From 87843e93821da00c6e7ffbc13c8da4d8766ab49c Mon Sep 17 00:00:00 2001 From: Yifan Yang <64255737+yfyeung@users.noreply.github.com> Date: Thu, 4 Apr 2024 23:29:16 +0800 Subject: [PATCH] k2SSL: a Faster and Better Framework for Self-Supervised Speech Representation Learning (#1500) * Add k2SSL * fix flake8 * fix for black * fix for black * fix for black * Update ssl_datamodule.py * Fix bugs in HubertDataset * update comments * add librilight * add checkpoint convert script * format --------- Co-authored-by: yifanyeung Co-authored-by: zzasdf <15218404468@163.com> --- .../SSL/zipformer/asr_datamodule.py | 1 + egs/librilight/SSL/zipformer/beam_search.py | 1 + egs/librilight/SSL/zipformer/dataset.py | 1 + egs/librilight/SSL/zipformer/decode.py | 1045 +++++++ egs/librilight/SSL/zipformer/decoder.py | 1 + .../SSL/zipformer/encoder_interface.py | 1 + egs/librilight/SSL/zipformer/finetune.py | 1552 +++++++++++ egs/librilight/SSL/zipformer/hubert_ce.py | 1 + egs/librilight/SSL/zipformer/joiner.py | 1 + egs/librilight/SSL/zipformer/model.py | 1 + egs/librilight/SSL/zipformer/optim.py | 1 + egs/librilight/SSL/zipformer/pretrain.py | 1366 +++++++++ egs/librilight/SSL/zipformer/scaling.py | 1 + .../SSL/zipformer/ssl_datamodule.py | 334 +++ egs/librilight/SSL/zipformer/utils.py | 1 + .../SSL/zipformer/wav2vec2_module.py | 1 + egs/librilight/SSL/zipformer/zipformer.py | 1 + egs/librispeech/SSL/hubert/asr_datamodule.py | 287 ++ .../SSL/hubert/attention_module.py | 840 ++++++ egs/librispeech/SSL/hubert/beam_search.py | 1 + egs/librispeech/SSL/hubert/dataset.py | 367 +++ egs/librispeech/SSL/hubert/decode.py | 1045 +++++++ egs/librispeech/SSL/hubert/decode_ce.py | 1045 +++++++ egs/librispeech/SSL/hubert/decoder.py | 1 + egs/librispeech/SSL/hubert/finetune.py | 1254 +++++++++ egs/librispeech/SSL/hubert/finetune_ce.py | 1254 +++++++++ egs/librispeech/SSL/hubert/hubert.py | 984 +++++++ egs/librispeech/SSL/hubert/hubert_ce.py | 940 +++++++ egs/librispeech/SSL/hubert/joiner.py | 1 + egs/librispeech/SSL/hubert/model.py | 344 +++ egs/librispeech/SSL/hubert/optim.py | 1 + egs/librispeech/SSL/hubert/pretrain.py | 1082 ++++++++ egs/librispeech/SSL/hubert/pretrain_ce.py | 1082 ++++++++ egs/librispeech/SSL/hubert/scaling.py | 1 + egs/librispeech/SSL/hubert/ssl_datamodule.py | 341 +++ egs/librispeech/SSL/hubert/utils.py | 338 +++ egs/librispeech/SSL/hubert/wav2vec2_module.py | 593 ++++ .../local/attach_kmeans_to_supervisions.py | 52 + .../local/convert_checkpoint_from_fairseq.py | 18 + egs/librispeech/SSL/local/prepare_char.py | 259 ++ egs/librispeech/SSL/local/prepare_lang.py | 388 +++ .../SSL/local/process_librispeech4finetune.py | 107 + .../SSL/local/process_librispeech4pretrain.py | 104 + egs/librispeech/SSL/local/process_raw_cuts.py | 23 + egs/librispeech/SSL/shared | 1 + .../SSL/zipformer/asr_datamodule.py | 1 + egs/librispeech/SSL/zipformer/beam_search.py | 1 + egs/librispeech/SSL/zipformer/dataset.py | 1 + egs/librispeech/SSL/zipformer/decode.py | 1043 +++++++ egs/librispeech/SSL/zipformer/decoder.py | 1 + .../SSL/zipformer/encoder_interface.py | 1 + egs/librispeech/SSL/zipformer/finetune.py | 1551 +++++++++++ egs/librispeech/SSL/zipformer/hubert_ce.py | 601 ++++ egs/librispeech/SSL/zipformer/joiner.py | 1 + egs/librispeech/SSL/zipformer/model.py | 344 +++ egs/librispeech/SSL/zipformer/optim.py | 1 + egs/librispeech/SSL/zipformer/pretrain.py | 1380 ++++++++++ egs/librispeech/SSL/zipformer/scaling.py | 1 + .../SSL/zipformer/ssl_datamodule.py | 1 + egs/librispeech/SSL/zipformer/utils.py | 337 +++ .../SSL/zipformer/wav2vec2_module.py | 108 + egs/librispeech/SSL/zipformer/zipformer.py | 2438 +++++++++++++++++ 62 files changed, 24874 insertions(+) create mode 120000 egs/librilight/SSL/zipformer/asr_datamodule.py create mode 120000 egs/librilight/SSL/zipformer/beam_search.py create mode 120000 egs/librilight/SSL/zipformer/dataset.py create mode 100644 egs/librilight/SSL/zipformer/decode.py create mode 120000 egs/librilight/SSL/zipformer/decoder.py create mode 120000 egs/librilight/SSL/zipformer/encoder_interface.py create mode 100644 egs/librilight/SSL/zipformer/finetune.py create mode 120000 egs/librilight/SSL/zipformer/hubert_ce.py create mode 120000 egs/librilight/SSL/zipformer/joiner.py create mode 120000 egs/librilight/SSL/zipformer/model.py create mode 120000 egs/librilight/SSL/zipformer/optim.py create mode 100644 egs/librilight/SSL/zipformer/pretrain.py create mode 120000 egs/librilight/SSL/zipformer/scaling.py create mode 100644 egs/librilight/SSL/zipformer/ssl_datamodule.py create mode 120000 egs/librilight/SSL/zipformer/utils.py create mode 120000 egs/librilight/SSL/zipformer/wav2vec2_module.py create mode 120000 egs/librilight/SSL/zipformer/zipformer.py create mode 100644 egs/librispeech/SSL/hubert/asr_datamodule.py create mode 100644 egs/librispeech/SSL/hubert/attention_module.py create mode 120000 egs/librispeech/SSL/hubert/beam_search.py create mode 100644 egs/librispeech/SSL/hubert/dataset.py create mode 100644 egs/librispeech/SSL/hubert/decode.py create mode 100644 egs/librispeech/SSL/hubert/decode_ce.py create mode 120000 egs/librispeech/SSL/hubert/decoder.py create mode 100644 egs/librispeech/SSL/hubert/finetune.py create mode 100644 egs/librispeech/SSL/hubert/finetune_ce.py create mode 100644 egs/librispeech/SSL/hubert/hubert.py create mode 100644 egs/librispeech/SSL/hubert/hubert_ce.py create mode 120000 egs/librispeech/SSL/hubert/joiner.py create mode 100644 egs/librispeech/SSL/hubert/model.py create mode 120000 egs/librispeech/SSL/hubert/optim.py create mode 100644 egs/librispeech/SSL/hubert/pretrain.py create mode 100644 egs/librispeech/SSL/hubert/pretrain_ce.py create mode 120000 egs/librispeech/SSL/hubert/scaling.py create mode 100644 egs/librispeech/SSL/hubert/ssl_datamodule.py create mode 100644 egs/librispeech/SSL/hubert/utils.py create mode 100644 egs/librispeech/SSL/hubert/wav2vec2_module.py create mode 100644 egs/librispeech/SSL/local/attach_kmeans_to_supervisions.py create mode 100644 egs/librispeech/SSL/local/convert_checkpoint_from_fairseq.py create mode 100644 egs/librispeech/SSL/local/prepare_char.py create mode 100644 egs/librispeech/SSL/local/prepare_lang.py create mode 100644 egs/librispeech/SSL/local/process_librispeech4finetune.py create mode 100644 egs/librispeech/SSL/local/process_librispeech4pretrain.py create mode 100644 egs/librispeech/SSL/local/process_raw_cuts.py create mode 120000 egs/librispeech/SSL/shared create mode 120000 egs/librispeech/SSL/zipformer/asr_datamodule.py create mode 120000 egs/librispeech/SSL/zipformer/beam_search.py create mode 120000 egs/librispeech/SSL/zipformer/dataset.py create mode 100644 egs/librispeech/SSL/zipformer/decode.py create mode 120000 egs/librispeech/SSL/zipformer/decoder.py create mode 120000 egs/librispeech/SSL/zipformer/encoder_interface.py create mode 100644 egs/librispeech/SSL/zipformer/finetune.py create mode 100644 egs/librispeech/SSL/zipformer/hubert_ce.py create mode 120000 egs/librispeech/SSL/zipformer/joiner.py create mode 100644 egs/librispeech/SSL/zipformer/model.py create mode 120000 egs/librispeech/SSL/zipformer/optim.py create mode 100644 egs/librispeech/SSL/zipformer/pretrain.py create mode 120000 egs/librispeech/SSL/zipformer/scaling.py create mode 120000 egs/librispeech/SSL/zipformer/ssl_datamodule.py create mode 100644 egs/librispeech/SSL/zipformer/utils.py create mode 100644 egs/librispeech/SSL/zipformer/wav2vec2_module.py create mode 100644 egs/librispeech/SSL/zipformer/zipformer.py diff --git a/egs/librilight/SSL/zipformer/asr_datamodule.py b/egs/librilight/SSL/zipformer/asr_datamodule.py new file mode 120000 index 0000000000..b9313bffce --- /dev/null +++ b/egs/librilight/SSL/zipformer/asr_datamodule.py @@ -0,0 +1 @@ +../../../librispeech/SSL/zipformer/asr_datamodule.py \ No newline at end of file diff --git a/egs/librilight/SSL/zipformer/beam_search.py b/egs/librilight/SSL/zipformer/beam_search.py new file mode 120000 index 0000000000..3b02c21db2 --- /dev/null +++ b/egs/librilight/SSL/zipformer/beam_search.py @@ -0,0 +1 @@ +../../../librispeech/SSL/zipformer/beam_search.py \ No newline at end of file diff --git a/egs/librilight/SSL/zipformer/dataset.py b/egs/librilight/SSL/zipformer/dataset.py new file mode 120000 index 0000000000..5cd60d3b45 --- /dev/null +++ b/egs/librilight/SSL/zipformer/dataset.py @@ -0,0 +1 @@ +../../../librispeech/SSL/zipformer/dataset.py \ No newline at end of file diff --git a/egs/librilight/SSL/zipformer/decode.py b/egs/librilight/SSL/zipformer/decode.py new file mode 100644 index 0000000000..95643c5e14 --- /dev/null +++ b/egs/librilight/SSL/zipformer/decode.py @@ -0,0 +1,1045 @@ +#!/usr/bin/env python3 +# Copyright 2021-2024 Xiaomi Corp. (authors: Fangjun Kuang, +# Zengwei Yao, +# Yifan Yang) +# +# See ../../../../LICENSE for clarification regarding multiple authors +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +""" +Usage: +(1) greedy search +./zipformer/decode.py \ + --epoch 28 \ + --avg 15 \ + --exp-dir ./zipformer/exp \ + --max-duration 600 \ + --decoding-method greedy_search + +(2) beam search (not recommended) +./zipformer/decode.py \ + --epoch 28 \ + --avg 15 \ + --exp-dir ./zipformer/exp \ + --max-duration 600 \ + --decoding-method beam_search \ + --beam-size 4 + +(3) modified beam search +./zipformer/decode.py \ + --epoch 28 \ + --avg 15 \ + --exp-dir ./zipformer/exp \ + --max-duration 600 \ + --decoding-method modified_beam_search \ + --beam-size 4 + +(4) fast beam search (one best) +./zipformer/decode.py \ + --epoch 28 \ + --avg 15 \ + --exp-dir ./zipformer/exp \ + --max-duration 600 \ + --decoding-method fast_beam_search \ + --beam 20.0 \ + --max-contexts 8 \ + --max-states 64 + +(5) fast beam search (nbest) +./zipformer/decode.py \ + --epoch 28 \ + --avg 15 \ + --exp-dir ./zipformer/exp \ + --max-duration 600 \ + --decoding-method fast_beam_search_nbest \ + --beam 20.0 \ + --max-contexts 8 \ + --max-states 64 \ + --num-paths 200 \ + --nbest-scale 0.5 + +(6) fast beam search (nbest oracle WER) +./zipformer/decode.py \ + --epoch 28 \ + --avg 15 \ + --exp-dir ./zipformer/exp \ + --max-duration 600 \ + --decoding-method fast_beam_search_nbest_oracle \ + --beam 20.0 \ + --max-contexts 8 \ + --max-states 64 \ + --num-paths 200 \ + --nbest-scale 0.5 + +(7) fast beam search (with LG) +./zipformer/decode.py \ + --epoch 28 \ + --avg 15 \ + --exp-dir ./zipformer/exp \ + --max-duration 600 \ + --decoding-method fast_beam_search_nbest_LG \ + --beam 20.0 \ + --max-contexts 8 \ + --max-states 64 +""" + + +import argparse +import logging +import math +import os +from collections import defaultdict +from pathlib import Path +from typing import Dict, List, Optional, Tuple + +import k2 +import sentencepiece as spm +import torch +import torch.nn as nn +from asr_datamodule import LibriSpeechAsrDataModule +from beam_search import ( + beam_search, + fast_beam_search_nbest, + fast_beam_search_nbest_LG, + fast_beam_search_nbest_oracle, + fast_beam_search_one_best, + greedy_search, + greedy_search_batch, + modified_beam_search, + modified_beam_search_lm_rescore, + modified_beam_search_lm_rescore_LODR, + modified_beam_search_lm_shallow_fusion, + modified_beam_search_LODR, +) +from finetune import add_model_arguments, get_model, get_params + +from icefall import ContextGraph, LmScorer, NgramLm +from icefall.checkpoint import ( + average_checkpoints, + average_checkpoints_with_averaged_model, + find_checkpoints, + load_checkpoint, +) +from icefall.lexicon import Lexicon +from icefall.utils import ( + AttributeDict, + make_pad_mask, + setup_logger, + store_transcripts, + str2bool, + write_error_stats, +) + + +def get_parser(): + parser = argparse.ArgumentParser( + formatter_class=argparse.ArgumentDefaultsHelpFormatter + ) + + parser.add_argument( + "--epoch", + type=int, + default=30, + help="""It specifies the checkpoint to use for decoding. + Note: Epoch counts from 1. + You can specify --avg to use more checkpoints for model averaging.""", + ) + + parser.add_argument( + "--iter", + type=int, + default=0, + help="""If positive, --epoch is ignored and it + will use the checkpoint exp_dir/checkpoint-iter.pt. + You can specify --avg to use more checkpoints for model averaging. + """, + ) + + parser.add_argument( + "--avg", + type=int, + default=15, + help="Number of checkpoints to average. Automatically select " + "consecutive checkpoints before the checkpoint specified by " + "'--epoch' and '--iter'", + ) + + parser.add_argument( + "--use-averaged-model", + type=str2bool, + default=True, + help="Whether to load averaged model. Currently it only supports " + "using --epoch. If True, it would decode with the averaged model " + "over the epoch range from `epoch-avg` (excluded) to `epoch`." + "Actually only the models with epoch number of `epoch-avg` and " + "`epoch` are loaded for averaging. ", + ) + + parser.add_argument( + "--exp-dir", + type=str, + default="zipformer/exp", + help="The experiment dir", + ) + + parser.add_argument( + "--bpe-model", + type=str, + default="data/lang_bpe_500/bpe.model", + help="Path to the BPE model", + ) + + parser.add_argument( + "--lang-dir", + type=Path, + default="data/lang_bpe_500", + help="The lang dir containing word table and LG graph", + ) + + parser.add_argument( + "--decoding-method", + type=str, + default="greedy_search", + help="""Possible values are: + - greedy_search + - beam_search + - modified_beam_search + - modified_beam_search_LODR + - fast_beam_search + - fast_beam_search_nbest + - fast_beam_search_nbest_oracle + - fast_beam_search_nbest_LG + If you use fast_beam_search_nbest_LG, you have to specify + `--lang-dir`, which should contain `LG.pt`. + """, + ) + + parser.add_argument( + "--beam-size", + type=int, + default=4, + help="""An integer indicating how many candidates we will keep for each + frame. Used only when --decoding-method is beam_search or + modified_beam_search.""", + ) + + parser.add_argument( + "--beam", + type=float, + default=20.0, + help="""A floating point value to calculate the cutoff score during beam + search (i.e., `cutoff = max-score - beam`), which is the same as the + `beam` in Kaldi. + Used only when --decoding-method is fast_beam_search, + fast_beam_search_nbest, fast_beam_search_nbest_LG, + and fast_beam_search_nbest_oracle + """, + ) + + parser.add_argument( + "--ngram-lm-scale", + type=float, + default=0.01, + help=""" + Used only when --decoding-method is fast_beam_search_nbest_LG. + It specifies the scale for n-gram LM scores. + """, + ) + + parser.add_argument( + "--max-contexts", + type=int, + default=8, + help="""Used only when --decoding-method is + fast_beam_search, fast_beam_search_nbest, fast_beam_search_nbest_LG, + and fast_beam_search_nbest_oracle""", + ) + + parser.add_argument( + "--max-states", + type=int, + default=64, + help="""Used only when --decoding-method is + fast_beam_search, fast_beam_search_nbest, fast_beam_search_nbest_LG, + and fast_beam_search_nbest_oracle""", + ) + + parser.add_argument( + "--context-size", + type=int, + default=2, + help="The context size in the decoder. 1 means bigram; " "2 means tri-gram", + ) + parser.add_argument( + "--max-sym-per-frame", + type=int, + default=1, + help="""Maximum number of symbols per frame. + Used only when --decoding-method is greedy_search""", + ) + + parser.add_argument( + "--num-paths", + type=int, + default=200, + help="""Number of paths for nbest decoding. + Used only when the decoding method is fast_beam_search_nbest, + fast_beam_search_nbest_LG, and fast_beam_search_nbest_oracle""", + ) + + parser.add_argument( + "--nbest-scale", + type=float, + default=0.5, + help="""Scale applied to lattice scores when computing nbest paths. + Used only when the decoding method is fast_beam_search_nbest, + fast_beam_search_nbest_LG, and fast_beam_search_nbest_oracle""", + ) + + parser.add_argument( + "--use-shallow-fusion", + type=str2bool, + default=False, + help="""Use neural network LM for shallow fusion. + If you want to use LODR, you will also need to set this to true + """, + ) + + parser.add_argument( + "--lm-type", + type=str, + default="rnn", + help="Type of NN lm", + choices=["rnn", "transformer"], + ) + + parser.add_argument( + "--lm-scale", + type=float, + default=0.3, + help="""The scale of the neural network LM + Used only when `--use-shallow-fusion` is set to True. + """, + ) + + parser.add_argument( + "--tokens-ngram", + type=int, + default=2, + help="""The order of the ngram lm. + """, + ) + + parser.add_argument( + "--backoff-id", + type=int, + default=500, + help="ID of the backoff symbol in the ngram LM", + ) + + parser.add_argument( + "--context-score", + type=float, + default=2, + help=""" + The bonus score of each token for the context biasing words/phrases. + Used only when --decoding-method is modified_beam_search and + modified_beam_search_LODR. + """, + ) + + parser.add_argument( + "--context-file", + type=str, + default="", + help=""" + The path of the context biasing lists, one word/phrase each line + Used only when --decoding-method is modified_beam_search and + modified_beam_search_LODR. + """, + ) + + add_model_arguments(parser) + + return parser + + +def decode_one_batch( + params: AttributeDict, + model: nn.Module, + sp: spm.SentencePieceProcessor, + batch: dict, + word_table: Optional[k2.SymbolTable] = None, + decoding_graph: Optional[k2.Fsa] = None, + context_graph: Optional[ContextGraph] = None, + LM: Optional[LmScorer] = None, + ngram_lm=None, + ngram_lm_scale: float = 0.0, +) -> Dict[str, List[List[str]]]: + """Decode one batch and return the result in a dict. The dict has the + following format: + + - key: It indicates the setting used for decoding. For example, + if greedy_search is used, it would be "greedy_search" + If beam search with a beam size of 7 is used, it would be + "beam_7" + - value: It contains the decoding result. `len(value)` equals to + batch size. `value[i]` is the decoding result for the i-th + utterance in the given batch. + Args: + params: + It's the return value of :func:`get_params`. + model: + The neural model. + sp: + The BPE model. + batch: + It is the return value from iterating + `lhotse.dataset.K2SpeechRecognitionDataset`. See its documentation + for the format of the `batch`. + word_table: + The word symbol table. + decoding_graph: + The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used + only when --decoding-method is fast_beam_search, fast_beam_search_nbest, + fast_beam_search_nbest_oracle, and fast_beam_search_nbest_LG. + LM: + A neural network language model. + ngram_lm: + A ngram language model + ngram_lm_scale: + The scale for the ngram language model. + Returns: + Return the decoding result. See above description for the format of + the returned dict. + """ + device = next(model.parameters()).device + audio = batch["audio"].to(device) + padding_mask = batch["padding_mask"].to(device) + + encoder_out, encoder_out_lens = model.forward_encoder(audio, padding_mask) + + 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(batch["supervisions"]["text"]), + nbest_scale=params.nbest_scale, + ) + for hyp in sp.decode(hyp_tokens): + hyps.append(hyp.split()) + elif params.decoding_method == "greedy_search" and params.max_sym_per_frame == 1: + hyp_tokens = greedy_search_batch( + model=model, + encoder_out=encoder_out, + encoder_out_lens=encoder_out_lens, + ) + for hyp in sp.decode(hyp_tokens): + hyps.append(hyp.split()) + elif params.decoding_method == "modified_beam_search": + hyp_tokens = modified_beam_search( + model=model, + encoder_out=encoder_out, + encoder_out_lens=encoder_out_lens, + beam=params.beam_size, + context_graph=context_graph, + ) + for hyp in sp.decode(hyp_tokens): + hyps.append(hyp.split()) + elif params.decoding_method == "modified_beam_search_lm_shallow_fusion": + hyp_tokens = modified_beam_search_lm_shallow_fusion( + model=model, + encoder_out=encoder_out, + encoder_out_lens=encoder_out_lens, + beam=params.beam_size, + LM=LM, + ) + for hyp in sp.decode(hyp_tokens): + hyps.append(hyp.split()) + elif params.decoding_method == "modified_beam_search_LODR": + hyp_tokens = modified_beam_search_LODR( + model=model, + encoder_out=encoder_out, + encoder_out_lens=encoder_out_lens, + beam=params.beam_size, + LODR_lm=ngram_lm, + LODR_lm_scale=ngram_lm_scale, + LM=LM, + context_graph=context_graph, + ) + for hyp in sp.decode(hyp_tokens): + hyps.append(hyp.split()) + elif params.decoding_method == "modified_beam_search_lm_rescore": + lm_scale_list = [0.01 * i for i in range(10, 50)] + ans_dict = modified_beam_search_lm_rescore( + model=model, + encoder_out=encoder_out, + encoder_out_lens=encoder_out_lens, + beam=params.beam_size, + LM=LM, + lm_scale_list=lm_scale_list, + ) + elif params.decoding_method == "modified_beam_search_lm_rescore_LODR": + lm_scale_list = [0.02 * i for i in range(2, 30)] + ans_dict = modified_beam_search_lm_rescore_LODR( + model=model, + encoder_out=encoder_out, + encoder_out_lens=encoder_out_lens, + beam=params.beam_size, + LM=LM, + LODR_lm=ngram_lm, + sp=sp, + lm_scale_list=lm_scale_list, + ) + else: + batch_size = encoder_out.size(0) + + for i in range(batch_size): + # fmt: off + encoder_out_i = encoder_out[i:i+1, :encoder_out_lens[i]] + # fmt: on + if params.decoding_method == "greedy_search": + hyp = greedy_search( + model=model, + encoder_out=encoder_out_i, + max_sym_per_frame=params.max_sym_per_frame, + ) + elif params.decoding_method == "beam_search": + hyp = beam_search( + model=model, + encoder_out=encoder_out_i, + beam=params.beam_size, + ) + else: + raise ValueError( + f"Unsupported decoding method: {params.decoding_method}" + ) + hyps.append(sp.decode(hyp).split()) + + if params.decoding_method == "greedy_search": + return {"greedy_search": hyps} + elif "fast_beam_search" in params.decoding_method: + key = f"beam_{params.beam}_" + key += f"max_contexts_{params.max_contexts}_" + key += f"max_states_{params.max_states}" + if "nbest" in params.decoding_method: + key += f"_num_paths_{params.num_paths}_" + key += f"nbest_scale_{params.nbest_scale}" + if "LG" in params.decoding_method: + key += f"_ngram_lm_scale_{params.ngram_lm_scale}" + + return {key: hyps} + elif "modified_beam_search" in params.decoding_method: + prefix = f"beam_size_{params.beam_size}" + if params.decoding_method in ( + "modified_beam_search_lm_rescore", + "modified_beam_search_lm_rescore_LODR", + ): + ans = dict() + assert ans_dict is not None + for key, hyps in ans_dict.items(): + hyps = [sp.decode(hyp).split() for hyp in hyps] + ans[f"{prefix}_{key}"] = hyps + return ans + else: + if params.has_contexts: + prefix += f"-context-score-{params.context_score}" + return {prefix: hyps} + else: + return {f"beam_size_{params.beam_size}": hyps} + + +def decode_dataset( + dl: torch.utils.data.DataLoader, + params: AttributeDict, + model: nn.Module, + sp: spm.SentencePieceProcessor, + word_table: Optional[k2.SymbolTable] = None, + decoding_graph: Optional[k2.Fsa] = None, + context_graph: Optional[ContextGraph] = None, + LM: Optional[LmScorer] = None, + ngram_lm=None, + ngram_lm_scale: float = 0.0, +) -> Dict[str, List[Tuple[str, List[str], List[str]]]]: + """Decode dataset. + + Args: + dl: + PyTorch's dataloader containing the dataset to decode. + params: + It is returned by :func:`get_params`. + model: + The neural model. + sp: + The BPE model. + word_table: + The word symbol table. + decoding_graph: + The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used + only when --decoding-method is fast_beam_search, fast_beam_search_nbest, + fast_beam_search_nbest_oracle, and fast_beam_search_nbest_LG. + Returns: + Return a dict, whose key may be "greedy_search" if greedy search + is used, or it may be "beam_7" if beam size of 7 is used. + Its value is a list of tuples. Each tuple contains two elements: + The first is the reference transcript, and the second is the + predicted result. + """ + num_cuts = 0 + + try: + num_batches = len(dl) + except TypeError: + num_batches = "?" + + if params.decoding_method == "greedy_search": + log_interval = 50 + else: + log_interval = 20 + + results = defaultdict(list) + for batch_idx, batch in enumerate(dl): + texts = batch["supervisions"]["text"] + cut_ids = [cut.id for cut in batch["cuts"]] + + hyps_dict = decode_one_batch( + params=params, + model=model, + sp=sp, + decoding_graph=decoding_graph, + context_graph=context_graph, + word_table=word_table, + batch=batch, + LM=LM, + ngram_lm=ngram_lm, + ngram_lm_scale=ngram_lm_scale, + ) + + for name, hyps in hyps_dict.items(): + this_batch = [] + assert len(hyps) == len(texts) + for cut_id, hyp_words, ref_text in zip(cut_ids, hyps, texts): + ref_words = ref_text.split() + this_batch.append((cut_id, ref_words, hyp_words)) + + results[name].extend(this_batch) + + num_cuts += len(texts) + + if batch_idx % log_interval == 0: + batch_str = f"{batch_idx}/{num_batches}" + + logging.info(f"batch {batch_str}, cuts processed until now is {num_cuts}") + return results + + +def save_results( + params: AttributeDict, + test_set_name: str, + results_dict: Dict[str, List[Tuple[str, List[str], List[str]]]], +): + test_set_wers = dict() + for key, results in results_dict.items(): + recog_path = ( + params.res_dir / f"recogs-{test_set_name}-{key}-{params.suffix}.txt" + ) + results = sorted(results) + store_transcripts(filename=recog_path, texts=results) + logging.info(f"The transcripts are stored in {recog_path}") + + # The following prints out WERs, per-word error statistics and aligned + # ref/hyp pairs. + errs_filename = ( + params.res_dir / f"errs-{test_set_name}-{key}-{params.suffix}.txt" + ) + with open(errs_filename, "w") as f: + wer = write_error_stats( + f, f"{test_set_name}-{key}", results, enable_log=True + ) + test_set_wers[key] = wer + + logging.info("Wrote detailed error stats to {}".format(errs_filename)) + + test_set_wers = sorted(test_set_wers.items(), key=lambda x: x[1]) + errs_info = ( + params.res_dir / f"wer-summary-{test_set_name}-{key}-{params.suffix}.txt" + ) + with open(errs_info, "w") as f: + print("settings\tWER", file=f) + for key, val in test_set_wers: + print("{}\t{}".format(key, val), file=f) + + s = "\nFor {}, WER of different settings are:\n".format(test_set_name) + note = "\tbest for {}".format(test_set_name) + for key, val in test_set_wers: + s += "{}\t{}{}\n".format(key, val, note) + note = "" + logging.info(s) + + +@torch.no_grad() +def main(): + parser = get_parser() + LibriSpeechAsrDataModule.add_arguments(parser) + LmScorer.add_arguments(parser) + args = parser.parse_args() + args.exp_dir = Path(args.exp_dir) + + params = get_params() + params.update(vars(args)) + + assert params.decoding_method in ( + "greedy_search", + "beam_search", + "fast_beam_search", + "fast_beam_search_nbest", + "fast_beam_search_nbest_LG", + "fast_beam_search_nbest_oracle", + "modified_beam_search", + "modified_beam_search_LODR", + "modified_beam_search_lm_shallow_fusion", + "modified_beam_search_lm_rescore", + "modified_beam_search_lm_rescore_LODR", + ) + params.res_dir = params.exp_dir / params.decoding_method + + if os.path.exists(params.context_file): + params.has_contexts = True + else: + params.has_contexts = False + + if params.iter > 0: + params.suffix = f"iter-{params.iter}-avg-{params.avg}" + else: + params.suffix = f"epoch-{params.epoch}-avg-{params.avg}" + + if "fast_beam_search" in params.decoding_method: + params.suffix += f"-beam-{params.beam}" + params.suffix += f"-max-contexts-{params.max_contexts}" + params.suffix += f"-max-states-{params.max_states}" + if "nbest" in params.decoding_method: + params.suffix += f"-nbest-scale-{params.nbest_scale}" + params.suffix += f"-num-paths-{params.num_paths}" + if "LG" in params.decoding_method: + params.suffix += f"-ngram-lm-scale-{params.ngram_lm_scale}" + elif "beam_search" in params.decoding_method: + params.suffix += f"-{params.decoding_method}-beam-size-{params.beam_size}" + if params.decoding_method in ( + "modified_beam_search", + "modified_beam_search_LODR", + ): + if params.has_contexts: + params.suffix += f"-context-score-{params.context_score}" + else: + params.suffix += f"-context-{params.context_size}" + params.suffix += f"-max-sym-per-frame-{params.max_sym_per_frame}" + + if params.use_shallow_fusion: + params.suffix += f"-{params.lm_type}-lm-scale-{params.lm_scale}" + + if "LODR" in params.decoding_method: + params.suffix += ( + f"-LODR-{params.tokens_ngram}gram-scale-{params.ngram_lm_scale}" + ) + + if params.use_averaged_model: + params.suffix += "-use-averaged-model" + + setup_logger(f"{params.res_dir}/log-decode-{params.suffix}") + logging.info("Decoding started") + + device = torch.device("cpu") + if torch.cuda.is_available(): + device = torch.device("cuda", 0) + + logging.info(f"Device: {device}") + + sp = spm.SentencePieceProcessor() + sp.load(params.bpe_model) + + # and are defined in local/train_bpe_model.py + params.blank_id = sp.piece_to_id("") + params.unk_id = sp.piece_to_id("") + params.vocab_size = sp.get_piece_size() + + logging.info(params) + + logging.info("About to create model") + model = get_model(params) + + if not params.use_averaged_model: + if params.iter > 0: + filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[ + : params.avg + ] + if len(filenames) == 0: + raise ValueError( + f"No checkpoints found for" + f" --iter {params.iter}, --avg {params.avg}" + ) + elif len(filenames) < params.avg: + raise ValueError( + f"Not enough checkpoints ({len(filenames)}) found for" + f" --iter {params.iter}, --avg {params.avg}" + ) + logging.info(f"averaging {filenames}") + model.to(device) + model.load_state_dict(average_checkpoints(filenames, device=device)) + elif params.avg == 1: + load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model) + else: + start = params.epoch - params.avg + 1 + filenames = [] + for i in range(start, params.epoch + 1): + if i >= 1: + filenames.append(f"{params.exp_dir}/epoch-{i}.pt") + logging.info(f"averaging {filenames}") + model.to(device) + model.load_state_dict(average_checkpoints(filenames, device=device)) + else: + if params.iter > 0: + filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[ + : params.avg + 1 + ] + if len(filenames) == 0: + raise ValueError( + f"No checkpoints found for" + f" --iter {params.iter}, --avg {params.avg}" + ) + elif len(filenames) < params.avg + 1: + raise ValueError( + f"Not enough checkpoints ({len(filenames)}) found for" + f" --iter {params.iter}, --avg {params.avg}" + ) + filename_start = filenames[-1] + filename_end = filenames[0] + logging.info( + "Calculating the averaged model over iteration checkpoints" + f" from {filename_start} (excluded) to {filename_end}" + ) + model.to(device) + model.load_state_dict( + average_checkpoints_with_averaged_model( + filename_start=filename_start, + filename_end=filename_end, + device=device, + ) + ) + else: + assert params.avg > 0, params.avg + start = params.epoch - params.avg + assert start >= 1, start + filename_start = f"{params.exp_dir}/epoch-{start}.pt" + filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt" + logging.info( + f"Calculating the averaged model over epoch range from " + f"{start} (excluded) to {params.epoch}" + ) + model.to(device) + model.load_state_dict( + average_checkpoints_with_averaged_model( + filename_start=filename_start, + filename_end=filename_end, + device=device, + ) + ) + + model.to(device) + model.eval() + + # only load the neural network LM if required + if params.use_shallow_fusion or params.decoding_method in ( + "modified_beam_search_lm_rescore", + "modified_beam_search_lm_rescore_LODR", + "modified_beam_search_lm_shallow_fusion", + "modified_beam_search_LODR", + ): + LM = LmScorer( + lm_type=params.lm_type, + params=params, + device=device, + lm_scale=params.lm_scale, + ) + LM.to(device) + LM.eval() + else: + LM = None + + # only load N-gram LM when needed + if params.decoding_method == "modified_beam_search_lm_rescore_LODR": + try: + import kenlm + except ImportError: + print("Please install kenlm first. You can use") + print(" pip install https://github.com/kpu/kenlm/archive/master.zip") + print("to install it") + import sys + + sys.exit(-1) + ngram_file_name = str(params.lang_dir / f"{params.tokens_ngram}gram.arpa") + logging.info(f"lm filename: {ngram_file_name}") + ngram_lm = kenlm.Model(ngram_file_name) + ngram_lm_scale = None # use a list to search + + elif params.decoding_method == "modified_beam_search_LODR": + lm_filename = f"{params.tokens_ngram}gram.fst.txt" + logging.info(f"Loading token level lm: {lm_filename}") + ngram_lm = NgramLm( + str(params.lang_dir / lm_filename), + backoff_id=params.backoff_id, + is_binary=False, + ) + logging.info(f"num states: {ngram_lm.lm.num_states}") + ngram_lm_scale = params.ngram_lm_scale + else: + ngram_lm = None + ngram_lm_scale = None + + if "fast_beam_search" in params.decoding_method: + if params.decoding_method == "fast_beam_search_nbest_LG": + lexicon = Lexicon(params.lang_dir) + word_table = lexicon.word_table + lg_filename = params.lang_dir / "LG.pt" + logging.info(f"Loading {lg_filename}") + decoding_graph = k2.Fsa.from_dict( + torch.load(lg_filename, map_location=device) + ) + decoding_graph.scores *= params.ngram_lm_scale + else: + word_table = None + decoding_graph = k2.trivial_graph(params.vocab_size - 1, device=device) + else: + decoding_graph = None + word_table = None + + if "modified_beam_search" in params.decoding_method: + if os.path.exists(params.context_file): + contexts = [] + for line in open(params.context_file).readlines(): + contexts.append((sp.encode(line.strip()), 0.0)) + context_graph = ContextGraph(params.context_score) + context_graph.build(contexts) + else: + context_graph = None + else: + context_graph = None + + num_param = sum([p.numel() for p in model.parameters()]) + logging.info(f"Number of model parameters: {num_param}") + + # we need cut ids to display recognition results. + args.return_cuts = True + librispeech = LibriSpeechAsrDataModule(args) + + dev_clean_cuts = librispeech.dev_clean_cuts() + dev_other_cuts = librispeech.dev_other_cuts() + + dev_clean_dl = librispeech.test_dataloaders( + dev_clean_cuts, + do_normalize=params.do_normalize, + ) + dev_other_dl = librispeech.test_dataloaders( + dev_other_cuts, + do_normalize=params.do_normalize, + ) + + test_clean_cuts = librispeech.test_clean_cuts() + test_other_cuts = librispeech.test_other_cuts() + + test_clean_dl = librispeech.test_dataloaders( + test_clean_cuts, + do_normalize=params.do_normalize, + ) + test_other_dl = librispeech.test_dataloaders( + test_other_cuts, + do_normalize=params.do_normalize, + ) + + test_sets = ["dev-clean", "dev-other", "test-clean", "test-other"] + test_dl = [dev_clean_dl, dev_other_dl, test_clean_dl, test_other_dl] + # test_sets = ["dev-clean", "dev-other"] + # test_dl = [dev_clean_dl, dev_other_dl] + + for test_set, test_dl in zip(test_sets, test_dl): + results_dict = decode_dataset( + dl=test_dl, + params=params, + model=model, + sp=sp, + word_table=word_table, + decoding_graph=decoding_graph, + context_graph=context_graph, + LM=LM, + ngram_lm=ngram_lm, + ngram_lm_scale=ngram_lm_scale, + ) + + save_results( + params=params, + test_set_name=test_set, + results_dict=results_dict, + ) + + logging.info("Done!") + + +if __name__ == "__main__": + main() diff --git a/egs/librilight/SSL/zipformer/decoder.py b/egs/librilight/SSL/zipformer/decoder.py new file mode 120000 index 0000000000..96dbfc5cd2 --- /dev/null +++ b/egs/librilight/SSL/zipformer/decoder.py @@ -0,0 +1 @@ +../../../librispeech/SSL/zipformer/decoder.py \ No newline at end of file diff --git a/egs/librilight/SSL/zipformer/encoder_interface.py b/egs/librilight/SSL/zipformer/encoder_interface.py new file mode 120000 index 0000000000..30859c51b1 --- /dev/null +++ b/egs/librilight/SSL/zipformer/encoder_interface.py @@ -0,0 +1 @@ +../../../librispeech/SSL/zipformer/encoder_interface.py \ No newline at end of file diff --git a/egs/librilight/SSL/zipformer/finetune.py b/egs/librilight/SSL/zipformer/finetune.py new file mode 100644 index 0000000000..50dbd5f2d7 --- /dev/null +++ b/egs/librilight/SSL/zipformer/finetune.py @@ -0,0 +1,1552 @@ +#!/usr/bin/env python3 +# Copyright 2021-2024 Xiaomi Corp. (authors: Fangjun Kuang, +# Wei Kang, +# Mingshuang Luo, +# Zengwei Yao, +# Yifan Yang, +# Daniel Povey) +# +# Copyright 2024 Shanghai Jiao Tong University (authors: Jianheng Zhuo) +# +# 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="0,1,2,3,4,5,6,7" + +# For HuBERT model finetuning: +./hubert/finetune.py \ + --world-size 8 \ + --num-epochs 200 \ + --start-epoch 1 \ + --use-fp16 1 \ + --exp-dir hubert/exp \ + --full-libri 0 \ + --max-duration 1000 + +It supports finetuning with: + - transducer loss (default), with `--use-transducer True --use-ctc False` + - ctc loss (not recommended), with `--use-transducer False --use-ctc True` + - transducer loss & ctc loss, with `--use-transducer True --use-ctc True` +""" + + +import argparse +import copy +import logging +import warnings +from pathlib import Path +from shutil import copyfile +from typing import Any, Dict, Optional, Tuple, Union + +import k2 +import optim +import sentencepiece as spm +import torch +import torch.multiprocessing as mp +import torch.nn as nn +from asr_datamodule import LibriSpeechAsrDataModule +from decoder import Decoder +from hubert_ce import HubertModel +from joiner import Joiner +from lhotse.cut import Cut +from lhotse.dataset.sampling.base import CutSampler +from lhotse.utils import fix_random_seed +from model import AsrModel +from optim import Eden, ScaledAdam +from 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.hooks import register_inf_check_hooks +from icefall.utils import ( + AttributeDict, + MetricsTracker, + get_parameter_groups_with_lrs, + setup_logger, + str2bool, +) + +LRSchedulerType = Union[torch.optim.lr_scheduler._LRScheduler, optim.LRScheduler] + + +def get_adjusted_batch_count(params: AttributeDict) -> float: + # returns the number of batches we would have used so far if we had used the reference + # duration. This is for purposes of set_batch_count(). + return ( + params.batch_idx_train + * params.accum_grad + * (params.max_duration * params.world_size) + / params.ref_duration + ) + + +def set_batch_count(model: Union[nn.Module, DDP], batch_count: float) -> None: + if isinstance(model, DDP): + # get underlying nn.Module + model = model.module + for name, module in model.named_modules(): + if hasattr(module, "batch_count"): + module.batch_count = batch_count + if hasattr(module, "name"): + module.name = name + + +def add_model_arguments(parser: argparse.ArgumentParser): + parser.add_argument( + "--num-encoder-layers", + type=str, + default="2,2,3,4,3,2", + help="Number of zipformer encoder layers per stack, comma separated.", + ) + + parser.add_argument( + "--downsampling-factor", + type=str, + default="1,2,4,8,4,2", + help="Downsampling factor for each stack of encoder layers.", + ) + + parser.add_argument( + "--feedforward-dim", + type=str, + default="512,768,1024,1536,1024,768", + help="Feedforward dimension of the zipformer encoder layers, per stack, comma separated.", + ) + + parser.add_argument( + "--num-heads", + type=str, + default="4,4,4,8,4,4", + help="Number of attention heads in the zipformer encoder layers: a single int or comma-separated list.", + ) + + parser.add_argument( + "--encoder-dim", + type=str, + default="192,256,384,512,384,256", + help="Embedding dimension in encoder stacks: a single int or comma-separated list.", + ) + + parser.add_argument( + "--query-head-dim", + type=str, + default="32", + help="Query/key dimension per head in encoder stacks: a single int or comma-separated list.", + ) + + parser.add_argument( + "--value-head-dim", + type=str, + default="12", + help="Value dimension per head in encoder stacks: a single int or comma-separated list.", + ) + + parser.add_argument( + "--pos-head-dim", + type=str, + default="4", + help="Positional-encoding dimension per head in encoder stacks: a single int or comma-separated list.", + ) + + parser.add_argument( + "--pos-dim", + type=int, + default="48", + help="Positional-encoding embedding dimension", + ) + + parser.add_argument( + "--encoder-unmasked-dim", + type=str, + default="192,192,256,256,256,192", + help="Unmasked dimensions in the encoders, relates to augmentation during training. " + "A single int or comma-separated list. Must be <= each corresponding encoder_dim.", + ) + + parser.add_argument( + "--cnn-module-kernel", + type=str, + default="31,31,15,15,15,31", + help="Sizes of convolutional kernels in convolution modules in each encoder stack: " + "a single int or comma-separated list.", + ) + + # hubert parameters + parser.add_argument( + "--label-rate", + type=float, + default=50, + ) + + parser.add_argument( + "--sample-rate", + type=float, + default=16000, + ) + + parser.add_argument( + "--extractor-mode", + type=str, + default="default", + help="""mode for feature extractor, should in EXTRACTOR_MODE_CHOICES. default has a single group + norm with d groups in the first conv block, whereas layer_norm + has layer norms in every block (meant to use with normalize=True)""", + ) + + parser.add_argument( + "--conv-feature-layers", + type=str, + default="[(512,10,5)] + [(512,3,2)] * 4 + [(512,2,2)] * 2", + help="string describing convolutional feature extraction layers in form of a python list that contains [(dim, kernel_size, stride), ...]", + ) + + parser.add_argument( + "--conv-bias", type=bool, default=False, help="include bias in conv encoder" + ) + + parser.add_argument( + "--feature-grad-mult", + type=float, + default=1.0, + help="multiply feature extractor var grads by this", + ) + + # masking + parser.add_argument("--mask-length", type=int, default=10, help="mask_length") + + parser.add_argument( + "--mask-prob", + type=float, + default=0.65, + help="probability of replacing a token with mask", + ) + + parser.add_argument( + "--mask-selection", + type=str, + choices=["static", "uniform", "normal", "poisson"], + default="static", + help="how to choose mask length", + ) + + parser.add_argument( + "--mask-other", + type=float, + default=0, + help="secondary mask argument (used for more complex distributions),see help in compute_mask_indicesh", + ) + + parser.add_argument( + "--no-mask-overlap", + type=bool, + default=False, + help="whether to allow masks to overlap", + ) + + parser.add_argument( + "--mask-min-space", + type=int, + default=1, + help="min space between spans (if no overlap is enabled)", + ) + + # channel masking + parser.add_argument( + "--mask-channel-length", + type=int, + default=10, + help="length of the mask for features (channels)", + ) + + parser.add_argument( + "--mask-channel-prob", + type=float, + default=0.0, + help="probability of replacing a feature with 0", + ) + + parser.add_argument( + "--mask-channel-selection", + type=str, + choices=["static", "uniform", "normal", "poisson"], + default="static", + help="how to choose mask length for channel masking", + ) + + parser.add_argument( + "--mask-channel-other", + type=float, + default=0, + help="secondary mask argument (used for more complex distributions), see help in compute_mask_indicesh", + ) + + parser.add_argument( + "--no-mask-channel-overlap", + type=bool, + default=False, + help="whether to allow channel masks to overlap", + ) + + parser.add_argument( + "--mask-channel-min-space", + type=int, + default=1, + help="min space between spans (if no overlap is enabled)", + ) + + # loss computation + parser.add_argument( + "--skip-masked", + type=bool, + default=False, + help="skip computing losses over masked frames", + ) + + parser.add_argument( + "--skip-nomask", + type=bool, + default=False, + help="skip computing losses over unmasked frames", + ) + + parser.add_argument( + "--checkpoint-activations", + type=bool, + default=False, + help="recompute activations and save memory for extra compute", + ) + + parser.add_argument( + "--pred-masked-weight", + type=float, + default=1, + help="weight for masked part in ssl loss", + ) + + parser.add_argument( + "--pred-nomask-weight", + type=float, + default=0, + help="weight for masked part in ssl loss", + ) + + parser.add_argument( + "--loss-weights", + type=float, + nargs="*", + default=[10], + help="weight for masked part in ssl loss", + ) + + # FP16 optimization + parser.add_argument( + "--required-seq-len-multiple", + type=int, + default=2, + help="pad the input to encoder such that the sequence length is divisible by multiple", + ) + + parser.add_argument( + "--attn-type", type=str, default="", help="if espnet use ESPNET MHA" + ) + + parser.add_argument( + "--pos-enc-type", + type=str, + default="abs", + help="Positional encoding type to use in conformer", + ) + + parser.add_argument( + "--logit-temp", type=float, default=0.1, help="temperature to divide logits by" + ) + + parser.add_argument( + "--dropout-input", + type=float, + default=0.0, + help="dropout to apply to the input (after feat extr)", + ) + + parser.add_argument( + "--dropout-features", + type=float, + default=0.0, + help="dropout to apply to the features (after feat extr)", + ) + + parser.add_argument( + "--num-classes", + type=int, + nargs="*", + default=[504], + help="""num class, a little larger than the number of cluster, + the largest is for padding, + and the value should be the multiple of 4, for faster computation""", + ) + + parser.add_argument( + "--untie-final-proj", + type=bool, + default=False, + help="use separate projection for each target", + ) + + parser.add_argument( + "--decoder-dim", + type=int, + default=512, + help="Embedding dimension in the decoder model.", + ) + + parser.add_argument( + "--joiner-dim", + type=int, + default=512, + help="""Dimension used in the joiner model. + Outputs from the encoder and decoder model are projected + to this dimension before adding. + """, + ) + + parser.add_argument( + "--use-transducer", + type=str2bool, + default=True, + help="If True, use Transducer head.", + ) + + parser.add_argument( + "--use-ctc", + type=str2bool, + default=False, + help="If True, use CTC head.", + ) + + +def get_parser(): + parser = argparse.ArgumentParser( + formatter_class=argparse.ArgumentDefaultsHelpFormatter + ) + + parser.add_argument( + "--world-size", + type=int, + default=1, + help="Number of GPUs for DDP training.", + ) + + parser.add_argument( + "--master-port", + type=int, + default=12354, + help="Master port to use for DDP training.", + ) + + parser.add_argument( + "--tensorboard", + type=str2bool, + default=True, + help="Should various information be logged in tensorboard.", + ) + + parser.add_argument( + "--num-epochs", + type=int, + default=222, + 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="hubert/exp", + help="""The experiment dir. + It specifies the directory where all training related + files, e.g., checkpoints, log, etc, are saved + """, + ) + + parser.add_argument( + "--pretrained-dir", + type=str, + help="""The pretrained model dir. + It specifies the directory where the pretrained checkpoint is saved.""", + ) + + parser.add_argument( + "--bpe-model", + type=str, + default="data/lang_bpe_500/bpe.model", + help="Path to the BPE model", + ) + + parser.add_argument( + "--base-lr", type=float, default=0.001, help="The base learning rate." + ) + + parser.add_argument( + "--lr-batches", + type=float, + default=100000, + 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=100, + help="""Number of epochs that affects how rapidly the learning rate decreases. + """, + ) + + parser.add_argument( + "--ref-duration", + type=float, + default=600, + help="Reference batch duration for purposes of adjusting batch counts for setting various " + "schedules inside the model", + ) + + parser.add_argument( + "--context-size", + type=int, + default=2, + help="The context size in the decoder. 1 means bigram; " "2 means tri-gram", + ) + + parser.add_argument( + "--prune-range", + type=int, + default=5, + help="The prune range for rnnt loss, it means how many symbols(context)" + "we are using to compute the loss", + ) + + parser.add_argument( + "--lm-scale", + type=float, + default=0.25, + help="The scale to smooth the loss with lm " + "(output of prediction network) part.", + ) + + parser.add_argument( + "--am-scale", + type=float, + default=0.0, + help="The scale to smooth the loss with am (output of encoder network)" "part.", + ) + + parser.add_argument( + "--simple-loss-scale", + type=float, + default=0.5, + help="To get pruning ranges, we will calculate a simple version" + "loss(joiner is just addition), this simple loss also uses for" + "training (as a regularization item). We will scale the simple loss" + "with this parameter before adding to the final loss.", + ) + + parser.add_argument( + "--ctc-loss-scale", + type=float, + default=0.2, + help="Scale for CTC loss.", + ) + + parser.add_argument( + "--seed", + type=int, + default=42, + help="The seed for random generators intended for reproducibility", + ) + + parser.add_argument( + "--print-diagnostics", + type=str2bool, + default=False, + help="Accumulate stats on activations, print them and exit.", + ) + + parser.add_argument( + "--sanity-check", + type=str2bool, + default=False, + help="Check if any of the batches in epoch 1 would cause OOM.", + ) + + parser.add_argument( + "--inf-check", + type=str2bool, + default=False, + help="Add hooks to check for infinite module outputs and gradients.", + ) + + parser.add_argument( + "--save-every-n", + type=int, + default=100000, + help="""Save checkpoint after processing this number of batches" + periodically. We save checkpoint to exp-dir/ whenever + params.batch_idx_train % save_every_n == 0. The checkpoint filename + has the form: f'exp-dir/checkpoint-{params.batch_idx_train}.pt' + Note: It also saves checkpoint to `exp-dir/epoch-xxx.pt` at the + end of each epoch where `xxx` is the epoch number counting from 1. + """, + ) + + parser.add_argument( + "--keep-last-k", + type=int, + default=30, + help="""Only keep this number of checkpoints on disk. + For instance, if it is 3, there are only 3 checkpoints + in the exp-dir with filenames `checkpoint-xxx.pt`. + It does not affect checkpoints with name `epoch-xxx.pt`. + """, + ) + + parser.add_argument( + "--average-period", + type=int, + default=200, + help="""Update the averaged model, namely `model_avg`, after processing + this number of batches. `model_avg` is a separate version of model, + in which each floating-point parameter is the average of all the + parameters from the start of training. Each time we take the average, + we do: `model_avg = model * (average_period / batch_idx_train) + + model_avg * ((batch_idx_train - average_period) / batch_idx_train)`. + """, + ) + + parser.add_argument( + "--accum-grad", + type=int, + default=1, + help="""update gradient when batch_idx_train % accum_grad == 0. + """, + ) + + 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 updates happen to the model so far across + epochs. + + - sub_batch_idx_train: It contains number of batch 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 + + - warm_step: The warmup period that dictates the decay of the + scale on "simple" (un-pruned) loss. + """ + params = AttributeDict( + { + "best_train_loss": float("inf"), + "best_valid_loss": float("inf"), + "best_train_epoch": -1, + "best_valid_epoch": -1, + "batch_idx_train": 0, + "sub_batch_idx_train": 0, + "log_interval": 50, + "reset_interval": 200, + "valid_interval": 3000, # For the 100h subset, use 800 + # parameters for pruned RNN-T loss + "warm_step": 2000, + "env_info": get_env_info(), + } + ) + + return params + + +def _to_int_tuple(s: str): + return tuple(map(int, s.split(","))) + + +def get_encoder_model(params: AttributeDict) -> nn.Module: + if hasattr(params, "pretrained_dir"): + logging.info(f"Loading {params.pretrained_dir}") + pretrained = torch.load(params.pretrained_dir) + encoder = HubertModel(params) + encoder.load_state_dict(pretrained["model"]) + else: + encoder = HubertModel(params) + return encoder + + +def get_decoder_model(params: AttributeDict) -> nn.Module: + decoder = Decoder( + vocab_size=params.vocab_size, + decoder_dim=params.decoder_dim, + blank_id=params.blank_id, + context_size=params.context_size, + ) + return decoder + + +def get_joiner_model(params: AttributeDict) -> nn.Module: + joiner = Joiner( + encoder_dim=max(_to_int_tuple(params.encoder_dim)), + decoder_dim=params.decoder_dim, + joiner_dim=params.joiner_dim, + vocab_size=params.vocab_size, + ) + return joiner + + +def get_model(params: AttributeDict) -> nn.Module: + assert params.use_transducer or params.use_ctc, ( + f"At least one of them should be True, " + f"but got params.use_transducer={params.use_transducer}, " + f"params.use_ctc={params.use_ctc}" + ) + + encoder = get_encoder_model(params) + + if params.use_transducer: + decoder = get_decoder_model(params) + joiner = get_joiner_model(params) + else: + decoder = None + joiner = None + + model = AsrModel( + encoder=encoder, + decoder=decoder, + joiner=joiner, + encoder_dim=max(_to_int_tuple(params.encoder_dim)), + decoder_dim=params.decoder_dim, + vocab_size=params.vocab_size, + use_transducer=params.use_transducer, + use_ctc=params.use_ctc, + ) + return model + + +def load_checkpoint_if_available( + params: AttributeDict, + model: nn.Module, + model_avg: nn.Module = None, + optimizer: Optional[torch.optim.Optimizer] = None, + scheduler: Optional[LRSchedulerType] = None, +) -> Optional[Dict[str, Any]]: + """Load checkpoint from file. + + If params.start_batch is positive, it will load the checkpoint from + `params.exp_dir/checkpoint-{params.start_batch}.pt`. Otherwise, if + params.start_epoch is larger than 1, it will load the checkpoint from + `params.start_epoch - 1`. + + Apart from loading state dict for `model` and `optimizer` it also updates + `best_train_epoch`, `best_train_loss`, `best_valid_epoch`, + and `best_valid_loss` in `params`. + + Args: + params: + The return value of :func:`get_params`. + model: + The training model. + model_avg: + The stored model averaged from the start of training. + optimizer: + The optimizer that we are using. + scheduler: + The scheduler that we are using. + Returns: + Return a dict containing previously saved training info. + """ + if params.start_batch > 0: + filename = params.exp_dir / f"checkpoint-{params.start_batch}.pt" + elif params.start_epoch > 1: + filename = params.exp_dir / f"epoch-{params.start_epoch-1}.pt" + else: + return None + + assert filename.is_file(), f"{filename} does not exist!" + + saved_params = load_checkpoint( + filename, + model=model, + model_avg=model_avg, + optimizer=optimizer, + scheduler=scheduler, + ) + + keys = [ + "best_train_epoch", + "best_valid_epoch", + "batch_idx_train", + "best_train_loss", + "best_valid_loss", + ] + for k in keys: + params[k] = saved_params[k] + + if params.start_batch > 0: + if "cur_epoch" in saved_params: + params["start_epoch"] = saved_params["cur_epoch"] + + return saved_params + + +def save_checkpoint( + params: AttributeDict, + model: Union[nn.Module, DDP], + model_avg: Optional[nn.Module] = None, + optimizer: Optional[torch.optim.Optimizer] = None, + scheduler: Optional[LRSchedulerType] = None, + sampler: Optional[CutSampler] = None, + scaler: Optional[GradScaler] = None, + rank: int = 0, +) -> None: + """Save model, optimizer, scheduler and training stats to file. + + Args: + params: + It is returned by :func:`get_params`. + model: + The training model. + model_avg: + The stored model averaged from the start of training. + optimizer: + The optimizer used in the training. + sampler: + The sampler for the training dataset. + scaler: + The scaler used for mix precision training. + """ + if rank != 0: + return + filename = params.exp_dir / f"epoch-{params.cur_epoch}.pt" + save_checkpoint_impl( + filename=filename, + model=model, + model_avg=model_avg, + params=params, + optimizer=optimizer, + scheduler=scheduler, + sampler=sampler, + scaler=scaler, + rank=rank, + ) + + if params.best_train_epoch == params.cur_epoch: + best_train_filename = params.exp_dir / "best-train-loss.pt" + copyfile(src=filename, dst=best_train_filename) + + if params.best_valid_epoch == params.cur_epoch: + best_valid_filename = params.exp_dir / "best-valid-loss.pt" + copyfile(src=filename, dst=best_valid_filename) + + +def compute_loss( + params: AttributeDict, + model: Union[nn.Module, DDP], + sp: spm.SentencePieceProcessor, + batch: dict, + is_training: bool, +) -> Tuple[Tensor, MetricsTracker]: + """ + Compute loss given the model and its inputs. + + Args: + params: + Parameters for training. See :func:`get_params`. + model: + The model for training. It is an instance of Zipformer in our case. + batch: + A batch of data. See `dataset.HubertAsrDataset()` + 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 + audio = batch["audio"].to(device) + padding_mask = batch["padding_mask"].to(device) + + batch_idx_train = params.batch_idx_train + warm_step = params.warm_step + + texts = batch["supervisions"]["text"] + y = sp.encode(texts, out_type=int) + y = k2.RaggedTensor(y) + + with torch.set_grad_enabled(is_training): + simple_loss, pruned_loss, ctc_loss, num_frames = model( + x=audio, + padding_mask=padding_mask, + y=y, + prune_range=params.prune_range, + am_scale=params.am_scale, + lm_scale=params.lm_scale, + ) + + loss = 0.0 + + if params.use_transducer: + s = params.simple_loss_scale + # take down the scale on the simple loss from 1.0 at the start + # to params.simple_loss scale by warm_step. + simple_loss_scale = ( + s + if batch_idx_train >= warm_step + else 1.0 - (batch_idx_train / warm_step) * (1.0 - s) + ) + pruned_loss_scale = ( + 1.0 + if batch_idx_train >= warm_step + else 0.1 + 0.9 * (batch_idx_train / warm_step) + ) + loss += simple_loss_scale * simple_loss + pruned_loss_scale * pruned_loss + + if params.use_ctc: + loss += params.ctc_loss_scale * ctc_loss + + assert loss.requires_grad == is_training + + info = MetricsTracker() + with warnings.catch_warnings(): + warnings.simplefilter("ignore") + info["frames"] = num_frames.sum().item() + + # Note: We use reduction=sum while computing the loss. + info["loss"] = loss.detach().cpu().item() + if params.use_transducer: + info["simple_loss"] = simple_loss.detach().cpu().item() + info["pruned_loss"] = pruned_loss.detach().cpu().item() + if params.use_ctc: + info["ctc_loss"] = ctc_loss.detach().cpu().item() + + return loss, info + + +def compute_validation_loss( + params: AttributeDict, + model: Union[nn.Module, DDP], + sp: spm.SentencePieceProcessor, + valid_dl: torch.utils.data.DataLoader, + world_size: int = 1, +) -> MetricsTracker: + """Run the validation process.""" + model.eval() + + tot_loss = MetricsTracker() + + for batch_idx, batch in enumerate(valid_dl): + loss, loss_info = compute_loss( + params=params, + model=model, + sp=sp, + batch=batch, + is_training=False, + ) + assert loss.requires_grad is False + tot_loss = tot_loss + loss_info + + if world_size > 1: + tot_loss.reduce(loss.device) + + loss_value = tot_loss["loss"] / tot_loss["frames"] + if loss_value < params.best_valid_loss: + params.best_valid_epoch = params.cur_epoch + params.best_valid_loss = loss_value + + return tot_loss + + +def train_one_epoch( + params: AttributeDict, + model: Union[nn.Module, DDP], + optimizer: torch.optim.Optimizer, + scheduler: LRSchedulerType, + sp: spm.SentencePieceProcessor, + train_dl: torch.utils.data.DataLoader, + valid_dl: torch.utils.data.DataLoader, + scaler: GradScaler, + model_avg: Optional[nn.Module] = None, + tb_writer: Optional[SummaryWriter] = None, + world_size: int = 1, + rank: int = 0, +) -> None: + """Train the model for one epoch. + + The training loss from the mean of all frames is saved in + `params.train_loss`. It runs the validation process every + `params.valid_interval` batches. + + Args: + params: + It is returned by :func:`get_params`. + model: + The model for training. + optimizer: + The optimizer we are using. + scheduler: + The learning rate scheduler, we call step() every step. + train_dl: + Dataloader for the training dataset. + valid_dl: + Dataloader for the validation dataset. + scaler: + The scaler used for mix precision training. + model_avg: + The stored model averaged from the start of training. + tb_writer: + Writer to write log messages to tensorboard. + world_size: + Number of nodes in DDP training. If it is 1, DDP is disabled. + rank: + The rank of the node in DDP training. If no DDP is used, it should + be set to 0. + """ + model.train() + + tot_loss = MetricsTracker() + + saved_bad_model = False + + def save_bad_model(suffix: str = ""): + save_checkpoint_impl( + filename=params.exp_dir / f"bad-model{suffix}-{rank}.pt", + model=model, + model_avg=model_avg, + params=params, + optimizer=optimizer, + scheduler=scheduler, + sampler=train_dl.sampler, + scaler=scaler, + rank=0, + ) + + for sub_batch_idx, batch in enumerate(train_dl): + params.sub_batch_idx_train += 1 + batch_idx = sub_batch_idx // params.accum_grad + + if batch_idx % 10 == 0: + set_batch_count(model, get_adjusted_batch_count(params)) + + batch_size = len(batch["supervisions"]["text"]) + + try: + with torch.cuda.amp.autocast(enabled=params.use_fp16): + loss, loss_info = compute_loss( + params=params, + model=model, + sp=sp, + batch=batch, + is_training=True, + ) + # summary stats + tot_loss = (tot_loss * (1 - 1 / params.reset_interval)) + loss_info + + # NOTE: We use reduction==sum and loss is computed over utterances + # in the batch and there is no normalization to it so far. + scaler.scale(loss / params.accum_grad).backward() + + if sub_batch_idx % params.accum_grad == params.accum_grad - 1: + params.batch_idx_train += 1 + scheduler.step_batch(params.batch_idx_train) + + scaler.step(optimizer) + scaler.update() + optimizer.zero_grad() + else: + continue + + except: # noqa + save_bad_model() + display_and_save_batch(batch, params=params, sp=sp) + raise + + if params.print_diagnostics and batch_idx == 5: + return + + if ( + rank == 0 + and params.batch_idx_train > 0 + and params.batch_idx_train % params.average_period == 0 + ): + update_averaged_model( + params=params, + model_cur=model, + model_avg=model_avg, + ) + + if ( + params.batch_idx_train > 0 + and params.batch_idx_train % params.save_every_n == 0 + ): + save_checkpoint_with_global_batch_idx( + out_dir=params.exp_dir, + global_batch_idx=params.batch_idx_train, + model=model, + model_avg=model_avg, + params=params, + optimizer=optimizer, + scheduler=scheduler, + sampler=train_dl.sampler, + scaler=scaler, + rank=rank, + ) + remove_checkpoints( + out_dir=params.exp_dir, + topk=params.keep_last_k, + rank=rank, + ) + + if batch_idx % 100 == 0 and params.use_fp16: + # If the grad scale was less than 1, try increasing it. The _growth_interval + # of the grad scaler is configurable, but we can't configure it to have different + # behavior depending on the current grad scale. + cur_grad_scale = scaler._scale.item() + + if cur_grad_scale < 8.0 or (cur_grad_scale < 32.0 and batch_idx % 400 == 0): + scaler.update(cur_grad_scale * 2.0) + if cur_grad_scale < 0.01: + if not saved_bad_model: + save_bad_model(suffix="-first-warning") + saved_bad_model = True + logging.warning(f"Grad scale is small: {cur_grad_scale}") + if cur_grad_scale < 1.0e-05: + save_bad_model() + raise RuntimeError( + f"grad_scale is too small, exiting: {cur_grad_scale}" + ) + + if batch_idx % params.log_interval == 0: + cur_lr = max(scheduler.get_last_lr()) + cur_grad_scale = scaler._scale.item() if params.use_fp16 else 1.0 + + logging.info( + f"Epoch {params.cur_epoch}, " + f"batch {batch_idx}, loss[{loss_info}], " + f"tot_loss[{tot_loss}], batch size: {batch_size}, " + f"lr: {cur_lr:.2e}, " + + (f"grad_scale: {scaler._scale.item()}" if params.use_fp16 else "") + ) + + if tb_writer is not None: + tb_writer.add_scalar( + "train/learning_rate", cur_lr, params.batch_idx_train + ) + + loss_info.write_summary( + tb_writer, "train/current_", params.batch_idx_train + ) + tot_loss.write_summary(tb_writer, "train/tot_", params.batch_idx_train) + if params.use_fp16: + tb_writer.add_scalar( + "train/grad_scale", cur_grad_scale, params.batch_idx_train + ) + + if batch_idx % params.valid_interval == 0 and not params.print_diagnostics: + logging.info("Computing validation loss") + valid_info = compute_validation_loss( + params=params, + model=model, + sp=sp, + valid_dl=valid_dl, + world_size=world_size, + ) + model.train() + logging.info(f"Epoch {params.cur_epoch}, validation: {valid_info}") + logging.info( + f"Maximum memory allocated so far is {torch.cuda.max_memory_allocated()//1000000}MB" + ) + if tb_writer is not None: + valid_info.write_summary( + tb_writer, "train/valid_", params.batch_idx_train + ) + + if batch_idx % params.accum_grad != params.accum_grad - 1: + optimizer.zero_grad() + loss_value = tot_loss["loss"] / tot_loss["frames"] + params.train_loss = loss_value + if params.train_loss < params.best_train_loss: + params.best_train_epoch = params.cur_epoch + params.best_train_loss = params.train_loss + + +def run(rank, world_size, args): + """ + Args: + rank: + It is a value between 0 and `world_size-1`, which is + passed automatically by `mp.spawn()` in :func:`main`. + The node with rank 0 is responsible for saving checkpoint. + world_size: + Number of GPUs for DDP training. + args: + The return value of get_parser().parse_args() + """ + params = get_params() + params.update(vars(args)) + + fix_random_seed(params.seed) + if world_size > 1: + setup_dist(rank, world_size, params.master_port) + + setup_logger(f"{params.exp_dir}/log/log-train") + logging.info("Training started") + + if args.tensorboard and rank == 0: + tb_writer = SummaryWriter(log_dir=f"{params.exp_dir}/tensorboard") + else: + tb_writer = None + + device = torch.device("cpu") + if torch.cuda.is_available(): + device = torch.device("cuda", rank) + logging.info(f"Device: {device}") + + sp = spm.SentencePieceProcessor() + sp.load(params.bpe_model) + + # is defined in local/train_bpe_model.py + params.blank_id = sp.piece_to_id("") + params.vocab_size = sp.get_piece_size() + + if not params.use_transducer: + params.ctc_loss_scale = 1.0 + + logging.info(params) + + logging.info("About to create model") + model = get_model(params) + + num_param = sum([p.numel() for p in model.parameters()]) + logging.info(f"Number of model parameters: {num_param}") + + assert params.save_every_n >= params.average_period + model_avg: Optional[nn.Module] = None + if rank == 0: + # model_avg is only used with rank 0 + model_avg = copy.deepcopy(model).to(torch.float64) + + assert params.start_epoch > 0, params.start_epoch + checkpoints = load_checkpoint_if_available( + params=params, model=model, model_avg=model_avg + ) + + model.to(device) + if world_size > 1: + logging.info("Using DDP") + model = DDP(model, device_ids=[rank], find_unused_parameters=True) + + optimizer = ScaledAdam( + get_parameter_groups_with_lrs(model, lr=params.base_lr, include_names=True), + lr=params.base_lr, # should have no effect + clipping_scale=2.0, + ) + + scheduler = Eden(optimizer, params.lr_batches, params.lr_epochs, warmup_batches=0) + + if checkpoints and "optimizer" in checkpoints: + logging.info("Loading optimizer state dict") + optimizer.load_state_dict(checkpoints["optimizer"]) + + if ( + checkpoints + and "scheduler" in checkpoints + and checkpoints["scheduler"] is not None + ): + logging.info("Loading scheduler state dict") + scheduler.load_state_dict(checkpoints["scheduler"]) + + if params.print_diagnostics: + opts = diagnostics.TensorDiagnosticOptions( + 512 + ) # allow 4 megabytes per sub-module + diagnostic = diagnostics.attach_diagnostics(model, opts) + + if params.inf_check: + register_inf_check_hooks(model) + + librispeech = LibriSpeechAsrDataModule(args) + + train_cuts = ( + librispeech.train_all_shuf_cuts() + if params.full_libri + else librispeech.train_clean_100_cuts() + ) + + def remove_short_and_long_utt(c: Cut): + # Keep only utterances with duration between 1 second and 20 seconds + # + # Caution: There is a reason to select 20.0 here. Please see + # ../local/display_manifest_statistics.py + # + # You should use ../local/display_manifest_statistics.py to get + # an utterance duration distribution for your dataset to select + # the threshold + if c.duration < 1.0 or c.duration > 20.0: + # logging.warning( + # f"Exclude cut with ID {c.id} from training. Duration: {c.duration}" + # ) + return False + + return True + + train_cuts = train_cuts.filter(remove_short_and_long_utt) + + if params.start_batch > 0 and checkpoints and "sampler" in checkpoints: + # We only load the sampler's state dict when it loads a checkpoint + # saved in the middle of an epoch + sampler_state_dict = checkpoints["sampler"] + else: + sampler_state_dict = None + + train_dl = librispeech.train_dataloaders( + train_cuts, + do_normalize=params.do_normalize, + 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, + do_normalize=params.do_normalize, + ) + + if params.sanity_check and not params.print_diagnostics: + scan_pessimistic_batches_for_oom( + model=model, + train_dl=train_dl, + optimizer=optimizer, + sp=sp, + params=params, + ) + + scaler = GradScaler(enabled=params.use_fp16, init_scale=1.0) + if checkpoints and "grad_scaler" in checkpoints: + logging.info("Loading grad scaler state dict") + scaler.load_state_dict(checkpoints["grad_scaler"]) + + for epoch in range(params.start_epoch, params.num_epochs + 1): + scheduler.step_epoch(epoch - 1) + fix_random_seed(params.seed + epoch - 1) + train_dl.sampler.set_epoch(epoch - 1) + + if tb_writer is not None: + tb_writer.add_scalar("train/epoch", epoch, params.batch_idx_train) + + params.cur_epoch = epoch + + train_one_epoch( + params=params, + model=model, + model_avg=model_avg, + optimizer=optimizer, + scheduler=scheduler, + sp=sp, + train_dl=train_dl, + valid_dl=valid_dl, + scaler=scaler, + tb_writer=tb_writer, + world_size=world_size, + rank=rank, + ) + + if params.print_diagnostics: + diagnostic.print_diagnostics() + break + + save_checkpoint( + params=params, + model=model, + model_avg=model_avg, + optimizer=optimizer, + scheduler=scheduler, + sampler=train_dl.sampler, + scaler=scaler, + rank=rank, + ) + + logging.info("Done!") + + if world_size > 1: + torch.distributed.barrier() + cleanup_dist() + + +def display_and_save_batch( + batch: dict, + params: AttributeDict, + sp: spm.SentencePieceProcessor, +) -> None: + """Display the batch statistics and save the batch into disk. + + Args: + batch: + A batch of data. See `dataset.HubertAsrDataset()` + for the content in it. + params: + Parameters for training. See :func:`get_params`. + sp: + The BPE model. + """ + from lhotse.utils import uuid4 + + filename = f"{params.exp_dir}/batch-{uuid4()}.pt" + logging.info(f"Saving batch to {filename}") + torch.save(batch, filename) + + audio = batch["audio"] + logging.info(f"audio shape: {audio.shape}") + + y = sp.encode(batch["supervisions"]["text"], out_type=int) + num_tokens = sum(len(i) for i in y) + logging.info(f"num tokens: {num_tokens}") + + +def scan_pessimistic_batches_for_oom( + model: Union[nn.Module, DDP], + train_dl: torch.utils.data.DataLoader, + optimizer: torch.optim.Optimizer, + sp: spm.SentencePieceProcessor, + params: AttributeDict, +): + from lhotse.dataset import find_pessimistic_batches + + logging.info( + "Sanity check -- see if any of the batches in epoch 1 would cause OOM." + ) + batches, crit_values = find_pessimistic_batches(train_dl.sampler) + for criterion, cuts in batches.items(): + batch = train_dl.dataset[cuts] + try: + with torch.cuda.amp.autocast(enabled=params.use_fp16): + loss, _ = compute_loss( + params=params, + model=model, + sp=sp, + batch=batch, + is_training=True, + ) + loss.backward() + optimizer.zero_grad() + except Exception as e: + if "CUDA out of memory" in str(e): + logging.error( + "Your GPU ran out of memory with the current " + "max_duration setting. We recommend decreasing " + "max_duration and trying again.\n" + f"Failing criterion: {criterion} " + f"(={crit_values[criterion]}) ..." + ) + display_and_save_batch(batch, params=params, sp=sp) + raise + logging.info( + f"Maximum memory allocated so far is {torch.cuda.max_memory_allocated()//1000000}MB" + ) + + +def main(): + parser = get_parser() + LibriSpeechAsrDataModule.add_arguments(parser) + args = parser.parse_args() + args.exp_dir = Path(args.exp_dir) + + world_size = args.world_size + assert world_size >= 1 + if world_size > 1: + mp.spawn(run, args=(world_size, args), nprocs=world_size, join=True) + else: + run(rank=0, world_size=1, args=args) + + +torch.set_num_threads(1) +torch.set_num_interop_threads(1) + +if __name__ == "__main__": + main() diff --git a/egs/librilight/SSL/zipformer/hubert_ce.py b/egs/librilight/SSL/zipformer/hubert_ce.py new file mode 120000 index 0000000000..2b8482f786 --- /dev/null +++ b/egs/librilight/SSL/zipformer/hubert_ce.py @@ -0,0 +1 @@ +../../../librispeech/SSL/zipformer/hubert_ce.py \ No newline at end of file diff --git a/egs/librilight/SSL/zipformer/joiner.py b/egs/librilight/SSL/zipformer/joiner.py new file mode 120000 index 0000000000..587823e654 --- /dev/null +++ b/egs/librilight/SSL/zipformer/joiner.py @@ -0,0 +1 @@ +../../../librispeech/SSL/zipformer/joiner.py \ No newline at end of file diff --git a/egs/librilight/SSL/zipformer/model.py b/egs/librilight/SSL/zipformer/model.py new file mode 120000 index 0000000000..ca3daacca6 --- /dev/null +++ b/egs/librilight/SSL/zipformer/model.py @@ -0,0 +1 @@ +../../../librispeech/SSL/zipformer/model.py \ No newline at end of file diff --git a/egs/librilight/SSL/zipformer/optim.py b/egs/librilight/SSL/zipformer/optim.py new file mode 120000 index 0000000000..bd2153ebf4 --- /dev/null +++ b/egs/librilight/SSL/zipformer/optim.py @@ -0,0 +1 @@ +../../../librispeech/SSL/zipformer/optim.py \ No newline at end of file diff --git a/egs/librilight/SSL/zipformer/pretrain.py b/egs/librilight/SSL/zipformer/pretrain.py new file mode 100644 index 0000000000..5728dbe752 --- /dev/null +++ b/egs/librilight/SSL/zipformer/pretrain.py @@ -0,0 +1,1366 @@ +#!