diff --git a/egs/librispeech/ASR/RESULTS.md b/egs/librispeech/ASR/RESULTS.md index a1808edd3..cc809860d 100644 --- a/egs/librispeech/ASR/RESULTS.md +++ b/egs/librispeech/ASR/RESULTS.md @@ -375,6 +375,38 @@ for m in greedy_search modified_beam_search fast_beam_search; do done ``` +### zipformer ctc streaming + +| decoding method | test-clean | test-other | comment | +|----------------------|------------|------------|--------------------| +| greedy_search | 4.07 | 10.51 | --epoch 30 --avg 15| +| greedy_search | 4.0 | 10.3 | --epoch 30 --avg 9 | + +The training command is: +```bash +export CUDA_VISIBLE_DEVICES="0,1" +python ./zipformer_ctc_streaming/train.py \ + --world-size 2 \ + --num-epochs 30 \ + --start-epoch 1 \ + --exp-dir ./zipformer_ctc_streaming/exp \ + --max-duration 100 +``` + +The decoding command is: +```bash +export CUDA_VISIBLE_DEVICES="0" +./zipformer_ctc_streaming/decode.py \ +--epoch 30 \ +--avg 15 \ +--exp-dir ./zipformer_ctc_streaming/exp \ +--max-duration 300 \ +--decode-chunk-len 32 \ +--method ctc-decoding \ +--lm-dir data/lm \ +--lang-dir data/lang_bpe_500 +``` + ### pruned_transducer_stateless7 (Fine-tune with mux) See for more details. @@ -456,6 +488,39 @@ The decoding commands are: --max-states 64 ``` +### zipformer ctc streaming + +| decoding method | test-clean | test-other | comment | +|----------------------|------------|------------|--------------------| +| greedy_search | 4.07 | 10.51 | --epoch 30 --avg 15| +| greedy_search | 4.0 | 10.3 | --epoch 30 --avg 9 | + +The training command is: +```bash +export CUDA_VISIBLE_DEVICES="0,1" +python ./zipformer_ctc_streaming/train.py \ + --world-size 2 \ + --num-epochs 30 \ + --start-epoch 1 \ + --exp-dir ./zipformer_ctc_streaming/exp \ + --max-duration 100 +``` + +The decoding command is: +```bash +export CUDA_VISIBLE_DEVICES="0" +./zipformer_ctc_streaming/decode.py \ +--epoch 30 \ +--avg 15 \ +--exp-dir ./zipformer_ctc_streaming/exp \ +--max-duration 300 \ +--decode-chunk-len 32 \ +--method ctc-decoding \ +--lm-dir data/lm \ +--lang-dir data/lang_bpe_500 +``` + + ### pruned_transducer_stateless7 (zipformer + multidataset(LibriSpeech + GigaSpeech + CommonVoice 13.0)) See for more details. diff --git a/egs/librispeech/ASR/zipformer_ctc_streaming/__init__.py b/egs/librispeech/ASR/zipformer_ctc_streaming/__init__.py new file mode 100644 index 000000000..e69de29bb diff --git a/egs/librispeech/ASR/zipformer_ctc_streaming/asr_datamodule.py b/egs/librispeech/ASR/zipformer_ctc_streaming/asr_datamodule.py new file mode 100644 index 000000000..c47964b07 --- /dev/null +++ b/egs/librispeech/ASR/zipformer_ctc_streaming/asr_datamodule.py @@ -0,0 +1,475 @@ +# Copyright 2021 Piotr Żelasko +# Copyright 2022 Xiaomi Corporation (Author: Mingshuang Luo) +# +# See ../../../../LICENSE for clarification regarding multiple authors +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + + +import argparse +import inspect +import logging +from functools import lru_cache +from pathlib import Path +from typing import Any, Dict, Optional + +import torch +from lhotse import CutSet, Fbank, FbankConfig, load_manifest, load_manifest_lazy +from lhotse.dataset import ( # noqa F401 for PrecomputedFeatures + CutConcatenate, + CutMix, + DynamicBucketingSampler, + K2SpeechRecognitionDataset, + PrecomputedFeatures, + SingleCutSampler, + SpecAugment, +) +from lhotse.dataset.input_strategies import ( # noqa F401 For AudioSamples + AudioSamples, + OnTheFlyFeatures, +) +from lhotse.utils import fix_random_seed +from torch.utils.data import DataLoader + +from icefall.utils import str2bool + + +class _SeedWorkers: + def __init__(self, seed: int): + self.seed = seed + + def __call__(self, worker_id: int): + fix_random_seed(self.seed + worker_id) + + +class LibriSpeechAsrDataModule: + """ + DataModule for k2 ASR experiments. + It assumes there is always one train and valid dataloader, + but there can be multiple test dataloaders (e.g. LibriSpeech test-clean + and test-other). + + It contains all the common data pipeline modules used in ASR + experiments, e.g.: + - dynamic batch size, + - bucketing samplers, + - cut concatenation, + - augmentation, + - on-the-fly feature extraction + + This class should be derived for specific corpora used in ASR tasks. + """ + + def __init__(self, args: argparse.Namespace): + self.args = args + + @classmethod + def add_arguments(cls, parser: argparse.ArgumentParser): + group = parser.add_argument_group( + title="ASR data related options", + description="These options are used for the preparation of " + "PyTorch DataLoaders from Lhotse CutSet's -- they control the " + "effective batch sizes, sampling strategies, applied data " + "augmentations, etc.", + ) + group.add_argument( + "--full-libri", + type=str2bool, + default=True, + help="""Used only when --mini-libri is False.When enabled, + use 960h LibriSpeech. Otherwise, use 100h subset.""", + ) + group.add_argument( + "--mini-libri", + type=str2bool, + default=False, + help="True for mini librispeech", + ) + + group.add_argument( + "--manifest-dir", + type=Path, + default=Path("data/fbank"), + help="Path to directory with train/valid/test cuts.", + ) + group.add_argument( + "--max-duration", + type=int, + default=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( + "--concatenate-cuts", + type=str2bool, + default=False, + help="When enabled, utterances (cuts) will be concatenated " + "to minimize the amount of padding.", + ) + group.add_argument( + "--duration-factor", + type=float, + default=1.0, + help="Determines the maximum duration of a concatenated cut " + "relative to the duration of the longest cut in a batch.", + ) + group.add_argument( + "--gap", + type=float, + default=1.0, + help="The amount of padding (in seconds) inserted between " + "concatenated cuts. This padding is filled with noise when " + "noise augmentation is used.", + ) + group.add_argument( + "--on-the-fly-feats", + type=str2bool, + default=False, + help="When enabled, use on-the-fly cut mixing and feature " + "extraction. Will drop existing precomputed feature manifests " + "if available.", + ) + group.add_argument( + "--shuffle", + type=str2bool, + default=True, + help="When enabled (=default), the examples will be " + "shuffled for each epoch.", + ) + group.add_argument( + "--drop-last", + type=str2bool, + default=True, + help="Whether to drop last batch. Used by sampler.", + ) + group.add_argument( + "--return-cuts", + type=str2bool, + default=True, + help="When enabled, each batch will have the " + "field: batch['supervisions']['cut'] with the cuts that " + "were used to construct it.", + ) + + group.add_argument( + "--num-workers", + type=int, + default=2, + help="The number of training dataloader workers that " + "collect the batches.", + ) + + group.add_argument( + "--enable-spec-aug", + type=str2bool, + default=True, + help="When enabled, use SpecAugment for training dataset.", + ) + + group.add_argument( + "--spec-aug-time-warp-factor", + type=int, + default=80, + help="Used only when --enable-spec-aug is True. " + "It specifies the factor for time warping in SpecAugment. " + "Larger values mean more warping. " + "A value less than 1 means to disable time warp.", + ) + + group.add_argument( + "--enable-musan", + type=str2bool, + default=True, + help="When enabled, select noise from MUSAN and mix it" + "with training dataset. ", + ) + + group.add_argument( + "--input-strategy", + type=str, + default="PrecomputedFeatures", + help="AudioSamples or PrecomputedFeatures", + ) + + def train_dataloaders( + self, + cuts_train: CutSet, + sampler_state_dict: Optional[Dict[str, Any]] = None, + ) -> DataLoader: + """ + Args: + cuts_train: + CutSet for training. + sampler_state_dict: + The state dict for the training sampler. + """ + transforms = [] + if self.args.enable_musan: + logging.info("Enable MUSAN") + logging.info("About to get Musan cuts") + cuts_musan = load_manifest(self.args.manifest_dir / "musan_cuts.jsonl.gz") + transforms.append( + CutMix(cuts=cuts_musan, prob=0.5, snr=(10, 20), preserve_id=True) + ) + else: + logging.info("Disable MUSAN") + + if self.args.concatenate_cuts: + logging.info( + f"Using cut concatenation with duration factor " + f"{self.args.duration_factor} and gap {self.args.gap}." + ) + # Cut concatenation should be the first transform in the list, + # so that if we e.g. mix noise in, it will fill the gaps between + # different utterances. + transforms = [ + CutConcatenate( + duration_factor=self.args.duration_factor, gap=self.args.gap + ) + ] + transforms + + input_transforms = [] + if self.args.enable_spec_aug: + logging.info("Enable SpecAugment") + logging.info(f"Time warp factor: {self.args.spec_aug_time_warp_factor}") + # Set the value of num_frame_masks according to Lhotse's version. + # In different Lhotse's versions, the default of num_frame_masks is + # different. + num_frame_masks = 10 + num_frame_masks_parameter = inspect.signature( + SpecAugment.