_size).tolist() + + for k in range(len(topk_hyp_indexes)): + hyp_idx = topk_hyp_indexes[k] + hyp = A[i][hyp_idx] + + new_ys = hyp.ys[:] + new_token = topk_token_indexes[k] + if new_token != blank_id: + new_ys.append(new_token) + + new_log_prob = topk_log_probs[k] + new_hyp = Hypothesis(ys=new_ys, log_prob=new_log_prob) + B[i].add(new_hyp) + + for i in range(batch_size): + streams[i].hyps = B[i] + + +def fast_beam_search_one_best( + model: nn.Module, + streams: List[Stream], + encoder_out: torch.Tensor, + processed_lens: torch.Tensor, + beam: float, + max_states: int, + max_contexts: int, +) -> None: + """It limits the maximum number of symbols per frame to 1. + + A lattice is first obtained using modified beam search, and then + the shortest path within the lattice is used as the final output. + + Args: + model: + An instance of `Transducer`. + streams: + A list of stream objects. + encoder_out: + A tensor of shape (N, T, C) from the encoder. + processed_lens: + A tensor of shape (N,) containing the number of processed frames + in `encoder_out` before padding. + beam: + Beam value, similar to the beam used in Kaldi.. + max_states: + Max states per stream per frame. + max_contexts: + Max contexts pre stream per frame. + """ + assert encoder_out.ndim == 3 + + context_size = model.decoder.context_size + vocab_size = model.decoder.vocab_size + + B, T, C = encoder_out.shape + assert B == len(streams) + + config = k2.RnntDecodingConfig( + vocab_size=vocab_size, + decoder_history_len=context_size, + beam=beam, + max_contexts=max_contexts, + max_states=max_states, + ) + individual_streams = [] + for i in range(B): + individual_streams.append(streams[i].rnnt_decoding_stream) + decoding_streams = k2.RnntDecodingStreams(individual_streams, config) + + encoder_out = model.joiner.encoder_proj(encoder_out) + + for t in range(T): + # shape is a RaggedShape of shape (B, context) + # contexts is a Tensor of shape (shape.NumElements(), context_size) + shape, contexts = decoding_streams.get_contexts() + # `nn.Embedding()` in torch below v1.7.1 supports only torch.int64 + contexts = contexts.to(torch.int64) + # decoder_out is of shape (shape.NumElements(), 1, decoder_out_dim) + decoder_out = model.decoder(contexts, need_pad=False) + decoder_out = model.joiner.decoder_proj(decoder_out) + # current_encoder_out is of shape + # (shape.NumElements(), 1, joiner_dim) + # fmt: off + current_encoder_out = torch.index_select( + encoder_out[:, t:t + 1, :], 0, shape.row_ids(1).to(torch.int64) + ) + # fmt: on + logits = model.joiner( + current_encoder_out.unsqueeze(2), + decoder_out.unsqueeze(1), + project_input=False, + ) + logits = logits.squeeze(1).squeeze(1) + log_probs = logits.log_softmax(dim=-1) + decoding_streams.advance(log_probs) + + decoding_streams.terminate_and_flush_to_streams() + + lattice = decoding_streams.format_output(processed_lens.tolist()) + + best_path = one_best_decoding(lattice) + hyps = get_texts(best_path) + + for i in range(B): + streams[i].hyp = hyps[i] + + +def decode_one_chunk( + model: nn.Module, + streams: List[Stream], + params: AttributeDict, + decoding_graph: Optional[k2.Fsa] = None, +) -> List[int]: + """ + Args: + model: + The Transducer model. + streams: + A list of Stream objects. + params: + It is returned by :func:`get_params`. + decoding_graph: + The decoding graph. Can be either a `k2.trivial_graph` or LG, Used + only when --decoding_method is fast_beam_search. + + Returns: + A list of indexes indicating the finished streams. + """ + device = next(model.parameters()).device + + feature_list = [] + feature_len_list = [] + state_list = [] + num_processed_frames_list = [] + + for stream in streams: + # We should first get `stream.num_processed_frames` + # before calling `stream.get_feature_chunk()` + # since `stream.num_processed_frames` would be updated + num_processed_frames_list.append(stream.num_processed_frames) + feature = stream.get_feature_chunk() + feature_len = feature.size(0) + feature_list.append(feature) + feature_len_list.append(feature_len) + state_list.append(stream.states) + + features = pad_sequence( + feature_list, batch_first=True, padding_value=LOG_EPSILON + ).to(device) + feature_lens = torch.tensor(feature_len_list, device=device) + num_processed_frames = torch.tensor( + num_processed_frames_list, device=device + ) + + # Make sure it has at least 1 frame after subsampling + tail_length = params.subsampling_factor + 5 + if features.size(1) < tail_length: + pad_length = tail_length - features.size(1) + feature_lens += pad_length + features = torch.nn.functional.pad( + features, + (0, 0, 0, pad_length), + mode="constant", + value=LOG_EPSILON, + ) + + # Stack states of all streams + states = stack_states(state_list) + + encoder_out, encoder_out_lens, states = model.encoder( + x=features, + x_lens=feature_lens, + states=states, + ) + + if params.decoding_method == "greedy_search": + greedy_search( + model=model, + streams=streams, + encoder_out=encoder_out, + ) + elif params.decoding_method == "modified_beam_search": + modified_beam_search( + model=model, + streams=streams, + encoder_out=encoder_out, + beam=params.beam_size, + ) + elif params.decoding_method == "fast_beam_search": + # feature_len is needed to get partial results. + # The rnnt_decoding_stream for fast_beam_search. + with warnings.catch_warnings(): + warnings.simplefilter("ignore") + processed_lens = ( + num_processed_frames // params.subsampling_factor + + encoder_out_lens + ) + fast_beam_search_one_best( + model=model, + streams=streams, + encoder_out=encoder_out, + processed_lens=processed_lens, + beam=params.beam, + max_contexts=params.max_contexts, + max_states=params.max_states, + ) + else: + raise ValueError( + f"Unsupported decoding method: {params.decoding_method}" + ) + + # Update cached states of each stream + state_list = unstack_states(states) + for i, s in enumerate(state_list): + streams[i].states = s + + finished_streams = [i for i, stream in enumerate(streams) if stream.done] + return finished_streams + + +def create_streaming_feature_extractor() -> Fbank: + """Create a CPU streaming feature extractor. + + At present, we assume it returns a fbank feature extractor with + fixed options. In the future, we will support passing in the options + from outside. + + Returns: + Return a CPU streaming feature extractor. + """ + opts = FbankOptions() + opts.device = "cpu" + opts.frame_opts.dither = 0 + opts.frame_opts.snip_edges = False + opts.frame_opts.samp_freq = 16000 + opts.mel_opts.num_bins = 80 + return Fbank(opts) + + +def decode_dataset( + cuts: CutSet, + model: nn.Module, + params: AttributeDict, + sp: spm.SentencePieceProcessor, + decoding_graph: Optional[k2.Fsa] = None, +): + """Decode dataset. + + Args: + cuts: + Lhotse Cutset containing the dataset to decode. + params: + It is returned by :func:`get_params`. + model: + The Transducer model. + sp: + The BPE model. + decoding_graph: + The decoding graph. Can be either a `k2.trivial_graph` or LG, Used + only when --decoding_method is fast_beam_search. + + Returns: + Return a dict, whose key may be "greedy_search" if greedy search + is used, or it may be "beam_7" if beam size of 7 is used. + Its value is a list of tuples. Each tuple contains two elements: + The first is the reference transcript, and the second is the + predicted result. + """ + device = next(model.parameters()).device + + log_interval = 300 + + fbank = create_streaming_feature_extractor() + + decode_results = [] + streams = [] + for num, cut in enumerate(cuts): + # Each utterance has a Stream. + stream = Stream( + params=params, + cut_id=cut.id, + decoding_graph=decoding_graph, + device=device, + LOG_EPS=LOG_EPSILON, + ) + + stream.states = model.encoder.get_init_states(device=device) + + audio: np.ndarray = cut.load_audio() + # audio.shape: (1, num_samples) + assert len(audio.shape) == 2 + assert audio.shape[0] == 1, "Should be single channel" + assert audio.dtype == np.float32, audio.dtype + # The trained model is using normalized samples + assert audio.max() <= 1, "Should be normalized to [-1, 1])" + + samples = torch.from_numpy(audio).squeeze(0) + feature = fbank(samples) + stream.set_feature(feature) + stream.ground_truth = cut.supervisions[0].text + + streams.append(stream) + + while len(streams) >= params.num_decode_streams: + finished_streams = decode_one_chunk( + model=model, + streams=streams, + params=params, + decoding_graph=decoding_graph, + ) + + for i in sorted(finished_streams, reverse=True): + decode_results.append( + ( + streams[i].id, + streams[i].ground_truth.split(), + sp.decode(streams[i].decoding_result()).split(), + ) + ) + del streams[i] + + if num % log_interval == 0: + logging.info(f"Cuts processed until now is {num}.") + + while len(streams) > 0: + finished_streams = decode_one_chunk( + model=model, + streams=streams, + params=params, + decoding_graph=decoding_graph, + ) + + for i in sorted(finished_streams, reverse=True): + decode_results.append( + ( + streams[i].id, + streams[i].ground_truth.split(), + sp.decode(streams[i].decoding_result()).split(), + ) + ) + del streams[i] + + if params.decoding_method == "greedy_search": + key = "greedy_search" + elif params.decoding_method == "fast_beam_search": + key = ( + f"beam_{params.beam}_" + f"max_contexts_{params.max_contexts}_" + f"max_states_{params.max_states}" + ) + else: + key = f"beam_size_{params.beam_size}" + + return {key: decode_results} + + +def save_results( + params: AttributeDict, + test_set_name: str, + results_dict: Dict[str, List[Tuple[List[str], List[str]]]], +): + test_set_wers = dict() + for key, results in results_dict.items(): + recog_path = ( + params.res_dir / f"recogs-{test_set_name}-{key}-{params.suffix}.txt" + ) + store_transcripts(filename=recog_path, texts=sorted(results)) + logging.info(f"The transcripts are stored in {recog_path}") + + # The following prints out WERs, per-word error statistics and aligned + # ref/hyp pairs. + errs_filename = ( + params.res_dir / f"errs-{test_set_name}-{key}-{params.suffix}.txt" + ) + with open(errs_filename, "w") as f: + wer = write_error_stats( + f, f"{test_set_name}-{key}", results, enable_log=True + ) + test_set_wers[key] = wer + + logging.info("Wrote detailed error stats to {}".format(errs_filename)) + + test_set_wers = sorted(test_set_wers.items(), key=lambda x: x[1]) + errs_info = ( + params.res_dir + / f"wer-summary-{test_set_name}-{key}-{params.suffix}.txt" + ) + with open(errs_info, "w") as f: + print("settings\tWER", file=f) + for key, val in test_set_wers: + print("{}\t{}".format(key, val), file=f) + + s = "\nFor {}, WER of different settings are:\n".format(test_set_name) + note = "\tbest for {}".format(test_set_name) + for key, val in test_set_wers: + s += "{}\t{}{}\n".format(key, val, note) + note = "" + logging.info(s) + + +@torch.no_grad() +def main(): + parser = get_parser() + LibriSpeechAsrDataModule.add_arguments(parser) + args = parser.parse_args() + args.exp_dir = Path(args.exp_dir) + + params = get_params() + params.update(vars(args)) + + assert params.decoding_method in ( + "greedy_search", + "fast_beam_search", + "modified_beam_search", + ) + params.res_dir = params.exp_dir / "streaming" / params.decoding_method + + if params.iter > 0: + params.suffix = f"iter-{params.iter}-avg-{params.avg}" + else: + params.suffix = f"epoch-{params.epoch}-avg-{params.avg}" + + if "fast_beam_search" in params.decoding_method: + params.suffix += f"-beam-{params.beam}" + params.suffix += f"-max-contexts-{params.max_contexts}" + params.suffix += f"-max-states-{params.max_states}" + elif "beam_search" in params.decoding_method: + params.suffix += ( + f"-{params.decoding_method}-beam-size-{params.beam_size}" + ) + else: + params.suffix += f"-context-{params.context_size}" + params.suffix += f"-max-sym-per-frame-{params.max_sym_per_frame}" + + if params.use_averaged_model: + params.suffix += "-use-averaged-model" + + setup_logger(f"{params.res_dir}/log-streaming-decode") + 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() + + params.device = device + + logging.info(params) + + logging.info("About to create model") + model = get_transducer_model(params) + + if not params.use_averaged_model: + if params.iter > 0: + filenames = find_checkpoints( + params.exp_dir, iteration=-params.iter + )[: params.avg] + if len(filenames) == 0: + raise ValueError( + f"No checkpoints found for" + f" --iter {params.iter}, --avg {params.avg}" + ) + elif len(filenames) < params.avg: + raise ValueError( + f"Not enough checkpoints ({len(filenames)}) found for" + f" --iter {params.iter}, --avg {params.avg}" + ) + logging.info(f"averaging {filenames}") + model.to(device) + model.load_state_dict(average_checkpoints(filenames, device=device)) + elif params.avg == 1: + load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model) + else: + start = params.epoch - params.avg + 1 + filenames = [] + for i in range(start, params.epoch + 1): + if i >= 1: + filenames.append(f"{params.exp_dir}/epoch-{i}.pt") + logging.info(f"averaging {filenames}") + model.to(device) + model.load_state_dict(average_checkpoints(filenames, device=device)) + else: + if params.iter > 0: + filenames = find_checkpoints( + params.exp_dir, iteration=-params.iter + )[: params.avg + 1] + if len(filenames) == 0: + raise ValueError( + f"No checkpoints found for" + f" --iter {params.iter}, --avg {params.avg}" + ) + elif len(filenames) < params.avg + 1: + raise ValueError( + f"Not enough checkpoints ({len(filenames)}) found for" + f" --iter {params.iter}, --avg {params.avg}" + ) + filename_start = filenames[-1] + filename_end = filenames[0] + logging.info( + "Calculating the averaged model over iteration checkpoints" + f" from {filename_start} (excluded) to {filename_end}" + ) + model.to(device) + model.load_state_dict( + average_checkpoints_with_averaged_model( + filename_start=filename_start, + filename_end=filename_end, + device=device, + ) + ) + else: + assert params.avg > 0, 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.eval() + + if params.decoding_method == "fast_beam_search": + decoding_graph = k2.trivial_graph(params.vocab_size - 1, device=device) + else: + decoding_graph = None + + num_param = sum([p.numel() for p in model.parameters()]) + logging.info(f"Number of model parameters: {num_param}") + + librispeech = LibriSpeechAsrDataModule(args) + + test_clean_cuts = librispeech.test_clean_cuts() + test_other_cuts = librispeech.test_other_cuts() + + test_sets = ["test-clean", "test-other"] + test_cuts = [test_clean_cuts, test_other_cuts] + + for test_set, test_cut in zip(test_sets, test_cuts): + results_dict = decode_dataset( + cuts=test_cut, + model=model, + params=params, + sp=sp, + decoding_graph=decoding_graph, + ) + + save_results( + params=params, + test_set_name=test_set, + results_dict=results_dict, + ) + + logging.info("Done!") + + +if __name__ == "__main__": + torch.manual_seed(20220810) + main() diff --git a/egs/librispeech/ASR/lstm_transducer_stateless3/test_model.py b/egs/librispeech/ASR/lstm_transducer_stateless3/test_model.py new file mode 100755 index 000000000..03dfe1997 --- /dev/null +++ b/egs/librispeech/ASR/lstm_transducer_stateless3/test_model.py @@ -0,0 +1,92 @@ +#!/usr/bin/env python3 +# Copyright 2022 Xiaomi Corp. (authors: Fangjun Kuang) +# +# See ../../../../LICENSE for clarification regarding multiple authors +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + + +""" +To run this file, do: + + cd icefall/egs/librispeech/ASR + python ./lstm_transducer_stateless/test_model.py +""" + +import os +from pathlib import Path + +import torch +from export import ( + export_decoder_model_jit_trace, + export_encoder_model_jit_trace, + export_joiner_model_jit_trace, +) +from lstm import stack_states, unstack_states +from scaling_converter import convert_scaled_to_non_scaled +from train import get_params, get_transducer_model + + +def test_model(): + params = get_params() + params.vocab_size = 500 + params.blank_id = 0 + params.context_size = 2 + params.unk_id = 2 + params.encoder_dim = 512 + params.rnn_hidden_size = 1024 + params.num_encoder_layers = 12 + params.aux_layer_period = 0 + params.exp_dir = Path("exp_test_model") + + model = get_transducer_model(params) + model.eval() + + num_param = sum([p.numel() for p in