/usr/bin/env python3 +# Copyright 2021-2024 Xiaomi Corp. (authors: Fangjun Kuang, +# Wei Kang, +# Mingshuang Luo, +# Zengwei Yao, +# Yifan Yang, +# Daniel Povey) +# +# Copyright 2024 Shanghai Jiao Tong University (authors: Jianheng Zhuo) +# +# 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="0,1,2,3,4,5,6,7" + +# For hubert model pretraining: +./zipformer/pretrain.py \ + --world-size 8 \ + --num-epochs 400 \ + --start-epoch 1 \ + --use-fp16 1 \ + --exp-dir zipformer/exp \ + --max-duration 87.5 \ + --accum-grad 4 +""" + + +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 optim +import torch +import torch.multiprocessing as mp +import torch.nn as nn +from hubert_ce import HubertModel +from lhotse.cut import Cut +from lhotse.dataset.sampling.base import CutSampler +from lhotse.utils import fix_random_seed +from optim import Eden, ScaledAdam +from ssl_datamodule import LibriLightDataModule +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.hooks import register_inf_check_hooks +from icefall.utils import ( + AttributeDict, + MetricsTracker, + get_parameter_groups_with_lrs, + setup_logger, + str2bool, +) + +LRSchedulerType = Union[torch.optim.lr_scheduler._LRScheduler, optim.LRScheduler] + + +def get_adjusted_batch_count(params: AttributeDict) -> float: + # returns the number of batches we would have used so far if we had used the reference + # duration. This is for purposes of set_batch_count(). + return ( + params.batch_idx_train + * params.accum_grad + * (params.max_duration * params.world_size) + / params.ref_duration + ) + + +def set_batch_count(model: Union[nn.Module, DDP], batch_count: float) -> None: + if isinstance(model, DDP): + # get underlying nn.Module + model = model.module + for name, module in model.named_modules(): + if hasattr(module, "batch_count"): + module.batch_count = batch_count + if hasattr(module, "name"): + module.name = name + + +def add_model_arguments(parser: argparse.ArgumentParser): + parser.add_argument( + "--num-encoder-layers", + type=str, + default="2,2,3,4,3,2", + help="Number of zipformer encoder layers per stack, comma separated.", + ) + + parser.add_argument( + "--downsampling-factor", + type=str, + default="1,2,4,8,4,2", + help="Downsampling factor for each stack of encoder layers.", + ) + + parser.add_argument( + "--feedforward-dim", + type=str, + default="512,768,1024,1536,1024,768", + help="Feedforward dimension of the zipformer encoder layers, per stack, comma separated.", + ) + + parser.add_argument( + "--num-heads", + type=str, + default="4,4,4,8,4,4", + help="Number of attention heads in the zipformer encoder layers: a single int or comma-separated list.", + ) + + parser.add_argument( + "--encoder-dim", + type=str, + default="192,256,384,512,384,256", + help="Embedding dimension in encoder stacks: a single int or comma-separated list.", + ) + + parser.add_argument( + "--query-head-dim", + type=str, + default="32", + help="Query/key dimension per head in encoder stacks: a single int or comma-separated list.", + ) + + parser.add_argument( + "--value-head-dim", + type=str, + default="12", + help="Value dimension per head in encoder stacks: a single int or comma-separated list.", + ) + + parser.add_argument( + "--pos-head-dim", + type=str, + default="4", + help="Positional-encoding dimension per head in encoder stacks: a single int or comma-separated list.", + ) + + parser.add_argument( + "--pos-dim", + type=int, + default="48", + help="Positional-encoding embedding dimension", + ) + + parser.add_argument( + "--encoder-unmasked-dim", + type=str, + default="192,192,256,256,256,192", + help="Unmasked dimensions in the encoders, relates to augmentation during training. " + "A single int or comma-separated list. Must be <= each corresponding encoder_dim.", + ) + + parser.add_argument( + "--cnn-module-kernel", + type=str, + default="31,31,15,15,15,31", + help="Sizes of convolutional kernels in convolution modules in each encoder stack: " + "a single int or comma-separated list.", + ) + + # hubert parameters + parser.add_argument( + "--label-rate", + type=float, + default=50, + ) + + parser.add_argument( + "--sample-rate", + type=float, + default=16000, + ) + + parser.add_argument( + "--extractor-mode", + type=str, + default="default", + help="""mode for feature extractor, should in EXTRACTOR_MODE_CHOICES. default has a single group + norm with d groups in the first conv block, whereas layer_norm + has layer norms in every block (meant to use with normalize=True)""", + ) + + parser.add_argument( + "--conv-feature-layers", + type=str, + default="[(512,10,5)] + [(512,3,2)] * 4 + [(512,2,2)] * 2", + help="string describing convolutional feature extraction layers in form of a python list that contains [(dim, kernel_size, stride), ...]", + ) + + parser.add_argument( + "--conv-bias", type=bool, default=False, help="include bias in conv encoder" + ) + + parser.add_argument( + "--feature-grad-mult", + type=float, + default=1.0, + help="multiply feature extractor var grads by this", + ) + + # masking + parser.add_argument("--mask-length", type=int, default=10, help="mask_length") + + parser.add_argument( + "--mask-prob", + type=float, + default=0.65, + help="probability of replacing a token with mask", + ) + + parser.add_argument( + "--mask-selection", + type=str, + choices=["static", "uniform", "normal", "poisson"], + default="static", + help="how to choose mask length", + ) + + parser.add_argument( + "--mask-other", + type=float, + default=0, + help="secondary mask argument (used for more complex distributions),see help in compute_mask_indicesh", + ) + + parser.add_argument( + "--no-mask-overlap", + type=bool, + default=False, + help="whether to allow masks to overlap", + ) + + parser.add_argument( + "--mask-min-space", + type=int, + default=1, + help="min space between spans (if no overlap is enabled)", + ) + + # channel masking + parser.add_argument( + "--mask-channel-length", + type=int, + default=10, + help="length of the mask for features (channels)", + ) + + parser.add_argument( + "--mask-channel-prob", + type=float, + default=0.0, + help="probability of replacing a feature with 0", + ) + + parser.add_argument( + "--mask-channel-selection", + type=str, + choices=["static", "uniform", "normal", "poisson"], + default="static", + help="how to choose mask length for channel masking", + ) + + parser.add_argument( + "--mask-channel-other", + type=float, + default=0, + help="secondary mask argument (used for more complex distributions), see help in compute_mask_indicesh", + ) + + parser.add_argument( + "--no-mask-channel-overlap", + type=bool, + default=False, + help="whether to allow channel masks to overlap", + ) + + parser.add_argument( + "--mask-channel-min-space", + type=int, + default=1, + help="min space between spans (if no overlap is enabled)", + ) + + # loss computation + parser.add_argument( + "--skip-masked", + type=bool, + default=False, + help="skip computing losses over masked frames", + ) + + parser.add_argument( + "--skip-nomask", + type=bool, + default=False, + help="skip computing losses over unmasked frames", + ) + + parser.add_argument( + "--checkpoint-activations", + type=bool, + default=False, + help="recompute activations and save memory for extra compute", + ) + + parser.add_argument( + "--pred-masked-weight", + type=float, + default=1, + help="weight for masked part in ssl loss", + ) + + parser.add_argument( + "--pred-nomask-weight", + type=float, + default=0, + help="weight for masked part in ssl loss", + ) + + parser.add_argument( + "--loss-weights", + type=float, + nargs="*", + default=[10], + help="weight for masked part in ssl loss", + ) + + # FP16 optimization + parser.add_argument( + "--required-seq-len-multiple", + type=int, + default=2, + help="pad the input to encoder such that the sequence length is divisible by multiple", + ) + + parser.add_argument( + "--attn-type", type=str, default="", help="if espnet use ESPNET MHA" + ) + + parser.add_argument( + "--pos-enc-type", + type=str, + default="abs", + help="Positional encoding type to use in conformer", + ) + + parser.add_argument( + "--logit-temp", type=float, default=0.1, help="temperature to divide logits by" + ) + + parser.add_argument( + "--dropout-input", + type=float, + default=0.0, + help="dropout to apply to the input (after feat extr)", + ) + + parser.add_argument( + "--dropout-features", + type=float, + default=0.0, + help="dropout to apply to the features (after feat extr)", + ) + + parser.add_argument( + "--num-classes", + type=int, + nargs="*", + default=[504], + help="""num class, a little larger than the number of cluster, + the largest is for padding, + and the value should be the multiple of 4, for faster computation""", + ) + + parser.add_argument( + "--untie-final-proj", + type=bool, + default=False, + help="use separate projection for each target", + ) + + +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=400, + help="Number of epochs to train.", + ) + + parser.add_argument( + "--start-epoch", + type=int, + default=1, + help="""Resume training from this epoch. It should be positive. + If larger than 1, it will load checkpoint from + exp-dir/epoch-{start_epoch-1}.pt + """, + ) + + parser.add_argument( + "--start-batch", + type=int, + default=0, + help="""If positive, --start-epoch is ignored and + it loads the checkpoint from exp-dir/checkpoint-{start_batch}.pt + """, + ) + + parser.add_argument( + "--exp-dir", + type=str, + default="zipformer/exp", + help="""The experiment dir. + It specifies the directory where all training related + files, e.g., checkpoints, log, etc, are saved + """, + ) + + parser.add_argument( + "--base-lr", type=float, default=0.045, help="The base learning rate." + ) + + parser.add_argument( + "--lr-batches", + type=float, + default=7500, + help="""Number of steps that affects how rapidly the learning rate + decreases. We suggest not to change this.""", + ) + + parser.add_argument( + "--lr-epochs", + type=float, + default=10.5, + help="""Number of epochs that affects how rapidly the learning rate decreases. + """, + ) + + parser.add_argument( + "--warmup-batches", + type=float, + default=5000, + help="Eden warmup steps", + ) + + parser.add_argument( + "--warmup-start", + type=float, + default=0, + help="Eden warmup start learning rate", + ) + + parser.add_argument( + "--ref-duration", + type=float, + default=600, + help="Reference batch duration for purposes of adjusting batch counts for setting various " + "schedules inside the model", + ) + + parser.add_argument( + "--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( + "--sanity-check", + type=str2bool, + default=False, + help="Check if any of the batches in epoch 1 would cause OOM.", + ) + + parser.add_argument( + "--inf-check", + type=str2bool, + default=False, + help="Add hooks to check for infinite module outputs and gradients.", + ) + + parser.add_argument( + "--save-every-n", + type=int, + default=100000, + help="""Save checkpoint after processing this number of batches" + periodically. We save checkpoint to exp-dir/ whenever + params.batch_idx_train % save_every_n == 0. The checkpoint filename + has the form: f'exp-dir/checkpoint-{params.batch_idx_train}.pt' + Note: It also saves checkpoint to `exp-dir/epoch-xxx.pt` at the + end of each epoch where `xxx` is the epoch number counting from 1. + """, + ) + + parser.add_argument( + "--keep-last-k", + type=int, + default=30, + help="""Only keep this number of checkpoints on disk. + For instance, if it is 3, there are only 3 checkpoints + in the exp-dir with filenames `checkpoint-xxx.pt`. + It does not affect checkpoints with name `epoch-xxx.pt`. + """, + ) + + parser.add_argument( + "--average-period", + type=int, + default=200, + help="""Update the averaged model, namely `model_avg`, after processing + this number of batches. `model_avg` is a separate version of model, + in which each floating-point parameter is the average of all the + parameters from the start of training. Each time we take the average, + we do: `model_avg = model * (average_period / batch_idx_train) + + model_avg * ((batch_idx_train - average_period) / batch_idx_train)`. + """, + ) + + parser.add_argument( + "--accum-grad", + type=int, + default=4, + help="""update gradient when batch_idx_train % accum_grad == 0. + """, + ) + + parser.add_argument( + "--use-fp16", + type=str2bool, + default=False, + help="Whether to use half precision training.", + ) + + parser.add_argument( + "--max-sample-size", + type=float, + default=250000, + help="max sample size", + ) + + parser.add_argument( + "--min-sample-size", + type=float, + default=32000, + help="min sample size", + ) + + 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 updates happen to the model so far across + epochs. + + - sub_batch_idx_train: It contains number of batch 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 + """ + params = AttributeDict( + { + "best_train_loss": float("inf"), + "best_valid_loss": float("inf"), + "best_train_epoch": -1, + "best_valid_epoch": -1, + "batch_idx_train": 0, + "sub_batch_idx_train": 0, + "log_interval": 50, + "reset_interval": 200, + "valid_interval": 3000, # For the 100h subset, use 800 + "env_info": get_env_info(), + } + ) + + return params + + +def _to_int_tuple(s: str): + return tuple(map(int, s.split(","))) + + +def get_model(params: AttributeDict) -> nn.Module: + model = HubertModel(params) + 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], + batch: dict, + is_training: bool, +) -> Tuple[Tensor, MetricsTracker]: + """ + Compute loss given the model and its inputs. + + Args: + params: + Parameters for training. See :func:`get_params`. + model: + The model for training. It is an instance of Zipformer in our case. + batch: + A batch of data. See `dataset.HubertDataset()` + for the content in it. + is_training: + True for training. False for validation. When it is True, this + function enables autograd during computation; when it is False, it + disables autograd. + """ + device = model.device if isinstance(model, DDP) else next(model.parameters()).device + audio = batch["audio"].to(device) + padding_mask = batch["padding_mask"].to(device) + kmeans = batch["kmeans"].to(device) + + with torch.set_grad_enabled(is_training): + loss, num_masked_tokens, logging_output = model( + source=audio, target_list=[kmeans], padding_mask=padding_mask + ) + + assert loss.requires_grad == is_training + + info = MetricsTracker() + with warnings.catch_warnings(): + warnings.simplefilter("ignore") + info["frames"] = num_masked_tokens + for item in logging_output: + info[item] = logging_output[item] + return loss, info + + +def compute_validation_loss( + params: AttributeDict, + model: Union[nn.Module, DDP], + valid_dl: torch.utils.data.DataLoader, + world_size: int = 1, +) -> MetricsTracker: + """Run the validation process.""" + model.eval() + + tot_loss = MetricsTracker() + + for batch_idx, batch in enumerate(valid_dl): + loss, loss_info = compute_loss( + params=params, + model=model, + batch=batch, + is_training=False, + ) + assert loss.requires_grad is False + tot_loss = tot_loss + loss_info + + if world_size > 1: + tot_loss.reduce(loss.device) + + loss_value = tot_loss["loss"] / tot_loss["frames"] + if loss_value < params.best_valid_loss: + params.best_valid_epoch = params.cur_epoch + params.best_valid_loss = loss_value + + return tot_loss + + +def train_one_epoch( + params: AttributeDict, + model: Union[nn.Module, DDP], + optimizer: torch.optim.Optimizer, + scheduler: LRSchedulerType, + train_dl: torch.utils.data.DataLoader, + valid_dl: torch.utils.data.DataLoader, + scaler: GradScaler, + model_avg: Optional[nn.Module] = None, + tb_writer: Optional[SummaryWriter] = None, + world_size: int = 1, + rank: int = 0, +) -> None: + """Train the model for one epoch. + + The training loss from the mean of all frames is saved in + `params.train_loss`. It runs the validation process every + `params.valid_interval` batches. + + Args: + params: + It is returned by :func:`get_params`. + model: + The model for training. + optimizer: + The optimizer we are using. + scheduler: + The learning rate scheduler, we call step() every step. + train_dl: + Dataloader for the training dataset. + valid_dl: + Dataloader for the validation dataset. + scaler: + The scaler used for mix precision training. + model_avg: + The stored model averaged from the start of training. + tb_writer: + Writer to write log messages to tensorboard. + world_size: + Number of nodes in DDP training. If it is 1, DDP is disabled. + rank: + The rank of the node in DDP training. If no DDP is used, it should + be set to 0. + """ + model.train() + + tot_loss = MetricsTracker() + + saved_bad_model = False + + def save_bad_model(suffix: str = ""): + save_checkpoint_impl( + filename=params.exp_dir / f"bad-model{suffix}-{rank}.pt", + model=model, + model_avg=model_avg, + params=params, + optimizer=optimizer, + scheduler=scheduler, + sampler=train_dl.sampler, + scaler=scaler, + rank=0, + ) + + for sub_batch_idx, batch in enumerate(train_dl): + params.sub_batch_idx_train += 1 + batch_idx = sub_batch_idx // params.accum_grad + + if batch_idx % 10 == 0: + set_batch_count(model, get_adjusted_batch_count(params)) + + batch_size = batch["kmeans"].shape[0] + + try: + with torch.cuda.amp.autocast(enabled=params.use_fp16): + loss, loss_info = compute_loss( + params=params, + model=model, + batch=batch, + is_training=True, + ) + # summary stats + tot_loss = (tot_loss * (1 - 1 / params.reset_interval)) + loss_info + + # NOTE: We use reduction==sum and loss is computed over utterances + # in the batch and there is no normalization to it so far. + scaler.scale(loss / params.accum_grad).backward() + + if sub_batch_idx % params.accum_grad == params.accum_grad - 1: + params.batch_idx_train += 1 + scheduler.step_batch(params.batch_idx_train) + + scaler.step(optimizer) + scaler.update() + optimizer.zero_grad() + else: + continue + + except: # noqa + save_bad_model() + display_and_save_batch(batch, params=params) + raise + + if params.print_diagnostics and batch_idx == 5: + return + + if ( + rank == 0 + and params.batch_idx_train > 0 + and params.batch_idx_train % params.average_period == 0 + ): + update_averaged_model( + params=params, + model_cur=model, + model_avg=model_avg, + ) + + if ( + params.batch_idx_train > 0 + and params.batch_idx_train % params.save_every_n == 0 + ): + save_checkpoint_with_global_batch_idx( + out_dir=params.exp_dir, + global_batch_idx=params.batch_idx_train, + model=model, + model_avg=model_avg, + params=params, + optimizer=optimizer, + scheduler=scheduler, + sampler=train_dl.sampler, + scaler=scaler, + rank=rank, + ) + remove_checkpoints( + out_dir=params.exp_dir, + topk=params.keep_last_k, + rank=rank, + ) + + if batch_idx % 100 == 0 and params.use_fp16: + # If the grad scale was less than 1, try increasing it. The _growth_interval + # of the grad scaler is configurable, but we can't configure it to have different + # behavior depending on the current grad scale. + cur_grad_scale = scaler._scale.item() + + if cur_grad_scale < 8.0 or (cur_grad_scale < 32.0 and batch_idx % 400 == 0): + scaler.update(cur_grad_scale * 2.0) + if cur_grad_scale < 0.01: + if not saved_bad_model: + save_bad_model(suffix="-first-warning") + saved_bad_model = True + logging.warning(f"Grad scale is small: {cur_grad_scale}") + if cur_grad_scale < 1.0e-05: + save_bad_model() + raise RuntimeError( + f"grad_scale is too small, exiting: {cur_grad_scale}" + ) + + if batch_idx % params.log_interval == 0: + cur_lr = max(scheduler.get_last_lr()) + cur_grad_scale = scaler._scale.item() if params.use_fp16 else 1.0 + + logging.info( + f"Epoch {params.cur_epoch}, " + f"batch {batch_idx}, loss[{loss_info}], " + f"tot_loss[{tot_loss}], batch size: {batch_size}, " + f"lr: {cur_lr:.2e}, " + + (f"grad_scale: {scaler._scale.item()}" if params.use_fp16 else "") + ) + + if tb_writer is not None: + tb_writer.add_scalar( + "train/learning_rate", cur_lr, params.batch_idx_train + ) + + loss_info.write_summary( + tb_writer, "train/current_", params.batch_idx_train + ) + tot_loss.write_summary(tb_writer, "train/tot_", params.batch_idx_train) + if params.use_fp16: + tb_writer.add_scalar( + "train/grad_scale", cur_grad_scale, params.batch_idx_train + ) + + if batch_idx % params.valid_interval == 0 and not params.print_diagnostics: + logging.info("Computing validation loss") + valid_info = compute_validation_loss( + params=params, + model=model, + valid_dl=valid_dl, + world_size=world_size, + ) + model.train() + logging.info(f"Epoch {params.cur_epoch}, validation: {valid_info}") + logging.info( + f"Maximum memory allocated so far is {torch.cuda.max_memory_allocated()//1000000}MB" + ) + if tb_writer is not None: + valid_info.write_summary( + tb_writer, "train/valid_", params.batch_idx_train + ) + + if batch_idx % params.accum_grad != params.accum_grad - 1: + optimizer.zero_grad() + loss_value = tot_loss["loss"] / tot_loss["frames"] + params.train_loss = loss_value + if params.train_loss < params.best_train_loss: + params.best_train_epoch = params.cur_epoch + params.best_train_loss = params.train_loss + + +def run(rank, world_size, args): + """ + Args: + rank: + It is a value between 0 and `world_size-1`, which is + passed automatically by `mp.spawn()` in :func:`main`. + The node with rank 0 is responsible for saving checkpoint. + world_size: + Number of GPUs for DDP training. + args: + The return value of get_parser().parse_args() + """ + params = get_params() + params.update(vars(args)) + + fix_random_seed(params.seed) + if world_size > 1: + setup_dist(rank, world_size, params.master_port) + + setup_logger(f"{params.exp_dir}/log/log-train") + logging.info("Training started") + + if args.tensorboard and rank == 0: + tb_writer = SummaryWriter(log_dir=f"{params.exp_dir}/tensorboard") + else: + tb_writer = None + + device = torch.device("cpu") + if torch.cuda.is_available(): + device = torch.device("cuda", rank) + logging.info(f"Device: {device}") + logging.info(params) + + logging.info("About to create model") + model = get_model(params) + + num_param = sum([p.numel() for p in model.parameters()]) + logging.info(f"Number of model parameters: {num_param}") + + assert params.save_every_n >= params.average_period + model_avg: Optional[nn.Module] = None + if rank == 0: + # model_avg is only used with rank 0 + model_avg = copy.deepcopy(model).to(torch.float64) + + assert params.start_epoch > 0, params.start_epoch + checkpoints = load_checkpoint_if_available( + params=params, model=model, model_avg=model_avg + ) + + model.to(device) + if world_size > 1: + logging.info("Using DDP") + model = DDP(model, device_ids=[rank], find_unused_parameters=True) + + optimizer = ScaledAdam( + get_parameter_groups_with_lrs(model, lr=params.base_lr, include_names=True), + lr=params.base_lr, # should have no effect + clipping_scale=2.0, + ) + + scheduler = Eden( + optimizer, + params.lr_batches, + params.lr_epochs, + params.warmup_batches, + params.warmup_start, + ) + + if checkpoints and "optimizer" in checkpoints: + logging.info("Loading optimizer state dict") + optimizer.load_state_dict(checkpoints["optimizer"]) + + if ( + checkpoints + and "scheduler" in checkpoints + and checkpoints["scheduler"] is not None + ): + logging.info("Loading scheduler state dict") + scheduler.load_state_dict(checkpoints["scheduler"]) + + if params.print_diagnostics: + opts = diagnostics.TensorDiagnosticOptions( + 512 + ) # allow 4 megabytes per sub-module + diagnostic = diagnostics.attach_diagnostics(model, opts) + + if params.inf_check: + register_inf_check_hooks(model) + + librilight = LibriLightDataModule(args) + + train_cuts = librilight.train_all_shuf_cuts() + + def remove_short_and_long_utt(c: Cut): + # Keep only utterances with duration between 1 second and 20 seconds + # + # Caution: There is a reason to select 20.0 here. Please see + # ../local/display_manifest_statistics.py + # + # You should use ../local/display_manifest_statistics.py to get + # an utterance duration distribution for your dataset to select + # the threshold + if ( + c.duration < params.min_sample_size / params.sample_rate + or c.duration > params.max_sample_size / params.sample_rate + ): + # logging.warning( + # f"Exclude cut with ID {c.id} from training. Duration: {c.duration}" + # ) + return False + + return True + + train_cuts = train_cuts.filter(remove_short_and_long_utt) + + if params.start_batch > 0 and checkpoints and "sampler" in checkpoints: + # We only load the sampler's state dict when it loads a checkpoint + # saved in the middle of an epoch + sampler_state_dict = checkpoints["sampler"] + else: + sampler_state_dict = None + + train_dl = librilight.train_dataloaders( + train_cuts, + sample_rate=params.sample_rate, + label_rate=params.label_rate, + random_crop=params.random_crop, + pad_audio=False, + num_classes=params.num_classes, + do_normalize=params.do_normalize, + sampler_state_dict=sampler_state_dict, + ) + + valid_cuts = librilight.dev_clean_cuts() + # valid_cuts += librilight.dev_other_cuts() + valid_cuts = valid_cuts.filter(remove_short_and_long_utt) + + valid_dl = librilight.valid_dataloaders( + valid_cuts, + sample_rate=params.sample_rate, + label_rate=params.label_rate, + random_crop=params.random_crop, + pad_audio=False, + num_classes=params.num_classes, + do_normalize=params.do_normalize, + ) + + if params.sanity_check and not params.print_diagnostics: + scan_pessimistic_batches_for_oom( + model=model, + train_dl=train_dl, + optimizer=optimizer, + params=params, + ) + + scaler = GradScaler(enabled=params.use_fp16, init_scale=1.0) + if checkpoints and "grad_scaler" in checkpoints: + logging.info("Loading grad scaler state dict") + scaler.load_state_dict(checkpoints["grad_scaler"]) + + for epoch in range(params.start_epoch, params.num_epochs + 1): + scheduler.step_epoch(epoch - 1) + fix_random_seed(params.seed + epoch - 1) + train_dl.sampler.set_epoch(epoch - 1) + + if tb_writer is not None: + tb_writer.add_scalar("train/epoch", epoch, params.batch_idx_train) + + params.cur_epoch = epoch + + train_one_epoch( + params=params, + model=model, + model_avg=model_avg, + optimizer=optimizer, + scheduler=scheduler, + train_dl=train_dl, + valid_dl=valid_dl, + scaler=scaler, + tb_writer=tb_writer, + world_size=world_size, + rank=rank, + ) + + if params.print_diagnostics: + diagnostic.print_diagnostics() + break + + save_checkpoint( + params=params, + model=model, + model_avg=model_avg, + optimizer=optimizer, + scheduler=scheduler, + sampler=train_dl.sampler, + scaler=scaler, + rank=rank, + ) + + logging.info("Done!") + + if world_size > 1: + torch.distributed.barrier() + cleanup_dist() + + +def display_and_save_batch( + batch: dict, + params: AttributeDict, +) -> None: + """Display the batch statistics and save the batch into disk. + + Args: + batch: + A batch of data. See `dataset.HubertDataset()` + for the content in it. + params: + Parameters for training. See :func:`get_params`. + sp: + The BPE model. + """ + from lhotse.utils import uuid4 + + filename = f"{params.exp_dir}/batch-{uuid4()}.pt" + logging.info(f"Saving batch to {filename}") + torch.save(batch, filename) + + audio = batch["audio"] + logging.info(f"audio shape: {audio.shape}") + + +def scan_pessimistic_batches_for_oom( + model: Union[nn.Module, DDP], + train_dl: torch.utils.data.DataLoader, + optimizer: torch.optim.Optimizer, + params: AttributeDict, +): + from lhotse.dataset import find_pessimistic_batches + + logging.info( + "Sanity check -- see if any of the batches in epoch 1 would cause OOM." + ) + batches, crit_values = find_pessimistic_batches(train_dl.sampler) + for criterion, cuts in batches.items(): + batch = train_dl.dataset[cuts] + try: + with torch.cuda.amp.autocast(enabled=params.use_fp16): + loss, _ = compute_loss( + params=params, + model=model, + batch=batch, + is_training=True, + ) + loss.backward() + optimizer.zero_grad() + except Exception as e: + if "CUDA out of memory" in str(e): + logging.error( + "Your GPU ran out of memory with the current " + "max_duration setting. We recommend decreasing " + "max_duration and trying again.\n" + f"Failing criterion: {criterion} " + f"(={crit_values[criterion]}) ..." + ) + display_and_save_batch(batch, params=params) + raise + logging.info( + f"Maximum memory allocated so far is {torch.cuda.max_memory_allocated()//1000000}MB" + ) + + +def main(): + parser = get_parser() + LibriLightDataModule.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/librilight/SSL/zipformer/scaling.py b/egs/librilight/SSL/zipformer/scaling.py new file mode 120000 index 0000000000..24b661dfba --- /dev/null +++ b/egs/librilight/SSL/zipformer/scaling.py @@ -0,0 +1 @@ +../../../librispeech/SSL/zipformer/scaling.py \ No newline at end of file diff --git a/egs/librilight/SSL/zipformer/ssl_datamodule.py b/egs/librilight/SSL/zipformer/ssl_datamodule.py new file mode 100644 index 0000000000..dc0dbec6ca --- /dev/null +++ b/egs/librilight/SSL/zipformer/ssl_datamodule.py @@ -0,0 +1,334 @@ +# Copyright 2021 Piotr Żelasko +# Copyright 2023 Xiaomi Corporation (Author: Yifan Yang) +# +# See ../../../../LICENSE for clarification regarding multiple authors +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + + +import argparse +import glob +import logging +import re +from functools import lru_cache +from pathlib import Path +from typing import Any, Dict, Optional + +import torch +from dataset import HubertDataset +from lhotse import CutSet, combine, load_manifest_lazy +from lhotse.dataset import DynamicBucketingSampler, SimpleCutSampler +from lhotse.utils import fix_random_seed +from torch.utils.data import DataLoader + +from icefall.utils import str2bool + + +class _SeedWorkers: + def __init__(self, seed: int): + self.seed = seed + + def __call__(self, worker_id: int): + fix_random_seed(self.seed + worker_id) + + +class LibriLightDataModule: + """ + DataModule for SSL experiments. + It assumes there is always one train and valid dataloader, + but there can be multiple test dataloaders (e.g. LibriSpeech test-clean + and test-other). + + It contains all the common data pipeline modules used in SSL + experiments, e.g.: + - dynamic batch size, + - bucketing samplers, + + This class should be derived for specific corpora used in SSL tasks. + """ + + def __init__(self, args: argparse.Namespace): + self.args = args + + @classmethod + def add_arguments(cls, parser: argparse.ArgumentParser): + group = parser.add_argument_group( + title="ASR SSL related options", + description="These options are used for the preparation of " + "PyTorch DataLoaders from Lhotse CutSet's -- they control the " + "effective batch sizes, sampling strategies.", + ) + + group.add_argument( + "--manifest-dir", + type=Path, + default=Path("data/kmeans"), + help="Path to directory with train/valid/test cuts.", + ) + group.add_argument( + "--max-duration", + type=float, + default=200.0, + help="Maximum pooled recordings duration (seconds) in a " + "single batch. You can reduce it if it causes CUDA OOM.", + ) + group.add_argument( + "--bucketing-sampler", + type=str2bool, + default=True, + help="When enabled, the batches will come from buckets of " + "similar duration (saves padding frames).", + ) + group.add_argument( + "--num-buckets", + type=int, + default=30, + help="The number of buckets for the DynamicBucketingSampler" + "(you might want to increase it for larger datasets).", + ) + group.add_argument( + "--shuffle", + type=str2bool, + default=True, + help="When enabled (=default), the examples will be " + "shuffled for each epoch.", + ) + group.add_argument( + "--drop-last", + type=str2bool, + default=True, + help="Whether to drop last batch. Used by sampler.", + ) + group.add_argument( + "--num-workers", + type=int, + default=2, + help="The number of training dataloader workers that " + "collect the batches.", + ) + group.add_argument( + "--do-normalize", + type=str2bool, + default=True, + help="whether to normalize the data", + ) + group.add_argument( + "--random-crop", + type=str2bool, + default=True, + help="audio sample rate", + ) + + def train_dataloaders( + self, + cuts_train: CutSet, + sample_rate: float = 16000, + label_rate: float = 50, + random_crop: bool = True, + pad_audio: bool = False, + num_classes: list = [504], + do_normalize: bool = True, + sampler_state_dict: Optional[Dict[str, Any]] = None, + ) -> DataLoader: + """ + Args: + cuts_train: + CutSet for training. + sampler_state_dict: + The state dict for the training sampler. + """ + logging.info("About to create train dataset") + train = HubertDataset( + sample_rate=sample_rate, + label_rate=label_rate, + random_crop=random_crop, + pad_audio=pad_audio, + num_classes=num_classes, + do_normalize=do_normalize, + ) + + if self.args.bucketing_sampler: + logging.info("Using DynamicBucketingSampler.") + train_sampler = DynamicBucketingSampler( + cuts_train, + max_duration=self.args.max_duration, + shuffle=self.args.shuffle, + num_buckets=self.args.num_buckets, + drop_last=self.args.drop_last, + ) + else: + logging.info("Using SimpleCutSampler.") + train_sampler = SimpleCutSampler( + cuts_train, + max_duration=self.args.max_duration, + shuffle=self.args.shuffle, + ) + logging.info("About to create train dataloader") + + if sampler_state_dict is not None: + logging.info("Loading sampler state dict") + train_sampler.load_state_dict(sampler_state_dict) + + # 'seed' is derived from the current random state, which will have + # previously been set in the main process. + seed = torch.randint(0, 100000, ()).item() + worker_init_fn = _SeedWorkers(seed) + + train_dl = DataLoader( + train, + sampler=train_sampler, + batch_size=None, + num_workers=self.args.num_workers, + persistent_workers=False, + worker_init_fn=worker_init_fn, + ) + + return train_dl + + def valid_dataloaders( + self, + cuts_valid: CutSet, + sample_rate: float = 16000, + label_rate: float = 50, + random_crop: bool = True, + pad_audio: bool = False, + num_classes: list = [504], + do_normalize: bool = True, + ) -> DataLoader: + logging.info("About to create dev dataset") + validate = HubertDataset( + sample_rate=sample_rate, + label_rate=label_rate, + random_crop=random_crop, + pad_audio=pad_audio, + num_classes=num_classes, + do_normalize=do_normalize, + ) + valid_sampler = DynamicBucketingSampler( + cuts_valid, + max_duration=self.args.max_duration, + shuffle=False, + ) + logging.info("About to create dev dataloader") + valid_dl = DataLoader( + validate, + sampler=valid_sampler, + batch_size=None, + num_workers=2, + persistent_workers=False, + ) + + return valid_dl + + def test_dataloaders( + self, + cuts: CutSet, + sample_rate: float = 16000, + label_rate: float = 50, + random_crop: bool = True, + pad_audio: bool = False, + num_classes: list = [504], + do_normalize: bool = True, + ) -> DataLoader: + logging.debug("About to create test dataset") + test = HubertDataset( + sample_rate=sample_rate, + label_rate=label_rate, + random_crop=random_crop, + pad_audio=pad_audio, + num_classes=num_classes, + do_normalize=do_normalize, + ) + sampler = DynamicBucketingSampler( + cuts, + max_duration=self.args.max_duration, + shuffle=False, + ) + logging.debug("About to create test dataloader") + test_dl = DataLoader( + test, + batch_size=None, + sampler=sampler, + num_workers=self.args.num_workers, + ) + return test_dl + + @lru_cache() + def small_cuts(self) -> CutSet: + logging.info("About to get small cuts") + return load_manifest_lazy( + self.args.manifest_dir / "librilight_cuts_small.jsonl.gz" + ) + + @lru_cache() + def medium_cuts(self) -> CutSet: + logging.info("About to get medium cuts") + filenames = glob.glob( + f"{self.args.manifest_dir}/medium_splits/librilight_cuts_medium.*.jsonl.gz" + ) + pattern = re.compile(r"librilight_cuts_medium.([0-9]+).jsonl.gz") + idx_filenames = ((int(pattern.search(f).group(1)), f) for f in filenames) + idx_filenames = sorted(idx_filenames, key=lambda x: x[0]) + sorted_filenames = [f[1] for f in idx_filenames] + logging.info( + f"Loading LibriLight medium {len(sorted_filenames)} splits in lazy mode" + ) + + return combine(load_manifest_lazy(p) for p in sorted_filenames) + + @lru_cache() + def large_cuts(self) -> CutSet: + logging.info("About to get large cuts") + filenames = glob.glob( + f"{self.args.manifest_dir}/large_splits/librilight_cuts_large.*.jsonl.gz" + ) + pattern = re.compile(r"librilight_cuts_large.([0-9]+).jsonl.gz") + idx_filenames = ((int(pattern.search(f).group(1)), f) for f in filenames) + idx_filenames = sorted(idx_filenames, key=lambda x: x[0]) + sorted_filenames = [f[1] for f in idx_filenames] + logging.info( + f"Loading LibriLight large {len(sorted_filenames)} splits in lazy mode" + ) + + return combine(load_manifest_lazy(p) for p in sorted_filenames) + + @lru_cache() + def train_all_shuf_cuts(self) -> CutSet: + logging.info("About to get the shuffled small, medium and large cuts") + small_cuts = self.small_cuts() + medium_cuts = self.medium_cuts() + large_cuts = self.large_cuts() + return CutSet.mux( + small_cuts, + medium_cuts, + large_cuts, + weights=[ + 122867, # len(small_cuts) + 1104071, # len(medium_cuts) + 11012085, # len(large_cuts) + ], + ) + + @lru_cache() + def dev_clean_cuts(self) -> CutSet: + logging.info("About to get dev-clean cuts") + return load_manifest_lazy( + self.args.manifest_dir / "librispeech_cuts_dev-clean.jsonl.gz" + ) + + @lru_cache() + def dev_other_cuts(self) -> CutSet: + logging.info("About to get dev-other cuts") + return load_manifest_lazy( + self.args.manifest_dir / "librispeech_cuts_dev-other.jsonl.gz" + ) diff --git a/egs/librilight/SSL/zipformer/utils.py b/egs/librilight/SSL/zipformer/utils.py new file mode 120000 index 0000000000..119992bdbf --- /dev/null +++ b/egs/librilight/SSL/zipformer/utils.py @@ -0,0 +1 @@ +../../../librispeech/SSL/zipformer/utils.py \ No newline at end of file diff --git a/egs/librilight/SSL/zipformer/wav2vec2_module.py b/egs/librilight/SSL/zipformer/wav2vec2_module.py new file mode 120000 index 0000000000..81ad701e47 --- /dev/null +++ b/egs/librilight/SSL/zipformer/wav2vec2_module.py @@ -0,0 +1 @@ +../../../librispeech/SSL/zipformer/wav2vec2_module.py \ No newline at end of file diff --git a/egs/librilight/SSL/zipformer/zipformer.py b/egs/librilight/SSL/zipformer/zipformer.py new file mode 120000 index 0000000000..5b3da8cd5b --- /dev/null +++ b/egs/librilight/SSL/zipformer/zipformer.py @@ -0,0 +1 @@ +../../../librispeech/SSL/zipformer/zipformer.py \ No newline at end of file diff --git a/egs/librispeech/SSL/hubert/asr_datamodule.py b/egs/librispeech/SSL/hubert/asr_datamodule.py new file mode 100644 index 0000000000..3746d8a3a3 --- /dev/null +++ b/egs/librispeech/SSL/hubert/asr_datamodule.py @@ -0,0 +1,287 @@ +# Copyright 2021 Piotr Żelasko +# Copyright 2024 Xiaomi Corporation (Author: Yifan Yang) +# +# See ../../../../LICENSE for clarification regarding multiple authors +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + + +import argparse +import logging +from functools import lru_cache +from pathlib import Path +from typing import Any, Dict, Optional + +import torch +from dataset import HubertAsrDataset +from lhotse import CutSet, load_manifest_lazy +from lhotse.dataset import DynamicBucketingSampler, SimpleCutSampler +from lhotse.utils import fix_random_seed +from torch.utils.data import DataLoader + +from icefall.utils import str2bool + + +class _SeedWorkers: + def __init__(self, seed: int): + self.seed = seed + + def __call__(self, worker_id: int): + fix_random_seed(self.seed + worker_id) + + +class LibriSpeechAsrDataModule: + """ + DataModule for ASR experiments. + It assumes there is always one train and valid dataloader, + but there can be multiple test dataloaders (e.g. LibriSpeech test-clean + and test-other). + + It contains all the common data pipeline modules used in ASR + experiments, e.g.: + - dynamic batch size, + - bucketing samplers, + + This class should be derived for specific corpora used in ASR tasks. + """ + + def __init__(self, args: argparse.Namespace): + self.args = args + + @classmethod + def add_arguments(cls, parser: argparse.ArgumentParser): + group = parser.add_argument_group( + title="ASR data related options", + description="These options are used for the preparation of " + "PyTorch DataLoaders from Lhotse CutSet's -- they control the " + "effective batch sizes, sampling strategies.", + ) + group.add_argument( + "--full-libri", + type=str2bool, + default=True, + help="When enabled use 960h LibriSpeech. " "Otherwise, use 100h subset.", + ) + + group.add_argument( + "--manifest-dir", + type=Path, + default=Path("data/wav"), + help="Path to directory with train/valid/test cuts.", + ) + group.add_argument( + "--max-duration", + type=float, + default=200.0, + help="Maximum pooled recordings duration (seconds) in a " + "single batch. You can reduce it if it causes CUDA OOM.", + ) + group.add_argument( + "--bucketing-sampler", + type=str2bool, + default=True, + help="When enabled, the batches will come from buckets of " + "similar duration (saves padding frames).", + ) + group.add_argument( + "--num-buckets", + type=int, + default=30, + help="The number of buckets for the DynamicBucketingSampler" + "(you might want to increase it for larger datasets).", + ) + group.add_argument( + "--shuffle", + type=str2bool, + default=True, + help="When enabled (=default), the examples will be " + "shuffled for each epoch.", + ) + group.add_argument( + "--drop-last", + type=str2bool, + default=True, + help="Whether to drop last batch. Used by sampler.", + ) + group.add_argument( + "--num-workers", + type=int, + default=2, + help="The number of training dataloader workers that " + "collect the batches.", + ) + group.add_argument( + "--do-normalize", + type=str2bool, + default=True, + help="whether to normalize the data", + ) + + def train_dataloaders( + self, + cuts_train: CutSet, + do_normalize: bool, + sampler_state_dict: Optional[Dict[str, Any]] = None, + ) -> DataLoader: + """ + Args: + cuts_train: + CutSet for training. + sampler_state_dict: + The state dict for the training sampler. + """ + logging.info("About to create train dataset") + train = HubertAsrDataset(do_normalize=do_normalize) + + if self.args.bucketing_sampler: + logging.info("Using DynamicBucketingSampler.") + train_sampler = DynamicBucketingSampler( + cuts_train, + max_duration=self.args.max_duration, + shuffle=self.args.shuffle, + num_buckets=self.args.num_buckets, + drop_last=self.args.drop_last, + ) + else: + logging.info("Using SimpleCutSampler.") + train_sampler = SimpleCutSampler( + cuts_train, + max_duration=self.args.max_duration, + shuffle=self.args.shuffle, + ) + logging.info("About to create train dataloader") + + if sampler_state_dict is not None: + logging.info("Loading sampler state dict") + train_sampler.load_state_dict(sampler_state_dict) + + # 'seed' is derived from the current random state, which will have + # previously been set in the main process. + seed = torch.randint(0, 100000, ()).item() + worker_init_fn = _SeedWorkers(seed) + + train_dl = DataLoader( + train, + sampler=train_sampler, + batch_size=None, + num_workers=self.args.num_workers, + persistent_workers=False, + worker_init_fn=worker_init_fn, + ) + + return train_dl + + def valid_dataloaders(self, cuts_valid: CutSet, do_normalize: bool) -> DataLoader: + logging.info("About to create dev dataset") + validate = HubertAsrDataset(do_normalize=do_normalize) + valid_sampler = DynamicBucketingSampler( + cuts_valid, + max_duration=self.args.max_duration, + shuffle=False, + ) + logging.info("About to create dev dataloader") + valid_dl = DataLoader( + validate, + sampler=valid_sampler, + batch_size=None, + num_workers=2, + persistent_workers=False, + ) + + return valid_dl + + def test_dataloaders(self, cuts: CutSet, do_normalize: bool) -> DataLoader: + logging.debug("About to create test dataset") + test = HubertAsrDataset(do_normalize=do_normalize) + sampler = DynamicBucketingSampler( + cuts, + max_duration=self.args.max_duration, + shuffle=False, + ) + logging.debug("About to create test dataloader") + test_dl = DataLoader( + test, + batch_size=None, + sampler=sampler, + num_workers=self.args.num_workers, + ) + return test_dl + + @lru_cache() + def train_clean_100_cuts(self) -> CutSet: + logging.info("About to get train-clean-100 cuts") + return load_manifest_lazy( + self.args.manifest_dir / "librispeech_cuts_train-clean-100.jsonl.gz" + ) + + @lru_cache() + def train_clean_360_cuts(self) -> CutSet: + logging.info("About to get train-clean-360 cuts") + return load_manifest_lazy( + self.args.manifest_dir / "librispeech_cuts_train-clean-360.jsonl.gz" + ) + + @lru_cache() + def train_other_500_cuts(self) -> CutSet: + logging.info("About to get train-other-500 cuts") + return load_manifest_lazy( + self.args.manifest_dir / "librispeech_cuts_train-other-500.jsonl.gz" + ) + + @lru_cache() + def train_all_shuf_cuts(self) -> CutSet: + logging.info( + "About to get the shuffled train-clean-100, \ + train-clean-360 and train-other-500 cuts" + ) + train_clean_100_cuts = self.train_clean_100_cuts() + train_clean_360_cuts = self.train_clean_360_cuts() + train_other_500_cuts = self.train_other_500_cuts() + return CutSet.mux( + train_clean_100_cuts, + train_clean_360_cuts, + train_other_500_cuts, + weights=[ + 28539, # len(train_clean_100_cuts) + 104014, # len(train_clean_360_cuts) + 148688, # len(train_other_500_cuts) + ], + ) + + @lru_cache() + def dev_clean_cuts(self) -> CutSet: + logging.info("About to get dev-clean cuts") + return load_manifest_lazy( + self.args.manifest_dir / "librispeech_cuts_dev-clean.jsonl.gz" + ) + + @lru_cache() + def dev_other_cuts(self) -> CutSet: + logging.info("About to get dev-other cuts") + return load_manifest_lazy( + self.args.manifest_dir / "librispeech_cuts_dev-other.jsonl.gz" + ) + + @lru_cache() + def test_clean_cuts(self) -> CutSet: + logging.info("About to get test-clean cuts") + return load_manifest_lazy( + self.args.manifest_dir / "librispeech_cuts_test-clean.jsonl.gz" + ) + + @lru_cache() + def test_other_cuts(self) -> CutSet: + logging.info("About to get test-other cuts") + return load_manifest_lazy( + self.args.manifest_dir / "librispeech_cuts_test-other.jsonl.gz" + ) diff --git a/egs/librispeech/SSL/hubert/attention_module.py b/egs/librispeech/SSL/hubert/attention_module.py new file mode 100644 index 0000000000..39ef8698ee --- /dev/null +++ b/egs/librispeech/SSL/hubert/attention_module.py @@ -0,0 +1,840 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# Permission is hereby granted, free of charge, to any person obtaining a copy +# of this software and associated documentation files (the "Software"), to deal +# in the Software without restriction, including without limitation the rights +# to use, copy, modify, merge, publish, distribute, sublicense, and/or sell +# copies of the Software, and to permit persons to whom the Software is +# furnished to do so, subject to the following conditions: +# +# The above copyright notice and this permission notice shall be included in all +# copies or substantial portions of the Software. +# +# THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR +# IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, +# FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE +# AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER +# LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, +# OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE +# SOFTWARE. + +import math +from typing import Dict, List, Optional, Tuple + +import torch +import torch.nn.functional as F +import utils +from torch import Tensor, nn +from torch.nn import Parameter +from utils import FairseqDropout, quant_noise + +_xformers_available = False + + +# TODO: move this into xformers? +# TODO: uint8 input type should just output a bool +def _mask_for_xformers(mask: Tensor, to_dtype: Optional[torch.dtype] = None): + """ + call to pytorch multihead accepts three mask types: + - ByteTensor where non-zero means to mask + - FloatTensor which is an additive mask + - BoolTensor where True means to mask + xFormers currently accepts boolean and additive maks. For boolean masks + the values have opposite meaning. For a BoolTensor True mean to keep the value. + """ + float_types = [torch.float, torch.float16] + # If an input mask is a float it is an additive mask. Otherwise it is either uint8 or bool. + additive = mask.dtype in float_types + # If to_dype is not specified, keep same dtype as mask. + to_dtype = mask.dtype if to_dtype is None else to_dtype + to_additive = to_dtype in float_types + + if additive: + if to_additive: + return mask.to(to_dtype) + mask = mask < 0 + + if to_additive: + # return additive mask + new_mask = torch.zeros_like(mask, dtype=to_dtype) + new_mask = new_mask.masked_fill_(mask, -float("inf")) + return new_mask + + # In xFormers True is value to keep rather than value to mask + mask = ~mask.to(torch.bool) + mask = mask.to(to_dtype) + return mask + + +def init_bert_params(module): + """ + Initialize the weights specific to the BERT Model. + This overrides the default initializations depending on the specified arguments. + 1. If normal_init_linear_weights is set then weights of linear + layer will be initialized using the normal distribution and + bais will be set to the specified value. + 2. If normal_init_embed_weights is set then weights of embedding + layer will be initialized using the normal distribution. + 3. If normal_init_proj_weights is set then weights of + in_project_weight for MultiHeadAttention initialized using + the normal distribution (to be validated). + """ + + def normal_(data): + # with FSDP, module params will be on CUDA, so we cast them back to CPU + # so that the RNG is consistent with and without FSDP + data.copy_(data.cpu().normal_(mean=0.0, std=0.02).to(data.device)) + + if isinstance(module, nn.Linear): + normal_(module.weight.data) + if module.bias is not None: + module.bias.data.zero_() + if isinstance(module, nn.Embedding): + normal_(module.weight.data) + if module.padding_idx is not None: + module.weight.data[module.padding_idx].zero_() + if isinstance(module, MultiheadAttention): + normal_(module.q_proj.weight.data) + normal_(module.k_proj.weight.data) + normal_(module.v_proj.weight.data) + + +class MultiheadAttention(nn.Module): + """Multi-headed attention. + + See "Attention Is All You Need" for more details. + """ + + def __init__( + self, + embed_dim, + num_heads, + kdim=None, + vdim=None, + dropout=0.0, + bias=True, + add_bias_kv=False, + add_zero_attn=False, + self_attention=False, + encoder_decoder_attention=False, + dictionary=None, + q_noise=0.0, + qn_block_size=8, + # TODO: pass in config rather than string. + # config defined in xformers.components.attention.AttentionConfig + xformers_att_config: Optional[str] = None, + xformers_blocksparse_layout: Optional[ + torch.Tensor + ] = None, # This should be part of the config + xformers_blocksparse_blocksize: Optional[ + int + ] = 16, # This should be part of the config + ): + super().__init__() + + self.use_xformers = False + if self.use_xformers and not _xformers_available: + raise ImportError("\n\n Please install xFormers.") + self.embed_dim = embed_dim + self.kdim = kdim if kdim is not None else embed_dim + self.vdim = vdim if vdim is not None else embed_dim + self.qkv_same_dim = self.kdim == embed_dim and self.vdim == embed_dim + + self.num_heads = num_heads + self.dropout_module = FairseqDropout( + dropout, module_name=self.__class__.__name__ + ) + + self.head_dim = embed_dim // num_heads + assert ( + self.head_dim * num_heads == self.embed_dim + ), "embed_dim must be divisible by num_heads" + self.scaling = self.head_dim**-0.5 + + self.self_attention = self_attention + self.encoder_decoder_attention = encoder_decoder_attention + + assert ( + not self.self_attention or self.qkv_same_dim + ), "Self-attention requires query, key and value to be of the same size" + + self.k_proj = quant_noise( + nn.Linear(self.kdim, embed_dim, bias=bias), q_noise, qn_block_size + ) + self.v_proj = quant_noise( + nn.Linear(self.vdim, embed_dim, bias=bias), q_noise, qn_block_size + ) + self.q_proj = quant_noise( + nn.Linear(embed_dim, embed_dim, bias=bias), q_noise, qn_block_size + ) + + self.out_proj = quant_noise( + nn.Linear(embed_dim, embed_dim, bias=bias), q_noise, qn_block_size + ) + + if add_bias_kv: + self.bias_k = Parameter(torch.Tensor(1, 1, embed_dim)) + self.bias_v = Parameter(torch.Tensor(1, 1, embed_dim)) + else: + self.bias_k = self.bias_v = None + + self.add_zero_attn = add_zero_attn + self.beam_size = 1 + self.reset_parameters() + + self.onnx_trace = False + self.skip_embed_dim_check = False + + def prepare_for_onnx_export_(self): + self.onnx_trace = True + + def reset_parameters(self): + if self.qkv_same_dim: + # Empirically observed the convergence to be much better with + # the scaled initialization + nn.init.xavier_uniform_(self.k_proj.weight, gain=1 / math.sqrt(2)) + nn.init.xavier_uniform_(self.v_proj.weight, gain=1 / math.sqrt(2)) + nn.init.xavier_uniform_(self.q_proj.weight, gain=1 / math.sqrt(2)) + else: + nn.init.xavier_uniform_(self.k_proj.weight) + nn.init.xavier_uniform_(self.v_proj.weight) + nn.init.xavier_uniform_(self.q_proj.weight) + + nn.init.xavier_uniform_(self.out_proj.weight) + if self.out_proj.bias is not None: + nn.init.constant_(self.out_proj.bias, 0.0) + if self.bias_k is not None: + nn.init.xavier_normal_(self.bias_k) + if self.bias_v is not None: + nn.init.xavier_normal_(self.bias_v) + + def _get_reserve_head_index(self, num_heads_to_keep: int): + k_proj_heads_norm = [] + q_proj_heads_norm = [] + v_proj_heads_norm = [] + + for i in range(self.num_heads): + start_idx = i * self.head_dim + end_idx = (i + 1) * self.head_dim + k_proj_heads_norm.append( + torch.sum( + torch.abs( + self.k_proj.weight[ + start_idx:end_idx, + ] + ) + ).tolist() + + torch.sum(torch.abs(self.k_proj.bias[start_idx:end_idx])).tolist() + ) + q_proj_heads_norm.append( + torch.sum( + torch.abs( + self.q_proj.weight[ + start_idx:end_idx, + ] + ) + ).tolist() + + torch.sum(torch.abs(self.q_proj.bias[start_idx:end_idx])).tolist() + ) + v_proj_heads_norm.append( + torch.sum( + torch.abs( + self.v_proj.weight[ + start_idx:end_idx, + ] + ) + ).tolist() + + torch.sum(torch.abs(self.v_proj.bias[start_idx:end_idx])).tolist() + ) + + heads_norm = [] + for i in range(self.num_heads): + heads_norm.append( + k_proj_heads_norm[i] + q_proj_heads_norm[i] + v_proj_heads_norm[i] + ) + + sorted_head_index = sorted( + range(self.num_heads), key=lambda k: heads_norm[k], reverse=True + ) + reserve_head_index = [] + for i in range(num_heads_to_keep): + start = sorted_head_index[i] * self.head_dim + end = (sorted_head_index[i] + 1) * self.head_dim + reserve_head_index.append((start, end)) + return reserve_head_index + + def _adaptive_prune_heads(self, reserve_head_index: List[Tuple[int, int]]): + new_q_weight = [] + new_q_bias = [] + new_k_weight = [] + new_k_bias = [] + new_v_weight = [] + new_v_bias = [] + new_out_proj_weight = [] + + for ele in reserve_head_index: + start_idx, end_idx = ele + new_q_weight.append( + self.q_proj.weight[ + start_idx:end_idx, + ] + ) + new_q_bias.append(self.q_proj.bias[start_idx:end_idx]) + + new_k_weight.append( + self.k_proj.weight[ + start_idx:end_idx, + ] + ) + + new_k_bias.append(self.k_proj.bias[start_idx:end_idx]) + + new_v_weight.append( + self.v_proj.weight[ + start_idx:end_idx, + ] + ) + new_v_bias.append(self.v_proj.bias[start_idx:end_idx]) + + new_out_proj_weight.append(self.out_proj.weight[:, start_idx:end_idx]) + + new_q_weight = torch.cat(new_q_weight).detach() + new_k_weight = torch.cat(new_k_weight).detach() + new_v_weight = torch.cat(new_v_weight).detach() + new_out_proj_weight = torch.cat(new_out_proj_weight, dim=-1).detach() + new_q_weight.requires_grad = True + new_k_weight.requires_grad = True + new_v_weight.requires_grad = True + new_out_proj_weight.requires_grad = True + + new_q_bias = torch.cat(new_q_bias).detach() + new_q_bias.requires_grad = True + + new_k_bias = torch.cat(new_k_bias).detach() + new_k_bias.requires_grad = True + + new_v_bias = torch.cat(new_v_bias).detach() + new_v_bias.requires_grad = True + + self.q_proj.weight = torch.nn.Parameter(new_q_weight) + self.q_proj.bias = torch.nn.Parameter(new_q_bias) + + self.k_proj.weight = torch.nn.Parameter(new_k_weight) + self.k_proj.bias = torch.nn.Parameter(new_k_bias) + + self.v_proj.weight = torch.nn.Parameter(new_v_weight) + self.v_proj.bias = torch.nn.Parameter(new_v_bias) + + self.out_proj.weight = torch.nn.Parameter(new_out_proj_weight) + + self.num_heads = len(reserve_head_index) + self.embed_dim = self.head_dim * self.num_heads + self.q_proj.out_features = self.embed_dim + self.k_proj.out_features = self.embed_dim + self.v_proj.out_features = self.embed_dim + + def _set_skip_embed_dim_check(self): + self.skip_embed_dim_check = True + + def _pad_masks( + self, + key_padding_mask: Optional[Tensor], + attn_mask: Optional[Tensor], + ) -> Tuple[Optional[Tensor], Optional[Tensor]]: + if attn_mask is not None: + shape = attn_mask.size()[:-1] + torch.Size([1]) + attn_mask = torch.cat([attn_mask, attn_mask.new_zeros(shape)], dim=-1) + if key_padding_mask is not None: + shape = key_padding_mask.size()[:-1] + torch.Size([1]) + key_padding_mask = torch.cat( + [ + key_padding_mask, + key_padding_mask.new_zeros(shape), + ], + dim=-1, + ) + return key_padding_mask, attn_mask + + def _add_bias( + self, + k: Tensor, + v: Tensor, + key_padding_mask: Optional[Tensor], + attn_mask: Optional[Tensor], + bsz: int, + ) -> Tuple[Tensor, Tensor, Optional[Tensor], Optional[Tensor]]: + assert self.bias_k is not None + assert self.bias_v is not None + k = torch.cat([k, self.bias_k.repeat(1, bsz, 1)]) + v = torch.cat([v, self.bias_v.repeat(1, bsz, 1)]) + key_padding_mask, attn_mask = self._pad_masks( + key_padding_mask=key_padding_mask, attn_mask=attn_mask + ) + return k, v, key_padding_mask, attn_mask + + def _append_zero_attn( + self, + k: Tensor, + v: Tensor, + key_padding_mask: Optional[Tensor], + attn_mask: Optional[Tensor], + ) -> Tuple[Tensor, Tensor, Optional[Tensor], Optional[Tensor]]: + zero_attn_shape = k.size()[:-2] + torch.Size([1]) + k.size()[-1:] + k = torch.cat( + [k, torch.zeros(zero_attn_shape, dtype=k.dtype, device=k.device)], + dim=-2, + ) + v = torch.cat( + [v, torch.zeros(zero_attn_shape, dtype=v.dtype, device=v.device)], + dim=-2, + ) + key_padding_mask, attn_mask = self._pad_masks( + key_padding_mask=key_padding_mask, attn_mask=attn_mask + ) + return k, v, key_padding_mask, attn_mask + + def forward( + self, + query: Tensor, + key: Optional[Tensor], + value: Optional[Tensor], + key_padding_mask: Optional[Tensor] = None, + incremental_state: Optional[Dict[str, Dict[str, Optional[Tensor]]]] = None, + need_weights: bool = True, + static_kv: bool = False, + attn_mask: Optional[Tensor] = None, + before_softmax: bool = False, + need_head_weights: bool = False, + ) -> Tuple[Tensor, Optional[Tensor]]: + """Input shape: Time x Batch x Channel + + Args: + key_padding_mask (ByteTensor, optional): mask to exclude + keys that are pads, of shape `(batch, src_len)`, where + padding elements are indicated by 1s. + need_weights (bool, optional): return the attention weights, + averaged over heads (default: False). + attn_mask (ByteTensor, optional): typically used to + implement causal attention, where the mask prevents the + attention from looking forward in time (default: None). + before_softmax (bool, optional): return the raw attention + weights and values before the attention softmax. + need_head_weights (bool, optional): return the attention + weights for each head. Implies *need_weights*. Default: + return the average attention weights over all heads. + """ + if need_head_weights: + need_weights = True + + is_tpu = query.device.type == "xla" + + tgt_len, bsz, embed_dim = query.size() + src_len = tgt_len + if not self.skip_embed_dim_check: + assert ( + embed_dim == self.embed_dim + ), f"query dim {embed_dim} != {self.embed_dim}" + assert list(query.size()) == [tgt_len, bsz, embed_dim] + if key is not None: + src_len, key_bsz, _ = key.size() + if not torch.jit.is_scripting(): + assert value is not None + assert src_len, key_bsz == value.shape[:2] + + if ( + not self.onnx_trace + and not is_tpu # don't use PyTorch version on TPUs + and incremental_state is None + and not static_kv + # A workaround for quantization to work. Otherwise JIT compilation + # treats bias in linear module as method. + and not torch.jit.is_scripting() + # The Multihead attention implemented in pytorch forces strong dimension check + # for input embedding dimention and K,Q,V projection dimension. + # Since pruning will break the dimension check and it is not easy to modify the pytorch API, + # it is preferred to bypass the pytorch MHA when we need to skip embed_dim_check + and not self.skip_embed_dim_check + ): + assert key is not None and value is not None + + return F.multi_head_attention_forward( + query, + key, + value, + self.embed_dim, + self.num_heads, + torch.empty([0]), + torch.cat((self.q_proj.bias, self.k_proj.bias, self.v_proj.bias)), + self.bias_k, + self.bias_v, + self.add_zero_attn, + self.dropout_module.p, + self.out_proj.weight, + self.out_proj.bias, + self.training or self.dropout_module.apply_during_inference, + key_padding_mask.bool() if key_padding_mask is not None else None, + need_weights, + attn_mask, + use_separate_proj_weight=True, + q_proj_weight=self.q_proj.weight, + k_proj_weight=self.k_proj.weight, + v_proj_weight=self.v_proj.weight, + ) + + if incremental_state is not None: + saved_state = self._get_input_buffer(incremental_state) + if saved_state is not None and "prev_key" in saved_state: + # previous time steps are cached - no need to recompute + # key and value if they are static + if static_kv: + assert self.encoder_decoder_attention and not self.self_attention + key = value = None + else: + saved_state = None + + if self.self_attention: + q = self.q_proj(query) + k = self.k_proj(query) + v = self.v_proj(query) + elif self.encoder_decoder_attention: + # encoder-decoder attention + q = self.q_proj(query) + if key is None: + assert value is None + k = v = None + else: + if self.beam_size > 1 and bsz == key.size(1): + # key is [T, bsz*beam_size, C], reduce to [T, bsz, C] + key = key.view(key.size(0), -1, self.beam_size, key.size(2))[ + :, :, 0, : + ] + if key_padding_mask is not None: + key_padding_mask = key_padding_mask.view( + -1, self.beam_size, key_padding_mask.size(1) + )[:, 0, :] + k = self.k_proj(key) + v = self.v_proj(key) + + else: + assert key is not None and value is not None + q = self.q_proj(query) + k = self.k_proj(key) + v = self.v_proj(value) + q *= self.scaling + + if self.bias_k is not None: + assert self.bias_v is not None + k, v, attn_mask, key_padding_mask = self._add_bias( + k, v, attn_mask, key_padding_mask, bsz + ) + + q = ( + q.contiguous() + .view(tgt_len, bsz * self.num_heads, self.head_dim) + .transpose(0, 1) + ) + kv_bsz = bsz # need default value for scripting + if k is not None: + kv_bsz = k.size(1) + k = ( + k.contiguous() + .view(-1, kv_bsz * self.num_heads, self.head_dim) + .transpose(0, 1) + ) + if v is not None: + v = ( + v.contiguous() + .view(-1, kv_bsz * self.num_heads, self.head_dim) + .transpose(0, 1) + ) + + if saved_state is not None: + # saved states are stored with shape (bsz, num_heads, seq_len, head_dim) + if "prev_key" in saved_state: + _prev_key = saved_state["prev_key"] + assert _prev_key is not None + kv_bsz = _prev_key.size(0) + prev_key = _prev_key.view(kv_bsz * self.num_heads, -1, self.head_dim) + if static_kv: + k = prev_key + else: + assert k is not None + k = torch.cat([prev_key, k], dim=1) + src_len = k.size(1) + if "prev_value" in saved_state: + _prev_value = saved_state["prev_value"] + assert _prev_value is not None + assert kv_bsz == _prev_value.size(0) + prev_value = _prev_value.view( + kv_bsz * self.num_heads, -1, self.head_dim + ) + if static_kv: + v = prev_value + else: + assert v is not None + v = torch.cat([prev_value, v], dim=1) + prev_key_padding_mask: Optional[Tensor] = None + if "prev_key_padding_mask" in saved_state: + prev_key_padding_mask = saved_state["prev_key_padding_mask"] + assert k is not None and v is not None + key_padding_mask = MultiheadAttention._append_prev_key_padding_mask( + key_padding_mask=key_padding_mask, + prev_key_padding_mask=prev_key_padding_mask, + batch_size=kv_bsz, + src_len=k.size(1), + static_kv=static_kv, + ) + + saved_state["prev_key"] = k.view(kv_bsz, self.num_heads, -1, self.head_dim) + saved_state["prev_value"] = v.view( + kv_bsz, self.num_heads, -1, self.head_dim + ) + saved_state["prev_key_padding_mask"] = key_padding_mask + # In this branch incremental_state is never None + assert incremental_state is not None + incremental_state = self._set_input_buffer(incremental_state, saved_state) + assert k is not None + assert k.size(1) == src_len + + # This is part of a workaround to get around fork/join parallelism + # not supporting Optional types. + if key_padding_mask is not None and key_padding_mask.dim() == 0: + key_padding_mask = None + + if key_padding_mask is not None: + assert key_padding_mask.size(0) == kv_bsz + assert key_padding_mask.size(1) == src_len + + if self.add_zero_attn: + assert v is not None + src_len += 1 + k, v, key_padding_mask, attn_mask = self._append_zero_attn( + k=k, v=v, key_padding_mask=key_padding_mask, attn_mask=attn_mask + ) + + if self.encoder_decoder_attention and bsz != kv_bsz: + attn_weights = torch.einsum( + "bxhtd,bhsd->bxhts", + q.view((kv_bsz, -1, self.num_heads) + q.size()[1:]), + k.view((kv_bsz, self.num_heads) + k.size()[1:]), + ) + attn_weights = attn_weights.reshape((-1,) + attn_weights.size()[-2:]) + else: + attn_weights = torch.bmm(q, k.transpose(1, 2)) + attn_weights = self.apply_sparse_mask(attn_weights, tgt_len, src_len, bsz) + + assert list(attn_weights.size()) == [ + bsz * self.num_heads, + tgt_len, + src_len, + ] + + if attn_mask is not None: + attn_mask = attn_mask.unsqueeze(0) + if self.onnx_trace: + attn_mask = attn_mask.repeat(attn_weights.size(0), 1, 1) + attn_weights += attn_mask + + if key_padding_mask is not None: + # don't attend to padding symbols + attn_weights = attn_weights.view(bsz, self.num_heads, tgt_len, src_len) + if not is_tpu: + attn_weights = attn_weights.view( + kv_bsz, -1, self.num_heads, tgt_len, src_len + ) + attn_weights = attn_weights.masked_fill( + key_padding_mask.unsqueeze(1) + .unsqueeze(2) + .unsqueeze(3) + .to(torch.bool), + float("-inf"), + ) + else: + attn_weights = attn_weights.transpose(0, 2) + attn_weights = attn_weights.masked_fill(key_padding_mask, float("-inf")) + attn_weights = attn_weights.transpose(0, 2) + attn_weights = attn_weights.view(bsz * self.num_heads, tgt_len, src_len) + + if before_softmax: + return attn_weights, v + + attn_weights_float = utils.softmax( + attn_weights, dim=-1, onnx_trace=self.onnx_trace + ) + attn_weights = attn_weights_float.type_as(attn_weights) + attn_probs = self.dropout_module(attn_weights) + + assert v is not None + attn: Optional[Tensor] = None + if self.encoder_decoder_attention and bsz != kv_bsz: + attn = torch.einsum( + "bxhts,bhsd->bxhtd", + attn_probs.view( + ( + kv_bsz, + -1, + self.num_heads, + ) + + attn_probs.size()[1:] + ), + v.view( + ( + kv_bsz, + self.num_heads, + ) + + v.size()[1:] + ), + ) + attn = attn.reshape((-1,) + attn.size()[-2:]) + else: + attn = torch.bmm(attn_probs, v) + assert list(attn.size()) == [ + bsz * self.num_heads, + tgt_len, + self.head_dim, + ] + if self.onnx_trace and attn.size(1) == 1: + # when ONNX tracing a single decoder step (sequence length == 1) + # the transpose is a no-op copy before view, thus unnecessary + attn = attn.contiguous().view(tgt_len, bsz, self.embed_dim) + else: + attn = attn.transpose(0, 1).contiguous().view(tgt_len, bsz, self.embed_dim) + attn = self.out_proj(attn) + attn_weights: Optional[Tensor] = None + if need_weights: + attn_weights = attn_weights_float.view( + bsz, self.num_heads, tgt_len, src_len + ).transpose(1, 0) + if not need_head_weights: + # average attention weights over heads + attn_weights = attn_weights.mean(dim=0) + + return attn, attn_weights + + @staticmethod + def _append_prev_key_padding_mask( + key_padding_mask: Optional[Tensor], + prev_key_padding_mask: Optional[Tensor], + batch_size: int, + src_len: int, + static_kv: bool, + ) -> Optional[Tensor]: + # saved key padding masks have shape (bsz, seq_len) + if prev_key_padding_mask is not None and static_kv: + new_key_padding_mask = prev_key_padding_mask + elif prev_key_padding_mask is not None and key_padding_mask is not None: + new_key_padding_mask = torch.cat( + [prev_key_padding_mask.float(), key_padding_mask.float()], dim=1 + ) + # During incremental decoding, as the padding token enters and + # leaves the frame, there will be a time when prev or current + # is None + elif prev_key_padding_mask is not None: + if src_len > prev_key_padding_mask.size(1): + filler = torch.zeros( + (batch_size, src_len - prev_key_padding_mask.size(1)), + device=prev_key_padding_mask.device, + ) + new_key_padding_mask = torch.cat( + [prev_key_padding_mask.float(), filler.float()], dim=1 + ) + else: + new_key_padding_mask = prev_key_padding_mask.float() + elif key_padding_mask is not None: + if src_len > key_padding_mask.size(1): + filler = torch.zeros( + (batch_size, src_len - key_padding_mask.size(1)), + device=key_padding_mask.device, + ) + new_key_padding_mask = torch.cat( + [filler.float(), key_padding_mask.float()], dim=1 + ) + else: + new_key_padding_mask = key_padding_mask.float() + else: + new_key_padding_mask = prev_key_padding_mask + return new_key_padding_mask + + @torch.jit.export + def reorder_incremental_state( + self, + incremental_state: Optional[Dict[str, Dict[str, Optional[Tensor]]]], + new_order: Tensor, + ): + """Reorder buffered internal state (for incremental generation).""" + input_buffer = self._get_input_buffer(incremental_state) + if input_buffer is not None: + for k in input_buffer.keys(): + input_buffer_k = input_buffer[k] + if input_buffer_k is not None: + if self.encoder_decoder_attention: + if input_buffer_k.size(0) * self.beam_size == new_order.size(0): + return incremental_state + elif self.beam_size > 1: + input_buffer[k] = input_buffer_k.index_select( + 0, + new_order.reshape(-1, self.beam_size)[:, 0] + // self.beam_size, + ) + else: + input_buffer[k] = input_buffer_k.index_select(0, new_order) + else: + input_buffer[k] = input_buffer_k.index_select(0, new_order) + incremental_state = self._set_input_buffer(incremental_state, input_buffer) + return incremental_state + + def set_beam_size(self, beam_size): + """Used for effiecient beamable enc-dec attention""" + self.beam_size = beam_size + + def _get_input_buffer( + self, + incremental_state: Optional[Dict[str, Dict[str, Optional[Tensor]]]], + ) -> Dict[str, Optional[Tensor]]: + result = self.get_incremental_state(incremental_state, "attn_state") + if result is not None: + return result + else: + empty_result: Dict[str, Optional[Tensor]] = {} + return empty_result + + def _set_input_buffer( + self, + incremental_state: Optional[Dict[str, Dict[str, Optional[Tensor]]]], + buffer: Dict[str, Optional[Tensor]], + ): + return self.set_incremental_state(incremental_state, "attn_state", buffer) + + def apply_sparse_mask(self, attn_weights, tgt_len: int, src_len: int, bsz: int): + return attn_weights + + def upgrade_state_dict_named(self, state_dict, name): + prefix = name + "." if name != "" else "" + items_to_add = {} + keys_to_remove = [] + for k in state_dict.keys(): + if k.endswith(prefix + "in_proj_weight"): + # in_proj_weight used to be q + k + v with same dimensions + dim = int(state_dict[k].shape[0] / 3) + items_to_add[prefix + "q_proj.weight"] = state_dict[k][:dim] + items_to_add[prefix + "k_proj.weight"] = state_dict[k][dim : 2 * dim] + items_to_add[prefix + "v_proj.weight"] = state_dict[k][2 * dim :] + + keys_to_remove.append(k) + + k_bias = prefix + "in_proj_bias" + if k_bias in state_dict.keys(): + dim = int(state_dict[k].shape[0] / 3) + items_to_add[prefix + "q_proj.bias"] = state_dict[k_bias][:dim] + items_to_add[prefix + "k_proj.bias"] = state_dict[k_bias][ + dim : 2 * dim + ] + items_to_add[prefix + "v_proj.bias"] = state_dict[k_bias][2 * dim :] + + keys_to_remove.append(prefix + "in_proj_bias") + + for k in keys_to_remove: + del state_dict[k] + + for key, value in items_to_add.items(): + state_dict[key] = value diff --git a/egs/librispeech/SSL/hubert/beam_search.py b/egs/librispeech/SSL/hubert/beam_search.py new file mode 120000 index 0000000000..f4d4b57326 --- /dev/null +++ b/egs/librispeech/SSL/hubert/beam_search.py @@ -0,0 +1 @@ +../../ASR/zipformer/beam_search.py \ No newline at end of file diff --git a/egs/librispeech/SSL/hubert/dataset.py b/egs/librispeech/SSL/hubert/dataset.py new file mode 100644 index 0000000000..76edfb340a --- /dev/null +++ b/egs/librispeech/SSL/hubert/dataset.py @@ -0,0 +1,367 @@ +# Copyright 2024 Xiaomi Corporation (authors: Yifan Yang) +# Copyright 2024 Shanghai Jiao Tong University (authors: Jianheng Zhuo) +# +# 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 sys +from typing import Any, Dict, Optional + +import numpy as np +import torch +import torch.nn.functional as F +from lhotse import validate +from lhotse.cut import CutSet +from lhotse.dataset.collation import read_audio_from_cuts +from torch.utils.data.dataloader import default_collate + + +class HubertDataset(torch.utils.data.Dataset): + """ + In this implementation, there will always be a single channel. + + Returns: + + .. code-block:: + + { + 'audio': (B x NumSamples) float tensor + } + """ + + def __init__( + self, + max_sample_size: Optional[int] = None, + sample_rate: float = 16000, + label_rate: float = 50, + random_crop: bool = True, + pad_audio: bool = False, + num_classes: list = [504], + do_normalize: bool = True, + ) -> None: + super().__init__() + self.sample_rate = sample_rate + self.label_rate = label_rate + self.random_crop = random_crop + self.pad_audio = pad_audio + self.num_classes = num_classes + self.normalize = do_normalize + self.max_sample_size = ( + max_sample_size if max_sample_size is not None else sys.maxsize + ) + + def __getitem__(self, cuts: CutSet) -> Dict[str, Any]: + self._validate(cuts) + audio, _ = read_audio_from_cuts(cuts) + for i, item in enumerate(audio): + audio[i] = self.postprocess(item, self.sample_rate) + audio_lens = [cut.num_samples for cut in cuts] + + if self.pad_audio: + audio_size = min(max(audio_lens), self.max_sample_size) + else: + audio_size = min(min(audio_lens), self.max_sample_size) + + audio, padding_mask, audio_starts = self.collater_audio( + audio, audio_lens, audio_size + ) + + kmeans = [cut.custom["kmeans"] for cut in cuts] + kmeans = [ + torch.tensor([int(item) for item in label.split()], dtype=torch.int64) + for label in kmeans + ] + kmeans, _ = self.collater_frm_label(kmeans, audio_size, audio_starts) + + return { + "cuts": cuts, + "audio": audio, + "padding_mask": padding_mask, + "kmeans": kmeans, + } + + def postprocess(self, wav, cur_sample_rate): + if wav.dim() == 2: + wav = wav.mean(-1) + assert wav.dim() == 1, wav.dim() + + if cur_sample_rate != self.sample_rate: + raise Exception(f"sr {cur_sample_rate} != {self.sample_rate}") + + if self.normalize: + with torch.no_grad(): + wav = F.layer_norm(wav, wav.shape) + return wav + + def _validate(self, cuts: CutSet) -> None: + validate(cuts) + assert all(cut.has_recording for cut in cuts) + + def crop_to_max_size(self, wav, target_size): + size = len(wav) + diff = size - target_size + if diff <= 0: + return wav, 0 + + start, end = 0, target_size + if self.random_crop: + start = np.random.randint(0, diff + 1) + end = size - diff + start + return wav[start:end], start + + def collater_audio(self, audios, audio_lens, audio_size): + collated_audios = audios[0].new_zeros(len(audios), audio_size) + padding_mask = ( + torch.BoolTensor(collated_audios.shape).fill_(False) + # if self.pad_audio else None + ) + audio_starts = [0 for _ in audios] + for i, (audio, audio_len) in enumerate(zip(audios, audio_lens)): + audio = audio[:audio_len] + diff = audio_len - audio_size + if diff == 0: + collated_audios[i] = audio + elif diff < 0: + assert self.pad_audio + collated_audios[i] = torch.cat([audio, audio.new_full((-diff,), 0.0)]) + padding_mask[i, diff:] = True + else: + collated_audios[i], audio_starts[i] = self.crop_to_max_size( + audio, audio_size + ) + return collated_audios, padding_mask, audio_starts + + def collate_tokens( + self, + values, + pad_idx, + eos_idx=None, + left_pad=False, + move_eos_to_beginning=False, + pad_to_length=None, + pad_to_multiple=1, + pad_to_bsz=None, + ): + """Convert a list of 1d tensors into a padded 2d tensor.""" + size = max(v.size(0) for v in values) + size = size if pad_to_length is None else max(size, pad_to_length) + if pad_to_multiple != 1 and size % pad_to_multiple != 0: + size = int(((size - 0.1) // pad_to_multiple + 1) * pad_to_multiple) + + batch_size = len(values) if pad_to_bsz is None else max(len(values), pad_to_bsz) + res = values[0].new(batch_size, size).fill_(pad_idx) + + def copy_tensor(src, dst): + assert dst.numel() == src.numel() + if move_eos_to_beginning: + if eos_idx is None: + # if no eos_idx is specified, then use the last token in src + dst[0] = src[-1] + else: + dst[0] = eos_idx + dst[1:] = src[:-1] + else: + dst.copy_(src) + + for i, v in enumerate(values): + copy_tensor(v, res[i][size - len(v) :] if left_pad else res[i][: len(v)]) + return res + + def collater_frm_label(self, targets, audio_size, audio_starts): + label_rate = self.label_rate + pad = self.num_classes[0] - 1 + assert label_rate > 0 + s2f = label_rate / self.sample_rate + frm_starts = [int(round(s * s2f)) for s in audio_starts] + frm_size = int(round(audio_size * s2f)) + if not self.pad_audio: + rem_size = [len(t) - s for t, s in zip(targets, frm_starts)] + frm_size = min(frm_size, *rem_size) + targets = [t[s : s + frm_size] for t, s in zip(targets, frm_starts)] + + lengths = torch.LongTensor([len(t) for t in targets]) + targets = self.collate_tokens(targets, pad_idx=pad, left_pad=False) + return targets, lengths + + +class HubertAsrDataset(torch.utils.data.Dataset): + """ + In this implementation, there will always be a single channel. + + Returns: + + .. code-block:: + + { + 'audio': (B x NumSamples) float tensor + } + """ + + def __init__( + self, + max_sample_size: Optional[int] = None, + sample_rate: float = 16000, + random_crop: bool = True, + pad_audio: bool = True, + do_normalize: bool = True, + ) -> None: + super().__init__() + self.sample_rate = sample_rate + self.random_crop = random_crop + self.pad_audio = pad_audio + self.normalize = do_normalize + self.max_sample_size = ( + max_sample_size if max_sample_size is not None else sys.maxsize + ) + + def __getitem__(self, cuts: CutSet) -> Dict[str, Any]: + self._validate(cuts) + audio, _ = read_audio_from_cuts(cuts) + for i, item in enumerate(audio): + audio[i] = self.postprocess(item, self.sample_rate) + audio_lens = [cut.num_samples for cut in cuts] + if self.pad_audio: + audio_size = min(max(audio_lens), self.max_sample_size) + else: + audio_size = min(min(audio_lens), self.max_sample_size) + + audio, padding_mask, audio_starts = self.collater_audio( + audio, audio_lens, audio_size + ) + + return { + "cuts": cuts, + "audio": audio, + "padding_mask": padding_mask, + "supervisions": default_collate( + [ + { + "text": supervision.text, + } + for sequence_idx, cut in enumerate(cuts) + for supervision in cut.supervisions + ] + ), + } + + def postprocess(self, wav, cur_sample_rate): + if wav.dim() == 2: + wav = wav.mean(-1) + assert wav.dim() == 1, wav.dim() + + if cur_sample_rate != self.sample_rate: + raise Exception(f"sr {cur_sample_rate} != {self.sample_rate}") + + if self.normalize: + with torch.no_grad(): + wav = F.layer_norm(wav, wav.shape) + return wav + + def _validate(self, cuts: CutSet) -> None: + validate(cuts) + assert all(cut.has_recording for cut in cuts) + + def crop_to_max_size(self, wav, target_size): + size = len(wav) + diff = size - target_size + if diff <= 0: + return wav, 0 + + start, end = 0, target_size + if self.random_crop: + start = np.random.randint(0, diff + 1) + end = size - diff + start + return wav[start:end], start + + def collater_audio(self, audios, audio_lens, audio_size): + collated_audios = audios[0].new_zeros(len(audios), audio_size) + padding_mask = ( + torch.BoolTensor(collated_audios.shape).fill_(False) + # if self.pad_audio else None + ) + audio_starts = [0 for _ in audios] + for i, (audio, audio_len) in enumerate(zip(audios, audio_lens)): + audio = audio[:audio_len] + diff = audio_len - audio_size + if diff == 0: + collated_audios[i] = audio + elif diff < 0: + assert self.pad_audio + collated_audios[i] = torch.cat([audio, audio.new_full((-diff,), 0.0)]) + padding_mask[i, diff:] = True + else: + collated_audios[i], audio_starts[i] = self.crop_to_max_size( + audio, audio_size + ) + return collated_audios, padding_mask, audio_starts + + def collate_tokens( + self, + values, + pad_idx, + eos_idx=None, + left_pad=False, + move_eos_to_beginning=False, + pad_to_length=None, + pad_to_multiple=1, + pad_to_bsz=None, + ): + """Convert a list of 1d tensors into a padded 2d tensor.""" + size = max(v.size(0) for v in values) + size = size if pad_to_length is None else max(size, pad_to_length) + if pad_to_multiple != 1 and size % pad_to_multiple != 0: + size = int(((size - 0.1) // pad_to_multiple + 1) * pad_to_multiple) + + batch_size = len(values) if pad_to_bsz is None else max(len(values), pad_to_bsz) + res = values[0].new(batch_size, size).fill_(pad_idx) + + def copy_tensor(src, dst): + assert dst.numel() == src.numel() + if move_eos_to_beginning: + if eos_idx is None: + # if no eos_idx is specified, then use the last token in src + dst[0] = src[-1] + else: + dst[0] = eos_idx + dst[1:] = src[:-1] + else: + dst.copy_(src) + + for i, v in enumerate(values): + copy_tensor(v, res[i][size - len(v) :] if left_pad else res[i][: len(v)]) + return res + + +if __name__ == "__main__": + from lhotse import load_manifest_lazy + from lhotse.dataset import DynamicBucketingSampler + from torch.utils.data import DataLoader + + dataset = HubertDataset() + cuts = load_manifest_lazy("data/fbank2/librispeech_cuts_train-clean-100.jsonl.gz") + sampler = DynamicBucketingSampler( + cuts, + max_duration=100, + shuffle=False, + ) + dl = DataLoader( + dataset, + batch_size=None, + sampler=sampler, + num_workers=2, + ) + + for batch_idx, batch in enumerate(dl): + print(batch) + break diff --git a/egs/librispeech/SSL/hubert/decode.py b/egs/librispeech/SSL/hubert/decode.py new file mode 100644 index 0000000000..837061b8cd --- /dev/null +++ b/egs/librispeech/SSL/hubert/decode.py @@ -0,0 +1,1045 @@ +#!/usr/bin/env python3 +# Copyright 2021-2024 Xiaomi Corp. (authors: Fangjun Kuang, +# Zengwei Yao, +# Yifan Yang) +# +# See ../../../../LICENSE for clarification regarding multiple authors +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +""" +Usage: +(1) greedy search +./hubert/decode.py \ + --epoch 28 \ + --avg 15 \ + --exp-dir ./hubert/exp \ + --max-duration 600 \ + --decoding-method greedy_search + +(2) beam search (not recommended) +./hubert/decode.py \ + --epoch 28 \ + --avg 15 \ + --exp-dir ./hubert/exp \ + --max-duration 600 \ + --decoding-method beam_search \ + --beam-size 4 + +(3) modified beam search +./hubert/decode.py \ + --epoch 28 \ + --avg 15 \ + --exp-dir ./hubert/exp \ + --max-duration 600 \ + --decoding-method modified_beam_search \ + --beam-size 4 + +(4) fast beam search (one best) +./hubert/decode.py \ + --epoch 28 \ + --avg 15 \ + --exp-dir ./hubert/exp \ + --max-duration 600 \ + --decoding-method fast_beam_search \ + --beam 20.0 \ + --max-contexts 8 \ + --max-states 64 + +(5) fast beam search (nbest) +./hubert/decode.py \ + --epoch 28 \ + --avg 15 \ + --exp-dir ./hubert/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) +./hubert/decode.py \ + --epoch 28 \ + --avg 15 \ + --exp-dir ./hubert/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) +./hubert/decode.py \ + --epoch 28 \ + --avg 15 \ + --exp-dir ./hubert/exp \ + --max-duration 600 \ + --decoding-method fast_beam_search_nbest_LG \ + --beam 20.0 \ + --max-contexts 8 \ + --max-states 64 +""" + + +import argparse +import logging +import math +import os +from collections import defaultdict +from pathlib import Path +from typing import Dict, List, Optional, Tuple + +import k2 +import sentencepiece as spm +import torch +import torch.nn as nn +from asr_datamodule import LibriSpeechAsrDataModule +from beam_search import ( + beam_search, + fast_beam_search_nbest, + fast_beam_search_nbest_LG, + fast_beam_search_nbest_oracle, + fast_beam_search_one_best, + greedy_search, + greedy_search_batch, + modified_beam_search, + modified_beam_search_lm_rescore, + modified_beam_search_lm_rescore_LODR, + modified_beam_search_lm_shallow_fusion, + modified_beam_search_LODR, +) +from finetune import add_model_arguments, get_model, get_params +from hubert import add_hubert_arguments + +from icefall import ContextGraph, LmScorer, NgramLm +from icefall.checkpoint import ( + average_checkpoints, + average_checkpoints_with_averaged_model, + find_checkpoints, + load_checkpoint, +) +from icefall.lexicon import Lexicon +from icefall.utils import ( + AttributeDict, + make_pad_mask, + setup_logger, + store_transcripts, + str2bool, + write_error_stats, +) + + +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="hubert/exp", + help="The experiment dir", + ) + + parser.add_argument( + "--bpe-model", + type=str, + default="data/lang_bpe_500/bpe.model", + help="Path to the BPE model", + ) + + parser.add_argument( + "--lang-dir", + type=Path, + default="data/lang_bpe_500", + help="The lang dir containing word table and LG graph", + ) + + parser.add_argument( + "--decoding-method", + type=str, + default="greedy_search", + help="""Possible values are: + - greedy_search + - beam_search + - modified_beam_search + - modified_beam_search_LODR + - fast_beam_search + - fast_beam_search_nbest + - fast_beam_search_nbest_oracle + - fast_beam_search_nbest_LG + If you use fast_beam_search_nbest_LG, you have to specify + `--lang-dir`, which should contain `LG.pt`. + """, + ) + + parser.add_argument( + "--beam-size", + type=int, + default=4, + help="""An integer indicating how many candidates we will keep for each + frame. Used only when --decoding-method is beam_search or + modified_beam_search.""", + ) + + parser.add_argument( + "--beam", + type=float, + default=20.0, + help="""A floating point value to calculate the cutoff score during beam + search (i.e., `cutoff = max-score - beam`), which is the same as the + `beam` in Kaldi. + Used only when --decoding-method is fast_beam_search, + fast_beam_search_nbest, fast_beam_search_nbest_LG, + and fast_beam_search_nbest_oracle + """, + ) + + parser.add_argument( + "--ngram-lm-scale", + type=float, + default=0.01, + help=""" + Used only when --decoding-method is fast_beam_search_nbest_LG. + It specifies the scale for n-gram LM scores. + """, + ) + + parser.add_argument( + "--max-contexts", + type=int, + default=8, + help="""Used only when --decoding-method is + fast_beam_search, fast_beam_search_nbest, fast_beam_search_nbest_LG, + and fast_beam_search_nbest_oracle""", + ) + + parser.add_argument( + "--max-states", + type=int, + default=64, + help="""Used only when --decoding-method is + fast_beam_search, fast_beam_search_nbest, fast_beam_search_nbest_LG, + and fast_beam_search_nbest_oracle""", + ) + + parser.add_argument( + "--context-size", + type=int, + default=2, + help="The context size in the decoder. 