__init__ + ).parameters["num_frame_masks"] + if num_frame_masks_parameter.default == 1: + num_frame_masks = 2 + logging.info(f"Num frame mask: {num_frame_masks}") + input_transforms.append( + SpecAugment( + time_warp_factor=self.args.spec_aug_time_warp_factor, + num_frame_masks=num_frame_masks, + features_mask_size=27, + num_feature_masks=2, + frames_mask_size=100, + ) + ) + else: + logging.info("Disable SpecAugment") + + logging.info("About to create train dataset") + train = K2SpeechRecognitionDataset( + input_strategy=eval(self.args.input_strategy)(), + cut_transforms=transforms, + input_transforms=input_transforms, + return_cuts=self.args.return_cuts, + ) + + if self.args.on_the_fly_feats: + # NOTE: the PerturbSpeed transform should be added only if we + # remove it from data prep stage. + # Add on-the-fly speed perturbation; since originally it would + # have increased epoch size by 3, we will apply prob 2/3 and use + # 3x more epochs. + # Speed perturbation probably should come first before + # concatenation, but in principle the transforms order doesn't have + # to be strict (e.g. could be randomized) + # transforms = [PerturbSpeed(factors=[0.9, 1.1], p=2/3)] + transforms # noqa + # Drop feats to be on the safe side. + train = K2SpeechRecognitionDataset( + cut_transforms=transforms, + input_strategy=OnTheFlyFeatures(Fbank(FbankConfig(num_mel_bins=80))), + input_transforms=input_transforms, + return_cuts=self.args.return_cuts, + ) + + if self.args.bucketing_sampler: + logging.info("Using DynamicBucketingSampler.") + train_sampler = DynamicBucketingSampler( + cuts_train, + max_duration=self.args.max_duration, + shuffle=self.args.shuffle, + num_buckets=self.args.num_buckets, + drop_last=self.args.drop_last, + ) + else: + logging.info("Using SingleCutSampler.") + train_sampler = SingleCutSampler( + 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) -> DataLoader: + transforms = [] + if self.args.concatenate_cuts: + transforms = [ + CutConcatenate( + duration_factor=self.args.duration_factor, gap=self.args.gap + ) + ] + transforms + + logging.info("About to create dev dataset") + if self.args.on_the_fly_feats: + validate = K2SpeechRecognitionDataset( + cut_transforms=transforms, + input_strategy=OnTheFlyFeatures(Fbank(FbankConfig(num_mel_bins=80))), + return_cuts=self.args.return_cuts, + ) + else: + validate = K2SpeechRecognitionDataset( + cut_transforms=transforms, + return_cuts=self.args.return_cuts, + ) + valid_sampler = DynamicBucketingSampler( + cuts_valid, + max_duration=self.args.max_duration, + shuffle=False, + ) + logging.info("About to create dev dataloader") + valid_dl = DataLoader( + validate, + sampler=valid_sampler, + batch_size=None, + num_workers=2, + persistent_workers=False, + ) + + return valid_dl + + def test_dataloaders(self, cuts: CutSet) -> DataLoader: + logging.debug("About to create test dataset") + test = K2SpeechRecognitionDataset( + input_strategy=OnTheFlyFeatures(Fbank(FbankConfig(num_mel_bins=80))) + if self.args.on_the_fly_feats + else eval(self.args.input_strategy)(), + return_cuts=self.args.return_cuts, + ) + sampler = DynamicBucketingSampler( + cuts, + max_duration=self.args.max_duration, + shuffle=False, + ) + logging.debug("About to create test dataloader") + test_dl = DataLoader( + test, + batch_size=None, + sampler=sampler, + num_workers=self.args.num_workers, + ) + return test_dl + + @lru_cache() + def train_clean_5_cuts(self) -> CutSet: + logging.info("mini_librispeech: About to get train-clean-5 cuts") + return load_manifest_lazy( + self.args.manifest_dir / "librispeech_cuts_train-clean-5.jsonl.gz" + ) + + @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" + ) + return load_manifest_lazy( + self.args.manifest_dir / "librispeech_cuts_train-all-shuf.jsonl.gz" + ) + + @lru_cache() + def dev_clean_2_cuts(self) -> CutSet: + logging.info("mini_librispeech: About to get dev-clean-2 cuts") + return load_manifest_lazy( + self.args.manifest_dir / "librispeech_cuts_dev-clean-2.jsonl.gz" + ) + + @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/ASR/zipformer_ctc_streaming/decode.py b/egs/librispeech/ASR/zipformer_ctc_streaming/decode.py new file mode 100755 index 000000000..3954df3ed --- /dev/null +++ b/egs/librispeech/ASR/zipformer_ctc_streaming/decode.py @@ -0,0 +1,891 @@ +#!/usr/bin/env python3 +# Copyright 2021 Xiaomi Corporation (Author: Liyong Guo, Fangjun Kuang) +# +# See ../../../../LICENSE for clarification regarding multiple authors +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + + +import argparse +import logging +from 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 train import add_model_arguments, get_ctc_model, get_params +from transformer import encoder_padding_mask + +from icefall.bpe_graph_compiler import BpeCtcTrainingGraphCompiler +from icefall.checkpoint import ( + average_checkpoints, + average_checkpoints_with_averaged_model, + find_checkpoints, + load_checkpoint, +) +from icefall.decode import ( + get_lattice, + nbest_decoding, + nbest_oracle, + one_best_decoding, + rescore_with_attention_decoder, + rescore_with_n_best_list, + rescore_with_rnn_lm, + rescore_with_whole_lattice, +) +from icefall.lexicon import Lexicon +from icefall.rnn_lm.model import RnnLmModel +from icefall.utils import ( + AttributeDict, + get_texts, + load_averaged_model, + 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=77, + help="It specifies the checkpoint to use for decoding." + "Note: Epoch counts from 0.", + ) + + parser.add_argument( + "--iter", + type=int, + default=0, + help="""If positive, --epoch is ignored and it + will use the checkpoint exp_dir/checkpoint-iter.pt. + You can specify --avg to use more checkpoints for model averaging. + """, + ) + + parser.add_argument( + "--avg", + type=int, + default=55, + help="Number of checkpoints to average. Automatically select " + "consecutive checkpoints before the checkpoint specified by " + "'--epoch'. ", + ) + + parser.add_argument( + "--use-averaged-model", + type=str2bool, + default=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( + "--method", + type=str, + default="attention-decoder", + help="""Decoding method. + Supported values are: + - (0) ctc-decoding. Use CTC decoding. It uses a sentence piece + model, i.e., lang_dir/bpe.model, to convert word pieces to words. + It needs neither a lexicon nor an n-gram LM. + - (1) 1best. Extract the best path from the decoding lattice as the + decoding result. + - (2) nbest. Extract n paths from the decoding lattice; the path + with the highest score is the decoding result. + - (3) nbest-rescoring. Extract n paths from the decoding lattice, + rescore them with an n-gram LM (e.g., a 4-gram LM), the path with + the highest score is the decoding result. + - (4) whole-lattice-rescoring. Rescore the decoding lattice with an + n-gram LM (e.g., a 4-gram LM), the best path of rescored lattice + is the decoding result. + - (5) attention-decoder. Extract n paths from the LM rescored + lattice, the path with the highest score is the decoding result. + - (6) rnn-lm. Rescoring with attention-decoder and RNN LM. We assume + you have trained an RNN LM using ./rnn_lm/train.py + - (7) nbest-oracle. Its WER is the lower bound of any n-best + rescoring method can achieve. Useful for debugging n-best + rescoring method. + """, + ) + + parser.add_argument( + "--num-paths", + type=int, + default=100, + help="""Number of paths for n-best based decoding method. + Used only when "method" is one of the following values: + nbest, nbest-rescoring, attention-decoder, rnn-lm, and nbest-oracle + """, + ) + + parser.add_argument( + "--nbest-scale", + type=float, + default=0.5, + help="""The scale to be applied to `lattice.scores`. + It's needed if you use any kinds of n-best based rescoring. + Used only when "method" is one of the following values: + nbest, nbest-rescoring, attention-decoder, rnn-lm, and nbest-oracle + A smaller value results in more unique paths. + """, + ) + + parser.add_argument( + "--exp-dir", + type=str, + default="zipformer_ctc/exp", + help="The experiment dir", + ) + + parser.add_argument( + "--lang-dir", + type=str, + default="data/lang_bpe_500", + help="The lang dir", + ) + + parser.add_argument( + "--lm-dir", + type=str, + default="data/lm", + help="""The n-gram LM dir. + It should contain either G_3_gram.pt or G_3_gram.fst.txt + """, + ) + + parser.add_argument( + "--rnn-lm-exp-dir", + type=str, + default="rnn_lm/exp", + help="""Used only when --method is rnn-lm. + It specifies the path to RNN LM exp dir. + """, + ) + + parser.add_argument( + "--rnn-lm-epoch", + type=int, + default=7, + help="""Used only when --method is rnn-lm. + It specifies the checkpoint to use. + """, + ) + + parser.add_argument( + "--rnn-lm-avg", + type=int, + default=2, + help="""Used only when --method is rnn-lm. + It specifies the number of checkpoints to average. + """, + ) + + parser.add_argument( + "--rnn-lm-embedding-dim", + type=int, + default=2048, + help="Embedding dim of the model", + ) + + parser.add_argument( + "--rnn-lm-hidden-dim", + type=int, + default=2048, + help="Hidden dim of the model", + ) + + parser.add_argument( + "--rnn-lm-num-layers", + type=int, + default=4, + help="Number of RNN layers the model", + ) + parser.add_argument( + "--rnn-lm-tie-weights", + type=str2bool, + default=False, + help="""True to share the weights between the input embedding layer and the + last output linear layer + """, + ) + + add_model_arguments(parser) + + return parser + + +def decode_one_batch( + params: AttributeDict, + model: nn.Module, + rnn_lm_model: Optional[nn.Module], + HLG: Optional[k2.Fsa], + H: Optional[k2.Fsa], + bpe_model: Optional[spm.SentencePieceProcessor], + batch: dict, + word_table: k2.SymbolTable, + sos_id: int, + eos_id: int, + G: Optional[k2.Fsa] = None, +) -> Dict[str, List[List[str]]]: + """Decode one batch and return the result in a dict. The dict has the + following format: + + - key: It indicates the setting used for decoding. For example, + if no rescoring is used, the key is the string `no_rescore`. + If LM rescoring is used, the key is the string `lm_scale_xxx`, + where `xxx` is the value of `lm_scale`. An example key is + `lm_scale_0.7` + - 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`. + + - params.method is "1best", it uses 1best decoding without LM rescoring. + - params.method is "nbest", it uses nbest decoding without LM rescoring. + - params.method is "nbest-rescoring", it uses nbest LM rescoring. + - params.method is "whole-lattice-rescoring", it uses whole lattice LM + rescoring. + + model: + The neural model. + rnn_lm_model: + The neural model for RNN LM. + HLG: + The decoding graph. Used only when params.method is NOT ctc-decoding. + H: + The ctc topo. Used only when params.method is ctc-decoding. + bpe_model: + The BPE model. Used only when params.method is ctc-decoding. + 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. + sos_id: + The token ID of the SOS. + eos_id: + The token ID of the EOS. + G: + An LM. It is not None when params.method is "nbest-rescoring" + or "whole-lattice-rescoring". In general, the G in HLG + is a 3-gram LM, while this G is a 4-gram LM. + Returns: + Return the decoding result. See above description for the format of + the returned dict. Note: If it decodes to nothing, then return None. + """ + if HLG is not None: + device = HLG.device + else: + device = H.device + feature = batch["inputs"] + assert feature.ndim == 3 + feature = feature.to(device) + # at entry, feature is (N, T, C) + + supervisions = batch["supervisions"] + feature_lens = supervisions["num_frames"].to(device) + + nnet_output, _ = model.encoder(feature, feature_lens) + ctc_output = model.ctc_output(nnet_output) + # nnet_output is (N, T, C) + + supervision_segments = torch.stack( + ( + supervisions["sequence_idx"], + supervisions["start_frame"] // params.subsampling_factor, + supervisions["num_frames"] // params.subsampling_factor, + ), + 1, + ).to(torch.int32) + + if H is None: + assert HLG is not None + decoding_graph = HLG + else: + assert HLG is None + assert bpe_model is not None + decoding_graph = H + + lattice = get_lattice( + nnet_output=ctc_output, + decoding_graph=decoding_graph, + supervision_segments=supervision_segments, + search_beam=params.search_beam, + output_beam=params.output_beam, + min_active_states=params.min_active_states, + max_active_states=params.max_active_states, + subsampling_factor=params.subsampling_factor, + ) + + if params.method == "ctc-decoding": + best_path = one_best_decoding( + lattice=lattice, use_double_scores=params.use_double_scores + ) + # Note: `best_path.aux_labels` contains token IDs, not word IDs + # since we are using H, not HLG here. + # + # token_ids is a lit-of-list of IDs + token_ids = get_texts(best_path) + + # hyps is a list of str, e.g., ['xxx yyy zzz', ...] + hyps = bpe_model.decode(token_ids) + + # hyps is a list of list of str, e.g., [['xxx', 'yyy', 'zzz'], ... ] + hyps = [s.split() for s in hyps] + key = "ctc-decoding" + return {key: hyps} + + if params.method == "nbest-oracle": + # Note: You can also pass rescored lattices to it. + # We choose the HLG decoded lattice for speed reasons + # as HLG decoding is faster and the oracle WER + # is only slightly worse than that of rescored lattices. + best_path = nbest_oracle( + lattice=lattice, + num_paths=params.num_paths, + ref_texts=supervisions["text"], + word_table=word_table, + nbest_scale=params.nbest_scale, + oov="", + ) + hyps = get_texts(best_path) + hyps = [[word_table[i] for i in ids] for ids in hyps] + key = f"oracle_{params.num_paths}_nbest_scale_{params.nbest_scale}" # noqa + return {key: hyps} + + if params.method in ["1best", "nbest"]: + if params.method == "1best": + best_path = one_best_decoding( + lattice=lattice, use_double_scores=params.use_double_scores + ) + key = "no_rescore" + else: + best_path = nbest_decoding( + lattice=lattice, + num_paths=params.num_paths, + use_double_scores=params.use_double_scores, + nbest_scale=params.nbest_scale, + ) + key = f"no_rescore-nbest-scale-{params.nbest_scale}-{params.num_paths}" # noqa + + hyps = get_texts(best_path) + hyps = [[word_table[i] for i in ids] for ids in hyps] + return {key: hyps} + + assert params.method in [ + "nbest-rescoring", + "whole-lattice-rescoring", + "attention-decoder", + "rnn-lm", + ] + + lm_scale_list = [0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7] + + nnet_output = nnet_output.permute(1, 0, 2) # (N, T, C) -> (T, N, C) + mask = encoder_padding_mask(nnet_output.size(0), supervisions) + mask = mask.to(nnet_output.device) if mask is not None else None + mmodel = model.decoder.module if hasattr(model.decoder, "module") else model.decoder + + if params.method == "nbest-rescoring": + best_path_dict = rescore_with_n_best_list( + lattice=lattice, + G=G, + num_paths=params.num_paths, + lm_scale_list=lm_scale_list, + nbest_scale=params.nbest_scale, + ) + elif params.method == "whole-lattice-rescoring": + best_path_dict = rescore_with_whole_lattice( + lattice=lattice, + G_with_epsilon_loops=G, + lm_scale_list=lm_scale_list, + ) + elif params.method == "attention-decoder": + # lattice uses a 3-gram Lm. We rescore it with a 4-gram LM. + rescored_lattice = rescore_with_whole_lattice( + lattice=lattice, + G_with_epsilon_loops=G, + lm_scale_list=None, + ) + + best_path_dict = rescore_with_attention_decoder( + lattice=rescored_lattice, + num_paths=params.num_paths, + model=mmodel, + memory=nnet_output, + memory_key_padding_mask=mask, + sos_id=sos_id, + eos_id=eos_id, + nbest_scale=params.nbest_scale, + ) + elif params.method == "rnn-lm": + # lattice uses a 3-gram Lm. We rescore it with a 4-gram LM. + rescored_lattice = rescore_with_whole_lattice( + lattice=lattice, + G_with_epsilon_loops=G, + lm_scale_list=None, + ) + + best_path_dict = rescore_with_rnn_lm( + lattice=rescored_lattice, + num_paths=params.num_paths, + rnn_lm_model=rnn_lm_model, + model=mmodel, + memory=nnet_output, + memory_key_padding_mask=mask, + sos_id=sos_id, + eos_id=eos_id, + blank_id=0, + nbest_scale=params.nbest_scale, + ) + else: + assert False, f"Unsupported decoding method: {params.method}" + + ans = dict() + if best_path_dict is not None: + for lm_scale_str, best_path in best_path_dict.items(): + hyps = get_texts(best_path) + hyps = [[word_table[i] for i in ids] for ids in hyps] + ans[lm_scale_str] = hyps + else: + ans = None + return ans + + +def decode_dataset( + dl: torch.utils.data.DataLoader, + params: AttributeDict, + model: nn.Module, + rnn_lm_model: Optional[nn.Module], + HLG: Optional[k2.Fsa], + H: Optional[k2.Fsa], + bpe_model: Optional[spm.SentencePieceProcessor], + word_table: k2.SymbolTable, + sos_id: int, + eos_id: int, + G: Optional[k2.Fsa] = None, +) -> Dict[str, List[Tuple[str, List[str], List[str]]]]: + """Decode dataset. + + Args: + dl: + PyTorch's dataloader containing the dataset to decode. + params: + It is returned by :func:`get_params`. + model: + The neural model. + rnn_lm_model: + The neural model for RNN LM. + HLG: + The decoding graph. Used only when params.method is NOT ctc-decoding. + H: + The ctc topo. Used only when params.method is ctc-decoding. + bpe_model: + The BPE model. Used only when params.method is ctc-decoding. + word_table: + It is the word symbol table. + sos_id: + The token ID for SOS. + eos_id: + The token ID for EOS. + G: + An LM. It is not None when params.method is "nbest-rescoring" + or "whole-lattice-rescoring". In general, the G in HLG + is a 3-gram LM, while this G is a 4-gram LM. + Returns: + Return a dict, whose key may be "no-rescore" if no LM rescoring + is used, or it may be "lm_scale_0.7" if LM rescoring 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 = "?" + + results = defaultdict(list) + for batch_idx, batch in enumerate(dl): + texts = batch["supervisions"]["text"] + cut_ids = [cut.id for cut in batch["supervisions"]["cut"]] + + hyps_dict = decode_one_batch( + params=params, + model=model, + rnn_lm_model=rnn_lm_model, + HLG=HLG, + H=H, + bpe_model=bpe_model, + batch=batch, + word_table=word_table, + G=G, + sos_id=sos_id, + eos_id=eos_id, + ) + + if hyps_dict is not None: + for lm_scale, 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[lm_scale].extend(this_batch) + else: + assert len(results) > 0, "It should not decode to empty in the first batch!" + this_batch = [] + hyp_words = [] + for ref_text in texts: + ref_words = ref_text.split() + this_batch.append((ref_words, hyp_words)) + + for lm_scale in results.keys(): + results[lm_scale].extend(this_batch) + + num_cuts += len(texts) + + if batch_idx % 100 == 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[int], List[int]]]], +): + if params.method in ("attention-decoder", "rnn-lm"): + # Set it to False since there are too many logs. + enable_log = False + else: + enable_log = True + test_set_wers = dict() + for key, results in results_dict.items(): + recog_path = params.exp_dir / f"recogs-{test_set_name}-{key}.txt" + results = sorted(results) + store_transcripts(filename=recog_path, texts=results) + if enable_log: + 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.exp_dir / f"errs-{test_set_name}-{key}.txt" + with open(errs_filename, "w") as f: + wer = write_error_stats( + f, f"{test_set_name}-{key}", results, enable_log=enable_log + ) + test_set_wers[key] = wer + + if enable_log: + 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.exp_dir / f"wer-summary-{test_set_name}.txt" + with open(errs_info, "w") as f: + print("settings\tWER", file=f) + for key, val in test_set_wers: + print("{}\t{}".format(key, val), file=f) + + s = "\nFor {}, WER of different settings are:\n".format(test_set_name) + note = "\tbest for {}".format(test_set_name) + for key, val in test_set_wers: + s += "{}\t{}{}\n".format(key, val, note) + note = "" + logging.info(s) + + +@torch.no_grad() +def main(): + parser = get_parser() + LibriSpeechAsrDataModule.add_arguments(parser) + args = parser.parse_args() + args.exp_dir = Path(args.exp_dir) + args.lang_dir = Path(args.lang_dir) + args.lm_dir = Path(args.lm_dir) + + params = get_params() + params.update(vars(args)) + + setup_logger(f"{params.exp_dir}/log-{params.method}/log-decode") + logging.info("Decoding started") + logging.info(params) + + lexicon = Lexicon(params.lang_dir) + max_token_id = max(lexicon.tokens) + num_classes = max_token_id + 1 # +1 for the blank + params.vocab_size = num_classes + + device = torch.device("cpu") + if torch.cuda.is_available(): + device = torch.device("cuda", 0) + + logging.info(f"device: {device}") + + graph_compiler = BpeCtcTrainingGraphCompiler( + params.lang_dir, + device=device, + sos_token="", + eos_token="", + ) + sos_id = graph_compiler.sos_id + eos_id = graph_compiler.eos_id + + params.num_classes = num_classes + params.sos_id = sos_id + params.eos_id = eos_id + + if params.method == "ctc-decoding": + HLG = None + H = k2.ctc_topo( + max_token=max_token_id, + modified=False, + device=device, + ) + bpe_model = spm.SentencePieceProcessor() + bpe_model.load(str(params.lang_dir / "bpe.model")) + else: + H = None + bpe_model = None + HLG = k2.Fsa.from_dict( + torch.load(f"{params.lang_dir}/HLG.pt", map_location=device) + ) + assert HLG.requires_grad is False + + if not hasattr(HLG, "lm_scores"): + HLG.lm_scores = HLG.scores.clone() + + if params.method in ( + "nbest-rescoring", + "whole-lattice-rescoring", + "attention-decoder", + "rnn-lm", + ): + if not (params.lm_dir / "G_3_gram.pt").is_file(): + logging.info("Loading G_3_gram.fst.txt") + logging.warning("It may take 8 minutes.") + with open(params.lm_dir / "G_3_gram.fst.txt") as f: + #with open(params.lm_dir / "G_3_gram.fst.txt") as f: + first_word_disambig_id = lexicon.word_table["#0"] + + G = k2.Fsa.from_openfst(f.read(), acceptor=False) + # G.aux_labels is not needed in later computations, so + # remove it here. + del G.aux_labels + # CAUTION: The following line is crucial. + # Arcs entering the back-off state have label equal to #0. + # We have to change it to 0 here. + G.labels[G.labels >= first_word_disambig_id] = 0 + # See https://github.com/k2-fsa/k2/issues/874 + # for why we need to set G.properties to None + G.