model.parameters()]) + print(f"Number of model parameters: {num_param}") + + convert_scaled_to_non_scaled(model, inplace=True) + + if not os.path.exists(params.exp_dir): + os.path.mkdir(params.exp_dir) + + encoder_filename = params.exp_dir / "encoder_jit_trace.pt" + export_encoder_model_jit_trace(model.encoder, encoder_filename) + + decoder_filename = params.exp_dir / "decoder_jit_trace.pt" + export_decoder_model_jit_trace(model.decoder, decoder_filename) + + joiner_filename = params.exp_dir / "joiner_jit_trace.pt" + export_joiner_model_jit_trace(model.joiner, joiner_filename) + + print("The model has been successfully exported using jit.trace.") + + +def test_states_stack_and_unstack(): + layer, batch, hidden, cell = 12, 100, 512, 1024 + states = ( + torch.randn(layer, batch, hidden), + torch.randn(layer, batch, cell), + ) + states2 = stack_states(unstack_states(states)) + assert torch.allclose(states[0], states2[0]) + assert torch.allclose(states[1], states2[1]) + + +def main(): + test_model() + test_states_stack_and_unstack() + + +if __name__ == "__main__": + main() diff --git a/egs/librispeech/ASR/lstm_transducer_stateless3/test_scaling_converter.py b/egs/librispeech/ASR/lstm_transducer_stateless3/test_scaling_converter.py new file mode 100644 index 000000000..7567dd58c --- /dev/null +++ b/egs/librispeech/ASR/lstm_transducer_stateless3/test_scaling_converter.py @@ -0,0 +1,257 @@ +#!/usr/bin/env python3 +# Copyright 2022 Xiaomi Corp. (authors: Fangjun Kuang) +# +# See ../../../../LICENSE for clarification regarding multiple authors +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +""" +To run this file, do: + + cd icefall/egs/librispeech/ASR + python ./lstm_transducer_stateless/test_scaling_converter.py +""" + +import copy + +import torch +from scaling import ( + ScaledConv1d, + ScaledConv2d, + ScaledEmbedding, + ScaledLinear, + ScaledLSTM, +) +from scaling_converter import ( + convert_scaled_to_non_scaled, + scaled_conv1d_to_conv1d, + scaled_conv2d_to_conv2d, + scaled_embedding_to_embedding, + scaled_linear_to_linear, + scaled_lstm_to_lstm, +) +from train import get_params, get_transducer_model + + +def get_model(): + params = get_params() + params.vocab_size = 500 + params.blank_id = 0 + params.context_size = 2 + params.unk_id = 2 + params.encoder_dim = 512 + params.rnn_hidden_size = 1024 + params.num_encoder_layers = 12 + params.aux_layer_period = -1 + + model = get_transducer_model(params) + return model + + +def test_scaled_linear_to_linear(): + N = 5 + in_features = 10 + out_features = 20 + for bias in [True, False]: + scaled_linear = ScaledLinear( + in_features=in_features, + out_features=out_features, + bias=bias, + ) + linear = scaled_linear_to_linear(scaled_linear) + x = torch.rand(N, in_features) + + y1 = scaled_linear(x) + y2 = linear(x) + assert torch.allclose(y1, y2) + + jit_scaled_linear = torch.jit.script(scaled_linear) + jit_linear = torch.jit.script(linear) + + y3 = jit_scaled_linear(x) + y4 = jit_linear(x) + + assert torch.allclose(y3, y4) + assert torch.allclose(y1, y4) + + +def test_scaled_conv1d_to_conv1d(): + in_channels = 3 + for bias in [True, False]: + scaled_conv1d = ScaledConv1d( + in_channels, + 6, + kernel_size=1, + stride=1, + padding=0, + bias=bias, + ) + + conv1d = scaled_conv1d_to_conv1d(scaled_conv1d) + + x = torch.rand(20, in_channels, 10) + y1 = scaled_conv1d(x) + y2 = conv1d(x) + assert torch.allclose(y1, y2) + + jit_scaled_conv1d = torch.jit.script(scaled_conv1d) + jit_conv1d = torch.jit.script(conv1d) + + y3 = jit_scaled_conv1d(x) + y4 = jit_conv1d(x) + + assert torch.allclose(y3, y4) + assert torch.allclose(y1, y4) + + +def test_scaled_conv2d_to_conv2d(): + in_channels = 1 + for bias in [True, False]: + scaled_conv2d = ScaledConv2d( + in_channels=in_channels, + out_channels=3, + kernel_size=3, + padding=1, + bias=bias, + ) + + conv2d = scaled_conv2d_to_conv2d(scaled_conv2d) + + x = torch.rand(20, in_channels, 10, 20) + y1 = scaled_conv2d(x) + y2 = conv2d(x) + assert torch.allclose(y1, y2) + + jit_scaled_conv2d = torch.jit.script(scaled_conv2d) + jit_conv2d = torch.jit.script(conv2d) + + y3 = jit_scaled_conv2d(x) + y4 = jit_conv2d(x) + + assert torch.allclose(y3, y4) + assert torch.allclose(y1, y4) + + +def test_scaled_embedding_to_embedding(): + scaled_embedding = ScaledEmbedding( + num_embeddings=500, + embedding_dim=10, + padding_idx=0, + ) + embedding = scaled_embedding_to_embedding(scaled_embedding) + + for s in [10, 100, 300, 500, 800, 1000]: + x = torch.randint(low=0, high=500, size=(s,)) + scaled_y = scaled_embedding(x) + y = embedding(x) + assert torch.equal(scaled_y, y) + + +def test_scaled_lstm_to_lstm(): + input_size = 512 + batch_size = 20 + for bias in [True, False]: + for hidden_size in [512, 1024]: + scaled_lstm = ScaledLSTM( + input_size=input_size, + hidden_size=hidden_size, + num_layers=1, + bias=bias, + proj_size=0 if hidden_size == input_size else input_size, + ) + + lstm = scaled_lstm_to_lstm(scaled_lstm) + + x = torch.rand(200, batch_size, input_size) + h0 = torch.randn(1, batch_size, input_size) + c0 = torch.randn(1, batch_size, hidden_size) + + y1, (h1, c1) = scaled_lstm(x, (h0, c0)) + y2, (h2, c2) = lstm(x, (h0, c0)) + assert torch.allclose(y1, y2) + assert torch.allclose(h1, h2) + assert torch.allclose(c1, c2) + + jit_scaled_lstm = torch.jit.trace(lstm, (x, (h0, c0))) + y3, (h3, c3) = jit_scaled_lstm(x, (h0, c0)) + assert torch.allclose(y1, y3) + assert