1 means bigram; " "2 means tri-gram", + ) + parser.add_argument( + "--max-sym-per-frame", + type=int, + default=1, + help="""Maximum number of symbols per frame. + Used only when --decoding-method is greedy_search""", + ) + + parser.add_argument( + "--num-paths", + type=int, + default=200, + help="""Number of paths for nbest decoding. + Used only when the decoding method is fast_beam_search_nbest, + fast_beam_search_nbest_LG, and fast_beam_search_nbest_oracle""", + ) + + parser.add_argument( + "--nbest-scale", + type=float, + default=0.5, + help="""Scale applied to lattice scores when computing nbest paths. + Used only when the decoding method is fast_beam_search_nbest, + fast_beam_search_nbest_LG, and fast_beam_search_nbest_oracle""", + ) + + parser.add_argument( + "--use-shallow-fusion", + type=str2bool, + default=False, + help="""Use neural network LM for shallow fusion. + If you want to use LODR, you will also need to set this to true + """, + ) + + parser.add_argument( + "--lm-type", + type=str, + default="rnn", + help="Type of NN lm", + choices=["rnn", "transformer"], + ) + + parser.add_argument( + "--lm-scale", + type=float, + default=0.3, + help="""The scale of the neural network LM + Used only when `--use-shallow-fusion` is set to True. + """, + ) + + parser.add_argument( + "--tokens-ngram", + type=int, + default=2, + help="""The order of the ngram lm. + """, + ) + + parser.add_argument( + "--backoff-id", + type=int, + default=500, + help="ID of the backoff symbol in the ngram LM", + ) + + parser.add_argument( + "--context-score", + type=float, + default=2, + help=""" + The bonus score of each token for the context biasing words/phrases. + Used only when --decoding-method is modified_beam_search and + modified_beam_search_LODR. + """, + ) + + parser.add_argument( + "--context-file", + type=str, + default="", + help=""" + The path of the context biasing lists, one word/phrase each line + Used only when --decoding-method is modified_beam_search and + modified_beam_search_LODR. + """, + ) + + add_hubert_arguments(parser) + add_model_arguments(parser) + + return parser + + +def decode_one_batch( + params: AttributeDict, + model: nn.Module, + sp: spm.SentencePieceProcessor, + batch: dict, + word_table: Optional[k2.SymbolTable] = None, + decoding_graph: Optional[k2.Fsa] = None, + context_graph: Optional[ContextGraph] = None, + LM: Optional[LmScorer] = None, + ngram_lm=None, + ngram_lm_scale: float = 0.0, +) -> Dict[str, List[List[str]]]: + """Decode one batch and return the result in a dict. The dict has the + following format: + + - key: It indicates the setting used for decoding. For example, + if greedy_search is used, it would be "greedy_search" + If beam search with a beam size of 7 is used, it would be + "beam_7" + - value: It contains the decoding result. `len(value)` equals to + batch size. `value[i]` is the decoding result for the i-th + utterance in the given batch. + Args: + params: + It's the return value of :func:`get_params`. + model: + The neural model. + sp: + The BPE model. + batch: + It is the return value from iterating + `lhotse.dataset.K2SpeechRecognitionDataset`. See its documentation + for the format of the `batch`. + word_table: + The word symbol table. + decoding_graph: + The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used + only when --decoding-method is fast_beam_search, fast_beam_search_nbest, + fast_beam_search_nbest_oracle, and fast_beam_search_nbest_LG. + LM: + A neural network language model. + ngram_lm: + A ngram language model + ngram_lm_scale: + The scale for the ngram language model. + Returns: + Return the decoding result. See above description for the format of + the returned dict. + """ + device = next(model.parameters()).device + audio = batch["audio"].to(device) + padding_mask = batch["padding_mask"].to(device) + + encoder_out, encoder_out_lens = model.forward_encoder(audio, padding_mask) + + 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(batch["supervisions"]["text"]), + nbest_scale=params.nbest_scale, + ) + for hyp in sp.decode(hyp_tokens): + hyps.append(hyp.split()) + elif params.decoding_method == "greedy_search" and params.max_sym_per_frame == 1: + hyp_tokens = greedy_search_batch( + model=model, + encoder_out=encoder_out, + encoder_out_lens=encoder_out_lens, + ) + for hyp in sp.decode(hyp_tokens): + hyps.append(hyp.split()) + elif params.decoding_method == "modified_beam_search": + hyp_tokens = modified_beam_search( + model=model, + encoder_out=encoder_out, + encoder_out_lens=encoder_out_lens, + beam=params.beam_size, + context_graph=context_graph, + ) + for hyp in sp.decode(hyp_tokens): + hyps.append(hyp.split()) + elif params.decoding_method == "modified_beam_search_lm_shallow_fusion": + hyp_tokens = modified_beam_search_lm_shallow_fusion( + model=model, + encoder_out=encoder_out, + encoder_out_lens=encoder_out_lens, + beam=params.beam_size, + LM=LM, + ) + for hyp in sp.decode(hyp_tokens): + hyps.append(hyp.split()) + elif params.decoding_method == "modified_beam_search_LODR": + hyp_tokens = modified_beam_search_LODR( + model=model, + encoder_out=encoder_out, + encoder_out_lens=encoder_out_lens, + beam=params.beam_size, + LODR_lm=ngram_lm, + LODR_lm_scale=ngram_lm_scale, + LM=LM, + context_graph=context_graph, + ) + for hyp in sp.decode(hyp_tokens): + hyps.append(hyp.split()) + elif params.decoding_method == "modified_beam_search_lm_rescore": + lm_scale_list = [0.01 * i for i in range(10, 50)] + ans_dict = modified_beam_search_lm_rescore( + model=model, + encoder_out=encoder_out, + encoder_out_lens=encoder_out_lens, + beam=params.beam_size, + LM=LM, + lm_scale_list=lm_scale_list, + ) + elif params.decoding_method == "modified_beam_search_lm_rescore_LODR": + lm_scale_list = [0.02 * i for i in range(2, 30)] + ans_dict = modified_beam_search_lm_rescore_LODR( + model=model, + encoder_out=encoder_out, + encoder_out_lens=encoder_out_lens, + beam=params.beam_size, + LM=LM, + LODR_lm=ngram_lm, + sp=sp, + lm_scale_list=lm_scale_list, + ) + else: + batch_size = encoder_out.size(0) + + for i in range(batch_size): + # fmt: off + encoder_out_i = encoder_out[i:i+1, :encoder_out_lens[i]] + # fmt: on + if params.decoding_method == "greedy_search": + hyp = greedy_search( + model=model, + encoder_out=encoder_out_i, + max_sym_per_frame=params.max_sym_per_frame, + ) + elif params.decoding_method == "beam_search": + hyp = beam_search( + model=model, + encoder_out=encoder_out_i, + beam=params.beam_size, + ) + else: + raise ValueError( + f"Unsupported decoding method: {params.decoding_method}" + ) + hyps.append(sp.decode(hyp).split()) + + if params.decoding_method == "greedy_search": + return {"greedy_search": hyps} + elif "fast_beam_search" in params.decoding_method: + key = f"beam_{params.beam}_" + key += f"max_contexts_{params.max_contexts}_" + key += f"max_states_{params.max_states}" + if "nbest" in params.decoding_method: + key += f"_num_paths_{params.num_paths}_" + key += f"nbest_scale_{params.nbest_scale}" + if "LG" in params.decoding_method: + key += f"_ngram_lm_scale_{params.ngram_lm_scale}" + + return {key: hyps} + elif "modified_beam_search" in params.decoding_method: + prefix = f"beam_size_{params.beam_size}" + if params.decoding_method in ( + "modified_beam_search_lm_rescore", + "modified_beam_search_lm_rescore_LODR", + ): + ans = dict() + assert ans_dict is not None + for key, hyps in ans_dict.items(): + hyps = [sp.decode(hyp).split() for hyp in hyps] + ans[f"{prefix}_{key}"] = hyps + return ans + else: + if params.has_contexts: + prefix += f"-context-score-{params.context_score}" + return {prefix: hyps} + else: + return {f"beam_size_{params.beam_size}": hyps} + + +def decode_dataset( + dl: torch.utils.data.DataLoader, + params: AttributeDict, + model: nn.Module, + sp: spm.SentencePieceProcessor, + word_table: Optional[k2.SymbolTable] = None, + decoding_graph: Optional[k2.Fsa] = None, + context_graph: Optional[ContextGraph] = None, + LM: Optional[LmScorer] = None, + ngram_lm=None, + ngram_lm_scale: float = 0.0, +) -> Dict[str, List[Tuple[str, List[str], List[str]]]]: + """Decode dataset. + + Args: + dl: + PyTorch's dataloader containing the dataset to decode. + params: + It is returned by :func:`get_params`. + model: + The neural model. + sp: + The BPE model. + word_table: + The word symbol table. + decoding_graph: + The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used + only when --decoding-method is fast_beam_search, fast_beam_search_nbest, + fast_beam_search_nbest_oracle, and fast_beam_search_nbest_LG. + Returns: + Return a dict, whose key may be "greedy_search" if greedy search + is used, or it may be "beam_7" if beam size of 7 is used. + Its value is a list of tuples. Each tuple contains two elements: + The first is the reference transcript, and the second is the + predicted result. + """ + num_cuts = 0 + + try: + num_batches = len(dl) + except TypeError: + num_batches = "?" + + if params.decoding_method == "greedy_search": + log_interval = 50 + else: + log_interval = 20 + + results = defaultdict(list) + for batch_idx, batch in enumerate(dl): + texts = batch["supervisions"]["text"] + cut_ids = [cut.id for cut in batch["cuts"]] + + hyps_dict = decode_one_batch( + params=params, + model=model, + sp=sp, + decoding_graph=decoding_graph, + context_graph=context_graph, + word_table=word_table, + batch=batch, + LM=LM, + ngram_lm=ngram_lm, + ngram_lm_scale=ngram_lm_scale, + ) + + for name, hyps in hyps_dict.items(): + this_batch = [] + assert len(hyps) == len(texts) + for cut_id, hyp_words, ref_text in zip(cut_ids, hyps, texts): + ref_words = ref_text.split() + this_batch.append((cut_id, ref_words, hyp_words)) + + results[name].extend(this_batch) + + num_cuts += len(texts) + + if batch_idx % log_interval == 0: + batch_str = f"{batch_idx}/{num_batches}" + + logging.info(f"batch {batch_str}, cuts processed until now is {num_cuts}") + return results + + +def save_results( + params: AttributeDict, + test_set_name: str, + results_dict: Dict[str, List[Tuple[str, List[str], List[str]]]], +): + test_set_wers = dict() + for key, results in results_dict.items(): + recog_path = ( + params.res_dir / f"recogs-{test_set_name}-{key}-{params.suffix}.txt" + ) + results = sorted(results) + store_transcripts(filename=recog_path, texts=results) + logging.info(f"The transcripts are stored in {recog_path}") + + # The following prints out WERs, per-word error statistics and aligned + # ref/hyp pairs. + errs_filename = ( + params.res_dir / f"errs-{test_set_name}-{key}-{params.suffix}.txt" + ) + with open(errs_filename, "w") as f: + wer = write_error_stats( + f, f"{test_set_name}-{key}", results, enable_log=True + ) + test_set_wers[key] = wer + + logging.info("Wrote detailed error stats to {}".format(errs_filename)) + + test_set_wers = sorted(test_set_wers.items(), key=lambda x: x[1]) + errs_info = ( + params.res_dir / f"wer-summary-{test_set_name}-{key}-{params.suffix}.txt" + ) + with open(errs_info, "w") as f: + print("settings\tWER", file=f) + for key, val in test_set_wers: + print("{}\t{}".format(key, val), file=f) + + s = "\nFor {}, WER of different settings are:\n".format(test_set_name) + note = "\tbest for {}".format(test_set_name) + for key, val in test_set_wers: + s += "{}\t{}{}\n".format(key, val, note) + note = "" + logging.info(s) + + +@torch.no_grad() +def main(): + parser = get_parser() + LibriSpeechAsrDataModule.add_arguments(parser) + LmScorer.add_arguments(parser) + args = parser.parse_args() + args.exp_dir = Path(args.exp_dir) + + params = get_params() + params.update(vars(args)) + + assert params.decoding_method in ( + "greedy_search", + "beam_search", + "fast_beam_search", + "fast_beam_search_nbest", + "fast_beam_search_nbest_LG", + "fast_beam_search_nbest_oracle", + "modified_beam_search", + "modified_beam_search_LODR", + "modified_beam_search_lm_shallow_fusion", + "modified_beam_search_lm_rescore", + "modified_beam_search_lm_rescore_LODR", + ) + params.res_dir = params.exp_dir / params.decoding_method + + if os.path.exists(params.context_file): + params.has_contexts = True + else: + params.has_contexts = False + + if params.iter > 0: + params.suffix = f"iter-{params.iter}-avg-{params.avg}" + else: + params.suffix = f"epoch-{params.epoch}-avg-{params.avg}" + + if "fast_beam_search" in params.decoding_method: + params.suffix += f"-beam-{params.beam}" + params.suffix += f"-max-contexts-{params.max_contexts}" + params.suffix += f"-max-states-{params.max_states}" + if "nbest" in params.decoding_method: + params.suffix += f"-nbest-scale-{params.nbest_scale}" + params.suffix += f"-num-paths-{params.num_paths}" + if "LG" in params.decoding_method: + params.suffix += f"-ngram-lm-scale-{params.ngram_lm_scale}" + elif "beam_search" in params.decoding_method: + params.suffix += f"-{params.decoding_method}-beam-size-{params.beam_size}" + if params.decoding_method in ( + "modified_beam_search", + "modified_beam_search_LODR", + ): + if params.has_contexts: + params.suffix += f"-context-score-{params.context_score}" + else: + params.suffix += f"-context-{params.context_size}" + params.suffix += f"-max-sym-per-frame-{params.max_sym_per_frame}" + + if params.use_shallow_fusion: + params.suffix += f"-{params.lm_type}-lm-scale-{params.lm_scale}" + + if "LODR" in params.decoding_method: + params.suffix += ( + f"-LODR-{params.tokens_ngram}gram-scale-{params.ngram_lm_scale}" + ) + + if params.use_averaged_model: + params.suffix += "-use-averaged-model" + + setup_logger(f"{params.res_dir}/log-decode-{params.suffix}") + logging.info("Decoding started") + + device = torch.device("cpu") + if torch.cuda.is_available(): + device = torch.device("cuda", 0) + + logging.info(f"Device: {device}") + + sp = spm.SentencePieceProcessor() + sp.load(params.bpe_model) + + # and are defined in local/train_bpe_model.py + params.blank_id = sp.piece_to_id("") + params.unk_id = sp.piece_to_id("") + params.vocab_size = sp.get_piece_size() + + logging.info(params) + + logging.info("About to create model") + model = get_model(params) + + if not params.use_averaged_model: + if params.iter > 0: + filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[ + : params.avg + ] + if len(filenames) == 0: + raise ValueError( + f"No checkpoints found for" + f" --iter {params.iter}, --avg {params.avg}" + ) + elif len(filenames) < params.avg: + raise ValueError( + f"Not enough checkpoints ({len(filenames)}) found for" + f" --iter {params.iter}, --avg {params.avg}" + ) + logging.info(f"averaging {filenames}") + model.to(device) + model.load_state_dict(average_checkpoints(filenames, device=device)) + elif params.avg == 1: + load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model) + else: + start = params.epoch - params.avg + 1 + filenames = [] + for i in range(start, params.epoch + 1): + if i >= 1: + filenames.append(f"{params.exp_dir}/epoch-{i}.pt") + logging.info(f"averaging {filenames}") + model.to(device) + model.load_state_dict(average_checkpoints(filenames, device=device)) + else: + if params.iter > 0: + filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[ + : params.avg + 1 + ] + if len(filenames) == 0: + raise ValueError( + f"No checkpoints found for" + f" --iter {params.iter}, --avg {params.avg}" + ) + elif len(filenames) < params.avg + 1: + raise ValueError( + f"Not enough checkpoints ({len(filenames)}) found for" + f" --iter {params.iter}, --avg {params.avg}" + ) + filename_start = filenames[-1] + filename_end = filenames[0] + logging.info( + "Calculating the averaged model over iteration checkpoints" + f" from {filename_start} (excluded) to {filename_end}" + ) + model.to(device) + model.load_state_dict( + average_checkpoints_with_averaged_model( + filename_start=filename_start, + filename_end=filename_end, + device=device, + ) + ) + else: + assert params.avg > 0, params.avg + start = params.epoch - params.avg + assert start >= 1, start + filename_start = f"{params.exp_dir}/epoch-{start}.pt" + filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt" + logging.info( + f"Calculating the averaged model over epoch range from " + f"{start} (excluded) to {params.epoch}" + ) + model.to(device) + model.load_state_dict( + average_checkpoints_with_averaged_model( + filename_start=filename_start, + filename_end=filename_end, + device=device, + ) + ) + + model.to(device) + model.eval() + + # only load the neural network LM if required + if params.use_shallow_fusion or params.decoding_method in ( + "modified_beam_search_lm_rescore", + "modified_beam_search_lm_rescore_LODR", + "modified_beam_search_lm_shallow_fusion", + "modified_beam_search_LODR", + ): + LM = LmScorer( + lm_type=params.lm_type, + params=params, + device=device, + lm_scale=params.lm_scale, + ) + LM.to(device) + LM.eval() + else: + LM = None + + # only load N-gram LM when needed + if params.decoding_method == "modified_beam_search_lm_rescore_LODR": + try: + import kenlm + except ImportError: + print("Please install kenlm first. You can use") + print(" pip install https://github.com/kpu/kenlm/archive/master.zip") + print("to install it") + import sys + + sys.exit(-1) + ngram_file_name = str(params.lang_dir / f"{params.tokens_ngram}gram.arpa") + logging.info(f"lm filename: {ngram_file_name}") + ngram_lm = kenlm.Model(ngram_file_name) + ngram_lm_scale = None # use a list to search + + elif params.decoding_method == "modified_beam_search_LODR": + lm_filename = f"{params.tokens_ngram}gram.fst.txt" + logging.info(f"Loading token level lm: {lm_filename}") + ngram_lm = NgramLm( + str(params.lang_dir / lm_filename), + backoff_id=params.backoff_id, + is_binary=False, + ) + logging.info(f"num states: {ngram_lm.lm.num_states}") + ngram_lm_scale = params.ngram_lm_scale + else: + ngram_lm = None + ngram_lm_scale = None + + if "fast_beam_search" in params.decoding_method: + if params.decoding_method == "fast_beam_search_nbest_LG": + lexicon = Lexicon(params.lang_dir) + word_table = lexicon.word_table + lg_filename = params.lang_dir / "LG.pt" + logging.info(f"Loading {lg_filename}") + decoding_graph = k2.Fsa.from_dict( + torch.load(lg_filename, map_location=device) + ) + decoding_graph.scores *= params.ngram_lm_scale + else: + word_table = None + decoding_graph = k2.trivial_graph(params.vocab_size - 1, device=device) + else: + decoding_graph = None + word_table = None + + if "modified_beam_search" in params.decoding_method: + if os.path.exists(params.context_file): + contexts = [] + for line in open(params.context_file).readlines(): + contexts.append((sp.encode(line.strip()), 0.0)) + context_graph = ContextGraph(params.context_score) + context_graph.build(contexts) + else: + context_graph = None + else: + context_graph = None + + num_param = sum([p.numel() for p in model.parameters()]) + logging.info(f"Number of model parameters: {num_param}") + + # we need cut ids to display recognition results. + args.return_cuts = True + librispeech = LibriSpeechAsrDataModule(args) + + dev_clean_cuts = librispeech.dev_clean_cuts() + dev_other_cuts = librispeech.dev_other_cuts() + + dev_clean_dl = librispeech.test_dataloaders( + dev_clean_cuts, + do_normalize=params.do_normalize, + ) + dev_other_dl = librispeech.test_dataloaders( + dev_other_cuts, + do_normalize=params.do_normalize, + ) + + test_clean_cuts = librispeech.test_clean_cuts() + test_other_cuts = librispeech.test_other_cuts() + + test_clean_dl = librispeech.test_dataloaders( + test_clean_cuts, + do_normalize=params.do_normalize, + ) + test_other_dl = librispeech.test_dataloaders( + test_other_cuts, + do_normalize=params.do_normalize, + ) + + test_sets = ["dev-clean", "dev-other", "test-clean", "test-other"] + test_dl = [dev_clean_dl, dev_other_dl, test_clean_dl, test_other_dl] + + for test_set, test_dl in zip(test_sets, test_dl): + results_dict = decode_dataset( + dl=test_dl, + params=params, + model=model, + sp=sp, + word_table=word_table, + decoding_graph=decoding_graph, + context_graph=context_graph, + LM=LM, + ngram_lm=ngram_lm, + ngram_lm_scale=ngram_lm_scale, + ) + + save_results( + params=params, + test_set_name=test_set, + results_dict=results_dict, + ) + + logging.info("Done!") + + +if __name__ == "__main__": + main() diff --git a/egs/librispeech/SSL/hubert/decode_ce.py b/egs/librispeech/SSL/hubert/decode_ce.py new file mode 100644 index 0000000000..a8d8bc9c2f --- /dev/null +++ b/egs/librispeech/SSL/hubert/decode_ce.py @@ -0,0 +1,1045 @@ +#!/usr/bin/env python3 +# Copyright 2021-2024 Xiaomi Corp. (authors: Fangjun Kuang, +# Zengwei Yao, +# Yifan Yang) +# +# See ../../../../LICENSE for clarification regarding multiple authors +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +""" +Usage: +(1) greedy search +./hubert/decode.py \ + --epoch 28 \ + --avg 15 \ + --exp-dir ./hubert/exp \ + --max-duration 600 \ + --decoding-method greedy_search + +(2) beam search (not recommended) +./hubert/decode.py \ + --epoch 28 \ + --avg 15 \ + --exp-dir ./hubert/exp \ + --max-duration 600 \ + --decoding-method beam_search \ + --beam-size 4 + +(3) modified beam search +./hubert/decode.py \ + --epoch 28 \ + --avg 15 \ + --exp-dir ./hubert/exp \ + --max-duration 600 \ + --decoding-method modified_beam_search \ + --beam-size 4 + +(4) fast beam search (one best) +./hubert/decode.py \ + --epoch 28 \ + --avg 15 \ + --exp-dir ./hubert/exp \ + --max-duration 600 \ + --decoding-method fast_beam_search \ + --beam 20.0 \ + --max-contexts 8 \ + --max-states 64 + +(5) fast beam search (nbest) +./hubert/decode.py \ + --epoch 28 \ + --avg 15 \ + --exp-dir ./hubert/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) +./hubert/decode.py \ + --epoch 28 \ + --avg 15 \ + --exp-dir ./hubert/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) +./hubert/decode.py \ + --epoch 28 \ + --avg 15 \ + --exp-dir ./hubert/exp \ + --max-duration 600 \ + --decoding-method fast_beam_search_nbest_LG \ + --beam 20.0 \ + --max-contexts 8 \ + --max-states 64 +""" + + +import argparse +import logging +import math +import os +from collections import defaultdict +from pathlib import Path +from typing import Dict, List, Optional, Tuple + +import k2 +import sentencepiece as spm +import torch +import torch.nn as nn +from asr_datamodule import LibriSpeechAsrDataModule +from beam_search import ( + beam_search, + fast_beam_search_nbest, + fast_beam_search_nbest_LG, + fast_beam_search_nbest_oracle, + fast_beam_search_one_best, + greedy_search, + greedy_search_batch, + modified_beam_search, + modified_beam_search_lm_rescore, + modified_beam_search_lm_rescore_LODR, + modified_beam_search_lm_shallow_fusion, + modified_beam_search_LODR, +) +from finetune_ce import add_model_arguments, get_model, get_params +from hubert_ce import add_hubert_arguments + +from icefall import ContextGraph, LmScorer, NgramLm +from icefall.checkpoint import ( + average_checkpoints, + average_checkpoints_with_averaged_model, + find_checkpoints, + load_checkpoint, +) +from icefall.lexicon import Lexicon +from icefall.utils import ( + AttributeDict, + make_pad_mask, + setup_logger, + store_transcripts, + str2bool, + write_error_stats, +) + + +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="hubert/exp", + help="The experiment dir", + ) + + parser.add_argument( + "--bpe-model", + type=str, + default="data/lang_bpe_500/bpe.model", + help="Path to the BPE model", + ) + + parser.add_argument( + "--lang-dir", + type=Path, + default="data/lang_bpe_500", + help="The lang dir containing word table and LG graph", + ) + + parser.add_argument( + "--decoding-method", + type=str, + default="greedy_search", + help="""Possible values are: + - greedy_search + - beam_search + - modified_beam_search + - modified_beam_search_LODR + - fast_beam_search + - fast_beam_search_nbest + - fast_beam_search_nbest_oracle + - fast_beam_search_nbest_LG + If you use fast_beam_search_nbest_LG, you have to specify + `--lang-dir`, which should contain `LG.pt`. + """, + ) + + parser.add_argument( + "--beam-size", + type=int, + default=4, + help="""An integer indicating how many candidates we will keep for each + frame. Used only when --decoding-method is beam_search or + modified_beam_search.""", + ) + + parser.add_argument( + "--beam", + type=float, + default=20.0, + help="""A floating point value to calculate the cutoff score during beam + search (i.e., `cutoff = max-score - beam`), which is the same as the + `beam` in Kaldi. + Used only when --decoding-method is fast_beam_search, + fast_beam_search_nbest, fast_beam_search_nbest_LG, + and fast_beam_search_nbest_oracle + """, + ) + + parser.add_argument( + "--ngram-lm-scale", + type=float, + default=0.01, + help=""" + Used only when --decoding-method is fast_beam_search_nbest_LG. + It specifies the scale for n-gram LM scores. + """, + ) + + parser.add_argument( + "--max-contexts", + type=int, + default=8, + help="""Used only when --decoding-method is + fast_beam_search, fast_beam_search_nbest, fast_beam_search_nbest_LG, + and fast_beam_search_nbest_oracle""", + ) + + parser.add_argument( + "--max-states", + type=int, + default=64, + help="""Used only when --decoding-method is + fast_beam_search, fast_beam_search_nbest, fast_beam_search_nbest_LG, + and fast_beam_search_nbest_oracle""", + ) + + parser.add_argument( + "--context-size", + type=int, + default=2, + help="The context size in the decoder. 1 means bigram; " "2 means tri-gram", + ) + parser.add_argument( + "--max-sym-per-frame", + type=int, + default=1, + help="""Maximum number of symbols per frame. + Used only when --decoding-method is greedy_search""", + ) + + parser.add_argument( + "--num-paths", + type=int, + default=200, + help="""Number of paths for nbest decoding. + Used only when the decoding method is fast_beam_search_nbest, + fast_beam_search_nbest_LG, and fast_beam_search_nbest_oracle""", + ) + + parser.add_argument( + "--nbest-scale", + type=float, + default=0.5, + help="""Scale applied to lattice scores when computing nbest paths. + Used only when the decoding method is fast_beam_search_nbest, + fast_beam_search_nbest_LG, and fast_beam_search_nbest_oracle""", + ) + + parser.add_argument( + "--use-shallow-fusion", + type=str2bool, + default=False, + help="""Use neural network LM for shallow fusion. + If you want to use LODR, you will also need to set this to true + """, + ) + + parser.add_argument( + "--lm-type", + type=str, + default="rnn", + help="Type of NN lm", + choices=["rnn", "transformer"], + ) + + parser.add_argument( + "--lm-scale", + type=float, + default=0.3, + help="""The scale of the neural network LM + Used only when `--use-shallow-fusion` is set to True. + """, + ) + + parser.add_argument( + "--tokens-ngram", + type=int, + default=2, + help="""The order of the ngram lm. + """, + ) + + parser.add_argument( + "--backoff-id", + type=int, + default=500, + help="ID of the backoff symbol in the ngram LM", + ) + + parser.add_argument( + "--context-score", + type=float, + default=2, + help=""" + The bonus score of each token for the context biasing words/phrases. + Used only when --decoding-method is modified_beam_search and + modified_beam_search_LODR. + """, + ) + + parser.add_argument( + "--context-file", + type=str, + default="", + help=""" + The path of the context biasing lists, one word/phrase each line + Used only when --decoding-method is modified_beam_search and + modified_beam_search_LODR. + """, + ) + + add_hubert_arguments(parser) + add_model_arguments(parser) + + return parser + + +def decode_one_batch( + params: AttributeDict, + model: nn.Module, + sp: spm.SentencePieceProcessor, + batch: dict, + word_table: Optional[k2.SymbolTable] = None, + decoding_graph: Optional[k2.Fsa] = None, + context_graph: Optional[ContextGraph] = None, + LM: Optional[LmScorer] = None, + ngram_lm=None, + ngram_lm_scale: float = 0.0, +) -> Dict[str, List[List[str]]]: + """Decode one batch and return the result in a dict. The dict has the + following format: + + - key: It indicates the setting used for decoding. For example, + if greedy_search is used, it would be "greedy_search" + If beam search with a beam size of 7 is used, it would be + "beam_7" + - value: It contains the decoding result. `len(value)` equals to + batch size. `value[i]` is the decoding result for the i-th + utterance in the given batch. + Args: + params: + It's the return value of :func:`get_params`. + model: + The neural model. + sp: + The BPE model. + batch: + It is the return value from iterating + `lhotse.dataset.K2SpeechRecognitionDataset`. See its documentation + for the format of the `batch`. + word_table: + The word symbol table. + decoding_graph: + The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used + only when --decoding-method is fast_beam_search, fast_beam_search_nbest, + fast_beam_search_nbest_oracle, and fast_beam_search_nbest_LG. + LM: + A neural network language model. + ngram_lm: + A ngram language model + ngram_lm_scale: + The scale for the ngram language model. + Returns: + Return the decoding result. See above description for the format of + the returned dict. + """ + device = next(model.parameters()).device + audio = batch["audio"].to(device) + padding_mask = batch["padding_mask"].to(device) + + encoder_out, encoder_out_lens = model.forward_encoder(audio, padding_mask) + + 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(batch["supervisions"]["text"]), + nbest_scale=params.nbest_scale, + ) + for hyp in sp.decode(hyp_tokens): + hyps.append(hyp.split()) + elif params.decoding_method == "greedy_search" and params.max_sym_per_frame == 1: + hyp_tokens = greedy_search_batch( + model=model, + encoder_out=encoder_out, + encoder_out_lens=encoder_out_lens, + ) + for hyp in sp.decode(hyp_tokens): + hyps.append(hyp.split()) + elif params.decoding_method == "modified_beam_search": + hyp_tokens = modified_beam_search( + model=model, + encoder_out=encoder_out, + encoder_out_lens=encoder_out_lens, + beam=params.beam_size, + context_graph=context_graph, + ) + for hyp in sp.decode(hyp_tokens): + hyps.append(hyp.split()) + elif params.decoding_method == "modified_beam_search_lm_shallow_fusion": + hyp_tokens = modified_beam_search_lm_shallow_fusion( + model=model, + encoder_out=encoder_out, + encoder_out_lens=encoder_out_lens, + beam=params.beam_size, + LM=LM, + ) + for hyp in sp.decode(hyp_tokens): + hyps.append(hyp.split()) + elif params.decoding_method == "modified_beam_search_LODR": + hyp_tokens = modified_beam_search_LODR( + model=model, + encoder_out=encoder_out, + encoder_out_lens=encoder_out_lens, + beam=params.beam_size, + LODR_lm=ngram_lm, + LODR_lm_scale=ngram_lm_scale, + LM=LM, + context_graph=context_graph, + ) + for hyp in sp.decode(hyp_tokens): + hyps.append(hyp.split()) + elif params.decoding_method == "modified_beam_search_lm_rescore": + lm_scale_list = [0.01 * i for i in range(10, 50)] + ans_dict = modified_beam_search_lm_rescore( + model=model, + encoder_out=encoder_out, + encoder_out_lens=encoder_out_lens, + beam=params.beam_size, + LM=LM, + lm_scale_list=lm_scale_list, + ) + elif params.decoding_method == "modified_beam_search_lm_rescore_LODR": + lm_scale_list = [0.02 * i for i in range(2, 30)] + ans_dict = modified_beam_search_lm_rescore_LODR( + model=model, + encoder_out=encoder_out, + encoder_out_lens=encoder_out_lens, + beam=params.beam_size, + LM=LM, + LODR_lm=ngram_lm, + sp=sp, + lm_scale_list=lm_scale_list, + ) + else: + batch_size = encoder_out.size(0) + + for i in range(batch_size): + # fmt: off + encoder_out_i = encoder_out[i:i+1, :encoder_out_lens[i]] + # fmt: on + if params.decoding_method == "greedy_search": + hyp = greedy_search( + model=model, + encoder_out=encoder_out_i, + max_sym_per_frame=params.max_sym_per_frame, + ) + elif params.decoding_method == "beam_search": + hyp = beam_search( + model=model, + encoder_out=encoder_out_i, + beam=params.beam_size, + ) + else: + raise ValueError( + f"Unsupported decoding method: {params.decoding_method}" + ) + hyps.append(sp.decode(hyp).split()) + + if params.decoding_method == "greedy_search": + return {"greedy_search": hyps} + elif "fast_beam_search" in params.decoding_method: + key = f"beam_{params.beam}_" + key += f"max_contexts_{params.max_contexts}_" + key += f"max_states_{params.max_states}" + if "nbest" in params.decoding_method: + key += f"_num_paths_{params.num_paths}_" + key += f"nbest_scale_{params.nbest_scale}" + if "LG" in params.decoding_method: + key += f"_ngram_lm_scale_{params.ngram_lm_scale}" + + return {key: hyps} + elif "modified_beam_search" in params.decoding_method: + prefix = f"beam_size_{params.beam_size}" + if params.decoding_method in ( + "modified_beam_search_lm_rescore", + "modified_beam_search_lm_rescore_LODR", + ): + ans = dict() + assert ans_dict is not None + for key, hyps in ans_dict.items(): + hyps = [sp.decode(hyp).split() for hyp in hyps] + ans[f"{prefix}_{key}"] = hyps + return ans + else: + if params.has_contexts: + prefix += f"-context-score-{params.context_score}" + return {prefix: hyps} + else: + return {f"beam_size_{params.beam_size}": hyps} + + +def decode_dataset( + dl: torch.utils.data.DataLoader, + params: AttributeDict, + model: nn.Module, + sp: spm.SentencePieceProcessor, + word_table: Optional[k2.SymbolTable] = None, + decoding_graph: Optional[k2.Fsa] = None, + context_graph: Optional[ContextGraph] = None, + LM: Optional[LmScorer] = None, + ngram_lm=None, + ngram_lm_scale: float = 0.0, +) -> Dict[str, List[Tuple[str, List[str], List[str]]]]: + """Decode dataset. + + Args: + dl: + PyTorch's dataloader containing the dataset to decode. + params: + It is returned by :func:`get_params`. + model: + The neural model. + sp: + The BPE model. + word_table: + The word symbol table. + decoding_graph: + The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used + only when --decoding-method is fast_beam_search, fast_beam_search_nbest, + fast_beam_search_nbest_oracle, and fast_beam_search_nbest_LG. + Returns: + Return a dict, whose key may be "greedy_search" if greedy search + is used, or it may be "beam_7" if beam size of 7 is used. + Its value is a list of tuples. Each tuple contains two elements: + The first is the reference transcript, and the second is the + predicted result. + """ + num_cuts = 0 + + try: + num_batches = len(dl) + except TypeError: + num_batches = "?" + + if params.decoding_method == "greedy_search": + log_interval = 50 + else: + log_interval = 20 + + results = defaultdict(list) + for batch_idx, batch in enumerate(dl): + texts = batch["supervisions"]["text"] + cut_ids = [cut.id for cut in batch["cuts"]] + + hyps_dict = decode_one_batch( + params=params, + model=model, + sp=sp, + decoding_graph=decoding_graph, + context_graph=context_graph, + word_table=word_table, + batch=batch, + LM=LM, + ngram_lm=ngram_lm, + ngram_lm_scale=ngram_lm_scale, + ) + + for name, hyps in hyps_dict.items(): + this_batch = [] + assert len(hyps) == len(texts) + for cut_id, hyp_words, ref_text in zip(cut_ids, hyps, texts): + ref_words = ref_text.split() + this_batch.append((cut_id, ref_words, hyp_words)) + + results[name].extend(this_batch) + + num_cuts += len(texts) + + if batch_idx % log_interval == 0: + batch_str = f"{batch_idx}/{num_batches}" + + logging.info(f"batch {batch_str}, cuts processed until now is {num_cuts}") + return results + + +def save_results( + params: AttributeDict, + test_set_name: str, + results_dict: Dict[str, List[Tuple[str, List[str], List[str]]]], +): + test_set_wers = dict() + for key, results in results_dict.items(): + recog_path = ( + params.res_dir / f"recogs-{test_set_name}-{key}-{params.suffix}.txt" + ) + results = sorted(results) + store_transcripts(filename=recog_path, texts=results) + logging.info(f"The transcripts are stored in {recog_path}") + + # The following prints out WERs, per-word error statistics and aligned + # ref/hyp pairs. + errs_filename = ( + params.res_dir / f"errs-{test_set_name}-{key}-{params.suffix}.txt" + ) + with open(errs_filename, "w") as f: + wer = write_error_stats( + f, f"{test_set_name}-{key}", results, enable_log=True + ) + test_set_wers[key] = wer + + logging.info("Wrote detailed error stats to {}".format(errs_filename)) + + test_set_wers = sorted(test_set_wers.items(), key=lambda x: x[1]) + errs_info = ( + params.res_dir / f"wer-summary-{test_set_name}-{key}-{params.suffix}.txt" + ) + with open(errs_info, "w") as f: + print("settings\tWER", file=f) + for key, val in test_set_wers: + print("{}\t{}".format(key, val), file=f) + + s = "\nFor {}, WER of different settings are:\n".format(test_set_name) + note = "\tbest for {}".format(test_set_name) + for key, val in test_set_wers: + s += "{}\t{}{}\n".format(key, val, note) + note = "" + logging.info(s) + + +@torch.no_grad() +def main(): + parser = get_parser() + LibriSpeechAsrDataModule.add_arguments(parser) + LmScorer.add_arguments(parser) + args = parser.parse_args() + args.exp_dir = Path(args.exp_dir) + + params = get_params() + params.update(vars(args)) + + assert params.decoding_method in ( + "greedy_search", + "beam_search", + "fast_beam_search", + "fast_beam_search_nbest", + "fast_beam_search_nbest_LG", + "fast_beam_search_nbest_oracle", + "modified_beam_search", + "modified_beam_search_LODR", + "modified_beam_search_lm_shallow_fusion", + "modified_beam_search_lm_rescore", + "modified_beam_search_lm_rescore_LODR", + ) + params.res_dir = params.exp_dir / params.decoding_method + + if os.path.exists(params.context_file): + params.has_contexts = True + else: + params.has_contexts = False + + if params.iter > 0: + params.suffix = f"iter-{params.iter}-avg-{params.avg}" + else: + params.suffix = f"epoch-{params.epoch}-avg-{params.avg}" + + if "fast_beam_search" in params.decoding_method: + params.suffix += f"-beam-{params.beam}" + params.suffix += f"-max-contexts-{params.max_contexts}" + params.suffix += f"-max-states-{params.max_states}" + if "nbest" in params.decoding_method: + params.suffix += f"-nbest-scale-{params.nbest_scale}" + params.suffix += f"-num-paths-{params.num_paths}" + if "LG" in params.decoding_method: + params.suffix += f"-ngram-lm-scale-{params.ngram_lm_scale}" + elif "beam_search" in params.decoding_method: + params.suffix += f"-{params.decoding_method}-beam-size-{params.beam_size}" + if params.decoding_method in ( + "modified_beam_search", + "modified_beam_search_LODR", + ): + if params.has_contexts: + params.suffix += f"-context-score-{params.context_score}" + else: + params.suffix += f"-context-{params.context_size}" + params.suffix += f"-max-sym-per-frame-{params.max_sym_per_frame}" + + if params.use_shallow_fusion: + params.suffix += f"-{params.lm_type}-lm-scale-{params.lm_scale}" + + if "LODR" in params.decoding_method: + params.suffix += ( + f"-LODR-{params.tokens_ngram}gram-scale-{params.ngram_lm_scale}" + ) + + if params.use_averaged_model: + params.suffix += "-use-averaged-model" + + setup_logger(f"{params.res_dir}/log-decode-{params.suffix}") + logging.info("Decoding started") + + device = torch.device("cpu") + if torch.cuda.is_available(): + device = torch.device("cuda", 0) + + logging.info(f"Device: {device}") + + sp = spm.SentencePieceProcessor() + sp.load(params.bpe_model) + + # and are defined in local/train_bpe_model.py + params.blank_id = sp.piece_to_id("") + params.unk_id = sp.piece_to_id("") + params.vocab_size = sp.get_piece_size() + + logging.info(params) + + logging.info("About to create model") + model = get_model(params) + + if not params.use_averaged_model: + if params.iter > 0: + filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[ + : params.avg + ] + if len(filenames) == 0: + raise ValueError( + f"No checkpoints found for" + f" --iter {params.iter}, --avg {params.avg}" + ) + elif len(filenames) < params.avg: + raise ValueError( + f"Not enough checkpoints ({len(filenames)}) found for" + f" --iter {params.iter}, --avg {params.avg}" + ) + logging.info(f"averaging {filenames}") + model.to(device) + model.load_state_dict(average_checkpoints(filenames, device=device)) + elif params.avg == 1: + load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model) + else: + start = params.epoch - params.avg + 1 + filenames = [] + for i in range(start, params.epoch + 1): + if i >= 1: + filenames.append(f"{params.exp_dir}/epoch-{i}.pt") + logging.info(f"averaging {filenames}") + model.to(device) + model.load_state_dict(average_checkpoints(filenames, device=device)) + else: + if params.iter > 0: + filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[ + : params.avg + 1 + ] + if len(filenames) == 0: + raise ValueError( + f"No checkpoints found for" + f" --iter {params.iter}, --avg {params.avg}" + ) + elif len(filenames) < params.avg + 1: + raise ValueError( + f"Not enough checkpoints ({len(filenames)}) found for" + f" --iter {params.iter}, --avg {params.avg}" + ) + filename_start = filenames[-1] + filename_end = filenames[0] + logging.info( + "Calculating the averaged model over iteration checkpoints" + f" from {filename_start} (excluded) to {filename_end}" + ) + model.to(device) + model.load_state_dict( + average_checkpoints_with_averaged_model( + filename_start=filename_start, + filename_end=filename_end, + device=device, + ) + ) + else: + assert params.avg > 0, params.avg + start = params.epoch - params.avg + assert start >= 1, start + filename_start = f"{params.exp_dir}/epoch-{start}.pt" + filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt" + logging.info( + f"Calculating the averaged model over epoch range from " + f"{start} (excluded) to {params.epoch}" + ) + model.to(device) + model.load_state_dict( + average_checkpoints_with_averaged_model( + filename_start=filename_start, + filename_end=filename_end, + device=device, + ) + ) + + model.to(device) + model.eval() + + # only load the neural network LM if required + if params.use_shallow_fusion or params.decoding_method in ( + "modified_beam_search_lm_rescore", + "modified_beam_search_lm_rescore_LODR", + "modified_beam_search_lm_shallow_fusion", + "modified_beam_search_LODR", + ): + LM = LmScorer( + lm_type=params.lm_type, + params=params, + device=device, + lm_scale=params.lm_scale, + ) + LM.to(device) + LM.eval() + else: + LM = None + + # only load N-gram LM when needed + if params.decoding_method == "modified_beam_search_lm_rescore_LODR": + try: + import kenlm + except ImportError: + print("Please install kenlm first. You can use") + print(" pip install https://github.com/kpu/kenlm/archive/master.zip") + print("to install it") + import sys + + sys.exit(-1) + ngram_file_name = str(params.lang_dir / f"{params.tokens_ngram}gram.arpa") + logging.info(f"lm filename: {ngram_file_name}") + ngram_lm = kenlm.Model(ngram_file_name) + ngram_lm_scale = None # use a list to search + + elif params.decoding_method == "modified_beam_search_LODR": + lm_filename = f"{params.tokens_ngram}gram.fst.txt" + logging.info(f"Loading token level lm: {lm_filename}") + ngram_lm = NgramLm( + str(params.lang_dir / lm_filename), + backoff_id=params.backoff_id, + is_binary=False, + ) + logging.info(f"num states: {ngram_lm.lm.num_states}") + ngram_lm_scale = params.ngram_lm_scale + else: + ngram_lm = None + ngram_lm_scale = None + + if "fast_beam_search" in params.decoding_method: + if params.decoding_method == "fast_beam_search_nbest_LG": + lexicon = Lexicon(params.lang_dir) + word_table = lexicon.word_table + lg_filename = params.lang_dir / "LG.pt" + logging.info(f"Loading {lg_filename}") + decoding_graph = k2.Fsa.from_dict( + torch.load(lg_filename, map_location=device) + ) + decoding_graph.scores *= params.ngram_lm_scale + else: + word_table = None + decoding_graph = k2.trivial_graph(params.vocab_size - 1, device=device) + else: + decoding_graph = None + word_table = None + + if "modified_beam_search" in params.decoding_method: + if os.path.exists(params.context_file): + contexts = [] + for line in open(params.context_file).readlines(): + contexts.append((sp.encode(line.strip()), 0.0)) + context_graph = ContextGraph(params.context_score) + context_graph.build(contexts) + else: + context_graph = None + else: + context_graph = None + + num_param = sum([p.numel() for p in model.parameters()]) + logging.info(f"Number of model parameters: {num_param}") + + # we need cut ids to display recognition results. + args.return_cuts = True + librispeech = LibriSpeechAsrDataModule(args) + + dev_clean_cuts = librispeech.dev_clean_cuts() + dev_other_cuts = librispeech.dev_other_cuts() + + dev_clean_dl = librispeech.test_dataloaders( + dev_clean_cuts, + do_normalize=params.do_normalize, + ) + dev_other_dl = librispeech.test_dataloaders( + dev_other_cuts, + do_normalize=params.do_normalize, + ) + + test_clean_cuts = librispeech.test_clean_cuts() + test_other_cuts = librispeech.test_other_cuts() + + test_clean_dl = librispeech.test_dataloaders( + test_clean_cuts, + do_normalize=params.do_normalize, + ) + test_other_dl = librispeech.test_dataloaders( + test_other_cuts, + do_normalize=params.do_normalize, + ) + + test_sets = ["dev-clean", "dev-other", "test-clean", "test-other"] + test_dl = [dev_clean_dl, dev_other_dl, test_clean_dl, test_other_dl] + + for test_set, test_dl in zip(test_sets, test_dl): + results_dict = decode_dataset( + dl=test_dl, + params=params, + model=model, + sp=sp, + word_table=word_table, + decoding_graph=decoding_graph, + context_graph=context_graph, + LM=LM, + ngram_lm=ngram_lm, + ngram_lm_scale=ngram_lm_scale, + ) + + save_results( + params=params, + test_set_name=test_set, + results_dict=results_dict, + ) + + logging.info("Done!") + + +if __name__ == "__main__": + main() diff --git a/egs/librispeech/SSL/hubert/decoder.py b/egs/librispeech/SSL/hubert/decoder.py new file mode 120000 index 0000000000..a2138e5da4 --- /dev/null +++ b/egs/librispeech/SSL/hubert/decoder.py @@ -0,0 +1 @@ +../../ASR/zipformer/decoder.py \ No newline at end of file diff --git a/egs/librispeech/SSL/hubert/finetune.py b/egs/librispeech/SSL/hubert/finetune.py new file mode 100644 index 0000000000..201847aed3 --- /dev/null +++ b/egs/librispeech/SSL/hubert/finetune.py @@ -0,0 +1,1254 @@ +#!/usr/bin/env python3 +# Copyright 2021-2024 Xiaomi Corp. (authors: Fangjun Kuang, +# Wei Kang, +# Mingshuang Luo, +# Zengwei Yao, +# Yifan Yang, +# Daniel Povey) +# Copyright 2024 Shanghai Jiao Tong University (authors: Jianheng Zhuo) +# +# 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="0,1,2,3,4,5,6,7" + +# For HuBERT model finetuning: +./hubert/finetune.py \ + --world-size 8 \ + --num-epochs 200 \ + --start-epoch 1 \ + --use-fp16 1 \ + --exp-dir hubert/exp \ + --full-libri 0 \ + --max-duration 200 + +It supports finetuning with: + - transducer loss (default), with `--use-transducer True --use-ctc False` + - ctc loss (not recommended), with `--use-transducer False --use-ctc True` + - transducer loss & ctc loss, with `--use-transducer True --use-ctc True` +""" + + +import argparse +import copy +import logging +import warnings +from pathlib import Path +from shutil import copyfile +from typing import Any, Dict, Optional, Tuple, Union + +import k2 +import optim +import sentencepiece as spm +import torch +import torch.multiprocessing as mp +import torch.nn as nn +from asr_datamodule import LibriSpeechAsrDataModule +from decoder import Decoder +from hubert import HubertModel, add_hubert_arguments +from joiner import Joiner +from lhotse.cut import Cut +from lhotse.dataset.sampling.base import CutSampler +from lhotse.utils import fix_random_seed +from model import AsrModel +from optim import Eden, ScaledAdam +from 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.hooks import register_inf_check_hooks +from icefall.utils import ( + AttributeDict, + MetricsTracker, + get_parameter_groups_with_lrs, + setup_logger, + str2bool, +) + +LRSchedulerType = Union[torch.optim.lr_scheduler._LRScheduler, optim.LRScheduler] + + +def get_adjusted_batch_count(params: AttributeDict) -> float: + # returns the number of batches we would have used so far if we had used the reference + # duration. This is for purposes of set_batch_count(). + return ( + params.batch_idx_train + * params.accum_grad + * (params.max_duration * params.world_size) + / params.ref_duration + ) + + +def set_batch_count(model: Union[nn.Module, DDP], batch_count: float) -> None: + if isinstance(model, DDP): + # get underlying nn.Module + model = model.module + for name, module in model.named_modules(): + if hasattr(module, "batch_count"): + module.batch_count = batch_count + if hasattr(module, "name"): + module.name = name + + +def add_model_arguments(parser: argparse.ArgumentParser): + parser.add_argument( + "--decoder-dim", + type=int, + default=512, + help="Embedding dimension in the decoder model.", + ) + + parser.add_argument( + "--joiner-dim", + type=int, + default=512, + help="""Dimension used in the joiner model. + Outputs from the encoder and decoder model are projected + to this dimension before adding. + """, + ) + + parser.add_argument( + "--use-transducer", + type=str2bool, + default=True, + help="If True, use Transducer head.", + ) + + parser.add_argument( + "--use-ctc", + type=str2bool, + default=False, + help="If True, use CTC head.", + ) + + +def get_parser(): + parser = argparse.ArgumentParser( + formatter_class=argparse.ArgumentDefaultsHelpFormatter + ) + + parser.add_argument( + "--world-size", + type=int, + default=1, + help="Number of GPUs for DDP training.", + ) + + parser.add_argument( + "--master-port", + type=int, + default=12354, + help="Master port to use for DDP training.", + ) + + parser.add_argument( + "--tensorboard", + type=str2bool, + default=True, + help="Should various information be logged in tensorboard.", + ) + + parser.add_argument( + "--num-epochs", + type=int, + default=222, + 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="hubert/exp", + help="""The experiment dir. + It specifies the directory where all training related + files, e.g., checkpoints, log, etc, are saved + """, + ) + + parser.add_argument( + "--pretrained-dir", + type=str, + help="""The pretrained model dir. + It specifies the directory where the pretrained checkpoint is saved.""", + ) + + parser.add_argument( + "--bpe-model", + type=str, + default="data/lang_bpe_500/bpe.model", + help="Path to the BPE model", + ) + + parser.add_argument( + "--base-lr", type=float, default=0.001, help="The base learning rate." + ) + + parser.add_argument( + "--lr-batches", + type=float, + default=100000, + 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=100, + help="""Number of epochs that affects how rapidly the learning rate decreases. + """, + ) + + parser.add_argument( + "--ref-duration", + type=float, + default=600, + help="Reference batch duration for purposes of adjusting batch counts for setting various " + "schedules inside the model", + ) + + parser.add_argument( + "--context-size", + type=int, + default=2, + help="The context size in the decoder. 1 means bigram; " "2 means tri-gram", + ) + + parser.add_argument( + "--prune-range", + type=int, + default=5, + help="The prune range for rnnt loss, it means how many symbols(context)" + "we are using to compute the loss", + ) + + parser.add_argument( + "--lm-scale", + type=float, + default=0.25, + help="The scale to smooth the loss with lm " + "(output of prediction network) part.", + ) + + parser.add_argument( + "--am-scale", + type=float, + default=0.0, + help="The scale to smooth the loss with am (output of encoder network)" "part.", + ) + + parser.add_argument( + "--simple-loss-scale", + type=float, + default=0.5, + help="To get pruning ranges, we will calculate a simple version" + "loss(joiner is just addition), this simple loss also uses for" + "training (as a regularization item). We will scale the simple loss" + "with this parameter before adding to the final loss.", + ) + + parser.add_argument( + "--ctc-loss-scale", + type=float, + default=0.2, + help="Scale for CTC loss.", + ) + + parser.add_argument( + "--seed", + type=int, + default=42, + help="The seed for random generators intended for reproducibility", + ) + + parser.add_argument( + "--print-diagnostics", + type=str2bool, + default=False, + help="Accumulate stats on activations, print them and exit.", + ) + + parser.add_argument( + "--sanity-check", + type=str2bool, + default=False, + help="Check if any of the batches in epoch 1 would cause OOM.", + ) + + parser.add_argument( + "--inf-check", + type=str2bool, + default=False, + help="Add hooks to check for infinite module outputs and gradients.", + ) + + parser.add_argument( + "--save-every-n", + type=int, + default=100000, + help="""Save checkpoint after processing this number of batches" + periodically. We save checkpoint to exp-dir/ whenever + params.batch_idx_train % save_every_n == 0. The checkpoint filename + has the form: f'exp-dir/checkpoint-{params.batch_idx_train}.pt' + Note: It also saves checkpoint to `exp-dir/epoch-xxx.pt` at the + end of each epoch where `xxx` is the epoch number counting from 1. + """, + ) + + parser.add_argument( + "--keep-last-k", + type=int, + default=30, + help="""Only keep this number of checkpoints on disk. + For instance, if it is 3, there are only 3 checkpoints + in the exp-dir with filenames `checkpoint-xxx.pt`. + It does not affect checkpoints with name `epoch-xxx.pt`. + """, + ) + + parser.add_argument( + "--average-period", + type=int, + default=200, + help="""Update the averaged model, namely `model_avg`, after processing + this number of batches. `model_avg` is a separate version of model, + in which each floating-point parameter is the average of all the + parameters from the start of training. Each time we take the average, + we do: `model_avg = model * (average_period / batch_idx_train) + + model_avg * ((batch_idx_train - average_period) / batch_idx_train)`. + """, + ) + + parser.add_argument( + "--accum-grad", + type=int, + default=1, + help="""update gradient when batch_idx_train % accum_grad == 0. + """, + ) + + parser.add_argument( + "--use-fp16", + type=str2bool, + default=False, + help="Whether to use half precision training.", + ) + + add_hubert_arguments(parser) + 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 updates happen to the model so far across + epochs. + + - sub_batch_idx_train: It contains number of batch 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 + + - warm_step: The warmup period that dictates the decay of the + scale on "simple" (un-pruned) loss. + """ + params = AttributeDict( + { + "best_train_loss": float("inf"), + "best_valid_loss": float("inf"), + "best_train_epoch": -1, + "best_valid_epoch": -1, + "batch_idx_train": 0, + "sub_batch_idx_train": 0, + "log_interval": 50, + "reset_interval": 200, + "valid_interval": 3000, # For the 100h subset, use 800 + # parameters for pruned RNN-T loss + "warm_step": 2000, + "env_info": get_env_info(), + } + ) + + return params + + +def _to_int_tuple(s: str): + return tuple(map(int, s.split(","))) + + +def get_encoder_model(params: AttributeDict) -> nn.Module: + if hasattr(params, "pretrained_dir"): + logging.info(f"Loading {params.pretrained_dir}") + pretrained = torch.load(params.pretrained_dir) + encoder = HubertModel(params) + encoder.load_state_dict(pretrained["model"]) + else: + encoder = HubertModel(params) + 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_embed_dim, + decoder_dim=params.decoder_dim, + joiner_dim=params.joiner_dim, + vocab_size=params.vocab_size, + ) + return joiner + + +def get_model(params: AttributeDict) -> nn.Module: + assert params.use_transducer or params.use_ctc, ( + f"At least one of them should be True, " + f"but got params.use_transducer={params.use_transducer}, " + f"params.use_ctc={params.use_ctc}" + ) + + encoder = get_encoder_model(params) + + if params.use_transducer: + decoder = get_decoder_model(params) + joiner = get_joiner_model(params) + else: + decoder = None + joiner = None + + model = AsrModel( + encoder=encoder, + decoder=decoder, + joiner=joiner, + encoder_dim=params.encoder_embed_dim, + decoder_dim=params.decoder_dim, + vocab_size=params.vocab_size, + use_transducer=params.use_transducer, + use_ctc=params.use_ctc, + ) + return model + + +def load_checkpoint_if_available( + params: AttributeDict, + model: nn.Module, + model_avg: nn.Module = None, + optimizer: Optional[torch.optim.Optimizer] = None, + scheduler: Optional[LRSchedulerType] = None, +) -> Optional[Dict[str, Any]]: + """Load checkpoint from file. + + If params.start_batch is positive, it will load the checkpoint from + `params.exp_dir/checkpoint-{params.start_batch}.pt`. Otherwise, if + params.start_epoch is larger than 1, it will load the checkpoint from + `params.start_epoch - 1`. + + Apart from loading state dict for `model` and `optimizer` it also updates + `best_train_epoch`, `best_train_loss`, `best_valid_epoch`, + and `best_valid_loss` in `params`. + + Args: + params: + The return value of :func:`get_params`. + model: + The training model. + model_avg: + The stored model averaged from the start of training. + optimizer: + The optimizer that we are using. + scheduler: + The scheduler that we are using. + Returns: + Return a dict containing previously saved training info. + """ + if params.start_batch > 0: + filename = params.exp_dir / f"checkpoint-{params.start_batch}.pt" + elif params.start_epoch > 1: + filename = params.exp_dir / f"epoch-{params.start_epoch-1}.pt" + else: + return None + + assert filename.is_file(), f"{filename} does not exist!" + + saved_params = load_checkpoint( + filename, + model=model, + model_avg=model_avg, + optimizer=optimizer, + scheduler=scheduler, + ) + + keys = [ + "best_train_epoch", + "best_valid_epoch", + "batch_idx_train", + "best_train_loss", + "best_valid_loss", + ] + for k in keys: + params[k] = saved_params[k] + + if params.start_batch > 0: + if "cur_epoch" in saved_params: + params["start_epoch"] = saved_params["cur_epoch"] + + return saved_params + + +def save_checkpoint( + params: AttributeDict, + model: Union[nn.Module, DDP], + model_avg: Optional[nn.Module] = None, + optimizer: Optional[torch.optim.Optimizer] = None, + scheduler: Optional[LRSchedulerType] = None, + sampler: Optional[CutSampler] = None, + scaler: Optional[GradScaler] = None, + rank: int = 0, +) -> None: + """Save model, optimizer, scheduler and training stats to file. + + Args: + params: + It is returned by :func:`get_params`. + model: + The training model. + model_avg: + The stored model averaged from the start of training. + optimizer: + The optimizer used in the training. + sampler: + The sampler for the training dataset. + scaler: + The scaler used for mix precision training. + """ + if rank != 0: + return + filename = params.exp_dir / f"epoch-{params.cur_epoch}.pt" + save_checkpoint_impl( + filename=filename, + model=model, + model_avg=model_avg, + params=params, + optimizer=optimizer, + scheduler=scheduler, + sampler=sampler, + scaler=scaler, + rank=rank, + ) + + if params.best_train_epoch == params.cur_epoch: + best_train_filename = params.exp_dir / "best-train-loss.pt" + copyfile(src=filename, dst=best_train_filename) + + if params.best_valid_epoch == params.cur_epoch: + best_valid_filename = params.exp_dir / "best-valid-loss.pt" + copyfile(src=filename, dst=best_valid_filename) + + +def compute_loss( + params: AttributeDict, + model: Union[nn.Module, DDP], + sp: spm.SentencePieceProcessor, + batch: dict, + is_training: bool, +) -> Tuple[Tensor, MetricsTracker]: + """ + Compute loss given the model and its inputs. + + Args: + params: + Parameters for training. See :func:`get_params`. + model: + The model for training. It is an instance of Zipformer in our case. + batch: + A batch of data. See `dataset.HubertAsrDataset()` + 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 + audio = batch["audio"].to(device) + padding_mask = batch["padding_mask"].to(device) + + batch_idx_train = params.batch_idx_train + warm_step = params.warm_step + + texts = batch["supervisions"]["text"] + y = sp.encode(texts, out_type=int) + y = k2.RaggedTensor(y) + + with torch.set_grad_enabled(is_training): + simple_loss, pruned_loss, ctc_loss, num_frames = model( + x=audio, + padding_mask=padding_mask, + y=y, + prune_range=params.prune_range, + am_scale=params.am_scale, + lm_scale=params.lm_scale, + ) + + loss = 0.0 + + if params.use_transducer: + s = params.simple_loss_scale + # take down the scale on the simple loss from 1.0 at the start + # to params.simple_loss scale by warm_step. + simple_loss_scale = ( + s + if batch_idx_train >= warm_step + else 1.0 - (batch_idx_train / warm_step) * (1.0 - s) + ) + pruned_loss_scale = ( + 1.0 + if batch_idx_train >= warm_step + else 0.1 + 0.9 * (batch_idx_train / warm_step) + ) + loss += simple_loss_scale * simple_loss + pruned_loss_scale * pruned_loss + + if params.use_ctc: + loss += params.ctc_loss_scale * ctc_loss + + assert loss.requires_grad == is_training + + info = MetricsTracker() + with warnings.catch_warnings(): + warnings.simplefilter("ignore") + info["frames"] = num_frames.sum().item() + + # Note: We use reduction=sum while computing the loss. + info["loss"] = loss.detach().cpu().item() + if params.use_transducer: + info["simple_loss"] = simple_loss.detach().cpu().item() + info["pruned_loss"] = pruned_loss.detach().cpu().item() + if params.use_ctc: + info["ctc_loss"] = ctc_loss.detach().cpu().item() + + return loss, info + + +def compute_validation_loss( + params: AttributeDict, + model: Union[nn.Module, DDP], + sp: spm.SentencePieceProcessor, + valid_dl: torch.utils.data.DataLoader, + world_size: int = 1, +) -> MetricsTracker: + """Run the validation process.""" + model.eval() + + tot_loss = MetricsTracker() + + for batch_idx, batch in enumerate(valid_dl): + loss, loss_info = compute_loss( + params=params, + model=model, + sp=sp, + batch=batch, + is_training=False, + ) + assert loss.requires_grad is False + tot_loss = tot_loss + loss_info + + if world_size > 1: + tot_loss.reduce(loss.device) + + loss_value = tot_loss["loss"] / tot_loss["frames"] + if loss_value < params.best_valid_loss: + params.best_valid_epoch = params.cur_epoch + params.best_valid_loss = loss_value + + return tot_loss + + +def train_one_epoch( + params: AttributeDict, + model: Union[nn.Module, DDP], + optimizer: torch.optim.Optimizer, + scheduler: LRSchedulerType, + sp: spm.SentencePieceProcessor, + train_dl: torch.utils.data.DataLoader, + valid_dl: torch.utils.data.DataLoader, + scaler: GradScaler, + model_avg: Optional[nn.Module] = None, + tb_writer: Optional[SummaryWriter] = None, + world_size: int = 1, + rank: int = 0, +) -> None: + """Train the model for one epoch. + + The training loss from the mean of all frames is saved in + `params.train_loss`. It runs the validation process every + `params.valid_interval` batches. + + Args: + params: + It is returned by :func:`get_params`. + model: + The model for training. + optimizer: + The optimizer we are using. + scheduler: + The learning rate scheduler, we call step() every step. + train_dl: + Dataloader for the training dataset. + valid_dl: + Dataloader for the validation dataset. + scaler: + The scaler used for mix precision training. + model_avg: + The stored model averaged from the start of training. + tb_writer: + Writer to write log messages to tensorboard. + world_size: + Number of nodes in DDP training. If it is 1, DDP is disabled. + rank: + The rank of the node in DDP training. If no DDP is used, it should + be set to 0. + """ + model.train() + + tot_loss = MetricsTracker() + + saved_bad_model = False + + def save_bad_model(suffix: str = ""): + save_checkpoint_impl( + filename=params.exp_dir / f"bad-model{suffix}-{rank}.pt", + model=model, + model_avg=model_avg, + params=params, + optimizer=optimizer, + scheduler=scheduler, + sampler=train_dl.sampler, + scaler=scaler, + rank=0, + ) + + for sub_batch_idx, batch in enumerate(train_dl): + params.sub_batch_idx_train += 1 + batch_idx = sub_batch_idx // params.accum_grad + + if batch_idx % 10 == 0: + set_batch_count(model, get_adjusted_batch_count(params)) + + batch_size = len(batch["supervisions"]["text"]) + + try: + with torch.cuda.amp.autocast(enabled=params.use_fp16): + loss, loss_info = compute_loss( + params=params, + model=model, + sp=sp, + batch=batch, + is_training=True, + ) + # summary stats + tot_loss = (tot_loss * (1 - 1 / params.reset_interval)) + loss_info + + # NOTE: We use reduction==sum and loss is computed over utterances + # in the batch and there is no normalization to it so far. + scaler.scale(loss / params.accum_grad).backward() + + if sub_batch_idx % params.accum_grad == params.accum_grad - 1: + params.batch_idx_train += 1 + scheduler.step_batch(params.batch_idx_train) + + scaler.step(optimizer) + scaler.update() + optimizer.zero_grad() + else: + continue + + except: # noqa + save_bad_model() + display_and_save_batch(batch, params=params, sp=sp) + raise + + if params.print_diagnostics and batch_idx == 5: + return + + if ( + rank == 0 + and params.batch_idx_train > 0 + and params.batch_idx_train % params.average_period == 0 + ): + update_averaged_model( + params=params, + model_cur=model, + model_avg=model_avg, + ) + + if ( + params.batch_idx_train > 0 + and params.batch_idx_train % params.save_every_n == 0 + ): + save_checkpoint_with_global_batch_idx( + out_dir=params.exp_dir, + global_batch_idx=params.batch_idx_train, + model=model, + model_avg=model_avg, + params=params, + optimizer=optimizer, + scheduler=scheduler, + sampler=train_dl.sampler, + scaler=scaler, + rank=rank, + ) + remove_checkpoints( + out_dir=params.exp_dir, + topk=params.keep_last_k, + rank=rank, + ) + + if batch_idx % 100 == 0 and params.use_fp16: + # If the grad scale was less than 1, try increasing it. The _growth_interval + # of the grad scaler is configurable, but we can't configure it to have different + # behavior depending on the current grad scale. + cur_grad_scale = scaler._scale.item() + + if cur_grad_scale < 8.0 or (cur_grad_scale < 32.0 and batch_idx % 400 == 0): + scaler.update(cur_grad_scale * 2.0) + if cur_grad_scale < 0.01: + if not saved_bad_model: + save_bad_model(suffix="-first-warning") + saved_bad_model = True + logging.warning(f"Grad scale is small: {cur_grad_scale}") + if cur_grad_scale < 1.0e-05: + save_bad_model() + raise RuntimeError( + f"grad_scale is too small, exiting: {cur_grad_scale}" + ) + + if batch_idx % params.log_interval == 0: + cur_lr = max(scheduler.get_last_lr()) + cur_grad_scale = scaler._scale.item() if params.use_fp16 else 1.0 + + logging.info( + f"Epoch {params.cur_epoch}, " + f"batch {batch_idx}, loss[{loss_info}], " + f"tot_loss[{tot_loss}], batch size: {batch_size}, " + f"lr: {cur_lr:.2e}, " + + (f"grad_scale: {scaler._scale.item()}" if params.use_fp16 else "") + ) + + if tb_writer is not None: + tb_writer.add_scalar( + "train/learning_rate", cur_lr, params.batch_idx_train + ) + + loss_info.write_summary( + tb_writer, "train/current_", params.batch_idx_train + ) + tot_loss.write_summary(tb_writer, "train/tot_", params.batch_idx_train) + if params.use_fp16: + tb_writer.add_scalar( + "train/grad_scale", + cur_grad_scale, + params.batch_idx_train, + ) + + if batch_idx % params.valid_interval == 0 and not params.print_diagnostics: + logging.info("Computing validation loss") + valid_info = compute_validation_loss( + params=params, + model=model, + sp=sp, + valid_dl=valid_dl, + world_size=world_size, + ) + model.train() + logging.info(f"Epoch {params.cur_epoch}, validation: {valid_info}") + logging.info( + f"Maximum memory allocated so far is {torch.cuda.max_memory_allocated()//1000000}MB" + ) + if tb_writer is not None: + valid_info.write_summary( + tb_writer, "train/valid_", params.batch_idx_train + ) + + if batch_idx % params.accum_grad != params.accum_grad - 1: + optimizer.zero_grad() + loss_value = tot_loss["loss"] / tot_loss["frames"] + params.train_loss = loss_value + if params.train_loss < params.best_train_loss: + params.best_train_epoch = params.cur_epoch + params.best_train_loss = params.train_loss + + +def run(rank, world_size, args): + """ + Args: + rank: + It is a value between 0 and `world_size-1`, which is + passed automatically by `mp.spawn()` in :func:`main`. + The node with rank 0 is responsible for saving checkpoint. + world_size: + Number of GPUs for DDP training. + args: + The return value of get_parser().parse_args() + """ + params = get_params() + params.update(vars(args)) + + fix_random_seed(params.seed) + if world_size > 1: + setup_dist(rank, world_size, params.master_port) + + setup_logger(f"{params.exp_dir}/log/log-train") + logging.info("Training started") + + if args.tensorboard and rank == 0: + tb_writer = SummaryWriter(log_dir=f"{params.exp_dir}/tensorboard") + else: + tb_writer = None + + device = torch.device("cpu") + if torch.cuda.is_available(): + device = torch.device("cuda", rank) + logging.info(f"Device: {device}") + + sp = spm.SentencePieceProcessor() + sp.load(params.bpe_model) + + # is defined in local/train_bpe_model.py + params.blank_id = sp.piece_to_id("") + params.vocab_size = sp.get_piece_size() + + if not params.use_transducer: + params.ctc_loss_scale = 1.0 + + logging.info(params) + + logging.info("About to create model") + model = get_model(params) + + num_param = sum([p.numel() for p in model.parameters()]) + logging.info(f"Number of model parameters: {num_param}") + + assert params.save_every_n >= params.average_period + model_avg: Optional[nn.Module] = None + if rank == 0: + # model_avg is only used with rank 0 + model_avg = copy.deepcopy(model).to(torch.float64) + + assert params.start_epoch > 0, params.start_epoch + checkpoints = load_checkpoint_if_available( + params=params, model=model, model_avg=model_avg + ) + + model.to(device) + if world_size > 1: + logging.info("Using DDP") + model = DDP(model, device_ids=[rank], find_unused_parameters=True) + + optimizer = ScaledAdam( + get_parameter_groups_with_lrs(model, lr=params.base_lr, include_names=True), + lr=params.base_lr, # should have no effect + clipping_scale=2.0, + ) + + scheduler = Eden(optimizer, params.lr_batches, params.lr_epochs) + + if checkpoints and "optimizer" in checkpoints: + logging.info("Loading optimizer state dict") + optimizer.load_state_dict(checkpoints["optimizer"]) + + if ( + checkpoints + and "scheduler" in checkpoints + and checkpoints["scheduler"] is not None + ): + logging.info("Loading scheduler state dict") + scheduler.load_state_dict(checkpoints["scheduler"]) + + if params.print_diagnostics: + opts = diagnostics.TensorDiagnosticOptions( + 512 + ) # allow 4 megabytes per sub-module + diagnostic = diagnostics.attach_diagnostics(model, opts) + + if params.inf_check: + register_inf_check_hooks(model) + + librispeech = LibriSpeechAsrDataModule(args) + + train_cuts = ( + librispeech.train_all_shuf_cuts() + if params.full_libri + else librispeech.train_clean_100_cuts() + ) + + def remove_short_and_long_utt(c: Cut): + # Keep only utterances with duration between 1 second and 20 seconds + # + # Caution: There is a reason to select 20.0 here. Please see + # ../local/display_manifest_statistics.py + # + # You should use ../local/display_manifest_statistics.py to get + # an utterance duration distribution for your dataset to select + # the threshold + if c.duration < 1.0 or c.duration > 20.0: + # logging.warning( + # f"Exclude cut with ID {c.id} from training. Duration: {c.duration}" + # ) + return False + + return True + + train_cuts = train_cuts.filter(remove_short_and_long_utt) + + if params.start_batch > 0 and checkpoints and "sampler" in checkpoints: + # We only load the sampler's state dict when it loads a checkpoint + # saved in the middle of an epoch + sampler_state_dict = checkpoints["sampler"] + else: + sampler_state_dict = None + + train_dl = librispeech.train_dataloaders( + train_cuts, + do_normalize=params.do_normalize, + 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, + do_normalize=params.do_normalize, + ) + + if params.sanity_check and not params.print_diagnostics: + scan_pessimistic_batches_for_oom( + model=model, + train_dl=train_dl, + optimizer=optimizer, + sp=sp, + params=params, + ) + + scaler = GradScaler(enabled=params.use_fp16, init_scale=1.0) + if checkpoints and "grad_scaler" in checkpoints: + logging.info("Loading grad scaler state dict") + scaler.load_state_dict(checkpoints["grad_scaler"]) + + for epoch in range(params.start_epoch, params.num_epochs + 1): + scheduler.step_epoch(epoch - 1) + fix_random_seed(params.seed + epoch - 1) + train_dl.sampler.set_epoch(epoch - 1) + + if tb_writer is not None: + tb_writer.add_scalar("train/epoch", epoch, params.batch_idx_train) + + params.cur_epoch = epoch + + train_one_epoch( + params=params, + model=model, + model_avg=model_avg, + optimizer=optimizer, + scheduler=scheduler, + sp=sp, + train_dl=train_dl, + valid_dl=valid_dl, + scaler=scaler, + tb_writer=tb_writer, + world_size=world_size, + rank=rank, + ) + + if params.print_diagnostics: + diagnostic.print_diagnostics() + break + + save_checkpoint( + params=params, + model=model, + model_avg=model_avg, + optimizer=optimizer, + scheduler=scheduler, + sampler=train_dl.sampler, + scaler=scaler, + rank=rank, + ) + + logging.info("Done!") + + if world_size > 1: + torch.distributed.barrier() + cleanup_dist() + + +def display_and_save_batch( + batch: dict, + params: AttributeDict, + sp: spm.SentencePieceProcessor, +) -> None: + """Display the batch statistics and save the batch into disk. + + Args: + batch: + A batch of data. See `dataset.HubertAsrDataset()` + for the content in it. + params: + Parameters for training. See :func:`get_params`. + sp: + The BPE model. + """ + from lhotse.utils import uuid4 + + filename = f"{params.exp_dir}/batch-{uuid4()}.pt" + logging.info(f"Saving batch to {filename}") + torch.save(batch, filename) + + audio = batch["audio"] + logging.info(f"audio shape: {audio.shape}") + + y = sp.encode(batch["supervisions"]["text"], out_type=int) + num_tokens = sum(len(i) for i in y) + logging.info(f"num tokens: {num_tokens}") + + +def scan_pessimistic_batches_for_oom( + model: Union[nn.Module, DDP], + train_dl: torch.utils.data.DataLoader, + optimizer: torch.optim.Optimizer, + sp: spm.SentencePieceProcessor, + params: AttributeDict, +): + from lhotse.dataset import find_pessimistic_batches + + logging.info( + "Sanity check -- see if any of the batches in epoch 1 would cause OOM." + ) + batches, crit_values = find_pessimistic_batches(train_dl.sampler) + for criterion, cuts in batches.items(): + batch = train_dl.dataset[cuts] + try: + with torch.cuda.amp.autocast(enabled=params.use_fp16): + loss, _ = compute_loss( + params=params, + model=model, + sp=sp, + batch=batch, + is_training=True, + ) + loss.backward() + optimizer.zero_grad() + except Exception as e: + if "CUDA out of memory" in str(e): + logging.error( + "Your GPU ran out of memory with the current " + "max_duration setting. We recommend decreasing " + "max_duration and trying again.\n" + f"Failing criterion: {criterion} " + f"(={crit_values[criterion]}) ..." + ) + display_and_save_batch(batch, params=params, sp=sp) + raise + logging.info( + f"Maximum memory allocated so far is {torch.cuda.max_memory_allocated()//1000000}MB" + ) + + +def main(): + parser = get_parser() + LibriSpeechAsrDataModule.add_arguments(parser) + args = parser.parse_args() + args.exp_dir = Path(args.exp_dir) + + world_size = args.world_size + assert world_size >= 1 + if world_size > 1: + mp.spawn(run, args=(world_size, args), nprocs=world_size, join=True) + else: + run(rank=0, world_size=1, args=args) + + +torch.set_num_threads(1) +torch.set_num_interop_threads(1) + +if __name__ == "__main__": + main() diff --git a/egs/librispeech/SSL/hubert/finetune_ce.py b/egs/librispeech/SSL/hubert/finetune_ce.py new file mode 100644 index 0000000000..e69a5a8cd1 --- /dev/null +++ b/egs/librispeech/SSL/hubert/finetune_ce.py @@ -0,0 +1,1254 @@ +#!/usr/bin/env python3 +# Copyright 2021-2024 Xiaomi Corp. (authors: Fangjun Kuang, +# Wei Kang, +# Mingshuang Luo, +# Zengwei Yao, +# Yifan Yang, +# Daniel Povey) +# Copyright 2024 Shanghai Jiao Tong University (authors: Jianheng Zhuo) +# +# 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="0,1,2,3,4,5,6,7" + +# For HuBERT model finetuning: +./hubert/finetune.py \ + --world-size 8 \ + --num-epochs 200 \ + --start-epoch 1 \ + --use-fp16 1 \ + --exp-dir hubert/exp \ + --full-libri 0 \ + --max-duration 200 + +It supports finetuning with: + - transducer loss (default), with `--use-transducer True --use-ctc False` + - ctc loss (not recommended), with `--use-transducer False --use-ctc True` + - transducer loss & ctc loss, with `--use-transducer True --use-ctc True` +""" + + +import argparse +import copy +import logging +import warnings +from pathlib import Path +from shutil import copyfile +from typing import Any, Dict, Optional, Tuple, Union + +import k2 +import optim +import sentencepiece as spm +import torch +import torch.multiprocessing as mp +import torch.nn as nn +from asr_datamodule import LibriSpeechAsrDataModule +from decoder import Decoder +from hubert_ce import HubertModel, add_hubert_arguments +from joiner import Joiner +from lhotse.cut import Cut +from lhotse.dataset.sampling.base import CutSampler +from lhotse.utils import fix_random_seed +from model import AsrModel +from optim import Eden, ScaledAdam +from 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.hooks import register_inf_check_hooks +from icefall.utils import ( + AttributeDict, + MetricsTracker, + get_parameter_groups_with_lrs, + setup_logger, + str2bool, +) + +LRSchedulerType = Union[torch.optim.lr_scheduler._LRScheduler, optim.LRScheduler] + + +def get_adjusted_batch_count(params: AttributeDict) -> float: + # returns the number of batches we would have used so far if we had used the reference + # duration. This is for purposes of set_batch_count(). + return ( + params.batch_idx_train + * params.accum_grad + * (params.max_duration * params.world_size) + / params.ref_duration + ) + + +def set_batch_count(model: Union[nn.Module, DDP], batch_count: float) -> None: + if isinstance(model, DDP): + # get underlying nn.Module + model = model.module + for name, module in model.named_modules(): + if hasattr(module, "batch_count"): + module.batch_count = batch_count + if hasattr(module, "name"): + module.name = name + + +def add_model_arguments(parser: argparse.ArgumentParser): + parser.add_argument( + "--decoder-dim", + type=int, + default=512, + help="Embedding dimension in the decoder model.", + ) + + parser.add_argument( + "--joiner-dim", + type=int, + default=512, + help="""Dimension used in the joiner model. + Outputs from the encoder and decoder model are projected + to this dimension before adding. + """, + ) + + parser.add_argument( + "--use-transducer", + type=str2bool, + default=True, + help="If True, use Transducer head.", + ) + + parser.add_argument( + "--use-ctc", + type=str2bool, + default=False, + help="If True, use CTC head.", + ) + + +def get_parser(): + parser = argparse.ArgumentParser( + formatter_class=argparse.ArgumentDefaultsHelpFormatter + ) + + parser.add_argument( + "--world-size", + type=int, + default=1, + help="Number of GPUs for DDP training.", + ) + + parser.add_argument( + "--master-port", + type=int, + default=12354, + help="Master port to use for DDP training.", + ) + + parser.add_argument( + "--tensorboard", + type=str2bool, + default=True, + help="Should various information be logged in tensorboard.", + ) + + parser.add_argument( + "--num-epochs", + type=int, + default=222, + 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="hubert/exp", + help="""The experiment dir. + It specifies the directory where all training related + files, e.g., checkpoints, log, etc, are saved + """, + ) + + parser.add_argument( + "--pretrained-dir", + type=str, + help="""The pretrained model dir. + It specifies the directory where the pretrained checkpoint is saved.""", + ) + + parser.add_argument( + "--bpe-model", + type=str, + default="data/lang_bpe_500/bpe.model", + help="Path to the BPE model", + ) + + parser.add_argument( + "--base-lr", type=float, default=0.001, help="The base learning rate." + ) + + parser.add_argument( + "--lr-batches", + type=float, + default=100000, + 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=100, + help="""Number of epochs that affects how rapidly the learning rate decreases. + """, + ) + + parser.add_argument( + "--ref-duration", + type=float, + default=600, + help="Reference batch duration for purposes of adjusting batch counts for setting various " + "schedules inside the model", + ) + + parser.add_argument( + "--context-size", + type=int, + default=2, + help="The context size in the decoder. 