__dict__["_properties"] = None + G = k2.Fsa.from_fsas([G]).to(device) + G = k2.arc_sort(G) + # Save a dummy value so that it can be loaded in C++. + # See https://github.com/pytorch/pytorch/issues/67902 + # for why we need to do this. + G.dummy = 1 + + torch.save(G.as_dict(), params.lm_dir / "G_3_gram.pt") + else: + logging.info("Loading pre-compiled G_3_gram.pt") + d = torch.load(params.lm_dir / "G_3_gram.pt", map_location=device) + G = k2.Fsa.from_dict(d) + + if params.method in [ + "whole-lattice-rescoring", + "attention-decoder", + "rnn-lm", + ]: + # Add epsilon self-loops to G as we will compose + # it with the whole lattice later + G = k2.add_epsilon_self_loops(G) + G = k2.arc_sort(G) + G = G.to(device) + + # G.lm_scores is used to replace HLG.lm_scores during + # LM rescoring. + G.lm_scores = G.scores.clone() + else: + G = None + + logging.info("About to create model") + model = get_ctc_model(params) + assert model.encoder.decode_chunk_size == params.decode_chunk_len // 2, ( + model.encoder.decode_chunk_size, + params.decode_chunk_len, + ) + + 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() + num_param = sum([p.numel() for p in model.parameters()]) + logging.info(f"Number of model parameters: {num_param}") + + rnn_lm_model = None + if params.method == "rnn-lm": + rnn_lm_model = RnnLmModel( + vocab_size=params.num_classes, + embedding_dim=params.rnn_lm_embedding_dim, + hidden_dim=params.rnn_lm_hidden_dim, + num_layers=params.rnn_lm_num_layers, + tie_weights=params.rnn_lm_tie_weights, + ) + if params.rnn_lm_avg == 1: + load_checkpoint( + f"{params.rnn_lm_exp_dir}/epoch-{params.rnn_lm_epoch}.pt", + rnn_lm_model, + ) + rnn_lm_model.to(device) + else: + rnn_lm_model = load_averaged_model( + params.rnn_lm_exp_dir, + rnn_lm_model, + params.rnn_lm_epoch, + params.rnn_lm_avg, + device, + ) + rnn_lm_model.eval() + + # we need cut ids to display recognition results. + args.return_cuts = True + librispeech = LibriSpeechAsrDataModule(args) + + test_clean_cuts = librispeech.test_clean_cuts() + test_other_cuts = librispeech.test_other_cuts() + + test_clean_dl = librispeech.test_dataloaders(test_clean_cuts) + test_other_dl = librispeech.test_dataloaders(test_other_cuts) + + test_sets = ["test-clean", "test-other"] + test_dl = [test_clean_dl, test_other_dl] + + for test_set, test_dl in zip(test_sets, test_dl): + results_dict = decode_dataset( + dl=test_dl, + params=params, + model=model, + rnn_lm_model=rnn_lm_model, + HLG=HLG, + H=H, + bpe_model=bpe_model, + word_table=lexicon.word_table, + G=G, + sos_id=sos_id, + eos_id=eos_id, + ) + + save_results(params=params, test_set_name=test_set, results_dict=results_dict) + + logging.info("Done!") + + +if __name__ == "__main__": + main() diff --git a/egs/librispeech/ASR/zipformer_ctc_streaming/decode_G_4gram.py b/egs/librispeech/ASR/zipformer_ctc_streaming/decode_G_4gram.py new file mode 100755 index 000000000..5fd212fd7 --- /dev/null +++ b/egs/librispeech/ASR/zipformer_ctc_streaming/decode_G_4gram.py @@ -0,0 +1,887 @@ +#!/usr/bin/env python3 +# Copyright 2021 Xiaomi Corporation (Author: Liyong Guo, Fangjun Kuang) +# +# See ../../../../LICENSE for clarification regarding multiple authors +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + + +import argparse +import logging +from 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 train import add_model_arguments, get_ctc_model, get_params +from transformer import encoder_padding_mask + +from icefall.bpe_graph_compiler import BpeCtcTrainingGraphCompiler +from icefall.checkpoint import ( + average_checkpoints, + average_checkpoints_with_averaged_model, + find_checkpoints, + load_checkpoint, +) +from icefall.decode import ( + get_lattice, + nbest_decoding, + nbest_oracle, + one_best_decoding, + rescore_with_attention_decoder, + rescore_with_n_best_list, + rescore_with_rnn_lm, + rescore_with_whole_lattice, +) +from icefall.lexicon import Lexicon +from icefall.rnn_lm.model import RnnLmModel +from icefall.utils import ( + AttributeDict, + get_texts, + load_averaged_model, + 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=77, + help="It specifies the checkpoint to use for decoding." + "Note: Epoch counts from 0.", + ) + + parser.add_argument( + "--iter", + type=int, + default=0, + help="""If positive, --epoch is ignored and it + will use the checkpoint exp_dir/checkpoint-iter.pt. + You can specify --avg to use more checkpoints for model averaging. + """, + ) + + parser.add_argument( + "--avg", + type=int, + default=55, + help="Number of checkpoints to average. Automatically select " + "consecutive checkpoints before the checkpoint specified by " + "'--epoch'. ", + ) + + parser.add_argument( + "--use-averaged-model", + type=str2bool, + default=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( + "--method", + type=str, + default="attention-decoder", + help="""Decoding method. + Supported values are: + - (0) ctc-decoding. Use CTC decoding. It uses a sentence piece + model, i.e., lang_dir/bpe.model, to convert word pieces to words. + It needs neither a lexicon nor an n-gram LM. + - (1) 1best. Extract the best path from the decoding lattice as the + decoding result. + - (2) nbest. Extract n paths from the decoding lattice; the path + with the highest score is the decoding result. + - (3) nbest-rescoring. Extract n paths from the decoding lattice, + rescore them with an n-gram LM (e.g., a 4-gram LM), the path with + the highest score is the decoding result. + - (4) whole-lattice-rescoring. Rescore the decoding lattice with an + n-gram LM (e.g., a 4-gram LM), the best path of rescored lattice + is the decoding result. + - (5) attention-decoder. Extract n paths from the LM rescored + lattice, the path with the highest score is the decoding result. + - (6) rnn-lm. Rescoring with attention-decoder and RNN LM. We assume + you have trained an RNN LM using ./rnn_lm/train.py + - (7) nbest-oracle. Its WER is the lower bound of any n-best + rescoring method can achieve. Useful for debugging n-best + rescoring method. + """, + ) + + parser.add_argument( + "--num-paths", + type=int, + default=100, + help="""Number of paths for n-best based decoding method. + Used only when "method" is one of the following values: + nbest, nbest-rescoring, attention-decoder, rnn-lm, and nbest-oracle + """, + ) + + parser.add_argument( + "--nbest-scale", + type=float, + default=0.5, + help="""The scale to be applied to `lattice.scores`. + It's needed if you use any kinds of n-best based rescoring. + Used only when "method" is one of the following values: + nbest, nbest-rescoring, attention-decoder, rnn-lm, and nbest-oracle + A smaller value results in more unique paths. + """, + ) + + parser.add_argument( + "--exp-dir", + type=str, + default="zipformer_ctc/exp", + help="The experiment dir", + ) + + parser.add_argument( + "--lang-dir", + type=str, + default="data/lang_bpe_500", + help="The lang dir", + ) + + parser.add_argument( + "--lm-dir", + type=str, + default="data/lm", + help="""The n-gram LM dir. + It should contain either G_4_gram.pt or G_4_gram.fst.txt + """, + ) + + parser.add_argument( + "--rnn-lm-exp-dir", + type=str, + default="rnn_lm/exp", + help="""Used only when --method is rnn-lm. + It specifies the path to RNN LM exp dir. + """, + ) + + parser.add_argument( + "--rnn-lm-epoch", + type=int, + default=7, + help="""Used only when --method is rnn-lm. + It specifies the checkpoint to use. + """, + ) + + parser.add_argument( + "--rnn-lm-avg", + type=int, + default=2, + help="""Used only when --method is rnn-lm. + It specifies the number of checkpoints to average. + """, + ) + + parser.add_argument( + "--rnn-lm-embedding-dim", + type=int, + default=2048, + help="Embedding dim of the model", + ) + + parser.add_argument( + "--rnn-lm-hidden-dim", + type=int, + default=2048, + help="Hidden dim of the model", + ) + + parser.add_argument( + "--rnn-lm-num-layers", + type=int, + default=4, + help="Number of RNN layers the model", + ) + parser.add_argument( + "--rnn-lm-tie-weights", + type=str2bool, + default=False, + help="""True to share the weights between the input embedding layer and the + last output linear layer + """, + ) + + add_model_arguments(parser) + + return parser + + +def decode_one_batch( + params: AttributeDict, + model: nn.Module, + rnn_lm_model: Optional[nn.Module], + HLG: Optional[k2.Fsa], + H: Optional[k2.Fsa], + bpe_model: Optional[spm.SentencePieceProcessor], + batch: dict, + word_table: k2.SymbolTable, + sos_id: int, + eos_id: int, + G: Optional[k2.Fsa] = None, +) -> Dict[str, List[List[str]]]: + """Decode one batch and return the result in a dict. The dict has the + following format: + + - key: It indicates the setting used for decoding. For example, + if no rescoring is used, the key is the string `no_rescore`. + If LM rescoring is used, the key is the string `lm_scale_xxx`, + where `xxx` is the value of `lm_scale`. An example key is + `lm_scale_0.7` + - 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`. + + - params.method is "1best", it uses 1best decoding without LM rescoring. + - params.method is "nbest", it uses nbest decoding without LM rescoring. + - params.method is "nbest-rescoring", it uses nbest LM rescoring. + - params.method is "whole-lattice-rescoring", it uses whole lattice LM + rescoring. + + model: + The neural model. + rnn_lm_model: + The neural model for RNN LM. + HLG: + The decoding graph. Used only when params.method is NOT ctc-decoding. + H: + The ctc topo. Used only when params.method is ctc-decoding. + bpe_model: + The BPE model. Used only when params.method