torch.allclose(h1, h3) + assert torch.allclose(c1, c3) + + +def test_convert_scaled_to_non_scaled(): + for inplace in [False, True]: + model = get_model() + model.eval() + + orig_model = copy.deepcopy(model) + + converted_model = convert_scaled_to_non_scaled(model, inplace=inplace) + + model = orig_model + + # test encoder + N = 2 + T = 100 + vocab_size = model.decoder.vocab_size + + x = torch.randn(N, T, 80, dtype=torch.float32) + x_lens = torch.full((N,), x.size(1)) + + e1, e1_lens, _ = model.encoder(x, x_lens) + e2, e2_lens, _ = converted_model.encoder(x, x_lens) + + assert torch.all(torch.eq(e1_lens, e2_lens)) + assert torch.allclose(e1, e2), (e1 - e2).abs().max() + + # test decoder + U = 50 + y = torch.randint(low=1, high=vocab_size - 1, size=(N, U)) + + d1 = model.decoder(y) + d2 = model.decoder(y) + + assert torch.allclose(d1, d2) + + # test simple projection + lm1 = model.simple_lm_proj(d1) + am1 = model.simple_am_proj(e1) + + lm2 = converted_model.simple_lm_proj(d2) + am2 = converted_model.simple_am_proj(e2) + + assert torch.allclose(lm1, lm2) + assert torch.allclose(am1, am2) + + # test joiner + e = torch.rand(2, 3, 4, 512) + d = torch.rand(2, 3, 4, 512) + + j1 = model.joiner(e, d) + j2 = converted_model.joiner(e, d) + assert torch.allclose(j1, j2) + + +@torch.no_grad() +def main(): + test_scaled_linear_to_linear() + test_scaled_conv1d_to_conv1d() + test_scaled_conv2d_to_conv2d() + test_scaled_embedding_to_embedding() + test_scaled_lstm_to_lstm() + test_convert_scaled_to_non_scaled() + + +if __name__ == "__main__": + torch.manual_seed(20220730) + main() diff --git a/egs/librispeech/ASR/lstm_transducer_stateless3/train.py b/egs/librispeech/ASR/lstm_transducer_stateless3/train.py new file mode 100755 index 000000000..dc3697ae7 --- /dev/null +++ b/egs/librispeech/ASR/lstm_transducer_stateless3/train.py @@ -0,0 +1,1138 @@ +#!/usr/bin/env python3 +# Copyright 2021 Xiaomi Corp. (authors: Fangjun Kuang, +# Wei Kang, +# Mingshuang Luo,) +# 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: + +export CUDA_VISIBLE_DEVICES="0,1,2,3" + +./lstm_transducer_stateless3/train.py \ + --world-size 4 \ + --num-epochs 40 \ + --start-epoch 1 \ + --exp-dir lstm_transducer_stateless3/exp \ + --full-libri 1 \ + --max-duration 500 + +# For mix precision training: + +./lstm_transducer_stateless3/train.py \ + --world-size 4 \ + --num-epochs 40 \ + --start-epoch 1 \ + --use-fp16 1 \ + --exp-dir lstm_transducer_stateless3/exp \ + --full-libri 1 \ + --max-duration 550 +""" + +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 LibriSpeechAsrDataModule +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 lstm import RNN +from model import Transducer +from optim import Eden, Eve +from torch import Tensor +from torch.cuda.amp import GradScaler +from torch.nn.parallel import DistributedDataParallel as DDP +from torch.utils.tensorboard import SummaryWriter + +from icefall import diagnostics +from icefall.checkpoint import load_checkpoint, remove_checkpoints +from icefall.checkpoint import save_checkpoint as save_checkpoint_impl +from icefall.checkpoint import ( + save_checkpoint_with_global_batch_idx, + update_averaged_model, +) +from icefall.dist import cleanup_dist, setup_dist +from icefall.env import get_env_info +from icefall.utils import ( + AttributeDict, + MetricsTracker, + display_and_save_batch, + setup_logger, + str2bool, +) + +LRSchedulerType = Union[ + torch.optim.lr_scheduler._LRScheduler, optim.LRScheduler +] + + +def add_model_arguments(parser: argparse.ArgumentParser): + parser.add_argument( + "--num-encoder-layers", + type=int, + default=12, + help="Number of RNN encoder layers..", + ) + + parser.add_argument( + "--encoder-dim", + type=int, + default=512, + help="Encoder output dimesion.", + ) + + parser.add_argument( + "--rnn-hidden-size", + type=int, + default=1024, + help="Hidden dim for LSTM layers.", + ) + + parser.add_argument( + "--aux-layer-period", + type=int, + default=0, + help="""Peroid of auxiliary layers used for randomly combined during training. + If set to 0, will not use the random combiner (Default). + You can set a positive integer to use the random combiner, e.g., 3. + """, + ) + + parser.add_argument( + "--grad-norm-threshold", + type=float, + default=25.0, + help="""For each sequence 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 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=40, + 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="lstm_transducer_stateless/exp", + help="""The experiment dir. + It specifies the directory where all training related + files, e.g., checkpoints, log, etc, are saved + """, + ) + + parser.add_argument( + "--bpe-model", + type=str, + default="data/lang_bpe_500/bpe.model", + help="Path to the BPE model", + ) + + parser.add_argument( + "--initial-lr", + type=float, + default=0.003, + help="""The initial learning rate. This value should not need to be + changed.""", + ) + + parser.add_argument( + "--lr-batches", + type=float, + default=5000, + help="""Number of steps that affects how rapidly the learning rate decreases. + We suggest not to change this.""", + ) + + parser.add_argument( + "--lr-epochs", + type=float, + default=10, + help="""Number of epochs that affects how rapidly the learning rate decreases. + """, + ) + + parser.add_argument( + "--context-size", + type=int, + default=2, + help="The context size in the decoder. 