1 means bigram; " "2 means tri-gram", + ) + + parser.add_argument( + "--prune-range", + type=int, + default=5, + help="The prune range for rnnt loss, it means how many symbols(context)" + "we are using to compute the loss", + ) + + parser.add_argument( + "--lm-scale", + type=float, + default=0.25, + help="The scale to smooth the loss with lm " + "(output of prediction network) part.", + ) + + parser.add_argument( + "--am-scale", + type=float, + default=0.0, + help="The scale to smooth the loss with am (output of encoder network)" "part.", + ) + + parser.add_argument( + "--simple-loss-scale", + type=float, + default=0.5, + help="To get pruning ranges, we will calculate a simple version" + "loss(joiner is just addition), this simple loss also uses for" + "training (as a regularization item). We will scale the simple loss" + "with this parameter before adding to the final loss.", + ) + + parser.add_argument( + "--ctc-loss-scale", + type=float, + default=0.2, + help="Scale for CTC loss.", + ) + + parser.add_argument( + "--seed", + type=int, + default=42, + help="The seed for random generators intended for reproducibility", + ) + + parser.add_argument( + "--print-diagnostics", + type=str2bool, + default=False, + help="Accumulate stats on activations, print them and exit.", + ) + + parser.add_argument( + "--sanity-check", + type=str2bool, + default=False, + help="Check if any of the batches in epoch 1 would cause OOM.", + ) + + parser.add_argument( + "--inf-check", + type=str2bool, + default=False, + help="Add hooks to check for infinite module outputs and gradients.", + ) + + parser.add_argument( + "--save-every-n", + type=int, + default=100000, + help="""Save checkpoint after processing this number of batches" + periodically. We save checkpoint to exp-dir/ whenever + params.batch_idx_train % save_every_n == 0. The checkpoint filename + has the form: f'exp-dir/checkpoint-{params.batch_idx_train}.pt' + Note: It also saves checkpoint to `exp-dir/epoch-xxx.pt` at the + end of each epoch where `xxx` is the epoch number counting from 1. + """, + ) + + parser.add_argument( + "--keep-last-k", + type=int, + default=30, + help="""Only keep this number of checkpoints on disk. + For instance, if it is 3, there are only 3 checkpoints + in the exp-dir with filenames `checkpoint-xxx.pt`. + It does not affect checkpoints with name `epoch-xxx.pt`. + """, + ) + + parser.add_argument( + "--average-period", + type=int, + default=200, + help="""Update the averaged model, namely `model_avg`, after processing + this number of batches. `model_avg` is a separate version of model, + in which each floating-point parameter is the average of all the + parameters from the start of training. Each time we take the average, + we do: `model_avg = model * (average_period / batch_idx_train) + + model_avg * ((batch_idx_train - average_period) / batch_idx_train)`. + """, + ) + + parser.add_argument( + "--accum-grad", + type=int, + default=1, + help="""update gradient when batch_idx_train % accum_grad == 0. + """, + ) + + parser.add_argument( + "--use-fp16", + type=str2bool, + default=False, + help="Whether to use half precision training.", + ) + + add_hubert_arguments(parser) + 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 updates happen to the model so far across + epochs. + + - sub_batch_idx_train: It contains number of batch 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 + + - warm_step: The warmup period that dictates the decay of the + scale on "simple" (un-pruned) loss. + """ + params = AttributeDict( + { + "best_train_loss": float("inf"), + "best_valid_loss": float("inf"), + "best_train_epoch": -1, + "best_valid_epoch": -1, + "batch_idx_train": 0, + "sub_batch_idx_train": 0, + "log_interval": 50, + "reset_interval": 200, + "valid_interval": 3000, # For the 100h subset, use 800 + # parameters for pruned RNN-T loss + "warm_step": 2000, + "env_info": get_env_info(), + } + ) + + return params + + +def _to_int_tuple(s: str): + return tuple(map(int, s.split(","))) + + +def get_encoder_model(params: AttributeDict) -> nn.Module: + if hasattr(params, "pretrained_dir"): + logging.info(f"Loading {params.pretrained_dir}") + pretrained = torch.load(params.pretrained_dir) + encoder = HubertModel(params) + encoder.load_state_dict(pretrained["model"]) + else: + encoder = HubertModel(params) + 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_embed_dim, + decoder_dim=params.decoder_dim, + joiner_dim=params.joiner_dim, + vocab_size=params.vocab_size, + ) + return joiner + + +def get_model(params: AttributeDict) -> nn.Module: + assert params.use_transducer or params.use_ctc, ( + f"At least one of them should be True, " + f"but got params.use_transducer={params.use_transducer}, " + f"params.use_ctc={params.use_ctc}" + ) + + encoder = get_encoder_model(params) + + if params.use_transducer: + decoder = get_decoder_model(params) + joiner = get_joiner_model(params) + else: + decoder = None + joiner = None + + model = AsrModel( + encoder=encoder, + decoder=decoder, + joiner=joiner, + encoder_dim=params.encoder_embed_dim, + decoder_dim=params.decoder_dim, + vocab_size=params.vocab_size, + use_transducer=params.use_transducer, + use_ctc=params.use_ctc, + ) + return model + + +def load_checkpoint_if_available( + params: AttributeDict, + model: nn.Module, + model_avg: nn.Module = None, + optimizer: Optional[torch.optim.Optimizer] = None, + scheduler: Optional[LRSchedulerType] = None, +) -> Optional[Dict[str, Any]]: + """Load checkpoint from file. + + If params.start_batch is positive, it will load the checkpoint from + `params.exp_dir/checkpoint-{params.start_batch}.pt`. Otherwise, if + params.start_epoch is larger than 1, it will load the checkpoint from + `params.start_epoch - 1`. + + Apart from loading state dict for `model` and `optimizer` it also updates + `best_train_epoch`, `best_train_loss`, `best_valid_epoch`, + and `best_valid_loss` in `params`. + + Args: + params: + The return value of :func:`get_params`. + model: + The training model. + model_avg: + The stored model averaged from the start of training. + optimizer: + The optimizer that we are using. + scheduler: + The scheduler that we are using. + Returns: + Return a dict containing previously saved training info. + """ + if params.start_batch > 0: + filename = params.exp_dir / f"checkpoint-{params.start_batch}.pt" + elif params.start_epoch > 1: + filename = params.exp_dir / f"epoch-{params.start_epoch-1}.pt" + else: + return None + + assert filename.is_file(), f"{filename} does not exist!" + + saved_params = load_checkpoint( + filename, + model=model, + model_avg=model_avg, + optimizer=optimizer, + scheduler=scheduler, + ) + + keys = [ + "best_train_epoch", + "best_valid_epoch", + "batch_idx_train", + "best_train_loss", + "best_valid_loss", + ] + for k in keys: + params[k] = saved_params[k] + + if params.start_batch > 0: + if "cur_epoch" in saved_params: + params["start_epoch"] = saved_params["cur_epoch"] + + return saved_params + + +def save_checkpoint( + params: AttributeDict, + model: Union[nn.Module, DDP], + model_avg: Optional[nn.Module] = None, + optimizer: Optional[torch.optim.Optimizer] = None, + scheduler: Optional[LRSchedulerType] = None, + sampler: Optional[CutSampler] = None, + scaler: Optional[GradScaler] = None, + rank: int = 0, +) -> None: + """Save model, optimizer, scheduler and training stats to file. + + Args: + params: + It is returned by :func:`get_params`. + model: + The training model. + model_avg: + The stored model averaged from the start of training. + optimizer: + The optimizer used in the training. + sampler: + The sampler for the training dataset. + scaler: + The scaler used for mix precision training. + """ + if rank != 0: + return + filename = params.exp_dir / f"epoch-{params.cur_epoch}.pt" + save_checkpoint_impl( + filename=filename, + model=model, + model_avg=model_avg, + params=params, + optimizer=optimizer, + scheduler=scheduler, + sampler=sampler, + scaler=scaler, + rank=rank, + ) + + if params.best_train_epoch == params.cur_epoch: + best_train_filename = params.exp_dir / "best-train-loss.pt" + copyfile(src=filename, dst=best_train_filename) + + if params.best_valid_epoch == params.cur_epoch: + best_valid_filename = params.exp_dir / "best-valid-loss.pt" + copyfile(src=filename, dst=best_valid_filename) + + +def compute_loss( + params: AttributeDict, + model: Union[nn.Module, DDP], + sp: spm.SentencePieceProcessor, + batch: dict, + is_training: bool, +) -> Tuple[Tensor, MetricsTracker]: + """ + Compute loss given the model and its inputs. + + Args: + params: + Parameters for training. See :func:`get_params`. + model: + The model for training. It is an instance of Zipformer in our case. + batch: + A batch of data. See `dataset.HubertAsrDataset()` + 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 + audio = batch["audio"].to(device) + padding_mask = batch["padding_mask"].to(device) + + batch_idx_train = params.batch_idx_train + warm_step = params.warm_step + + texts = batch["supervisions"]["text"] + y = sp.encode(texts, out_type=int) + y = k2.RaggedTensor(y) + + with torch.set_grad_enabled(is_training): + simple_loss, pruned_loss, ctc_loss, num_frames = model( + x=audio, + padding_mask=padding_mask, + y=y, + prune_range=params.prune_range, + am_scale=params.am_scale, + lm_scale=params.lm_scale, + ) + + loss = 0.0 + + if params.use_transducer: + s = params.simple_loss_scale + # take down the scale on the simple loss from 1.0 at the start + # to params.simple_loss scale by warm_step. + simple_loss_scale = ( + s + if batch_idx_train >= warm_step + else 1.0 - (batch_idx_train / warm_step) * (1.0 - s) + ) + pruned_loss_scale = ( + 1.0 + if batch_idx_train >= warm_step + else 0.1 + 0.9 * (batch_idx_train / warm_step) + ) + loss += simple_loss_scale * simple_loss + pruned_loss_scale * pruned_loss + + if params.use_ctc: + loss += params.ctc_loss_scale * ctc_loss + + assert loss.requires_grad == is_training + + info = MetricsTracker() + with warnings.catch_warnings(): + warnings.simplefilter("ignore") + info["frames"] = num_frames.sum().item() + + # Note: We use reduction=sum while computing the loss. + info["loss"] = loss.detach().cpu().item() + if params.use_transducer: + info["simple_loss"] = simple_loss.detach().cpu().item() + info["pruned_loss"] = pruned_loss.detach().cpu().item() + if params.use_ctc: + info["ctc_loss"] = ctc_loss.detach().cpu().item() + + return loss, info + + +def compute_validation_loss( + params: AttributeDict, + model: Union[nn.Module, DDP], + sp: spm.SentencePieceProcessor, + valid_dl: torch.utils.data.DataLoader, + world_size: int = 1, +) -> MetricsTracker: + """Run the validation process.""" + model.eval() + + tot_loss = MetricsTracker() + + for batch_idx, batch in enumerate(valid_dl): + loss, loss_info = compute_loss( + params=params, + model=model, + sp=sp, + batch=batch, + is_training=False, + ) + assert loss.requires_grad is False + tot_loss = tot_loss + loss_info + + if world_size > 1: + tot_loss.reduce(loss.device) + + loss_value = tot_loss["loss"] / tot_loss["frames"] + if loss_value < params.best_valid_loss: + params.best_valid_epoch = params.cur_epoch + params.best_valid_loss = loss_value + + return tot_loss + + +def train_one_epoch( + params: AttributeDict, + model: Union[nn.Module, DDP], + optimizer: torch.optim.Optimizer, + scheduler: LRSchedulerType, + sp: spm.SentencePieceProcessor, + train_dl: torch.utils.data.DataLoader, + valid_dl: torch.utils.data.DataLoader, + scaler: GradScaler, + model_avg: Optional[nn.Module] = None, + tb_writer: Optional[SummaryWriter] = None, + world_size: int = 1, + rank: int = 0, +) -> None: + """Train the model for one epoch. + + The training loss from the mean of all frames is saved in + `params.train_loss`. It runs the validation process every + `params.valid_interval` batches. + + Args: + params: + It is returned by :func:`get_params`. + model: + The model for training. + optimizer: + The optimizer we are using. + scheduler: + The learning rate scheduler, we call step() every step. + train_dl: + Dataloader for the training dataset. + valid_dl: + Dataloader for the validation dataset. + scaler: + The scaler used for mix precision training. + model_avg: + The stored model averaged from the start of training. + tb_writer: + Writer to write log messages to tensorboard. + world_size: + Number of nodes in DDP training. If it is 1, DDP is disabled. + rank: + The rank of the node in DDP training. If no DDP is used, it should + be set to 0. + """ + model.train() + + tot_loss = MetricsTracker() + + saved_bad_model = False + + def save_bad_model(suffix: str = ""): + save_checkpoint_impl( + filename=params.exp_dir / f"bad-model{suffix}-{rank}.pt", + model=model, + model_avg=model_avg, + params=params, + optimizer=optimizer, + scheduler=scheduler, + sampler=train_dl.sampler, + scaler=scaler, + rank=0, + ) + + for sub_batch_idx, batch in enumerate(train_dl): + params.sub_batch_idx_train += 1 + batch_idx = sub_batch_idx // params.accum_grad + + if batch_idx % 10 == 0: + set_batch_count(model, get_adjusted_batch_count(params)) + + batch_size = len(batch["supervisions"]["text"]) + + try: + with torch.cuda.amp.autocast(enabled=params.use_fp16): + loss, loss_info = compute_loss( + params=params, + model=model, + sp=sp, + batch=batch, + is_training=True, + ) + # summary stats + tot_loss = (tot_loss * (1 - 1 / params.reset_interval)) + loss_info + + # NOTE: We use reduction==sum and loss is computed over utterances + # in the batch and there is no normalization to it so far. + scaler.scale(loss / params.accum_grad).backward() + + if sub_batch_idx % params.accum_grad == params.accum_grad - 1: + params.batch_idx_train += 1 + scheduler.step_batch(params.batch_idx_train) + + scaler.step(optimizer) + scaler.update() + optimizer.zero_grad() + else: + continue + + except: # noqa + save_bad_model() + display_and_save_batch(batch, params=params, sp=sp) + raise + + if params.print_diagnostics and batch_idx == 5: + return + + if ( + rank == 0 + and params.batch_idx_train > 0 + and params.batch_idx_train % params.average_period == 0 + ): + update_averaged_model( + params=params, + model_cur=model, + model_avg=model_avg, + ) + + if ( + params.batch_idx_train > 0 + and params.batch_idx_train % params.save_every_n == 0 + ): + save_checkpoint_with_global_batch_idx( + out_dir=params.exp_dir, + global_batch_idx=params.batch_idx_train, + model=model, + model_avg=model_avg, + params=params, + optimizer=optimizer, + scheduler=scheduler, + sampler=train_dl.sampler, + scaler=scaler, + rank=rank, + ) + remove_checkpoints( + out_dir=params.exp_dir, + topk=params.keep_last_k, + rank=rank, + ) + + if batch_idx % 100 == 0 and params.use_fp16: + # If the grad scale was less than 1, try increasing it. The _growth_interval + # of the grad scaler is configurable, but we can't configure it to have different + # behavior depending on the current grad scale. + cur_grad_scale = scaler._scale.item() + + if cur_grad_scale < 8.0 or (cur_grad_scale < 32.0 and batch_idx % 400 == 0): + scaler.update(cur_grad_scale * 2.0) + if cur_grad_scale < 0.01: + if not saved_bad_model: + save_bad_model(suffix="-first-warning") + saved_bad_model = True + logging.warning(f"Grad scale is small: {cur_grad_scale}") + if cur_grad_scale < 1.0e-05: + save_bad_model() + raise RuntimeError( + f"grad_scale is too small, exiting: {cur_grad_scale}" + ) + + if batch_idx % params.log_interval == 0: + cur_lr = max(scheduler.get_last_lr()) + cur_grad_scale = scaler._scale.item() if params.use_fp16 else 1.0 + + logging.info( + f"Epoch {params.cur_epoch}, " + f"batch {batch_idx}, loss[{loss_info}], " + f"tot_loss[{tot_loss}], batch size: {batch_size}, " + f"lr: {cur_lr:.2e}, " + + (f"grad_scale: {scaler._scale.item()}" if params.use_fp16 else "") + ) + + if tb_writer is not None: + tb_writer.add_scalar( + "train/learning_rate", cur_lr, params.batch_idx_train + ) + + loss_info.write_summary( + tb_writer, "train/current_", params.batch_idx_train + ) + tot_loss.write_summary(tb_writer, "train/tot_", params.batch_idx_train) + if params.use_fp16: + tb_writer.add_scalar( + "train/grad_scale", + cur_grad_scale, + params.batch_idx_train, + ) + + if batch_idx % params.valid_interval == 0 and not params.print_diagnostics: + logging.info("Computing validation loss") + valid_info = compute_validation_loss( + params=params, + model=model, + sp=sp, + valid_dl=valid_dl, + world_size=world_size, + ) + model.train() + logging.info(f"Epoch {params.cur_epoch}, validation: {valid_info}") + logging.info( + f"Maximum memory allocated so far is {torch.cuda.max_memory_allocated()//1000000}MB" + ) + if tb_writer is not None: + valid_info.write_summary( + tb_writer, "train/valid_", params.batch_idx_train + ) + + if batch_idx % params.accum_grad != params.accum_grad - 1: + optimizer.zero_grad() + loss_value = tot_loss["loss"] / tot_loss["frames"] + params.train_loss = loss_value + if params.train_loss < params.best_train_loss: + params.best_train_epoch = params.cur_epoch + params.best_train_loss = params.train_loss + + +def run(rank, world_size, args): + """ + Args: + rank: + It is a value between 0 and `world_size-1`, which is + passed automatically by `mp.spawn()` in :func:`main`. + The node with rank 0 is responsible for saving checkpoint. + world_size: + Number of GPUs for DDP training. + args: + The return value of get_parser().parse_args() + """ + params = get_params() + params.update(vars(args)) + + fix_random_seed(params.seed) + if world_size > 1: + setup_dist(rank, world_size, params.master_port) + + setup_logger(f"{params.exp_dir}/log/log-train") + logging.info("Training started") + + if args.tensorboard and rank == 0: + tb_writer = SummaryWriter(log_dir=f"{params.exp_dir}/tensorboard") + else: + tb_writer = None + + device = torch.device("cpu") + if torch.cuda.is_available(): + device = torch.device("cuda", rank) + logging.info(f"Device: {device}") + + sp = spm.SentencePieceProcessor() + sp.load(params.bpe_model) + + # is defined in local/train_bpe_model.py + params.blank_id = sp.piece_to_id("") + params.vocab_size = sp.get_piece_size() + + if not params.use_transducer: + params.ctc_loss_scale = 1.0 + + logging.info(params) + + logging.info("About to create model") + model = get_model(params) + + num_param = sum([p.numel() for p in model.parameters()]) + logging.info(f"Number of model parameters: {num_param}") + + assert params.save_every_n >= params.average_period + model_avg: Optional[nn.Module] = None + if rank == 0: + # model_avg is only used with rank 0 + model_avg = copy.deepcopy(model).to(torch.float64) + + assert params.start_epoch > 0, params.start_epoch + checkpoints = load_checkpoint_if_available( + params=params, model=model, model_avg=model_avg + ) + + model.to(device) + if world_size > 1: + logging.info("Using DDP") + model = DDP(model, device_ids=[rank], find_unused_parameters=True) + + optimizer = ScaledAdam( + get_parameter_groups_with_lrs(model, lr=params.base_lr, include_names=True), + lr=params.base_lr, # should have no effect + clipping_scale=2.0, + ) + + scheduler = Eden(optimizer, params.lr_batches, params.lr_epochs) + + if checkpoints and "optimizer" in checkpoints: + logging.info("Loading optimizer state dict") + optimizer.load_state_dict(checkpoints["optimizer"]) + + if ( + checkpoints + and "scheduler" in checkpoints + and checkpoints["scheduler"] is not None + ): + logging.info("Loading scheduler state dict") + scheduler.load_state_dict(checkpoints["scheduler"]) + + if params.print_diagnostics: + opts = diagnostics.TensorDiagnosticOptions( + 512 + ) # allow 4 megabytes per sub-module + diagnostic = diagnostics.attach_diagnostics(model, opts) + + if params.inf_check: + register_inf_check_hooks(model) + + librispeech = LibriSpeechAsrDataModule(args) + + train_cuts = ( + librispeech.train_all_shuf_cuts() + if params.full_libri + else librispeech.train_clean_100_cuts() + ) + + def remove_short_and_long_utt(c: Cut): + # Keep only utterances with duration between 1 second and 20 seconds + # + # Caution: There is a reason to select 20.0 here. Please see + # ../local/display_manifest_statistics.py + # + # You should use ../local/display_manifest_statistics.py to get + # an utterance duration distribution for your dataset to select + # the threshold + if c.duration < 1.0 or c.duration > 20.0: + # logging.warning( + # f"Exclude cut with ID {c.id} from training. Duration: {c.duration}" + # ) + return False + + return True + + train_cuts = train_cuts.filter(remove_short_and_long_utt) + + if params.start_batch > 0 and checkpoints and "sampler" in checkpoints: + # We only load the sampler's state dict when it loads a checkpoint + # saved in the middle of an epoch + sampler_state_dict = checkpoints["sampler"] + else: + sampler_state_dict = None + + train_dl = librispeech.train_dataloaders( + train_cuts, + do_normalize=params.do_normalize, + 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, + do_normalize=params.do_normalize, + ) + + if params.sanity_check and not params.print_diagnostics: + scan_pessimistic_batches_for_oom( + model=model, + train_dl=train_dl, + optimizer=optimizer, + sp=sp, + params=params, + ) + + scaler = GradScaler(enabled=params.use_fp16, init_scale=1.0) + if checkpoints and "grad_scaler" in checkpoints: + logging.info("Loading grad scaler state dict") + scaler.load_state_dict(checkpoints["grad_scaler"]) + + for epoch in range(params.start_epoch, params.num_epochs + 1): + scheduler.step_epoch(epoch - 1) + fix_random_seed(params.seed + epoch - 1) + train_dl.sampler.set_epoch(epoch - 1) + + if tb_writer is not None: + tb_writer.add_scalar("train/epoch", epoch, params.batch_idx_train) + + params.cur_epoch = epoch + + train_one_epoch( + params=params, + model=model, + model_avg=model_avg, + optimizer=optimizer, + scheduler=scheduler, + sp=sp, + train_dl=train_dl, + valid_dl=valid_dl, + scaler=scaler, + tb_writer=tb_writer, + world_size=world_size, + rank=rank, + ) + + if params.print_diagnostics: + diagnostic.print_diagnostics() + break + + save_checkpoint( + params=params, + model=model, + model_avg=model_avg, + optimizer=optimizer, + scheduler=scheduler, + sampler=train_dl.sampler, + scaler=scaler, + rank=rank, + ) + + logging.info("Done!") + + if world_size > 1: + torch.distributed.barrier() + cleanup_dist() + + +def display_and_save_batch( + batch: dict, + params: AttributeDict, + sp: spm.SentencePieceProcessor, +) -> None: + """Display the batch statistics and save the batch into disk. + + Args: + batch: + A batch of data. See `dataset.HubertAsrDataset()` + for the content in it. + params: + Parameters for training. See :func:`get_params`. + sp: + The BPE model. + """ + from lhotse.utils import uuid4 + + filename = f"{params.exp_dir}/batch-{uuid4()}.pt" + logging.info(f"Saving batch to {filename}") + torch.save(batch, filename) + + audio = batch["audio"] + logging.info(f"audio shape: {audio.shape}") + + y = sp.encode(batch["supervisions"]["text"], out_type=int) + num_tokens = sum(len(i) for i in y) + logging.info(f"num tokens: {num_tokens}") + + +def scan_pessimistic_batches_for_oom( + model: Union[nn.Module, DDP], + train_dl: torch.utils.data.DataLoader, + optimizer: torch.optim.Optimizer, + sp: spm.SentencePieceProcessor, + params: AttributeDict, +): + from lhotse.dataset import find_pessimistic_batches + + logging.info( + "Sanity check -- see if any of the batches in epoch 1 would cause OOM." + ) + batches, crit_values = find_pessimistic_batches(train_dl.sampler) + for criterion, cuts in batches.items(): + batch = train_dl.dataset[cuts] + try: + with torch.cuda.amp.autocast(enabled=params.use_fp16): + loss, _ = compute_loss( + params=params, + model=model, + sp=sp, + batch=batch, + is_training=True, + ) + loss.backward() + optimizer.zero_grad() + except Exception as e: + if "CUDA out of memory" in str(e): + logging.error( + "Your GPU ran out of memory with the current " + "max_duration setting. We recommend decreasing " + "max_duration and trying again.\n" + f"Failing criterion: {criterion} " + f"(={crit_values[criterion]}) ..." + ) + display_and_save_batch(batch, params=params, sp=sp) + raise + logging.info( + f"Maximum memory allocated so far is {torch.cuda.max_memory_allocated()//1000000}MB" + ) + + +def main(): + parser = get_parser() + LibriSpeechAsrDataModule.add_arguments(parser) + args = parser.parse_args() + args.exp_dir = Path(args.exp_dir) + + world_size = args.world_size + assert world_size >= 1 + if world_size > 1: + mp.spawn(run, args=(world_size, args), nprocs=world_size, join=True) + else: + run(rank=0, world_size=1, args=args) + + +torch.set_num_threads(1) +torch.set_num_interop_threads(1) + +if __name__ == "__main__": + main() diff --git a/egs/librispeech/SSL/hubert/hubert.py b/egs/librispeech/SSL/hubert/hubert.py new file mode 100644 index 0000000000..f800044f48 --- /dev/null +++ b/egs/librispeech/SSL/hubert/hubert.py @@ -0,0 +1,984 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# Permission is hereby granted, free of charge, to any person obtaining a copy +# of this software and associated documentation files (the "Software"), to deal +# in the Software without restriction, including without limitation the rights +# to use, copy, modify, merge, publish, distribute, sublicense, and/or sell +# copies of the Software, and to permit persons to whom the Software is +# furnished to do so, subject to the following conditions: +# +# The above copyright notice and this permission notice shall be included in all +# copies or substantial portions of the Software. +# +# THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR +# IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, +# FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE +# AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER +# LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, +# OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE +# SOFTWARE. + +import argparse +import logging +from typing import Dict, List, Optional, Tuple + +import numpy as np +import torch +import torch.nn as nn +import torch.nn.functional as F +from utils import GradMultiply, LayerNorm +from wav2vec2_module import ConvFeatureExtractionModel, TransformerEncoder + + +def compute_mask_indices( + shape: Tuple[int, int], + padding_mask: Optional[torch.Tensor], + mask_prob: float, + mask_length: int, + mask_type: str = "static", + mask_other: float = 0.0, + min_masks: int = 0, + no_overlap: bool = False, + min_space: int = 0, + require_same_masks: bool = True, + mask_dropout: float = 0.0, + add_masks: bool = False, + seed: Optional[int] = None, + epoch: Optional[int] = None, + indices: Optional[torch.Tensor] = None, + idc_select_ver: int = 1, # 2 to reproduce mask_tokens_dataset + num_mask_ver: int = 2, # 2 to reproduce mask_tokens_dataset +) -> np.ndarray: + """ + Computes random mask spans for a given shape + + Args: + shape: the the shape for which to compute masks. + should be of size 2 where first element is batch size and 2nd is timesteps + padding_mask: optional padding mask of the same size as shape, which will prevent masking padded elements + mask_prob: probability for each token to be chosen as start of the span to be masked. this will be multiplied by + number of timesteps divided by length of mask span to mask approximately this percentage of all elements. + however due to overlaps, the actual number will be smaller (unless no_overlap is True) + mask_type: how to compute mask lengths + static = fixed size + uniform = sample from uniform distribution [mask_other, mask_length*2] + normal = sample from normal distribution with mean mask_length and stdev mask_other. mask is min 1 element + poisson = sample from possion distribution with lambda = mask length + min_masks: minimum number of masked spans + no_overlap: if false, will switch to an alternative recursive algorithm that prevents spans from overlapping + min_space: only used if no_overlap is True, this is how many elements to keep unmasked between spans + require_same_masks: if true, will randomly drop out masks until same amount of masks remains in each sample + mask_dropout: randomly dropout this percentage of masks in each example + """ + + bsz, all_sz = shape + mask = np.full((bsz, all_sz), False) + + if num_mask_ver == 1: + all_num_mask = int( + # add a random number for probabilistic rounding + mask_prob * all_sz / float(mask_length) + + np.random.rand() + ) + all_num_mask = max(min_masks, all_num_mask) + + mask_idcs = [] + for i in range(bsz): + if seed is not None and epoch is not None and indices is not None: + seed_i = int(hash((seed, epoch, indices[i].item())) % 1e6) + else: + seed_i = None + + rng = np.random.default_rng(seed_i) + + if padding_mask is not None: + sz = all_sz - padding_mask[i].long().sum().item() + assert sz >= 0, sz + else: + sz = all_sz + + if num_mask_ver == 1: + if padding_mask is not None: + num_mask = int( + # add a random number for probabilistic rounding + mask_prob * sz / float(mask_length) + + np.random.rand() + ) + num_mask = max(min_masks, num_mask) + else: + num_mask = all_num_mask + elif num_mask_ver == 2: + num_mask = int( + # add a random number for probabilistic rounding + mask_prob * sz / float(mask_length) + + rng.random() + ) + num_mask = max(min_masks, num_mask) + else: + raise ValueError() + + if mask_type == "static": + lengths = np.full(num_mask, mask_length) + elif mask_type == "uniform": + lengths = rng.randint(mask_other, mask_length * 2 + 1, size=num_mask) + elif mask_type == "normal": + lengths = rng.normal(mask_length, mask_other, size=num_mask) + lengths = [max(1, int(round(x))) for x in lengths] + elif mask_type == "poisson": + lengths = rng.poisson(mask_length, size=num_mask) + lengths = [int(round(x)) for x in lengths] + else: + raise Exception("unknown mask selection " + mask_type) + + if sum(lengths) == 0: + if mask_type == "static": + raise ValueError(f"this should never happens") + else: + lengths = [min(mask_length, sz - 1)] + + if no_overlap: + mask_idc = [] + + def arrange(s, e, length, keep_length): + span_start = rng.randint(s, e - length) + mask_idc.extend(span_start + i for i in range(length)) + + new_parts = [] + if span_start - s - min_space >= keep_length: + new_parts.append((s, span_start - min_space + 1)) + if e - span_start - length - min_space > keep_length: + new_parts.append((span_start + length + min_space, e)) + return new_parts + + parts = [(0, sz)] + min_length = min(lengths) + for length in sorted(lengths, reverse=True): + lens = np.fromiter( + (e - s if e - s >= length + min_space else 0 for s, e in parts), + np.int, + ) + l_sum = np.sum(lens) + if l_sum == 0: + break + probs = lens / np.sum(lens) + c = rng.choice(len(parts), p=probs) + s, e = parts.pop(c) + parts.extend(arrange(s, e, length, min_length)) + mask_idc = np.asarray(mask_idc) + else: + if idc_select_ver == 1: + min_len = min(lengths) + if sz - min_len <= num_mask: + min_len = sz - num_mask - 1 + mask_idc = rng.choice(sz - min_len, num_mask, replace=False) + elif idc_select_ver == 2: + mask_idc = rng.choice(sz, num_mask, replace=False) + else: + raise ValueError() + + mask_idc = np.asarray( + [ + mask_idc[j] + offset + for j in range(len(mask_idc)) + for offset in range(lengths[j]) + ] + ) + + mask_idc = np.unique(mask_idc[mask_idc < sz]) + if len(mask_idc) >= sz: + raise ValueError( + ( + f"the entire sequence is masked. " + f"sz={sz}; mask_idc[mask_idc]; " + f"index={indices[i] if indices is not None else None}" + ) + ) + mask_idcs.append(mask_idc) + + target_len = None + if require_same_masks: + if add_masks: + target_len = max([len(m) for m in mask_idcs]) + else: + target_len = min([len(m) for m in mask_idcs]) + + for i, mask_idc in enumerate(mask_idcs): + if target_len is not None and len(mask_idc) > target_len: + mask_idc = rng.choice(mask_idc, target_len, replace=False) + + mask[i, mask_idc] = True + + if target_len is not None and len(mask_idc) < target_len: + unmasked = np.flatnonzero(~mask[i]) + to_mask = rng.choice(unmasked, target_len - len(mask_idc), replace=False) + mask[i, to_mask] = True + + if mask_dropout > 0: + masked = np.flatnonzero(mask[i]) + num_holes = np.rint(len(masked) * mask_dropout).astype(int) + to_drop = rng.choice(masked, num_holes, replace=False) + mask[i, to_drop] = False + + return mask + + +def add_hubert_arguments(parser: argparse.ArgumentParser): + parser.add_argument( + "--label-rate", + type=float, + default=50, + ) + + parser.add_argument( + "--sample-rate", + type=float, + default=16000, + ) + + parser.add_argument( + "--extractor-mode", + type=str, + default="default", + help="""mode for feature extractor, should in EXTRACTOR_MODE_CHOICES. default has a single group + norm with d groups in the first conv block, whereas layer_norm + has layer norms in every block (meant to use with normalize=True)""", + ) + parser.add_argument( + "--encoder-layers", + type=int, + default=12, + help="num encoder layers in the transformer", + ) + + parser.add_argument( + "--encoder-embed-dim", + type=int, + default=768, + help="encoder embedding dimension", + ) + + parser.add_argument( + "--encoder-ffn-embed-dim", + type=int, + default=3072, + help="encoder embedding dimension for FFN", + ) + + parser.add_argument( + "--encoder-attention-heads", + type=int, + default=12, + help="num encoder attention heads", + ) + + parser.add_argument( + "--activation-fn", + type=str, + choices=[ + "relu", + "gelu", + "gelu_fast", + "gelu_accurate", + "tanh", + "linear", + ], + default="gelu", + help="activation function to use", + ) + + parser.add_argument( + "--layer-type", + type=str, + choices=["transformer", "conformer", "trf_adp"], + default="transformer", + help="layer type in encoder", + ) + + # dropouts + parser.add_argument( + "--dropout", + type=float, + default=0.1, + help="dropout probability for the transformer", + ) + + parser.add_argument( + "--attention-dropout", + type=float, + default=0.1, + help="dropout probability for attention weights", + ) + + parser.add_argument( + "--activation-dropout", + type=float, + default=0.0, + help="dropout probability after activation in FFN", + ) + + parser.add_argument( + "--encoder-layerdrop", + type=float, + default=0.0, + help="probability of dropping a tarnsformer layer", + ) + + parser.add_argument( + "--dropout-input", + type=float, + default=0.0, + help="dropout to apply to the input (after feat extr)", + ) + + parser.add_argument( + "--dropout-features", + type=float, + default=0.0, + help="dropout to apply to the features (after feat extr)", + ) + + parser.add_argument( + "--final-dim", + type=int, + default=0, + help="project final representations and targets to this many dimensions. set to encoder_embed_dim is <= 0", + ) + + parser.add_argument( + "--untie-final-proj", + type=bool, + default=False, + help="use separate projection for each target", + ) + + parser.add_argument( + "--layer-norm-first", + type=bool, + default=False, + help="apply layernorm first in the transformer", + ) + + parser.add_argument( + "--conv-feature-layers", + type=str, + default="[(512,10,5)] + [(512,3,2)] * 4 + [(512,2,2)] * 2", + help="string describing convolutional feature extraction layers in form of a python list that contains [(dim, kernel_size, stride), ...]", + ) + + parser.add_argument( + "--conv-bias", + type=bool, + default=False, + help="include bias in conv encoder", + ) + + parser.add_argument( + "--logit-temp", + type=float, + default=0.1, + help="temperature to divide logits by", + ) + + parser.add_argument( + "--target-glu", + type=bool, + default=False, + help="adds projection + glu to targets", + ) + + parser.add_argument( + "--feature-grad-mult", + type=float, + default=1.0, + help="multiply feature extractor var grads by this", + ) + + # masking + parser.add_argument("--mask-length", type=int, default=10, help="mask_length") + + parser.add_argument( + "--mask-prob", + type=float, + default=0.65, + help="probability of replacing a token with mask", + ) + + parser.add_argument( + "--mask-selection", + type=str, + choices=["static", "uniform", "normal", "poisson"], + default="static", + help="how to choose mask length", + ) + + parser.add_argument( + "--mask-other", + type=float, + default=0, + help="secondary mask argument (used for more complex distributions),see help in compute_mask_indicesh", + ) + + parser.add_argument( + "--no-mask-overlap", + type=bool, + default=False, + help="whether to allow masks to overlap", + ) + + parser.add_argument( + "--mask-min-space", + type=int, + default=1, + help="min space between spans (if no overlap is enabled)", + ) + + # channel masking + parser.add_argument( + "--mask-channel-length", + type=int, + default=10, + help="length of the mask for features (channels)", + ) + + parser.add_argument( + "--mask-channel-prob", + type=float, + default=0.0, + help="probability of replacing a feature with 0", + ) + + parser.add_argument( + "--mask-channel-selection", + type=str, + choices=["static", "uniform", "normal", "poisson"], + default="static", + help="how to choose mask length for channel masking", + ) + + parser.add_argument( + "--mask-channel-other", + type=float, + default=0, + help="secondary mask argument (used for more complex distributions), see help in compute_mask_indicesh", + ) + + parser.add_argument( + "--no-mask-channel-overlap", + type=bool, + default=False, + help="whether to allow channel masks to overlap", + ) + + parser.add_argument( + "--mask-channel-min-space", + type=int, + default=1, + help="min space between spans (if no overlap is enabled)", + ) + + # positional embeddings + parser.add_argument( + "--conv-pos", + type=int, + default=128, + help="number of filters for convolutional positional embeddings", + ) + + parser.add_argument( + "--conv-pos-groups", + type=int, + default=16, + help="number of groups for convolutional positional embedding", + ) + + parser.add_argument( + "--conv-pos-batch-norm", + type=bool, + default=False, + help="use batch norm instead of weight norm in conv_pos (for bf16 models)", + ) + + parser.add_argument( + "--latent-temp", + type=float, + nargs="*", + default=[2, 0.5, 0.999995], + help="legacy (to be removed)", + ) + + # loss computation + parser.add_argument( + "--skip-masked", + type=bool, + default=False, + help="skip computing losses over masked frames", + ) + + parser.add_argument( + "--skip-nomask", + type=bool, + default=False, + help="skip computing losses over unmasked frames", + ) + + parser.add_argument( + "--checkpoint-activations", + type=bool, + default=False, + help="recompute activations and save memory for extra compute", + ) + + parser.add_argument( + "--pred-masked-weight", + type=float, + default=1, + help="weight for masked part in ssl loss", + ) + + parser.add_argument( + "--pred-nomask-weight", + type=float, + default=0, + help="weight for masked part in ssl loss", + ) + + parser.add_argument( + "--loss-weights", + type=float, + nargs="*", + default=[10], + help="weight for masked part in ssl loss", + ) + + # FP16 optimization + parser.add_argument( + "--required-seq-len-multiple", + type=int, + default=2, + help="pad the input to encoder such that the sequence length is divisible by multiple", + ) + + parser.add_argument( + "--attn-type", type=str, default="", help="if espnet use ESPNET MHA" + ) + + parser.add_argument( + "--pos-enc-type", + type=str, + default="abs", + help="Positional encoding type to use in conformer", + ) + + parser.add_argument( + "--num-classes", + type=int, + nargs="*", + default=[504], + help="""num class, a little larger than the number of cluster, + the largest is for padding, + and the value should be the multiple of 4, for faster computation""", + ) + + +class HubertModel(nn.Module): + def __init__( + self, + cfg, + ) -> None: + super().__init__() + feature_enc_layers = eval(cfg.conv_feature_layers) # noqa + self.embed = feature_enc_layers[-1][0] + + self.feature_extractor = ConvFeatureExtractionModel( + conv_layers=feature_enc_layers, + dropout=0.0, + mode=cfg.extractor_mode, + conv_bias=cfg.conv_bias, + ) + feature_ds_rate = np.prod([s for _, _, s in feature_enc_layers]) + self.feat2tar_ratio = cfg.label_rate * feature_ds_rate / cfg.sample_rate + + self.post_extract_proj = ( + nn.Linear(self.embed, cfg.encoder_embed_dim) + if self.embed != cfg.encoder_embed_dim + else None + ) + + self.mask_prob = cfg.mask_prob + self.mask_selection = cfg.mask_selection + self.mask_other = cfg.mask_other + self.mask_length = cfg.mask_length + self.no_mask_overlap = cfg.no_mask_overlap + self.mask_min_space = cfg.mask_min_space + + self.mask_channel_prob = cfg.mask_channel_prob + self.mask_channel_selection = cfg.mask_channel_selection + self.mask_channel_other = cfg.mask_channel_other + self.mask_channel_length = cfg.mask_channel_length + self.no_mask_channel_overlap = cfg.no_mask_channel_overlap + self.mask_channel_min_space = cfg.mask_channel_min_space + + self.dropout_input = nn.Dropout(cfg.dropout_input) + self.dropout_features = nn.Dropout(cfg.dropout_features) + + self.feature_grad_mult = cfg.feature_grad_mult + self.logit_temp = cfg.logit_temp + self.skip_masked = cfg.skip_masked + self.skip_nomask = cfg.skip_nomask + + final_dim = cfg.final_dim if cfg.final_dim > 0 else cfg.encoder_embed_dim + + self.mask_emb = nn.Parameter( + torch.FloatTensor(cfg.encoder_embed_dim).uniform_() + ) + + self.encoder = TransformerEncoder(cfg) + self.layer_norm = LayerNorm(self.embed) + + self.target_glu = None + if cfg.target_glu: + self.target_glu = nn.Sequential( + nn.Linear(final_dim, final_dim * 2), nn.GLU() + ) + + self.untie_final_proj = cfg.untie_final_proj + if self.untie_final_proj: + self.final_proj = nn.Linear( + cfg.encoder_embed_dim, final_dim * len(cfg.num_classes) + ) + else: + self.final_proj = nn.Linear(cfg.encoder_embed_dim, final_dim) + + # modules below are not needed during fine-tuning + self.num_classes = cfg.num_classes + self.label_embs_concat = nn.Parameter( + torch.FloatTensor(sum(self.num_classes), final_dim) + ) + self.pred_masked_weight = cfg.pred_masked_weight + self.pred_nomask_weight = cfg.pred_nomask_weight + self.loss_weights = cfg.loss_weights + nn.init.uniform_(self.label_embs_concat) + + def upgrade_state_dict_named(self, state_dict, name): + """Upgrade a (possibly old) state dict for new versions of fairseq.""" + + super().upgrade_state_dict_named(state_dict, name) + return state_dict + + def apply_mask(self, x, padding_mask, target_list): + B, T, C = x.shape + if self.mask_prob > 0: + mask_indices = compute_mask_indices( + (B, T), + padding_mask, + self.mask_prob, + self.mask_length, + self.mask_selection, + self.mask_other, + min_masks=2, + no_overlap=self.no_mask_overlap, + min_space=self.mask_min_space, + ) + mask_indices = torch.from_numpy(mask_indices).to(x.device) + x[mask_indices] = self.mask_emb.to(x.dtype) + else: + mask_indices = None + + if self.mask_channel_prob > 0: + mask_channel_indices = compute_mask_indices( + (B, C), + None, + self.mask_channel_prob, + self.mask_channel_length, + self.mask_channel_selection, + self.mask_channel_other, + no_overlap=self.no_mask_channel_overlap, + min_space=self.mask_channel_min_space, + ) + mask_channel_indices = ( + torch.from_numpy(mask_channel_indices) + .to(x.device) + .unsqueeze(1) + .expand(-1, T, -1) + ) + x[mask_channel_indices] = 0 + + return x, mask_indices + + def compute_nce(self, x, pos, negs): + neg_is_pos = (pos == negs).all(-1) + pos = pos.unsqueeze(0) + targets = torch.cat([pos, negs], dim=0) + + logits = torch.cosine_similarity(x.float(), targets.float(), dim=-1).type_as(x) + logits /= self.logit_temp + if neg_is_pos.any(): + logits[1:][neg_is_pos] = float("-inf") + logits = logits.transpose(0, 1) # (num_x, num_cls+1) + return logits + + def forward_features(self, source: torch.Tensor) -> torch.Tensor: + if self.feature_grad_mult > 0: + features = self.feature_extractor(source) + if self.feature_grad_mult != 1.0: + features = GradMultiply.apply(features, self.feature_grad_mult) + else: + with torch.no_grad(): + features = self.feature_extractor(source) + return features + + def forward_targets( + self, + features: torch.Tensor, + target_list: List[torch.Tensor], + ) -> Tuple[torch.Tensor, torch.Tensor]: + # Trim features to ensure labels exist and then get aligned labels + feat_tsz = features.size(2) + targ_tsz = min([t.size(1) for t in target_list]) + if self.feat2tar_ratio * feat_tsz > targ_tsz: + feat_tsz = int(targ_tsz / self.feat2tar_ratio) + features = features[..., :feat_tsz] + target_inds = torch.arange(feat_tsz).float() * self.feat2tar_ratio + target_list = [t[:, target_inds.long()] for t in target_list] + return features, target_list + + def forward_padding_mask( + self, + features: torch.Tensor, + padding_mask: torch.Tensor, + ) -> torch.Tensor: + extra = padding_mask.size(1) % features.size(1) + if extra > 0: + padding_mask = padding_mask[:, :-extra] + padding_mask = padding_mask.view(padding_mask.size(0), features.size(1), -1) + padding_mask = padding_mask.all(-1) + return padding_mask + + def forward( + self, + source: torch.Tensor, + target_list: Optional[List[torch.Tensor]] = None, + padding_mask: Optional[torch.Tensor] = None, + mask: bool = True, + features_only: bool = False, + output_layer: Optional[int] = None, + ): + """output layer is 1-based""" + features = self.forward_features(source) + if target_list is not None: + features, target_list = self.forward_targets(features, target_list) + + features_pen = features.float().pow(2).mean() + + features = features.transpose(1, 2) + features = self.layer_norm(features) + unmasked_features = features.clone() + + if padding_mask is not None: + padding_mask = self.forward_padding_mask(features, padding_mask) + + if self.post_extract_proj is not None: + features = self.post_extract_proj(features) + + features = self.dropout_input(features) + unmasked_features = self.dropout_features(unmasked_features) + + if mask: + x, mask_indices = self.apply_mask(features, padding_mask, target_list) + else: + x = features + mask_indices = None + + # feature: (B, T, D), float + # target: (B, T), long + # x: (B, T, D), float + # padding_mask: (B, T), bool + # mask_indices: (B, T), bool + x, _ = self.encoder( + x, + padding_mask=padding_mask, + layer=None if output_layer is None else output_layer - 1, + ) + + if features_only: + return {"x": x, "padding_mask": padding_mask, "features": features} + + def compute_pred(proj_x, target, label_embs): + # compute logits for the i-th label set + y = torch.index_select(label_embs, 0, target.long()) + negs = label_embs.unsqueeze(1).expand(-1, proj_x.size(0), -1) + if self.target_glu: + y = self.target_glu(y) + negs = self.target_glu(negs) + # proj_x: (S, D) + # y: (S, D) + # negs: (Neg, S, D) + return self.compute_nce(proj_x, y, negs) + + label_embs_list = self.label_embs_concat.split(self.num_classes, 0) + + if not self.skip_masked: + masked_indices = torch.logical_and(~padding_mask, mask_indices) + proj_x_m = self.final_proj(x[masked_indices]) + if self.untie_final_proj: + proj_x_m_list = proj_x_m.chunk(len(target_list), dim=-1) + else: + proj_x_m_list = [proj_x_m for _ in range(len(target_list))] + logit_m_list = [ + compute_pred(proj_x_m, t[masked_indices], label_embs_list[i]) + for i, (proj_x_m, t) in enumerate(zip(proj_x_m_list, target_list)) + ] + else: + logit_m_list = [None for _ in target_list] + + if not self.skip_nomask: + nomask_indices = torch.logical_and(~padding_mask, ~mask_indices) + proj_x_u = self.final_proj(x[nomask_indices]) + if self.untie_final_proj: + proj_x_u_list = proj_x_u.chunk(len(target_list), dim=-1) + else: + proj_x_u_list = [proj_x_u for _ in range(len(target_list))] + + logit_u_list = [ + compute_pred(proj_x_u, t[nomask_indices], label_embs_list[i]) + for i, (proj_x_u, t) in enumerate(zip(proj_x_u_list, target_list)) + ] + else: + logit_u_list = [None for _ in target_list] + + # result = { + # "logit_m_list": logit_m_list, + # "logit_u_list": logit_u_list, + # "padding_mask": padding_mask, + # "features_pen": features_pen, + # } + return self.compute_loss(logit_m_list, logit_u_list, features_pen) + + def extract_features( + self, + source: torch.Tensor, + padding_mask: Optional[torch.Tensor] = None, + mask: bool = False, + ret_conv: bool = False, + output_layer: Optional[int] = None, + ) -> Tuple[torch.Tensor, torch.Tensor]: + res = self.forward( + source, + padding_mask=padding_mask, + mask=mask, + features_only=True, + output_layer=output_layer, + ) + feature = res["features"] if ret_conv else res["x"] + return feature, res["padding_mask"] + + def get_logits(self, net_output, is_masked=True): + if is_masked: + logits_list = net_output["logit_m_list"] + else: + logits_list = net_output["logit_u_list"] + logits_list = [x.float() for x in logits_list if x is not None] + return logits_list + + def get_targets(self, net_output, is_masked=True): + logits_list = self.get_logits(net_output, is_masked) + targets_list = [x.new_zeros(x.size(0), dtype=torch.long) for x in logits_list] + return targets_list + + def get_extra_losses(self, net_output): + extra_losses = [] + names = [] + + if "features_pen" in net_output: + extra_losses.append(net_output["features_pen"]) + names.append("features_pen") + + return extra_losses, names + + def remove_pretraining_modules(self): + self.target_glu = None + self.final_proj = None + + def compute_loss(self, logit_m_list, logit_u_list, features_pen): + loss = 0.0 + sample_size = 0 + logging_output = {} + reduce = True + reduction = "sum" if reduce else "none" + + loss_m_list = [] + logp_m_list = [x.float() for x in logit_m_list if x is not None] + targ_m_list = [x.new_zeros(x.size(0), dtype=torch.long) for x in logp_m_list] + assert self.pred_masked_weight == 0 or len(logp_m_list) > 0 + for i, (logp_m, targ_m) in enumerate(zip(logp_m_list, targ_m_list)): + loss_m = F.cross_entropy(logp_m, targ_m, reduction=reduction) + loss_m_list.append(loss_m) + logging_output[f"loss_m_{i}"] = loss_m.detach().item() + if self.pred_masked_weight > 0: + loss += self.pred_masked_weight * sum(loss_m_list) + sample_size += targ_m_list[0].numel() + + loss_u_list = [] + logp_u_list = [x.float() for x in logit_u_list if x is not None] + targ_u_list = [x.new_zeros(x.size(0), dtype=torch.long) for x in logp_u_list] + assert self.pred_nomask_weight == 0 or len(logp_u_list) > 0 + for i, (logp_u, targ_u) in enumerate(zip(logp_u_list, targ_u_list)): + loss_u = F.cross_entropy(logp_u, targ_u, reduction=reduction) + loss_u_list.append(loss_u) + logging_output[f"loss_u_{i}"] = loss_u.detach().item() + if self.pred_nomask_weight > 0: + loss += self.pred_nomask_weight * sum(loss_u_list) + sample_size += targ_u_list[0].numel() + + if self.loss_weights is not None: + extra_losses = [] + names = [] + extra_losses.append(features_pen) + names.append("features_pen") + if torch.is_tensor(extra_losses): + extra_losses = [extra_losses] + names = [names] + if len(self.loss_weights) == 1 and len(extra_losses) != 1: + self.loss_weights = [self.loss_weights[0]] * len(extra_losses) + assert len(extra_losses) == len( + self.loss_weights + ), f"{len(extra_losses)}, {len(self.loss_weights)}" + for p, n, coef in zip(extra_losses, names, self.loss_weights): + if coef != 0 and p is not None: + p = coef * p.float() * sample_size + loss += p + logging_output[f"loss_{n}"] = p.item() + + logging_output = { + "loss": loss.item() if reduce else loss, + **logging_output, + } + + # for lk in self.log_keys: + # if lk in net_output: + # logging_output[lk] = float((net_output[lk])) + + def compute_correct(logits): + if logits.numel() == 0: + return 0, 0 + else: + assert logits.dim() > 1, logits.shape + max = logits.argmax(-1) == 0 + min = logits.argmin(-1) == 0 + both = max & min + corr = max.long().sum().item() - both.long().sum().item() + count = max.numel() + return corr, count + + with torch.no_grad(): + for i, logp_m in enumerate(logp_m_list): + corr_m, count_m = compute_correct(logp_m) + logging_output[f"correct_m_{i}"] = corr_m + logging_output[f"count_m_{i}"] = count_m + + for i, logp_u in enumerate(logp_u_list): + corr_u, count_u = compute_correct(logp_u) + logging_output[f"correct_u_{i}"] = corr_u + logging_output[f"count_u_{i}"] = count_u + + return loss, sample_size, logging_output diff --git a/egs/librispeech/SSL/hubert/hubert_ce.py b/egs/librispeech/SSL/hubert/hubert_ce.py new file mode 100644 index 0000000000..ccdd63efd5 --- /dev/null +++ b/egs/librispeech/SSL/hubert/hubert_ce.py @@ -0,0 +1,940 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# Permission is hereby granted, free of charge, to any person obtaining a copy +# of this software and associated documentation files (the "Software"), to deal +# in the Software without restriction, including without limitation the rights +# to use, copy, modify, merge, publish, distribute, sublicense, and/or sell +# copies of the Software, and to permit persons to whom the Software is +# furnished to do so, subject to the following conditions: +# +# The above copyright notice and this permission notice shall be included in all +# copies or substantial portions of the Software. +# +# THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR +# IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, +# FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE +# AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER +# LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, +# OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE +# SOFTWARE. + +import argparse +import logging +from typing import Dict, List, Optional, Tuple + +import numpy as np +import torch +import torch.nn as nn +import torch.nn.functional as F +from utils import GradMultiply, LayerNorm +from wav2vec2_module import ConvFeatureExtractionModel, TransformerEncoder + + +def compute_mask_indices( + shape: Tuple[int, int], + padding_mask: Optional[torch.Tensor], + mask_prob: float, + mask_length: int, + mask_type: str = "static", + mask_other: float = 0.0, + min_masks: int = 0, + no_overlap: bool = False, + min_space: int = 0, + require_same_masks: bool = True, + mask_dropout: float = 0.0, + add_masks: bool = False, + seed: Optional[int] = None, + epoch: Optional[int] = None, + indices: Optional[torch.Tensor] = None, + idc_select_ver: int = 1, # 2 to reproduce mask_tokens_dataset + num_mask_ver: int = 2, # 2 to reproduce mask_tokens_dataset +) -> np.ndarray: + """ + Computes random mask spans for a given shape + + Args: + shape: the the shape for which to compute masks. + should be of size 2 where first element is batch size and 2nd is timesteps + padding_mask: optional padding mask of the same size as shape, which will prevent masking padded elements + mask_prob: probability for each token to be chosen as start of the span to be masked. this will be multiplied by + number of timesteps divided by length of mask span to mask approximately this percentage of all elements. + however due to overlaps, the actual number will be smaller (unless no_overlap is True) + mask_type: how to compute mask lengths + static = fixed size + uniform = sample from uniform distribution [mask_other, mask_length*2] + normal = sample from normal distribution with mean mask_length and stdev mask_other. mask is min 1 element + poisson = sample from possion distribution with lambda = mask length + min_masks: minimum number of masked spans + no_overlap: if false, will switch to an alternative recursive algorithm that prevents spans from overlapping + min_space: only used if no_overlap is True, this is how many elements to keep unmasked between spans + require_same_masks: if true, will randomly drop out masks until same amount of masks remains in each sample + mask_dropout: randomly dropout this percentage of masks in each example + """ + + bsz, all_sz = shape + mask = np.full((bsz, all_sz), False) + + if num_mask_ver == 1: + all_num_mask = int( + # add a random number for probabilistic rounding + mask_prob * all_sz / float(mask_length) + + np.random.rand() + ) + all_num_mask = max(min_masks, all_num_mask) + + mask_idcs = [] + for i in range(bsz): + if seed is not None and epoch is not None and indices is not None: + seed_i = int(hash((seed, epoch, indices[i].item())) % 1e6) + else: + seed_i = None + + rng = np.random.default_rng(seed_i) + + if padding_mask is not None: + sz = all_sz - padding_mask[i].long().sum().item() + assert sz >= 0, sz + else: + sz = all_sz + + if num_mask_ver == 1: + if padding_mask is not None: + num_mask = int( + # add a random number for probabilistic rounding + mask_prob * sz / float(mask_length) + + np.random.rand() + ) + num_mask = max(min_masks, num_mask) + else: + num_mask = all_num_mask + elif num_mask_ver == 2: + num_mask = int( + # add a random number for probabilistic rounding + mask_prob * sz / float(mask_length) + + rng.random() + ) + num_mask = max(min_masks, num_mask) + else: + raise ValueError() + + if mask_type == "static": + lengths = np.full(num_mask, mask_length) + elif mask_type == "uniform": + lengths = rng.randint(mask_other, mask_length * 2 + 1, size=num_mask) + elif mask_type == "normal": + lengths = rng.normal(mask_length, mask_other, size=num_mask) + lengths = [max(1, int(round(x))) for x in lengths] + elif mask_type == "poisson": + lengths = rng.poisson(mask_length, size=num_mask) + lengths = [int(round(x)) for x in lengths] + else: + raise Exception("unknown mask selection " + mask_type) + + if sum(lengths) == 0: + if mask_type == "static": + raise ValueError(f"this should never happens") + else: + lengths = [min(mask_length, sz - 1)] + + if no_overlap: + mask_idc = [] + + def arrange(s, e, length, keep_length): + span_start = rng.randint(s, e - length) + mask_idc.extend(span_start + i for i in range(length)) + + new_parts = [] + if span_start - s - min_space >= keep_length: + new_parts.append((s, span_start - min_space + 1)) + if e - span_start - length - min_space > keep_length: + new_parts.append((span_start + length + min_space, e)) + return new_parts + + parts = [(0, sz)] + min_length = min(lengths) + for length in sorted(lengths, reverse=True): + lens = np.fromiter( + (e - s if e - s >= length + min_space else 0 for s, e in parts), + np.int, + ) + l_sum = np.sum(lens) + if l_sum == 0: + break + probs = lens / np.sum(lens) + c = rng.choice(len(parts), p=probs) + s, e = parts.pop(c) + parts.extend(arrange(s, e, length, min_length)) + mask_idc = np.asarray(mask_idc) + else: + if idc_select_ver == 1: + min_len = min(lengths) + if sz - min_len <= num_mask: + min_len = sz - num_mask - 1 + mask_idc = rng.choice(sz - min_len, num_mask, replace=False) + elif idc_select_ver == 2: + mask_idc = rng.choice(sz, num_mask, replace=False) + else: + raise ValueError() + + mask_idc = np.asarray( + [ + mask_idc[j] + offset + for j in range(len(mask_idc)) + for offset in range(lengths[j]) + ] + ) + + mask_idc = np.unique(mask_idc[mask_idc < sz]) + if len(mask_idc) >= sz: + raise ValueError( + ( + f"the entire sequence is masked. " + f"sz={sz}; mask_idc[mask_idc]; " + f"index={indices[i] if indices is not None else None}" + ) + ) + mask_idcs.append(mask_idc) + + target_len = None + if require_same_masks: + if add_masks: + target_len = max([len(m) for m in mask_idcs]) + else: + target_len = min([len(m) for m in mask_idcs]) + + for i, mask_idc in enumerate(mask_idcs): + if target_len is not None and len(mask_idc) > target_len: + mask_idc = rng.choice(mask_idc, target_len, replace=False) + + mask[i, mask_idc] = True + + if target_len is not None and len(mask_idc) < target_len: + unmasked = np.flatnonzero(~mask[i]) + to_mask = rng.choice(unmasked, target_len - len(mask_idc), replace=False) + mask[i, to_mask] = True + + if mask_dropout > 0: + masked = np.flatnonzero(mask[i]) + num_holes = np.rint(len(masked) * mask_dropout).astype(int) + to_drop = rng.choice(masked, num_holes, replace=False) + mask[i, to_drop] = False + + return mask + + +def add_hubert_arguments(parser: argparse.ArgumentParser): + parser.add_argument( + "--label-rate", + type=float, + default=50, + ) + + parser.add_argument( + "--sample-rate", + type=float, + default=16000, + ) + + parser.add_argument( + "--extractor-mode", + type=str, + default="default", + help="""mode for feature extractor, should in EXTRACTOR_MODE_CHOICES. default has a single group + norm with d groups in the first conv block, whereas layer_norm + has layer norms in every block (meant to use with normalize=True)""", + ) + parser.add_argument( + "--encoder-layers", + type=int, + default=12, + help="num encoder layers in the transformer", + ) + + parser.add_argument( + "--encoder-embed-dim", + type=int, + default=768, + help="encoder embedding dimension", + ) + + parser.add_argument( + "--encoder-ffn-embed-dim", + type=int, + default=3072, + help="encoder embedding dimension for FFN", + ) + + parser.add_argument( + "--encoder-attention-heads", + type=int, + default=12, + help="num encoder attention heads", + ) + + parser.add_argument( + "--activation-fn", + type=str, + choices=[ + "relu", + "gelu", + "gelu_fast", + "gelu_accurate", + "tanh", + "linear", + ], + default="gelu", + help="activation function to use", + ) + + parser.add_argument( + "--layer-type", + type=str, + choices=["transformer", "conformer", "trf_adp"], + default="transformer", + help="layer type in encoder", + ) + + # dropouts + parser.add_argument( + "--dropout", + type=float, + default=0.1, + help="dropout probability for the transformer", + ) + + parser.add_argument( + "--attention-dropout", + type=float, + default=0.1, + help="dropout probability for attention weights", + ) + + parser.add_argument( + "--activation-dropout", + type=float, + default=0.0, + help="dropout probability after activation in FFN", + ) + + parser.add_argument( + "--encoder-layerdrop", + type=float, + default=0.0, + help="probability of dropping a tarnsformer layer", + ) + + parser.add_argument( + "--dropout-input", + type=float, + default=0.0, + help="dropout to apply to the input (after feat extr)", + ) + + parser.add_argument( + "--dropout-features", + type=float, + default=0.0, + help="dropout to apply to the features (after feat extr)", + ) + + parser.add_argument( + "--final-dim", + type=int, + default=0, + help="project final representations and targets to this many dimensions. set to encoder_embed_dim is <= 0", + ) + + parser.add_argument( + "--untie-final-proj", + type=bool, + default=False, + help="use separate projection for each target", + ) + + parser.add_argument( + "--layer-norm-first", + type=bool, + default=False, + help="apply layernorm first in the transformer", + ) + + parser.add_argument( + "--conv-feature-layers", + type=str, + default="[(512,10,5)] + [(512,3,2)] * 4 + [(512,2,2)] * 2", + help="string describing convolutional feature extraction layers in form of a python list that contains [(dim, kernel_size, stride), ...]", + ) + + parser.add_argument( + "--conv-bias", + type=bool, + default=False, + help="include bias in conv encoder", + ) + + parser.add_argument( + "--logit-temp", + type=float, + default=0.1, + help="temperature to divide logits by", + ) + + parser.add_argument( + "--target-glu", + type=bool, + default=False, + help="adds projection + glu to targets", + ) + + parser.add_argument( + "--feature-grad-mult", + type=float, + default=1.0, + help="multiply feature extractor var grads by this", + ) + + # masking + parser.add_argument("--mask-length", type=int, default=10, help="mask_length") + + parser.add_argument( + "--mask-prob", + type=float, + default=0.65, + help="probability of replacing a token with mask", + ) + + parser.add_argument( + "--mask-selection", + type=str, + choices=["static", "uniform", "normal", "poisson"], + default="static", + help="how to choose mask length", + ) + + parser.add_argument( + "--mask-other", + type=float, + default=0, + help="secondary mask argument (used for more complex distributions),see help in compute_mask_indicesh", + ) + + parser.add_argument( + "--no-mask-overlap", + type=bool, + default=False, + help="whether to allow masks to overlap", + ) + + parser.add_argument( + "--mask-min-space", + type=int, + default=1, + help="min space between spans (if no overlap is enabled)", + ) + + # channel masking + parser.add_argument( + "--mask-channel-length", + type=int, + default=10, + help="length of the mask for features (channels)", + ) + + parser.add_argument( + "--mask-channel-prob", + type=float, + default=0.0, + help="probability of replacing a feature with 0", + ) + + parser.add_argument( + "--mask-channel-selection", + type=str, + choices=["static", "uniform", "normal", "poisson"], + default="static", + help="how to choose mask length for channel masking", + ) + + parser.add_argument( + "--mask-channel-other", + type=float, + default=0, + help="secondary mask argument (used for more complex distributions), see help in compute_mask_indicesh", + ) + + parser.add_argument( + "--no-mask-channel-overlap", + type=bool, + default=False, + help="whether to allow channel masks to overlap", + ) + + parser.add_argument( + "--mask-channel-min-space", + type=int, + default=1, + help="min space between spans (if no overlap is enabled)", + ) + + # positional embeddings + parser.add_argument( + "--conv-pos", + type=int, + default=128, + help="number of filters for convolutional positional embeddings", + ) + + parser.add_argument( + "--conv-pos-groups", + type=int, + default=16, + help="number of groups for convolutional positional embedding", + ) + + parser.add_argument( + "--conv-pos-batch-norm", + type=bool, + default=False, + help="use batch norm instead of weight norm in conv_pos (for bf16 models)", + ) + + parser.add_argument( + "--latent-temp", + type=float, + nargs="*", + default=[2, 0.5, 0.999995], + help="legacy (to be removed)", + ) + + # loss computation + parser.add_argument( + "--skip-masked", + type=bool, + default=False, + help="skip computing losses over masked frames", + ) + + parser.add_argument( + "--skip-nomask", + type=bool, + default=False, + help="skip computing losses over unmasked frames", + ) + + parser.add_argument( + "--checkpoint-activations", + type=bool, + default=False, + help="recompute activations and save memory for extra compute", + ) + + parser.add_argument( + "--pred-masked-weight", + type=float, + default=1, + help="weight for masked part in ssl loss", + ) + + parser.add_argument( + "--pred-nomask-weight", + type=float, + default=0, + help="weight for masked part in ssl loss", + ) + + parser.add_argument( + "--loss-weights", + type=float, + nargs="*", + default=[10], + help="weight for masked part in ssl loss", + ) + + # FP16 optimization + parser.add_argument( + "--required-seq-len-multiple", + type=int, + default=2, + help="pad the input to encoder such that the sequence length is divisible by multiple", + ) + + parser.add_argument( + "--attn-type", type=str, default="", help="if espnet use ESPNET MHA" + ) + + parser.add_argument( + "--pos-enc-type", + type=str, + default="abs", + help="Positional encoding type to use in conformer", + ) + + parser.add_argument( + "--num-classes", + type=int, + nargs="*", + default=[504], + help="""num class, a little larger than the number of cluster, + the largest is for padding, + and the value should be the multiple of 4, for faster computation""", + ) + + +class HubertModel(nn.Module): + def __init__( + self, + cfg, + ) -> None: + super().__init__() + feature_enc_layers = eval(cfg.conv_feature_layers) # noqa + self.embed = feature_enc_layers[-1][0] + + self.feature_extractor = ConvFeatureExtractionModel( + conv_layers=feature_enc_layers, + dropout=0.0, + mode=cfg.extractor_mode, + conv_bias=cfg.conv_bias, + ) + feature_ds_rate = np.prod([s for _, _, s in feature_enc_layers]) + self.feat2tar_ratio = cfg.label_rate * feature_ds_rate / cfg.sample_rate + + self.post_extract_proj = ( + nn.Linear(self.embed, cfg.encoder_embed_dim) + if self.embed != cfg.encoder_embed_dim + else None + ) + + self.mask_prob = cfg.mask_prob + self.mask_selection = cfg.mask_selection + self.mask_other = cfg.mask_other + self.mask_length = cfg.mask_length + self.no_mask_overlap = cfg.no_mask_overlap + self.mask_min_space = cfg.mask_min_space + + self.mask_channel_prob = cfg.mask_channel_prob + self.mask_channel_selection = cfg.mask_channel_selection + self.mask_channel_other = cfg.mask_channel_other + self.mask_channel_length = cfg.mask_channel_length + self.no_mask_channel_overlap = cfg.no_mask_channel_overlap + self.mask_channel_min_space = cfg.mask_channel_min_space + + self.dropout_input = nn.Dropout(cfg.dropout_input) + self.dropout_features = nn.Dropout(cfg.dropout_features) + + self.feature_grad_mult = cfg.feature_grad_mult + self.logit_temp = cfg.logit_temp + self.skip_masked = cfg.skip_masked + self.skip_nomask = cfg.skip_nomask + + self.mask_emb = nn.Parameter( + torch.FloatTensor(cfg.encoder_embed_dim).uniform_() + ) + + self.encoder = TransformerEncoder(cfg) + self.layer_norm = LayerNorm(self.embed) + + self.untie_final_proj = cfg.untie_final_proj + self.final_proj = nn.Linear(cfg.encoder_embed_dim, sum(cfg.num_classes)) + + # modules below are not needed during fine-tuning + self.num_classes = cfg.num_classes + self.pred_masked_weight = cfg.pred_masked_weight + self.pred_nomask_weight = cfg.pred_nomask_weight + self.loss_weights = cfg.loss_weights + + def upgrade_state_dict_named(self, state_dict, name): + """Upgrade a (possibly old) state dict for new versions of fairseq.""" + + super().upgrade_state_dict_named(state_dict, name) + return state_dict + + def apply_mask(self, x, padding_mask, target_list): + B, T, C = x.shape + if self.mask_prob > 0: + mask_indices = compute_mask_indices( + (B, T), + padding_mask, + self.mask_prob, + self.mask_length, + self.mask_selection, + self.mask_other, + min_masks=2, + no_overlap=self.no_mask_overlap, + min_space=self.mask_min_space, + ) + mask_indices = torch.from_numpy(mask_indices).to(x.device) + x[mask_indices] = self.mask_emb.to(x.dtype) + else: + mask_indices = None + + if self.mask_channel_prob > 0: + mask_channel_indices = compute_mask_indices( + (B, C), + None, + self.mask_channel_prob, + self.mask_channel_length, + self.mask_channel_selection, + self.mask_channel_other, + no_overlap=self.no_mask_channel_overlap, + min_space=self.mask_channel_min_space, + ) + mask_channel_indices = ( + torch.from_numpy(mask_channel_indices) + .to(x.device) + .unsqueeze(1) + .expand(-1, T, -1) + ) + x[mask_channel_indices] = 0 + + return x, mask_indices + + def forward_features(self, source: torch.Tensor) -> torch.Tensor: + if self.feature_grad_mult > 0: + features = self.feature_extractor(source) + if self.feature_grad_mult != 1.0: + features = GradMultiply.apply(features, self.feature_grad_mult) + else: + with torch.no_grad(): + features = self.feature_extractor(source) + return features + + def forward_targets( + self, + features: torch.Tensor, + target_list: List[torch.Tensor], + ) -> Tuple[torch.Tensor, torch.Tensor]: + # Trim features to ensure labels exist and then get aligned labels + feat_tsz = features.size(2) + targ_tsz = min([t.size(1) for t in target_list]) + if self.feat2tar_ratio * feat_tsz > targ_tsz: + feat_tsz = int(targ_tsz / self.feat2tar_ratio) + features = features[..., :feat_tsz] + target_inds = torch.arange(feat_tsz).float() * self.feat2tar_ratio + target_list = [t[:, target_inds.long()] for t in target_list] + return features, target_list + + def forward_padding_mask( + self, + features: torch.Tensor, + padding_mask: torch.Tensor, + ) -> torch.Tensor: + extra = padding_mask.size(1) % features.size(1) + if extra > 0: + padding_mask = padding_mask[:, :-extra] + padding_mask = padding_mask.view(padding_mask.size(0), features.size(1), -1) + padding_mask = padding_mask.all(-1) + return padding_mask + + def forward( + self, + source: torch.Tensor, + target_list: Optional[List[torch.Tensor]] = None, + padding_mask: Optional[torch.Tensor] = None, + mask: bool = True, + features_only: bool = False, + output_layer: Optional[int] = None, + ): + """output layer is 1-based""" + features = self.forward_features(source) + if target_list is not None: + features, target_list = self.forward_targets(features, target_list) + + features_pen = features.float().pow(2).mean() + + features = features.transpose(1, 2) + features = self.layer_norm(features) + unmasked_features = features.clone() + + if padding_mask is not None: + padding_mask = self.forward_padding_mask(features, padding_mask) + + if self.post_extract_proj is not None: + features = self.post_extract_proj(features) + + features = self.dropout_input(features) + unmasked_features = self.dropout_features(unmasked_features) + + if mask: + x, mask_indices = self.apply_mask(features, padding_mask, target_list) + else: + x = features + mask_indices = None + + # feature: (B, T, D), float + # target: (B, T), long + # x: (B, T, D), float + # padding_mask: (B, T), bool + # mask_indices: (B, T), bool + x, _ = self.encoder( + x, + padding_mask=padding_mask, + layer=None if output_layer is None else output_layer - 1, + ) + + if features_only: + return {"x": x, "padding_mask": padding_mask, "features": features} + + if not self.skip_masked: + masked_indices = torch.logical_and(~padding_mask, mask_indices) + proj_x_m = self.final_proj(x[masked_indices]) + proj_x_m /= self.logit_temp + logit_m_list = [proj_x_m for _ in range(len(target_list))] + else: + logit_m_list = [None for _ in target_list] + + if not self.skip_nomask: + nomask_indices = torch.logical_and(~padding_mask, ~mask_indices) + proj_x_u = self.final_proj(x[nomask_indices]) + proj_x_u /= self.logit_temp + logit_u_list = [proj_x_u for _ in range(len(target_list))] + else: + logit_u_list = [None for _ in target_list] + + # result = { + # "logit_m_list": logit_m_list, + # "logit_u_list": logit_u_list, + # "padding_mask": padding_mask, + # "features_pen": features_pen, + # } + targ_m_list = target_list[0][masked_indices] + targ_m_list = targ_m_list.long() + targ_m_list = [targ_m_list for _ in range(len(target_list))] + + targ_u_list = target_list[0][nomask_indices] + targ_u_list = targ_u_list.long() + targ_u_list = [targ_u_list for _ in range(len(target_list))] + return self.compute_loss( + logit_m_list, logit_u_list, targ_m_list, targ_u_list, features_pen + ) + + def extract_features( + self, + source: torch.Tensor, + padding_mask: Optional[torch.Tensor] = None, + mask: bool = False, + ret_conv: bool = False, + output_layer: Optional[int] = None, + ) -> Tuple[torch.Tensor, torch.Tensor]: + res = self.forward( + source, + padding_mask=padding_mask, + mask=mask, + features_only=True, + output_layer=output_layer, + ) + feature = res["features"] if ret_conv else res["x"] + return feature, res["padding_mask"] + + def get_logits(self, net_output, is_masked=True): + if is_masked: + logits_list = net_output["logit_m_list"] + else: + logits_list = net_output["logit_u_list"] + logits_list = [x.float() for x in logits_list if x is not None] + return logits_list + + def get_targets(self, net_output, is_masked=True): + logits_list = self.get_logits(net_output, is_masked) + targets_list = [x.new_zeros(x.size(0), dtype=torch.long) for x in logits_list] + return targets_list + + def get_extra_losses(self, net_output): + extra_losses = [] + names = [] + + if "features_pen" in net_output: + extra_losses.append(net_output["features_pen"]) + names.append("features_pen") + + return extra_losses, names + + def remove_pretraining_modules(self): + self.final_proj = None + + def compute_loss( + self, logit_m_list, logit_u_list, targ_m_list, targ_u_list, features_pen + ): + loss = 0.0 + sample_size = 0 + logging_output = {} + reduce = True + reduction = "sum" if reduce else "none" + + loss_m_list = [] + logp_m_list = [x.float() for x in logit_m_list if x is not None] + logp_m_list = torch.cat(logp_m_list) + targ_m_list = torch.cat(targ_m_list) + + loss_m = F.cross_entropy(logp_m_list, targ_m_list, reduction=reduction) + loss_m_list.append(loss_m) + logging_output[f"loss_m_0"] = loss_m.detach().item() + + assert self.pred_masked_weight == 0 or len(logp_m_list) > 0 + if self.pred_masked_weight > 0: + loss += self.pred_masked_weight * sum(loss_m_list) + sample_size += len(targ_m_list) + + loss_u_list = [] + logp_u_list = [x.float() for x in logit_u_list if x is not None] + logp_u_list = torch.cat(logp_u_list) + targ_u_list = torch.cat(targ_u_list) + + loss_u = F.cross_entropy(logp_u_list, targ_u_list, reduction=reduction) + loss_u_list.append(loss_u) + logging_output[f"loss_u_0"] = loss_u.detach().item() + + assert self.pred_nomask_weight == 0 or len(logp_u_list) > 0 + if self.pred_nomask_weight > 0: + loss += self.pred_nomask_weight * sum(loss_u_list) + sample_size += len(targ_u_list) + + if self.loss_weights is not None: + extra_losses = [] + names = [] + extra_losses.append(features_pen) + names.append("features_pen") + if torch.is_tensor(extra_losses): + extra_losses = [extra_losses] + names = [names] + if len(self.loss_weights) == 1 and len(extra_losses) != 1: + self.loss_weights = [self.loss_weights[0]] * len(extra_losses) + assert len(extra_losses) == len( + self.loss_weights + ), f"{len(extra_losses)}, {len(self.loss_weights)}" + for p, n, coef in zip(extra_losses, names, self.loss_weights): + if coef != 0 and p is not None: + p = coef * p.float() * sample_size + loss += p + logging_output[f"loss_{n}"] = p.item() + + logging_output = { + "loss": loss.item() if reduce else loss, + **logging_output, + } + + # for lk in self.log_keys: + # if lk in net_output: + # logging_output[lk] = float((net_output[lk])) + + def compute_correct(logits, target): + if logits.numel() == 0: + return 0, 0 + else: + assert logits.dim() > 1, logits.shape + max = logits.argmax(-1) == target + min = logits.argmin(-1) == target + both = max & min + corr = max.long().sum().item() - both.long().sum().item() + count = max.numel() + return corr, count + + with torch.no_grad(): + corr_m, count_m = compute_correct(logp_m_list, targ_m_list) + logging_output[f"correct_m_0"] = corr_m + logging_output[f"count_m_0"] = count_m + + corr_u, count_u = compute_correct(logp_u_list, targ_u_list) + logging_output[f"correct_u_0"] = corr_u + logging_output[f"count_u_0"] = count_u + + return loss, sample_size, logging_output diff --git a/egs/librispeech/SSL/hubert/joiner.py b/egs/librispeech/SSL/hubert/joiner.py new file mode 120000 index 0000000000..aa3362cda4 --- /dev/null +++ b/egs/librispeech/SSL/hubert/joiner.py @@ -0,0 +1 @@ +../../ASR/zipformer/joiner.py \ No newline at end of file diff --git a/egs/librispeech/SSL/hubert/model.py b/egs/librispeech/SSL/hubert/model.py new file mode 100644 index 0000000000..46a968b69e --- /dev/null +++ b/egs/librispeech/SSL/hubert/model.py @@ -0,0 +1,344 @@ +# Copyright 2021-2024 Xiaomi Corp. (authors: Fangjun Kuang, +# Wei Kang, +# Zengwei Yao, +# Yifan Yang) +# +# See ../../../../LICENSE for clarification regarding multiple authors +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +from typing import Optional, Tuple + +import k2 +import torch +import torch.nn as nn +from scaling import ScaledLinear + +from icefall.utils import add_sos + + +class AsrModel(nn.Module): + def __init__( + self, + encoder, + decoder: Optional[nn.Module] = None, + joiner: Optional[nn.Module] = None, + encoder_dim: int = 768, + decoder_dim: int = 512, + vocab_size: int = 500, + use_transducer: bool = True, + use_ctc: bool = False, + ): + """A joint CTC & Transducer ASR model. + + - Connectionist temporal classification: labelling unsegmented sequence data with recurrent neural networks (http://imagine.enpc.fr/~obozinsg/teaching/mva_gm/papers/ctc.pdf) + - Sequence Transduction with Recurrent Neural Networks (https://arxiv.org/pdf/1211.3711.pdf) + - Pruned RNN-T for fast, memory-efficient ASR training (https://arxiv.org/pdf/2206.13236.pdf) + + Args: + encoder: + It is the transcription network in the paper. Its accepts + inputs: `x` of (N, T, encoder_dim). + It returns two tensors: `logits` of shape (N, T, encoder_dim) and + `logit_lens` of shape (N,). + decoder: + It is the prediction network in the paper. Its input shape + is (N, U) and its output shape is (N, U, decoder_dim). + It should contain one attribute: `blank_id`. + It is used when use_transducer is True. + joiner: + It has two inputs with shapes: (N, T, encoder_dim) and (N, U, decoder_dim). + Its output shape is (N, T, U, vocab_size). Note that its output contains + unnormalized probs, i.e., not processed by log-softmax. + It is used when use_transducer is True. + use_transducer: + Whether use transducer head. Default: True. + use_ctc: + Whether use CTC head. Default: False. + """ + super().