is ctc-decoding. + 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. + sos_id: + The token ID of the SOS. + eos_id: + The token ID of the EOS. + G: + An LM. It is not None when params.method is "nbest-rescoring" + or "whole-lattice-rescoring". In general, the G in HLG + is a 3-gram LM, while this G is a 4-gram LM. + Returns: + Return the decoding result. See above description for the format of + the returned dict. Note: If it decodes to nothing, then return None. + """ + if HLG is not None: + device = HLG.device + else: + device = H.device + feature = batch["inputs"] + assert feature.ndim == 3 + feature = feature.to(device) + # at entry, feature is (N, T, C) + + supervisions = batch["supervisions"] + feature_lens = supervisions["num_frames"].to(device) + + nnet_output, _ = model.encoder(feature, feature_lens) + ctc_output = model.ctc_output(nnet_output) + # nnet_output is (N, T, C) + + supervision_segments = torch.stack( + ( + supervisions["sequence_idx"], + supervisions["start_frame"] // params.subsampling_factor, + supervisions["num_frames"] // params.subsampling_factor, + ), + 1, + ).to(torch.int32) + + if H is None: + assert HLG is not None + decoding_graph = HLG + else: + assert HLG is None + assert bpe_model is not None + decoding_graph = H + + lattice = get_lattice( + nnet_output=ctc_output, + decoding_graph=decoding_graph, + supervision_segments=supervision_segments, + search_beam=params.search_beam, + output_beam=params.output_beam, + min_active_states=params.min_active_states, + max_active_states=params.max_active_states, + subsampling_factor=params.subsampling_factor, + ) + + if params.method == "ctc-decoding": + best_path = one_best_decoding( + lattice=lattice, use_double_scores=params.use_double_scores + ) + # Note: `best_path.aux_labels` contains token IDs, not word IDs + # since we are using H, not HLG here. + # + # token_ids is a lit-of-list of IDs + token_ids = get_texts(best_path) + + # hyps is a list of str, e.g., ['xxx yyy zzz', ...] + hyps = bpe_model.decode(token_ids) + + # hyps is a list of list of str, e.g., [['xxx', 'yyy', 'zzz'], ... ] + hyps = [s.split() for s in hyps] + key = "ctc-decoding" + return {key: hyps} + + if params.method == "nbest-oracle": + # Note: You can also pass rescored lattices to it. + # We choose the HLG decoded lattice for speed reasons + # as HLG decoding is faster and the oracle WER + # is only slightly worse than that of rescored lattices. + best_path = nbest_oracle( + lattice=lattice, + num_paths=params.num_paths, + ref_texts=supervisions["text"], + word_table=word_table, + nbest_scale=params.nbest_scale, + oov="", + ) + hyps = get_texts(best_path) + hyps = [[word_table[i] for i in ids] for ids in hyps] + key = f"oracle_{params.num_paths}_nbest_scale_{params.nbest_scale}" # noqa + return {key: hyps} + + if params.method in ["1best", "nbest"]: + if params.method == "1best": + best_path = one_best_decoding( + lattice=lattice, use_double_scores=params.use_double_scores + ) + key = "no_rescore" + else: + best_path = nbest_decoding( + lattice=lattice, + num_paths=params.num_paths, + use_double_scores=params.use_double_scores, + nbest_scale=params.nbest_scale, + ) + key = f"no_rescore-nbest-scale-{params.nbest_scale}-{params.num_paths}" # noqa + + hyps = get_texts(best_path) + hyps = [[word_table[i] for i in ids] for ids in hyps] + return {key: hyps} + + assert params.method in [ + "nbest-rescoring", + "whole-lattice-rescoring", + "attention-decoder", + "rnn-lm", + ] + + lm_scale_list = [0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7] + + nnet_output = nnet_output.permute(1, 0, 2) # (N, T, C) -> (T, N, C) + mask = encoder_padding_mask(nnet_output.size(0), supervisions) + mask = mask.to(nnet_output.device) if mask is not None else None + mmodel = model.decoder.module if hasattr(model.decoder, "module") else model.decoder + + if params.method == "nbest-rescoring": + best_path_dict = rescore_with_n_best_list( + lattice=lattice, + G=G, + num_paths=params.num_paths, + lm_scale_list=lm_scale_list, + nbest_scale=params.nbest_scale, + ) + elif params.method == "whole-lattice-rescoring": + best_path_dict = rescore_with_whole_lattice( + lattice=lattice, + G_with_epsilon_loops=G, + lm_scale_list=lm_scale_list, + ) + elif params.method == "attention-decoder": + # lattice uses a 3-gram Lm. We rescore it with a 4-gram LM. + rescored_lattice = rescore_with_whole_lattice( + lattice=lattice, + G_with_epsilon_loops=G, + lm_scale_list=None, + ) + + best_path_dict = rescore_with_attention_decoder( + lattice=rescored_lattice, + num_paths=params.num_paths, + model=mmodel, + memory=nnet_output, + memory_key_padding_mask=mask, + sos_id=sos_id, + eos_id=eos_id, + nbest_scale=params.nbest_scale, + ) + elif params.method == "rnn-lm": + # lattice uses a 3-gram Lm. We rescore it with a 4-gram LM. + rescored_lattice = rescore_with_whole_lattice( + lattice=lattice, + G_with_epsilon_loops=G, + lm_scale_list=None, + ) + + best_path_dict = rescore_with_rnn_lm( + lattice=rescored_lattice, + num_paths=params.num_paths, + rnn_lm_model=rnn_lm_model, + model=mmodel, + memory=nnet_output, + memory_key_padding_mask=mask, + sos_id=sos_id, + eos_id=eos_id, + blank_id=0, + nbest_scale=params.nbest_scale, + ) + else: + assert False, f"Unsupported decoding method: {params.method}" + + ans = dict() + if best_path_dict is not None: + for lm_scale_str, best_path in best_path_dict.items(): + hyps = get_texts(best_path) + hyps = [[word_table[i] for i in ids] for ids in hyps] + ans[lm_scale_str] = hyps + else: + ans = None + return ans + + +def decode_dataset( + dl: torch.utils.data.DataLoader, + params: AttributeDict, + model: nn.Module, + rnn_lm_model: Optional[nn.Module], + HLG: Optional[k2.Fsa], + H: Optional[k2.Fsa], + bpe_model: Optional[spm.SentencePieceProcessor], + word_table: k2.SymbolTable, + sos_id: int, + eos_id: int, + G: Optional[k2.Fsa] = None, +) -> Dict[str, List[Tuple[str, List[str], List[str]]]]: + """Decode dataset. + + Args: + dl: + PyTorch's dataloader containing the dataset to decode. + params: + It is returned by :func:`get_params`. + model: + The neural model. + rnn_lm_model: + The neural model for RNN LM. + HLG: + The decoding graph. Used only when params.method is NOT ctc-decoding. + H: + The ctc topo. Used only when params.method is ctc-decoding. + bpe_model: + The BPE model. Used only when params.method is ctc-decoding. + word_table: + It is the word symbol table. + sos_id: + The token ID for SOS. + eos_id: + The token ID for EOS. + G: + An LM. It is not None when params.method is "nbest-rescoring" + or "whole-lattice-rescoring". In general, the G in HLG + is a 3-gram LM, while this G is a 4-gram LM. + Returns: + Return a dict, whose key may be "no-rescore" if no LM rescoring + is used, or it may be "lm_scale_0.7" if LM rescoring 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 = "?" + + results = defaultdict(list) + for batch_idx, batch in enumerate(dl): + texts = batch["supervisions"]["text"] + cut_ids = [cut.id for cut in batch["supervisions"]["cut"]] + + hyps_dict = decode_one_batch( + params=params, + model=model, + rnn_lm_model=rnn_lm_model, + HLG=HLG, + H=H, + bpe_model=bpe_model, + batch=batch, + word_table=word_table, + G=G, + sos_id=sos_id, + eos_id=eos_id, + ) + + if hyps_dict is not None: + for lm_scale, 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[lm_scale].extend(this_batch) + else: + assert len(results) > 0, "It should not decode to empty in the first batch!" + this_batch = [] + hyp_words = [] + for ref_text in texts: + ref_words = ref_text.split() + this_batch.append((ref_words, hyp_words)) + + for lm_scale in results.keys(): + results[lm_scale].extend(this_batch) + + num_cuts += len(texts) + + if batch_idx % 100 == 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[int], List[int]]]], +): + if params.method in ("attention-decoder", "rnn-lm"): + # Set it to False since there are too many logs. + enable_log = False + else: + enable_log = True + test_set_wers = dict() + for key, results in results_dict.items(): + recog_path = params.exp_dir / f"recogs-{test_set_name}-{key}.txt" + results = sorted(results) + store_transcripts(filename=recog_path, texts=results) + if enable_log: + 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.exp_dir / f"errs-{test_set_name}-{key}.txt" + with open(errs_filename, "w") as f: + wer = write_error_stats( + f, f"{test_set_name}-{key}", results, enable_log=enable_log + ) + test_set_wers[key] = wer + + if enable_log: + 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.exp_dir / f"wer-summary-{test_set_name}.txt" + with open(errs_info, "w") as f: + print("settings\tWER", file=f) + for key, val in test_set_wers: + print("{}\t{}".format(key, val), file=f) + + s = "\nFor {}, WER of different settings are:\n".format(test_set_name) + note = "\tbest for {}".format(test_set_name) + for key, val in test_set_wers: + s += "{}\t{}{}\n".format(key, val, note) + note = "" + logging.info(s) + + +@torch.no_grad() +def main(): + parser = get_parser() + LibriSpeechAsrDataModule.add_arguments(parser) + args = parser.parse_args() + args.exp_dir = Path(args.exp_dir) + args.lang_dir = Path(args.lang_dir) + args.lm_dir = Path(args.lm_dir) + + params = get_params() + params.update(vars(args)) + + setup_logger(f"{params.exp_dir}/log-{params.method}/log-decode") + logging.info("Decoding started") + logging.info(params) + + lexicon = Lexicon(params.lang_dir) + max_token_id = max(lexicon.tokens) + num_classes = max_token_id + 1 # +1 for the blank + params.vocab_size = num_classes + + device = torch.device("cpu") + if torch.cuda.is_available(): + device = torch.device("cuda", 0) + + logging.info(f"device: {device}") + + graph_compiler = BpeCtcTrainingGraphCompiler( + params.lang_dir, + device=device, + sos_token="", + eos_token="", + ) + sos_id = graph_compiler.sos_id + eos_id = graph_compiler.eos_id + + params.num_classes = num_classes + params.sos_id = sos_id + params.eos_id = eos_id + + if params.method == "ctc-decoding": + HLG = None + H = k2.ctc_topo( + max_token=max_token_id, + modified=False, + device=device, + ) + bpe_model = spm.SentencePieceProcessor() + bpe_model.load(str(params.lang_dir / "bpe.model")) + else: + H = None + bpe_model = None + HLG = k2.Fsa.from_dict( + torch.load(f"{params.lang_dir}/HLG.pt", map_location=device) + ) + assert HLG.requires_grad is False + + if not hasattr(HLG, "lm_scores"): + HLG.lm_scores = HLG.scores.clone() + + if params.method in ( + "nbest-rescoring", + "whole-lattice-rescoring", + "attention-decoder", + "rnn-lm", + ): + if not (params.lm_dir / "G_4_gram.pt").is_file(): + logging.info("Loading G_4_gram.fst.txt") + logging.warning("It may take 8 minutes.") + with open(params.lm_dir / "G_4_gram.fst.txt") as f: + #with open(params.lm_dir / "G_3_gram.fst.txt") as f: + first_word_disambig_id = lexicon.word_table["#0"] + + G = k2.Fsa.from_openfst(f.read(), acceptor=False) + # G.aux_labels is not needed in later computations, so + # remove it here. + del G.aux_labels + # CAUTION: The following line is crucial. + # Arcs entering the back-off state have label equal to #0. + # We have to change it to 0 here. + G.labels[G.labels >= first_word_disambig_id] = 0 + # See https://github.com/k2-fsa/k2/issues/874 + # for why we need to set G.properties to None + G.