1 means bigram; " + "2 means tri-gram", + ) + + parser.add_argument( + "--prune-range", + type=int, + default=5, + help="The prune range for rnnt loss, it means how many symbols(context)" + "we are using to compute the loss", + ) + + parser.add_argument( + "--lm-scale", + type=float, + default=0.25, + help="The scale to smooth the loss with lm " + "(output of prediction network) part.", + ) + + parser.add_argument( + "--am-scale", + type=float, + default=0.0, + help="The scale to smooth the loss with am (output of encoder network)" + "part.", + ) + + parser.add_argument( + "--simple-loss-scale", + type=float, + default=0.5, + help="To get pruning ranges, we will calculate a simple version" + "loss(joiner is just addition), this simple loss also uses for" + "training (as a regularization item). We will scale the simple loss" + "with this parameter before adding to the final loss.", + ) + + parser.add_argument( + "--seed", + type=int, + default=42, + help="The seed for random generators intended for reproducibility", + ) + + parser.add_argument( + "--print-diagnostics", + type=str2bool, + default=False, + help="Accumulate stats on activations, print them and exit.", + ) + + parser.add_argument( + "--save-every-n", + type=int, + default=4000, + help="""Save checkpoint after processing this number of batches" + periodically. We save checkpoint to exp-dir/ whenever + params.batch_idx_train % save_every_n == 0. The checkpoint filename + has the form: f'exp-dir/checkpoint-{params.batch_idx_train}.pt' + Note: It also saves checkpoint to `exp-dir/epoch-xxx.pt` at the + end of each epoch where `xxx` is the epoch number counting from 0. + """, + ) + + parser.add_argument( + "--keep-last-k", + type=int, + default=20, + help="""Only keep this number of checkpoints on disk. + For instance, if it is 3, there are only 3 checkpoints + in the exp-dir with filenames `checkpoint-xxx.pt`. + It does not affect checkpoints with name `epoch-xxx.pt`. + """, + ) + + parser.add_argument( + "--average-period", + type=int, + default=100, + 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. + + - num_decoder_layers: Number of decoder layer of transformer decoder. + + - warm_step: The warm_step for Noam optimizer. + """ + params = AttributeDict( + { + "best_train_loss": float("inf"), + "best_valid_loss": float("inf"), + "best_train_epoch": -1, + "best_valid_epoch": -1, + "batch_idx_train": 0, + "log_interval": 50, + "reset_interval": 200, + "valid_interval": 3000, # For the 100h subset, use 800 + # parameters for conformer + "feature_dim": 80, + "subsampling_factor": 4, + "dim_feedforward": 2048, + # parameters for decoder + "decoder_dim": 512, + # parameters for joiner + "joiner_dim": 512, + # parameters for Noam + "model_warm_step": 3000, # arg given to model, not for lrate + "env_info": get_env_info(), + } + ) + + return params + + +def get_encoder_model(params: AttributeDict) -> nn.Module: + encoder = RNN( + num_features=params.feature_dim, + subsampling_factor=params.subsampling_factor, + d_model=params.encoder_dim, + rnn_hidden_size=params.rnn_hidden_size, + grad_norm_threshold=params.grad_norm_threshold, + dim_feedforward=params.dim_feedforward, + num_encoder_layers=params.num_encoder_layers, + aux_layer_period=params.aux_layer_period, + ) + return encoder + + +def get_decoder_model(params: AttributeDict) -> nn.Module: + decoder = Decoder( + vocab_size=params.vocab_size, + decoder_dim=params.decoder_dim, + blank_id=params.blank_id, + context_size=params.context_size, + ) + return decoder + + +def get_joiner_model(params: AttributeDict) -> nn.Module: + joiner = Joiner( + encoder_dim=params.encoder_dim, + decoder_dim=params.decoder_dim, + joiner_dim=params.joiner_dim, + vocab_size=params.vocab_size, + ) + return joiner + + +def get_transducer_model(params: AttributeDict) -> nn.Module: + encoder = get_encoder_model(params) + decoder = get_decoder_model(params) + joiner = get_joiner_model(params) + + model = Transducer( + encoder=encoder, + decoder=decoder, + joiner=joiner, + encoder_dim=params.encoder_dim, + decoder_dim=params.decoder_dim, + joiner_dim=params.joiner_dim, + vocab_size=params.vocab_size, + ) + return model + + +def load_checkpoint_if_available( + params: AttributeDict, + model: nn.Module, + 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, + warmup: float = 1.0, +) -> Tuple[Tensor, MetricsTracker]: + """ + Compute RNN-T loss given the model and its inputs. + + Args: + params: + Parameters for training. See :func:`get_params`. + model: + The model for training. It is an instance of Conformer in our case. + batch: + A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()` + for the content in it. + is_training: + True for training. False for validation. When it is True, this + function enables autograd during computation; when it is False, it + disables autograd. + warmup: a floating point value which increases throughout training; + values >= 1.0 are fully warmed up and have all modules present. + """ + device = ( + model.device + 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) + + texts = batch["supervisions"]["text"] + y = sp.encode(texts, out_type=int) + y = k2.RaggedTensor(y).to(device) + + with torch.set_grad_enabled(is_training): + simple_loss, pruned_loss = model( + x=feature, + x_lens=feature_lens, + y=y, + prune_range=params.prune_range, + am_scale=params.am_scale, + lm_scale=params.lm_scale, + warmup=warmup, + reduction="none", + ) + simple_loss_is_finite = torch.isfinite(simple_loss) + pruned_loss_is_finite = torch.isfinite(pruned_loss) + is_finite = simple_loss_is_finite & pruned_loss_is_finite + if not torch.all(is_finite): + logging.info( + "Not all losses are finite!