__init__() + + assert ( + use_transducer or use_ctc + ), f"At least one of them should be True, but got use_transducer={use_transducer}, use_ctc={use_ctc}" + + self.encoder = encoder + + self.use_transducer = use_transducer + if use_transducer: + # Modules for Transducer head + assert decoder is not None + assert hasattr(decoder, "blank_id") + assert joiner is not None + + self.decoder = decoder + self.joiner = joiner + + self.simple_am_proj = ScaledLinear( + encoder_dim, vocab_size, initial_scale=0.25 + ) + self.simple_lm_proj = ScaledLinear( + decoder_dim, vocab_size, initial_scale=0.25 + ) + else: + assert decoder is None + assert joiner is None + + self.use_ctc = use_ctc + if use_ctc: + # Modules for CTC head + self.ctc_output = nn.Sequential( + nn.Dropout(p=0.1), + nn.Linear(encoder_dim, vocab_size), + nn.LogSoftmax(dim=-1), + ) + + def forward_encoder( + self, + x: torch.Tensor, + padding_mask: Optional[torch.Tensor] = None, + ) -> Tuple[torch.Tensor, torch.Tensor]: + """Compute encoder outputs. + Args: + x: + A 2-D tensor of shape (N, T). + + Returns: + encoder_out: + Encoder output, of shape (N, T, C). + encoder_out_lens: + Encoder output lengths, of shape (N,). + """ + if padding_mask is None: + padding_mask = torch.zeros_like(x, dtype=torch.bool) + + encoder_out, padding_mask = self.encoder.extract_features( + source=x, + padding_mask=padding_mask, + mask=self.encoder.training, + ) + encoder_out_lens = torch.sum(~padding_mask, dim=1) + assert torch.all(encoder_out_lens > 0), encoder_out_lens + + return encoder_out, encoder_out_lens + + def forward_ctc( + self, + encoder_out: torch.Tensor, + encoder_out_lens: torch.Tensor, + targets: torch.Tensor, + target_lengths: torch.Tensor, + ) -> torch.Tensor: + """Compute CTC loss. + Args: + encoder_out: + Encoder output, of shape (N, T, C). + encoder_out_lens: + Encoder output lengths, of shape (N,). + targets: + Target Tensor of shape (sum(target_lengths)). The targets are assumed + to be un-padded and concatenated within 1 dimension. + """ + # Compute CTC log-prob + ctc_output = self.ctc_output(encoder_out) # (N, T, C) + + ctc_loss = torch.nn.functional.ctc_loss( + log_probs=ctc_output.permute(1, 0, 2), # (T, N, C) + targets=targets, + input_lengths=encoder_out_lens, + target_lengths=target_lengths, + reduction="sum", + ) + return ctc_loss + + def forward_transducer( + self, + encoder_out: torch.Tensor, + encoder_out_lens: torch.Tensor, + y: k2.RaggedTensor, + y_lens: torch.Tensor, + prune_range: int = 5, + am_scale: float = 0.0, + lm_scale: float = 0.0, + ) -> Tuple[torch.Tensor, torch.Tensor]: + """Compute Transducer loss. + Args: + encoder_out: + Encoder output, of shape (N, T, C). + encoder_out_lens: + Encoder output lengths, of shape (N,). + y: + A ragged tensor with 2 axes [utt][label]. It contains labels of each + utterance. + prune_range: + The prune range for rnnt loss, it means how many symbols(context) + we are considering for each frame to compute the loss. + am_scale: + The scale to smooth the loss with am (output of encoder network) + part + lm_scale: + The scale to smooth the loss with lm (output of predictor network) + part + """ + # Now for the decoder, i.e., the prediction network + blank_id = self.decoder.blank_id + sos_y = add_sos(y, sos_id=blank_id) + + # sos_y_padded: [B, S + 1], start with SOS. + sos_y_padded = sos_y.pad(mode="constant", padding_value=blank_id) + + # decoder_out: [B, S + 1, decoder_dim] + decoder_out = self.decoder(sos_y_padded) + + # Note: y does not start with SOS + # y_padded : [B, S] + y_padded = y.pad(mode="constant", padding_value=0) + + y_padded = y_padded.to(torch.int64) + boundary = torch.zeros( + (encoder_out.size(0), 4), + dtype=torch.int64, + device=encoder_out.device, + ) + boundary[:, 2] = y_lens + boundary[:, 3] = encoder_out_lens + + lm = self.simple_lm_proj(decoder_out) + am = self.simple_am_proj(encoder_out) + + # if self.training and random.random() < 0.25: + # lm = penalize_abs_values_gt(lm, 100.0, 1.0e-04) + # if self.training and random.random() < 0.25: + # am = penalize_abs_values_gt(am, 30.0, 1.0e-04) + + with torch.cuda.amp.autocast(enabled=False): + simple_loss, (px_grad, py_grad) = k2.rnnt_loss_smoothed( + lm=lm.float(), + am=am.float(), + symbols=y_padded, + termination_symbol=blank_id, + lm_only_scale=lm_scale, + am_only_scale=am_scale, + boundary=boundary, + reduction="sum", + return_grad=True, + ) + + # ranges : [B, T, prune_range] + ranges = k2.get_rnnt_prune_ranges( + px_grad=px_grad, + py_grad=py_grad, + boundary=boundary, + s_range=prune_range, + ) + + # am_pruned : [B, T, prune_range, encoder_dim] + # lm_pruned : [B, T, prune_range, decoder_dim] + am_pruned, lm_pruned = k2.do_rnnt_pruning( + am=self.joiner.encoder_proj(encoder_out), + lm=self.joiner.decoder_proj(decoder_out), + ranges=ranges, + ) + + # logits : [B, T, prune_range, vocab_size] + + # project_input=False since we applied the decoder's input projections + # prior to do_rnnt_pruning (this is an optimization for speed). + logits = self.joiner(am_pruned, lm_pruned, project_input=False) + + with torch.cuda.amp.autocast(enabled=False): + pruned_loss = k2.rnnt_loss_pruned( + logits=logits.float(), + symbols=y_padded, + ranges=ranges, + termination_symbol=blank_id, + boundary=boundary, + reduction="sum", + ) + + return simple_loss, pruned_loss + + def forward( + self, + x: torch.Tensor, + y: k2.RaggedTensor, + padding_mask: Optional[torch.Tensor] = None, + prune_range: int = 5, + am_scale: float = 0.0, + lm_scale: float = 0.0, + ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]: + """ + Args: + x: + A 2-D tensor of shape (N, T). + y: + A ragged tensor with 2 axes [utt][label]. It contains labels of each + utterance. + prune_range: + The prune range for rnnt loss, it means how many symbols(context) + we are considering for each frame to compute the loss. + am_scale: + The scale to smooth the loss with am (output of encoder network) + part + lm_scale: + The scale to smooth the loss with lm (output of predictor network) + part + Returns: + Return the transducer losses and CTC loss, + in form of (simple_loss, pruned_loss, ctc_loss) + + Note: + Regarding am_scale & lm_scale, it will make the loss-function one of + the form: + lm_scale * lm_probs + am_scale * am_probs + + (1-lm_scale-am_scale) * combined_probs + """ + assert x.ndim == 2, x.shape + assert y.num_axes == 2, y.num_axes + + assert x.size(0) == y.dim0, (x.shape, y.dim0) + + # Compute encoder outputs + encoder_out, encoder_out_lens = self.forward_encoder(x, padding_mask) + + row_splits = y.shape.row_splits(1) + y_lens = row_splits[1:] - row_splits[:-1] + + if self.use_transducer: + # Compute transducer loss + simple_loss, pruned_loss = self.forward_transducer( + encoder_out=encoder_out, + encoder_out_lens=encoder_out_lens, + y=y.to(x.device), + y_lens=y_lens, + prune_range=prune_range, + am_scale=am_scale, + lm_scale=lm_scale, + ) + else: + simple_loss = torch.empty(0) + pruned_loss = torch.empty(0) + + if self.use_ctc: + # Compute CTC loss + targets = y.values + ctc_loss = self.forward_ctc( + encoder_out=encoder_out, + encoder_out_lens=encoder_out_lens, + targets=targets, + target_lengths=y_lens, + ) + else: + ctc_loss = torch.empty(0) + + return simple_loss, pruned_loss, ctc_loss, encoder_out_lens diff --git a/egs/librispeech/SSL/hubert/optim.py b/egs/librispeech/SSL/hubert/optim.py new file mode 120000 index 0000000000..56b827b8ae --- /dev/null +++ b/egs/librispeech/SSL/hubert/optim.py @@ -0,0 +1 @@ +../../ASR/zipformer/optim.py \ No newline at end of file diff --git a/egs/librispeech/SSL/hubert/pretrain.py b/egs/librispeech/SSL/hubert/pretrain.py new file mode 100644 index 0000000000..d9bda88570 --- /dev/null +++ b/egs/librispeech/SSL/hubert/pretrain.py @@ -0,0 +1,1082 @@ +#!/usr/bin/env python3 +# Copyright 2021-2024 Xiaomi Corp. (authors: Fangjun Kuang, +# Wei Kang, +# Mingshuang Luo, +# Zengwei Yao, +# Yifan Yang, +# Daniel Povey) +# Copyright 2024 Shanghai Jiao Tong University (authors: Jianheng Zhuo) +# +# 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="0,1,2,3,4,5,6,7" + +# For hubert model pretraining: +./hubert/pretrain.py \ + --world-size 8 \ + --num-epochs 400 \ + --start-epoch 1 \ + --use-fp16 1 \ + --exp-dir hubert/exp \ + --full-libri 1 \ + --max-duration 87.5 \ + --accum-grad 4 +""" + + +import argparse +import copy +import logging +import sys +import warnings +from pathlib import Path +from shutil import copyfile +from typing import Any, Dict, Optional, Tuple, Union + +import k2 +import optim +import torch +import torch.multiprocessing as mp +import torch.nn as nn +from hubert import HubertModel, add_hubert_arguments +from lhotse.cut import Cut +from lhotse.dataset.sampling.base import CutSampler +from lhotse.utils import fix_random_seed +from optim import Eden, ScaledAdam +from ssl_datamodule import LibriSpeechDataModule +from torch import Tensor +from torch.cuda.amp import GradScaler +from torch.nn.functional import pad +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.hooks import register_inf_check_hooks +from icefall.utils import ( + AttributeDict, + MetricsTracker, + get_parameter_groups_with_lrs, + setup_logger, + str2bool, +) + +LRSchedulerType = Union[torch.optim.lr_scheduler._LRScheduler, optim.LRScheduler] + + +def get_adjusted_batch_count(params: AttributeDict) -> float: + # returns the number of batches we would have used so far if we had used the reference + # duration. This is for purposes of set_batch_count(). + return ( + params.batch_idx_train + * params.accum_grad + * (params.max_duration * params.world_size) + / params.ref_duration + ) + + +def set_batch_count(model: Union[nn.Module, DDP], batch_count: float) -> None: + if isinstance(model, DDP): + # get underlying nn.Module + model = model.module + for name, module in model.named_modules(): + if hasattr(module, "batch_count"): + module.batch_count = batch_count + if hasattr(module, "name"): + module.name = name + + +def 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=400, + 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="hubert/exp", + help="""The experiment dir. + It specifies the directory where all training related + files, e.g., checkpoints, log, etc, are saved + """, + ) + + parser.add_argument( + "--base-lr", type=float, default=0.045, help="The base learning rate." + ) + + parser.add_argument( + "--lr-batches", + type=float, + default=7500, + help="""Number of steps that affects how rapidly the learning rate + decreases. We suggest not to change this.""", + ) + + parser.add_argument( + "--lr-epochs", + type=float, + default=10.5, + help="""Number of epochs that affects how rapidly the learning rate decreases. + """, + ) + + parser.add_argument( + "--warmup-batches", + type=float, + default=5000, + help="Eden warmup steps", + ) + + parser.add_argument( + "--warmup-start", + type=float, + default=0, + help="Eden warmup start learning rate", + ) + + parser.add_argument( + "--ref-duration", + type=float, + default=80, + help="Reference batch duration for purposes of adjusting batch counts for setting various " + "schedules inside the model", + ) + + 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( + "--sanity-check", + type=str2bool, + default=False, + help="Check if any of the batches in epoch 1 would cause OOM.", + ) + + parser.add_argument( + "--inf-check", + type=str2bool, + default=False, + help="Add hooks to check for infinite module outputs and gradients.", + ) + + parser.add_argument( + "--save-every-n", + type=int, + default=100000, + help="""Save checkpoint after processing this number of batches" + periodically. We save checkpoint to exp-dir/ whenever + params.batch_idx_train % save_every_n == 0. The checkpoint filename + has the form: f'exp-dir/checkpoint-{params.batch_idx_train}.pt' + Note: It also saves checkpoint to `exp-dir/epoch-xxx.pt` at the + end of each epoch where `xxx` is the epoch number counting from 1. + """, + ) + + parser.add_argument( + "--keep-last-k", + type=int, + default=30, + help="""Only keep this number of checkpoints on disk. + For instance, if it is 3, there are only 3 checkpoints + in the exp-dir with filenames `checkpoint-xxx.pt`. + It does not affect checkpoints with name `epoch-xxx.pt`. + """, + ) + + parser.add_argument( + "--average-period", + type=int, + default=200, + help="""Update the averaged model, namely `model_avg`, after processing + this number of batches. `model_avg` is a separate version of model, + in which each floating-point parameter is the average of all the + parameters from the start of training. Each time we take the average, + we do: `model_avg = model * (average_period / batch_idx_train) + + model_avg * ((batch_idx_train - average_period) / batch_idx_train)`. + """, + ) + + parser.add_argument( + "--accum-grad", + type=int, + default=4, + help="""update gradient when batch_idx_train % accum_grad == 0. + """, + ) + + parser.add_argument( + "--use-fp16", + type=str2bool, + default=False, + help="Whether to use half precision training.", + ) + + parser.add_argument( + "--max-keep-size", + type=int, + default=sys.maxsize, + help="exclude sample longer than this.", + ) + + parser.add_argument( + "--min-keep-size", + type=float, + default=32000, + help="exclude sample longer less than this.", + ) + + parser.add_argument( + "--max-sample-size", + type=float, + default=250000, + help="max sample size to crop to for batching.", + ) + + add_hubert_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 updates happen to the model so far across + epochs. + + - sub_batch_idx_train: It contains number of batch 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 + """ + params = AttributeDict( + { + "best_train_loss": float("inf"), + "best_valid_loss": float("inf"), + "best_train_epoch": -1, + "best_valid_epoch": -1, + "batch_idx_train": 0, + "sub_batch_idx_train": 0, + "log_interval": 50, + "reset_interval": 200, + "valid_interval": 3000, # For the 100h subset, use 800 + "env_info": get_env_info(), + } + ) + + return params + + +def _to_int_tuple(s: str): + return tuple(map(int, s.split(","))) + + +def get_model(params: AttributeDict) -> nn.Module: + model = HubertModel(params) + 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], + batch: dict, + is_training: bool, +) -> Tuple[Tensor, MetricsTracker]: + """ + Compute loss given the model and its inputs. + + Args: + params: + Parameters for training. See :func:`get_params`. + model: + The model for training. It is an instance of Zipformer in our case. + batch: + A batch of data. See `dataset.HubertDataset()` + for the content in it. + is_training: + True for training. False for validation. When it is True, this + function enables autograd during computation; when it is False, it + disables autograd. + """ + device = model.device if isinstance(model, DDP) else next(model.parameters()).device + audio = batch["audio"].to(device) + padding_mask = batch["padding_mask"].to(device) + kmeans = batch["kmeans"].to(device) + + with torch.set_grad_enabled(is_training): + loss, num_masked_tokens, logging_output = model( + source=audio, target_list=[kmeans], padding_mask=padding_mask + ) + + assert loss.requires_grad == is_training + + info = MetricsTracker() + with warnings.catch_warnings(): + warnings.simplefilter("ignore") + info["frames"] = num_masked_tokens + for item in logging_output: + info[item] = logging_output[item] + return loss, info + + +def compute_validation_loss( + params: AttributeDict, + model: Union[nn.Module, DDP], + valid_dl: torch.utils.data.DataLoader, + world_size: int = 1, +) -> MetricsTracker: + """Run the validation process.""" + model.eval() + + tot_loss = MetricsTracker() + + for batch_idx, batch in enumerate(valid_dl): + loss, loss_info = compute_loss( + params=params, + model=model, + batch=batch, + is_training=False, + ) + assert loss.requires_grad is False + tot_loss = tot_loss + loss_info + + if world_size > 1: + tot_loss.reduce(loss.device) + + loss_value = tot_loss["loss"] / tot_loss["frames"] + if loss_value < params.best_valid_loss: + params.best_valid_epoch = params.cur_epoch + params.best_valid_loss = loss_value + + return tot_loss + + +def train_one_epoch( + params: AttributeDict, + model: Union[nn.Module, DDP], + optimizer: torch.optim.Optimizer, + scheduler: LRSchedulerType, + train_dl: torch.utils.data.DataLoader, + valid_dl: torch.utils.data.DataLoader, + scaler: GradScaler, + model_avg: Optional[nn.Module] = None, + tb_writer: Optional[SummaryWriter] = None, + world_size: int = 1, + rank: int = 0, +) -> None: + """Train the model for one epoch. + + The training loss from the mean of all frames is saved in + `params.train_loss`. It runs the validation process every + `params.valid_interval` batches. + + Args: + params: + It is returned by :func:`get_params`. + model: + The model for training. + optimizer: + The optimizer we are using. + scheduler: + The learning rate scheduler, we call step() every step. + train_dl: + Dataloader for the training dataset. + valid_dl: + Dataloader for the validation dataset. + scaler: + The scaler used for mix precision training. + model_avg: + The stored model averaged from the start of training. + tb_writer: + Writer to write log messages to tensorboard. + world_size: + Number of nodes in DDP training. If it is 1, DDP is disabled. + rank: + The rank of the node in DDP training. If no DDP is used, it should + be set to 0. + """ + model.train() + + tot_loss = MetricsTracker() + + saved_bad_model = False + + def save_bad_model(suffix: str = ""): + save_checkpoint_impl( + filename=params.exp_dir / f"bad-model{suffix}-{rank}.pt", + model=model, + model_avg=model_avg, + params=params, + optimizer=optimizer, + scheduler=scheduler, + sampler=train_dl.sampler, + scaler=scaler, + rank=0, + ) + + for sub_batch_idx, batch in enumerate(train_dl): + params.sub_batch_idx_train += 1 + batch_idx = sub_batch_idx // params.accum_grad + + if batch_idx % 10 == 0: + set_batch_count(model, get_adjusted_batch_count(params)) + + batch_size = batch["kmeans"].shape[0] + + try: + with torch.cuda.amp.autocast(enabled=params.use_fp16): + loss, loss_info = compute_loss( + params=params, + model=model, + batch=batch, + is_training=True, + ) + # summary stats + tot_loss = (tot_loss * (1 - 1 / params.reset_interval)) + loss_info + + # NOTE: We use reduction==sum and loss is computed over utterances + # in the batch and there is no normalization to it so far. + scaler.scale(loss / params.accum_grad).backward() + + if sub_batch_idx % params.accum_grad == params.accum_grad - 1: + params.batch_idx_train += 1 + scheduler.step_batch(params.batch_idx_train) + + scaler.step(optimizer) + scaler.update() + optimizer.zero_grad() + else: + continue + + except: # noqa + save_bad_model() + display_and_save_batch(batch, params=params) + raise + + if params.print_diagnostics and batch_idx == 5: + return + + if ( + rank == 0 + and params.batch_idx_train > 0 + and params.batch_idx_train % params.average_period == 0 + ): + update_averaged_model( + params=params, + model_cur=model, + model_avg=model_avg, + ) + + if ( + params.batch_idx_train > 0 + and params.batch_idx_train % params.save_every_n == 0 + ): + save_checkpoint_with_global_batch_idx( + out_dir=params.exp_dir, + global_batch_idx=params.batch_idx_train, + model=model, + model_avg=model_avg, + params=params, + optimizer=optimizer, + scheduler=scheduler, + sampler=train_dl.sampler, + scaler=scaler, + rank=rank, + ) + remove_checkpoints( + out_dir=params.exp_dir, + topk=params.keep_last_k, + rank=rank, + ) + + if batch_idx % 100 == 0 and params.use_fp16: + # If the grad scale was less than 1, try increasing it. The _growth_interval + # of the grad scaler is configurable, but we can't configure it to have different + # behavior depending on the current grad scale. + cur_grad_scale = scaler._scale.item() + + if cur_grad_scale < 8.0 or (cur_grad_scale < 32.0 and batch_idx % 400 == 0): + scaler.update(cur_grad_scale * 2.0) + if cur_grad_scale < 0.01: + if not saved_bad_model: + save_bad_model(suffix="-first-warning") + saved_bad_model = True + logging.warning(f"Grad scale is small: {cur_grad_scale}") + if cur_grad_scale < 1.0e-05: + save_bad_model() + raise RuntimeError( + f"grad_scale is too small, exiting: {cur_grad_scale}" + ) + + if batch_idx % params.log_interval == 0: + cur_lr = max(scheduler.get_last_lr()) + cur_grad_scale = scaler._scale.item() if params.use_fp16 else 1.0 + + logging.info( + f"Epoch {params.cur_epoch}, " + f"batch {batch_idx}, loss[{loss_info}], " + f"tot_loss[{tot_loss}], batch size: {batch_size}, " + f"lr: {cur_lr:.2e}, " + + (f"grad_scale: {scaler._scale.item()}" if params.use_fp16 else "") + ) + + if tb_writer is not None: + tb_writer.add_scalar( + "train/learning_rate", cur_lr, params.batch_idx_train + ) + + loss_info.write_summary( + tb_writer, "train/current_", params.batch_idx_train + ) + tot_loss.write_summary(tb_writer, "train/tot_", params.batch_idx_train) + if params.use_fp16: + tb_writer.add_scalar( + "train/grad_scale", + cur_grad_scale, + params.batch_idx_train, + ) + + if batch_idx % params.valid_interval == 0 and not params.print_diagnostics: + logging.info("Computing validation loss") + valid_info = compute_validation_loss( + params=params, + model=model, + valid_dl=valid_dl, + world_size=world_size, + ) + model.train() + logging.info(f"Epoch {params.cur_epoch}, validation: {valid_info}") + logging.info( + f"Maximum memory allocated so far is {torch.cuda.max_memory_allocated()//1000000}MB" + ) + if tb_writer is not None: + valid_info.write_summary( + tb_writer, "train/valid_", params.batch_idx_train + ) + + if batch_idx % params.accum_grad != params.accum_grad - 1: + optimizer.zero_grad() + loss_value = tot_loss["loss"] / tot_loss["frames"] + params.train_loss = loss_value + if params.train_loss < params.best_train_loss: + params.best_train_epoch = params.cur_epoch + params.best_train_loss = params.train_loss + + +def run(rank, world_size, args): + """ + Args: + rank: + It is a value between 0 and `world_size-1`, which is + passed automatically by `mp.spawn()` in :func:`main`. + The node with rank 0 is responsible for saving checkpoint. + world_size: + Number of GPUs for DDP training. + args: + The return value of get_parser().parse_args() + """ + params = get_params() + params.update(vars(args)) + + fix_random_seed(params.seed) + if world_size > 1: + setup_dist(rank, world_size, params.master_port) + + setup_logger(f"{params.exp_dir}/log/log-train") + logging.info("Training started") + + if args.tensorboard and rank == 0: + tb_writer = SummaryWriter(log_dir=f"{params.exp_dir}/tensorboard") + else: + tb_writer = None + + device = torch.device("cpu") + if torch.cuda.is_available(): + device = torch.device("cuda", rank) + logging.info(f"Device: {device}") + logging.info(params) + + logging.info("About to create model") + model = get_model(params) + + num_param = sum([p.numel() for p in model.parameters()]) + logging.info(f"Number of model parameters: {num_param}") + + assert params.save_every_n >= params.average_period + model_avg: Optional[nn.Module] = None + if rank == 0: + # model_avg is only used with rank 0 + model_avg = copy.deepcopy(model).to(torch.float64) + + assert params.start_epoch > 0, params.start_epoch + checkpoints = load_checkpoint_if_available( + params=params, model=model, model_avg=model_avg + ) + + model.to(device) + if world_size > 1: + logging.info("Using DDP") + model = DDP(model, device_ids=[rank], find_unused_parameters=True) + + optimizer = ScaledAdam( + get_parameter_groups_with_lrs(model, lr=params.base_lr, include_names=True), + lr=params.base_lr, # should have no effect + clipping_scale=2.0, + ) + + scheduler = Eden( + optimizer, + params.lr_batches, + params.lr_epochs, + params.warmup_batches, + params.warmup_start, + ) + + if checkpoints and "optimizer" in checkpoints: + logging.info("Loading optimizer state dict") + optimizer.load_state_dict(checkpoints["optimizer"]) + + if ( + checkpoints + and "scheduler" in checkpoints + and checkpoints["scheduler"] is not None + ): + logging.info("Loading scheduler state dict") + scheduler.load_state_dict(checkpoints["scheduler"]) + + if params.print_diagnostics: + opts = diagnostics.TensorDiagnosticOptions( + 512 + ) # allow 4 megabytes per sub-module + diagnostic = diagnostics.attach_diagnostics(model, opts) + + if params.inf_check: + register_inf_check_hooks(model) + + librispeech = LibriSpeechDataModule(args) + + train_cuts = ( + librispeech.train_all_shuf_cuts() + if params.full_libri + else librispeech.train_clean_100_cuts() + ) + + def remove_short_and_long_utt(c: Cut): + # Keep only utterances with duration between 1 second and 20 seconds + # + # Caution: There is a reason to select 20.0 here. Please see + # ../local/display_manifest_statistics.py + # + # You should use ../local/display_manifest_statistics.py to get + # an utterance duration distribution for your dataset to select + # the threshold + if ( + c.duration < params.min_keep_size / params.sample_rate + or c.duration > params.max_keep_size / params.sample_rate + ): + logging.warning( + f"Exclude cut with ID {c.id} from training. Duration: {c.duration}" + ) + return False + + return True + + train_cuts = train_cuts.filter(remove_short_and_long_utt) + + if params.start_batch > 0 and checkpoints and "sampler" in checkpoints: + # We only load the sampler's state dict when it loads a checkpoint + # saved in the middle of an epoch + sampler_state_dict = checkpoints["sampler"] + else: + sampler_state_dict = None + + train_dl = librispeech.train_dataloaders( + train_cuts, + max_sample_size=params.max_sample_size, + sample_rate=params.sample_rate, + label_rate=params.label_rate, + random_crop=params.random_crop, + pad_audio=False, + num_classes=params.num_classes, + do_normalize=params.do_normalize, + sampler_state_dict=sampler_state_dict, + ) + + valid_cuts = librispeech.dev_clean_cuts() + # valid_cuts += librispeech.dev_other_cuts() + valid_cuts = valid_cuts.filter(remove_short_and_long_utt) + + valid_dl = librispeech.valid_dataloaders( + valid_cuts, + max_sample_size=params.max_sample_size, + sample_rate=params.sample_rate, + label_rate=params.label_rate, + random_crop=params.random_crop, + pad_audio=False, + num_classes=params.num_classes, + do_normalize=params.do_normalize, + ) + + if params.sanity_check and not params.print_diagnostics: + scan_pessimistic_batches_for_oom( + model=model, + train_dl=train_dl, + optimizer=optimizer, + params=params, + ) + + scaler = GradScaler(enabled=params.use_fp16, init_scale=1.0) + if checkpoints and "grad_scaler" in checkpoints: + logging.info("Loading grad scaler state dict") + scaler.load_state_dict(checkpoints["grad_scaler"]) + + for epoch in range(params.start_epoch, params.num_epochs + 1): + scheduler.step_epoch(epoch - 1) + fix_random_seed(params.seed + epoch - 1) + train_dl.sampler.set_epoch(epoch - 1) + + if tb_writer is not None: + tb_writer.add_scalar("train/epoch", epoch, params.batch_idx_train) + + params.cur_epoch = epoch + + train_one_epoch( + params=params, + model=model, + model_avg=model_avg, + optimizer=optimizer, + scheduler=scheduler, + train_dl=train_dl, + valid_dl=valid_dl, + scaler=scaler, + tb_writer=tb_writer, + world_size=world_size, + rank=rank, + ) + + if params.print_diagnostics: + diagnostic.print_diagnostics() + break + + save_checkpoint( + params=params, + model=model, + model_avg=model_avg, + optimizer=optimizer, + scheduler=scheduler, + sampler=train_dl.sampler, + scaler=scaler, + rank=rank, + ) + + logging.info("Done!") + + if world_size > 1: + torch.distributed.barrier() + cleanup_dist() + + +def display_and_save_batch( + batch: dict, + params: AttributeDict, +) -> None: + """Display the batch statistics and save the batch into disk. + + Args: + batch: + A batch of data. See `dataset.HubertDataset()` + for the content in it. + params: + Parameters for training. See :func:`get_params`. + sp: + The BPE model. + """ + from lhotse.utils import uuid4 + + filename = f"{params.exp_dir}/batch-{uuid4()}.pt" + logging.info(f"Saving batch to {filename}") + torch.save(batch, filename) + + audio = batch["audio"] + logging.info(f"audio shape: {audio.shape}") + + +def scan_pessimistic_batches_for_oom( + model: Union[nn.Module, DDP], + train_dl: torch.utils.data.DataLoader, + optimizer: torch.optim.Optimizer, + params: AttributeDict, +): + from lhotse.dataset import find_pessimistic_batches + + logging.info( + "Sanity check -- see if any of the batches in epoch 1 would cause OOM." + ) + batches, crit_values = find_pessimistic_batches(train_dl.sampler) + for criterion, cuts in batches.items(): + batch = train_dl.dataset[cuts] + try: + with torch.cuda.amp.autocast(enabled=params.use_fp16): + loss, _ = compute_loss( + params=params, + model=model, + batch=batch, + is_training=True, + ) + loss.backward() + optimizer.zero_grad() + except Exception as e: + if "CUDA out of memory" in str(e): + logging.error( + "Your GPU ran out of memory with the current " + "max_duration setting. We recommend decreasing " + "max_duration and trying again.\n" + f"Failing criterion: {criterion} " + f"(={crit_values[criterion]}) ..." + ) + display_and_save_batch(batch, params=params) + raise + logging.info( + f"Maximum memory allocated so far is {torch.cuda.max_memory_allocated()//1000000}MB" + ) + + +def main(): + parser = get_parser() + LibriSpeechDataModule.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/SSL/hubert/pretrain_ce.py b/egs/librispeech/SSL/hubert/pretrain_ce.py new file mode 100644 index 0000000000..24c0d4d3a0 --- /dev/null +++ b/egs/librispeech/SSL/hubert/pretrain_ce.py @@ -0,0 +1,1082 @@ +#!/usr/bin/env python3 +# Copyright 2021-2024 Xiaomi Corp. (authors: Fangjun Kuang, +# Wei Kang, +# Mingshuang Luo, +# Zengwei Yao, +# Yifan Yang, +# Daniel Povey) +# Copyright 2024 Shanghai Jiao Tong University (authors: Jianheng Zhuo) +# +# 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="0,1,2,3,4,5,6,7" + +# For hubert model pretraining: +./hubert/pretrain.py \ + --world-size 8 \ + --num-epochs 400 \ + --start-epoch 1 \ + --use-fp16 1 \ + --exp-dir hubert/exp \ + --full-libri 1 \ + --max-duration 87.5 \ + --accum-grad 4 +""" + + +import argparse +import copy +import logging +import sys +import warnings +from pathlib import Path +from shutil import copyfile +from typing import Any, Dict, Optional, Tuple, Union + +import k2 +import optim +import torch +import torch.multiprocessing as mp +import torch.nn as nn +from hubert_ce import HubertModel, add_hubert_arguments +from lhotse.cut import Cut +from lhotse.dataset.sampling.base import CutSampler +from lhotse.utils import fix_random_seed +from optim import Eden, ScaledAdam +from ssl_datamodule import LibriSpeechDataModule +from torch import Tensor +from torch.cuda.amp import GradScaler +from torch.nn.functional import pad +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.hooks import register_inf_check_hooks +from icefall.utils import ( + AttributeDict, + MetricsTracker, + get_parameter_groups_with_lrs, + setup_logger, + str2bool, +) + +LRSchedulerType = Union[torch.optim.lr_scheduler._LRScheduler, optim.LRScheduler] + + +def get_adjusted_batch_count(params: AttributeDict) -> float: + # returns the number of batches we would have used so far if we had used the reference + # duration. This is for purposes of set_batch_count(). + return ( + params.batch_idx_train + * params.accum_grad + * (params.max_duration * params.world_size) + / params.ref_duration + ) + + +def set_batch_count(model: Union[nn.Module, DDP], batch_count: float) -> None: + if isinstance(model, DDP): + # get underlying nn.Module + model = model.module + for name, module in model.named_modules(): + if hasattr(module, "batch_count"): + module.batch_count = batch_count + if hasattr(module, "name"): + module.name = name + + +def 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=400, + 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="hubert/exp", + help="""The experiment dir. + It specifies the directory where all training related + files, e.g., checkpoints, log, etc, are saved + """, + ) + + parser.add_argument( + "--base-lr", type=float, default=0.045, help="The base learning rate." + ) + + parser.add_argument( + "--lr-batches", + type=float, + default=7500, + help="""Number of steps that affects how rapidly the learning rate + decreases. We suggest not to change this.""", + ) + + parser.add_argument( + "--lr-epochs", + type=float, + default=10.5, + help="""Number of epochs that affects how rapidly the learning rate decreases. + """, + ) + + parser.add_argument( + "--warmup-batches", + type=float, + default=5000, + help="Eden warmup steps", + ) + + parser.add_argument( + "--warmup-start", + type=float, + default=0, + help="Eden warmup start learning rate", + ) + + parser.add_argument( + "--ref-duration", + type=float, + default=80, + help="Reference batch duration for purposes of adjusting batch counts for setting various " + "schedules inside the model", + ) + + 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( + "--sanity-check", + type=str2bool, + default=False, + help="Check if any of the batches in epoch 1 would cause OOM.", + ) + + parser.add_argument( + "--inf-check", + type=str2bool, + default=False, + help="Add hooks to check for infinite module outputs and gradients.", + ) + + parser.add_argument( + "--save-every-n", + type=int, + default=100000, + help="""Save checkpoint after processing this number of batches" + periodically. We save checkpoint to exp-dir/ whenever + params.batch_idx_train % save_every_n == 0. The checkpoint filename + has the form: f'exp-dir/checkpoint-{params.batch_idx_train}.pt' + Note: It also saves checkpoint to `exp-dir/epoch-xxx.pt` at the + end of each epoch where `xxx` is the epoch number counting from 1. + """, + ) + + parser.add_argument( + "--keep-last-k", + type=int, + default=30, + help="""Only keep this number of checkpoints on disk. + For instance, if it is 3, there are only 3 checkpoints + in the exp-dir with filenames `checkpoint-xxx.pt`. + It does not affect checkpoints with name `epoch-xxx.pt`. + """, + ) + + parser.add_argument( + "--average-period", + type=int, + default=200, + help="""Update the averaged model, namely `model_avg`, after processing + this number of batches. `model_avg` is a separate version of model, + in which each floating-point parameter is the average of all the + parameters from the start of training. Each time we take the average, + we do: `model_avg = model * (average_period / batch_idx_train) + + model_avg * ((batch_idx_train - average_period) / batch_idx_train)`. + """, + ) + + parser.add_argument( + "--accum-grad", + type=int, + default=4, + help="""update gradient when batch_idx_train % accum_grad == 0. + """, + ) + + parser.add_argument( + "--use-fp16", + type=str2bool, + default=False, + help="Whether to use half precision training.", + ) + + parser.add_argument( + "--max-keep-size", + type=int, + default=sys.maxsize, + help="exclude sample longer than this.", + ) + + parser.add_argument( + "--min-keep-size", + type=float, + default=32000, + help="exclude sample longer less than this.", + ) + + parser.add_argument( + "--max-sample-size", + type=float, + default=250000, + help="max sample size to crop to for batching.", + ) + + add_hubert_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 updates happen to the model so far across + epochs. + + - sub_batch_idx_train: It contains number of batch 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 + """ + params = AttributeDict( + { + "best_train_loss": float("inf"), + "best_valid_loss": float("inf"), + "best_train_epoch": -1, + "best_valid_epoch": -1, + "batch_idx_train": 0, + "sub_batch_idx_train": 0, + "log_interval": 50, + "reset_interval": 200, + "valid_interval": 3000, # For the 100h subset, use 800 + "env_info": get_env_info(), + } + ) + + return params + + +def _to_int_tuple(s: str): + return tuple(map(int, s.split(","))) + + +def get_model(params: AttributeDict) -> nn.Module: + model = HubertModel(params) + 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], + batch: dict, + is_training: bool, +) -> Tuple[Tensor, MetricsTracker]: + """ + Compute loss given the model and its inputs. + + Args: + params: + Parameters for training. See :func:`get_params`. + model: + The model for training. It is an instance of Zipformer in our case. + batch: + A batch of data. See `dataset.HubertDataset()` + for the content in it. + is_training: + True for training. False for validation. When it is True, this + function enables autograd during computation; when it is False, it + disables autograd. + """ + device = model.device if isinstance(model, DDP) else next(model.parameters()).device + audio = batch["audio"].to(device) + padding_mask = batch["padding_mask"].to(device) + kmeans = batch["kmeans"].to(device) + + with torch.set_grad_enabled(is_training): + loss, num_masked_tokens, logging_output = model( + source=audio, target_list=[kmeans], padding_mask=padding_mask + ) + + assert loss.requires_grad == is_training + + info = MetricsTracker() + with warnings.catch_warnings(): + warnings.simplefilter("ignore") + info["frames"] = num_masked_tokens + for item in logging_output: + info[item] = logging_output[item] + return loss, info + + +def compute_validation_loss( + params: AttributeDict, + model: Union[nn.Module, DDP], + valid_dl: torch.utils.data.DataLoader, + world_size: int = 1, +) -> MetricsTracker: + """Run the validation process.""" + model.eval() + + tot_loss = MetricsTracker() + + for batch_idx, batch in enumerate(valid_dl): + loss, loss_info = compute_loss( + params=params, + model=model, + batch=batch, + is_training=False, + ) + assert loss.requires_grad is False + tot_loss = tot_loss + loss_info + + if world_size > 1: + tot_loss.reduce(loss.device) + + loss_value = tot_loss["loss"] / tot_loss["frames"] + if loss_value < params.best_valid_loss: + params.best_valid_epoch = params.cur_epoch + params.best_valid_loss = loss_value + + return tot_loss + + +def train_one_epoch( + params: AttributeDict, + model: Union[nn.Module, DDP], + optimizer: torch.optim.Optimizer, + scheduler: LRSchedulerType, + train_dl: torch.utils.data.DataLoader, + valid_dl: torch.utils.data.DataLoader, + scaler: GradScaler, + model_avg: Optional[nn.Module] = None, + tb_writer: Optional[SummaryWriter] = None, + world_size: int = 1, + rank: int = 0, +) -> None: + """Train the model for one epoch. + + The training loss from the mean of all frames is saved in + `params.train_loss`. It runs the validation process every + `params.valid_interval` batches. + + Args: + params: + It is returned by :func:`get_params`. + model: + The model for training. + optimizer: + The optimizer we are using. + scheduler: + The learning rate scheduler, we call step() every step. + train_dl: + Dataloader for the training dataset. + valid_dl: + Dataloader for the validation dataset. + scaler: + The scaler used for mix precision training. + model_avg: + The stored model averaged from the start of training. + tb_writer: + Writer to write log messages to tensorboard. + world_size: + Number of nodes in DDP training. If it is 1, DDP is disabled. + rank: + The rank of the node in DDP training. If no DDP is used, it should + be set to 0. + """ + model.train() + + tot_loss = MetricsTracker() + + saved_bad_model = False + + def save_bad_model(suffix: str = ""): + save_checkpoint_impl( + filename=params.exp_dir / f"bad-model{suffix}-{rank}.pt", + model=model, + model_avg=model_avg, + params=params, + optimizer=optimizer, + scheduler=scheduler, + sampler=train_dl.sampler, + scaler=scaler, + rank=0, + ) + + for sub_batch_idx, batch in enumerate(train_dl): + params.sub_batch_idx_train += 1 + batch_idx = sub_batch_idx // params.accum_grad + + if batch_idx % 10 == 0: + set_batch_count(model, get_adjusted_batch_count(params)) + + batch_size = batch["kmeans"].shape[0] + + try: + with torch.cuda.amp.autocast(enabled=params.use_fp16): + loss, loss_info = compute_loss( + params=params, + model=model, + batch=batch, + is_training=True, + ) + # summary stats + tot_loss = (tot_loss * (1 - 1 / params.reset_interval)) + loss_info + + # NOTE: We use reduction==sum and loss is computed over utterances + # in the batch and there is no normalization to it so far. + scaler.scale(loss / params.accum_grad).backward() + + if sub_batch_idx % params.accum_grad == params.accum_grad - 1: + params.batch_idx_train += 1 + scheduler.step_batch(params.batch_idx_train) + + scaler.step(optimizer) + scaler.update() + optimizer.zero_grad() + else: + continue + + except: # noqa + save_bad_model() + display_and_save_batch(batch, params=params) + raise + + if params.print_diagnostics and batch_idx == 5: + return + + if ( + rank == 0 + and params.batch_idx_train > 0 + and params.batch_idx_train % params.average_period == 0 + ): + update_averaged_model( + params=params, + model_cur=model, + model_avg=model_avg, + ) + + if ( + params.batch_idx_train > 0 + and params.batch_idx_train % params.save_every_n == 0 + ): + save_checkpoint_with_global_batch_idx( + out_dir=params.exp_dir, + global_batch_idx=params.batch_idx_train, + model=model, + model_avg=model_avg, + params=params, + optimizer=optimizer, + scheduler=scheduler, + sampler=train_dl.sampler, + scaler=scaler, + rank=rank, + ) + remove_checkpoints( + out_dir=params.exp_dir, + topk=params.keep_last_k, + rank=rank, + ) + + if batch_idx % 100 == 0 and params.use_fp16: + # If the grad scale was less than 1, try increasing it. The _growth_interval + # of the grad scaler is configurable, but we can't configure it to have different + # behavior depending on the current grad scale. + cur_grad_scale = scaler._scale.item() + + if cur_grad_scale < 8.0 or (cur_grad_scale < 32.0 and batch_idx % 400 == 0): + scaler.update(cur_grad_scale * 2.0) + if cur_grad_scale < 0.01: + if not saved_bad_model: + save_bad_model(suffix="-first-warning") + saved_bad_model = True + logging.warning(f"Grad scale is small: {cur_grad_scale}") + if cur_grad_scale < 1.0e-05: + save_bad_model() + raise RuntimeError( + f"grad_scale is too small, exiting: {cur_grad_scale}" + ) + + if batch_idx % params.log_interval == 0: + cur_lr = max(scheduler.get_last_lr()) + cur_grad_scale = scaler._scale.item() if params.use_fp16 else 1.0 + + logging.info( + f"Epoch {params.cur_epoch}, " + f"batch {batch_idx}, loss[{loss_info}], " + f"tot_loss[{tot_loss}], batch size: {batch_size}, " + f"lr: {cur_lr:.2e}, " + + (f"grad_scale: {scaler._scale.item()}" if params.use_fp16 else "") + ) + + if tb_writer is not None: + tb_writer.add_scalar( + "train/learning_rate", cur_lr, params.batch_idx_train + ) + + loss_info.write_summary( + tb_writer, "train/current_", params.batch_idx_train + ) + tot_loss.write_summary(tb_writer, "train/tot_", params.batch_idx_train) + if params.use_fp16: + tb_writer.add_scalar( + "train/grad_scale", + cur_grad_scale, + params.batch_idx_train, + ) + + if batch_idx % params.valid_interval == 0 and not params.print_diagnostics: + logging.info("Computing validation loss") + valid_info = compute_validation_loss( + params=params, + model=model, + valid_dl=valid_dl, + world_size=world_size, + ) + model.train() + logging.info(f"Epoch {params.cur_epoch}, validation: {valid_info}") + logging.info( + f"Maximum memory allocated so far is {torch.cuda.max_memory_allocated()//1000000}MB" + ) + if tb_writer is not None: + valid_info.write_summary( + tb_writer, "train/valid_", params.batch_idx_train + ) + + if batch_idx % params.accum_grad != params.accum_grad - 1: + optimizer.zero_grad() + loss_value = tot_loss["loss"] / tot_loss["frames"] + params.train_loss = loss_value + if params.train_loss < params.best_train_loss: + params.best_train_epoch = params.cur_epoch + params.best_train_loss = params.train_loss + + +def run(rank, world_size, args): + """ + Args: + rank: + It is a value between 0 and `world_size-1`, which is + passed automatically by `mp.spawn()` in :func:`main`. + The node with rank 0 is responsible for saving checkpoint. + world_size: + Number of GPUs for DDP training. + args: + The return value of get_parser().parse_args() + """ + params = get_params() + params.update(vars(args)) + + fix_random_seed(params.seed) + if world_size > 1: + setup_dist(rank, world_size, params.master_port) + + setup_logger(f"{params.exp_dir}/log/log-train") + logging.info("Training started") + + if args.tensorboard and rank == 0: + tb_writer = SummaryWriter(log_dir=f"{params.exp_dir}/tensorboard") + else: + tb_writer = None + + device = torch.device("cpu") + if torch.cuda.is_available(): + device = torch.device("cuda", rank) + logging.info(f"Device: {device}") + logging.info(params) + + logging.info("About to create model") + model = get_model(params) + + num_param = sum([p.numel() for p in model.parameters()]) + logging.info(f"Number of model parameters: {num_param}") + + assert params.save_every_n >= params.average_period + model_avg: Optional[nn.Module] = None + if rank == 0: + # model_avg is only used with rank 0 + model_avg = copy.deepcopy(model).to(torch.float64) + + assert params.start_epoch > 0, params.start_epoch + checkpoints = load_checkpoint_if_available( + params=params, model=model, model_avg=model_avg + ) + + model.to(device) + if world_size > 1: + logging.info("Using DDP") + model = DDP(model, device_ids=[rank], find_unused_parameters=True) + + optimizer = ScaledAdam( + get_parameter_groups_with_lrs(model, lr=params.base_lr, include_names=True), + lr=params.base_lr, # should have no effect + clipping_scale=2.0, + ) + + scheduler = Eden( + optimizer, + params.lr_batches, + params.lr_epochs, + params.warmup_batches, + params.warmup_start, + ) + + if checkpoints and "optimizer" in checkpoints: + logging.info("Loading optimizer state dict") + optimizer.load_state_dict(checkpoints["optimizer"]) + + if ( + checkpoints + and "scheduler" in checkpoints + and checkpoints["scheduler"] is not None + ): + logging.info("Loading scheduler state dict") + scheduler.load_state_dict(checkpoints["scheduler"]) + + if params.print_diagnostics: + opts = diagnostics.TensorDiagnosticOptions( + 512 + ) # allow 4 megabytes per sub-module + diagnostic = diagnostics.attach_diagnostics(model, opts) + + if params.inf_check: + register_inf_check_hooks(model) + + librispeech = LibriSpeechDataModule(args) + + train_cuts = ( + librispeech.train_all_shuf_cuts() + if params.full_libri + else librispeech.train_clean_100_cuts() + ) + + def remove_short_and_long_utt(c: Cut): + # Keep only utterances with duration between 1 second and 20 seconds + # + # Caution: There is a reason to select 20.0 here. Please see + # ../local/display_manifest_statistics.py + # + # You should use ../local/display_manifest_statistics.py to get + # an utterance duration distribution for your dataset to select + # the threshold + if ( + c.duration < params.min_keep_size / params.sample_rate + or c.duration > params.max_keep_size / params.sample_rate + ): + logging.warning( + f"Exclude cut with ID {c.id} from training. Duration: {c.duration}" + ) + return False + + return True + + train_cuts = train_cuts.filter(remove_short_and_long_utt) + + if params.start_batch > 0 and checkpoints and "sampler" in checkpoints: + # We only load the sampler's state dict when it loads a checkpoint + # saved in the middle of an epoch + sampler_state_dict = checkpoints["sampler"] + else: + sampler_state_dict = None + + train_dl = librispeech.train_dataloaders( + train_cuts, + max_sample_size=params.max_sample_size, + sample_rate=params.sample_rate, + label_rate=params.label_rate, + random_crop=params.random_crop, + pad_audio=False, + num_classes=params.num_classes, + do_normalize=params.do_normalize, + sampler_state_dict=sampler_state_dict, + ) + + valid_cuts = librispeech.dev_clean_cuts() + # valid_cuts += librispeech.dev_other_cuts() + valid_cuts = valid_cuts.filter(remove_short_and_long_utt) + + valid_dl = librispeech.valid_dataloaders( + valid_cuts, + max_sample_size=params.max_sample_size, + sample_rate=params.sample_rate, + label_rate=params.label_rate, + random_crop=params.random_crop, + pad_audio=False, + num_classes=params.num_classes, + do_normalize=params.do_normalize, + ) + + if params.sanity_check and not params.print_diagnostics: + scan_pessimistic_batches_for_oom( + model=model, + train_dl=train_dl, + optimizer=optimizer, + params=params, + ) + + scaler = GradScaler(enabled=params.use_fp16, init_scale=1.0) + if checkpoints and "grad_scaler" in checkpoints: + logging.info("Loading grad scaler state dict") + scaler.load_state_dict(checkpoints["grad_scaler"]) + + for epoch in range(params.start_epoch, params.num_epochs + 1): + scheduler.step_epoch(epoch - 1) + fix_random_seed(params.seed + epoch - 1) + train_dl.sampler.set_epoch(epoch - 1) + + if tb_writer is not None: + tb_writer.add_scalar("train/epoch", epoch, params.batch_idx_train) + + params.cur_epoch = epoch + + train_one_epoch( + params=params, + model=model, + model_avg=model_avg, + optimizer=optimizer, + scheduler=scheduler, + train_dl=train_dl, + valid_dl=valid_dl, + scaler=scaler, + tb_writer=tb_writer, + world_size=world_size, + rank=rank, + ) + + if params.print_diagnostics: + diagnostic.print_diagnostics() + break + + save_checkpoint( + params=params, + model=model, + model_avg=model_avg, + optimizer=optimizer, + scheduler=scheduler, + sampler=train_dl.sampler, + scaler=scaler, + rank=rank, + ) + + logging.info("Done!") + + if world_size > 1: + torch.distributed.barrier() + cleanup_dist() + + +def display_and_save_batch( + batch: dict, + params: AttributeDict, +) -> None: + """Display the batch statistics and save the batch into disk. + + Args: + batch: + A batch of data. See `dataset.HubertDataset()` + for the content in it. + params: + Parameters for training. See :func:`get_params`. + sp: + The BPE model. + """ + from lhotse.utils import uuid4 + + filename = f"{params.exp_dir}/batch-{uuid4()}.pt" + logging.info(f"Saving batch to {filename}") + torch.save(batch, filename) + + audio = batch["audio"] + logging.info(f"audio shape: {audio.shape}") + + +def scan_pessimistic_batches_for_oom( + model: Union[nn.Module, DDP], + train_dl: torch.utils.data.DataLoader, + optimizer: torch.optim.Optimizer, + params: AttributeDict, +): + from lhotse.dataset import find_pessimistic_batches + + logging.info( + "Sanity check -- see if any of the batches in epoch 1 would cause OOM." + ) + batches, crit_values = find_pessimistic_batches(train_dl.sampler) + for criterion, cuts in batches.items(): + batch = train_dl.dataset[cuts] + try: + with torch.cuda.amp.autocast(enabled=params.use_fp16): + loss, _ = compute_loss( + params=params, + model=model, + batch=batch, + is_training=True, + ) + loss.backward() + optimizer.zero_grad() + except Exception as e: + if "CUDA out of memory" in str(e): + logging.error( + "Your GPU ran out of memory with the current " + "max_duration setting. We recommend decreasing " + "max_duration and trying again.\n" + f"Failing criterion: {criterion} " + f"(={crit_values[criterion]}) ..." + ) + display_and_save_batch(batch, params=params) + raise + logging.info( + f"Maximum memory allocated so far is {torch.cuda.max_memory_allocated()//1000000}MB" + ) + + +def main(): + parser = get_parser() + LibriSpeechDataModule.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/SSL/hubert/scaling.py b/egs/librispeech/SSL/hubert/scaling.py new file mode 120000 index 0000000000..e30bd99de2 --- /dev/null +++ b/egs/librispeech/SSL/hubert/scaling.py @@ -0,0 +1 @@ +../../ASR/zipformer/scaling.py \ No newline at end of file diff --git a/egs/librispeech/SSL/hubert/ssl_datamodule.py b/egs/librispeech/SSL/hubert/ssl_datamodule.py new file mode 100644 index 0000000000..ac1a0997da --- /dev/null +++ b/egs/librispeech/SSL/hubert/ssl_datamodule.py @@ -0,0 +1,341 @@ +# Copyright 2021 Piotr Żelasko +# Copyright 2023 Xiaomi Corporation (Author: Yifan Yang) +# +# See ../../../../LICENSE for clarification regarding multiple authors +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + + +import argparse +import logging +from functools import lru_cache +from pathlib import Path +from typing import Any, Dict, Optional + +import torch +from dataset import HubertDataset +from lhotse import CutSet, load_manifest_lazy +from lhotse.dataset import DynamicBucketingSampler, SimpleCutSampler +from lhotse.utils import fix_random_seed +from torch.utils.data import DataLoader + +from icefall.utils import str2bool + + +class _SeedWorkers: + def __init__(self, seed: int): + self.seed = seed + + def __call__(self, worker_id: int): + fix_random_seed(self.seed + worker_id) + + +class LibriSpeechDataModule: + """ + DataModule for SSL experiments. + It assumes there is always one train and valid dataloader, + but there can be multiple test dataloaders (e.g. LibriSpeech test-clean + and test-other). + + It contains all the common data pipeline modules used in SSL + experiments, e.g.: + - dynamic batch size, + - bucketing samplers, + + This class should be derived for specific corpora used in SSL tasks. + """ + + def __init__(self, args: argparse.Namespace): + self.args = args + + @classmethod + def add_arguments(cls, parser: argparse.ArgumentParser): + group = parser.add_argument_group( + title="SSL data related options", + description="These options are used for the preparation of " + "PyTorch DataLoaders from Lhotse CutSet's -- they control the " + "effective batch sizes, sampling strategies.", + ) + group.add_argument( + "--full-libri", + type=str2bool, + default=True, + help="When enabled use 960h LibriSpeech. " "Otherwise, use 100h subset.", + ) + + group.add_argument( + "--manifest-dir", + type=Path, + default=Path("data/kmeans"), + help="Path to directory with train/valid/test cuts.", + ) + group.add_argument( + "--max-duration", + type=float, + default=200.0, + help="Maximum pooled recordings duration (seconds) in a " + "single batch. You can reduce it if it causes CUDA OOM.", + ) + group.add_argument( + "--bucketing-sampler", + type=str2bool, + default=True, + help="When enabled, the batches will come from buckets of " + "similar duration (saves padding frames).", + ) + group.add_argument( + "--num-buckets", + type=int, + default=30, + help="The number of buckets for the DynamicBucketingSampler" + "(you might want to increase it for larger datasets).", + ) + group.add_argument( + "--shuffle", + type=str2bool, + default=True, + help="When enabled (=default), the examples will be " + "shuffled for each epoch.", + ) + group.add_argument( + "--drop-last", + type=str2bool, + default=True, + help="Whether to drop last batch. Used by sampler.", + ) + group.add_argument( + "--num-workers", + type=int, + default=2, + help="The number of training dataloader workers that " + "collect the batches.", + ) + group.add_argument( + "--do-normalize", + type=str2bool, + default=True, + help="whether to normalize the data", + ) + group.add_argument( + "--random-crop", + type=str2bool, + default=True, + help="always crop from the beginning if false", + ) + + def train_dataloaders( + self, + cuts_train: CutSet, + max_sample_size: Optional[int] = None, + sample_rate: float = 16000, + label_rate: float = 50, + random_crop: bool = True, + pad_audio: bool = False, + num_classes: list = [504], + do_normalize: bool = True, + sampler_state_dict: Optional[Dict[str, Any]] = None, + ) -> DataLoader: + """ + Args: + cuts_train: + CutSet for training. + sampler_state_dict: + The state dict for the training sampler. + """ + logging.info("About to create train dataset") + train = HubertDataset( + max_sample_size=max_sample_size, + sample_rate=sample_rate, + label_rate=label_rate, + random_crop=random_crop, + pad_audio=pad_audio, + num_classes=num_classes, + do_normalize=do_normalize, + ) + + if self.args.bucketing_sampler: + logging.info("Using DynamicBucketingSampler.") + train_sampler = DynamicBucketingSampler( + cuts_train, + max_duration=self.args.max_duration, + shuffle=self.args.shuffle, + num_buckets=self.args.num_buckets, + drop_last=self.args.drop_last, + ) + else: + logging.info("Using SimpleCutSampler.") + train_sampler = SimpleCutSampler( + cuts_train, + max_duration=self.args.max_duration, + shuffle=self.args.shuffle, + ) + logging.info("About to create train dataloader") + + if sampler_state_dict is not None: + logging.info("Loading sampler state dict") + train_sampler.load_state_dict(sampler_state_dict) + + # 'seed' is derived from the current random state, which will have + # previously been set in the main process. + seed = torch.randint(0, 100000, ()).item() + worker_init_fn = _SeedWorkers(seed) + + train_dl = DataLoader( + train, + sampler=train_sampler, + batch_size=None, + num_workers=self.args.num_workers, + persistent_workers=False, + worker_init_fn=worker_init_fn, + ) + + return train_dl + + def valid_dataloaders( + self, + cuts_valid: CutSet, + max_sample_size: Optional[int] = None, + sample_rate: float = 16000, + label_rate: float = 50, + random_crop: bool = True, + pad_audio: bool = False, + num_classes: list = [504], + do_normalize: bool = True, + ) -> DataLoader: + logging.info("About to create dev dataset") + validate = HubertDataset( + max_sample_size=max_sample_size, + sample_rate=sample_rate, + label_rate=label_rate, + random_crop=random_crop, + pad_audio=pad_audio, + num_classes=num_classes, + do_normalize=do_normalize, + ) + valid_sampler = DynamicBucketingSampler( + cuts_valid, + max_duration=self.args.max_duration, + shuffle=False, + ) + logging.info("About to create dev dataloader") + valid_dl = DataLoader( + validate, + sampler=valid_sampler, + batch_size=None, + num_workers=2, + persistent_workers=False, + ) + + return valid_dl + + def test_dataloaders( + self, + cuts: CutSet, + sample_rate: float = 16000, + label_rate: float = 50, + random_crop: bool = True, + pad_audio: bool = False, + num_classes: list = [504], + do_normalize: bool = True, + ) -> DataLoader: + logging.debug("About to create test dataset") + test = HubertDataset( + sample_rate=sample_rate, + label_rate=label_rate, + random_crop=random_crop, + pad_audio=pad_audio, + num_classes=num_classes, + do_normalize=do_normalize, + ) + sampler = DynamicBucketingSampler( + cuts, + max_duration=self.args.max_duration, + shuffle=False, + ) + logging.debug("About to create test dataloader") + test_dl = DataLoader( + test, + batch_size=None, + sampler=sampler, + num_workers=self.args.num_workers, + ) + return test_dl + + @lru_cache() + def train_clean_100_cuts(self) -> CutSet: + logging.info("About to get train-clean-100 cuts") + return load_manifest_lazy( + self.args.manifest_dir / "librispeech_cuts_train-clean-100.jsonl.gz" + ) + + @lru_cache() + def train_clean_360_cuts(self) -> CutSet: + logging.info("About to get train-clean-360 cuts") + return load_manifest_lazy( + self.args.manifest_dir / "librispeech_cuts_train-clean-360.jsonl.gz" + ) + + @lru_cache() + def train_other_500_cuts(self) -> CutSet: + logging.info("About to get train-other-500 cuts") + return load_manifest_lazy( + self.args.manifest_dir / "librispeech_cuts_train-other-500.jsonl.gz" + ) + + @lru_cache() + def train_all_shuf_cuts(self) -> CutSet: + logging.info( + "About to get the shuffled train-clean-100, \ + train-clean-360 and train-other-500 cuts" + ) + train_clean_100_cuts = self.train_clean_100_cuts() + train_clean_360_cuts = self.train_clean_360_cuts() + train_other_500_cuts = self.train_other_500_cuts() + return CutSet.mux( + train_clean_100_cuts, + train_clean_360_cuts, + train_other_500_cuts, + weights=[ + 28539, # len(train_clean_100_cuts) + 104014, # len(train_clean_360_cuts) + 148688, # len(train_other_500_cuts) + ], + ) + + @lru_cache() + def dev_clean_cuts(self) -> CutSet: + logging.info("About to get dev-clean cuts") + return load_manifest_lazy( + self.args.manifest_dir / "librispeech_cuts_dev-clean.jsonl.gz" + ) + + @lru_cache() + def dev_other_cuts(self) -> CutSet: + logging.info("About to get dev-other cuts") + return load_manifest_lazy( + self.args.manifest_dir / "librispeech_cuts_dev-other.jsonl.gz" + ) + + @lru_cache() + def test_clean_cuts(self) -> CutSet: + logging.info("About to get test-clean cuts") + return load_manifest_lazy( + self.args.manifest_dir / "librispeech_cuts_test-clean.jsonl.gz" + ) + + @lru_cache() + def test_other_cuts(self) -> CutSet: + logging.info("About to get test-other cuts") + return load_manifest_lazy( + self.args.manifest_dir / "librispeech_cuts_test-other.jsonl.gz" + ) diff --git a/egs/librispeech/SSL/hubert/utils.py b/egs/librispeech/SSL/hubert/utils.py new file mode 100644 index 0000000000..de980ba621 --- /dev/null +++ b/egs/librispeech/SSL/hubert/utils.py @@ -0,0 +1,338 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# Permission is hereby granted, free of charge, to any person obtaining a copy +# of this software and associated documentation files (the "Software"), to deal +# in the Software without restriction, including without limitation the rights +# to use, copy, modify, merge, publish, distribute, sublicense, and/or sell +# copies of the Software, and to permit persons to whom the Software is +# furnished to do so, subject to the following conditions: +# +# The above copyright notice and this permission notice shall be included in all +# copies or substantial portions of the Software. +# +# THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR +# IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, +# FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE +# AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER +# LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, +# OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE +# SOFTWARE. + +import math +from typing import Callable, List, Optional + +import torch +import torch.nn as nn +import torch.nn.functional as F + + +def relu_squared(x: torch.Tensor): + return F.relu(x).pow(2) + + +def gelu_accurate(x): + if not hasattr(gelu_accurate, "_a"): + gelu_accurate._a = math.sqrt(2 / math.pi) + return ( + 0.5 * x * (1 + torch.tanh(gelu_accurate._a * (x + 0.044715 * torch.pow(x, 3)))) + ) + + +def is_xla_tensor(tensor): + return torch.is_tensor(tensor) and tensor.device.type == "xla" + + +def index_put(tensor, indices, value): + if is_xla_tensor(tensor): + for _ in range(indices.dim(), tensor.dim()): + indices = indices.unsqueeze(-1) + if indices.size(-1) < tensor.size(-1): + indices = indices.expand_as(tensor) + tensor = torch.mul(tensor, ~indices) + torch.mul(value, indices) + else: + tensor[indices] = value + return tensor + + +def pad_to_multiple(x, multiple, dim=-1, value=0): + # Inspired from https://github.com/lucidrains/local-attention/blob/master/local_attention/local_attention.py#L41 + if x is None: + return None, 0 + tsz = x.size(dim) + m = tsz / multiple + remainder = math.ceil(m) * multiple - tsz + if m.is_integer(): + return x, 0 + pad_offset = (0,) * (-1 - dim) * 2 + + return F.pad(x, (*pad_offset, 0, remainder), value=value), remainder + + +def gelu(x: torch.Tensor) -> torch.Tensor: + return torch.nn.functional.gelu(x.float()).type_as(x) + + +def get_activation_fn(activation: str) -> Callable: + """Returns the activation function corresponding to `activation`""" + if activation == "relu": + return F.relu + elif activation == "relu_squared": + return relu_squared + elif activation == "gelu": + return gelu + elif activation == "gelu_fast": + return gelu_accurate + elif activation == "gelu_accurate": + return gelu_accurate + elif activation == "tanh": + return torch.tanh + elif activation == "linear": + return lambda x: x + elif activation == "swish": + return torch.nn.SiLU + else: + raise RuntimeError("--activation-fn {} not supported".format(activation)) + + +class SamePad(nn.Module): + def __init__(self, kernel_size, causal=False): + super().__init__() + if causal: + self.remove = kernel_size - 1 + else: + self.remove = 1 if kernel_size % 2 == 0 else 0 + + def forward(self, x): + if self.remove > 0: + x = x[:, :, : -self.remove] + return x + + +class SamePad2d(nn.Module): + def __init__(self, kernel_size): + super().__init__() + self.remove = 1 if kernel_size % 2 == 0 else 0 + + def forward(self, x): + assert len(x.size()) == 4 + if self.remove > 0: + x = x[:, :, : -self.remove, : -self.remove] + return x + + +class TransposeLast(nn.Module): + def __init__(self, deconstruct_idx=None, tranpose_dim=-2): + super().__init__() + self.deconstruct_idx = deconstruct_idx + self.tranpose_dim = tranpose_dim + + def forward(self, x): + if self.deconstruct_idx is not None: + x = x[self.deconstruct_idx] + return x.transpose(self.tranpose_dim, -1) + + +try: + from apex.normalization import FusedLayerNorm as _FusedLayerNorm + + has_fused_layernorm = True + + class FusedLayerNorm(_FusedLayerNorm): + @torch.jit.unused + def forward(self, x): + if not x.is_cuda: + return super().forward(x) + else: + with torch.cuda.device(x.device): + return super().forward(x) + +except ImportError: + has_fused_layernorm = False + + +def LayerNorm(normalized_shape, eps=1e-5, elementwise_affine=True, export=False): + if torch.jit.is_scripting() or torch.jit.is_tracing(): + export = True + if not export and torch.cuda.is_available() and has_fused_layernorm: + return FusedLayerNorm(normalized_shape, eps, elementwise_affine) + return torch.nn.LayerNorm(normalized_shape, eps, elementwise_affine) + + +class Fp32LayerNorm(nn.LayerNorm): + def __init__(self, *args, **kwargs): + super().__init__(*args, **kwargs) + + def forward(self, input): + output = F.layer_norm( + input.float(), + self.normalized_shape, + self.weight.float() if self.weight is not None else None, + self.bias.float() if self.bias is not None else None, + self.eps, + ) + return output.type_as(input) + + +class Fp32GroupNorm(nn.GroupNorm): + def __init__(self, *args, **kwargs): + super().__init__(*args, **kwargs) + + def forward(self, input): + output = F.group_norm( + input.float(), + self.num_groups, + self.weight.float() if self.weight is not None else None, + self.bias.float() if self.bias is not None else None, + self.eps, + ) + return output.type_as(input) + + +def softmax(x, dim: int, onnx_trace: bool = False): + if onnx_trace: + return F.softmax(x.float(), dim=dim) + else: + return F.softmax(x, dim=dim, dtype=torch.float32) + + +def quant_noise(module, p, block_size): + """ + Wraps modules and applies quantization noise to the weights for + subsequent quantization with Iterative Product Quantization as + described in "Training with Quantization Noise for Extreme Model Compression" + + Args: + - module: nn.Module + - p: amount of Quantization Noise + - block_size: size of the blocks for subsequent quantization with iPQ + + Remarks: + - Module weights must have the right sizes wrt the block size + - Only Linear, Embedding and Conv2d modules are supported for the moment + - For more detail on how to quantize by blocks with convolutional weights, + see "And the Bit Goes Down: Revisiting the Quantization of Neural Networks" + - We implement the simplest form of noise here as stated in the paper + which consists in randomly dropping blocks + """ + + # if no quantization noise, don't register hook + if p <= 0: + return module + + # supported modules + assert isinstance(module, (nn.Linear, nn.Embedding, nn.Conv2d)) + + # test whether module.weight has the right sizes wrt block_size + is_conv = module.weight.ndim == 4 + + # 2D matrix + if not is_conv: + assert ( + module.weight.size(1) % block_size == 0 + ), "Input features must be a multiple of block sizes" + + # 4D matrix + else: + # 1x1 convolutions + if module.kernel_size == (1, 1): + assert ( + module.in_channels % block_size == 0 + ), "Input channels must be a multiple of block sizes" + # regular convolutions + else: + k = module.kernel_size[0] * module.kernel_size[1] + assert k % block_size == 0, "Kernel size must be a multiple of block size" + + def _forward_pre_hook(mod, input): + # no noise for evaluation + if mod.training: + if not is_conv: + # gather weight and sizes + weight = mod.weight + in_features = weight.size(1) + out_features = weight.size(0) + + # split weight matrix into blocks and randomly drop selected blocks + mask = torch.zeros( + in_features // block_size * out_features, + device=weight.device, + ) + mask.bernoulli_(p) + mask = mask.repeat_interleave(block_size, -1).view(-1, in_features) + + else: + # gather weight and sizes + weight = mod.weight + in_channels = mod.in_channels + out_channels = mod.out_channels + + # split weight matrix into blocks and randomly drop selected blocks + if mod.kernel_size == (1, 1): + mask = torch.zeros( + int(in_channels // block_size * out_channels), + device=weight.device, + ) + mask.bernoulli_(p) + mask = mask.repeat_interleave(block_size, -1).view(-1, in_channels) + else: + mask = torch.zeros( + weight.size(0), weight.size(1), device=weight.device + ) + mask.bernoulli_(p) + mask = ( + mask.unsqueeze(2) + .unsqueeze(3) + .repeat(1, 1, mod.kernel_size[0], mod.kernel_size[1]) + ) + + # scale weights and apply mask + mask = mask.to( + torch.bool + ) # x.bool() is not currently supported in TorchScript + s = 1 / (1 - p) + mod.weight.data = s * weight.masked_fill(mask, 0) + + module.register_forward_pre_hook(_forward_pre_hook) + return module + + +class FairseqDropout(nn.Module): + def __init__(self, p, module_name=None): + super().__init__() + self.p = p + self.module_name = module_name + self.apply_during_inference = False + + def forward(self, x, inplace: bool = False): + if self.p > 0 and (self.training or self.apply_during_inference): + return F.dropout(x, p=self.p, training=True, inplace=inplace) + else: + return x + + def make_generation_fast_( + self, + name: str, + retain_dropout: bool = False, + retain_dropout_modules: Optional[List[str]] = None, + **kwargs + ): + if retain_dropout: + if retain_dropout_modules is not None and self.module_name is None: + pass + elif ( + retain_dropout_modules is None # if None, apply to all modules + or self.module_name in retain_dropout_modules + ): + self.apply_during_inference = True + + +class GradMultiply(torch.autograd.Function): + @staticmethod + def forward(ctx, x, scale): + ctx.scale = scale + res = x.new(x) + return res + + @staticmethod + def backward(ctx, grad): + return grad * ctx.scale, None diff --git a/egs/librispeech/SSL/hubert/wav2vec2_module.py b/egs/librispeech/SSL/hubert/wav2vec2_module.py new file mode 100644 index 0000000000..4c2e1ce989 --- /dev/null +++ b/egs/librispeech/SSL/hubert/wav2vec2_module.py @@ -0,0 +1,593 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# Permission is hereby granted, free of charge, to any person obtaining a copy +# of this software and associated documentation files (the "Software"), to deal +# in the Software without restriction, including without limitation the rights +# to use, copy, modify, merge, publish, distribute, sublicense, and/or sell +# copies of the Software, and to permit persons to whom the Software is +# furnished to do so, subject to the following conditions: +# +# The above copyright notice and this permission notice shall be included in all +# copies or substantial portions of the Software. +# +# THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR +# IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, +# FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE +# AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER +# LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, +# OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE +# SOFTWARE. + +import math +from typing import List, Tuple + +import numpy as np +import torch +import torch.nn as nn +import torch.nn.functional as F +from attention_module import MultiheadAttention, init_bert_params +from utils import ( + Fp32GroupNorm, + Fp32LayerNorm, + LayerNorm, + SamePad, + TransposeLast, + get_activation_fn, + index_put, + pad_to_multiple, +) + + +class ConvFeatureExtractionModel(nn.Module): + def __init__( + self, + conv_layers: List[Tuple[int, int, int]], + dropout: float = 0.0, + mode: str = "default", + conv_bias: bool = False, + ): + super().__init__() + + assert mode in {"default", "layer_norm"} + + def block( + n_in, + n_out, + k, + stride, + is_layer_norm=False, + is_group_norm=False, + conv_bias=False, + ): + def make_conv(): + conv = nn.Conv1d(n_in, n_out, k, stride=stride, bias=conv_bias) + nn.init.kaiming_normal_(conv.weight) + return conv + + assert ( + is_layer_norm and is_group_norm + ) == False, "layer norm and group norm are exclusive" + + if is_layer_norm: + return nn.Sequential( + make_conv(), + nn.Dropout(p=dropout), + nn.Sequential( + TransposeLast(), + Fp32LayerNorm(dim, elementwise_affine=True), + TransposeLast(), + ), + nn.GELU(), + ) + elif is_group_norm: + return nn.Sequential( + make_conv(), + nn.Dropout(p=dropout), + Fp32GroupNorm(dim, dim, affine=True), + nn.GELU(), + ) + else: + return nn.Sequential(make_conv(), nn.Dropout(p=dropout), nn.GELU()) + + in_d = 1 + self.conv_layers = nn.ModuleList() + for i, cl in enumerate(conv_layers): + assert len(cl) == 3, "invalid conv definition: " + str(cl) + (dim, k, stride) = cl + + self.conv_layers.append( + block( + in_d, + dim, + k, + stride, + is_layer_norm=mode == "layer_norm", + is_group_norm=mode == "default" and i == 0, + conv_bias=conv_bias, + ) + ) + in_d = dim + + def forward(self, x): + # BxT -> BxCxT + x = x.unsqueeze(1) + + for conv in self.conv_layers: + x = conv(x) + + return x + + +def make_conv_pos(e, k, g, is_batch_norm=False): + pos_conv = nn.Conv1d( + e, + e, + kernel_size=k, + padding=k // 2, + groups=g, + ) + dropout = 0 + std = math.sqrt((4 * (1.0 - dropout)) / (k * e)) + nn.init.normal_(pos_conv.weight, mean=0, std=std) + nn.init.constant_(pos_conv.bias, 0) + + if not is_batch_norm: + pos_conv = nn.utils.parametrizations.weight_norm(pos_conv, name="weight", dim=2) + pos_conv = nn.Sequential(pos_conv, SamePad(k), nn.GELU()) + else: + batch_norm = nn.BatchNorm1d(e) + pos_conv = nn.Sequential(batch_norm, pos_conv, SamePad(k), nn.GELU()) + + return pos_conv + + +class TransformerEncoder(nn.Module): + def build_encoder_layer(self, args, **kwargs): + if args.layer_type == "transformer": + layer = TransformerSentenceEncoderLayer( + embedding_dim=self.embedding_dim, + ffn_embedding_dim=args.encoder_ffn_embed_dim, + num_attention_heads=args.encoder_attention_heads, + dropout=self.dropout, + attention_dropout=args.attention_dropout, + activation_dropout=args.activation_dropout, + activation_fn=args.activation_fn, + layer_norm_first=args.layer_norm_first, + ) + elif args.layer_type == "trf_adp": + use_adp = False + if args.adp_trf_idx == "all": + use_adp = True + else: + adp_trf_idx = list( + range(*[int(g) for g in args.adp_trf_idx.split(":")]) + ) + if kwargs.get("layer_idx", None) in adp_trf_idx: + use_adp = True + if use_adp: + layer = TransformerSentenceEncoderWithAdapterLayer( + embedding_dim=self.embedding_dim, + ffn_embedding_dim=args.encoder_ffn_embed_dim, + num_attention_heads=args.encoder_attention_heads, + dropout=self.dropout, + attention_dropout=args.attention_dropout, + activation_dropout=args.activation_dropout, + activation_fn=args.activation_fn, + layer_norm_first=args.layer_norm_first, + adapter_num=args.adp_num, + adapter_dim=args.adp_dim, + adapter_act_fn=args.adp_act_fn, + ) + else: + layer = TransformerSentenceEncoderLayer( + embedding_dim=self.embedding_dim, + ffn_embedding_dim=args.encoder_ffn_embed_dim, + num_attention_heads=args.encoder_attention_heads, + dropout=self.dropout, + attention_dropout=args.attention_dropout, + activation_dropout=args.activation_dropout, + activation_fn=args.activation_fn, + layer_norm_first=args.layer_norm_first, + ) + + # layer = fsdp_wrap(layer) + # if args.checkpoint_activations: + # layer = checkpoint_wrapper(layer) + return layer + + def __init__(self, args): + super().__init__() + + self.dropout = args.dropout + self.embedding_dim = args.encoder_embed_dim + self.required_seq_len_multiple = args.required_seq_len_multiple + + pos_conv_depth = getattr(args, "pos_conv_depth", 1) + if pos_conv_depth > 1: + num_layers = args.pos_conv_depth + k = max(3, args.conv_pos // num_layers) + + def make_conv_block(e, k, g, l): + return nn.Sequential( + *[ + nn.Sequential( + nn.Conv1d( + e, + e, + kernel_size=k, + padding=k // 2, + groups=g, + ), + SamePad(k), + TransposeLast(), + LayerNorm(e, elementwise_affine=False), + TransposeLast(), + nn.GELU(), + ) + for _ in range(l) + ] + ) + + self.pos_conv = make_conv_block( + self.embedding_dim, k, args.conv_pos_groups, num_layers + ) + + else: + self.pos_conv = make_conv_pos( + self.embedding_dim, + args.conv_pos, + args.conv_pos_groups, + is_batch_norm=args.conv_pos_batch_norm + if hasattr(args, "conv_pos_batch_norm") + else False, + ) + + self.layers = nn.ModuleList( + [ + self.build_encoder_layer(args, layer_idx=ii) + for ii in range(args.encoder_layers) + ] + ) + self.layer_norm_first = args.layer_norm_first + self.layer_norm = LayerNorm(self.embedding_dim) + self.layerdrop = args.encoder_layerdrop + + self.apply(init_bert_params) + + def forward(self, x, padding_mask=None, layer=None, corpus_key=None): + x, layer_results = self.extract_features( + x, padding_mask, layer, corpus_key=corpus_key + ) + + if self.layer_norm_first and layer is None: + x = self.layer_norm(x) + + return x, layer_results + + def extract_features( + self, + x, + padding_mask=None, + tgt_layer=None, + min_layer=0, + corpus_key=None, + ): + if padding_mask is not None: + x = index_put(x, padding_mask, 0) + + x_conv = self.pos_conv(x.transpose(1, 2)) + x_conv = x_conv.transpose(1, 2) + x = x + x_conv + + if not self.layer_norm_first: + x = self.layer_norm(x) + + # pad to the sequence length dimension + x, pad_length = pad_to_multiple( + x, self.required_seq_len_multiple, dim=-2, value=0 + ) + if pad_length > 0 and padding_mask is None: + padding_mask = x.new_zeros((x.size(0), x.size(1)), dtype=torch.bool) + padding_mask[:, -pad_length:] = True + else: + padding_mask, _ = pad_to_multiple( + padding_mask, self.required_seq_len_multiple, dim=-1, value=True + ) + x = F.dropout(x, p=self.dropout, training=self.training) + + # B x T x C -> T x B x C + x = x.transpose(0, 1) + + layer_results = [] + r = None + + for i, layer in enumerate(self.layers): + dropout_probability = np.random.random() if self.layerdrop > 0 else 1 + if not self.training or (dropout_probability > self.layerdrop): + layer_check = layer + # if isinstance(layer, FullyShardedDataParallel): + # layer_check = layer.unwrapped_module + if (corpus_key is None) or ( + not isinstance( + layer_check, + (TransformerSentenceEncoderWithAdapterLayer,), + ) + ): + x, (z, lr) = layer( + x, + self_attn_padding_mask=padding_mask, + need_weights=False, + ) + else: + x, (z, lr) = layer( + x, + self_attn_padding_mask=padding_mask, + need_weights=False, + corpus_key=corpus_key, + ) + if i >= min_layer: + layer_results.append((x, z, lr)) + if i == tgt_layer: + r = x + break + + if r is not None: + x = r + + # T x B x C -> B x T x C + x = x.transpose(0, 1) + + # undo paddding + if pad_length > 0: + x = x[:, :-pad_length] + + def undo_pad(a, b, c): + return ( + a[:-pad_length], + b[:-pad_length] if b is not None else b, + c[:-pad_length], + ) + + layer_results = [undo_pad(*u) for u in layer_results] + + return x, layer_results + + def max_positions(self): + """Maximum output length supported by the encoder.""" + return self.args.max_positions + + def upgrade_state_dict_named(self, state_dict, name): + """Upgrade a (possibly old) state dict for new versions of fairseq.""" + return state_dict + + +class TransformerSentenceEncoderLayer(nn.Module): + """ + Implements a Transformer Encoder Layer used in BERT/XLM style pre-trained + models. + """ + + def __init__( + self, + embedding_dim: float = 768, + ffn_embedding_dim: float = 3072, + num_attention_heads: int = 8, + dropout: float = 0.1, + attention_dropout: float = 0.1, + activation_dropout: float = 0.1, + activation_fn: str = "relu", + layer_norm_first: bool = False, + ) -> None: + super().__init__() + # Initialize parameters + self.embedding_dim = embedding_dim + self.dropout = dropout + self.activation_dropout = activation_dropout + + # Initialize blocks + self.activation_fn = get_activation_fn(activation_fn) + self.self_attn = MultiheadAttention( + self.embedding_dim, + num_attention_heads, + dropout=attention_dropout, + self_attention=True, + ) + + self.dropout1 = nn.Dropout(dropout) + self.dropout2 = nn.Dropout(self.activation_dropout) + self.dropout3 = nn.Dropout(dropout) + + self.layer_norm_first = layer_norm_first + + # layer norm associated with the self attention layer + self.self_attn_layer_norm = LayerNorm(self.embedding_dim) + self.fc1 = nn.Linear(self.embedding_dim, ffn_embedding_dim) + self.fc2 = nn.Linear(ffn_embedding_dim, self.embedding_dim) + + # layer norm associated with the position wise feed-forward NN + self.final_layer_norm = LayerNorm(self.embedding_dim) + + def forward( + self, + x: torch.Tensor, + self_attn_mask: torch.Tensor = None, + self_attn_padding_mask: torch.Tensor = None, + need_weights: bool = False, + att_args=None, + ): + """ + LayerNorm is applied either before or after the self-attention/ffn + modules similar to the original Transformer imlementation. + """ + residual = x + + if self.layer_norm_first: + x = self.self_attn_layer_norm(x) + x, attn = self.self_attn( + query=x, + key=x, + value=x, + key_padding_mask=self_attn_padding_mask, + attn_mask=self_attn_mask, + need_weights=False, + ) + x = self.dropout1(x) + x = residual + x + + residual = x + x = self.final_layer_norm(x) + x = self.activation_fn(self.fc1(x)) + x = self.dropout2(x) + x = self.fc2(x) + + layer_result = x + + x = self.dropout3(x) + x = residual + x + else: + x, attn = self.self_attn( + query=x, + key=x, + value=x, + key_padding_mask=self_attn_padding_mask, + need_weights=False, + ) + + x = self.dropout1(x) + x = residual + x + + x = self.self_attn_layer_norm(x) + + residual = x + x = self.activation_fn(self.fc1(x)) + x = self.dropout2(x) + x = self.fc2(x) + + layer_result = x + + x = self.dropout3(x) + x = residual + x + x = self.final_layer_norm(x) + + return x, (attn, layer_result) + + +class AdapterFast(nn.Module): + def __init__(self, adapter_num, input_dim, hidden_dim, act_fn): + """ + Implements adapter modules directly with 3D tensor weight as parameters + and without using ModuleList orto speed up training throughput. + """ + super().