__dict__["_properties"] = None + G = k2.Fsa.from_fsas([G]).to(device) + G = k2.arc_sort(G) + # Save a dummy value so that it can be loaded in C++. + # See https://github.com/pytorch/pytorch/issues/67902 + # for why we need to do this. + G.dummy = 1 + + torch.save(G.as_dict(), params.lm_dir / "G_4_gram.pt") + else: + logging.info("Loading pre-compiled G_4_gram.pt") + d = torch.load(params.lm_dir / "G_4_gram.pt", map_location=device) + G = k2.Fsa.from_dict(d) + + if params.method in [ + "whole-lattice-rescoring", + "attention-decoder", + "rnn-lm", + ]: + # Add epsilon self-loops to G as we will compose + # it with the whole lattice later + G = k2.add_epsilon_self_loops(G) + G = k2.arc_sort(G) + G = G.to(device) + + # G.lm_scores is used to replace HLG.lm_scores during + # LM rescoring. + G.lm_scores = G.scores.clone() + else: + G = None + + logging.info("About to create model") + model = get_ctc_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() + num_param = sum([p.numel() for p in model.parameters()]) + logging.info(f"Number of model parameters: {num_param}") + + rnn_lm_model = None + if params.method == "rnn-lm": + rnn_lm_model = RnnLmModel( + vocab_size=params.num_classes, + embedding_dim=params.rnn_lm_embedding_dim, + hidden_dim=params.rnn_lm_hidden_dim, + num_layers=params.rnn_lm_num_layers, + tie_weights=params.rnn_lm_tie_weights, + ) + if params.rnn_lm_avg == 1: + load_checkpoint( + f"{params.rnn_lm_exp_dir}/epoch-{params.rnn_lm_epoch}.pt", + rnn_lm_model, + ) + rnn_lm_model.to(device) + else: + rnn_lm_model = load_averaged_model( + params.rnn_lm_exp_dir, + rnn_lm_model, + params.rnn_lm_epoch, + params.rnn_lm_avg, + device, + ) + rnn_lm_model.eval() + + # we need cut ids to display recognition results. + args.return_cuts = True + librispeech = LibriSpeechAsrDataModule(args) + + test_clean_cuts = librispeech.test_clean_cuts() + test_other_cuts = librispeech.test_other_cuts() + + test_clean_dl = librispeech.test_dataloaders(test_clean_cuts) + test_other_dl = librispeech.test_dataloaders(test_other_cuts) + + test_sets = ["test-clean", "test-other"] + test_dl = [test_clean_dl, test_other_dl] + + for test_set, test_dl in zip(test_sets, test_dl): + results_dict = decode_dataset( + dl=test_dl, + params=params, + model=model, + rnn_lm_model=rnn_lm_model, + HLG=HLG, + H=H, + bpe_model=bpe_model, + word_table=lexicon.word_table, + G=G, + sos_id=sos_id, + eos_id=eos_id, + ) + + save_results(params=params, test_set_name=test_set, results_dict=results_dict) + + logging.info("Done!") + + +if __name__ == "__main__": + main() diff --git a/egs/librispeech/ASR/zipformer_ctc_streaming/decoder.py b/egs/librispeech/ASR/zipformer_ctc_streaming/decoder.py new file mode 100644 index 000000000..8dec048a1 --- /dev/null +++ b/egs/librispeech/ASR/zipformer_ctc_streaming/decoder.py @@ -0,0 +1,298 @@ +# 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. + +from typing import List + +import torch +import torch.nn as nn +import torch.nn.functional as F +from label_smoothing import LabelSmoothingLoss +from torch.nn.utils.rnn import pad_sequence +from transformer import PositionalEncoding, TransformerDecoderLayer + + +class Decoder(nn.Module): + """This class implements Transformer based decoder for an attention-based encoder-decoder + model. + """ + + def __init__( + self, + num_layers: int, + num_classes: int, + d_model: int = 256, + nhead: int = 4, + dim_feedforward: int = 2048, + dropout: float = 0.1, + normalize_before: bool = True, + ): + """ + Args: + num_layers: + Number of layers. + num_classes: + Number of tokens of the modeling unit including blank. + d_model: + Dimension of the input embedding, and of the decoder output. + """ + super().__init__() + + if num_layers > 0: + self.decoder_num_class = num_classes # bpe model already has sos/eos symbol + + self.decoder_embed = nn.Embedding( + num_embeddings=self.decoder_num_class, embedding_dim=d_model + ) + self.decoder_pos = PositionalEncoding(d_model, dropout) + + decoder_layer = TransformerDecoderLayer( + d_model=d_model, + nhead=nhead, + dim_feedforward=dim_feedforward, + dropout=dropout, + normalize_before=normalize_before, + ) + + if normalize_before: + decoder_norm = nn.LayerNorm(d_model) + else: + decoder_norm = None + + self.decoder = nn.TransformerDecoder( + decoder_layer=decoder_layer, + num_layers=num_layers, + norm=decoder_norm, + ) + + self.decoder_output_layer = torch.nn.Linear(d_model, self.decoder_num_class) + self.decoder_criterion = LabelSmoothingLoss() + else: + self.decoder_criterion = None + + @torch.jit.export + def forward( + self, + memory: torch.Tensor, + memory_key_padding_mask: torch.Tensor, + token_ids: List[List[int]], + sos_id: int, + eos_id: int, + ) -> torch.Tensor: + """ + Args: + memory: + It's the output of the encoder with shape (T, N, C) + memory_key_padding_mask: + The padding mask from the encoder. + token_ids: + A list-of-list IDs. Each sublist contains IDs for an utterance. + The IDs can be either phone IDs or word piece IDs. + sos_id: + sos token id + eos_id: + eos token id + Returns: + A scalar, the **sum** of label smoothing loss over utterances + in the batch without any normalization. + """ + ys_in = add_sos(token_ids, sos_id=sos_id) + ys_in = [torch.tensor(y) for y in ys_in] + ys_in_pad = pad_sequence(ys_in, batch_first=True, padding_value=float(eos_id)) + + ys_out = add_eos(token_ids, eos_id=eos_id) + ys_out = [torch.tensor(y) for y in ys_out] + ys_out_pad = pad_sequence(ys_out, batch_first=True, padding_value=float(-1)) + + device = memory.device + ys_in_pad = ys_in_pad.to(device) + ys_out_pad = ys_out_pad.to(device) + + tgt_mask = generate_square_subsequent_mask(ys_in_pad.shape[-1]).to(device) + + tgt_key_padding_mask = decoder_padding_mask(ys_in_pad, ignore_id=eos_id) + # TODO: Use length information to create the decoder padding mask + # We set the first column to False since the first column in ys_in_pad + # contains sos_id, which is the same as eos_id in our current setting. + tgt_key_padding_mask[:, 0] = False + + tgt = self.decoder_embed(ys_in_pad) # (N, T) -> (N, T, C) + tgt = self.decoder_pos(tgt) + tgt = tgt.permute(1, 0, 2) # (N, T, C) -> (T, N, C) + pred_pad = self.decoder( + tgt=tgt, + memory=memory, + tgt_mask=tgt_mask, + tgt_key_padding_mask=tgt_key_padding_mask, + memory_key_padding_mask=memory_key_padding_mask, + ) # (T, N, C) + pred_pad = pred_pad.permute(1, 0, 2) # (T, N, C) -> (N, T, C) + pred_pad = self.decoder_output_layer(pred_pad) # (N, T, C) + + decoder_loss = self.decoder_criterion(pred_pad, ys_out_pad) + + return decoder_loss + + @torch.jit.export + def decoder_nll( + self, + memory: torch.Tensor, + memory_key_padding_mask: torch.Tensor, + token_ids: List[torch.Tensor], + sos_id: int, + eos_id: int, + ) -> torch.Tensor: + """ + Args: + memory: + It's the output of the encoder with shape (T, N, C) + memory_key_padding_mask: + The padding mask from the encoder. + token_ids: + A list-of-list IDs (e.g., word piece IDs). + Each sublist represents an utterance. + sos_id: + The token ID for SOS. + eos_id: + The token ID for EOS. + Returns: + A 2-D tensor of shape (len(token_ids), max_token_length) + representing the cross entropy loss (i.e., negative log-likelihood). + """ + # The common part between this function and decoder_forward could be + # extracted as a separate function. + if isinstance(token_ids[0], torch.Tensor): + # This branch is executed by torchscript in C++. + # See https://github.com/k2-fsa/k2/pull/870 + # https://github.com/k2-fsa/k2/blob/3c1c18400060415b141ccea0115fd4bf0ad6234e/k2/torch/bin/attention_rescore.cu#L286 + token_ids = [tolist(t) for t in token_ids] + + ys_in = add_sos(token_ids, sos_id=sos_id) + ys_in = [torch.tensor(y) for y in ys_in] + ys_in_pad = pad_sequence(ys_in, batch_first=True, padding_value=float(eos_id)) + + ys_out = add_eos(token_ids, eos_id=eos_id) + ys_out = [torch.tensor(y) for y in ys_out] + ys_out_pad = pad_sequence(ys_out, batch_first=True, padding_value=float(-1)) + + device = memory.device + ys_in_pad = ys_in_pad.to(device, dtype=torch.int64) + ys_out_pad = ys_out_pad.to(device, dtype=torch.int64) + + tgt_mask = generate_square_subsequent_mask(ys_in_pad.shape[-1]).to(device) + + tgt_key_padding_mask = decoder_padding_mask(ys_in_pad, ignore_id=eos_id) + # TODO: Use length information to create the decoder padding mask + # We set the first column to False since the first column in ys_in_pad + # contains sos_id, which is the same as eos_id in our current setting. + tgt_key_padding_mask[:, 0] = False + + tgt = self.decoder_embed(ys_in_pad) # (B, T) -> (B, T, F) + tgt = self.decoder_pos(tgt) + tgt = tgt.permute(1, 0, 2) # (B, T, F) -> (T, B, F) + pred_pad = self.decoder( + tgt=tgt, + memory=memory, + tgt_mask=tgt_mask, + tgt_key_padding_mask=tgt_key_padding_mask, + memory_key_padding_mask=memory_key_padding_mask, + ) # (T, B, F) + pred_pad = pred_pad.permute(1, 0, 2) # (T, B, F) -> (B, T, F) + pred_pad = self.decoder_output_layer(pred_pad) # (B, T, F) + # nll: negative log-likelihood + nll = torch.nn.functional.cross_entropy( + pred_pad.view(-1, self.decoder_num_class), + ys_out_pad.view(-1), + ignore_index=-1, + reduction="none", + ) + + nll = nll.view(pred_pad.shape[0], -1) + + return nll + + +def add_sos(token_ids: List[List[int]], sos_id: int) -> List[List[int]]: + """Prepend sos_id to each utterance. + Args: + token_ids: + A list-of-list of token IDs. Each sublist contains + token IDs (e.g., word piece IDs) of an utterance. + sos_id: + The ID of the SOS token. + Return: + Return a new list-of-list, where each sublist starts + with SOS ID. + """ + return [[sos_id] + utt for utt in token_ids] + + +def add_eos(token_ids: List[List[int]], eos_id: int) -> List[List[int]]: + """Append eos_id to each utterance. + Args: + token_ids: + A list-of-list of token IDs. Each sublist contains + token IDs (e.g., word piece IDs) of an utterance. + eos_id: + The ID of the EOS token. + Return: + Return a new list-of-list, where each sublist ends + with EOS ID. + """ + return [utt + [eos_id] for utt in token_ids] + + +def decoder_padding_mask(ys_pad: torch.Tensor, ignore_id: int = -1) -> torch.Tensor: + """Generate a length mask for input. + The masked position are filled with True, + Unmasked positions are filled with False. + Args: + ys_pad: + padded tensor of dimension (batch_size, input_length). + ignore_id: + the ignored number (the padding number) in ys_pad + Returns: + Tensor: + a bool tensor of the same shape as the input tensor. + """ + ys_mask = ys_pad == ignore_id + return ys_mask + + +def generate_square_subsequent_mask(sz: int) -> torch.Tensor: + """Generate a square mask for the sequence. The masked positions are + filled with float('-inf'). Unmasked positions are filled with float(0.0). + The mask can be used for masked self-attention. + For instance, if sz is 3, it returns:: + tensor([[0., -inf, -inf], + [0., 0., -inf], + [0., 0., 0]]) + Args: + sz: mask size + Returns: + A square mask of dimension (sz, sz) + """ + mask = (torch.triu(torch.ones(sz, sz)) == 1).transpose(0, 1) + mask = ( + mask.float() + .masked_fill(mask == 0, float("-inf")) + .masked_fill(mask == 1, float(0.0)) + ) + return mask + + +def tolist(t: torch.Tensor) -> List[int]: + """Used by jit""" + return torch.jit.annotate(List[int], t.tolist()) diff --git a/egs/librispeech/ASR/zipformer_ctc_streaming/encoder_interface.py b/egs/librispeech/ASR/zipformer_ctc_streaming/encoder_interface.py new file mode 100644 index 000000000..257facce4 --- /dev/null +++ b/egs/librispeech/ASR/zipformer_ctc_streaming/encoder_interface.py @@ -0,0 +1,43 @@ +# 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. + +from typing import Tuple + +import torch +import torch.nn as nn + + +class EncoderInterface(nn.Module): + def forward( + self, x: torch.Tensor, x_lens: torch.Tensor + ) -> Tuple[torch.Tensor, torch.Tensor]: + """ + Args: + x: + A tensor of shape (batch_size, input_seq_len, num_features) + containing the input features. + x_lens: + A tensor of shape (batch_size,) containing the number of frames + in `x` before padding. + Returns: + Return a tuple containing two tensors: + - encoder_out, a tensor of (batch_size, out_seq_len, output_dim) + containing unnormalized probabilities, i.e., the output of a + linear layer. + - encoder_out_lens, a tensor of shape (batch_size,) containing + the number of frames in `encoder_out` before padding. + """ + raise NotImplementedError("Please implement it in a subclass") diff --git a/egs/librispeech/ASR/zipformer_ctc_streaming/export.py b/egs/librispeech/ASR/zipformer_ctc_streaming/export.py new file mode 100755 index 000000000..0ff50f128 --- /dev/null +++ b/egs/librispeech/ASR/zipformer_ctc_streaming/export.py @@ -0,0 +1,240 @@ +#!/usr/bin/env python3 +# +# Copyright 2021 Xiaomi Corporation (Author: Fangjun Kuang) +# +# See ../../../../LICENSE for clarification regarding multiple authors +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +# This script converts several saved checkpoints +# to a single one using model averaging. + +import argparse +import logging +from pathlib import Path + +import torch +from scaling_converter import convert_scaled_to_non_scaled +from train import add_model_arguments, get_ctc_model, get_params + +from icefall.checkpoint import ( + average_checkpoints, + average_checkpoints_with_averaged_model, + find_checkpoints, + load_checkpoint, +) +from icefall.lexicon import Lexicon +from icefall.utils import str2bool + + +def get_parser(): + parser = argparse.ArgumentParser( + formatter_class=argparse.ArgumentDefaultsHelpFormatter + ) + + parser.add_argument( + "--epoch", + type=int, + default=30, + help="It specifies the checkpoint to use for decoding." + "Note: Epoch counts from 0.", + ) + + parser.add_argument( + "--iter", + type=int, + default=0, + help="""If positive, --epoch is ignored and it + will use the checkpoint exp_dir/checkpoint-iter.pt. + You can specify --avg to use more checkpoints for model averaging. + """, + ) + + parser.add_argument( + "--avg", + type=int, + default=9, + help="Number of checkpoints to average. Automatically select " + "consecutive checkpoints before the checkpoint specified by " + "'--epoch'. ", + ) + + 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_ctc/exp", + help="""It specifies the directory where all training related + files, e.g., checkpoints, log, etc, are saved + """, + ) + + parser.add_argument( + "--lang-dir", + type=str, + default="data/lang_bpe_500", + help="""It contains language related input files such as "lexicon.txt" + """, + ) + + parser.add_argument( + "--jit", + type=str2bool, + default=True, + help="""True to save a model after applying torch.jit.script. + """, + ) + + add_model_arguments(parser) + + return parser + + +def main(): + args = get_parser().parse_args() + args.exp_dir = Path(args.exp_dir) + args.lang_dir = Path(args.lang_dir) + + params = get_params() + params.update(vars(args)) + + logging.info(params) + + lexicon = Lexicon(params.lang_dir) + max_token_id = max(lexicon.tokens) + num_classes = max_token_id + 1 # +1 for the blank + params.vocab_size = num_classes + + device = torch.device("cpu") + if torch.cuda.is_available(): + device = torch.device("cuda", 0) + + logging.info(f"device: {device}") + + model = get_ctc_model(params) + + if not params.use_averaged_model: + if params.iter > 0: + filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[ + : params.avg + ] + if len(filenames) == 0: + raise ValueError( + f"No checkpoints found for" + f" --iter {params.iter}, --avg {params.avg}" + ) + elif len(filenames) < params.avg: + raise ValueError( + f"Not enough checkpoints ({len(filenames)}) found for" + f" --iter {params.iter}, --avg {params.avg}" + ) + logging.info(f"averaging {filenames}") + model.to(device) + model.load_state_dict(average_checkpoints(filenames, device=device)) + elif params.avg == 1: + load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model) + else: + start = params.epoch - params.avg + 1 + filenames = [] + for i in range(start, params.epoch + 1): + if i >= 1: + filenames.append(f"{params.exp_dir}/epoch-{i}.pt") + logging.info(f"averaging {filenames}") + model.to(device) + model.load_state_dict(average_checkpoints(filenames, device=device)) + else: + if params.iter > 0: + filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[ + : params.avg + 1 + ] + if len(filenames) == 0: + raise ValueError( + f"No checkpoints found for" + f" --iter {params.iter}, --avg {params.avg}" + ) + elif len(filenames) < params.avg + 1: + raise ValueError( + f"Not enough checkpoints ({len(filenames)}) found for" + f" --iter {params.iter}, --avg {params.avg}" + ) + filename_start = filenames[-1] + filename_end = filenames[0] + logging.info( + "Calculating the averaged model over iteration checkpoints" + f" from {filename_start} (excluded) to {filename_end}" + ) + model.to(device) + model.load_state_dict( + average_checkpoints_with_averaged_model( + filename_start=filename_start, + filename_end=filename_end, + device=device, + ) + ) + else: + assert params.avg > 0, params.avg + start = params.epoch - params.avg + assert start >= 1, start + filename_start = f"{params.exp_dir}/epoch-{start}.pt" + filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt" + logging.info( + f"Calculating the averaged model over epoch range from " + f"{start} (excluded) to {params.epoch}" + ) + model.to(device) + model.load_state_dict( + average_checkpoints_with_averaged_model( + filename_start=filename_start, + filename_end=filename_end, + device=device, + ) + ) + + model.to("cpu") + model.eval() + + if params.jit: + logging.info("Using torch.jit.script") + # We won't use the forward() method of the model in C++, so just ignore + # it here. + # Otherwise, one of its arguments is a ragged tensor and is not + # torch scriptabe. + convert_scaled_to_non_scaled(model, inplace=True) + model = torch.jit.script(model) + filename = params.exp_dir / "cpu_jit.pt" + model.save(str(filename)) + logging.info(f"Saved to {filename}") + else: + logging.info("Not using torch.jit.script") + # Save it using a format so that it can be loaded + # by :func:`load_checkpoint` + filename = params.exp_dir / "pretrained.pt" + torch.save({"model": model.state_dict()}, str(filename)) + logging.info(f"Saved to {filename}") + + +if __name__ == "__main__": + formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s" + + logging.basicConfig(format=formatter, level=logging.INFO) + main() diff --git a/egs/librispeech/ASR/zipformer_ctc_streaming/label_smoothing.py b/egs/librispeech/ASR/zipformer_ctc_streaming/label_smoothing.py new file mode 100644 index 000000000..52d2eda3b --- /dev/null +++ b/egs/librispeech/ASR/zipformer_ctc_streaming/label_smoothing.py @@ -0,0 +1,109 @@ +# 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 torch + + +class LabelSmoothingLoss(torch.nn.Module): + """ + Implement the LabelSmoothingLoss proposed in the following paper + https://arxiv.org/pdf/1512.00567.pdf + (Rethinking the Inception Architecture for Computer Vision) + + """ + + def __init__( + self, + ignore_index: int = -1, + label_smoothing: float = 0.1, + reduction: str = "sum", + ) -> None: + """ + Args: + ignore_index: + ignored class id + label_smoothing: + smoothing rate (0.0 means the conventional cross entropy loss) + reduction: + It has the same meaning as the reduction in + `torch.nn.CrossEntropyLoss`. It can be one of the following three + values: (1) "none": No reduction will be applied. (2) "mean": the + mean of the output is taken. (3) "sum": the output will be summed. + """ + super().