\n" + f"simple_loss: {simple_loss}\n" + f"pruned_loss: {pruned_loss}" + ) + display_and_save_batch(batch, params=params, sp=sp) + simple_loss = simple_loss[simple_loss_is_finite] + pruned_loss = pruned_loss[pruned_loss_is_finite] + + # If either all simple_loss or pruned_loss is inf or nan, + # we stop the training process by raising an exception + if torch.all(~simple_loss_is_finite) or torch.all( + ~pruned_loss_is_finite + ): + raise ValueError( + "There are too many utterances in this batch " + "leading to inf or nan losses." + ) + + simple_loss = simple_loss.sum() + pruned_loss = pruned_loss.sum() + # after the main warmup step, we keep pruned_loss_scale small + # for the same amount of time (model_warm_step), to avoid + # overwhelming the simple_loss and causing it to diverge, + # in case it had not fully learned the alignment yet. + pruned_loss_scale = ( + 0.0 + if warmup < 1.0 + else (0.1 if warmup > 1.0 and warmup < 2.0 else 1.0) + ) + loss = ( + params.simple_loss_scale * simple_loss + + pruned_loss_scale * pruned_loss + ) + + assert loss.requires_grad == is_training + + info = MetricsTracker() + with warnings.catch_warnings(): + warnings.simplefilter("ignore") + # info["frames"] is an approximate number for two reasons: + # (1) The acutal subsampling factor is ((lens - 1) // 2 - 1) // 2 + # (2) If some utterances in the batch lead to inf/nan loss, they + # are filtered out. + info["frames"] = ( + (feature_lens // params.subsampling_factor).sum().item() + ) + + # `utt_duration` and `utt_pad_proportion` would be normalized by `utterances` # noqa + info["utterances"] = feature.size(0) + # averaged input duration in frames over utterances + info["utt_duration"] = feature_lens.sum().item() + # averaged padding proportion over utterances + info["utt_pad_proportion"] = ( + ((feature.size(1) - feature_lens) / feature.size(1)).sum().item() + ) + + # Note: We use reduction=sum while computing the loss. + info["loss"] = loss.detach().cpu().item() + info["simple_loss"] = simple_loss.detach().cpu().item() + info["pruned_loss"] = pruned_loss.detach().cpu().item() + + return loss, info + + +def compute_validation_loss( + params: AttributeDict, + model: Union[nn.Module, DDP], + 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() + + for batch_idx, batch in enumerate(train_dl): + 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, + warmup=(params.batch_idx_train / params.model_warm_step), + ) + # summary stats + tot_loss = (tot_loss * (1 - 1 / params.reset_interval)) + loss_info + + # NOTE: We use reduction==sum and loss is computed over utterances + # in the batch and there is no normalization to it so far. + scaler.scale(loss).backward() + + scheduler.step_batch(params.batch_idx_train) + scaler.step(optimizer) + scaler.update() + optimizer.zero_grad() + except: # noqa + display_and_save_batch(batch, params=params, sp=sp) + raise + + if params.print_diagnostics and batch_idx == 30: + 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 % params.log_interval == 0 + and not params.print_diagnostics + ): + cur_lr = scheduler.get_last_lr()[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}" + ) + + 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 ( + batch_idx > 0 + and 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}") + 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)) + if params.full_libri is False: + params.valid_interval = 800 + + 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() + + logging.info(params) + + logging.info("About to create model") + model = get_transducer_model(params) + + num_param = sum([p.numel() for p in model.parameters()]) + logging.info(f"Number of model parameters: {num_param}") + + 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) + + 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]) + + optimizer = Eve(model.parameters(), lr=params.initial_lr) + + scheduler = Eden(optimizer, params.lr_batches, params.lr_epochs) + + if checkpoints and "optimizer" in checkpoints: + logging.info("Loading optimizer state dict") + optimizer.load_state_dict(checkpoints["optimizer"]) + + if ( + checkpoints + and "scheduler" in checkpoints + and checkpoints["scheduler"] is not None + ): + logging.info("Loading scheduler state dict") + scheduler.load_state_dict(checkpoints["scheduler"]) + + # # overwrite it + # scheduler.base_lrs = [params.initial_lr for _ in scheduler.base_lrs] + # print(scheduler.base_lrs) + + if params.print_diagnostics: + diagnostic = diagnostics.attach_diagnostics(model) + + librispeech = LibriSpeechAsrDataModule(args) + + train_cuts = librispeech.train_clean_100_cuts() + if params.full_libri: + train_cuts += librispeech.train_clean_360_cuts() + train_cuts += librispeech.train_other_500_cuts() + + def remove_short_and_long_utt(c: Cut): + # Keep only utterances with duration between 1 second and 20 seconds + # + # Caution: There is a reason to select 20.0 here. Please see + # ../local/display_manifest_statistics.py + # + # You should use ../local/display_manifest_statistics.py to get + # an utterance duration distribution for your dataset to select + # the threshold + return 1.0 <= c.duration <= 20.0 + + train_cuts = train_cuts.filter(remove_short_and_long_utt) + + if params.start_batch > 0 and checkpoints and "sampler" in checkpoints: + # We only load the sampler's state dict when it loads a checkpoint + # saved in the middle of an epoch + sampler_state_dict = checkpoints["sampler"] + else: + sampler_state_dict = None + + train_dl = librispeech.train_dataloaders( + train_cuts, sampler_state_dict=sampler_state_dict + ) + + valid_cuts = librispeech.dev_clean_cuts() + valid_cuts += librispeech.dev_other_cuts() + valid_dl = librispeech.