__init__() + + self.adapter_num = adapter_num + self.input_dim = input_dim + self.hidden_dim = hidden_dim + self.W_a = nn.Parameter(torch.empty(adapter_num, hidden_dim, input_dim)) + self.W_b = nn.Parameter(torch.empty(adapter_num, input_dim, hidden_dim)) + self.b_a = nn.Parameter(torch.empty(adapter_num, hidden_dim)) + self.b_b = nn.Parameter(torch.empty(adapter_num, input_dim)) + + self.ln_W = nn.Parameter(torch.empty(adapter_num, input_dim)) + self.ln_b = nn.Parameter(torch.empty(adapter_num, input_dim)) + self.act_fn = nn.Identity() + if act_fn == "relu": + self.act_fn = nn.ReLU() + elif act_fn == "gelu": + self.act_fn = nn.GELU() + elif act_fn == "selu": + self.act_fn = nn.SELU() + else: + raise ValueError(f"unsupported {act_fn}") + + self.input_dim = input_dim + self.reset_parameters() + + def reset_parameters(self): + for ii in range(self.adapter_num): + nn.init.kaiming_uniform_(self.W_a[ii], a=math.sqrt(5)) + nn.init.kaiming_uniform_(self.W_b[ii], a=math.sqrt(5)) + fan_in, _ = nn.init._calculate_fan_in_and_fan_out(self.W_a[ii]) + bound = 1 / math.sqrt(fan_in) if fan_in > 0 else 0 + nn.init.uniform_(self.b_a[ii], -bound, bound) + fan_in, _ = nn.init._calculate_fan_in_and_fan_out(self.W_b[ii]) + bound = 1 / math.sqrt(fan_in) if fan_in > 0 else 0 + nn.init.uniform_(self.b_b[ii], -bound, bound) + + nn.init.ones_(self.ln_W) + nn.init.zeros_(self.ln_b) + + def forward(self, x, adapter_id): + ii = adapter_id + h = x + h = F.layer_norm(h, (self.input_dim,), self.ln_W[ii], self.ln_b[ii]) + h = F.linear(h, self.W_a[ii], self.b_a[ii]) + h = self.act_fn(h) + h = F.linear(h, self.W_b[ii], self.b_b[ii]) + outputs = h + return outputs + + def extra_repr(self): + return "adapter={}, input_dim={}, hidden_dim={}".format( + self.adapter_num, self.input_dim, self.hidden_dim + ) + + +class TransformerSentenceEncoderWithAdapterLayer(TransformerSentenceEncoderLayer): + """ + Implements a Transformer Encoder Layer with adapters used in BERT/XLM style pre-trained + models. An adapter module is added along with vanilla Transformer module. + """ + + def __init__( + self, + embedding_dim: float = 768, + ffn_embedding_dim: float = 3072, + num_attention_heads: int = 8, + dropout: float = 0.1, + attention_dropout: float = 0.1, + activation_dropout: float = 0.1, + activation_fn: str = "relu", + layer_norm_first: bool = False, + adapter_num=201, + adapter_dim=64, + adapter_act_fn="relu", + ) -> None: + super().__init__( + embedding_dim=embedding_dim, + ffn_embedding_dim=ffn_embedding_dim, + num_attention_heads=num_attention_heads, + dropout=dropout, + attention_dropout=attention_dropout, + activation_dropout=activation_dropout, + activation_fn=activation_fn, + layer_norm_first=layer_norm_first, + ) + + self.adapter_num = adapter_num + self.adapter_dim = adapter_dim + self.adapter_layer = AdapterFast( + adapter_num, self.embedding_dim, self.adapter_dim, adapter_act_fn + ) + + def forward( + self, + x: torch.Tensor, + self_attn_mask: torch.Tensor = None, + self_attn_padding_mask: torch.Tensor = None, + need_weights: bool = False, + att_args=None, + corpus_key=None, + ): + x, (attn, layer_result) = super().forward( + x=x, + self_attn_mask=self_attn_mask, + self_attn_padding_mask=self_attn_padding_mask, + need_weights=need_weights, + att_args=att_args, + ) + assert corpus_key is not None + assert len(set(corpus_key)) == 1, f"corpus_key items are not same {corpus_key}" + y = self.adapter_layer(x, corpus_key[0]) + x = x + y + return x, (attn, layer_result) diff --git a/egs/librispeech/SSL/local/attach_kmeans_to_supervisions.py b/egs/librispeech/SSL/local/attach_kmeans_to_supervisions.py new file mode 100644 index 0000000000..aa2f45f752 --- /dev/null +++ b/egs/librispeech/SSL/local/attach_kmeans_to_supervisions.py @@ -0,0 +1,52 @@ +import os + +import jsonlines +from tqdm import tqdm + +os.system( + "cp /userhome/user/yfy62/librispeech_data/data4ssl/manifests/librispeech_*_dev-clean* ." +) +os.system( + "cp /userhome/user/yfy62/librispeech_data/data4ssl/manifests/librispeech_*_train* ." +) +os.system("chmod -R 644 *.jsonl.gz") +os.system("gunzip *.gz") + +dataset_parts = ( + "dev-clean", + "train-clean-100", + "train-clean-360", + "train-other-500", +) + +kmeans_dir = "/userhome/user/yangguanrou/data/k500" +idx_dir = "/userhome/user/yangguanrou/data/shu" + +kmeans = [] +idxs = [] +for part in ["train", "valid"]: + with open(kmeans_dir + "/" + part + ".km", "r") as f: + kmeans += f.read().splitlines() + + with open(idx_dir + "/" + part + ".tsv", "r") as f: + lines = f.read().splitlines() + idxs += [ + line.split("\t", -1)[0].split("/", -1)[-1].replace(".flac", "") + for line in lines + if ".flac" in line + ] + +idx2kmeans = {} +for idx, km in zip(idxs, kmeans): + idx2kmeans[idx] = km + +for part in dataset_parts: + with jsonlines.open(f"librispeech_supervisions_{part}.jsonl") as reader: + with jsonlines.open( + f"librispeech_supervisions_{part}_new.jsonl", mode="w" + ) as writer: + for obj in tqdm(reader): + obj["custom"] = {"kmeans": idx2kmeans[obj["id"]]} + writer.write(obj) + +os.system('for file in *_new.jsonl; do mv "$file" "${file%_new.jsonl}.jsonl"; done') diff --git a/egs/librispeech/SSL/local/convert_checkpoint_from_fairseq.py b/egs/librispeech/SSL/local/convert_checkpoint_from_fairseq.py new file mode 100644 index 0000000000..4212cd9c60 --- /dev/null +++ b/egs/librispeech/SSL/local/convert_checkpoint_from_fairseq.py @@ -0,0 +1,18 @@ +# simple script to convert a fairseq checkpoint into pytorch parameter state dict +from argparse import ArgumentParser +from collections import OrderedDict + +import torch + +parser = ArgumentParser() +parser.add_argument("--src") +parser.add_argument("--tgt") + +args = parser.parse_args() +src = args.src +tgt = args.tgt + +old_checkpoint = torch.load(src) +new_checkpoint = OrderedDict() +new_checkpoint["model"] = old_checkpoint["model"] +torch.save(new_checkpoint, tgt) diff --git a/egs/librispeech/SSL/local/prepare_char.py b/egs/librispeech/SSL/local/prepare_char.py new file mode 100644 index 0000000000..8cc0502c28 --- /dev/null +++ b/egs/librispeech/SSL/local/prepare_char.py @@ -0,0 +1,259 @@ +#!/usr/bin/env python3 +# Copyright 2021 Xiaomi Corp. (authors: Fangjun Kuang, +# Wei Kang) +# +# See ../../../../LICENSE for clarification regarding multiple authors +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + + +""" + +This script takes as input `lang_dir`, which should contain:: + + - lang_dir/text, + - lang_dir/words.txt + +and generates the following files in the directory `lang_dir`: + + - lexicon.txt + - lexicon_disambig.txt + - L.pt + - L_disambig.pt + - tokens.txt +""" + +import argparse +import re +from pathlib import Path +from typing import Dict, List + +import k2 +import torch +from prepare_lang import ( + Lexicon, + add_disambig_symbols, + add_self_loops, + write_lexicon, + write_mapping, +) + + +def lexicon_to_fst_no_sil( + lexicon: Lexicon, + token2id: Dict[str, int], + word2id: Dict[str, int], + need_self_loops: bool = False, +) -> k2.Fsa: + """Convert a lexicon to an FST (in k2 format). + + Args: + lexicon: + The input lexicon. See also :func:`read_lexicon` + token2id: + A dict mapping tokens to IDs. + word2id: + A dict mapping words to IDs. + need_self_loops: + If True, add self-loop to states with non-epsilon output symbols + on at least one arc out of the state. The input label for this + self loop is `token2id["#0"]` and the output label is `word2id["#0"]`. + Returns: + Return an instance of `k2.Fsa` representing the given lexicon. + """ + loop_state = 0 # words enter and leave from here + next_state = 1 # the next un-allocated state, will be incremented as we go + + arcs = [] + + # The blank symbol is defined in local/train_bpe_model.py + assert token2id[""] == 0 + assert word2id[""] == 0 + + eps = 0 + + for word, pieces in lexicon: + assert len(pieces) > 0, f"{word} has no pronunciations" + cur_state = loop_state + + word = word2id[word] + pieces = [token2id[i] if i in token2id else token2id[""] for i in pieces] + + for i in range(len(pieces) - 1): + w = word if i == 0 else eps + arcs.append([cur_state, next_state, pieces[i], w, 0]) + + cur_state = next_state + next_state += 1 + + # now for the last piece of this word + i = len(pieces) - 1 + w = word if i == 0 else eps + arcs.append([cur_state, loop_state, pieces[i], w, 0]) + + if need_self_loops: + disambig_token = token2id["#0"] + disambig_word = word2id["#0"] + arcs = add_self_loops( + arcs, + disambig_token=disambig_token, + disambig_word=disambig_word, + ) + + final_state = next_state + arcs.append([loop_state, final_state, -1, -1, 0]) + arcs.append([final_state]) + + arcs = sorted(arcs, key=lambda arc: arc[0]) + arcs = [[str(i) for i in arc] for arc in arcs] + arcs = [" ".join(arc) for arc in arcs] + arcs = "\n".join(arcs) + + fsa = k2.Fsa.from_str(arcs, acceptor=False) + return fsa + + +def contain_oov(token_sym_table: Dict[str, int], tokens: List[str]) -> bool: + """Check if all the given tokens are in token symbol table. + + Args: + token_sym_table: + Token symbol table that contains all the valid tokens. + tokens: + A list of tokens. + Returns: + Return True if there is any token not in the token_sym_table, + otherwise False. + """ + for tok in tokens: + if tok not in token_sym_table: + return True + return False + + +def generate_lexicon(token_sym_table: Dict[str, int], words: List[str]) -> Lexicon: + """Generate a lexicon from a word list and token_sym_table. + + Args: + token_sym_table: + Token symbol table that mapping token to token ids. + words: + A list of strings representing words. + Returns: + Return a dict whose keys are words and values are the corresponding + tokens. + """ + lexicon = [] + for word in words: + chars = list(word.strip(" \t")) + if contain_oov(token_sym_table, chars): + continue + lexicon.append((word, chars)) + + # The OOV word is + lexicon.append(("", [""])) + return lexicon + + +def generate_tokens(text_file: str) -> Dict[str, int]: + """Generate tokens from the given text file. + + Args: + text_file: + A file that contains text lines to generate tokens. + Returns: + Return a dict whose keys are tokens and values are token ids ranged + from 0 to len(keys) - 1. + """ + tokens: Dict[str, int] = dict() + tokens[""] = 0 + tokens[""] = 1 + tokens[""] = 2 + whitespace = re.compile(r"([ \t\r\n]+)") + with open(text_file, "r", encoding="utf-8") as f: + for line in f: + line = re.sub(whitespace, "", line) + chars = list(line) + for char in chars: + if char not in tokens: + tokens[char] = len(tokens) + return tokens + + +def get_args(): + parser = argparse.ArgumentParser() + parser.add_argument( + "--lang-dir", + type=str, + help="""Input and output directory. + It should contain the bpe.model and words.txt + """, + ) + + return parser.parse_args() + + +def main(): + args = get_args() + lang_dir = Path(args.lang_dir) + text_file = lang_dir / "text" + + word_sym_table = k2.SymbolTable.from_file(lang_dir / "words.txt") + + words = word_sym_table.symbols + + excluded = ["", "!SIL", "", "", "#0", "", ""] + for w in excluded: + if w in words: + words.remove(w) + + token_sym_table = generate_tokens(text_file) + + lexicon = generate_lexicon(token_sym_table, words) + + lexicon_disambig, max_disambig = add_disambig_symbols(lexicon) + + next_token_id = max(token_sym_table.values()) + 1 + for i in range(max_disambig + 1): + disambig = f"#{i}" + assert disambig not in token_sym_table + token_sym_table[disambig] = next_token_id + next_token_id += 1 + + word_sym_table.add("#0") + word_sym_table.add("") + word_sym_table.add("") + + write_mapping(lang_dir / "tokens.txt", token_sym_table) + + write_lexicon(lang_dir / "lexicon.txt", lexicon) + write_lexicon(lang_dir / "lexicon_disambig.txt", lexicon_disambig) + + L = lexicon_to_fst_no_sil( + lexicon, + token2id=token_sym_table, + word2id=word_sym_table, + ) + + L_disambig = lexicon_to_fst_no_sil( + lexicon_disambig, + token2id=token_sym_table, + word2id=word_sym_table, + need_self_loops=True, + ) + torch.save(L.as_dict(), lang_dir / "L.pt") + torch.save(L_disambig.as_dict(), lang_dir / "L_disambig.pt") + + +if __name__ == "__main__": + main() diff --git a/egs/librispeech/SSL/local/prepare_lang.py b/egs/librispeech/SSL/local/prepare_lang.py new file mode 100644 index 0000000000..c8cf9b881b --- /dev/null +++ b/egs/librispeech/SSL/local/prepare_lang.py @@ -0,0 +1,388 @@ +#!/usr/bin/env python3 +# Copyright 2021 Xiaomi Corp. (authors: Fangjun Kuang) +# +# See ../../../../LICENSE for clarification regarding multiple authors +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + + +""" +This script takes as input a lexicon file "data/lang_phone/lexicon.txt" +consisting of words and tokens (i.e., phones) and does the following: + +1. Add disambiguation symbols to the lexicon and generate lexicon_disambig.txt + +2. Generate tokens.txt, the token table mapping a token to a unique integer. + +3. Generate words.txt, the word table mapping a word to a unique integer. + +4. Generate L.pt, in k2 format. It can be loaded by + + d = torch.load("L.pt") + lexicon = k2.Fsa.from_dict(d) + +5. Generate L_disambig.pt, in k2 format. +""" +import argparse +import math +from collections import defaultdict +from pathlib import Path +from typing import Any, Dict, List, Tuple + +import k2 +import torch + +from icefall.lexicon import read_lexicon, write_lexicon + +Lexicon = List[Tuple[str, List[str]]] + + +def write_mapping(filename: str, sym2id: Dict[str, int]) -> None: + """Write a symbol to ID mapping to a file. + + Note: + No need to implement `read_mapping` as it can be done + through :func:`k2.SymbolTable.from_file`. + + Args: + filename: + Filename to save the mapping. + sym2id: + A dict mapping symbols to IDs. + Returns: + Return None. + """ + with open(filename, "w", encoding="utf-8") as f: + for sym, i in sym2id.items(): + f.write(f"{sym} {i}\n") + + +def get_tokens(lexicon: Lexicon) -> List[str]: + """Get tokens from a lexicon. + + Args: + lexicon: + It is the return value of :func:`read_lexicon`. + Returns: + Return a list of unique tokens. + """ + ans = set() + for _, tokens in lexicon: + ans.update(tokens) + sorted_ans = sorted(list(ans)) + return sorted_ans + + +def get_words(lexicon: Lexicon) -> List[str]: + """Get words from a lexicon. + + Args: + lexicon: + It is the return value of :func:`read_lexicon`. + Returns: + Return a list of unique words. + """ + ans = set() + for word, _ in lexicon: + ans.add(word) + sorted_ans = sorted(list(ans)) + return sorted_ans + + +def add_disambig_symbols(lexicon: Lexicon) -> Tuple[Lexicon, int]: + """It adds pseudo-token disambiguation symbols #1, #2 and so on + at the ends of tokens to ensure that all pronunciations are different, + and that none is a prefix of another. + + See also add_lex_disambig.pl from kaldi. + + Args: + lexicon: + It is returned by :func:`read_lexicon`. + Returns: + Return a tuple with two elements: + + - The output lexicon with disambiguation symbols + - The ID of the max disambiguation symbol that appears + in the lexicon + """ + + # (1) Work out the count of each token-sequence in the + # lexicon. + count = defaultdict(int) + for _, tokens in lexicon: + count[" ".join(tokens)] += 1 + + # (2) For each left sub-sequence of each token-sequence, note down + # that it exists (for identifying prefixes of longer strings). + issubseq = defaultdict(int) + for _, tokens in lexicon: + tokens = tokens.copy() + tokens.pop() + while tokens: + issubseq[" ".join(tokens)] = 1 + tokens.pop() + + # (3) For each entry in the lexicon: + # if the token sequence is unique and is not a + # prefix of another word, no disambig symbol. + # Else output #1, or #2, #3, ... if the same token-seq + # has already been assigned a disambig symbol. + ans = [] + + # We start with #1 since #0 has its own purpose + first_allowed_disambig = 1 + max_disambig = first_allowed_disambig - 1 + last_used_disambig_symbol_of = defaultdict(int) + + for word, tokens in lexicon: + tokenseq = " ".join(tokens) + assert tokenseq != "" + if issubseq[tokenseq] == 0 and count[tokenseq] == 1: + ans.append((word, tokens)) + continue + + cur_disambig = last_used_disambig_symbol_of[tokenseq] + if cur_disambig == 0: + cur_disambig = first_allowed_disambig + else: + cur_disambig += 1 + + if cur_disambig > max_disambig: + max_disambig = cur_disambig + last_used_disambig_symbol_of[tokenseq] = cur_disambig + tokenseq += f" #{cur_disambig}" + ans.append((word, tokenseq.split())) + return ans, max_disambig + + +def generate_id_map(symbols: List[str]) -> Dict[str, int]: + """Generate ID maps, i.e., map a symbol to a unique ID. + + Args: + symbols: + A list of unique symbols. + Returns: + A dict containing the mapping between symbols and IDs. + """ + return {sym: i for i, sym in enumerate(symbols)} + + +def add_self_loops( + arcs: List[List[Any]], disambig_token: int, disambig_word: int +) -> List[List[Any]]: + """Adds self-loops to states of an FST to propagate disambiguation symbols + through it. They are added on each state with non-epsilon output symbols + on at least one arc out of the state. + + See also fstaddselfloops.pl from Kaldi. One difference is that + Kaldi uses OpenFst style FSTs and it has multiple final states. + This function uses k2 style FSTs and it does not need to add self-loops + to the final state. + + The input label of a self-loop is `disambig_token`, while the output + label is `disambig_word`. + + Args: + arcs: + A list-of-list. The sublist contains + `[src_state, dest_state, label, aux_label, score]` + disambig_token: + It is the token ID of the symbol `#0`. + disambig_word: + It is the word ID of the symbol `#0`. + + Return: + Return new `arcs` containing self-loops. + """ + states_needs_self_loops = set() + for arc in arcs: + src, dst, ilabel, olabel, score = arc + if olabel != 0: + states_needs_self_loops.add(src) + + ans = [] + for s in states_needs_self_loops: + ans.append([s, s, disambig_token, disambig_word, 0]) + + return arcs + ans + + +def lexicon_to_fst( + lexicon: Lexicon, + token2id: Dict[str, int], + word2id: Dict[str, int], + sil_token: str = "SIL", + sil_prob: float = 0.5, + need_self_loops: bool = False, +) -> k2.Fsa: + """Convert a lexicon to an FST (in k2 format) with optional silence at + the beginning and end of each word. + + Args: + lexicon: + The input lexicon. See also :func:`read_lexicon` + token2id: + A dict mapping tokens to IDs. + word2id: + A dict mapping words to IDs. + sil_token: + The silence token. + sil_prob: + The probability for adding a silence at the beginning and end + of the word. + need_self_loops: + If True, add self-loop to states with non-epsilon output symbols + on at least one arc out of the state. The input label for this + self loop is `token2id["#0"]` and the output label is `word2id["#0"]`. + Returns: + Return an instance of `k2.Fsa` representing the given lexicon. + """ + assert sil_prob > 0.0 and sil_prob < 1.0 + # CAUTION: we use score, i.e, negative cost. + sil_score = math.log(sil_prob) + no_sil_score = math.log(1.0 - sil_prob) + + start_state = 0 + loop_state = 1 # words enter and leave from here + sil_state = 2 # words terminate here when followed by silence; this state + # has a silence transition to loop_state. + next_state = 3 # the next un-allocated state, will be incremented as we go. + arcs = [] + + assert token2id[""] == 0 + assert word2id[""] == 0 + + eps = 0 + + sil_token = token2id[sil_token] + + arcs.append([start_state, loop_state, eps, eps, no_sil_score]) + arcs.append([start_state, sil_state, eps, eps, sil_score]) + arcs.append([sil_state, loop_state, sil_token, eps, 0]) + + for word, tokens in lexicon: + assert len(tokens) > 0, f"{word} has no pronunciations" + cur_state = loop_state + + word = word2id[word] + tokens = [token2id[i] for i in tokens] + + for i in range(len(tokens) - 1): + w = word if i == 0 else eps + arcs.append([cur_state, next_state, tokens[i], w, 0]) + + cur_state = next_state + next_state += 1 + + # now for the last token of this word + # It has two out-going arcs, one to the loop state, + # the other one to the sil_state. + i = len(tokens) - 1 + w = word if i == 0 else eps + arcs.append([cur_state, loop_state, tokens[i], w, no_sil_score]) + arcs.append([cur_state, sil_state, tokens[i], w, sil_score]) + + if need_self_loops: + disambig_token = token2id["#0"] + disambig_word = word2id["#0"] + arcs = add_self_loops( + arcs, + disambig_token=disambig_token, + disambig_word=disambig_word, + ) + + final_state = next_state + arcs.append([loop_state, final_state, -1, -1, 0]) + arcs.append([final_state]) + + arcs = sorted(arcs, key=lambda arc: arc[0]) + arcs = [[str(i) for i in arc] for arc in arcs] + arcs = [" ".join(arc) for arc in arcs] + arcs = "\n".join(arcs) + + fsa = k2.Fsa.from_str(arcs, acceptor=False) + return fsa + + +def get_args(): + parser = argparse.ArgumentParser() + parser.add_argument("--lang-dir", type=str, help="The lang dir, data/lang_phone") + return parser.parse_args() + + +def main(): + out_dir = Path(get_args().lang_dir) + lexicon_filename = out_dir / "lexicon.txt" + sil_token = "SIL" + sil_prob = 0.5 + + lexicon = read_lexicon(lexicon_filename) + tokens = get_tokens(lexicon) + words = get_words(lexicon) + + lexicon_disambig, max_disambig = add_disambig_symbols(lexicon) + + for i in range(max_disambig + 1): + disambig = f"#{i}" + assert disambig not in tokens + tokens.append(f"#{i}") + + assert "" not in tokens + tokens = [""] + tokens + + assert "" not in words + assert "#0" not in words + assert "" not in words + assert "" not in words + + words = [""] + words + ["#0", "", ""] + + token2id = generate_id_map(tokens) + word2id = generate_id_map(words) + + write_mapping(out_dir / "tokens.txt", token2id) + write_mapping(out_dir / "words.txt", word2id) + write_lexicon(out_dir / "lexicon_disambig.txt", lexicon_disambig) + + L = lexicon_to_fst( + lexicon, + token2id=token2id, + word2id=word2id, + sil_token=sil_token, + sil_prob=sil_prob, + ) + + L_disambig = lexicon_to_fst( + lexicon_disambig, + token2id=token2id, + word2id=word2id, + sil_token=sil_token, + sil_prob=sil_prob, + need_self_loops=True, + ) + torch.save(L.as_dict(), out_dir / "L.pt") + torch.save(L_disambig.as_dict(), out_dir / "L_disambig.pt") + + if False: + # Just for debugging, will remove it + L.labels_sym = k2.SymbolTable.from_file(out_dir / "tokens.txt") + L.aux_labels_sym = k2.SymbolTable.from_file(out_dir / "words.txt") + L_disambig.labels_sym = L.labels_sym + L_disambig.aux_labels_sym = L.aux_labels_sym + L.draw(out_dir / "L.png", title="L") + L_disambig.draw(out_dir / "L_disambig.png", title="L_disambig") + + +if __name__ == "__main__": + main() diff --git a/egs/librispeech/SSL/local/process_librispeech4finetune.py b/egs/librispeech/SSL/local/process_librispeech4finetune.py new file mode 100644 index 0000000000..09f4b8a3ed --- /dev/null +++ b/egs/librispeech/SSL/local/process_librispeech4finetune.py @@ -0,0 +1,107 @@ +#!/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. + +import argparse +import logging +import os +from pathlib import Path +from typing import Optional + +import torch +from lhotse import CutSet +from lhotse.recipes.utils import read_manifests_if_cached + +from icefall.utils import str2bool + +# Torch's multithreaded behavior needs to be disabled or +# it wastes a lot of CPU and slow things down. +# Do this outside of main() in case it needs to take effect +# even when we are not invoking the main (e.g. when spawning subprocesses). +torch.set_num_threads(1) +torch.set_num_interop_threads(1) + + +def get_args(): + parser = argparse.ArgumentParser() + + parser.add_argument( + "--dataset", + type=str, + help="""Dataset parts to compute fbank. If None, we will use all""", + ) + + return parser.parse_args() + + +def process_wav_librispeech( + dataset: Optional[str] = None, +): + src_dir = Path("data/manifests") + output_dir = Path("data/wav") + + if dataset is None: + dataset_parts = ( + "dev-clean", + "dev-other", + "test-clean", + "test-other", + "train-clean-100", + "train-clean-360", + "train-other-500", + ) + else: + dataset_parts = dataset.split(" ", -1) + + prefix = "librispeech" + suffix = "jsonl.gz" + manifests = read_manifests_if_cached( + dataset_parts=dataset_parts, + output_dir=src_dir, + prefix=prefix, + suffix=suffix, + ) + assert manifests is not None + + assert len(manifests) == len(dataset_parts), ( + len(manifests), + len(dataset_parts), + list(manifests.keys()), + dataset_parts, + ) + + for partition, m in manifests.items(): + cuts_filename = f"{prefix}_cuts_{partition}.{suffix}" + if (output_dir / cuts_filename).is_file(): + logging.info(f"{partition} already exists - skipping.") + continue + logging.info(f"Processing {partition}") + cut_set = CutSet.from_manifests( + recordings=m["recordings"], + supervisions=m["supervisions"], + ) + cut_set.to_file(output_dir / cuts_filename) + + +if __name__ == "__main__": + formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s" + + logging.basicConfig(format=formatter, level=logging.INFO) + args = get_args() + logging.info(vars(args)) + process_wav_librispeech( + dataset=args.dataset, + ) diff --git a/egs/librispeech/SSL/local/process_librispeech4pretrain.py b/egs/librispeech/SSL/local/process_librispeech4pretrain.py new file mode 100644 index 0000000000..c375a2df32 --- /dev/null +++ b/egs/librispeech/SSL/local/process_librispeech4pretrain.py @@ -0,0 +1,104 @@ +#!/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. + +import argparse +import logging +import os +from pathlib import Path +from typing import Optional + +import torch +from lhotse import CutSet +from lhotse.recipes.utils import read_manifests_if_cached + +from icefall.utils import str2bool + +# Torch's multithreaded behavior needs to be disabled or +# it wastes a lot of CPU and slow things down. +# Do this outside of main() in case it needs to take effect +# even when we are not invoking the main (e.g. when spawning subprocesses). +torch.set_num_threads(1) +torch.set_num_interop_threads(1) + + +def get_args(): + parser = argparse.ArgumentParser() + + parser.add_argument( + "--dataset", + type=str, + help="""Dataset parts to compute fbank. If None, we will use all""", + ) + + return parser.parse_args() + + +def process_kmeans_librispeech( + dataset: Optional[str] = None, +): + src_dir = Path(".") + output_dir = Path(".") + + if dataset is None: + dataset_parts = ( + "dev-clean", + "train-clean-100", + "train-clean-360", + "train-other-500", + ) + else: + dataset_parts = dataset.split(" ", -1) + + prefix = "librispeech" + suffix = "jsonl" + manifests = read_manifests_if_cached( + dataset_parts=dataset_parts, + output_dir=src_dir, + prefix=prefix, + suffix=suffix, + ) + assert manifests is not None + + assert len(manifests) == len(dataset_parts), ( + len(manifests), + len(dataset_parts), + list(manifests.keys()), + dataset_parts, + ) + + for partition, m in manifests.items(): + cuts_filename = f"{prefix}_cuts_{partition}_raw.{suffix}" + if (output_dir / cuts_filename).is_file(): + logging.info(f"{partition} already exists - skipping.") + continue + logging.info(f"Processing {partition}") + cut_set = CutSet.from_manifests( + recordings=m["recordings"], + supervisions=m["supervisions"], + ) + cut_set.to_file(output_dir / cuts_filename) + + +if __name__ == "__main__": + formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s" + + logging.basicConfig(format=formatter, level=logging.INFO) + args = get_args() + logging.info(vars(args)) + process_kmeans_librispeech( + dataset=args.dataset, + ) diff --git a/egs/librispeech/SSL/local/process_raw_cuts.py b/egs/librispeech/SSL/local/process_raw_cuts.py new file mode 100644 index 0000000000..9d2ee5945b --- /dev/null +++ b/egs/librispeech/SSL/local/process_raw_cuts.py @@ -0,0 +1,23 @@ +import os + +import jsonlines +from tqdm import tqdm + +dataset_parts = ( + "dev-clean", + "train-clean-100", + "train-clean-360", + "train-other-500", +) + +for part in dataset_parts: + with jsonlines.open(f"librispeech_cuts_{part}_raw.jsonl") as reader: + with jsonlines.open(f"librispeech_cuts_{part}.jsonl", mode="w") as writer: + for obj in tqdm(reader): + obj["custom"] = {"kmeans": obj["supervisions"][0]["custom"]["kmeans"]} + del obj["supervisions"][0]["custom"] + + writer.write(obj) + +os.system("rm *_raw.jsonl") +os.system("gzip *.jsonl") diff --git a/egs/librispeech/SSL/shared b/egs/librispeech/SSL/shared new file mode 120000 index 0000000000..4cbd91a7e9 --- /dev/null +++ b/egs/librispeech/SSL/shared @@ -0,0 +1 @@ +../../../icefall/shared \ No newline at end of file diff --git a/egs/librispeech/SSL/zipformer/asr_datamodule.py b/egs/librispeech/SSL/zipformer/asr_datamodule.py new file mode 120000 index 0000000000..21a701163d --- /dev/null +++ b/egs/librispeech/SSL/zipformer/asr_datamodule.py @@ -0,0 +1 @@ +../hubert/asr_datamodule.py \ No newline at end of file diff --git a/egs/librispeech/SSL/zipformer/beam_search.py b/egs/librispeech/SSL/zipformer/beam_search.py new file mode 120000 index 0000000000..f4d4b57326 --- /dev/null +++ b/egs/librispeech/SSL/zipformer/beam_search.py @@ -0,0 +1 @@ +../../ASR/zipformer/beam_search.py \ No newline at end of file diff --git a/egs/librispeech/SSL/zipformer/dataset.py b/egs/librispeech/SSL/zipformer/dataset.py new file mode 120000 index 0000000000..cb5aedde1a --- /dev/null +++ b/egs/librispeech/SSL/zipformer/dataset.py @@ -0,0 +1 @@ +../hubert/dataset.py \ No newline at end of file diff --git a/egs/librispeech/SSL/zipformer/decode.py b/egs/librispeech/SSL/zipformer/decode.py new file mode 100644 index 0000000000..1562c28b8b --- /dev/null +++ b/egs/librispeech/SSL/zipformer/decode.py @@ -0,0 +1,1043 @@ +#!/usr/bin/env python3 +# Copyright 2021-2024 Xiaomi Corp. (authors: Fangjun Kuang, +# Zengwei Yao, +# Yifan Yang) +# +# See ../../../../LICENSE for clarification regarding multiple authors +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +""" +Usage: +(1) greedy search +./zipformer/decode.py \ + --epoch 28 \ + --avg 15 \ + --exp-dir ./zipformer/exp \ + --max-duration 600 \ + --decoding-method greedy_search + +(2) beam search (not recommended) +./zipformer/decode.py \ + --epoch 28 \ + --avg 15 \ + --exp-dir ./zipformer/exp \ + --max-duration 600 \ + --decoding-method beam_search \ + --beam-size 4 + +(3) modified beam search +./zipformer/decode.py \ + --epoch 28 \ + --avg 15 \ + --exp-dir ./zipformer/exp \ + --max-duration 600 \ + --decoding-method modified_beam_search \ + --beam-size 4 + +(4) fast beam search (one best) +./zipformer/decode.py \ + --epoch 28 \ + --avg 15 \ + --exp-dir ./zipformer/exp \ + --max-duration 600 \ + --decoding-method fast_beam_search \ + --beam 20.0 \ + --max-contexts 8 \ + --max-states 64 + +(5) fast beam search (nbest) +./zipformer/decode.py \ + --epoch 28 \ + --avg 15 \ + --exp-dir ./zipformer/exp \ + --max-duration 600 \ + --decoding-method fast_beam_search_nbest \ + --beam 20.0 \ + --max-contexts 8 \ + --max-states 64 \ + --num-paths 200 \ + --nbest-scale 0.5 + +(6) fast beam search (nbest oracle WER) +./zipformer/decode.py \ + --epoch 28 \ + --avg 15 \ + --exp-dir ./zipformer/exp \ + --max-duration 600 \ + --decoding-method fast_beam_search_nbest_oracle \ + --beam 20.0 \ + --max-contexts 8 \ + --max-states 64 \ + --num-paths 200 \ + --nbest-scale 0.5 + +(7) fast beam search (with LG) +./zipformer/decode.py \ + --epoch 28 \ + --avg 15 \ + --exp-dir ./zipformer/exp \ + --max-duration 600 \ + --decoding-method fast_beam_search_nbest_LG \ + --beam 20.0 \ + --max-contexts 8 \ + --max-states 64 +""" + + +import argparse +import logging +import math +import os +from collections import defaultdict +from pathlib import Path +from typing import Dict, List, Optional, Tuple + +import k2 +import sentencepiece as spm +import torch +import torch.nn as nn +from asr_datamodule import LibriSpeechAsrDataModule +from beam_search import ( + beam_search, + fast_beam_search_nbest, + fast_beam_search_nbest_LG, + fast_beam_search_nbest_oracle, + fast_beam_search_one_best, + greedy_search, + greedy_search_batch, + modified_beam_search, + modified_beam_search_lm_rescore, + modified_beam_search_lm_rescore_LODR, + modified_beam_search_lm_shallow_fusion, + modified_beam_search_LODR, +) +from finetune import add_model_arguments, get_model, get_params + +from icefall import ContextGraph, LmScorer, NgramLm +from icefall.checkpoint import ( + average_checkpoints, + average_checkpoints_with_averaged_model, + find_checkpoints, + load_checkpoint, +) +from icefall.lexicon import Lexicon +from icefall.utils import ( + AttributeDict, + make_pad_mask, + setup_logger, + store_transcripts, + str2bool, + write_error_stats, +) + + +def get_parser(): + parser = argparse.ArgumentParser( + formatter_class=argparse.ArgumentDefaultsHelpFormatter + ) + + parser.add_argument( + "--epoch", + type=int, + default=30, + help="""It specifies the checkpoint to use for decoding. + Note: Epoch counts from 1. + You can specify --avg to use more checkpoints for model averaging.""", + ) + + parser.add_argument( + "--iter", + type=int, + default=0, + help="""If positive, --epoch is ignored and it + will use the checkpoint exp_dir/checkpoint-iter.pt. + You can specify --avg to use more checkpoints for model averaging. + """, + ) + + parser.add_argument( + "--avg", + type=int, + default=15, + help="Number of checkpoints to average. Automatically select " + "consecutive checkpoints before the checkpoint specified by " + "'--epoch' and '--iter'", + ) + + parser.add_argument( + "--use-averaged-model", + type=str2bool, + default=True, + help="Whether to load averaged model. Currently it only supports " + "using --epoch. If True, it would decode with the averaged model " + "over the epoch range from `epoch-avg` (excluded) to `epoch`." + "Actually only the models with epoch number of `epoch-avg` and " + "`epoch` are loaded for averaging. ", + ) + + parser.add_argument( + "--exp-dir", + type=str, + default="zipformer/exp", + help="The experiment dir", + ) + + parser.add_argument( + "--bpe-model", + type=str, + default="data/lang_bpe_500/bpe.model", + help="Path to the BPE model", + ) + + parser.add_argument( + "--lang-dir", + type=Path, + default="data/lang_bpe_500", + help="The lang dir containing word table and LG graph", + ) + + parser.add_argument( + "--decoding-method", + type=str, + default="greedy_search", + help="""Possible values are: + - greedy_search + - beam_search + - modified_beam_search + - modified_beam_search_LODR + - fast_beam_search + - fast_beam_search_nbest + - fast_beam_search_nbest_oracle + - fast_beam_search_nbest_LG + If you use fast_beam_search_nbest_LG, you have to specify + `--lang-dir`, which should contain `LG.pt`. + """, + ) + + parser.add_argument( + "--beam-size", + type=int, + default=4, + help="""An integer indicating how many candidates we will keep for each + frame. Used only when --decoding-method is beam_search or + modified_beam_search.""", + ) + + parser.add_argument( + "--beam", + type=float, + default=20.0, + help="""A floating point value to calculate the cutoff score during beam + search (i.e., `cutoff = max-score - beam`), which is the same as the + `beam` in Kaldi. + Used only when --decoding-method is fast_beam_search, + fast_beam_search_nbest, fast_beam_search_nbest_LG, + and fast_beam_search_nbest_oracle + """, + ) + + parser.add_argument( + "--ngram-lm-scale", + type=float, + default=0.01, + help=""" + Used only when --decoding-method is fast_beam_search_nbest_LG. + It specifies the scale for n-gram LM scores. + """, + ) + + parser.add_argument( + "--max-contexts", + type=int, + default=8, + help="""Used only when --decoding-method is + fast_beam_search, fast_beam_search_nbest, fast_beam_search_nbest_LG, + and fast_beam_search_nbest_oracle""", + ) + + parser.add_argument( + "--max-states", + type=int, + default=64, + help="""Used only when --decoding-method is + fast_beam_search, fast_beam_search_nbest, fast_beam_search_nbest_LG, + and fast_beam_search_nbest_oracle""", + ) + + parser.add_argument( + "--context-size", + type=int, + default=2, + help="The context size in the decoder. 1 means bigram; " "2 means tri-gram", + ) + parser.add_argument( + "--max-sym-per-frame", + type=int, + default=1, + help="""Maximum number of symbols per frame. + Used only when --decoding-method is greedy_search""", + ) + + parser.add_argument( + "--num-paths", + type=int, + default=200, + help="""Number of paths for nbest decoding. + Used only when the decoding method is fast_beam_search_nbest, + fast_beam_search_nbest_LG, and fast_beam_search_nbest_oracle""", + ) + + parser.add_argument( + "--nbest-scale", + type=float, + default=0.5, + help="""Scale applied to lattice scores when computing nbest paths. + Used only when the decoding method is fast_beam_search_nbest, + fast_beam_search_nbest_LG, and fast_beam_search_nbest_oracle""", + ) + + parser.add_argument( + "--use-shallow-fusion", + type=str2bool, + default=False, + help="""Use neural network LM for shallow fusion. + If you want to use LODR, you will also need to set this to true + """, + ) + + parser.add_argument( + "--lm-type", + type=str, + default="rnn", + help="Type of NN lm", + choices=["rnn", "transformer"], + ) + + parser.add_argument( + "--lm-scale", + type=float, + default=0.3, + help="""The scale of the neural network LM + Used only when `--use-shallow-fusion` is set to True. + """, + ) + + parser.add_argument( + "--tokens-ngram", + type=int, + default=2, + help="""The order of the ngram lm. + """, + ) + + parser.add_argument( + "--backoff-id", + type=int, + default=500, + help="ID of the backoff symbol in the ngram LM", + ) + + parser.add_argument( + "--context-score", + type=float, + default=2, + help=""" + The bonus score of each token for the context biasing words/phrases. + Used only when --decoding-method is modified_beam_search and + modified_beam_search_LODR. + """, + ) + + parser.add_argument( + "--context-file", + type=str, + default="", + help=""" + The path of the context biasing lists, one word/phrase each line + Used only when --decoding-method is modified_beam_search and + modified_beam_search_LODR. + """, + ) + + add_model_arguments(parser) + + return parser + + +def decode_one_batch( + params: AttributeDict, + model: nn.Module, + sp: spm.SentencePieceProcessor, + batch: dict, + word_table: Optional[k2.SymbolTable] = None, + decoding_graph: Optional[k2.Fsa] = None, + context_graph: Optional[ContextGraph] = None, + LM: Optional[LmScorer] = None, + ngram_lm=None, + ngram_lm_scale: float = 0.0, +) -> Dict[str, List[List[str]]]: + """Decode one batch and return the result in a dict. The dict has the + following format: + + - key: It indicates the setting used for decoding. For example, + if greedy_search is used, it would be "greedy_search" + If beam search with a beam size of 7 is used, it would be + "beam_7" + - value: It contains the decoding result. `len(value)` equals to + batch size. `value[i]` is the decoding result for the i-th + utterance in the given batch. + Args: + params: + It's the return value of :func:`get_params`. + model: + The neural model. + sp: + The BPE model. + batch: + It is the return value from iterating + `lhotse.dataset.K2SpeechRecognitionDataset`. See its documentation + for the format of the `batch`. + word_table: + The word symbol table. + decoding_graph: + The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used + only when --decoding-method is fast_beam_search, fast_beam_search_nbest, + fast_beam_search_nbest_oracle, and fast_beam_search_nbest_LG. + LM: + A neural network language model. + ngram_lm: + A ngram language model + ngram_lm_scale: + The scale for the ngram language model. + Returns: + Return the decoding result. See above description for the format of + the returned dict. + """ + device = next(model.parameters()).device + audio = batch["audio"].to(device) + padding_mask = batch["padding_mask"].to(device) + + encoder_out, encoder_out_lens = model.forward_encoder(audio, padding_mask) + + 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(batch["supervisions"]["text"]), + nbest_scale=params.nbest_scale, + ) + for hyp in sp.decode(hyp_tokens): + hyps.append(hyp.split()) + elif params.decoding_method == "greedy_search" and params.max_sym_per_frame == 1: + hyp_tokens = greedy_search_batch( + model=model, + encoder_out=encoder_out, + encoder_out_lens=encoder_out_lens, + ) + for hyp in sp.decode(hyp_tokens): + hyps.append(hyp.split()) + elif params.decoding_method == "modified_beam_search": + hyp_tokens = modified_beam_search( + model=model, + encoder_out=encoder_out, + encoder_out_lens=encoder_out_lens, + beam=params.beam_size, + context_graph=context_graph, + ) + for hyp in sp.decode(hyp_tokens): + hyps.append(hyp.split()) + elif params.decoding_method == "modified_beam_search_lm_shallow_fusion": + hyp_tokens = modified_beam_search_lm_shallow_fusion( + model=model, + encoder_out=encoder_out, + encoder_out_lens=encoder_out_lens, + beam=params.beam_size, + LM=LM, + ) + for hyp in sp.decode(hyp_tokens): + hyps.append(hyp.split()) + elif params.decoding_method == "modified_beam_search_LODR": + hyp_tokens = modified_beam_search_LODR( + model=model, + encoder_out=encoder_out, + encoder_out_lens=encoder_out_lens, + beam=params.beam_size, + LODR_lm=ngram_lm, + LODR_lm_scale=ngram_lm_scale, + LM=LM, + context_graph=context_graph, + ) + for hyp in sp.decode(hyp_tokens): + hyps.append(hyp.split()) + elif params.decoding_method == "modified_beam_search_lm_rescore": + lm_scale_list = [0.01 * i for i in range(10, 50)] + ans_dict = modified_beam_search_lm_rescore( + model=model, + encoder_out=encoder_out, + encoder_out_lens=encoder_out_lens, + beam=params.beam_size, + LM=LM, + lm_scale_list=lm_scale_list, + ) + elif params.decoding_method == "modified_beam_search_lm_rescore_LODR": + lm_scale_list = [0.02 * i for i in range(2, 30)] + ans_dict = modified_beam_search_lm_rescore_LODR( + model=model, + encoder_out=encoder_out, + encoder_out_lens=encoder_out_lens, + beam=params.beam_size, + LM=LM, + LODR_lm=ngram_lm, + sp=sp, + lm_scale_list=lm_scale_list, + ) + else: + batch_size = encoder_out.size(0) + + for i in range(batch_size): + # fmt: off + encoder_out_i = encoder_out[i:i+1, :encoder_out_lens[i]] + # fmt: on + if params.decoding_method == "greedy_search": + hyp = greedy_search( + model=model, + encoder_out=encoder_out_i, + max_sym_per_frame=params.max_sym_per_frame, + ) + elif params.decoding_method == "beam_search": + hyp = beam_search( + model=model, + encoder_out=encoder_out_i, + beam=params.beam_size, + ) + else: + raise ValueError( + f"Unsupported decoding method: {params.decoding_method}" + ) + hyps.append(sp.decode(hyp).split()) + + if params.decoding_method == "greedy_search": + return {"greedy_search": hyps} + elif "fast_beam_search" in params.decoding_method: + key = f"beam_{params.beam}_" + key += f"max_contexts_{params.max_contexts}_" + key += f"max_states_{params.max_states}" + if "nbest" in params.decoding_method: + key += f"_num_paths_{params.num_paths}_" + key += f"nbest_scale_{params.nbest_scale}" + if "LG" in params.decoding_method: + key += f"_ngram_lm_scale_{params.ngram_lm_scale}" + + return {key: hyps} + elif "modified_beam_search" in params.decoding_method: + prefix = f"beam_size_{params.beam_size}" + if params.decoding_method in ( + "modified_beam_search_lm_rescore", + "modified_beam_search_lm_rescore_LODR", + ): + ans = dict() + assert ans_dict is not None + for key, hyps in ans_dict.items(): + hyps = [sp.decode(hyp).split() for hyp in hyps] + ans[f"{prefix}_{key}"] = hyps + return ans + else: + if params.has_contexts: + prefix += f"-context-score-{params.context_score}" + return {prefix: hyps} + else: + return {f"beam_size_{params.beam_size}": hyps} + + +def decode_dataset( + dl: torch.utils.data.DataLoader, + params: AttributeDict, + model: nn.Module, + sp: spm.SentencePieceProcessor, + word_table: Optional[k2.SymbolTable] = None, + decoding_graph: Optional[k2.Fsa] = None, + context_graph: Optional[ContextGraph] = None, + LM: Optional[LmScorer] = None, + ngram_lm=None, + ngram_lm_scale: float = 0.0, +) -> Dict[str, List[Tuple[str, List[str], List[str]]]]: + """Decode dataset. + + Args: + dl: + PyTorch's dataloader containing the dataset to decode. + params: + It is returned by :func:`get_params`. + model: + The neural model. + sp: + The BPE model. + word_table: + The word symbol table. + decoding_graph: + The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used + only when --decoding-method is fast_beam_search, fast_beam_search_nbest, + fast_beam_search_nbest_oracle, and fast_beam_search_nbest_LG. + Returns: + Return a dict, whose key may be "greedy_search" if greedy search + is used, or it may be "beam_7" if beam size of 7 is used. + Its value is a list of tuples. Each tuple contains two elements: + The first is the reference transcript, and the second is the + predicted result. + """ + num_cuts = 0 + + try: + num_batches = len(dl) + except TypeError: + num_batches = "?" + + if params.decoding_method == "greedy_search": + log_interval = 50 + else: + log_interval = 20 + + results = defaultdict(list) + for batch_idx, batch in enumerate(dl): + texts = batch["supervisions"]["text"] + cut_ids = [cut.id for cut in batch["cuts"]] + + hyps_dict = decode_one_batch( + params=params, + model=model, + sp=sp, + decoding_graph=decoding_graph, + context_graph=context_graph, + word_table=word_table, + batch=batch, + LM=LM, + ngram_lm=ngram_lm, + ngram_lm_scale=ngram_lm_scale, + ) + + for name, hyps in hyps_dict.items(): + this_batch = [] + assert len(hyps) == len(texts) + for cut_id, hyp_words, ref_text in zip(cut_ids, hyps, texts): + ref_words = ref_text.split() + this_batch.append((cut_id, ref_words, hyp_words)) + + results[name].extend(this_batch) + + num_cuts += len(texts) + + if batch_idx % log_interval == 0: + batch_str = f"{batch_idx}/{num_batches}" + + logging.info(f"batch {batch_str}, cuts processed until now is {num_cuts}") + return results + + +def save_results( + params: AttributeDict, + test_set_name: str, + results_dict: Dict[str, List[Tuple[str, List[str], List[str]]]], +): + test_set_wers = dict() + for key, results in results_dict.items(): + recog_path = ( + params.res_dir / f"recogs-{test_set_name}-{key}-{params.suffix}.txt" + ) + results = sorted(results) + store_transcripts(filename=recog_path, texts=results) + logging.info(f"The transcripts are stored in {recog_path}") + + # The following prints out WERs, per-word error statistics and aligned + # ref/hyp pairs. + errs_filename = ( + params.res_dir / f"errs-{test_set_name}-{key}-{params.suffix}.txt" + ) + with open(errs_filename, "w") as f: + wer = write_error_stats( + f, f"{test_set_name}-{key}", results, enable_log=True + ) + test_set_wers[key] = wer + + logging.info("Wrote detailed error stats to {}".format(errs_filename)) + + test_set_wers = sorted(test_set_wers.items(), key=lambda x: x[1]) + errs_info = ( + params.res_dir / f"wer-summary-{test_set_name}-{key}-{params.suffix}.txt" + ) + with open(errs_info, "w") as f: + print("settings\tWER", file=f) + for key, val in test_set_wers: + print("{}\t{}".format(key, val), file=f) + + s = "\nFor {}, WER of different settings are:\n".format(test_set_name) + note = "\tbest for {}".format(test_set_name) + for key, val in test_set_wers: + s += "{}\t{}{}\n".format(key, val, note) + note = "" + logging.info(s) + + +@torch.no_grad() +def main(): + parser = get_parser() + LibriSpeechAsrDataModule.add_arguments(parser) + LmScorer.add_arguments(parser) + args = parser.parse_args() + args.exp_dir = Path(args.exp_dir) + + params = get_params() + params.update(vars(args)) + + assert params.decoding_method in ( + "greedy_search", + "beam_search", + "fast_beam_search", + "fast_beam_search_nbest", + "fast_beam_search_nbest_LG", + "fast_beam_search_nbest_oracle", + "modified_beam_search", + "modified_beam_search_LODR", + "modified_beam_search_lm_shallow_fusion", + "modified_beam_search_lm_rescore", + "modified_beam_search_lm_rescore_LODR", + ) + params.res_dir = params.exp_dir / params.decoding_method + + if os.path.exists(params.context_file): + params.has_contexts = True + else: + params.has_contexts = False + + if params.iter > 0: + params.suffix = f"iter-{params.iter}-avg-{params.avg}" + else: + params.suffix = f"epoch-{params.epoch}-avg-{params.avg}" + + if "fast_beam_search" in params.decoding_method: + params.suffix += f"-beam-{params.beam}" + params.suffix += f"-max-contexts-{params.max_contexts}" + params.suffix += f"-max-states-{params.max_states}" + if "nbest" in params.decoding_method: + params.suffix += f"-nbest-scale-{params.nbest_scale}" + params.suffix += f"-num-paths-{params.num_paths}" + if "LG" in params.decoding_method: + params.suffix += f"-ngram-lm-scale-{params.ngram_lm_scale}" + elif "beam_search" in params.decoding_method: + params.suffix += f"-{params.decoding_method}-beam-size-{params.beam_size}" + if params.decoding_method in ( + "modified_beam_search", + "modified_beam_search_LODR", + ): + if params.has_contexts: + params.suffix += f"-context-score-{params.context_score}" + else: + params.suffix += f"-context-{params.context_size}" + params.suffix += f"-max-sym-per-frame-{params.max_sym_per_frame}" + + if params.use_shallow_fusion: + params.suffix += f"-{params.lm_type}-lm-scale-{params.lm_scale}" + + if "LODR" in params.decoding_method: + params.suffix += ( + f"-LODR-{params.tokens_ngram}gram-scale-{params.ngram_lm_scale}" + ) + + if params.use_averaged_model: + params.suffix += "-use-averaged-model" + + setup_logger(f"{params.res_dir}/log-decode-{params.suffix}") + logging.info("Decoding started") + + device = torch.device("cpu") + if torch.cuda.is_available(): + device = torch.device("cuda", 0) + + logging.info(f"Device: {device}") + + sp = spm.SentencePieceProcessor() + sp.load(params.bpe_model) + + # and are defined in local/train_bpe_model.py + params.blank_id = sp.piece_to_id("") + params.unk_id = sp.piece_to_id("") + params.vocab_size = sp.get_piece_size() + + logging.info(params) + + logging.info("About to create model") + model = get_model(params) + + if not params.use_averaged_model: + if params.iter > 0: + filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[ + : params.avg + ] + if len(filenames) == 0: + raise ValueError( + f"No checkpoints found for" + f" --iter {params.iter}, --avg {params.avg}" + ) + elif len(filenames) < params.avg: + raise ValueError( + f"Not enough checkpoints ({len(filenames)}) found for" + f" --iter {params.iter}, --avg {params.avg}" + ) + logging.info(f"averaging {filenames}") + model.to(device) + model.load_state_dict(average_checkpoints(filenames, device=device)) + elif params.avg == 1: + load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model) + else: + start = params.epoch - params.avg + 1 + filenames = [] + for i in range(start, params.epoch + 1): + if i >= 1: + filenames.append(f"{params.exp_dir}/epoch-{i}.pt") + logging.info(f"averaging {filenames}") + model.to(device) + model.load_state_dict(average_checkpoints(filenames, device=device)) + else: + if params.iter > 0: + filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[ + : params.avg + 1 + ] + if len(filenames) == 0: + raise ValueError( + f"No checkpoints found for" + f" --iter {params.iter}, --avg {params.avg}" + ) + elif len(filenames) < params.avg + 1: + raise ValueError( + f"Not enough checkpoints ({len(filenames)}) found for" + f" --iter {params.iter}, --avg {params.avg}" + ) + filename_start = filenames[-1] + filename_end = filenames[0] + logging.info( + "Calculating the averaged model over iteration checkpoints" + f" from {filename_start} (excluded) to {filename_end}" + ) + model.to(device) + model.load_state_dict( + average_checkpoints_with_averaged_model( + filename_start=filename_start, + filename_end=filename_end, + device=device, + ) + ) + else: + assert params.avg > 0, params.avg + start = params.epoch - params.avg + assert start >= 1, start + filename_start = f"{params.exp_dir}/epoch-{start}.pt" + filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt" + logging.info( + f"Calculating the averaged model over epoch range from " + f"{start} (excluded) to {params.epoch}" + ) + model.to(device) + model.load_state_dict( + average_checkpoints_with_averaged_model( + filename_start=filename_start, + filename_end=filename_end, + device=device, + ) + ) + + model.to(device) + model.eval() + + # only load the neural network LM if required + if params.use_shallow_fusion or params.decoding_method in ( + "modified_beam_search_lm_rescore", + "modified_beam_search_lm_rescore_LODR", + "modified_beam_search_lm_shallow_fusion", + "modified_beam_search_LODR", + ): + LM = LmScorer( + lm_type=params.lm_type, + params=params, + device=device, + lm_scale=params.lm_scale, + ) + LM.to(device) + LM.eval() + else: + LM = None + + # only load N-gram LM when needed + if params.decoding_method == "modified_beam_search_lm_rescore_LODR": + try: + import kenlm + except ImportError: + print("Please install kenlm first. You can use") + print(" pip install https://github.com/kpu/kenlm/archive/master.zip") + print("to install it") + import sys + + sys.exit(-1) + ngram_file_name = str(params.lang_dir / f"{params.tokens_ngram}gram.arpa") + logging.info(f"lm filename: {ngram_file_name}") + ngram_lm = kenlm.Model(ngram_file_name) + ngram_lm_scale = None # use a list to search + + elif params.decoding_method == "modified_beam_search_LODR": + lm_filename = f"{params.tokens_ngram}gram.fst.txt" + logging.info(f"Loading token level lm: {lm_filename}") + ngram_lm = NgramLm( + str(params.lang_dir / lm_filename), + backoff_id=params.backoff_id, + is_binary=False, + ) + logging.info(f"num states: {ngram_lm.lm.num_states}") + ngram_lm_scale = params.ngram_lm_scale + else: + ngram_lm = None + ngram_lm_scale = None + + if "fast_beam_search" in params.decoding_method: + if params.decoding_method == "fast_beam_search_nbest_LG": + lexicon = Lexicon(params.lang_dir) + word_table = lexicon.word_table + lg_filename = params.lang_dir / "LG.pt" + logging.info(f"Loading {lg_filename}") + decoding_graph = k2.Fsa.from_dict( + torch.load(lg_filename, map_location=device) + ) + decoding_graph.scores *= params.ngram_lm_scale + else: + word_table = None + decoding_graph = k2.trivial_graph(params.vocab_size - 1, device=device) + else: + decoding_graph = None + word_table = None + + if "modified_beam_search" in params.decoding_method: + if os.path.exists(params.context_file): + contexts = [] + for line in open(params.context_file).readlines(): + contexts.append((sp.encode(line.strip()), 0.0)) + context_graph = ContextGraph(params.context_score) + context_graph.build(contexts) + else: + context_graph = None + else: + context_graph = None + + num_param = sum([p.numel() for p in model.parameters()]) + logging.info(f"Number of model parameters: {num_param}") + + # we need cut ids to display recognition results. + args.return_cuts = True + librispeech = LibriSpeechAsrDataModule(args) + + dev_clean_cuts = librispeech.dev_clean_cuts() + dev_other_cuts = librispeech.dev_other_cuts() + + dev_clean_dl = librispeech.test_dataloaders( + dev_clean_cuts, + do_normalize=params.do_normalize, + ) + dev_other_dl = librispeech.test_dataloaders( + dev_other_cuts, + do_normalize=params.do_normalize, + ) + + test_clean_cuts = librispeech.test_clean_cuts() + test_other_cuts = librispeech.test_other_cuts() + + test_clean_dl = librispeech.test_dataloaders( + test_clean_cuts, + do_normalize=params.do_normalize, + ) + test_other_dl = librispeech.test_dataloaders( + test_other_cuts, + do_normalize=params.do_normalize, + ) + + test_sets = ["dev-clean", "dev-other", "test-clean", "test-other"] + test_dl = [dev_clean_dl, dev_other_dl, test_clean_dl, test_other_dl] + + for test_set, test_dl in zip(test_sets, test_dl): + results_dict = decode_dataset( + dl=test_dl, + params=params, + model=model, + sp=sp, + word_table=word_table, + decoding_graph=decoding_graph, + context_graph=context_graph, + LM=LM, + ngram_lm=ngram_lm, + ngram_lm_scale=ngram_lm_scale, + ) + + save_results( + params=params, + test_set_name=test_set, + results_dict=results_dict, + ) + + logging.info("Done!") + + +if __name__ == "__main__": + main() diff --git a/egs/librispeech/SSL/zipformer/decoder.py b/egs/librispeech/SSL/zipformer/decoder.py new file mode 120000 index 0000000000..a2138e5da4 --- /dev/null +++ b/egs/librispeech/SSL/zipformer/decoder.py @@ -0,0 +1 @@ +../../ASR/zipformer/decoder.py \ No newline at end of file diff --git a/egs/librispeech/SSL/zipformer/encoder_interface.py b/egs/librispeech/SSL/zipformer/encoder_interface.py new file mode 120000 index 0000000000..0afd669f2c --- /dev/null +++ b/egs/librispeech/SSL/zipformer/encoder_interface.py @@ -0,0 +1 @@ +../../ASR/zipformer/encoder_interface.py \ No newline at end of file diff --git a/egs/librispeech/SSL/zipformer/finetune.py b/egs/librispeech/SSL/zipformer/finetune.py new file mode 100644 index 0000000000..bbb4453202 --- /dev/null +++ b/egs/librispeech/SSL/zipformer/finetune.py @@ -0,0 +1,1551 @@ +#!/usr/bin/env python3 +# Copyright 2021-2024 Xiaomi Corp. (authors: Fangjun Kuang, +# Wei Kang, +# Mingshuang Luo, +# Zengwei Yao, +# Yifan Yang, +# Daniel Povey) +# Copyright 2024 Shanghai Jiao Tong University (authors: Jianheng Zhuo) +# +# 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="0,1,2,3,4,5,6,7" + +# For HuBERT model finetuning: +./hubert/finetune.py \ + --world-size 8 \ + --num-epochs 200 \ + --start-epoch 1 \ + --use-fp16 1 \ + --exp-dir hubert/exp \ + --full-libri 0 \ + --max-duration 1000 + +It supports finetuning with: + - transducer loss (default), with `--use-transducer True --use-ctc False` + - ctc loss (not recommended), with `--use-transducer False --use-ctc True` + - transducer loss & ctc loss, with `--use-transducer True --use-ctc True` +""" + + +import argparse +import copy +import logging +import warnings +from pathlib import Path +from shutil import copyfile +from typing import Any, Dict, Optional, Tuple, Union + +import k2 +import optim +import sentencepiece as spm +import torch +import torch.multiprocessing as mp +import torch.nn as nn +from asr_datamodule import LibriSpeechAsrDataModule +from decoder import Decoder +from hubert_ce import HubertModel +from joiner import Joiner +from lhotse.cut import Cut +from lhotse.dataset.sampling.base import CutSampler +from lhotse.utils import fix_random_seed +from model import AsrModel +from optim import Eden, ScaledAdam +from 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.hooks import register_inf_check_hooks +from icefall.utils import ( + AttributeDict, + MetricsTracker, + get_parameter_groups_with_lrs, + setup_logger, + str2bool, +) + +LRSchedulerType = Union[torch.optim.lr_scheduler._LRScheduler, optim.LRScheduler] + + +def get_adjusted_batch_count(params: AttributeDict) -> float: + # returns the number of batches we would have used so far if we had used the reference + # duration. This is for purposes of set_batch_count(). + return ( + params.batch_idx_train + * params.accum_grad + * (params.max_duration * params.world_size) + / params.ref_duration + ) + + +def set_batch_count(model: Union[nn.Module, DDP], batch_count: float) -> None: + if isinstance(model, DDP): + # get underlying nn.Module + model = model.module + for name, module in model.named_modules(): + if hasattr(module, "batch_count"): + module.batch_count = batch_count + if hasattr(module, "name"): + module.name = name + + +def add_model_arguments(parser: argparse.ArgumentParser): + parser.add_argument( + "--num-encoder-layers", + type=str, + default="2,2,3,4,3,2", + help="Number of zipformer encoder layers per stack, comma separated.", + ) + + parser.add_argument( + "--downsampling-factor", + type=str, + default="1,2,4,8,4,2", + help="Downsampling factor for each stack of encoder layers.", + ) + + parser.add_argument( + "--feedforward-dim", + type=str, + default="512,768,1024,1536,1024,768", + help="Feedforward dimension of the zipformer encoder layers, per stack, comma separated.", + ) + + parser.add_argument( + "--num-heads", + type=str, + default="4,4,4,8,4,4", + help="Number of attention heads in the zipformer encoder layers: a single int or comma-separated list.", + ) + + parser.add_argument( + "--encoder-dim", + type=str, + default="192,256,384,512,384,256", + help="Embedding dimension in encoder stacks: a single int or comma-separated list.", + ) + + parser.add_argument( + "--query-head-dim", + type=str, + default="32", + help="Query/key dimension per head in encoder stacks: a single int or comma-separated list.", + ) + + parser.add_argument( + "--value-head-dim", + type=str, + default="12", + help="Value dimension per head in encoder stacks: a single int or comma-separated list.", + ) + + parser.add_argument( + "--pos-head-dim", + type=str, + default="4", + help="Positional-encoding dimension per head in encoder stacks: a single int or comma-separated list.", + ) + + parser.add_argument( + "--pos-dim", + type=int, + default="48", + help="Positional-encoding embedding dimension", + ) + + parser.add_argument( + "--encoder-unmasked-dim", + type=str, + default="192,192,256,256,256,192", + help="Unmasked dimensions in the encoders, relates to augmentation during training. " + "A single int or comma-separated list. Must be <= each corresponding encoder_dim.", + ) + + parser.add_argument( + "--cnn-module-kernel", + type=str, + default="31,31,15,15,15,31", + help="Sizes of convolutional kernels in convolution modules in each encoder stack: " + "a single int or comma-separated list.", + ) + + # hubert parameters + parser.add_argument( + "--label-rate", + type=float, + default=50, + ) + + parser.add_argument( + "--sample-rate", + type=float, + default=16000, + ) + + parser.add_argument( + "--extractor-mode", + type=str, + default="default", + help="""mode for feature extractor, should in EXTRACTOR_MODE_CHOICES. default has a single group + norm with d groups in the first conv block, whereas layer_norm + has layer norms in every block (meant to use with normalize=True)""", + ) + + parser.add_argument( + "--conv-feature-layers", + type=str, + default="[(512,10,5)] + [(512,3,2)] * 4 + [(512,2,2)] * 2", + help="string describing convolutional feature extraction layers in form of a python list that contains [(dim, kernel_size, stride), ...]", + ) + + parser.add_argument( + "--conv-bias", type=bool, default=False, help="include bias in conv encoder" + ) + + parser.add_argument( + "--feature-grad-mult", + type=float, + default=1.0, + help="multiply feature extractor var grads by this", + ) + + # masking + parser.add_argument("--mask-length", type=int, default=10, help="mask_length") + + parser.add_argument( + "--mask-prob", + type=float, + default=0.65, + help="probability of replacing a token with mask", + ) + + parser.add_argument( + "--mask-selection", + type=str, + choices=["static", "uniform", "normal", "poisson"], + default="static", + help="how to choose mask length", + ) + + parser.add_argument( + "--mask-other", + type=float, + default=0, + help="secondary mask argument (used for more complex distributions),see help in compute_mask_indicesh", + ) + + parser.add_argument( + "--no-mask-overlap", + type=bool, + default=False, + help="whether to allow masks to overlap", + ) + + parser.add_argument( + "--mask-min-space", + type=int, + default=1, + help="min space between spans (if no overlap is enabled)", + ) + + # channel masking + parser.add_argument( + "--mask-channel-length", + type=int, + default=10, + help="length of the mask for features (channels)", + ) + + parser.add_argument( + "--mask-channel-prob", + type=float, + default=0.0, + help="probability of replacing a feature with 0", + ) + + parser.add_argument( + "--mask-channel-selection", + type=str, + choices=["static", "uniform", "normal", "poisson"], + default="static", + help="how to choose mask length for channel masking", + ) + + parser.add_argument( + "--mask-channel-other", + type=float, + default=0, + help="secondary mask argument (used for more complex distributions), see help in compute_mask_indicesh", + ) + + parser.add_argument( + "--no-mask-channel-overlap", + type=bool, + default=False, + help="whether to allow channel masks to overlap", + ) + + parser.add_argument( + "--mask-channel-min-space", + type=int, + default=1, + help="min space between spans (if no overlap is enabled)", + ) + + # loss computation + parser.add_argument( + "--skip-masked", + type=bool, + default=False, + help="skip computing losses over masked frames", + ) + + parser.add_argument( + "--skip-nomask", + type=bool, + default=False, + help="skip computing losses over unmasked frames", + ) + + parser.add_argument( + "--checkpoint-activations", + type=bool, + default=False, + help="recompute activations and save memory for extra compute", + ) + + parser.add_argument( + "--pred-masked-weight", + type=float, + default=1, + help="weight for masked part in ssl loss", + ) + + parser.add_argument( + "--pred-nomask-weight", + type=float, + default=0, + help="weight for masked part in ssl loss", + ) + + parser.add_argument( + "--loss-weights", + type=float, + nargs="*", + default=[10], + help="weight for masked part in ssl loss", + ) + + # FP16 optimization + parser.add_argument( + "--required-seq-len-multiple", + type=int, + default=2, + help="pad the input to encoder such that the sequence length is divisible by multiple", + ) + + parser.add_argument( + "--attn-type", type=str, default="", help="if espnet use ESPNET MHA" + ) + + parser.add_argument( + "--pos-enc-type", + type=str, + default="abs", + help="Positional encoding type to use in conformer", + ) + + parser.add_argument( + "--logit-temp", type=float, default=0.1, help="temperature to divide logits by" + ) + + parser.add_argument( + "--dropout-input", + type=float, + default=0.0, + help="dropout to apply to the input (after feat extr)", + ) + + parser.add_argument( + "--dropout-features", + type=float, + default=0.0, + help="dropout to apply to the features (after feat extr)", + ) + + parser.add_argument( + "--num-classes", + type=int, + nargs="*", + default=[504], + help="""num class, a little larger than the number of cluster, + the largest is for padding, + and the value should be the multiple of 4, for faster computation""", + ) + + parser.add_argument( + "--untie-final-proj", + type=bool, + default=False, + help="use separate projection for each target", + ) + + parser.add_argument( + "--decoder-dim", + type=int, + default=512, + help="Embedding dimension in the decoder model.", + ) + + parser.add_argument( + "--joiner-dim", + type=int, + default=512, + help="""Dimension used in the joiner model. + Outputs from the encoder and decoder model are projected + to this dimension before adding. + """, + ) + + parser.add_argument( + "--use-transducer", + type=str2bool, + default=True, + help="If True, use Transducer head.", + ) + + parser.add_argument( + "--use-ctc", + type=str2bool, + default=False, + help="If True, use CTC head.", + ) + + +def get_parser(): + parser = argparse.ArgumentParser( + formatter_class=argparse.ArgumentDefaultsHelpFormatter + ) + + parser.add_argument( + "--world-size", + type=int, + default=1, + help="Number of GPUs for DDP training.", + ) + + parser.add_argument( + "--master-port", + type=int, + default=12354, + help="Master port to use for DDP training.", + ) + + parser.add_argument( + "--tensorboard", + type=str2bool, + default=True, + help="Should various information be logged in tensorboard.", + ) + + parser.add_argument( + "--num-epochs", + type=int, + default=222, + 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="hubert/exp", + help="""The experiment dir. + It specifies the directory where all training related + files, e.g., checkpoints, log, etc, are saved + """, + ) + + parser.add_argument( + "--pretrained-dir", + type=str, + help="""The pretrained model dir. + It specifies the directory where the pretrained checkpoint is saved.""", + ) + + parser.add_argument( + "--bpe-model", + type=str, + default="data/lang_bpe_500/bpe.model", + help="Path to the BPE model", + ) + + parser.add_argument( + "--base-lr", type=float, default=0.001, help="The base learning rate." + ) + + parser.add_argument( + "--lr-batches", + type=float, + default=100000, + 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=100, + help="""Number of epochs that affects how rapidly the learning rate decreases. + """, + ) + + parser.add_argument( + "--ref-duration", + type=float, + default=600, + help="Reference batch duration for purposes of adjusting batch counts for setting various " + "schedules inside the model", + ) + + parser.add_argument( + "--context-size", + type=int, + default=2, + help="The context size in the decoder. 