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, + warmup=0.0 if params.start_epoch == 1 else 1.0, + ) + + scaler = GradScaler(enabled=params.use_fp16) + if checkpoints and "grad_scaler" in checkpoints: + logging.info("Loading grad scaler state dict") + scaler.load_state_dict(checkpoints["grad_scaler"]) + + for epoch in range(params.start_epoch, params.num_epochs + 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 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, + warmup: float, +): + 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, + warmup=warmup, + ) + loss.backward() + optimizer.step() + optimizer.zero_grad() + except RuntimeError as e: + if "CUDA out of memory" in str(e): + logging.error( + "Your GPU ran out of memory with the current " + "max_duration setting. We recommend decreasing " + "max_duration and trying again.\n" + f"Failing criterion: {criterion} " + f"(={crit_values[criterion]}) ..." + ) + raise + + +def main(): + parser = get_parser() + LibriSpeechAsrDataModule.add_arguments(parser) + args = parser.parse_args() + args.exp_dir = Path(args.exp_dir) + + world_size = args.world_size + assert world_size >= 1 + if world_size > 1: + mp.spawn(run, args=(world_size, args), nprocs=world_size, join=True) + else: + run(rank=0, world_size=1, args=args) + + +torch.set_num_threads(1) +torch.set_num_interop_threads(1) + +if __name__ == "__main__": + main() diff --git a/egs/librispeech/ASR/pruned_transducer_stateless2/scaling.py b/egs/librispeech/ASR/pruned_transducer_stateless2/scaling.py index cc3caecc7..9f839cbe0 100644 --- a/egs/librispeech/ASR/pruned_transducer_stateless2/scaling.py +++ b/egs/librispeech/ASR/pruned_transducer_stateless2/scaling.py @@ -111,6 +111,76 @@ class ActivationBalancerFunction(torch.autograd.Function): return x_grad - neg_delta_grad, None, None, 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 @@ -195,7 +265,7 @@ class ScaledLinear(nn.Linear): *args, initial_scale: float = 1.0, initial_speed: float = 1.0, - **kwargs + **kwargs, ): super(ScaledLinear, self).__init__(*args, **kwargs) initial_scale = torch.tensor(initial_scale).log() @@ -242,7 +312,7 @@ class ScaledConv1d(nn.Conv1d): *args, initial_scale: float = 1.0, initial_speed: float = 1.0, - **kwargs + **kwargs, ): super(ScaledConv1d, self).__init__(*args, **kwargs) initial_scale = torch.tensor(initial_scale).log() @@ -314,7 +384,7 @@ class ScaledConv2d(nn.Conv2d): *args, initial_scale: float = 1.0, initial_speed: float = 1.0, - **kwargs + **kwargs, ): super(ScaledConv2d, self).__init__(*args, **kwargs) initial_scale = torch.tensor(initial_scale).log() @@ -389,7 +459,8 @@ class ScaledLSTM(nn.LSTM): *args, initial_scale: float = 1.0, initial_speed: float = 1.0, - **kwargs + grad_norm_threshold: float = 10.0, + **kwargs, ): if "bidirectional" in kwargs: assert kwargs["bidirectional"] is False @@ -404,6 +475,10 @@ class ScaledLSTM(nn.LSTM): setattr(self, scale_name, param) self._scales.append(param) + self.grad_filter = GradientFilter( + batch_dim=1, threshold=grad_norm_threshold + ) + self._reset_parameters( initial_speed ) # Overrides the reset_parameters in base class @@ -513,10 +588,14 @@ class ScaledLSTM(nn.LSTM): 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, - self._get_flat_weights(), + flat_weights, self.bias, self.num_layers, self.dropout, @@ -891,9 +970,54 @@ def _test_scaled_lstm(): assert c.shape == (1, N, dim_hidden) +def _test_grad_filter(): + threshold = 50.0 + time, batch, channel = 200, 5, 128 + grad_filter = GradientFilter(batch_dim=1, threshold=threshold) + + for i in range(2): + x = torch.randn(time, batch, channel, requires_grad=True) + w = nn.Parameter(torch.ones(5)) + b = nn.Parameter(torch.zeros(5)) + + x_out, w_out, b_out = grad_filter(x, w, b) + + w_out_grad = torch.randn_like(w) + b_out_grad = torch.randn_like(b) + x_out_grad = torch.rand_like(x) + if i % 2 == 1: + # The gradient norm of the first element must be larger than + # `threshold * median`, where `median` is the median value + # of gradient norms of all elements in batch. + x_out_grad[:, 0, :] = torch.full((time, channel), threshold) + + torch.autograd.backward( + [x_out, w_out, b_out], [x_out_grad, w_out_grad, b_out_grad] + ) + + print( + "_test_grad_filter: for gradient norms, the first element > median * threshold ", # noqa + i % 2 == 1, + ) + + print( + "_test_grad_filter: x_out_grad norm = ", + (x_out_grad ** 2).mean(dim=(0, 2)).sqrt(), + ) + print( + "_test_grad_filter: x.grad norm = ", + (x.grad ** 2).mean(dim=(0, 2)).sqrt(), + ) + print("_test_grad_filter: w_out_grad = ", w_out_grad) + print("_test_grad_filter: w.grad = ", w.grad) + print("_test_grad_filter: b_out_grad = ", b_out_grad) + print("_test_grad_filter: b.grad = ", b.grad) + + if __name__ == "__main__": _test_activation_balancer_sign() _test_activation_balancer_magnitude() _test_basic_norm() _test_double_swish_deriv() _test_scaled_lstm() + _test_grad_filter()