1 means bigram; " "2 means tri-gram", + ) + + parser.add_argument( + "--prune-range", + type=int, + default=5, + help="The prune range for rnnt loss, it means how many symbols(context)" + "we are using to compute the loss", + ) + + parser.add_argument( + "--lm-scale", + type=float, + default=0.25, + help="The scale to smooth the loss with lm " + "(output of prediction network) part.", + ) + + parser.add_argument( + "--am-scale", + type=float, + default=0.0, + help="The scale to smooth the loss with am (output of encoder network)" "part.", + ) + + parser.add_argument( + "--simple-loss-scale", + type=float, + default=0.5, + help="To get pruning ranges, we will calculate a simple version" + "loss(joiner is just addition), this simple loss also uses for" + "training (as a regularization item). We will scale the simple loss" + "with this parameter before adding to the final loss.", + ) + + parser.add_argument( + "--ctc-loss-scale", + type=float, + default=0.2, + help="Scale for CTC loss.", + ) + + parser.add_argument( + "--seed", + type=int, + default=42, + help="The seed for random generators intended for reproducibility", + ) + + parser.add_argument( + "--print-diagnostics", + type=str2bool, + default=False, + help="Accumulate stats on activations, print them and exit.", + ) + + parser.add_argument( + "--sanity-check", + type=str2bool, + default=False, + help="Check if any of the batches in epoch 1 would cause OOM.", + ) + + parser.add_argument( + "--inf-check", + type=str2bool, + default=False, + help="Add hooks to check for infinite module outputs and gradients.", + ) + + parser.add_argument( + "--save-every-n", + type=int, + default=100000, + help="""Save checkpoint after processing this number of batches" + periodically. We save checkpoint to exp-dir/ whenever + params.batch_idx_train % save_every_n == 0. The checkpoint filename + has the form: f'exp-dir/checkpoint-{params.batch_idx_train}.pt' + Note: It also saves checkpoint to `exp-dir/epoch-xxx.pt` at the + end of each epoch where `xxx` is the epoch number counting from 1. + """, + ) + + parser.add_argument( + "--keep-last-k", + type=int, + default=30, + help="""Only keep this number of checkpoints on disk. + For instance, if it is 3, there are only 3 checkpoints + in the exp-dir with filenames `checkpoint-xxx.pt`. + It does not affect checkpoints with name `epoch-xxx.pt`. + """, + ) + + parser.add_argument( + "--average-period", + type=int, + default=200, + help="""Update the averaged model, namely `model_avg`, after processing + this number of batches. `model_avg` is a separate version of model, + in which each floating-point parameter is the average of all the + parameters from the start of training. Each time we take the average, + we do: `model_avg = model * (average_period / batch_idx_train) + + model_avg * ((batch_idx_train - average_period) / batch_idx_train)`. + """, + ) + + parser.add_argument( + "--accum-grad", + type=int, + default=1, + help="""update gradient when batch_idx_train % accum_grad == 0. + """, + ) + + 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 updates happen to the model so far across + epochs. + + - sub_batch_idx_train: It contains number of batch 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 + + - warm_step: The warmup period that dictates the decay of the + scale on "simple" (un-pruned) loss. + """ + params = AttributeDict( + { + "best_train_loss": float("inf"), + "best_valid_loss": float("inf"), + "best_train_epoch": -1, + "best_valid_epoch": -1, + "batch_idx_train": 0, + "sub_batch_idx_train": 0, + "log_interval": 50, + "reset_interval": 200, + "valid_interval": 3000, # For the 100h subset, use 800 + # parameters for pruned RNN-T loss + "warm_step": 2000, + "env_info": get_env_info(), + } + ) + + return params + + +def _to_int_tuple(s: str): + return tuple(map(int, s.split(","))) + + +def get_encoder_model(params: AttributeDict) -> nn.Module: + if hasattr(params, "pretrained_dir"): + logging.info(f"Loading {params.pretrained_dir}") + pretrained = torch.load(params.pretrained_dir) + encoder = HubertModel(params) + encoder.load_state_dict(pretrained["model"]) + else: + encoder = HubertModel(params) + return encoder + + +def get_decoder_model(params: AttributeDict) -> nn.Module: + decoder = Decoder( + vocab_size=params.vocab_size, + decoder_dim=params.decoder_dim, + blank_id=params.blank_id, + context_size=params.context_size, + ) + return decoder + + +def get_joiner_model(params: AttributeDict) -> nn.Module: + joiner = Joiner( + encoder_dim=max(_to_int_tuple(params.encoder_dim)), + decoder_dim=params.decoder_dim, + joiner_dim=params.joiner_dim, + vocab_size=params.vocab_size, + ) + return joiner + + +def get_model(params: AttributeDict) -> nn.Module: + assert params.use_transducer or params.use_ctc, ( + f"At least one of them should be True, " + f"but got params.use_transducer={params.use_transducer}, " + f"params.use_ctc={params.use_ctc}" + ) + + encoder = get_encoder_model(params) + + if params.use_transducer: + decoder = get_decoder_model(params) + joiner = get_joiner_model(params) + else: + decoder = None + joiner = None + + model = AsrModel( + encoder=encoder, + decoder=decoder, + joiner=joiner, + encoder_dim=max(_to_int_tuple(params.encoder_dim)), + decoder_dim=params.decoder_dim, + vocab_size=params.vocab_size, + use_transducer=params.use_transducer, + use_ctc=params.use_ctc, + ) + return model + + +def load_checkpoint_if_available( + params: AttributeDict, + model: nn.Module, + model_avg: nn.Module = None, + optimizer: Optional[torch.optim.Optimizer] = None, + scheduler: Optional[LRSchedulerType] = None, +) -> Optional[Dict[str, Any]]: + """Load checkpoint from file. + + If params.start_batch is positive, it will load the checkpoint from + `params.exp_dir/checkpoint-{params.start_batch}.pt`. Otherwise, if + params.start_epoch is larger than 1, it will load the checkpoint from + `params.start_epoch - 1`. + + Apart from loading state dict for `model` and `optimizer` it also updates + `best_train_epoch`, `best_train_loss`, `best_valid_epoch`, + and `best_valid_loss` in `params`. + + Args: + params: + The return value of :func:`get_params`. + model: + The training model. + model_avg: + The stored model averaged from the start of training. + optimizer: + The optimizer that we are using. + scheduler: + The scheduler that we are using. + Returns: + Return a dict containing previously saved training info. + """ + if params.start_batch > 0: + filename = params.exp_dir / f"checkpoint-{params.start_batch}.pt" + elif params.start_epoch > 1: + filename = params.exp_dir / f"epoch-{params.start_epoch-1}.pt" + else: + return None + + assert filename.is_file(), f"{filename} does not exist!" + + saved_params = load_checkpoint( + filename, + model=model, + model_avg=model_avg, + optimizer=optimizer, + scheduler=scheduler, + ) + + keys = [ + "best_train_epoch", + "best_valid_epoch", + "batch_idx_train", + "best_train_loss", + "best_valid_loss", + ] + for k in keys: + params[k] = saved_params[k] + + if params.start_batch > 0: + if "cur_epoch" in saved_params: + params["start_epoch"] = saved_params["cur_epoch"] + + return saved_params + + +def save_checkpoint( + params: AttributeDict, + model: Union[nn.Module, DDP], + model_avg: Optional[nn.Module] = None, + optimizer: Optional[torch.optim.Optimizer] = None, + scheduler: Optional[LRSchedulerType] = None, + sampler: Optional[CutSampler] = None, + scaler: Optional[GradScaler] = None, + rank: int = 0, +) -> None: + """Save model, optimizer, scheduler and training stats to file. + + Args: + params: + It is returned by :func:`get_params`. + model: + The training model. + model_avg: + The stored model averaged from the start of training. + optimizer: + The optimizer used in the training. + sampler: + The sampler for the training dataset. + scaler: + The scaler used for mix precision training. + """ + if rank != 0: + return + filename = params.exp_dir / f"epoch-{params.cur_epoch}.pt" + save_checkpoint_impl( + filename=filename, + model=model, + model_avg=model_avg, + params=params, + optimizer=optimizer, + scheduler=scheduler, + sampler=sampler, + scaler=scaler, + rank=rank, + ) + + if params.best_train_epoch == params.cur_epoch: + best_train_filename = params.exp_dir / "best-train-loss.pt" + copyfile(src=filename, dst=best_train_filename) + + if params.best_valid_epoch == params.cur_epoch: + best_valid_filename = params.exp_dir / "best-valid-loss.pt" + copyfile(src=filename, dst=best_valid_filename) + + +def compute_loss( + params: AttributeDict, + model: Union[nn.Module, DDP], + sp: spm.SentencePieceProcessor, + batch: dict, + is_training: bool, +) -> Tuple[Tensor, MetricsTracker]: + """ + Compute loss given the model and its inputs. + + Args: + params: + Parameters for training. See :func:`get_params`. + model: + The model for training. It is an instance of Zipformer in our case. + batch: + A batch of data. See `dataset.HubertAsrDataset()` + 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 + audio = batch["audio"].to(device) + padding_mask = batch["padding_mask"].to(device) + + batch_idx_train = params.batch_idx_train + warm_step = params.warm_step + + texts = batch["supervisions"]["text"] + y = sp.encode(texts, out_type=int) + y = k2.RaggedTensor(y) + + with torch.set_grad_enabled(is_training): + simple_loss, pruned_loss, ctc_loss, num_frames = model( + x=audio, + padding_mask=padding_mask, + y=y, + prune_range=params.prune_range, + am_scale=params.am_scale, + lm_scale=params.lm_scale, + ) + + loss = 0.0 + + if params.use_transducer: + s = params.simple_loss_scale + # take down the scale on the simple loss from 1.0 at the start + # to params.simple_loss scale by warm_step. + simple_loss_scale = ( + s + if batch_idx_train >= warm_step + else 1.0 - (batch_idx_train / warm_step) * (1.0 - s) + ) + pruned_loss_scale = ( + 1.0 + if batch_idx_train >= warm_step + else 0.1 + 0.9 * (batch_idx_train / warm_step) + ) + loss += simple_loss_scale * simple_loss + pruned_loss_scale * pruned_loss + + if params.use_ctc: + loss += params.ctc_loss_scale * ctc_loss + + assert loss.requires_grad == is_training + + info = MetricsTracker() + with warnings.catch_warnings(): + warnings.simplefilter("ignore") + info["frames"] = num_frames.sum().item() + + # Note: We use reduction=sum while computing the loss. + info["loss"] = loss.detach().cpu().item() + if params.use_transducer: + info["simple_loss"] = simple_loss.detach().cpu().item() + info["pruned_loss"] = pruned_loss.detach().cpu().item() + if params.use_ctc: + info["ctc_loss"] = ctc_loss.detach().cpu().item() + + return loss, info + + +def compute_validation_loss( + params: AttributeDict, + model: Union[nn.Module, DDP], + sp: spm.SentencePieceProcessor, + valid_dl: torch.utils.data.DataLoader, + world_size: int = 1, +) -> MetricsTracker: + """Run the validation process.""" + model.eval() + + tot_loss = MetricsTracker() + + for batch_idx, batch in enumerate(valid_dl): + loss, loss_info = compute_loss( + params=params, + model=model, + sp=sp, + batch=batch, + is_training=False, + ) + assert loss.requires_grad is False + tot_loss = tot_loss + loss_info + + if world_size > 1: + tot_loss.reduce(loss.device) + + loss_value = tot_loss["loss"] / tot_loss["frames"] + if loss_value < params.best_valid_loss: + params.best_valid_epoch = params.cur_epoch + params.best_valid_loss = loss_value + + return tot_loss + + +def train_one_epoch( + params: AttributeDict, + model: Union[nn.Module, DDP], + optimizer: torch.optim.Optimizer, + scheduler: LRSchedulerType, + sp: spm.SentencePieceProcessor, + train_dl: torch.utils.data.DataLoader, + valid_dl: torch.utils.data.DataLoader, + scaler: GradScaler, + model_avg: Optional[nn.Module] = None, + tb_writer: Optional[SummaryWriter] = None, + world_size: int = 1, + rank: int = 0, +) -> None: + """Train the model for one epoch. + + The training loss from the mean of all frames is saved in + `params.train_loss`. It runs the validation process every + `params.valid_interval` batches. + + Args: + params: + It is returned by :func:`get_params`. + model: + The model for training. + optimizer: + The optimizer we are using. + scheduler: + The learning rate scheduler, we call step() every step. + train_dl: + Dataloader for the training dataset. + valid_dl: + Dataloader for the validation dataset. + scaler: + The scaler used for mix precision training. + model_avg: + The stored model averaged from the start of training. + tb_writer: + Writer to write log messages to tensorboard. + world_size: + Number of nodes in DDP training. If it is 1, DDP is disabled. + rank: + The rank of the node in DDP training. If no DDP is used, it should + be set to 0. + """ + model.train() + + tot_loss = MetricsTracker() + + saved_bad_model = False + + def save_bad_model(suffix: str = ""): + save_checkpoint_impl( + filename=params.exp_dir / f"bad-model{suffix}-{rank}.pt", + model=model, + model_avg=model_avg, + params=params, + optimizer=optimizer, + scheduler=scheduler, + sampler=train_dl.sampler, + scaler=scaler, + rank=0, + ) + + for sub_batch_idx, batch in enumerate(train_dl): + params.sub_batch_idx_train += 1 + batch_idx = sub_batch_idx // params.accum_grad + + if batch_idx % 10 == 0: + set_batch_count(model, get_adjusted_batch_count(params)) + + batch_size = len(batch["supervisions"]["text"]) + + try: + with torch.cuda.amp.autocast(enabled=params.use_fp16): + loss, loss_info = compute_loss( + params=params, + model=model, + sp=sp, + batch=batch, + is_training=True, + ) + # summary stats + tot_loss = (tot_loss * (1 - 1 / params.reset_interval)) + loss_info + + # NOTE: We use reduction==sum and loss is computed over utterances + # in the batch and there is no normalization to it so far. + scaler.scale(loss / params.accum_grad).backward() + + if sub_batch_idx % params.accum_grad == params.accum_grad - 1: + params.batch_idx_train += 1 + scheduler.step_batch(params.batch_idx_train) + + scaler.step(optimizer) + scaler.update() + optimizer.zero_grad() + else: + continue + + except: # noqa + save_bad_model() + display_and_save_batch(batch, params=params, sp=sp) + raise + + if params.print_diagnostics and batch_idx == 5: + return + + if ( + rank == 0 + and params.batch_idx_train > 0 + and params.batch_idx_train % params.average_period == 0 + ): + update_averaged_model( + params=params, + model_cur=model, + model_avg=model_avg, + ) + + if ( + params.batch_idx_train > 0 + and params.batch_idx_train % params.save_every_n == 0 + ): + save_checkpoint_with_global_batch_idx( + out_dir=params.exp_dir, + global_batch_idx=params.batch_idx_train, + model=model, + model_avg=model_avg, + params=params, + optimizer=optimizer, + scheduler=scheduler, + sampler=train_dl.sampler, + scaler=scaler, + rank=rank, + ) + remove_checkpoints( + out_dir=params.exp_dir, + topk=params.keep_last_k, + rank=rank, + ) + + if batch_idx % 100 == 0 and params.use_fp16: + # If the grad scale was less than 1, try increasing it. The _growth_interval + # of the grad scaler is configurable, but we can't configure it to have different + # behavior depending on the current grad scale. + cur_grad_scale = scaler._scale.item() + + if cur_grad_scale < 8.0 or (cur_grad_scale < 32.0 and batch_idx % 400 == 0): + scaler.update(cur_grad_scale * 2.0) + if cur_grad_scale < 0.01: + if not saved_bad_model: + save_bad_model(suffix="-first-warning") + saved_bad_model = True + logging.warning(f"Grad scale is small: {cur_grad_scale}") + if cur_grad_scale < 1.0e-05: + save_bad_model() + raise RuntimeError( + f"grad_scale is too small, exiting: {cur_grad_scale}" + ) + + if batch_idx % params.log_interval == 0: + cur_lr = max(scheduler.get_last_lr()) + cur_grad_scale = scaler._scale.item() if params.use_fp16 else 1.0 + + logging.info( + f"Epoch {params.cur_epoch}, " + f"batch {batch_idx}, loss[{loss_info}], " + f"tot_loss[{tot_loss}], batch size: {batch_size}, " + f"lr: {cur_lr:.2e}, " + + (f"grad_scale: {scaler._scale.item()}" if params.use_fp16 else "") + ) + + if tb_writer is not None: + tb_writer.add_scalar( + "train/learning_rate", cur_lr, params.batch_idx_train + ) + + loss_info.write_summary( + tb_writer, "train/current_", params.batch_idx_train + ) + tot_loss.write_summary(tb_writer, "train/tot_", params.batch_idx_train) + if params.use_fp16: + tb_writer.add_scalar( + "train/grad_scale", cur_grad_scale, params.batch_idx_train + ) + + if batch_idx % params.valid_interval == 0 and not params.print_diagnostics: + logging.info("Computing validation loss") + valid_info = compute_validation_loss( + params=params, + model=model, + sp=sp, + valid_dl=valid_dl, + world_size=world_size, + ) + model.train() + logging.info(f"Epoch {params.cur_epoch}, validation: {valid_info}") + logging.info( + f"Maximum memory allocated so far is {torch.cuda.max_memory_allocated()//1000000}MB" + ) + if tb_writer is not None: + valid_info.write_summary( + tb_writer, "train/valid_", params.batch_idx_train + ) + + if batch_idx % params.accum_grad != params.accum_grad - 1: + optimizer.zero_grad() + loss_value = tot_loss["loss"] / tot_loss["frames"] + params.train_loss = loss_value + if params.train_loss < params.best_train_loss: + params.best_train_epoch = params.cur_epoch + params.best_train_loss = params.train_loss + + +def run(rank, world_size, args): + """ + Args: + rank: + It is a value between 0 and `world_size-1`, which is + passed automatically by `mp.spawn()` in :func:`main`. + The node with rank 0 is responsible for saving checkpoint. + world_size: + Number of GPUs for DDP training. + args: + The return value of get_parser().parse_args() + """ + params = get_params() + params.update(vars(args)) + + fix_random_seed(params.seed) + if world_size > 1: + setup_dist(rank, world_size, params.master_port) + + setup_logger(f"{params.exp_dir}/log/log-train") + logging.info("Training started") + + if args.tensorboard and rank == 0: + tb_writer = SummaryWriter(log_dir=f"{params.exp_dir}/tensorboard") + else: + tb_writer = None + + device = torch.device("cpu") + if torch.cuda.is_available(): + device = torch.device("cuda", rank) + logging.info(f"Device: {device}") + + sp = spm.SentencePieceProcessor() + sp.load(params.bpe_model) + + # is defined in local/train_bpe_model.py + params.blank_id = sp.piece_to_id("") + params.vocab_size = sp.get_piece_size() + + if not params.use_transducer: + params.ctc_loss_scale = 1.0 + + logging.info(params) + + logging.info("About to create model") + model = get_model(params) + + num_param = sum([p.numel() for p in model.parameters()]) + logging.info(f"Number of model parameters: {num_param}") + + assert params.save_every_n >= params.average_period + model_avg: Optional[nn.Module] = None + if rank == 0: + # model_avg is only used with rank 0 + model_avg = copy.deepcopy(model).to(torch.float64) + + assert params.start_epoch > 0, params.start_epoch + checkpoints = load_checkpoint_if_available( + params=params, model=model, model_avg=model_avg + ) + + model.to(device) + if world_size > 1: + logging.info("Using DDP") + model = DDP(model, device_ids=[rank], find_unused_parameters=True) + + optimizer = ScaledAdam( + get_parameter_groups_with_lrs(model, lr=params.base_lr, include_names=True), + lr=params.base_lr, # should have no effect + clipping_scale=2.0, + ) + + scheduler = Eden(optimizer, params.lr_batches, params.lr_epochs, warmup_batches=0) + + if checkpoints and "optimizer" in checkpoints: + logging.info("Loading optimizer state dict") + optimizer.load_state_dict(checkpoints["optimizer"]) + + if ( + checkpoints + and "scheduler" in checkpoints + and checkpoints["scheduler"] is not None + ): + logging.info("Loading scheduler state dict") + scheduler.load_state_dict(checkpoints["scheduler"]) + + if params.print_diagnostics: + opts = diagnostics.TensorDiagnosticOptions( + 512 + ) # allow 4 megabytes per sub-module + diagnostic = diagnostics.attach_diagnostics(model, opts) + + if params.inf_check: + register_inf_check_hooks(model) + + librispeech = LibriSpeechAsrDataModule(args) + + train_cuts = ( + librispeech.train_all_shuf_cuts() + if params.full_libri + else librispeech.train_clean_100_cuts() + ) + + def remove_short_and_long_utt(c: Cut): + # Keep only utterances with duration between 1 second and 20 seconds + # + # Caution: There is a reason to select 20.0 here. Please see + # ../local/display_manifest_statistics.py + # + # You should use ../local/display_manifest_statistics.py to get + # an utterance duration distribution for your dataset to select + # the threshold + if c.duration < 1.0 or c.duration > 20.0: + # logging.warning( + # f"Exclude cut with ID {c.id} from training. Duration: {c.duration}" + # ) + return False + + return True + + train_cuts = train_cuts.filter(remove_short_and_long_utt) + + if params.start_batch > 0 and checkpoints and "sampler" in checkpoints: + # We only load the sampler's state dict when it loads a checkpoint + # saved in the middle of an epoch + sampler_state_dict = checkpoints["sampler"] + else: + sampler_state_dict = None + + train_dl = librispeech.train_dataloaders( + train_cuts, + do_normalize=params.do_normalize, + 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, + do_normalize=params.do_normalize, + ) + + if params.sanity_check and not params.print_diagnostics: + scan_pessimistic_batches_for_oom( + model=model, + train_dl=train_dl, + optimizer=optimizer, + sp=sp, + params=params, + ) + + scaler = GradScaler(enabled=params.use_fp16, init_scale=1.0) + if checkpoints and "grad_scaler" in checkpoints: + logging.info("Loading grad scaler state dict") + scaler.load_state_dict(checkpoints["grad_scaler"]) + + for epoch in range(params.start_epoch, params.num_epochs + 1): + scheduler.step_epoch(epoch - 1) + fix_random_seed(params.seed + epoch - 1) + train_dl.sampler.set_epoch(epoch - 1) + + if tb_writer is not None: + tb_writer.add_scalar("train/epoch", epoch, params.batch_idx_train) + + params.cur_epoch = epoch + + train_one_epoch( + params=params, + model=model, + model_avg=model_avg, + optimizer=optimizer, + scheduler=scheduler, + sp=sp, + train_dl=train_dl, + valid_dl=valid_dl, + scaler=scaler, + tb_writer=tb_writer, + world_size=world_size, + rank=rank, + ) + + if params.print_diagnostics: + diagnostic.print_diagnostics() + break + + save_checkpoint( + params=params, + model=model, + model_avg=model_avg, + optimizer=optimizer, + scheduler=scheduler, + sampler=train_dl.sampler, + scaler=scaler, + rank=rank, + ) + + logging.info("Done!") + + if world_size > 1: + torch.distributed.barrier() + cleanup_dist() + + +def display_and_save_batch( + batch: dict, + params: AttributeDict, + sp: spm.SentencePieceProcessor, +) -> None: + """Display the batch statistics and save the batch into disk. + + Args: + batch: + A batch of data. See `dataset.HubertAsrDataset()` + for the content in it. + params: + Parameters for training. See :func:`get_params`. + sp: + The BPE model. + """ + from lhotse.utils import uuid4 + + filename = f"{params.exp_dir}/batch-{uuid4()}.pt" + logging.info(f"Saving batch to {filename}") + torch.save(batch, filename) + + audio = batch["audio"] + logging.info(f"audio shape: {audio.shape}") + + y = sp.encode(batch["supervisions"]["text"], out_type=int) + num_tokens = sum(len(i) for i in y) + logging.info(f"num tokens: {num_tokens}") + + +def scan_pessimistic_batches_for_oom( + model: Union[nn.Module, DDP], + train_dl: torch.utils.data.DataLoader, + optimizer: torch.optim.Optimizer, + sp: spm.SentencePieceProcessor, + params: AttributeDict, +): + from lhotse.dataset import find_pessimistic_batches + + logging.info( + "Sanity check -- see if any of the batches in epoch 1 would cause OOM." + ) + batches, crit_values = find_pessimistic_batches(train_dl.sampler) + for criterion, cuts in batches.items(): + batch = train_dl.dataset[cuts] + try: + with torch.cuda.amp.autocast(enabled=params.use_fp16): + loss, _ = compute_loss( + params=params, + model=model, + sp=sp, + batch=batch, + is_training=True, + ) + loss.backward() + optimizer.zero_grad() + except Exception as e: + if "CUDA out of memory" in str(e): + logging.error( + "Your GPU ran out of memory with the current " + "max_duration setting. We recommend decreasing " + "max_duration and trying again.\n" + f"Failing criterion: {criterion} " + f"(={crit_values[criterion]}) ..." + ) + display_and_save_batch(batch, params=params, sp=sp) + raise + logging.info( + f"Maximum memory allocated so far is {torch.cuda.max_memory_allocated()//1000000}MB" + ) + + +def main(): + parser = get_parser() + LibriSpeechAsrDataModule.add_arguments(parser) + args = parser.parse_args() + args.exp_dir = Path(args.exp_dir) + + world_size = args.world_size + assert world_size >= 1 + if world_size > 1: + mp.spawn(run, args=(world_size, args), nprocs=world_size, join=True) + else: + run(rank=0, world_size=1, args=args) + + +torch.set_num_threads(1) +torch.set_num_interop_threads(1) + +if __name__ == "__main__": + main() diff --git a/egs/librispeech/SSL/zipformer/hubert_ce.py b/egs/librispeech/SSL/zipformer/hubert_ce.py new file mode 100644 index 0000000000..ba4e1cdddd --- /dev/null +++ b/egs/librispeech/SSL/zipformer/hubert_ce.py @@ -0,0 +1,601 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# Permission is hereby granted, free of charge, to any person obtaining a copy +# of this software and associated documentation files (the "Software"), to deal +# in the Software without restriction, including without limitation the rights +# to use, copy, modify, merge, publish, distribute, sublicense, and/or sell +# copies of the Software, and to permit persons to whom the Software is +# furnished to do so, subject to the following conditions: +# +# The above copyright notice and this permission notice shall be included in all +# copies or substantial portions of the Software. +# +# THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR +# IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, +# FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE +# AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER +# LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, +# OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE +# SOFTWARE. + +import argparse +import logging +from typing import Dict, List, Optional, Tuple + +import numpy as np +import torch +import torch.nn as nn +import torch.nn.functional as F +from scaling import ScheduledFloat +from utils import GradMultiply, LayerNorm +from wav2vec2_module import ConvFeatureExtractionModel +from zipformer import Zipformer2 + + +def compute_mask_indices( + shape: Tuple[int, int], + padding_mask: Optional[torch.Tensor], + mask_prob: float, + mask_length: int, + mask_type: str = "static", + mask_other: float = 0.0, + min_masks: int = 0, + no_overlap: bool = False, + min_space: int = 0, + require_same_masks: bool = True, + mask_dropout: float = 0.0, + add_masks: bool = False, + seed: Optional[int] = None, + epoch: Optional[int] = None, + indices: Optional[torch.Tensor] = None, + idc_select_ver: int = 1, # 2 to reproduce mask_tokens_dataset + num_mask_ver: int = 2, # 2 to reproduce mask_tokens_dataset +) -> np.ndarray: + """ + Computes random mask spans for a given shape + + Args: + shape: the the shape for which to compute masks. + should be of size 2 where first element is batch size and 2nd is timesteps + padding_mask: optional padding mask of the same size as shape, which will prevent masking padded elements + mask_prob: probability for each token to be chosen as start of the span to be masked. this will be multiplied by + number of timesteps divided by length of mask span to mask approximately this percentage of all elements. + however due to overlaps, the actual number will be smaller (unless no_overlap is True) + mask_type: how to compute mask lengths + static = fixed size + uniform = sample from uniform distribution [mask_other, mask_length*2] + normal = sample from normal distribution with mean mask_length and stdev mask_other. mask is min 1 element + poisson = sample from possion distribution with lambda = mask length + min_masks: minimum number of masked spans + no_overlap: if false, will switch to an alternative recursive algorithm that prevents spans from overlapping + min_space: only used if no_overlap is True, this is how many elements to keep unmasked between spans + require_same_masks: if true, will randomly drop out masks until same amount of masks remains in each sample + mask_dropout: randomly dropout this percentage of masks in each example + """ + + bsz, all_sz = shape + mask = np.full((bsz, all_sz), False) + + if num_mask_ver == 1: + all_num_mask = int( + # add a random number for probabilistic rounding + mask_prob * all_sz / float(mask_length) + + np.random.rand() + ) + all_num_mask = max(min_masks, all_num_mask) + + mask_idcs = [] + for i in range(bsz): + if seed is not None and epoch is not None and indices is not None: + seed_i = int(hash((seed, epoch, indices[i].item())) % 1e6) + else: + seed_i = None + + rng = np.random.default_rng(seed_i) + + if padding_mask is not None: + sz = all_sz - padding_mask[i].long().sum().item() + assert sz >= 0, sz + else: + sz = all_sz + + if num_mask_ver == 1: + if padding_mask is not None: + num_mask = int( + # add a random number for probabilistic rounding + mask_prob * sz / float(mask_length) + + np.random.rand() + ) + num_mask = max(min_masks, num_mask) + else: + num_mask = all_num_mask + elif num_mask_ver == 2: + num_mask = int( + # add a random number for probabilistic rounding + mask_prob * sz / float(mask_length) + + rng.random() + ) + num_mask = max(min_masks, num_mask) + else: + raise ValueError() + + if mask_type == "static": + lengths = np.full(num_mask, mask_length) + elif mask_type == "uniform": + lengths = rng.randint(mask_other, mask_length * 2 + 1, size=num_mask) + elif mask_type == "normal": + lengths = rng.normal(mask_length, mask_other, size=num_mask) + lengths = [max(1, int(round(x))) for x in lengths] + elif mask_type == "poisson": + lengths = rng.poisson(mask_length, size=num_mask) + lengths = [int(round(x)) for x in lengths] + else: + raise Exception("unknown mask selection " + mask_type) + + if sum(lengths) == 0: + if mask_type == "static": + raise ValueError(f"this should never happens") + else: + lengths = [min(mask_length, sz - 1)] + + if no_overlap: + mask_idc = [] + + def arrange(s, e, length, keep_length): + span_start = rng.randint(s, e - length) + mask_idc.extend(span_start + i for i in range(length)) + + new_parts = [] + if span_start - s - min_space >= keep_length: + new_parts.append((s, span_start - min_space + 1)) + if e - span_start - length - min_space > keep_length: + new_parts.append((span_start + length + min_space, e)) + return new_parts + + parts = [(0, sz)] + min_length = min(lengths) + for length in sorted(lengths, reverse=True): + lens = np.fromiter( + (e - s if e - s >= length + min_space else 0 for s, e in parts), + np.int, + ) + l_sum = np.sum(lens) + if l_sum == 0: + break + probs = lens / np.sum(lens) + c = rng.choice(len(parts), p=probs) + s, e = parts.pop(c) + parts.extend(arrange(s, e, length, min_length)) + mask_idc = np.asarray(mask_idc) + else: + if idc_select_ver == 1: + min_len = min(lengths) + if sz - min_len <= num_mask: + min_len = sz - num_mask - 1 + mask_idc = rng.choice(sz - min_len, num_mask, replace=False) + elif idc_select_ver == 2: + mask_idc = rng.choice(sz, num_mask, replace=False) + else: + raise ValueError() + + mask_idc = np.asarray( + [ + mask_idc[j] + offset + for j in range(len(mask_idc)) + for offset in range(lengths[j]) + ] + ) + + mask_idc = np.unique(mask_idc[mask_idc < sz]) + if len(mask_idc) >= sz: + raise ValueError( + ( + f"the entire sequence is masked. " + f"sz={sz}; mask_idc[mask_idc]; " + f"index={indices[i] if indices is not None else None}" + ) + ) + mask_idcs.append(mask_idc) + + target_len = None + if require_same_masks: + if add_masks: + target_len = max([len(m) for m in mask_idcs]) + else: + target_len = min([len(m) for m in mask_idcs]) + + for i, mask_idc in enumerate(mask_idcs): + if target_len is not None and len(mask_idc) > target_len: + mask_idc = rng.choice(mask_idc, target_len, replace=False) + + mask[i, mask_idc] = True + + if target_len is not None and len(mask_idc) < target_len: + unmasked = np.flatnonzero(~mask[i]) + to_mask = rng.choice(unmasked, target_len - len(mask_idc), replace=False) + mask[i, to_mask] = True + + if mask_dropout > 0: + masked = np.flatnonzero(mask[i]) + num_holes = np.rint(len(masked) * mask_dropout).astype(int) + to_drop = rng.choice(masked, num_holes, replace=False) + mask[i, to_drop] = False + + return mask + + +def _to_int_tuple(s: str): + return tuple(map(int, s.split(","))) + + +class HubertModel(nn.Module): + def __init__( + self, + cfg, + ) -> None: + super().__init__() + feature_enc_layers = eval(cfg.conv_feature_layers) # noqa + self.embed = feature_enc_layers[-1][0] + + self.feature_extractor = ConvFeatureExtractionModel( + conv_layers=feature_enc_layers, + dropout=0.0, + mode=cfg.extractor_mode, + conv_bias=cfg.conv_bias, + ) + feature_ds_rate = np.prod([s for _, _, s in feature_enc_layers]) + self.feat2tar_ratio = cfg.label_rate * feature_ds_rate / cfg.sample_rate + encoder_input_dim = _to_int_tuple(cfg.encoder_dim)[0] + encoder_output_dim = max(_to_int_tuple(cfg.encoder_dim)) + self.post_extract_proj = ( + nn.Linear(self.embed, encoder_input_dim) + if self.embed != encoder_input_dim + else None + ) + + self.mask_prob = cfg.mask_prob + self.mask_selection = cfg.mask_selection + self.mask_other = cfg.mask_other + self.mask_length = cfg.mask_length + self.no_mask_overlap = cfg.no_mask_overlap + self.mask_min_space = cfg.mask_min_space + + self.mask_channel_prob = cfg.mask_channel_prob + self.mask_channel_selection = cfg.mask_channel_selection + self.mask_channel_other = cfg.mask_channel_other + self.mask_channel_length = cfg.mask_channel_length + self.no_mask_channel_overlap = cfg.no_mask_channel_overlap + self.mask_channel_min_space = cfg.mask_channel_min_space + + self.dropout_input = nn.Dropout(cfg.dropout_input) + self.dropout_features = nn.Dropout(cfg.dropout_features) + + self.feature_grad_mult = cfg.feature_grad_mult + self.logit_temp = cfg.logit_temp + self.skip_masked = cfg.skip_masked + self.skip_nomask = cfg.skip_nomask + + self.mask_emb = nn.Parameter(torch.FloatTensor(encoder_input_dim).uniform_()) + + self.encoder = Zipformer2( + output_downsampling_factor=1, + downsampling_factor=_to_int_tuple(cfg.downsampling_factor), + num_encoder_layers=_to_int_tuple(cfg.num_encoder_layers), + encoder_dim=_to_int_tuple(cfg.encoder_dim), + encoder_unmasked_dim=_to_int_tuple(cfg.encoder_unmasked_dim), + query_head_dim=_to_int_tuple(cfg.query_head_dim), + pos_head_dim=_to_int_tuple(cfg.pos_head_dim), + value_head_dim=_to_int_tuple(cfg.value_head_dim), + pos_dim=cfg.pos_dim, + num_heads=_to_int_tuple(cfg.num_heads), + feedforward_dim=_to_int_tuple(cfg.feedforward_dim), + cnn_module_kernel=_to_int_tuple(cfg.cnn_module_kernel), + dropout=ScheduledFloat((0.0, 0.3), (20000.0, 0.1)), + warmup_batches=4000.0, + ) + + self.layer_norm = LayerNorm(self.embed) + + self.untie_final_proj = cfg.untie_final_proj + self.final_proj = nn.Linear(encoder_output_dim, sum(cfg.num_classes)) + + # modules below are not needed during fine-tuning + self.num_classes = cfg.num_classes + self.pred_masked_weight = cfg.pred_masked_weight + self.pred_nomask_weight = cfg.pred_nomask_weight + self.loss_weights = cfg.loss_weights + + def upgrade_state_dict_named(self, state_dict, name): + """Upgrade a (possibly old) state dict for new versions of fairseq.""" + + super().upgrade_state_dict_named(state_dict, name) + return state_dict + + def apply_mask(self, x, padding_mask, target_list): + B, T, C = x.shape + if self.mask_prob > 0: + mask_indices = compute_mask_indices( + (B, T), + padding_mask, + self.mask_prob, + self.mask_length, + self.mask_selection, + self.mask_other, + min_masks=2, + no_overlap=self.no_mask_overlap, + min_space=self.mask_min_space, + ) + mask_indices = torch.from_numpy(mask_indices).to(x.device) + x[mask_indices] = self.mask_emb.to(x.dtype) + else: + mask_indices = None + + if self.mask_channel_prob > 0: + mask_channel_indices = compute_mask_indices( + (B, C), + None, + self.mask_channel_prob, + self.mask_channel_length, + self.mask_channel_selection, + self.mask_channel_other, + no_overlap=self.no_mask_channel_overlap, + min_space=self.mask_channel_min_space, + ) + mask_channel_indices = ( + torch.from_numpy(mask_channel_indices) + .to(x.device) + .unsqueeze(1) + .expand(-1, T, -1) + ) + x[mask_channel_indices] = 0 + + return x, mask_indices + + def forward_features(self, source: torch.Tensor) -> torch.Tensor: + if self.feature_grad_mult > 0: + features = self.feature_extractor(source) + if self.feature_grad_mult != 1.0: + features = GradMultiply.apply(features, self.feature_grad_mult) + else: + with torch.no_grad(): + features = self.feature_extractor(source) + return features + + def forward_targets( + self, + features: torch.Tensor, + target_list: List[torch.Tensor], + ) -> Tuple[torch.Tensor, torch.Tensor]: + # Trim features to ensure labels exist and then get aligned labels + feat_tsz = features.size(2) + targ_tsz = min([t.size(1) for t in target_list]) + if self.feat2tar_ratio * feat_tsz > targ_tsz: + feat_tsz = int(targ_tsz / self.feat2tar_ratio) + features = features[..., :feat_tsz] + target_inds = torch.arange(feat_tsz).float() * self.feat2tar_ratio + target_list = [t[:, target_inds.long()] for t in target_list] + return features, target_list + + def forward_padding_mask( + self, + features: torch.Tensor, + padding_mask: torch.Tensor, + ) -> torch.Tensor: + extra = padding_mask.size(1) % features.size(1) + if extra > 0: + padding_mask = padding_mask[:, :-extra] + padding_mask = padding_mask.view(padding_mask.size(0), features.size(1), -1) + padding_mask = padding_mask.all(-1) + return padding_mask + + def forward( + self, + source: torch.Tensor, + target_list: Optional[List[torch.Tensor]] = None, + padding_mask: Optional[torch.Tensor] = None, + mask: bool = True, + features_only: bool = False, + output_layer: Optional[int] = None, + ): + """output layer is 1-based""" + features = self.forward_features(source) + if target_list is not None: + features, target_list = self.forward_targets(features, target_list) + + features_pen = features.float().pow(2).mean() + + features = features.transpose(1, 2) + features = self.layer_norm(features) + unmasked_features = features.clone() + + if padding_mask is not None: + padding_mask = self.forward_padding_mask(features, padding_mask) + + if self.post_extract_proj is not None: + features = self.post_extract_proj(features) + + features = self.dropout_input(features) + unmasked_features = self.dropout_features(unmasked_features) + + if mask: + x, mask_indices = self.apply_mask(features, padding_mask, target_list) + else: + x = features + mask_indices = None + + # feature: (B, T, D), float + # target: (B, T), long + # x: (B, T, D), float -> (T, B, D), float + # padding_mask: (B, T), bool + # mask_indices: (B, T), bool + x = x.transpose(0, 1) + x, x_lens = self.encoder(x, ~padding_mask.sum(dim=-1)) + x = x.transpose(0, 1) + + if features_only: + return {"x": x, "padding_mask": padding_mask, "features": features} + + if not self.skip_masked: + masked_indices = torch.logical_and(~padding_mask, mask_indices) + proj_x_m = self.final_proj(x[masked_indices]) + proj_x_m /= self.logit_temp + logit_m_list = [proj_x_m for _ in range(len(target_list))] + else: + logit_m_list = [None for _ in target_list] + + if not self.skip_nomask: + nomask_indices = torch.logical_and(~padding_mask, ~mask_indices) + proj_x_u = self.final_proj(x[nomask_indices]) + proj_x_u /= self.logit_temp + logit_u_list = [proj_x_u for _ in range(len(target_list))] + else: + logit_u_list = [None for _ in target_list] + + # result = { + # "logit_m_list": logit_m_list, + # "logit_u_list": logit_u_list, + # "padding_mask": padding_mask, + # "features_pen": features_pen, + # } + targ_m_list = target_list[0][masked_indices] + targ_m_list = targ_m_list.long() + targ_m_list = [targ_m_list for _ in range(len(target_list))] + + targ_u_list = target_list[0][nomask_indices] + targ_u_list = targ_u_list.long() + targ_u_list = [targ_u_list for _ in range(len(target_list))] + return self.compute_loss( + logit_m_list, logit_u_list, targ_m_list, targ_u_list, features_pen + ) + + def extract_features( + self, + source: torch.Tensor, + padding_mask: Optional[torch.Tensor] = None, + mask: bool = False, + ret_conv: bool = False, + output_layer: Optional[int] = None, + ) -> Tuple[torch.Tensor, torch.Tensor]: + res = self.forward( + source, + padding_mask=padding_mask, + mask=mask, + features_only=True, + output_layer=output_layer, + ) + feature = res["features"] if ret_conv else res["x"] + return feature, res["padding_mask"] + + def get_logits(self, net_output, is_masked=True): + if is_masked: + logits_list = net_output["logit_m_list"] + else: + logits_list = net_output["logit_u_list"] + logits_list = [x.float() for x in logits_list if x is not None] + return logits_list + + def get_targets(self, net_output, is_masked=True): + logits_list = self.get_logits(net_output, is_masked) + targets_list = [x.new_zeros(x.size(0), dtype=torch.long) for x in logits_list] + return targets_list + + def get_extra_losses(self, net_output): + extra_losses = [] + names = [] + + if "features_pen" in net_output: + extra_losses.append(net_output["features_pen"]) + names.append("features_pen") + + return extra_losses, names + + def remove_pretraining_modules(self): + self.final_proj = None + + def compute_loss( + self, logit_m_list, logit_u_list, targ_m_list, targ_u_list, features_pen + ): + loss = 0.0 + sample_size = 0 + logging_output = {} + reduce = True + reduction = "sum" if reduce else "none" + + loss_m_list = [] + logp_m_list = [x.float() for x in logit_m_list if x is not None] + logp_m_list = torch.cat(logp_m_list) + targ_m_list = torch.cat(targ_m_list) + + loss_m = F.cross_entropy(logp_m_list, targ_m_list, reduction=reduction) + loss_m_list.append(loss_m) + logging_output[f"loss_m_0"] = loss_m.detach().item() + + assert self.pred_masked_weight == 0 or len(logp_m_list) > 0 + if self.pred_masked_weight > 0: + loss += self.pred_masked_weight * sum(loss_m_list) + sample_size += len(targ_m_list) + + loss_u_list = [] + logp_u_list = [x.float() for x in logit_u_list if x is not None] + logp_u_list = torch.cat(logp_u_list) + targ_u_list = torch.cat(targ_u_list) + + loss_u = F.cross_entropy(logp_u_list, targ_u_list, reduction=reduction) + loss_u_list.append(loss_u) + logging_output[f"loss_u_0"] = loss_u.detach().item() + + assert self.pred_nomask_weight == 0 or len(logp_u_list) > 0 + if self.pred_nomask_weight > 0: + loss += self.pred_nomask_weight * sum(loss_u_list) + sample_size += len(targ_u_list) + + if self.loss_weights is not None: + extra_losses = [] + names = [] + extra_losses.append(features_pen) + names.append("features_pen") + if torch.is_tensor(extra_losses): + extra_losses = [extra_losses] + names = [names] + if len(self.loss_weights) == 1 and len(extra_losses) != 1: + self.loss_weights = [self.loss_weights[0]] * len(extra_losses) + assert len(extra_losses) == len( + self.loss_weights + ), f"{len(extra_losses)}, {len(self.loss_weights)}" + for p, n, coef in zip(extra_losses, names, self.loss_weights): + if coef != 0 and p is not None: + p = coef * p.float() * sample_size + loss += p + logging_output[f"loss_{n}"] = p.item() + + logging_output = { + "loss": loss.item() if reduce else loss, + **logging_output, + } + + # for lk in self.log_keys: + # if lk in net_output: + # logging_output[lk] = float((net_output[lk])) + + def compute_correct(logits, target): + if logits.numel() == 0: + return 0, 0 + else: + assert logits.dim() > 1, logits.shape + max = logits.argmax(-1) == target + min = logits.argmin(-1) == target + both = max & min + corr = max.long().sum().item() - both.long().sum().item() + count = max.numel() + return corr, count + + with torch.no_grad(): + corr_m, count_m = compute_correct(logp_m_list, targ_m_list) + logging_output[f"correct_m_0"] = corr_m + logging_output[f"count_m_0"] = count_m + + corr_u, count_u = compute_correct(logp_u_list, targ_u_list) + logging_output[f"correct_u_0"] = corr_u + logging_output[f"count_u_0"] = count_u + + return loss, sample_size, logging_output diff --git a/egs/librispeech/SSL/zipformer/joiner.py b/egs/librispeech/SSL/zipformer/joiner.py new file mode 120000 index 0000000000..aa3362cda4 --- /dev/null +++ b/egs/librispeech/SSL/zipformer/joiner.py @@ -0,0 +1 @@ +../../ASR/zipformer/joiner.py \ No newline at end of file diff --git a/egs/librispeech/SSL/zipformer/model.py b/egs/librispeech/SSL/zipformer/model.py new file mode 100644 index 0000000000..46a968b69e --- /dev/null +++ b/egs/librispeech/SSL/zipformer/model.py @@ -0,0 +1,344 @@ +# Copyright 2021-2024 Xiaomi Corp. (authors: Fangjun Kuang, +# Wei Kang, +# Zengwei Yao, +# Yifan Yang) +# +# See ../../../../LICENSE for clarification regarding multiple authors +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +from typing import Optional, Tuple + +import k2 +import torch +import torch.nn as nn +from scaling import ScaledLinear + +from icefall.utils import add_sos + + +class AsrModel(nn.Module): + def __init__( + self, + encoder, + decoder: Optional[nn.Module] = None, + joiner: Optional[nn.Module] = None, + encoder_dim: int = 768, + decoder_dim: int = 512, + vocab_size: int = 500, + use_transducer: bool = True, + use_ctc: bool = False, + ): + """A joint CTC & Transducer ASR model. + + - Connectionist temporal classification: labelling unsegmented sequence data with recurrent neural networks (http://imagine.enpc.fr/~obozinsg/teaching/mva_gm/papers/ctc.pdf) + - Sequence Transduction with Recurrent Neural Networks (https://arxiv.org/pdf/1211.3711.pdf) + - Pruned RNN-T for fast, memory-efficient ASR training (https://arxiv.org/pdf/2206.13236.pdf) + + Args: + encoder: + It is the transcription network in the paper. Its accepts + inputs: `x` of (N, T, encoder_dim). + It returns two tensors: `logits` of shape (N, T, encoder_dim) and + `logit_lens` of shape (N,). + decoder: + It is the prediction network in the paper. Its input shape + is (N, U) and its output shape is (N, U, decoder_dim). + It should contain one attribute: `blank_id`. + It is used when use_transducer is True. + joiner: + It has two inputs with shapes: (N, T, encoder_dim) and (N, U, decoder_dim). + Its output shape is (N, T, U, vocab_size). Note that its output contains + unnormalized probs, i.e., not processed by log-softmax. + It is used when use_transducer is True. + use_transducer: + Whether use transducer head. Default: True. + use_ctc: + Whether use CTC head. Default: False. + """ + super().__init__() + + assert ( + use_transducer or use_ctc + ), f"At least one of them should be True, but got use_transducer={use_transducer}, use_ctc={use_ctc}" + + self.encoder = encoder + + self.use_transducer = use_transducer + if use_transducer: + # Modules for Transducer head + assert decoder is not None + assert hasattr(decoder, "blank_id") + assert joiner is not None + + self.decoder = decoder + self.joiner = joiner + + self.simple_am_proj = ScaledLinear( + encoder_dim, vocab_size, initial_scale=0.25 + ) + self.simple_lm_proj = ScaledLinear( + decoder_dim, vocab_size, initial_scale=0.25 + ) + else: + assert decoder is None + assert joiner is None + + self.use_ctc = use_ctc + if use_ctc: + # Modules for CTC head + self.ctc_output = nn.Sequential( + nn.Dropout(p=0.1), + nn.Linear(encoder_dim, vocab_size), + nn.LogSoftmax(dim=-1), + ) + + def forward_encoder( + self, + x: torch.Tensor, + padding_mask: Optional[torch.Tensor] = None, + ) -> Tuple[torch.Tensor, torch.Tensor]: + """Compute encoder outputs. + Args: + x: + A 2-D tensor of shape (N, T). + + Returns: + encoder_out: + Encoder output, of shape (N, T, C). + encoder_out_lens: + Encoder output lengths, of shape (N,). + """ + if padding_mask is None: + padding_mask = torch.zeros_like(x, dtype=torch.bool) + + encoder_out, padding_mask = self.encoder.extract_features( + source=x, + padding_mask=padding_mask, + mask=self.encoder.training, + ) + encoder_out_lens = torch.sum(~padding_mask, dim=1) + assert torch.all(encoder_out_lens > 0), encoder_out_lens + + return encoder_out, encoder_out_lens + + def forward_ctc( + self, + encoder_out: torch.Tensor, + encoder_out_lens: torch.Tensor, + targets: torch.Tensor, + target_lengths: torch.Tensor, + ) -> torch.Tensor: + """Compute CTC loss. + Args: + encoder_out: + Encoder output, of shape (N, T, C). + encoder_out_lens: + Encoder output lengths, of shape (N,). + targets: + Target Tensor of shape (sum(target_lengths)). The targets are assumed + to be un-padded and concatenated within 1 dimension. + """ + # Compute CTC log-prob + ctc_output = self.ctc_output(encoder_out) # (N, T, C) + + ctc_loss = torch.nn.functional.ctc_loss( + log_probs=ctc_output.permute(1, 0, 2), # (T, N, C) + targets=targets, + input_lengths=encoder_out_lens, + target_lengths=target_lengths, + reduction="sum", + ) + return ctc_loss + + def forward_transducer( + self, + encoder_out: torch.Tensor, + encoder_out_lens: torch.Tensor, + y: k2.RaggedTensor, + y_lens: torch.Tensor, + prune_range: int = 5, + am_scale: float = 0.0, + lm_scale: float = 0.0, + ) -> Tuple[torch.Tensor, torch.Tensor]: + """Compute Transducer loss. + Args: + encoder_out: + Encoder output, of shape (N, T, C). + encoder_out_lens: + Encoder output lengths, of shape (N,). + y: + A ragged tensor with 2 axes [utt][label]. It contains labels of each + utterance. + prune_range: + The prune range for rnnt loss, it means how many symbols(context) + we are considering for each frame to compute the loss. + am_scale: + The scale to smooth the loss with am (output of encoder network) + part + lm_scale: + The scale to smooth the loss with lm (output of predictor network) + part + """ + # Now for the decoder, i.e., the prediction network + blank_id = self.decoder.blank_id + sos_y = add_sos(y, sos_id=blank_id) + + # sos_y_padded: [B, S + 1], start with SOS. + sos_y_padded = sos_y.pad(mode="constant", padding_value=blank_id) + + # decoder_out: [B, S + 1, decoder_dim] + decoder_out = self.decoder(sos_y_padded) + + # Note: y does not start with SOS + # y_padded : [B, S] + y_padded = y.pad(mode="constant", padding_value=0) + + y_padded = y_padded.to(torch.int64) + boundary = torch.zeros( + (encoder_out.size(0), 4), + dtype=torch.int64, + device=encoder_out.device, + ) + boundary[:, 2] = y_lens + boundary[:, 3] = encoder_out_lens + + lm = self.simple_lm_proj(decoder_out) + am = self.simple_am_proj(encoder_out) + + # if self.training and random.random() < 0.25: + # lm = penalize_abs_values_gt(lm, 100.0, 1.0e-04) + # if self.training and random.random() < 0.25: + # am = penalize_abs_values_gt(am, 30.0, 1.0e-04) + + with torch.cuda.amp.autocast(enabled=False): + simple_loss, (px_grad, py_grad) = k2.rnnt_loss_smoothed( + lm=lm.float(), + am=am.float(), + symbols=y_padded, + termination_symbol=blank_id, + lm_only_scale=lm_scale, + am_only_scale=am_scale, + boundary=boundary, + reduction="sum", + return_grad=True, + ) + + # ranges : [B, T, prune_range] + ranges = k2.get_rnnt_prune_ranges( + px_grad=px_grad, + py_grad=py_grad, + boundary=boundary, + s_range=prune_range, + ) + + # am_pruned : [B, T, prune_range, encoder_dim] + # lm_pruned : [B, T, prune_range, decoder_dim] + am_pruned, lm_pruned = k2.do_rnnt_pruning( + am=self.joiner.encoder_proj(encoder_out), + lm=self.joiner.decoder_proj(decoder_out), + ranges=ranges, + ) + + # logits : [B, T, prune_range, vocab_size] + + # project_input=False since we applied the decoder's input projections + # prior to do_rnnt_pruning (this is an optimization for speed). + logits = self.joiner(am_pruned, lm_pruned, project_input=False) + + with torch.cuda.amp.autocast(enabled=False): + pruned_loss = k2.rnnt_loss_pruned( + logits=logits.float(), + symbols=y_padded, + ranges=ranges, + termination_symbol=blank_id, + boundary=boundary, + reduction="sum", + ) + + return simple_loss, pruned_loss + + def forward( + self, + x: torch.Tensor, + y: k2.RaggedTensor, + padding_mask: Optional[torch.Tensor] = None, + prune_range: int = 5, + am_scale: float = 0.0, + lm_scale: float = 0.0, + ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]: + """ + Args: + x: + A 2-D tensor of shape (N, T). + y: + A ragged tensor with 2 axes [utt][label]. It contains labels of each + utterance. + prune_range: + The prune range for rnnt loss, it means how many symbols(context) + we are considering for each frame to compute the loss. + am_scale: + The scale to smooth the loss with am (output of encoder network) + part + lm_scale: + The scale to smooth the loss with lm (output of predictor network) + part + Returns: + Return the transducer losses and CTC loss, + in form of (simple_loss, pruned_loss, ctc_loss) + + Note: + Regarding am_scale & lm_scale, it will make the loss-function one of + the form: + lm_scale * lm_probs + am_scale * am_probs + + (1-lm_scale-am_scale) * combined_probs + """ + assert x.ndim == 2, x.shape + assert y.num_axes == 2, y.num_axes + + assert x.size(0) == y.dim0, (x.shape, y.dim0) + + # Compute encoder outputs + encoder_out, encoder_out_lens = self.forward_encoder(x, padding_mask) + + row_splits = y.shape.row_splits(1) + y_lens = row_splits[1:] - row_splits[:-1] + + if self.use_transducer: + # Compute transducer loss + simple_loss, pruned_loss = self.forward_transducer( + encoder_out=encoder_out, + encoder_out_lens=encoder_out_lens, + y=y.to(x.device), + y_lens=y_lens, + prune_range=prune_range, + am_scale=am_scale, + lm_scale=lm_scale, + ) + else: + simple_loss = torch.empty(0) + pruned_loss = torch.empty(0) + + if self.use_ctc: + # Compute CTC loss + targets = y.values + ctc_loss = self.forward_ctc( + encoder_out=encoder_out, + encoder_out_lens=encoder_out_lens, + targets=targets, + target_lengths=y_lens, + ) + else: + ctc_loss = torch.empty(0) + + return simple_loss, pruned_loss, ctc_loss, encoder_out_lens diff --git a/egs/librispeech/SSL/zipformer/optim.py b/egs/librispeech/SSL/zipformer/optim.py new file mode 120000 index 0000000000..56b827b8ae --- /dev/null +++ b/egs/librispeech/SSL/zipformer/optim.py @@ -0,0 +1 @@ +../../ASR/zipformer/optim.py \ No newline at end of file diff --git a/egs/librispeech/SSL/zipformer/pretrain.py b/egs/librispeech/SSL/zipformer/pretrain.py new file mode 100644 index 0000000000..5f547e0b87 --- /dev/null +++ b/egs/librispeech/SSL/zipformer/pretrain.py @@ -0,0 +1,1380 @@ +#!/usr/bin/env python3 +# Copyright 2021-2024 Xiaomi Corp. (authors: Fangjun Kuang, +# Wei Kang, +# Mingshuang Luo, +# Zengwei Yao, +# Yifan Yang, +# Daniel Povey) +# Copyright 2024 Shanghai Jiao Tong University (authors: Jianheng Zhuo) +# +# 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="0,1,2,3,4,5,6,7" + +# For hubert model pretraining: +./zipformer/pretrain.py \ + --world-size 8 \ + --num-epochs 400 \ + --start-epoch 1 \ + --use-fp16 1 \ + --exp-dir hubert/exp \ + --full-libri 1 \ + --max-duration 87.5 \ + --accum-grad 4 +""" + + +import argparse +import copy +import logging +import sys +import warnings +from pathlib import Path +from shutil import copyfile +from typing import Any, Dict, Optional, Tuple, Union + +import optim +import torch +import torch.multiprocessing as mp +import torch.nn as nn +from hubert_ce import HubertModel +from lhotse.cut import Cut +from lhotse.dataset.sampling.base import CutSampler +from lhotse.utils import fix_random_seed +from optim import Eden, ScaledAdam +from ssl_datamodule import LibriSpeechDataModule +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.hooks import register_inf_check_hooks +from icefall.utils import ( + AttributeDict, + MetricsTracker, + get_parameter_groups_with_lrs, + setup_logger, + str2bool, +) + +LRSchedulerType = Union[torch.optim.lr_scheduler._LRScheduler, optim.LRScheduler] + + +def get_adjusted_batch_count(params: AttributeDict) -> float: + # returns the number of batches we would have used so far if we had used the reference + # duration. This is for purposes of set_batch_count(). + return ( + params.batch_idx_train + * params.accum_grad + * (params.max_duration * params.world_size) + / params.ref_duration + ) + + +def set_batch_count(model: Union[nn.Module, DDP], batch_count: float) -> None: + if isinstance(model, DDP): + # get underlying nn.Module + model = model.module + for name, module in model.named_modules(): + if hasattr(module, "batch_count"): + module.batch_count = batch_count + if hasattr(module, "name"): + module.name = name + + +def add_model_arguments(parser: argparse.ArgumentParser): + parser.add_argument( + "--num-encoder-layers", + type=str, + default="2,2,3,4,3,2", + help="Number of zipformer encoder layers per stack, comma separated.", + ) + + parser.add_argument( + "--downsampling-factor", + type=str, + default="1,2,4,8,4,2", + help="Downsampling factor for each stack of encoder layers.", + ) + + parser.add_argument( + "--feedforward-dim", + type=str, + default="512,768,1024,1536,1024,768", + help="Feedforward dimension of the zipformer encoder layers, per stack, comma separated.", + ) + + parser.add_argument( + "--num-heads", + type=str, + default="4,4,4,8,4,4", + help="Number of attention heads in the zipformer encoder layers: a single int or comma-separated list.", + ) + + parser.add_argument( + "--encoder-dim", + type=str, + default="192,256,384,512,384,256", + help="Embedding dimension in encoder stacks: a single int or comma-separated list.", + ) + + parser.add_argument( + "--query-head-dim", + type=str, + default="32", + help="Query/key dimension per head in encoder stacks: a single int or comma-separated list.", + ) + + parser.add_argument( + "--value-head-dim", + type=str, + default="12", + help="Value dimension per head in encoder stacks: a single int or comma-separated list.", + ) + + parser.add_argument( + "--pos-head-dim", + type=str, + default="4", + help="Positional-encoding dimension per head in encoder stacks: a single int or comma-separated list.", + ) + + parser.add_argument( + "--pos-dim", + type=int, + default="48", + help="Positional-encoding embedding dimension", + ) + + parser.add_argument( + "--encoder-unmasked-dim", + type=str, + default="192,192,256,256,256,192", + help="Unmasked dimensions in the encoders, relates to augmentation during training. " + "A single int or comma-separated list. Must be <= each corresponding encoder_dim.", + ) + + parser.add_argument( + "--cnn-module-kernel", + type=str, + default="31,31,15,15,15,31", + help="Sizes of convolutional kernels in convolution modules in each encoder stack: " + "a single int or comma-separated list.", + ) + + # hubert parameters + parser.add_argument( + "--label-rate", + type=float, + default=50, + ) + + parser.add_argument( + "--sample-rate", + type=float, + default=16000, + ) + + parser.add_argument( + "--extractor-mode", + type=str, + default="default", + help="""mode for feature extractor, should in EXTRACTOR_MODE_CHOICES. default has a single group + norm with d groups in the first conv block, whereas layer_norm + has layer norms in every block (meant to use with normalize=True)""", + ) + + parser.add_argument( + "--conv-feature-layers", + type=str, + default="[(512,10,5)] + [(512,3,2)] * 4 + [(512,2,2)] * 2", + help="string describing convolutional feature extraction layers in form of a python list that contains [(dim, kernel_size, stride), ...]", + ) + + parser.add_argument( + "--conv-bias", type=bool, default=False, help="include bias in conv encoder" + ) + + parser.add_argument( + "--feature-grad-mult", + type=float, + default=1.0, + help="multiply feature extractor var grads by this", + ) + + # masking + parser.add_argument("--mask-length", type=int, default=10, help="mask_length") + + parser.add_argument( + "--mask-prob", + type=float, + default=0.65, + help="probability of replacing a token with mask", + ) + + parser.add_argument( + "--mask-selection", + type=str, + choices=["static", "uniform", "normal", "poisson"], + default="static", + help="how to choose mask length", + ) + + parser.add_argument( + "--mask-other", + type=float, + default=0, + help="secondary mask argument (used for more complex distributions),see help in compute_mask_indicesh", + ) + + parser.add_argument( + "--no-mask-overlap", + type=bool, + default=False, + help="whether to allow masks to overlap", + ) + + parser.add_argument( + "--mask-min-space", + type=int, + default=1, + help="min space between spans (if no overlap is enabled)", + ) + + # channel masking + parser.add_argument( + "--mask-channel-length", + type=int, + default=10, + help="length of the mask for features (channels)", + ) + + parser.add_argument( + "--mask-channel-prob", + type=float, + default=0.0, + help="probability of replacing a feature with 0", + ) + + parser.add_argument( + "--mask-channel-selection", + type=str, + choices=["static", "uniform", "normal", "poisson"], + default="static", + help="how to choose mask length for channel masking", + ) + + parser.add_argument( + "--mask-channel-other", + type=float, + default=0, + help="secondary mask argument (used for more complex distributions), see help in compute_mask_indicesh", + ) + + parser.add_argument( + "--no-mask-channel-overlap", + type=bool, + default=False, + help="whether to allow channel masks to overlap", + ) + + parser.add_argument( + "--mask-channel-min-space", + type=int, + default=1, + help="min space between spans (if no overlap is enabled)", + ) + + # loss computation + parser.add_argument( + "--skip-masked", + type=bool, + default=False, + help="skip computing losses over masked frames", + ) + + parser.add_argument( + "--skip-nomask", + type=bool, + default=False, + help="skip computing losses over unmasked frames", + ) + + parser.add_argument( + "--checkpoint-activations", + type=bool, + default=False, + help="recompute activations and save memory for extra compute", + ) + + parser.add_argument( + "--pred-masked-weight", + type=float, + default=1, + help="weight for masked part in ssl loss", + ) + + parser.add_argument( + "--pred-nomask-weight", + type=float, + default=0, + help="weight for masked part in ssl loss", + ) + + parser.add_argument( + "--loss-weights", + type=float, + nargs="*", + default=[10], + help="weight for masked part in ssl loss", + ) + + # FP16 optimization + parser.add_argument( + "--required-seq-len-multiple", + type=int, + default=2, + help="pad the input to encoder such that the sequence length is divisible by multiple", + ) + + parser.add_argument( + "--attn-type", type=str, default="", help="if espnet use ESPNET MHA" + ) + + parser.add_argument( + "--pos-enc-type", + type=str, + default="abs", + help="Positional encoding type to use in conformer", + ) + + parser.add_argument( + "--logit-temp", type=float, default=0.1, help="temperature to divide logits by" + ) + + parser.add_argument( + "--dropout-input", + type=float, + default=0.0, + help="dropout to apply to the input (after feat extr)", + ) + + parser.add_argument( + "--dropout-features", + type=float, + default=0.0, + help="dropout to apply to the features (after feat extr)", + ) + + parser.add_argument( + "--num-classes", + type=int, + nargs="*", + default=[504], + help="""num class, a little larger than the number of cluster, + the largest is for padding, + and the value should be the multiple of 4, for faster computation""", + ) + + parser.add_argument( + "--untie-final-proj", + type=bool, + default=False, + help="use separate projection for each target", + ) + + +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=400, + help="Number of epochs to train.", + ) + + parser.add_argument( + "--start-epoch", + type=int, + default=1, + help="""Resume training from this epoch. It should be positive. + If larger than 1, it will load checkpoint from + exp-dir/epoch-{start_epoch-1}.pt + """, + ) + + parser.add_argument( + "--start-batch", + type=int, + default=0, + help="""If positive, --start-epoch is ignored and + it loads the checkpoint from exp-dir/checkpoint-{start_batch}.pt + """, + ) + + parser.add_argument( + "--exp-dir", + type=str, + default="zipformer/exp", + help="""The experiment dir. + It specifies the directory where all training related + files, e.g., checkpoints, log, etc, are saved + """, + ) + + parser.add_argument( + "--base-lr", type=float, default=0.045, help="The base learning rate." + ) + + parser.add_argument( + "--lr-batches", + type=float, + default=7500, + help="""Number of steps that affects how rapidly the learning rate + decreases. We suggest not to change this.""", + ) + + parser.add_argument( + "--lr-epochs", + type=float, + default=10.5, + help="""Number of epochs that affects how rapidly the learning rate decreases. + """, + ) + + parser.add_argument( + "--warmup-batches", + type=float, + default=5000, + help="Eden warmup steps", + ) + + parser.add_argument( + "--warmup-start", + type=float, + default=0, + help="Eden warmup start learning rate", + ) + + parser.add_argument( + "--ref-duration", + type=float, + default=600, + help="Reference batch duration for purposes of adjusting batch counts for setting various " + "schedules inside the model", + ) + + parser.add_argument( + "--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( + "--sanity-check", + type=str2bool, + default=False, + help="Check if any of the batches in epoch 1 would cause OOM.", + ) + + parser.add_argument( + "--inf-check", + type=str2bool, + default=False, + help="Add hooks to check for infinite module outputs and gradients.", + ) + + parser.add_argument( + "--save-every-n", + type=int, + default=100000, + help="""Save checkpoint after processing this number of batches" + periodically. We save checkpoint to exp-dir/ whenever + params.batch_idx_train % save_every_n == 0. The checkpoint filename + has the form: f'exp-dir/checkpoint-{params.batch_idx_train}.pt' + Note: It also saves checkpoint to `exp-dir/epoch-xxx.pt` at the + end of each epoch where `xxx` is the epoch number counting from 1. + """, + ) + + parser.add_argument( + "--keep-last-k", + type=int, + default=30, + help="""Only keep this number of checkpoints on disk. + For instance, if it is 3, there are only 3 checkpoints + in the exp-dir with filenames `checkpoint-xxx.pt`. + It does not affect checkpoints with name `epoch-xxx.pt`. + """, + ) + + parser.add_argument( + "--average-period", + type=int, + default=200, + help="""Update the averaged model, namely `model_avg`, after processing + this number of batches. `model_avg` is a separate version of model, + in which each floating-point parameter is the average of all the + parameters from the start of training. Each time we take the average, + we do: `model_avg = model * (average_period / batch_idx_train) + + model_avg * ((batch_idx_train - average_period) / batch_idx_train)`. + """, + ) + + parser.add_argument( + "--accum-grad", + type=int, + default=4, + help="""update gradient when batch_idx_train % accum_grad == 0. + """, + ) + + parser.add_argument( + "--use-fp16", + type=str2bool, + default=False, + help="Whether to use half precision training.", + ) + + parser.add_argument( + "--max-keep-size", + type=int, + default=sys.maxsize, + help="exclude sample longer than this.", + ) + + parser.add_argument( + "--min-keep-size", + type=float, + default=32000, + help="exclude sample longer less than this.", + ) + + parser.add_argument( + "--max-sample-size", + type=float, + default=250000, + help="max sample size to crop to for batching.", + ) + + 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 updates happen to the model so far across + epochs. + + - sub_batch_idx_train: It contains number of batch 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 + """ + params = AttributeDict( + { + "best_train_loss": float("inf"), + "best_valid_loss": float("inf"), + "best_train_epoch": -1, + "best_valid_epoch": -1, + "batch_idx_train": 0, + "sub_batch_idx_train": 0, + "log_interval": 50, + "reset_interval": 200, + "valid_interval": 3000, # For the 100h subset, use 800 + "env_info": get_env_info(), + } + ) + + return params + + +def _to_int_tuple(s: str): + return tuple(map(int, s.split(","))) + + +def get_model(params: AttributeDict) -> nn.Module: + model = HubertModel(params) + 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], + batch: dict, + is_training: bool, +) -> Tuple[Tensor, MetricsTracker]: + """ + Compute loss given the model and its inputs. + + Args: + params: + Parameters for training. See :func:`get_params`. + model: + The model for training. It is an instance of Zipformer in our case. + batch: + A batch of data. See `dataset.HubertDataset()` + for the content in it. + is_training: + True for training. False for validation. When it is True, this + function enables autograd during computation; when it is False, it + disables autograd. + """ + device = model.device if isinstance(model, DDP) else next(model.parameters()).device + audio = batch["audio"].to(device) + padding_mask = batch["padding_mask"].to(device) + kmeans = batch["kmeans"].to(device) + + with torch.set_grad_enabled(is_training): + loss, num_masked_tokens, logging_output = model( + source=audio, target_list=[kmeans], padding_mask=padding_mask + ) + + assert loss.requires_grad == is_training + + info = MetricsTracker() + with warnings.catch_warnings(): + warnings.simplefilter("ignore") + info["frames"] = num_masked_tokens + for item in logging_output: + info[item] = logging_output[item] + return loss, info + + +def compute_validation_loss( + params: AttributeDict, + model: Union[nn.Module, DDP], + valid_dl: torch.utils.data.DataLoader, + world_size: int = 1, +) -> MetricsTracker: + """Run the validation process.""" + model.eval() + + tot_loss = MetricsTracker() + + for batch_idx, batch in enumerate(valid_dl): + loss, loss_info = compute_loss( + params=params, + model=model, + batch=batch, + is_training=False, + ) + assert loss.requires_grad is False + tot_loss = tot_loss + loss_info + + if world_size > 1: + tot_loss.reduce(loss.device) + + loss_value = tot_loss["loss"] / tot_loss["frames"] + if loss_value < params.best_valid_loss: + params.best_valid_epoch = params.cur_epoch + params.best_valid_loss = loss_value + + return tot_loss + + +def train_one_epoch( + params: AttributeDict, + model: Union[nn.Module, DDP], + optimizer: torch.optim.Optimizer, + scheduler: LRSchedulerType, + train_dl: torch.utils.data.DataLoader, + valid_dl: torch.utils.data.DataLoader, + scaler: GradScaler, + model_avg: Optional[nn.Module] = None, + tb_writer: Optional[SummaryWriter] = None, + world_size: int = 1, + rank: int = 0, +) -> None: + """Train the model for one epoch. + + The training loss from the mean of all frames is saved in + `params.train_loss`. It runs the validation process every + `params.valid_interval` batches. + + Args: + params: + It is returned by :func:`get_params`. + model: + The model for training. + optimizer: + The optimizer we are using. + scheduler: + The learning rate scheduler, we call step() every step. + train_dl: + Dataloader for the training dataset. + valid_dl: + Dataloader for the validation dataset. + scaler: + The scaler used for mix precision training. + model_avg: + The stored model averaged from the start of training. + tb_writer: + Writer to write log messages to tensorboard. + world_size: + Number of nodes in DDP training. If it is 1, DDP is disabled. + rank: + The rank of the node in DDP training. If no DDP is used, it should + be set to 0. + """ + model.train() + + tot_loss = MetricsTracker() + + saved_bad_model = False + + def save_bad_model(suffix: str = ""): + save_checkpoint_impl( + filename=params.exp_dir / f"bad-model{suffix}-{rank}.pt", + model=model, + model_avg=model_avg, + params=params, + optimizer=optimizer, + scheduler=scheduler, + sampler=train_dl.sampler, + scaler=scaler, + rank=0, + ) + + for sub_batch_idx, batch in enumerate(train_dl): + params.sub_batch_idx_train += 1 + batch_idx = sub_batch_idx // params.accum_grad + + if batch_idx % 10 == 0: + set_batch_count(model, get_adjusted_batch_count(params)) + + batch_size = batch["kmeans"].shape[0] + + try: + with torch.cuda.amp.autocast(enabled=params.use_fp16): + loss, loss_info = compute_loss( + params=params, + model=model, + batch=batch, + is_training=True, + ) + # summary stats + tot_loss = (tot_loss * (1 - 1 / params.reset_interval)) + loss_info + + # NOTE: We use reduction==sum and loss is computed over utterances + # in the batch and there is no normalization to it so far. + scaler.scale(loss / params.accum_grad).backward() + + if sub_batch_idx % params.accum_grad == params.accum_grad - 1: + params.batch_idx_train += 1 + scheduler.step_batch(params.batch_idx_train) + + scaler.step(optimizer) + scaler.update() + optimizer.zero_grad() + else: + continue + + except: # noqa + save_bad_model() + display_and_save_batch(batch, params=params) + raise + + if params.print_diagnostics and batch_idx == 5: + return + + if ( + rank == 0 + and params.batch_idx_train > 0 + and params.batch_idx_train % params.average_period == 0 + ): + update_averaged_model( + params=params, + model_cur=model, + model_avg=model_avg, + ) + + if ( + params.batch_idx_train > 0 + and params.batch_idx_train % params.save_every_n == 0 + ): + save_checkpoint_with_global_batch_idx( + out_dir=params.exp_dir, + global_batch_idx=params.batch_idx_train, + model=model, + model_avg=model_avg, + params=params, + optimizer=optimizer, + scheduler=scheduler, + sampler=train_dl.sampler, + scaler=scaler, + rank=rank, + ) + remove_checkpoints( + out_dir=params.exp_dir, + topk=params.keep_last_k, + rank=rank, + ) + + if batch_idx % 100 == 0 and params.use_fp16: + # If the grad scale was less than 1, try increasing it. The _growth_interval + # of the grad scaler is configurable, but we can't configure it to have different + # behavior depending on the current grad scale. + cur_grad_scale = scaler._scale.item() + + if cur_grad_scale < 8.0 or (cur_grad_scale < 32.0 and batch_idx % 400 == 0): + scaler.update(cur_grad_scale * 2.0) + if cur_grad_scale < 0.01: + if not saved_bad_model: + save_bad_model(suffix="-first-warning") + saved_bad_model = True + logging.warning(f"Grad scale is small: {cur_grad_scale}") + if cur_grad_scale < 1.0e-05: + save_bad_model() + raise RuntimeError( + f"grad_scale is too small, exiting: {cur_grad_scale}" + ) + + if batch_idx % params.log_interval == 0: + cur_lr = max(scheduler.get_last_lr()) + cur_grad_scale = scaler._scale.item() if params.use_fp16 else 1.0 + + logging.info( + f"Epoch {params.cur_epoch}, " + f"batch {batch_idx}, loss[{loss_info}], " + f"tot_loss[{tot_loss}], batch size: {batch_size}, " + f"lr: {cur_lr:.2e}, " + + (f"grad_scale: {scaler._scale.item()}" if params.use_fp16 else "") + ) + + if tb_writer is not None: + tb_writer.add_scalar( + "train/learning_rate", cur_lr, params.batch_idx_train + ) + + loss_info.write_summary( + tb_writer, "train/current_", params.batch_idx_train + ) + tot_loss.write_summary(tb_writer, "train/tot_", params.batch_idx_train) + if params.use_fp16: + tb_writer.add_scalar( + "train/grad_scale", cur_grad_scale, params.batch_idx_train + ) + + if batch_idx % params.valid_interval == 0 and not params.print_diagnostics: + logging.info("Computing validation loss") + valid_info = compute_validation_loss( + params=params, + model=model, + valid_dl=valid_dl, + world_size=world_size, + ) + model.train() + logging.info(f"Epoch {params.cur_epoch}, validation: {valid_info}") + logging.info( + f"Maximum memory allocated so far is {torch.cuda.max_memory_allocated()//1000000}MB" + ) + if tb_writer is not None: + valid_info.write_summary( + tb_writer, "train/valid_", params.batch_idx_train + ) + + if batch_idx % params.accum_grad != params.accum_grad - 1: + optimizer.zero_grad() + loss_value = tot_loss["loss"] / tot_loss["frames"] + params.train_loss = loss_value + if params.train_loss < params.best_train_loss: + params.best_train_epoch = params.cur_epoch + params.best_train_loss = params.train_loss + + +def run(rank, world_size, args): + """ + Args: + rank: + It is a value between 0 and `world_size-1`, which is + passed automatically by `mp.spawn()` in :func:`main`. + The node with rank 0 is responsible for saving checkpoint. + world_size: + Number of GPUs for DDP training. + args: + The return value of get_parser().parse_args() + """ + params = get_params() + params.update(vars(args)) + + fix_random_seed(params.seed) + if world_size > 1: + setup_dist(rank, world_size, params.master_port) + + setup_logger(f"{params.exp_dir}/log/log-train") + logging.info("Training started") + + if args.tensorboard and rank == 0: + tb_writer = SummaryWriter(log_dir=f"{params.exp_dir}/tensorboard") + else: + tb_writer = None + + device = torch.device("cpu") + if torch.cuda.is_available(): + device = torch.device("cuda", rank) + logging.info(f"Device: {device}") + logging.info(params) + + logging.info("About to create model") + model = get_model(params) + + num_param = sum([p.numel() for p in model.parameters()]) + logging.info(f"Number of model parameters: {num_param}") + + assert params.save_every_n >= params.average_period + model_avg: Optional[nn.Module] = None + if rank == 0: + # model_avg is only used with rank 0 + model_avg = copy.deepcopy(model).to(torch.float64) + + assert params.start_epoch > 0, params.start_epoch + checkpoints = load_checkpoint_if_available( + params=params, model=model, model_avg=model_avg + ) + + model.to(device) + if world_size > 1: + logging.info("Using DDP") + model = DDP(model, device_ids=[rank], find_unused_parameters=True) + + optimizer = ScaledAdam( + get_parameter_groups_with_lrs(model, lr=params.base_lr, include_names=True), + lr=params.base_lr, # should have no effect + clipping_scale=2.0, + ) + + scheduler = Eden( + optimizer, + params.lr_batches, + params.lr_epochs, + params.warmup_batches, + params.warmup_start, + ) + + if checkpoints and "optimizer" in checkpoints: + logging.info("Loading optimizer state dict") + optimizer.load_state_dict(checkpoints["optimizer"]) + + if ( + checkpoints + and "scheduler" in checkpoints + and checkpoints["scheduler"] is not None + ): + logging.info("Loading scheduler state dict") + scheduler.load_state_dict(checkpoints["scheduler"]) + + if params.print_diagnostics: + opts = diagnostics.TensorDiagnosticOptions( + 512 + ) # allow 4 megabytes per sub-module + diagnostic = diagnostics.attach_diagnostics(model, opts) + + if params.inf_check: + register_inf_check_hooks(model) + + librispeech = LibriSpeechDataModule(args) + + train_cuts = ( + librispeech.train_all_shuf_cuts() + if params.full_libri + else librispeech.train_clean_100_cuts() + ) + + def remove_short_and_long_utt(c: Cut): + # Keep only utterances with duration between 1 second and 20 seconds + # + # Caution: There is a reason to select 20.0 here. Please see + # ../local/display_manifest_statistics.py + # + # You should use ../local/display_manifest_statistics.py to get + # an utterance duration distribution for your dataset to select + # the threshold + if ( + c.duration < params.min_keep_size / params.sample_rate + or c.duration > params.max_keep_size / params.sample_rate + ): + logging.warning( + f"Exclude cut with ID {c.id} from training. Duration: {c.duration}" + ) + return False + + return True + + train_cuts = train_cuts.filter(remove_short_and_long_utt) + + if params.start_batch > 0 and checkpoints and "sampler" in checkpoints: + # We only load the sampler's state dict when it loads a checkpoint + # saved in the middle of an epoch + sampler_state_dict = checkpoints["sampler"] + else: + sampler_state_dict = None + + train_dl = librispeech.train_dataloaders( + train_cuts, + max_sample_size=params.max_sample_size, + sample_rate=params.sample_rate, + label_rate=params.label_rate, + random_crop=params.random_crop, + pad_audio=False, + num_classes=params.num_classes, + do_normalize=params.do_normalize, + sampler_state_dict=sampler_state_dict, + ) + + valid_cuts = librispeech.dev_clean_cuts() + # valid_cuts += librispeech.dev_other_cuts() + valid_cuts = valid_cuts.filter(remove_short_and_long_utt) + + valid_dl = librispeech.valid_dataloaders( + valid_cuts, + max_sample_size=params.max_sample_size, + sample_rate=params.sample_rate, + label_rate=params.label_rate, + random_crop=params.random_crop, + pad_audio=False, + num_classes=params.num_classes, + do_normalize=params.do_normalize, + ) + + if params.sanity_check and not params.print_diagnostics: + scan_pessimistic_batches_for_oom( + model=model, + train_dl=train_dl, + optimizer=optimizer, + params=params, + ) + + scaler = GradScaler(enabled=params.use_fp16, init_scale=1.0) + if checkpoints and "grad_scaler" in checkpoints: + logging.info("Loading grad scaler state dict") + scaler.load_state_dict(checkpoints["grad_scaler"]) + + for epoch in range(params.start_epoch, params.num_epochs + 1): + scheduler.step_epoch(epoch - 1) + fix_random_seed(params.seed + epoch - 1) + train_dl.sampler.set_epoch(epoch - 1) + + if tb_writer is not None: + tb_writer.add_scalar("train/epoch", epoch, params.batch_idx_train) + + params.cur_epoch = epoch + + train_one_epoch( + params=params, + model=model, + model_avg=model_avg, + optimizer=optimizer, + scheduler=scheduler, + train_dl=train_dl, + valid_dl=valid_dl, + scaler=scaler, + tb_writer=tb_writer, + world_size=world_size, + rank=rank, + ) + + if params.print_diagnostics: + diagnostic.print_diagnostics() + break + + save_checkpoint( + params=params, + model=model, + model_avg=model_avg, + optimizer=optimizer, + scheduler=scheduler, + sampler=train_dl.sampler, + scaler=scaler, + rank=rank, + ) + + logging.info("Done!") + + if world_size > 1: + torch.distributed.barrier() + cleanup_dist() + + +def display_and_save_batch( + batch: dict, + params: AttributeDict, +) -> None: + """Display the batch statistics and save the batch into disk. + + Args: + batch: + A batch of data. See `dataset.HubertDataset()` + for the content in it. + params: + Parameters for training. See :func:`get_params`. + sp: + The BPE model. + """ + from lhotse.utils import uuid4 + + filename = f"{params.exp_dir}/batch-{uuid4()}.pt" + logging.info(f"Saving batch to {filename}") + torch.save(batch, filename) + + audio = batch["audio"] + logging.info(f"audio shape: {audio.shape}") + + +def scan_pessimistic_batches_for_oom( + model: Union[nn.Module, DDP], + train_dl: torch.utils.data.DataLoader, + optimizer: torch.optim.Optimizer, + params: AttributeDict, +): + from lhotse.dataset import find_pessimistic_batches + + logging.info( + "Sanity check -- see if any of the batches in epoch 1 would cause OOM." + ) + batches, crit_values = find_pessimistic_batches(train_dl.sampler) + for criterion, cuts in batches.items(): + batch = train_dl.dataset[cuts] + try: + with torch.cuda.amp.autocast(enabled=params.use_fp16): + loss, _ = compute_loss( + params=params, + model=model, + batch=batch, + is_training=True, + ) + loss.backward() + optimizer.zero_grad() + except Exception as e: + if "CUDA out of memory" in str(e): + logging.error( + "Your GPU ran out of memory with the current " + "max_duration setting. We recommend decreasing " + "max_duration and trying again.\n" + f"Failing criterion: {criterion} " + f"(={crit_values[criterion]}) ..." + ) + display_and_save_batch(batch, params=params) + raise + logging.info( + f"Maximum memory allocated so far is {torch.cuda.max_memory_allocated()//1000000}MB" + ) + + +def main(): + parser = get_parser() + LibriSpeechDataModule.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/SSL/zipformer/scaling.py b/egs/librispeech/SSL/zipformer/scaling.py new file mode 120000 index 0000000000..e30bd99de2 --- /dev/null +++ b/egs/librispeech/SSL/zipformer/scaling.py @@ -0,0 +1 @@ +../../ASR/zipformer/scaling.py \ No newline at end of file diff --git a/egs/librispeech/SSL/zipformer/ssl_datamodule.py b/egs/librispeech/SSL/zipformer/ssl_datamodule.py new file mode 120000 index 0000000000..9f5085e3af --- /dev/null +++ b/egs/librispeech/SSL/zipformer/ssl_datamodule.py @@ -0,0 +1 @@ +../hubert/ssl_datamodule.py \ No newline at end of file diff --git a/egs/librispeech/SSL/zipformer/utils.py b/egs/librispeech/SSL/zipformer/utils.py new file mode 100644 index 0000000000..748d3c96e3 --- /dev/null +++ b/egs/librispeech/SSL/zipformer/utils.py @@ -0,0 +1,337 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# Permission is hereby granted, free of charge, to any person obtaining a copy +# of this software and associated documentation files (the "Software"), to deal +# in the Software without restriction, including without limitation the rights +# to use, copy, modify, merge, publish, distribute, sublicense, and/or sell +# copies of the Software, and to permit persons to whom the Software is +# furnished to do so, subject to the following conditions: +# +# The above copyright notice and this permission notice shall be included in all +# copies or substantial portions of the Software. +# +# THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR +# IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, +# FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE +# AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER +# LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, +# OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE +# SOFTWARE. + +import math +from typing import Callable, List, Optional + +import torch +import torch.nn as nn +import torch.nn.functional as F + + +def relu_squared(x: torch.Tensor): + return F.relu(x).pow(2) + + +def gelu_accurate(x): + if not hasattr(gelu_accurate, "_a"): + gelu_accurate._a = math.sqrt(2 / math.pi) + return ( + 0.5 * x * (1 + torch.tanh(gelu_accurate._a * (x + 0.044715 * torch.pow(x, 3)))) + ) + + +def is_xla_tensor(tensor): + return torch.is_tensor(tensor) and tensor.device.type == "xla" + + +def index_put(tensor, indices, value): + if is_xla_tensor(tensor): + for _ in range(indices.dim(), tensor.dim()): + indices = indices.unsqueeze(-1) + if indices.size(-1) < tensor.size(-1): + indices = indices.expand_as(tensor) + tensor = torch.mul(tensor, ~indices) + torch.mul(value, indices) + else: + tensor[indices] = value + return tensor + + +def pad_to_multiple(x, multiple, dim=-1, value=0): + # Inspired from https://github.com/lucidrains/local-attention/blob/master/local_attention/local_attention.py#L41 + if x is None: + return None, 0 + tsz = x.size(dim) + m = tsz / multiple + remainder = math.ceil(m) * multiple - tsz + if m.is_integer(): + return x, 0 + pad_offset = (0,) * (-1 - dim) * 2 + + return F.pad(x, (*pad_offset, 0, remainder), value=value), remainder + + +def gelu(x: torch.Tensor) -> torch.Tensor: + return torch.nn.functional.gelu(x.float()).type_as(x) + + +def get_activation_fn(activation: str) -> Callable: + """Returns the activation function corresponding to `activation`""" + if activation == "relu": + return F.relu + elif activation == "relu_squared": + return relu_squared + elif activation == "gelu": + return gelu + elif activation == "gelu_fast": + return gelu_accurate + elif activation == "gelu_accurate": + return gelu_accurate + elif activation == "tanh": + return torch.tanh + elif activation == "linear": + return lambda x: x + elif activation == "swish": + return torch.nn.SiLU + else: + raise RuntimeError("--activation-fn {} not supported".format(activation)) + + +class SamePad(nn.Module): + def __init__(self, kernel_size, causal=False): + super().__init__() + if causal: + self.remove = kernel_size - 1 + else: + self.remove = 1 if kernel_size % 2 == 0 else 0 + + def forward(self, x): + if self.remove > 0: + x = x[:, :, : -self.remove] + return x + + +class SamePad2d(nn.Module): + def __init__(self, kernel_size): + super().__init__() + self.remove = 1 if kernel_size % 2 == 0 else 0 + + def forward(self, x): + assert len(x.size()) == 4 + if self.remove > 0: + x = x[:, :, : -self.remove, : -self.remove] + return x + + +class TransposeLast(nn.Module): + def __init__(self, deconstruct_idx=None, tranpose_dim=-2): + super().__init__() + self.deconstruct_idx = deconstruct_idx + self.tranpose_dim = tranpose_dim + + def forward(self, x): + if self.deconstruct_idx is not None: + x = x[self.deconstruct_idx] + return x.transpose(self.tranpose_dim, -1) + + +try: + from apex.normalization import FusedLayerNorm as _FusedLayerNorm + + has_fused_layernorm = True + + class FusedLayerNorm(_FusedLayerNorm): + @torch.jit.unused + def forward(self, x): + if not x.is_cuda: + return super().forward(x) + else: + with torch.cuda.device(x.device): + return super().forward(x) + +except ImportError: + has_fused_layernorm = False + + +def LayerNorm(normalized_shape, eps=1e-5, elementwise_affine=True, export=False): + if torch.jit.is_scripting() or torch.jit.is_tracing(): + export = True + if not export and torch.cuda.is_available() and has_fused_layernorm: + return FusedLayerNorm(normalized_shape, eps, elementwise_affine) + return torch.nn.LayerNorm(normalized_shape, eps, elementwise_affine) + + +class Fp32LayerNorm(nn.LayerNorm): + def __init__(self, *args, **kwargs): + super().__init__(*args, **kwargs) + + def forward(self, input): + output = F.layer_norm( + input.float(), + self.normalized_shape, + self.weight.float() if self.weight is not None else None, + self.bias.float() if self.bias is not None else None, + self.eps, + ) + return output.type_as(input) + + +class Fp32GroupNorm(nn.GroupNorm): + def __init__(self, *args, **kwargs): + super().__init__(*args, **kwargs) + + def forward(self, input): + output = F.group_norm( + input.float(), + self.num_groups, + self.weight.float() if self.weight is not None else None, + self.bias.float() if self.bias is not None else None, + self.eps, + ) + return output.type_as(input) + + +def softmax(x, dim: int, onnx_trace: bool = False): + if onnx_trace: + return F.softmax(x.float(), dim=dim) + else: + return F.softmax(x, dim=dim, dtype=torch.float32) + + +def quant_noise(module, p, block_size): + """ + Wraps modules and applies quantization noise to the weights for + subsequent quantization with Iterative Product Quantization as + described in "Training with Quantization Noise for Extreme Model Compression" + + Args: + - module: nn.Module + - p: amount of Quantization Noise + - block_size: size of the blocks for subsequent quantization with iPQ + + Remarks: + - Module weights must have the right sizes wrt the block size + - Only Linear, Embedding and Conv2d modules are supported for the moment + - For more detail on how to quantize by blocks with convolutional weights, + see "And the Bit Goes Down: Revisiting the Quantization of Neural Networks" + - We implement the simplest form of noise here as stated in the paper + which consists in randomly dropping blocks + """ + + # if no quantization noise, don't register hook + if p <= 0: + return module + + # supported modules + assert isinstance(module, (nn.Linear, nn.Embedding, nn.Conv2d)) + + # test whether module.weight has the right sizes wrt block_size + is_conv = module.weight.ndim == 4 + + # 2D matrix + if not is_conv: + assert ( + module.weight.size(1) % block_size == 0 + ), "Input features must be a multiple of block sizes" + + # 4D matrix + else: + # 1x1 convolutions + if module.kernel_size == (1, 1): + assert ( + module.in_channels % block_size == 0 + ), "Input channels must be a multiple of block sizes" + # regular convolutions + else: + k = module.kernel_size[0] * module.kernel_size[1] + assert k % block_size == 0, "Kernel size must be a multiple of block size" + + def _forward_pre_hook(mod, input): + # no noise for evaluation + if mod.training: + if not is_conv: + # gather weight and sizes + weight = mod.weight + in_features = weight.size(1) + out_features = weight.size(0) + + # split weight matrix into blocks and randomly drop selected blocks + mask = torch.zeros( + in_features // block_size * out_features, device=weight.device + ) + mask.bernoulli_(p) + mask = mask.repeat_interleave(block_size, -1).view(-1, in_features) + + else: + # gather weight and sizes + weight = mod.weight + in_channels = mod.in_channels + out_channels = mod.out_channels + + # split weight matrix into blocks and randomly drop selected blocks + if mod.kernel_size == (1, 1): + mask = torch.zeros( + int(in_channels // block_size * out_channels), + device=weight.device, + ) + mask.bernoulli_(p) + mask = mask.repeat_interleave(block_size, -1).view(-1, in_channels) + else: + mask = torch.zeros( + weight.size(0), weight.size(1), device=weight.device + ) + mask.bernoulli_(p) + mask = ( + mask.unsqueeze(2) + .unsqueeze(3) + .repeat(1, 1, mod.kernel_size[0], mod.kernel_size[1]) + ) + + # scale weights and apply mask + mask = mask.to( + torch.bool + ) # x.bool() is not currently supported in TorchScript + s = 1 / (1 - p) + mod.weight.data = s * weight.masked_fill(mask, 0) + + module.register_forward_pre_hook(_forward_pre_hook) + return module + + +class FairseqDropout(nn.Module): + def __init__(self, p, module_name=None): + super().__init__() + self.p = p + self.module_name = module_name + self.apply_during_inference = False + + def forward(self, x, inplace: bool = False): + if self.p > 0 and (self.training or self.apply_during_inference): + return F.dropout(x, p=self.p, training=True, inplace=inplace) + else: + return x + + def make_generation_fast_( + self, + name: str, + retain_dropout: bool = False, + retain_dropout_modules: Optional[List[str]] = None, + **kwargs + ): + if retain_dropout: + if retain_dropout_modules is not None and self.module_name is None: + pass + elif ( + retain_dropout_modules is None # if None, apply to all modules + or self.module_name in retain_dropout_modules + ): + self.apply_during_inference = True + + +class GradMultiply(torch.autograd.Function): + @staticmethod + def forward(ctx, x, scale): + ctx.scale = scale + res = x.new(x) + return res + + @staticmethod + def backward(ctx, grad): + return grad * ctx.scale, None diff --git a/egs/librispeech/SSL/zipformer/wav2vec2_module.py b/egs/librispeech/SSL/zipformer/wav2vec2_module.py new file mode 100644 index 0000000000..ab5ca005f4 --- /dev/null +++ b/egs/librispeech/SSL/zipformer/wav2vec2_module.py @@ -0,0 +1,108 @@ +# Copyright (c) Facebook, Inc. and its affiliates. +# +# Permission is hereby granted, free of charge, to any person obtaining a copy +# of this software and associated documentation files (the "Software"), to deal +# in the Software without restriction, including without limitation the rights +# to use, copy, modify, merge, publish, distribute, sublicense, and/or sell +# copies of the Software, and to permit persons to whom the Software is +# furnished to do so, subject to the following conditions: +# +# The above copyright notice and this permission notice shall be included in all +# copies or substantial portions of the Software. +# +# THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR +# IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, +# FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE +# AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER +# LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, +# OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE +# SOFTWARE. + +import math +from typing import List, Tuple + +import numpy as np +import torch +import torch.nn as nn +import torch.nn.functional as F +from utils import Fp32GroupNorm, Fp32LayerNorm, TransposeLast + + +class ConvFeatureExtractionModel(nn.Module): + def __init__( + self, + conv_layers: List[Tuple[int, int, int]], + dropout: float = 0.0, + mode: str = "default", + conv_bias: bool = False, + ): + super().__init__() + + assert mode in {"default", "layer_norm"} + + def block( + n_in, + n_out, + k, + stride, + is_layer_norm=False, + is_group_norm=False, + conv_bias=False, + ): + def make_conv(): + conv = nn.Conv1d(n_in, n_out, k, stride=stride, bias=conv_bias) + nn.init.kaiming_normal_(conv.weight) + return conv + + assert ( + is_layer_norm and is_group_norm + ) == False, "layer norm and group norm are exclusive" + + if is_layer_norm: + return nn.Sequential( + make_conv(), + nn.Dropout(p=dropout), + nn.Sequential( + TransposeLast(), + Fp32LayerNorm(dim, elementwise_affine=True), + TransposeLast(), + ), + nn.GELU(), + ) + elif is_group_norm: + return nn.Sequential( + make_conv(), + nn.Dropout(p=dropout), + Fp32GroupNorm(dim, dim, affine=True), + nn.GELU(), + ) + else: + return nn.Sequential(make_conv(), nn.Dropout(p=dropout), nn.GELU()) + + in_d = 1 + self.conv_layers = nn.ModuleList() + for i, cl in enumerate(conv_layers): + assert len(cl) == 3, "invalid conv definition: " + str(cl) + (dim, k, stride) = cl + + self.conv_layers.append( + block( + in_d, + dim, + k, + stride, + is_layer_norm=mode == "layer_norm", + is_group_norm=mode == "default" and i == 0, + conv_bias=conv_bias, + ) + ) + in_d = dim + + def forward(self, x): + # BxT -> BxCxT + x = x.unsqueeze(1) + + for conv in self.conv_layers: + x = conv(x) + + return x diff --git a/egs/librispeech/SSL/zipformer/zipformer.py b/egs/librispeech/SSL/zipformer/zipformer.py new file mode 100644 index 0000000000..e9eff3357e --- /dev/null +++ b/egs/librispeech/SSL/zipformer/zipformer.py @@ -0,0 +1,2438 @@ +#!/usr/bin/env python3 +# Copyright 2022-2023 Xiaomi Corp. (authors: Daniel Povey, +# 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 logging +import math +import random +import warnings +from typing import List, Optional, Tuple, Union + +import torch +from encoder_interface import EncoderInterface +from scaling import ( + Identity, # more friendly to backward hooks than nn.Identity(), for diagnostic reasons. +) +from scaling import ( + ScaledLinear, # not as in other dirs.. just scales down initial parameter values. +) +from scaling import ( + ActivationDropoutAndLinear, + Balancer, + BiasNorm, + ChunkCausalDepthwiseConv1d, + Dropout2, + FloatLike, + ScheduledFloat, + Whiten, + convert_num_channels, + limit_param_value, + penalize_abs_values_gt, + softmax, +) +from torch import Tensor, nn + + +class Zipformer2(EncoderInterface): + """ + Args: + + Note: all "int or Tuple[int]" arguments below will be treated as lists of the same length + as downsampling_factor if they are single ints or one-element tuples. The length of + downsampling_factor defines the number of stacks. + + output_downsampling_factor (int): how much to downsample at the output. Note: + we also downsample by a factor of 2 in the Conv2dSubsampling encoder. + You should probably leave this at 2. + downsampling_factor (Tuple[int]): downsampling factor for each encoder stack. + Note: this is in addition to the downsampling factor of 2 that is applied in + the frontend (self.encoder_embed). + encoder_dim (Tuple[int]): embedding dimension of each of the encoder stacks, one per + encoder stack. + num_encoder_layers (int or Tuple[int])): number of encoder layers for each stack + encoder_unmasked_dim (int or Tuple[int]): unmasked dimension in each of + the encoder stacks for purposes of per-frame dropout (recommend 256 for + now). + query_head_dim (int or Tuple[int]): dimension of query and key per attention + head: per stack, if a tuple.. + pos_head_dim (int or Tuple[int]): dimension of positional-encoding projection per + attention head + value_head_dim (int or Tuple[int]): dimension of value in each attention head + num_heads: (int or Tuple[int]): number of heads in the self-attention mechanism. + Must be at least 4. + feedforward_dim (int or Tuple[int]): hidden dimension in feedforward modules + cnn_module_kernel (int or Tuple[int])): Kernel size of convolution module + + pos_dim (int): the dimension of each positional-encoding vector prior to projection, + e.g. 128. + + dropout (float): dropout rate + warmup_batches (float): number of batches to warm up over; this controls + dropout of encoder layers. + causal (bool): if True, support chunkwise causal convolution. This should + not hurt WER as no modeling power is lost, but the convolution modules will be + slightly slower and use more memory. Enables use of the chunk_size and + left_context_chunks options in forward(), which simulates streaming + decoding. + chunk_size: (list of int): only set this to other than [-1] if causal; + the chunk size will be randomly chosen from this list. -1 means no chunking. + left_context_frames: (list of int): determines the number of left- + context chunks for causal training; will be rounded to a number of + chunks. Must not be less than cnn_module_kernel (after factoring in + rounding and downsampling); an error will be thrown if this is violated. + """ + + def __init__( + self, + output_downsampling_factor: int = 2, + downsampling_factor: Tuple[int] = (2, 4), + encoder_dim: Union[int, Tuple[int]] = 384, + num_encoder_layers: Union[int, Tuple[int]] = 4, + encoder_unmasked_dim: Union[int, Tuple[int]] = 256, + query_head_dim: Union[int, Tuple[int]] = 24, + pos_head_dim: Union[int, Tuple[int]] = 4, + value_head_dim: Union[int, Tuple[int]] = 12, + num_heads: Union[int, Tuple[int]] = 8, + feedforward_dim: Union[int, Tuple[int]] = 1536, + cnn_module_kernel: Union[int, Tuple[int]] = 31, + pos_dim: int = 192, + dropout: FloatLike = None, # see code below for default + warmup_batches: float = 4000.0, + causal: bool = False, + chunk_size: Tuple[int] = [-1], + left_context_frames: Tuple[int] = [-1], + ) -> None: + super(Zipformer2, self).__init__() + + if dropout is None: + dropout = ScheduledFloat((0.0, 0.3), (20000.0, 0.1)) + + def _to_tuple(x): + """Converts a single int or a 1-tuple of an int to a tuple with the same length + as downsampling_factor""" + if isinstance(x, int): + x = (x,) + if len(x) == 1: + x = x * len(downsampling_factor) + else: + assert len(x) == len(downsampling_factor) and isinstance(x[0], int) + return x + + self.output_downsampling_factor = output_downsampling_factor # int + self.downsampling_factor = downsampling_factor # tuple + self.encoder_dim = encoder_dim = _to_tuple(encoder_dim) # tuple + self.encoder_unmasked_dim = encoder_unmasked_dim = _to_tuple( + encoder_unmasked_dim + ) # tuple + num_encoder_layers = _to_tuple(num_encoder_layers) + self.num_encoder_layers = num_encoder_layers + self.query_head_dim = query_head_dim = _to_tuple(query_head_dim) + self.value_head_dim = value_head_dim = _to_tuple(value_head_dim) + pos_head_dim = _to_tuple(pos_head_dim) + self.num_heads = num_heads = _to_tuple(num_heads) + feedforward_dim = _to_tuple(feedforward_dim) + self.cnn_module_kernel = cnn_module_kernel = _to_tuple(cnn_module_kernel) + + self.causal = causal + self.chunk_size = chunk_size + self.left_context_frames = left_context_frames + + for u, d in zip(encoder_unmasked_dim, encoder_dim): + assert u <= d + + # each one will be Zipformer2Encoder or DownsampledZipformer2Encoder + encoders = [] + + num_encoders = len(downsampling_factor) + for i in range(num_encoders): + encoder_layer = Zipformer2EncoderLayer( + embed_dim=encoder_dim[i], + pos_dim=pos_dim, + num_heads=num_heads[i], + query_head_dim=query_head_dim[i], + pos_head_dim=pos_head_dim[i], + value_head_dim=value_head_dim[i], + feedforward_dim=feedforward_dim[i], + dropout=dropout, + cnn_module_kernel=cnn_module_kernel[i], + causal=causal, + ) + + # For the segment of the warmup period, we let the Conv2dSubsampling + # layer learn something. Then we start to warm up the other encoders. + encoder = Zipformer2Encoder( + encoder_layer, + num_encoder_layers[i], + pos_dim=pos_dim, + dropout=dropout, + warmup_begin=warmup_batches * (i + 1) / (num_encoders + 1), + warmup_end=warmup_batches * (i + 2) / (num_encoders + 1), + final_layerdrop_rate=0.035 * (downsampling_factor[i] ** 0.5), + ) + + if downsampling_factor[i] != 1: + encoder = DownsampledZipformer2Encoder( + encoder, + dim=encoder_dim[i], + downsample=downsampling_factor[i], + dropout=dropout, + ) + + encoders.append(encoder) + + self.encoders = nn.ModuleList(encoders) + + self.downsample_output = SimpleDownsample( + max(encoder_dim), downsample=output_downsampling_factor, dropout=dropout + ) + + def get_feature_masks(self, x: Tensor) -> Union[List[float], List[Tensor]]: + """ + In eval mode, returns [1.0] * num_encoders; in training mode, returns a number of + randomized feature masks, one per encoder. + On e.g. 15% of frames, these masks will zero out all enocder dims larger than + some supplied number, e.g. >256, so in effect on those frames we are using + a smaller encoer dim. + + We generate the random masks at this level because we want the 2 masks to 'agree' + all the way up the encoder stack. This will mean that the 1st mask will have + mask values repeated self.zipformer_subsampling_factor times. + + Args: + x: the embeddings (needed for the shape and dtype and device), of shape + (1, batch_size, encoder_dims0) + """ + num_encoders = len(self.encoder_dim) + if not self.training: + return [1.0] * num_encoders + + (num_frames0, batch_size, _encoder_dims0) = x.shape + + assert self.encoder_dim[0] == _encoder_dims0, ( + self.encoder_dim[0], + _encoder_dims0, + ) + + feature_mask_dropout_prob = 0.125 + + # mask1 shape: (1, batch_size, 1) + mask1 = ( + torch.rand(1, batch_size, 1, device=x.device) > feature_mask_dropout_prob + ).to(x.dtype) + + # mask2 has additional sequences masked, about twice the number. + mask2 = torch.logical_and( + mask1, + ( + torch.rand(1, batch_size, 1, device=x.device) + > feature_mask_dropout_prob + ).to(x.dtype), + ) + + # dim: (1, batch_size, 2) + mask = torch.cat((mask1, mask2), dim=-1) + + feature_masks = [] + for i in range(num_encoders): + channels = self.encoder_dim[i] + feature_mask = torch.ones( + 1, batch_size, channels, dtype=x.dtype, device=x.device + ) + u1 = self.encoder_unmasked_dim[i] + u2 = u1 + (channels - u1) // 2 + + feature_mask[:, :, u1:u2] *= mask[..., 0:1] + feature_mask[:, :, u2:] *= mask[..., 1:2] + + feature_masks.append(feature_mask) + + return feature_masks + + def get_chunk_info(self) -> Tuple[int, int]: + """ + Returns chunk_size and left_context_chunks. + """ + if not self.causal: + return -1, -1 + + if torch.jit.is_scripting() or torch.jit.is_tracing(): + assert len(self.chunk_size) == 1, self.chunk_size + chunk_size = self.chunk_size[0] + else: + chunk_size = random.choice(self.chunk_size) + + if chunk_size == -1: + left_context_chunks = -1 + else: + if torch.jit.is_scripting() or torch.jit.is_tracing(): + assert len(self.left_context_frames) == 1, self.left_context_frames + left_context_frames = self.left_context_frames[0] + else: + left_context_frames = random.choice(self.left_context_frames) + # Note: in Python, -1 // n == -1 for n > 0 + left_context_chunks = left_context_frames // chunk_size + if left_context_chunks == 0: + left_context_chunks = 1 + + return chunk_size, left_context_chunks + + def forward( + self, + x: Tensor, + x_lens: Tensor, + src_key_padding_mask: Optional[Tensor] = None, + ) -> Tuple[Tensor, Tensor]: + """ + Args: + x: + The input tensor. Its shape is (seq_len, batch_size, feature_dim). + x_lens: + A tensor of shape (batch_size,) containing the number of frames in + `x` before padding. + src_key_padding_mask: + The mask for padding, of shape (batch_size, seq_len); True means + masked position. May be None. + Returns: + Return a tuple containing 2 tensors: + - embeddings: its shape is (output_seq_len, batch_size, max(encoder_dim)) + - lengths, a tensor of shape (batch_size,) containing the number + of frames in `embeddings` before padding. + """ + outputs = [] + if torch.jit.is_scripting() or torch.jit.is_tracing(): + feature_masks = [1.0] * len(self.encoder_dim) + else: + feature_masks = self.get_feature_masks(x) + + chunk_size, left_context_chunks = self.get_chunk_info() + + if torch.jit.is_scripting() or torch.jit.is_tracing(): + # Not support exporting a model for simulating streaming decoding + attn_mask = None + else: + attn_mask = self._get_attn_mask(x, chunk_size, left_context_chunks) + + for i, module in enumerate(self.encoders): + ds = self.downsampling_factor[i] + x = convert_num_channels(x, self.encoder_dim[i]) + + x = module( + x, + chunk_size=chunk_size, + feature_mask=feature_masks[i], + src_key_padding_mask=( + None + if src_key_padding_mask is None + else src_key_padding_mask[..., ::ds] + ), + attn_mask=attn_mask, + ) + outputs.append(x) + + # if the last output has the largest dimension, x will be unchanged, + # it will be the same as outputs[-1]. Otherwise it will be concatenated + # from different pieces of 'outputs', taking each dimension from the + # most recent output that has it present. + x = self._get_full_dim_output(outputs) + # x = self.downsample_output(x) + # class Downsample has this rounding behavior.. + # assert self.output_downsampling_factor == 2, self.output_downsampling_factor + # if torch.jit.is_scripting() or torch.jit.is_tracing(): + # lengths = (x_lens + 1) // 2 + # else: + # with warnings.catch_warnings(): + # warnings.simplefilter("ignore") + # lengths = (x_lens + 1) // 2 + + # return x, lengths + return x, x_lens + + def _get_attn_mask( + self, x: Tensor, chunk_size: int, left_context_chunks: int + ) -> Optional[Tensor]: + """ + Return None if chunk_size == -1, else return attention mask of shape + (seq_len, seq_len), interpreted as (tgt_seq_len, src_seq_len). True + means a masked position. + Args: + x: embeddings after self.encoder_embed(), of shape (seq_len, batch_size, embed_dim). + chunk_size: chunk size, must divide + """ + if chunk_size <= 0: + return None + assert all(chunk_size % d == 0 for d in self.downsampling_factor) + if left_context_chunks >= 0: + num_encoders = len(self.encoder_dim) + assert all( + chunk_size * left_context_chunks + >= (self.cnn_module_kernel[i] // 2) * self.downsampling_factor[i] + for i in range(num_encoders) + ) + else: + left_context_chunks = 1000000 + + seq_len = x.shape[0] + + # t is frame index, shape (seq_len,) + t = torch.arange(seq_len, dtype=torch.int32, device=x.device) + # c is chunk index for each frame, shape (seq_len,) + if torch.jit.is_scripting() or torch.jit.is_tracing(): + c = t // chunk_size + else: + with warnings.catch_warnings(): + warnings.simplefilter("ignore") + c = t // chunk_size + src_c = c + tgt_c = c.unsqueeze(-1) + + attn_mask = torch.logical_or(src_c > tgt_c, src_c < tgt_c - left_context_chunks) + if __name__ == "__main__": + logging.info(f"attn_mask = {attn_mask}") + return attn_mask + + def _get_full_dim_output(self, outputs: List[Tensor]): + num_encoders = len(self.encoder_dim) + assert len(outputs) == num_encoders + output_dim = max(self.encoder_dim) + output_pieces = [outputs[-1]] + cur_dim = self.encoder_dim[-1] + for i in range(num_encoders - 2, -1, -1): + d = self.encoder_dim[i] + if d > cur_dim: + this_output = outputs[i] + output_pieces.append(this_output[..., cur_dim:d]) + cur_dim = d + assert cur_dim == output_dim + return torch.cat(output_pieces, dim=-1) + + def streaming_forward( + self, + x: Tensor, + x_lens: Tensor, + states: List[Tensor], + src_key_padding_mask: Tensor, + ) -> Tuple[Tensor, Tensor, List[Tensor]]: + """ + Args: + x: + The input tensor. Its shape is (seq_len, batch_size, feature_dim). + x_lens: + A tensor of shape (batch_size,) containing the number of frames in + `x` before padding. + states: list of cached tensors of all encoder layers. For layer-i, + states[i*6:(i+1)*6] is (cached_key, cached_nonlin_attn, cached_val1, cached_val2, + cached_conv1, cached_conv2). + src_key_padding_mask: + The mask for padding, of shape (batch_size, seq_len); True means + masked position. May be None. + Returns: + Return a tuple containing 2 tensors: + - embeddings: its shape is (output_seq_len, batch_size, max(encoder_dim)) + - lengths, a tensor of shape (batch_size,) containing the number + of frames in `embeddings` before padding. + - updated states + """ + outputs = [] + new_states = [] + layer_offset = 0 + + for i, module in enumerate(self.encoders): + num_layers = module.num_layers + ds = self.downsampling_factor[i] + x = convert_num_channels(x, self.encoder_dim[i]) + + x, new_layer_states = module.streaming_forward( + x, + states=states[layer_offset * 6 : (layer_offset + num_layers) * 6], + left_context_len=self.left_context_frames[0] // ds, + src_key_padding_mask=src_key_padding_mask[..., ::ds], + ) + layer_offset += num_layers + outputs.append(x) + new_states += new_layer_states + + # if the last output has the largest dimension, x will be unchanged, + # it will be the same as outputs[-1]. Otherwise it will be concatenated + # from different pieces of 'outputs', taking each dimension from the + # most recent output that has it present. + x = self._get_full_dim_output(outputs) + x = self.downsample_output(x) + # class Downsample has this rounding behavior.. + assert self.output_downsampling_factor == 2 + if torch.jit.is_scripting() or torch.jit.is_tracing(): + lengths = (x_lens + 1) // 2 + else: + with warnings.catch_warnings(): + warnings.simplefilter("ignore") + lengths = (x_lens + 1) // 2 + + return x, lengths, new_states + + @torch.jit.export + def get_init_states( + self, + batch_size: int = 1, + device: torch.device = torch.device("cpu"), + ) -> List[Tensor]: + """Get initial states. + + A list of cached tensors of all encoder layers. For layer-i, states[i*6:(i+1)*6] + is (cached_key, cached_nonlin_attn, cached_val1, cached_val2, cached_conv1, cached_conv2). + """ + states = [] + for i, module in enumerate(self.encoders): + num_layers = module.num_layers + embed_dim = self.encoder_dim[i] + ds = self.downsampling_factor[i] + num_heads = self.num_heads[i] + key_dim = self.query_head_dim[i] * num_heads + value_dim = self.value_head_dim[i] * num_heads + downsample_left = self.left_context_frames[0] // ds + nonlin_attn_head_dim = 3 * embed_dim // 4 + conv_left_pad = self.cnn_module_kernel[i] // 2 + for layer in range(num_layers): + cached_key = torch.zeros(downsample_left, batch_size, key_dim).to( + device + ) + cached_nonlin_attn = torch.zeros( + 1, batch_size, downsample_left, nonlin_attn_head_dim + ).to(device) + cached_val1 = torch.zeros(downsample_left, batch_size, value_dim).to( + device + ) + cached_val2 = torch.zeros(downsample_left, batch_size, value_dim).to( + device + ) + cached_conv1 = torch.zeros(batch_size, embed_dim, conv_left_pad).to( + device + ) + cached_conv2 = torch.zeros(batch_size, embed_dim, conv_left_pad).to( + device + ) + states += [ + cached_key, + cached_nonlin_attn, + cached_val1, + cached_val2, + cached_conv1, + cached_conv2, + ] + + return states + + +def _whitening_schedule(x: float, ratio: float = 2.0) -> ScheduledFloat: + return ScheduledFloat((0.0, x), (20000.0, ratio * x), default=x) + + +def _balancer_schedule(min_prob: float): + return ScheduledFloat((0.0, 0.4), (8000.0, min_prob)) + + +class Zipformer2EncoderLayer(nn.Module): + """ + Args: + embed_dim: the number of expected features in the input (required). + nhead: the number of heads in the multiheadattention models (required). + feedforward_dim: the dimension of the feedforward network model (default=2048). + dropout: the dropout value (default=0.1). + cnn_module_kernel (int): Kernel size of convolution module. + + Examples:: + >>> encoder_layer = Zipformer2EncoderLayer(embed_dim=512, nhead=8) + >>> src = torch.rand(10, 32, 512) + >>> pos_emb = torch.rand(32, 19, 512) + >>> out = encoder_layer(src, pos_emb) + """ + + def __init__( + self, + embed_dim: int, + pos_dim: int, + num_heads: int, + query_head_dim: int, + pos_head_dim: int, + value_head_dim: int, + feedforward_dim: int, + dropout: FloatLike = 0.1, + cnn_module_kernel: int = 31, + causal: bool = False, + attention_skip_rate: FloatLike = ScheduledFloat( + (0.0, 0.2), (4000.0, 0.05), (16000, 0.0), default=0 + ), + conv_skip_rate: FloatLike = ScheduledFloat( + (0.0, 0.2), (4000.0, 0.05), (16000, 0.0), default=0 + ), + const_attention_rate: FloatLike = ScheduledFloat( + (0.0, 0.25), (4000.0, 0.025), default=0 + ), + ff2_skip_rate: FloatLike = ScheduledFloat( + (0.0, 0.1), (4000.0, 0.01), (50000.0, 0.0) + ), + ff3_skip_rate: FloatLike = ScheduledFloat( + (0.0, 0.1), (4000.0, 0.01), (50000.0, 0.0) + ), + bypass_skip_rate: FloatLike = ScheduledFloat( + (0.0, 0.5), (4000.0, 0.02), default=0 + ), + ) -> None: + super(Zipformer2EncoderLayer, self).__init__() + self.embed_dim = embed_dim + + # self.bypass implements layer skipping as well as bypass; see its default values. + self.bypass = BypassModule( + embed_dim, skip_rate=bypass_skip_rate, straight_through_rate=0 + ) + # bypass_mid is bypass used in the middle of the layer. + self.bypass_mid = BypassModule(embed_dim, straight_through_rate=0) + + # skip probability for dynamic modules (meaning: anything but feedforward). + self.attention_skip_rate = copy.deepcopy(attention_skip_rate) + # an additional skip probability that applies to ConvModule to stop it from + # contributing too much early on. + self.conv_skip_rate = copy.deepcopy(conv_skip_rate) + + # ff2_skip_rate is to prevent the ff2 module from having output that's too big + # compared to its residual. + self.ff2_skip_rate = copy.deepcopy(ff2_skip_rate) + self.ff3_skip_rate = copy.deepcopy(ff3_skip_rate) + + self.const_attention_rate = copy.deepcopy(const_attention_rate) + + self.self_attn_weights = RelPositionMultiheadAttentionWeights( + embed_dim, + pos_dim=pos_dim, + num_heads=num_heads, + query_head_dim=query_head_dim, + pos_head_dim=pos_head_dim, + dropout=0.0, + ) + + self.self_attn1 = SelfAttention(embed_dim, num_heads, value_head_dim) + + self.self_attn2 = SelfAttention(embed_dim, num_heads, value_head_dim) + + self.feed_forward1 = FeedforwardModule( + embed_dim, (feedforward_dim * 3) // 4, dropout + ) + + self.feed_forward2 = FeedforwardModule(embed_dim, feedforward_dim, dropout) + + self.feed_forward3 = FeedforwardModule( + embed_dim, (feedforward_dim * 5) // 4, dropout + ) + + self.nonlin_attention = NonlinAttention( + embed_dim, hidden_channels=3 * embed_dim // 4 + ) + + self.conv_module1 = ConvolutionModule( + embed_dim, cnn_module_kernel, causal=causal + ) + + self.conv_module2 = ConvolutionModule( + embed_dim, cnn_module_kernel, causal=causal + ) + + # TODO: remove it + self.bypass_scale = nn.Parameter(torch.full((embed_dim,), 0.5)) + + self.norm = BiasNorm(embed_dim) + + self.balancer1 = Balancer( + embed_dim, + channel_dim=-1, + min_positive=0.45, + max_positive=0.55, + min_abs=0.2, + max_abs=4.0, + ) + + # balancer for output of NonlinAttentionModule + self.balancer_na = Balancer( + embed_dim, + channel_dim=-1, + min_positive=0.3, + max_positive=0.7, + min_abs=ScheduledFloat((0.0, 0.004), (4000.0, 0.02)), + prob=0.05, # out of concern for memory usage + ) + + # balancer for output of feedforward2, prevent it from staying too + # small. give this a very small probability, even at the start of + # training, it's to fix a rare problem and it's OK to fix it slowly. + self.balancer_ff2 = Balancer( + embed_dim, + channel_dim=-1, + min_positive=0.3, + max_positive=0.7, + min_abs=ScheduledFloat((0.0, 0.0), (4000.0, 0.1), default=0.0), + max_abs=2.0, + prob=0.05, + ) + + self.balancer_ff3 = Balancer( + embed_dim, + channel_dim=-1, + min_positive=0.3, + max_positive=0.7, + min_abs=ScheduledFloat((0.0, 0.0), (4000.0, 0.2), default=0.0), + max_abs=4.0, + prob=0.05, + ) + + self.whiten = Whiten( + num_groups=1, + whitening_limit=_whitening_schedule(4.0, ratio=3.0), + prob=(0.025, 0.25), + grad_scale=0.01, + ) + + self.balancer2 = Balancer( + embed_dim, + channel_dim=-1, + min_positive=0.45, + max_positive=0.55, + min_abs=0.1, + max_abs=4.0, + ) + + def get_sequence_dropout_mask( + self, x: Tensor, dropout_rate: float + ) -> Optional[Tensor]: + if ( + dropout_rate == 0.0 + or not self.training + or torch.jit.is_scripting() + or torch.jit.is_tracing() + ): + return None + batch_size = x.shape[1] + mask = (torch.rand(batch_size, 1, device=x.device) > dropout_rate).to(x.dtype) + return mask + + def sequence_dropout(self, x: Tensor, dropout_rate: float) -> Tensor: + """ + Apply sequence-level dropout to x. + x shape: (seq_len, batch_size, embed_dim) + """ + dropout_mask = self.get_sequence_dropout_mask(x, dropout_rate) + if dropout_mask is None: + return x + else: + return x * dropout_mask + + def forward( + self, + src: Tensor, + pos_emb: Tensor, + chunk_size: int = -1, + attn_mask: Optional[Tensor] = None, + src_key_padding_mask: Optional[Tensor] = None, + ) -> Tensor: + """ + Pass the input through the encoder layer. + Args: + src: the sequence to the encoder (required): shape (seq_len, batch_size, embedding_dim). + pos_emb: (1, 2*seq_len-1, pos_emb_dim) or (batch_size, 2*seq_len-1, pos_emb_dim) + chunk_size: the number of frames per chunk, of >= 0; if -1, no chunking. + feature_mask: something that broadcasts with src, that we'll multiply `src` + by at every layer: if a Tensor, likely of shape (seq_len, batch_size, embedding_dim) + attn_mask: the attention mask, of shape (batch_size, seq_len, seq_len) or (seq_len, seq_len), + interpreted as (batch_size, tgt_seq_len, src_seq_len) or (tgt_seq_len, src_seq_len). + True means masked position. May be None. + src_key_padding_mask: the mask for padding, of shape (batch_size, seq_len); True means + masked position. May be None. + + Returns: + A tensor which has the same shape as src + """ + src_orig = src + + # dropout rate for non-feedforward submodules + if torch.jit.is_scripting() or torch.jit.is_tracing(): + attention_skip_rate = 0.0 + else: + attention_skip_rate = ( + float(self.attention_skip_rate) if self.training else 0.0 + ) + + # attn_weights: (num_heads, batch_size, seq_len, seq_len) + attn_weights = self.self_attn_weights( + src, + pos_emb=pos_emb, + attn_mask=attn_mask, + key_padding_mask=src_key_padding_mask, + ) + + src = src + self.feed_forward1(src) + + self_attn_dropout_mask = self.get_sequence_dropout_mask( + src, attention_skip_rate + ) + + selected_attn_weights = attn_weights[0:1] + if torch.jit.is_scripting() or torch.jit.is_tracing(): + pass + elif not self.training and random.random() < float(self.const_attention_rate): + # Make attention weights constant. The intention is to + # encourage these modules to do something similar to an + # averaging-over-time operation. + # only need the mask, can just use the 1st one and expand later + selected_attn_weights = selected_attn_weights[0:1] + selected_attn_weights = (selected_attn_weights > 0.0).to( + selected_attn_weights.dtype + ) + selected_attn_weights = selected_attn_weights * ( + 1.0 / selected_attn_weights.sum(dim=-1, keepdim=True) + ) + + na = self.balancer_na(self.nonlin_attention(src, selected_attn_weights)) + + src = src + ( + na if self_attn_dropout_mask is None else na * self_attn_dropout_mask + ) + + self_attn = self.self_attn1(src, attn_weights) + + src = src + ( + self_attn + if self_attn_dropout_mask is None + else self_attn * self_attn_dropout_mask + ) + + if torch.jit.is_scripting() or torch.jit.is_tracing(): + conv_skip_rate = 0.0 + else: + conv_skip_rate = float(self.conv_skip_rate) if self.training else 0.0 + src = src + self.sequence_dropout( + self.conv_module1( + src, chunk_size=chunk_size, src_key_padding_mask=src_key_padding_mask + ), + conv_skip_rate, + ) + + if torch.jit.is_scripting() or torch.jit.is_tracing(): + ff2_skip_rate = 0.0 + else: + ff2_skip_rate = float(self.ff2_skip_rate) if self.training else 0.0 + src = src + self.sequence_dropout( + self.balancer_ff2(self.feed_forward2(src)), ff2_skip_rate + ) + + # bypass in the middle of the layer. + src = self.bypass_mid(src_orig, src) + + self_attn = self.self_attn2(src, attn_weights) + + src = src + ( + self_attn + if self_attn_dropout_mask is None + else self_attn * self_attn_dropout_mask + ) + + if torch.jit.is_scripting() or torch.jit.is_tracing(): + conv_skip_rate = 0.0 + else: + conv_skip_rate = float(self.conv_skip_rate) if self.training else 0.0 + src = src + self.sequence_dropout( + self.conv_module2( + src, chunk_size=chunk_size, src_key_padding_mask=src_key_padding_mask + ), + conv_skip_rate, + ) + + if torch.jit.is_scripting() or torch.jit.is_tracing(): + ff3_skip_rate = 0.0 + else: + ff3_skip_rate = float(self.ff3_skip_rate) if self.training else 0.0 + src = src + self.sequence_dropout( + self.balancer_ff3(self.feed_forward3(src)), ff3_skip_rate + ) + + src = self.balancer1(src) + src = self.norm(src) + + src = self.bypass(src_orig, src) + + src = self.balancer2(src) + src = self.whiten(src) + + return src + + def streaming_forward( + self, + src: Tensor, + pos_emb: Tensor, + cached_key: Tensor, + cached_nonlin_attn: Tensor, + cached_val1: Tensor, + cached_val2: Tensor, + cached_conv1: Tensor, + cached_conv2: Tensor, + left_context_len: int, + src_key_padding_mask: Tensor, + ) -> Tuple[Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor]: + """Pass the input through the encoder layer in streaming forward mode. + + Args: + src: the sequence to the encoder (required): shape (seq_len, batch_size, embedding_dim). + pos_emb: (1, left_context_len+2*seq_len-1, pos_emb_dim) or + (batch_size, left_context_len+2*seq_len-1, pos_emb_dim) + cached_key: cached attention key tensor of left context, + of shape (left_context_len, batch_size, key_dim) + cached_nonlin_attn: left context for nonlin_attention module, a Tensor of shape + (num_heads, batch_size, left_context_len, head_dim) + cached_val1: cached left context for the first attention module, + of shape (left_context_len, batch_size, value_dim) + cached_val2: cached left context for the second attention module, + of shape (left_context_len, batch_size, value_dim) + cached_conv1: cached left context for the first convolution module, + of shape (batch_size, channels, left_pad) + cached_conv2: cached left context for the second convolution module, + of shape (batch_size, channels, left_pad) + left_context_len: number of left context frames. + src_key_padding_mask: the mask for padding, of shape + (batch_size, left_context_len + seq_len); True means masked position. + May be None. + + Returns: + - x, with the same shape as src + - updated cached_key + - updated cached_nonlin_attn + - updated cached_val1 + - updated cached_val2 + - updated cached_conv1 + - updated cached_conv2 + """ + src_orig = src + + # attn_weights: (num_heads, batch_size, seq_len, seq_len) + attn_weights, cached_key = self.self_attn_weights.streaming_forward( + src, + pos_emb=pos_emb, + cached_key=cached_key, + left_context_len=left_context_len, + key_padding_mask=src_key_padding_mask, + ) + + src = src + self.feed_forward1(src) + + na, cached_nonlin_attn = self.nonlin_attention.streaming_forward( + src, + attn_weights[0:1], + cached_x=cached_nonlin_attn, + left_context_len=left_context_len, + ) + src = src + na + + self_attn, cached_val1 = self.self_attn1.streaming_forward( + src, + attn_weights=attn_weights, + cached_val=cached_val1, + left_context_len=left_context_len, + ) + src = src + self_attn + + src_conv, cached_conv1 = self.conv_module1.streaming_forward( + src, + cache=cached_conv1, + src_key_padding_mask=src_key_padding_mask[:, left_context_len:], + ) + src = src + src_conv + + src = src + self.feed_forward2(src) + + # bypass in the middle of the layer. + src = self.bypass_mid(src_orig, src) + + self_attn, cached_val2 = self.self_attn2.streaming_forward( + src, + attn_weights=attn_weights, + cached_val=cached_val2, + left_context_len=left_context_len, + ) + src = src + self_attn + + src_conv, cached_conv2 = self.conv_module2.streaming_forward( + src, + cache=cached_conv2, + src_key_padding_mask=src_key_padding_mask[:, left_context_len:], + ) + src = src + src_conv + + src = src + self.feed_forward3(src) + + src = self.norm(src) + + src = self.bypass(src_orig, src) + + return ( + src, + cached_key, + cached_nonlin_attn, + cached_val1, + cached_val2, + cached_conv1, + cached_conv2, + ) + + +class Zipformer2Encoder(nn.Module): + r"""Zipformer2Encoder is a stack of N encoder layers + + Args: + encoder_layer: an instance of the Zipformer2EncoderLayer() class (required). + num_layers: the number of sub-encoder-layers in the encoder (required). + pos_dim: the dimension for the relative positional encoding + + Examples:: + >>> encoder_layer = Zipformer2EncoderLayer(embed_dim=512, nhead=8) + >>> zipformer_encoder = Zipformer2Encoder(encoder_layer, num_layers=6) + >>> src = torch.rand(10, 32, 512) + >>> out = zipformer_encoder(src) + """ + + def __init__( + self, + encoder_layer: nn.Module, + num_layers: int, + pos_dim: int, + dropout: float, + warmup_begin: float, + warmup_end: float, + initial_layerdrop_rate: float = 0.5, + final_layerdrop_rate: float = 0.05, + ) -> None: + super().__init__() + self.encoder_pos = CompactRelPositionalEncoding( + pos_dim, dropout_rate=0.15, length_factor=1.0 + ) + + self.layers = nn.ModuleList( + [copy.deepcopy(encoder_layer) for i in range(num_layers)] + ) + self.num_layers = num_layers + + assert 0 <= warmup_begin <= warmup_end + + delta = (1.0 / num_layers) * (warmup_end - warmup_begin) + cur_begin = warmup_begin # interpreted as a training batch index + for i in range(num_layers): + cur_end = cur_begin + delta + self.layers[i].bypass.skip_rate = ScheduledFloat( + (cur_begin, initial_layerdrop_rate), + (cur_end, final_layerdrop_rate), + default=0.0, + ) + cur_begin = cur_end + + def forward( + self, + src: Tensor, + chunk_size: int = -1, + feature_mask: Union[Tensor, float] = 1.0, + attn_mask: Optional[Tensor] = None, + src_key_padding_mask: Optional[Tensor] = None, + ) -> Tensor: + r"""Pass the input through the encoder layers in turn. + + Args: + src: the sequence to the encoder (required): shape (seq_len, batch_size, embedding_dim). + chunk_size: the number of frames per chunk, of >= 0; if -1, no chunking. + feature_mask: something that broadcasts with src, that we'll multiply `src` + by at every layer: if a Tensor, likely of shape (seq_len, batch_size, embedding_dim) + attn_mask: the attention mask, of shape (batch_size, seq_len, seq_len) or (seq_len, seq_len), + interpreted as (batch_size, tgt_seq_len, src_seq_len) or (tgt_seq_len, src_seq_len). + True means masked position. May be None. + src_key_padding_mask: the mask for padding, of shape (batch_size, seq_len); True means + masked position. May be None. + + Returns: a Tensor with the same shape as src. + """ + pos_emb = self.encoder_pos(src) + output = src + + if not torch.jit.is_scripting() and not torch.jit.is_tracing(): + output = output * feature_mask + + for i, mod in enumerate(self.layers): + output = mod( + output, + pos_emb, + chunk_size=chunk_size, + attn_mask=attn_mask, + src_key_padding_mask=src_key_padding_mask, + ) + + if not torch.jit.is_scripting() and not torch.jit.is_tracing(): + output = output * feature_mask + + return output + + def streaming_forward( + self, + src: Tensor, + states: List[Tensor], + left_context_len: int, + src_key_padding_mask: Tensor, + ) -> Tuple[Tensor, List[Tensor]]: + r"""Pass the input through the encoder layers in turn. + + Args: + src: the sequence to the encoder (required): shape (seq_len, batch_size, embedding_dim). + states: list of cached tensors of N encoder layers. For layer-i, states[i*6:(i+1)*6] is + (cached_key, cached_nonlin_attn, cached_val1, cached_val2, cached_conv1, cached_conv2). + left_context_len: Number of left context frames. + src_key_padding_mask: the mask for padding, of shape + (batch_size, left_context_len + seq_len); True means masked position. + May be None. + + Returns: + - output, a Tensor with the same shape as src. + - updated states + """ + pos_emb = self.encoder_pos(src, left_context_len) + output = src + + new_states = [] + for i, mod in enumerate(self.layers): + ( + cached_key, + cached_nonlin_attn, + cached_val1, + cached_val2, + cached_conv1, + cached_conv2, + ) = states[i * 6 : (i + 1) * 6] + ( + output, + new_cached_key, + new_cached_nonlin_attn, + new_cached_val1, + new_cached_val2, + new_cached_conv1, + new_cached_conv2, + ) = mod.streaming_forward( + output, + pos_emb, + cached_key=cached_key, + cached_nonlin_attn=cached_nonlin_attn, + cached_val1=cached_val1, + cached_val2=cached_val2, + cached_conv1=cached_conv1, + cached_conv2=cached_conv2, + left_context_len=left_context_len, + src_key_padding_mask=src_key_padding_mask, + ) + new_states += [ + new_cached_key, + new_cached_nonlin_attn, + new_cached_val1, + new_cached_val2, + new_cached_conv1, + new_cached_conv2, + ] + + return output, new_states + + +class BypassModule(nn.Module): + """ + An nn.Module that implements a learnable bypass scale, and also randomized per-sequence + layer-skipping. The bypass is limited during early stages of training to be close to + "straight-through", i.e. to not do the bypass operation much initially, in order to + force all the modules to learn something. + """ + + def __init__( + self, + embed_dim: int, + skip_rate: FloatLike = 0.0, + straight_through_rate: FloatLike = 0.0, + scale_min: FloatLike = ScheduledFloat((0.0, 0.9), (20000.0, 0.2), default=0), + scale_max: FloatLike = 1.0, + ): + super().__init__() + self.bypass_scale = nn.Parameter(torch.full((embed_dim,), 0.5)) + self.skip_rate = copy.deepcopy(skip_rate) + self.straight_through_rate = copy.deepcopy(straight_through_rate) + self.scale_min = copy.deepcopy(scale_min) + self.scale_max = copy.deepcopy(scale_max) + + def _get_bypass_scale(self, batch_size: int): + # returns bypass-scale of shape (num_channels,), + # or (batch_size, num_channels,). This is actually the + # scale on the non-residual term, so 0 correponds to bypassing + # this module. + if torch.jit.is_scripting() or torch.jit.is_tracing() or not self.training: + return self.bypass_scale + else: + ans = limit_param_value( + self.bypass_scale, min=float(self.scale_min), max=float(self.scale_max) + ) + skip_rate = float(self.skip_rate) + if skip_rate != 0.0: + mask = torch.rand((batch_size, 1), device=ans.device) > skip_rate + ans = ans * mask + # now ans is of shape (batch_size, num_channels), and is zero for sequences + # on which we have randomly chosen to do layer-skipping. + straight_through_rate = float(self.straight_through_rate) + if straight_through_rate != 0.0: + mask = ( + torch.rand((batch_size, 1), device=ans.device) + < straight_through_rate + ) + ans = torch.maximum(ans, mask.to(ans.dtype)) + return ans + + def forward(self, src_orig: Tensor, src: Tensor): + """ + Args: src_orig and src are both of shape (seq_len, batch_size, num_channels) + Returns: something with the same shape as src and src_orig + """ + bypass_scale = self._get_bypass_scale(src.shape[1]) + return src_orig + (src - src_orig) * bypass_scale + + +class DownsampledZipformer2Encoder(nn.Module): + r""" + DownsampledZipformer2Encoder is a zipformer encoder evaluated at a reduced frame rate, + after convolutional downsampling, and then upsampled again at the output, and combined + with the origin input, so that the output has the same shape as the input. + """ + + def __init__( + self, encoder: nn.Module, dim: int, downsample: int, dropout: FloatLike + ): + super(DownsampledZipformer2Encoder, self).__init__() + self.downsample_factor = downsample + self.downsample = SimpleDownsample(dim, downsample, dropout) + self.num_layers = encoder.num_layers + self.encoder = encoder + self.upsample = SimpleUpsample(dim, downsample) + self.out_combiner = BypassModule(dim, straight_through_rate=0) + + def forward( + self, + src: Tensor, + chunk_size: int = -1, + feature_mask: Union[Tensor, float] = 1.0, + attn_mask: Optional[Tensor] = None, + src_key_padding_mask: Optional[Tensor] = None, + ) -> Tensor: + r"""Downsample, go through encoder, upsample. + + Args: + src: the sequence to the encoder (required): shape (seq_len, batch_size, embedding_dim). + feature_mask: something that broadcasts with src, that we'll multiply `src` + by at every layer: if a Tensor, likely of shape (seq_len, batch_size, embedding_dim) + attn_mask: the attention mask, of shape (batch_size, seq_len, seq_len) or (seq_len, seq_len), + interpreted as (batch_size, tgt_seq_len, src_seq_len) or (tgt_seq_len, src_seq_len). + True means masked position. May be None. + src_key_padding_mask: the mask for padding, of shape (batch_size, seq_len); True means + masked position. May be None. + + Returns: a Tensor with the same shape as src. + """ + src_orig = src + src = self.downsample(src) + ds = self.downsample_factor + if attn_mask is not None: + attn_mask = attn_mask[::ds, ::ds] + + src = self.encoder( + src, + chunk_size=chunk_size // ds, + feature_mask=feature_mask, + attn_mask=attn_mask, + src_key_padding_mask=src_key_padding_mask, + ) + src = self.upsample(src) + # remove any extra frames that are not a multiple of downsample_factor + src = src[: src_orig.shape[0]] + + return self.out_combiner(src_orig, src) + + def streaming_forward( + self, + src: Tensor, + states: List[Tensor], + left_context_len: int, + src_key_padding_mask: Tensor, + ) -> Tuple[Tensor, List[Tensor]]: + r"""Downsample, go through encoder, upsample, in streaming forward mode. + + Args: + src: the sequence to the encoder (required): shape (seq_len, batch_size, embedding_dim). + states: list of cached tensors of N encoder layers. For layer-i, states[i*6:(i+1)*6] is + (cached_key, cached_nonlin_attn, cached_val1, cached_val2, cached_conv1, cached_conv2). + left_context_len: Number of left context frames. + src_key_padding_mask: the mask for padding, of shape (batch_size, left_context_len+seq_len); + True means masked position. May be None. + + Returns: + - output, a Tensor with the same shape as src. + - updated states + """ + src_orig = src + src = self.downsample(src) + + src, new_states = self.encoder.streaming_forward( + src, + states=states, + left_context_len=left_context_len, + src_key_padding_mask=src_key_padding_mask, + ) + src = self.upsample(src) + # remove any extra frames that are not a multiple of downsample_factor + src = src[: src_orig.shape[0]] + + return self.out_combiner(src_orig, src), new_states + + +class SimpleDownsample(torch.nn.Module): + """ + Does downsampling with attention, by weighted sum, and a projection.. + """ + + def __init__(self, channels: int, downsample: int, dropout: FloatLike): + super(SimpleDownsample, self).__init__() + + self.bias = nn.Parameter(torch.zeros(downsample)) + + self.name = None # will be set from training code + self.dropout = copy.deepcopy(dropout) + + self.downsample = downsample + + def forward(self, src: Tensor) -> Tensor: + """ + x: (seq_len, batch_size, in_channels) + Returns a tensor of shape + ( (seq_len+downsample-1)//downsample, batch_size, channels) + """ + (seq_len, batch_size, in_channels) = src.shape + ds = self.downsample + d_seq_len = (seq_len + ds - 1) // ds + + # Pad to an exact multiple of self.downsample + # right-pad src, repeating the last element. + pad = d_seq_len * ds - seq_len + src_extra = src[src.shape[0] - 1 :].expand(pad, src.shape[1], src.shape[2]) + src = torch.cat((src, src_extra), dim=0) + assert src.shape[0] == d_seq_len * ds + + src = src.reshape(d_seq_len, ds, batch_size, in_channels) + + weights = self.bias.softmax(dim=0) + # weights: (downsample, 1, 1) + weights = weights.unsqueeze(-1).unsqueeze(-1) + + # ans1 is the first `in_channels` channels of the output + ans = (src * weights).sum(dim=1) + + return ans + + +class SimpleUpsample(torch.nn.Module): + """ + A very simple form of upsampling that mostly just repeats the input, but + also adds a position-specific bias. + """ + + def __init__(self, num_channels: int, upsample: int): + super(SimpleUpsample, self).__init__() + self.upsample = upsample + + def forward(self, src: Tensor) -> Tensor: + """ + x: (seq_len, batch_size, num_channels) + Returns a tensor of shape + ( (seq_len*upsample), batch_size, num_channels) + """ + upsample = self.upsample + (seq_len, batch_size, num_channels) = src.shape + src = src.unsqueeze(1).expand(seq_len, upsample, batch_size, num_channels) + src = src.reshape(seq_len * upsample, batch_size, num_channels) + return src + + +class CompactRelPositionalEncoding(torch.nn.Module): + """ + Relative positional encoding module. This version is "compact" meaning it is able to encode + the important information about the relative position in a relatively small number of dimensions. + The goal is to make it so that small differences between large relative offsets (e.g. 1000 vs. 1001) + make very little difference to the embedding. Such differences were potentially important + when encoding absolute position, but not important when encoding relative position because there + is now no need to compare two large offsets with each other. + + Our embedding works done by projecting the interval [-infinity,infinity] to a finite interval + using the atan() function, before doing the fourier transform of that fixed interval. The + atan() function would compress the "long tails" too small, + making it hard to distinguish between different magnitudes of large offsets, so we use a logarithmic + function to compress large offsets to a smaller range before applying atan(). + Scalings are chosen in such a way that the embedding can clearly distinguish invidual offsets as long + as they are quite close to the origin, e.g. abs(offset) <= about sqrt(embedding_dim) + + + Args: + embed_dim: Embedding dimension. + dropout_rate: Dropout rate. + max_len: Maximum input length: just a heuristic for initialization. + length_factor: a heuristic scale (should be >= 1.0) which, if larger, gives + less weight to small differences of offset near the origin. + """ + + def __init__( + self, + embed_dim: int, + dropout_rate: FloatLike, + max_len: int = 1000, + length_factor: float = 1.0, + ) -> None: + """Construct a CompactRelPositionalEncoding object.""" + super(CompactRelPositionalEncoding, self).__init__() + self.embed_dim = embed_dim + assert embed_dim % 2 == 0 + self.dropout = Dropout2(dropout_rate) + self.pe = None + assert length_factor >= 1.0 + self.length_factor = length_factor + self.extend_pe(torch.tensor(0.0).expand(max_len)) + + def extend_pe(self, x: Tensor, left_context_len: int = 0) -> None: + """Reset the positional encodings.""" + T = x.size(0) + left_context_len + + if self.pe is not None: + # self.pe contains both positive and negative parts + # the length of self.pe is 2 * input_len - 1 + if self.pe.size(0) >= T * 2 - 1: + self.pe = self.pe.to(dtype=x.dtype, device=x.device) + return + + # if T == 4, x would contain [ -3, -2, 1, 0, 1, 2, 3 ] + x = torch.arange(-(T - 1), T, device=x.device).to(torch.float32).unsqueeze(1) + + freqs = 1 + torch.arange(self.embed_dim // 2, device=x.device) + + # `compression_length` this is arbitrary/heuristic, if it is larger we have more resolution + # for small time offsets but less resolution for large time offsets. + compression_length = self.embed_dim**0.5 + # x_compressed, like X, goes from -infinity to infinity as T goes from -infinity to infinity; + # but it does so more slowly than T for large absolute values of T. + # The formula is chosen so that d(x_compressed )/dx is 1 around x == 0, which + # is important. + x_compressed = ( + compression_length + * x.sign() + * ((x.abs() + compression_length).log() - math.log(compression_length)) + ) + + # if self.length_factor == 1.0, then length_scale is chosen so that the + # FFT can exactly separate points close to the origin (T == 0). So this + # part of the formulation is not really heuristic. + # But empirically, for ASR at least, length_factor > 1.0 seems to work better. + length_scale = self.length_factor * self.embed_dim / (2.0 * math.pi) + + # note for machine implementations: if atan is not available, we can use: + # x.sign() * ((1 / (x.abs() + 1)) - 1) * (-math.pi/2) + # check on wolframalpha.com: plot(sign(x) * (1 / ( abs(x) + 1) - 1 ) * -pi/2 , atan(x)) + x_atan = (x_compressed / length_scale).atan() # results between -pi and pi + + cosines = (x_atan * freqs).cos() + sines = (x_atan * freqs).sin() + + pe = torch.zeros(x.shape[0], self.embed_dim, device=x.device) + pe[:, 0::2] = cosines + pe[:, 1::2] = sines + pe[:, -1] = 1.0 # for bias. + + self.pe = pe.to(dtype=x.dtype) + + def forward(self, x: Tensor, left_context_len: int = 0) -> Tensor: + """Create positional encoding. + + Args: + x (Tensor): Input tensor (time, batch, `*`). + left_context_len: (int): Length of cached left context. + + Returns: + positional embedding, of shape (batch, left_context_len + 2*time-1, `*`). + """ + self.extend_pe(x, left_context_len) + x_size_left = x.size(0) + left_context_len + # length of positive side: x.size(0) + left_context_len + # length of negative side: x.size(0) + pos_emb = self.pe[ + self.pe.size(0) // 2 + - x_size_left + + 1 : self.pe.size(0) // 2 # noqa E203 + + x.size(0), + :, + ] + pos_emb = pos_emb.unsqueeze(0) + return self.dropout(pos_emb) + + +class RelPositionMultiheadAttentionWeights(nn.Module): + r"""Module that computes multi-head attention weights with relative position encoding. + Various other modules consume the resulting attention weights: see, for example, the + SimpleAttention module which allows you to compute conventional attention. + + This is a quite heavily modified from: "Transformer-XL: Attentive Language Models Beyond a Fixed-Length Context", + we have to write up the differences. + + + Args: + embed_dim: number of channels at the input to this module, e.g. 256 + pos_dim: dimension of the positional encoding vectors, e.g. 128. + num_heads: number of heads to compute weights for, e.g. 8 + query_head_dim: dimension of the query (and key), per head. e.g. 24. + pos_head_dim: dimension of the projected positional encoding per head, e.g. 4. + dropout: dropout probability for attn_output_weights. Default: 0.0. + pos_emb_skip_rate: probability for skipping the pos_emb part of the scores on + any given call to forward(), in training time. + """ + + def __init__( + self, + embed_dim: int, + pos_dim: int, + num_heads: int, + query_head_dim: int, + pos_head_dim: int, + dropout: float = 0.0, + pos_emb_skip_rate: FloatLike = ScheduledFloat((0.0, 0.5), (4000.0, 0.0)), + ) -> None: + super().__init__() + self.embed_dim = embed_dim + self.num_heads = num_heads + self.query_head_dim = query_head_dim + self.pos_head_dim = pos_head_dim + self.dropout = dropout + self.pos_emb_skip_rate = copy.deepcopy(pos_emb_skip_rate) + self.name = None # will be overwritten in training code; for diagnostics. + + key_head_dim = query_head_dim + in_proj_dim = (query_head_dim + key_head_dim + pos_head_dim) * num_heads + + # the initial_scale is supposed to take over the "scaling" factor of + # head_dim ** -0.5 that has been used in previous forms of attention, + # dividing it between the query and key. Note: this module is intended + # to be used with the ScaledAdam optimizer; with most other optimizers, + # it would be necessary to apply the scaling factor in the forward function. + self.in_proj = ScaledLinear( + embed_dim, in_proj_dim, bias=True, initial_scale=query_head_dim**-0.25 + ) + + self.whiten_keys = Whiten( + num_groups=num_heads, + whitening_limit=_whitening_schedule(3.0), + prob=(0.025, 0.25), + grad_scale=0.025, + ) + + # add a balancer for the keys that runs with very small probability, and + # tries to enforce that all dimensions have mean around zero. The + # weights produced by this module are invariant to adding a constant to + # the keys, so the derivative of the bias is mathematically zero; but + # due to how Adam/ScaledAdam work, it can learn a fairly large nonzero + # bias because the small numerical roundoff tends to have a non-random + # sign. This module is intended to prevent that. Use a very small + # probability; that should be suffixient to fix the problem. + self.balance_keys = Balancer( + key_head_dim * num_heads, + channel_dim=-1, + min_positive=0.4, + max_positive=0.6, + min_abs=0.0, + max_abs=100.0, + prob=0.025, + ) + + # linear transformation for positional encoding. + self.linear_pos = ScaledLinear( + pos_dim, num_heads * pos_head_dim, bias=False, initial_scale=0.05 + ) + + # the following are for diagnosics only, see --print-diagnostics option + self.copy_pos_query = Identity() + self.copy_query = Identity() + + def forward( + self, + x: Tensor, + pos_emb: Tensor, + key_padding_mask: Optional[Tensor] = None, + attn_mask: Optional[Tensor] = None, + ) -> Tensor: + r""" + Args: + x: input of shape (seq_len, batch_size, embed_dim) + pos_emb: Positional embedding tensor, of shape (1, 2*seq_len - 1, pos_dim) + key_padding_mask: a bool tensor of shape (batch_size, seq_len). Positions that + are True in this mask will be ignored as sources in the attention weighting. + attn_mask: mask of shape (seq_len, seq_len) or (batch_size, seq_len, seq_len), + interpreted as ([batch_size,] tgt_seq_len, src_seq_len) + saying which positions are allowed to attend to which other positions. + Returns: + a tensor of attention weights, of shape (hum_heads, batch_size, seq_len, seq_len) + interpreted as (hum_heads, batch_size, tgt_seq_len, src_seq_len). + """ + x = self.in_proj(x) + query_head_dim = self.query_head_dim + pos_head_dim = self.pos_head_dim + num_heads = self.num_heads + + seq_len, batch_size, _ = x.shape + + query_dim = query_head_dim * num_heads + + # self-attention + q = x[..., 0:query_dim] + k = x[..., query_dim : 2 * query_dim] + # p is the position-encoding query + p = x[..., 2 * query_dim :] + assert p.shape[-1] == num_heads * pos_head_dim + + q = self.copy_query(q) # for diagnostics only, does nothing. + k = self.whiten_keys(self.balance_keys(k)) # does nothing in the forward pass. + p = self.copy_pos_query(p) # for diagnostics only, does nothing. + + q = q.reshape(seq_len, batch_size, num_heads, query_head_dim) + p = p.reshape(seq_len, batch_size, num_heads, pos_head_dim) + k = k.reshape(seq_len, batch_size, num_heads, query_head_dim) + + # time1 refers to target, time2 refers to source. + q = q.permute(2, 1, 0, 3) # (head, batch, time1, query_head_dim) + p = p.permute(2, 1, 0, 3) # (head, batch, time1, pos_head_dim) + k = k.permute(2, 1, 3, 0) # (head, batch, d_k, time2) + + attn_scores = torch.matmul(q, k) + + use_pos_scores = False + if torch.jit.is_scripting() or torch.jit.is_tracing(): + # We can't put random.random() in the same line + use_pos_scores = True + elif not self.training or random.random() >= float(self.pos_emb_skip_rate): + use_pos_scores = True + + if use_pos_scores: + pos_emb = self.linear_pos(pos_emb) + seq_len2 = 2 * seq_len - 1 + pos_emb = pos_emb.reshape(-1, seq_len2, num_heads, pos_head_dim).permute( + 2, 0, 3, 1 + ) + # pos shape now: (head, {1 or batch_size}, pos_dim, seq_len2) + + # (head, batch, time1, pos_dim) x (head, 1, pos_dim, seq_len2) -> (head, batch, time1, seq_len2) + # [where seq_len2 represents relative position.] + pos_scores = torch.matmul(p, pos_emb) + # the following .as_strided() expression converts the last axis of pos_scores from relative + # to absolute position. I don't know whether I might have got the time-offsets backwards or + # not, but let this code define which way round it is supposed to be. + if torch.jit.is_tracing(): + (num_heads, batch_size, time1, n) = pos_scores.shape + rows = torch.arange(start=time1 - 1, end=-1, step=-1) + cols = torch.arange(seq_len) + rows = rows.repeat(batch_size * num_heads).unsqueeze(-1) + indexes = rows + cols + pos_scores = pos_scores.reshape(-1, n) + pos_scores = torch.gather(pos_scores, dim=1, index=indexes) + pos_scores = pos_scores.reshape(num_heads, batch_size, time1, seq_len) + else: + pos_scores = pos_scores.as_strided( + (num_heads, batch_size, seq_len, seq_len), + ( + pos_scores.stride(0), + pos_scores.stride(1), + pos_scores.stride(2) - pos_scores.stride(3), + pos_scores.stride(3), + ), + storage_offset=pos_scores.stride(3) * (seq_len - 1), + ) + + attn_scores = attn_scores + pos_scores + + if torch.jit.is_scripting() or torch.jit.is_tracing(): + pass + elif self.training and random.random() < 0.1: + # This is a harder way of limiting the attention scores to not be + # too large. It incurs a penalty if any of them has an absolute + # value greater than 50.0. this should be outside the normal range + # of the attention scores. We use this mechanism instead of, say, + # something added to the loss function involving the entropy, + # because once the entropy gets very small gradients through the + # softmax can become very small, and we'd get zero derivatives. The + # choices of 1.0e-04 as the scale on the penalty makes this + # mechanism vulnerable to the absolute scale of the loss function, + # but we view this as a failsafe to avoid "implausible" parameter + # values rather than a regularization method that should be active + # under normal circumstances. + attn_scores = penalize_abs_values_gt( + attn_scores, limit=25.0, penalty=1.0e-04, name=self.name + ) + + assert attn_scores.shape == (num_heads, batch_size, seq_len, seq_len) + + if attn_mask is not None: + assert attn_mask.dtype == torch.bool + # use -1000 to avoid nan's where attn_mask and key_padding_mask make + # all scores zero. It's important that this be large enough that exp(-1000) + # is exactly zero, for reasons related to const_attention_rate, it + # compares the final weights with zero. + attn_scores = attn_scores.masked_fill(attn_mask, -1000) + + if key_padding_mask is not None: + assert key_padding_mask.shape == ( + batch_size, + seq_len, + ), key_padding_mask.shape + attn_scores = attn_scores.masked_fill( + key_padding_mask.unsqueeze(1), + -1000, + ) + + # We use our own version of softmax, defined in scaling.py, which should + # save a little of the memory used in backprop by, if we are in + # automatic mixed precision mode (amp / autocast), by only storing the + # half-precision output for backprop purposes. + attn_weights = softmax(attn_scores, dim=-1) + + if torch.jit.is_scripting() or torch.jit.is_tracing(): + pass + elif random.random() < 0.001 and not self.training: + self._print_attn_entropy(attn_weights) + + attn_weights = nn.functional.dropout( + attn_weights, p=self.dropout, training=self.training + ) + + return attn_weights + + def streaming_forward( + self, + x: Tensor, + pos_emb: Tensor, + cached_key: Tensor, + left_context_len: int, + key_padding_mask: Tensor, + ) -> Tuple[Tensor, Tensor]: + r""" + Args: + x: input of shape (seq_len, batch_size, embed_dim) + pos_emb: Positional embedding tensor, of shape (1, left_context_len+2*seq_len-1, pos_dim) + cached_key: cached attention key tensor of left context, + of shape (left_context_len, batch_size, key_dim) + left_context_len: number of left context frames. + key_padding_mask: a bool tensor of shape (batch_size, seq_len). Positions that + are True in this mask will be ignored as sources in the attention weighting. + + Returns: + - attention weights, of shape (hum_heads, batch_size, seq_len, seq_len2), + interpreted as (hum_heads, batch_size, tgt_seq_len, src_seq_len). + - updated cached attention key tensor of left context. + """ + x = self.in_proj(x) + query_head_dim = self.query_head_dim + pos_head_dim = self.pos_head_dim + num_heads = self.num_heads + + seq_len, batch_size, _ = x.shape + + query_dim = query_head_dim * num_heads + + # self-attention + q = x[..., 0:query_dim] + k = x[..., query_dim : 2 * query_dim] + # p is the position-encoding query + p = x[..., 2 * query_dim :] + assert p.shape[-1] == num_heads * pos_head_dim + + # Pad cached left contexts + assert cached_key.shape[0] == left_context_len, ( + cached_key.shape[0], + left_context_len, + ) + k = torch.cat([cached_key, k], dim=0) + # Update cached left contexts + cached_key = k[-left_context_len:, ...] + + # The length of key + k_len = k.shape[0] + + q = q.reshape(seq_len, batch_size, num_heads, query_head_dim) + p = p.reshape(seq_len, batch_size, num_heads, pos_head_dim) + k = k.reshape(k_len, batch_size, num_heads, query_head_dim) + + # time1 refers to target, time2 refers to source. + q = q.permute(2, 1, 0, 3) # (head, batch, time1, query_head_dim) + p = p.permute(2, 1, 0, 3) # (head, batch, time1, pos_head_dim) + k = k.permute(2, 1, 3, 0) # (head, batch, d_k, time2) + + attn_scores = torch.matmul(q, k) + + pos_emb = self.linear_pos(pos_emb) + seq_len2 = 2 * seq_len - 1 + left_context_len + pos_emb = pos_emb.reshape(-1, seq_len2, num_heads, pos_head_dim).permute( + 2, 0, 3, 1 + ) + # pos shape now: (head, {1 or batch_size}, pos_dim, seq_len2) + + # (head, batch, time1, pos_dim) x (head, 1, pos_dim, seq_len2) -> (head, batch, time1, seq_len2) + # [where seq_len2 represents relative position.] + pos_scores = torch.matmul(p, pos_emb) + + if torch.jit.is_tracing(): + (num_heads, batch_size, time1, n) = pos_scores.shape + rows = torch.arange(start=time1 - 1, end=-1, step=-1) + cols = torch.arange(k_len) + rows = rows.repeat(batch_size * num_heads).unsqueeze(-1) + indexes = rows + cols + pos_scores = pos_scores.reshape(-1, n) + pos_scores = torch.gather(pos_scores, dim=1, index=indexes) + pos_scores = pos_scores.reshape(num_heads, batch_size, time1, k_len) + # the following .as_strided() expression converts the last axis of pos_scores from relative + # to absolute position. I don't know whether I might have got the time-offsets backwards or + # not, but let this code define which way round it is supposed to be. + else: + pos_scores = pos_scores.as_strided( + (num_heads, batch_size, seq_len, k_len), + ( + pos_scores.stride(0), + pos_scores.stride(1), + pos_scores.stride(2) - pos_scores.stride(3), + pos_scores.stride(3), + ), + storage_offset=pos_scores.stride(3) * (seq_len - 1), + ) + + attn_scores = attn_scores + pos_scores + + assert attn_scores.shape == ( + num_heads, + batch_size, + seq_len, + k_len, + ), attn_scores.shape + + if key_padding_mask is not None: + assert key_padding_mask.shape == (batch_size, k_len), key_padding_mask.shape + attn_scores = attn_scores.masked_fill( + key_padding_mask.unsqueeze(1), + -1000, + ) + + attn_weights = attn_scores.softmax(dim=-1) + + return attn_weights, cached_key + + def _print_attn_entropy(self, attn_weights: Tensor): + # attn_weights: (num_heads, batch_size, seq_len, seq_len) + (num_heads, batch_size, seq_len, seq_len) = attn_weights.shape + + with torch.no_grad(): + with torch.cuda.amp.autocast(enabled=False): + attn_weights = attn_weights.to(torch.float32) + attn_weights_entropy = ( + -((attn_weights + 1.0e-20).log() * attn_weights) + .sum(dim=-1) + .mean(dim=(1, 2)) + ) + logging.info( + f"name={self.name}, attn_weights_entropy = {attn_weights_entropy}" + ) + + +class SelfAttention(nn.Module): + """ + The simplest possible attention module. This one works with already-computed attention + weights, e.g. as computed by RelPositionMultiheadAttentionWeights. + + Args: + embed_dim: the input and output embedding dimension + num_heads: the number of attention heads + value_head_dim: the value dimension per head + """ + + def __init__( + self, + embed_dim: int, + num_heads: int, + value_head_dim: int, + ) -> None: + super().__init__() + self.in_proj = nn.Linear(embed_dim, num_heads * value_head_dim, bias=True) + + self.out_proj = ScaledLinear( + num_heads * value_head_dim, embed_dim, bias=True, initial_scale=0.05 + ) + + self.whiten = Whiten( + num_groups=1, + whitening_limit=_whitening_schedule(7.5, ratio=3.0), + prob=(0.025, 0.25), + grad_scale=0.01, + ) + + def forward( + self, + x: Tensor, + attn_weights: Tensor, + ) -> Tensor: + """ + Args: + x: input tensor, of shape (seq_len, batch_size, embed_dim) + attn_weights: a tensor of shape (num_heads, batch_size, seq_len, seq_len), + with seq_len being interpreted as (tgt_seq_len, src_seq_len). Expect + attn_weights.sum(dim=-1) == 1. + Returns: + a tensor with the same shape as x. + """ + (seq_len, batch_size, embed_dim) = x.shape + num_heads = attn_weights.shape[0] + assert attn_weights.shape == (num_heads, batch_size, seq_len, seq_len) + + x = self.in_proj(x) # (seq_len, batch_size, num_heads * value_head_dim) + x = x.reshape(seq_len, batch_size, num_heads, -1).permute(2, 1, 0, 3) + # now x: (num_heads, batch_size, seq_len, value_head_dim) + value_head_dim = x.shape[-1] + + # todo: see whether there is benefit in overriding matmul + x = torch.matmul(attn_weights, x) + # v: (num_heads, batch_size, seq_len, value_head_dim) + + x = ( + x.permute(2, 1, 0, 3) + .contiguous() + .view(seq_len, batch_size, num_heads * value_head_dim) + ) + + # returned value is of shape (seq_len, batch_size, embed_dim), like the input. + x = self.out_proj(x) + x = self.whiten(x) + + return x + + def streaming_forward( + self, + x: Tensor, + attn_weights: Tensor, + cached_val: Tensor, + left_context_len: int, + ) -> Tuple[Tensor, Tensor]: + """ + Args: + x: input tensor, of shape (seq_len, batch_size, embed_dim) + attn_weights: a tensor of shape (num_heads, batch_size, seq_len, seq_len), + with seq_len being interpreted as (tgt_seq_len, src_seq_len). Expect + attn_weights.sum(dim=-1) == 1. + cached_val: cached attention value tensor of left context, + of shape (left_context_len, batch_size, value_dim) + left_context_len: number of left context frames. + + Returns: + - attention weighted output, a tensor with the same shape as x. + - updated cached attention value tensor of left context. + """ + (seq_len, batch_size, embed_dim) = x.shape + num_heads = attn_weights.shape[0] + seq_len2 = seq_len + left_context_len + assert attn_weights.shape == (num_heads, batch_size, seq_len, seq_len2) + + x = self.in_proj(x) # (seq_len, batch_size, num_heads * value_head_dim) + + # Pad cached left contexts + assert cached_val.shape[0] == left_context_len, ( + cached_val.shape[0], + left_context_len, + ) + x = torch.cat([cached_val, x], dim=0) + # Update cached left contexts + cached_val = x[-left_context_len:, ...] + + x = x.reshape(seq_len2, batch_size, num_heads, -1).permute(2, 1, 0, 3) + # now x: (num_heads, batch_size, seq_len, value_head_dim) + value_head_dim = x.shape[-1] + + # todo: see whether there is benefit in overriding matmul + x = torch.matmul(attn_weights, x) + # v: (num_heads, batch_size, seq_len, value_head_dim) + + x = ( + x.permute(2, 1, 0, 3) + .contiguous() + .view(seq_len, batch_size, num_heads * value_head_dim) + ) + + # returned value is of shape (seq_len, batch_size, embed_dim), like the input. + x = self.out_proj(x) + + return x, cached_val + + +class FeedforwardModule(nn.Module): + """Feedforward module in Zipformer2 model.""" + + def __init__(self, embed_dim: int, feedforward_dim: int, dropout: FloatLike): + super(FeedforwardModule, self).__init__() + self.in_proj = nn.Linear(embed_dim, feedforward_dim) + + self.hidden_balancer = Balancer( + feedforward_dim, + channel_dim=-1, + min_positive=0.3, + max_positive=1.0, + min_abs=0.75, + max_abs=5.0, + ) + + # shared_dim=0 means we share the dropout mask along the time axis + self.out_proj = ActivationDropoutAndLinear( + feedforward_dim, + embed_dim, + activation="SwooshL", + dropout_p=dropout, + dropout_shared_dim=0, + bias=True, + initial_scale=0.1, + ) + + self.out_whiten = Whiten( + num_groups=1, + whitening_limit=_whitening_schedule(7.5), + prob=(0.025, 0.25), + grad_scale=0.01, + ) + + def forward(self, x: Tensor): + x = self.in_proj(x) + x = self.hidden_balancer(x) + # out_proj contains SwooshL activation, then dropout, then linear. + x = self.out_proj(x) + x = self.out_whiten(x) + return x + + +class NonlinAttention(nn.Module): + """This is like the ConvolutionModule, but refactored so that we use multiplication by attention weights (borrowed + from the attention module) in place of actual convolution. We also took out the second nonlinearity, the + one after the attention mechanism. + + Args: + channels (int): The number of channels of conv layers. + """ + + def __init__( + self, + channels: int, + hidden_channels: int, + ) -> None: + super().__init__() + + self.hidden_channels = hidden_channels + + self.in_proj = nn.Linear(channels, hidden_channels * 3, bias=True) + + # balancer that goes before the sigmoid. Have quite a large min_abs value, at 2.0, + # because we noticed that well-trained instances of this module have abs-value before the sigmoid + # starting from about 3, and poorly-trained instances of the module have smaller abs values + # before the sigmoid. + self.balancer = Balancer( + hidden_channels, + channel_dim=-1, + min_positive=ScheduledFloat((0.0, 0.25), (20000.0, 0.05)), + max_positive=ScheduledFloat((0.0, 0.75), (20000.0, 0.95)), + min_abs=0.5, + max_abs=5.0, + ) + self.tanh = nn.Tanh() + + self.identity1 = Identity() # for diagnostics. + self.identity2 = Identity() # for diagnostics. + self.identity3 = Identity() # for diagnostics. + + self.out_proj = ScaledLinear( + hidden_channels, channels, bias=True, initial_scale=0.05 + ) + + self.whiten1 = Whiten( + num_groups=1, + whitening_limit=_whitening_schedule(5.0), + prob=(0.025, 0.25), + grad_scale=0.01, + ) + + self.whiten2 = Whiten( + num_groups=1, + whitening_limit=_whitening_schedule(5.0, ratio=3.0), + prob=(0.025, 0.25), + grad_scale=0.01, + ) + + def forward( + self, + x: Tensor, + attn_weights: Tensor, + ) -> Tensor: + """. + Args: + x: a Tensor of shape (seq_len, batch_size, num_channels) + attn_weights: a Tensor of shape (num_heads, batch_size, seq_len, seq_len) + Returns: + a Tensor with the same shape as x + """ + x = self.in_proj(x) + + (seq_len, batch_size, _) = x.shape + hidden_channels = self.hidden_channels + + s, x, y = x.chunk(3, dim=2) + + # s will go through tanh. + + s = self.balancer(s) + s = self.tanh(s) + + s = s.unsqueeze(-1).reshape(seq_len, batch_size, hidden_channels) + x = self.whiten1(x) + x = x * s + x = self.identity1(x) # diagnostics only, it's the identity. + + (seq_len, batch_size, embed_dim) = x.shape + num_heads = attn_weights.shape[0] + assert attn_weights.shape == (num_heads, batch_size, seq_len, seq_len) + + x = x.reshape(seq_len, batch_size, num_heads, -1).permute(2, 1, 0, 3) + # now x: (num_heads, batch_size, seq_len, head_dim) + x = torch.matmul(attn_weights, x) + # now x: (num_heads, batch_size, seq_len, head_dim) + x = x.permute(2, 1, 0, 3).reshape(seq_len, batch_size, -1) + + y = self.identity2(y) + x = x * y + x = self.identity3(x) + + x = self.out_proj(x) + x = self.whiten2(x) + return x + + def streaming_forward( + self, + x: Tensor, + attn_weights: Tensor, + cached_x: Tensor, + left_context_len: int, + ) -> Tuple[Tensor, Tensor]: + """. + Args: + x: a Tensor of shape (seq_len, batch_size, num_channels) + attn_weights: a Tensor of shape (num_heads, batch_size, seq_len, seq_len) + cached_x: left context, a Tensor of shape + (num_heads, batch_size, left_context_len, head_dim) + left_context_len: number of left context frames. + Returns: + - a Tensor with the same shape as x + - updated left context with same shape as cached_x + """ + x = self.in_proj(x) + + (seq_len, batch_size, _) = x.shape + hidden_channels = self.hidden_channels + + s, x, y = x.chunk(3, dim=2) + + # s will go through tanh. + s = self.tanh(s) + + s = s.unsqueeze(-1).reshape(seq_len, batch_size, hidden_channels) + x = x * s + + (seq_len, batch_size, embed_dim) = x.shape + num_heads = attn_weights.shape[0] + assert attn_weights.shape == ( + num_heads, + batch_size, + seq_len, + left_context_len + seq_len, + ) + + x = x.reshape(seq_len, batch_size, num_heads, -1).permute(2, 1, 0, 3) + # now x: (num_heads, batch_size, seq_len, head_dim) + + # Pad cached tensor + assert cached_x.shape[2] == left_context_len, ( + cached_x.shape[2], + left_context_len, + ) + x_pad = torch.cat([cached_x, x], dim=2) + # Update cached tensor + cached_x = x_pad[:, :, -left_context_len:, :] + + x = torch.matmul(attn_weights, x_pad) + # now x: (num_heads, batch_size, seq_len, head_dim) + x = x.permute(2, 1, 0, 3).reshape(seq_len, batch_size, -1) + + x = x * y + + x = self.out_proj(x) + return x, cached_x + + +class ConvolutionModule(nn.Module): + """ConvolutionModule in Zipformer2 model. + Modified from https://github.com/espnet/espnet/blob/master/espnet/nets/pytorch_backend/zipformer/convolution.py + + Args: + channels (int): The number of channels of conv layers. + kernel_size (int): Kernerl size of conv layers. + bias (bool): Whether to use bias in conv layers (default=True). + + """ + + def __init__( + self, + channels: int, + kernel_size: int, + causal: bool, + ) -> None: + """Construct a ConvolutionModule object.""" + super(ConvolutionModule, self).__init__() + # kernerl_size should be a odd number for 'SAME' padding + assert (kernel_size - 1) % 2 == 0 + + bottleneck_dim = channels + self.causal = causal + + self.in_proj = nn.Linear( + channels, + 2 * bottleneck_dim, + ) + # the gradients on in_proj are a little noisy, likely to do with the + # sigmoid in glu. + + # after in_proj we put x through a gated linear unit (nn.functional.glu). + # For most layers the normal rms value of channels of x seems to be in the range 1 to 4, + # but sometimes, for some reason, for layer 0 the rms ends up being very large, + # between 50 and 100 for different channels. This will cause very peaky and + # sparse derivatives for the sigmoid gating function, which will tend to make + # the loss function not learn effectively. (for most layers the average absolute values + # are in the range 0.5..9.0, and the average p(x>0), i.e. positive proportion, + # at the output of pointwise_conv1.output is around 0.35 to 0.45 for different + # layers, which likely breaks down as 0.5 for the "linear" half and + # 0.2 to 0.3 for the part that goes into the sigmoid. The idea is that if we + # constrain the rms values to a reasonable range via a constraint of max_abs=10.0, + # it will be in a better position to start learning something, i.e. to latch onto + # the correct range. + self.balancer1 = Balancer( + bottleneck_dim, + channel_dim=-1, + min_positive=ScheduledFloat((0.0, 0.05), (8000.0, 0.025)), + max_positive=1.0, + min_abs=1.5, + max_abs=ScheduledFloat((0.0, 5.0), (8000.0, 10.0), default=1.0), + ) + + self.activation1 = Identity() # for diagnostics + + self.sigmoid = nn.Sigmoid() + + self.activation2 = Identity() # for diagnostics + + assert kernel_size % 2 == 1 + + self.depthwise_conv = ( + ChunkCausalDepthwiseConv1d(channels=bottleneck_dim, kernel_size=kernel_size) + if causal + else nn.Conv1d( + in_channels=bottleneck_dim, + out_channels=bottleneck_dim, + groups=bottleneck_dim, + kernel_size=kernel_size, + padding=kernel_size // 2, + ) + ) + + self.balancer2 = Balancer( + bottleneck_dim, + channel_dim=1, + min_positive=ScheduledFloat((0.0, 0.1), (8000.0, 0.05)), + max_positive=1.0, + min_abs=ScheduledFloat((0.0, 0.2), (20000.0, 0.5)), + max_abs=10.0, + ) + + self.whiten = Whiten( + num_groups=1, + whitening_limit=_whitening_schedule(7.5), + prob=(0.025, 0.25), + grad_scale=0.01, + ) + + self.out_proj = ActivationDropoutAndLinear( + bottleneck_dim, + channels, + activation="SwooshR", + dropout_p=0.0, + initial_scale=0.05, + ) + + def forward( + self, + x: Tensor, + src_key_padding_mask: Optional[Tensor] = None, + chunk_size: int = -1, + ) -> Tensor: + """Compute convolution module. + + Args: + x: Input tensor (#time, batch, channels). + src_key_padding_mask: the mask for the src keys per batch (optional): + (batch, #time), contains True in masked positions. + + Returns: + Tensor: Output tensor (#time, batch, channels). + + """ + + x = self.in_proj(x) # (time, batch, 2*channels) + + x, s = x.chunk(2, dim=2) + s = self.balancer1(s) + s = self.sigmoid(s) + x = self.activation1(x) # identity. + x = x * s + x = self.activation2(x) # identity + + # (time, batch, channels) + + # exchange the temporal dimension and the feature dimension + x = x.permute(1, 2, 0) # (#batch, channels, time). + + if src_key_padding_mask is not None: + x = x.masked_fill(src_key_padding_mask.unsqueeze(1).expand_as(x), 0.0) + + if ( + not torch.jit.is_scripting() + and not torch.jit.is_tracing() + and chunk_size >= 0 + ): + # Not support exporting a model for simulated streaming decoding + assert ( + self.causal + ), "Must initialize model with causal=True if you use chunk_size" + x = self.depthwise_conv(x, chunk_size=chunk_size) + else: + x = self.depthwise_conv(x) + + x = self.balancer2(x) + x = x.permute(2, 0, 1) # (time, batch, channels) + + x = self.whiten(x) # (time, batch, channels) + x = self.out_proj(x) # (time, batch, channels) + + return x + + def streaming_forward( + self, + x: Tensor, + cache: Tensor, + src_key_padding_mask: Tensor, + ) -> Tuple[Tensor, Tensor]: + """Compute convolution module in streaming forward mode. + + Args: + x: Input tensor (#time, batch, channels). + cache: cached left context for depthwise_conv of shape + (#batch, channels, left_pad) + src_key_padding_mask: the mask for the src keys per batch (optional): + (batch, #time), contains True in masked positions. + + Returns: + - Output tensor (#time, batch, channels). + - Updated cache (#batch, channels, left_pad) + """ + + x = self.in_proj(x) # (time, batch, 2*channels) + + x, s = x.chunk(2, dim=2) + s = self.sigmoid(s) + x = x * s + # (time, batch, channels) + + # exchange the temporal dimension and the feature dimension + x = x.permute(1, 2, 0) # (#batch, channels, time). + + if src_key_padding_mask is not None: + x = x.masked_fill(src_key_padding_mask.unsqueeze(1).expand_as(x), 0.0) + + x, cache = self.depthwise_conv.streaming_forward(x, cache=cache) + + x = x.permute(2, 0, 1) # (time, batch, channels) + + x = self.out_proj(x) # (time, batch, channels) + + return x, cache + + +class ScalarMultiply(nn.Module): + def __init__(self, scale: float): + super().__init__() + self.scale = scale + + def forward(self, x): + return x * self.scale + + +def _test_zipformer_main(causal: bool = False): + batch_size = 5 + seq_len = 20 + # Just make sure the forward pass runs. + + c = Zipformer2( + encoder_dim=(64, 96), + encoder_unmasked_dim=(48, 64), + num_heads=(4, 4), + causal=causal, + chunk_size=(4,) if causal else (-1,), + left_context_frames=(64,), + ) + batch_size = 5 + seq_len = 20 + # Just make sure the forward pass runs. + f = c( + torch.randn(seq_len, batch_size, 64), + torch.full((batch_size,), seq_len, dtype=torch.int64), + ) + f[0].sum().backward() + c.eval() + f = c( + torch.randn(seq_len, batch_size, 64), + torch.full((batch_size,), seq_len, dtype=torch.int64), + ) + f # to remove flake8 warnings + + +if __name__ == "__main__": + logging.getLogger().setLevel(logging.INFO) + torch.set_num_threads(1) + torch.set_num_interop_threads(1) + _test_zipformer_main(False) + _test_zipformer_main(True)