__init__() + assert 0.0 <= label_smoothing < 1.0, f"{label_smoothing}" + assert reduction in ("none", "sum", "mean"), reduction + self.ignore_index = ignore_index + self.label_smoothing = label_smoothing + self.reduction = reduction + + def forward(self, x: torch.Tensor, target: torch.Tensor) -> torch.Tensor: + """ + Compute loss between x and target. + + Args: + x: + prediction of dimension + (batch_size, input_length, number_of_classes). + target: + target masked with self.ignore_index of + dimension (batch_size, input_length). + + Returns: + A scalar tensor containing the loss without normalization. + """ + assert x.ndim == 3 + assert target.ndim == 2 + assert x.shape[:2] == target.shape + num_classes = x.size(-1) + x = x.reshape(-1, num_classes) + # Now x is of shape (N*T, C) + + # We don't want to change target in-place below, + # so we make a copy of it here + target = target.clone().reshape(-1) + + ignored = target == self.ignore_index + + # See https://github.com/k2-fsa/icefall/issues/240 + # and https://github.com/k2-fsa/icefall/issues/297 + # for why we don't use target[ignored] = 0 here + target = torch.where(ignored, torch.zeros_like(target), target) + + true_dist = torch.nn.functional.one_hot(target, num_classes=num_classes).to(x) + + true_dist = ( + true_dist * (1 - self.label_smoothing) + self.label_smoothing / num_classes + ) + + # Set the value of ignored indexes to 0 + # + # See https://github.com/k2-fsa/icefall/issues/240 + # and https://github.com/k2-fsa/icefall/issues/297 + # for why we don't use true_dist[ignored] = 0 here + true_dist = torch.where( + ignored.unsqueeze(1).repeat(1, true_dist.shape[1]), + torch.zeros_like(true_dist), + true_dist, + ) + + loss = -1 * (torch.log_softmax(x, dim=1) * true_dist) + if self.reduction == "sum": + return loss.sum() + elif self.reduction == "mean": + return loss.sum() / (~ignored).sum() + else: + return loss.sum(dim=-1) diff --git a/egs/librispeech/ASR/zipformer_ctc_streaming/model.py b/egs/librispeech/ASR/zipformer_ctc_streaming/model.py new file mode 100644 index 000000000..2aeb8a072 --- /dev/null +++ b/egs/librispeech/ASR/zipformer_ctc_streaming/model.py @@ -0,0 +1,158 @@ +# Copyright 2021 Xiaomi Corp. (authors: Fangjun Kuang, Wei Kang) +# +# See ../../../../LICENSE for clarification regarding multiple authors +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +from typing import List + +import k2 +import torch +import torch.nn as nn +from encoder_interface import EncoderInterface +from transformer import encoder_padding_mask + +from icefall.bpe_graph_compiler import BpeCtcTrainingGraphCompiler +from icefall.utils import encode_supervisions + + +class CTCModel(nn.Module): + """It implements a CTC model with an auxiliary attention head.""" + + def __init__( + self, + encoder: EncoderInterface, + decoder: nn.Module, + encoder_dim: int, + vocab_size: int, + ): + """ + Args: + encoder: + An instance of `EncoderInterface`. The shared encoder for the CTC and attention + branches + decoder: + An instance of `nn.Module`. This is the decoder for the attention branch. + encoder_dim: + Dimension of the encoder output. + decoder_dim: + Dimension of the decoder output. + vocab_size: + Number of tokens of the modeling unit including blank. + """ + super().__init__() + assert isinstance(encoder, EncoderInterface), type(encoder) + + self.encoder = encoder + self.ctc_output = nn.Sequential( + nn.Dropout(p=0.1), + nn.Linear(encoder_dim, vocab_size), + nn.LogSoftmax(dim=-1), + ) + self.decoder = decoder + + @torch.jit.ignore + def forward( + self, + x: torch.Tensor, + x_lens: torch.Tensor, + supervisions: torch.Tensor, + graph_compiler: BpeCtcTrainingGraphCompiler, + subsampling_factor: int = 1, + beam_size: int = 10, + reduction: str = "sum", + use_double_scores: bool = False, + ) -> torch.Tensor: + """ + Args: + x: + Tensor of dimension (N, T, C) where N is the batch size, + T is the number of frames, and C is the feature dimension. + x_lens: + Tensor of dimension (N,) where N is the batch size. + supervisions: + Supervisions are used in training. + graph_compiler: + It is used to compile a decoding graph from texts. + subsampling_factor: + It is used to compute the `supervisions` for the encoder. + beam_size: + Beam size used in `k2.ctc_loss`. + reduction: + Reduction method used in `k2.ctc_loss`. + use_double_scores: + If True, use double precision in `k2.ctc_loss`. + Returns: + Return the CTC loss, attention loss, and the total number of frames. + """ + assert x.ndim == 3, x.shape + assert x_lens.ndim == 1, x_lens.shape + + nnet_output, x_lens = self.encoder(x, x_lens) + assert torch.all(x_lens > 0) + # compute ctc log-probs + ctc_output = self.ctc_output(nnet_output) + + # NOTE: We need `encode_supervisions` to sort sequences with + # different duration in decreasing order, required by + # `k2.intersect_dense` called in `k2.ctc_loss` + supervision_segments, texts = encode_supervisions( + supervisions, subsampling_factor=subsampling_factor + ) + num_frames = supervision_segments[:, 2].sum().item() + + # Works with a BPE model + token_ids = graph_compiler.texts_to_ids(texts) + decoding_graph = graph_compiler.compile(token_ids) + + dense_fsa_vec = k2.DenseFsaVec( + ctc_output, + supervision_segments.cpu(), + allow_truncate=subsampling_factor - 1, + ) + + ctc_loss = k2.ctc_loss( + decoding_graph=decoding_graph, + dense_fsa_vec=dense_fsa_vec, + output_beam=beam_size, + reduction=reduction, + use_double_scores=use_double_scores, + ) + + if self.decoder is not None: + nnet_output = nnet_output.permute(1, 0, 2) # (N, T, C) -> (T, N, C) + mmodel = ( + self.decoder.module if hasattr(self.decoder, "module") else self.decoder + ) + # Note: We need to generate an unsorted version of token_ids + # `encode_supervisions()` called above sorts text, but + # encoder_memory and memory_mask are not sorted, so we + # use an unsorted version `supervisions["text"]` to regenerate + # the token_ids + # + # See https://github.com/k2-fsa/icefall/issues/97 + # for more details + unsorted_token_ids = graph_compiler.texts_to_ids(supervisions["text"]) + mask = encoder_padding_mask(nnet_output.size(0), supervisions) + mask = mask.to(nnet_output.device) if mask is not None else None + att_loss = mmodel.forward( + nnet_output, + mask, + token_ids=unsorted_token_ids, + sos_id=graph_compiler.sos_id, + eos_id=graph_compiler.eos_id, + ) + else: + att_loss = torch.tensor([0]) + + return ctc_loss, att_loss, num_frames diff --git a/egs/librispeech/ASR/zipformer_ctc_streaming/optim.py b/egs/librispeech/ASR/zipformer_ctc_streaming/optim.py new file mode 100644 index 000000000..374b78cb3 --- /dev/null +++ b/egs/librispeech/ASR/zipformer_ctc_streaming/optim.py @@ -0,0 +1,1061 @@ +# Copyright 2022 Xiaomi Corp. (authors: Daniel Povey) +# +# See ../LICENSE for clarification regarding multiple authors +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import contextlib +import logging +import random +from collections import defaultdict +from typing import List, Optional, Tuple, Union + +import torch +from lhotse.utils import fix_random_seed +from scaling import ActivationBalancer +from torch import Tensor +from torch.optim import Optimizer + + +class BatchedOptimizer(Optimizer): + """ + This class adds to class Optimizer the capability to optimize parameters in batches: + it will stack the parameters and their grads for you so the optimizer can work + on tensors with an extra leading dimension. This is intended for speed with GPUs, + as it reduces the number of kernels launched in the optimizer. + + Args: + params: + """ + + def __init__(self, params, defaults): + super(BatchedOptimizer, self).__init__(params, defaults) + + @contextlib.contextmanager + def batched_params(self, param_group, group_params_names): + """ + This function returns (technically, yields) a list of + of tuples (p, state), where + p is a `fake` parameter that is stacked (over axis 0) from real parameters + that share the same shape, and its gradient is also stacked; + `state` is the state corresponding to this batch of parameters + (it will be physically located in the "state" for one of the real + parameters, the last one that has any particular shape and dtype). + + This function is decorated as a context manager so that it can + write parameters back to their "real" locations. + + The idea is, instead of doing: + + for p in group["params"]: + state = self.state[p] + ... + + you can do: + + with self.batched_params(group["params"]) as batches: + for p, state, p_names in batches: + ... + + + Args: + group: a parameter group, which is a list of parameters; should be + one of self.param_groups. + group_params_names: name for each parameter in group, + which is List[str]. + """ + batches = defaultdict( + list + ) # `batches` maps from tuple (dtype_as_str,*shape) to list of nn.Parameter + batches_names = defaultdict( + list + ) # `batches` maps from tuple (dtype_as_str,*shape) to list of str + + assert len(param_group) == len(group_params_names) + for p, named_p in zip(param_group, group_params_names): + key = (str(p.dtype), *p.shape) + batches[key].append(p) + batches_names[key].append(named_p) + + batches_names_keys = list(batches_names.keys()) + sorted_idx = sorted( + range(len(batches_names)), key=lambda i: batches_names_keys[i] + ) + batches_names = [batches_names[batches_names_keys[idx]] for idx in sorted_idx] + batches = [batches[batches_names_keys[idx]] for idx in sorted_idx] + + stacked_params_dict = dict() + + # turn batches into a list, in deterministic order. + # tuples will contain tuples of (stacked_param, state, stacked_params_names), + # one for each batch in `batches`. + tuples = [] + + for batch, batch_names in zip(batches, batches_names): + p = batch[0] + # we arbitrarily store the state in the + # state corresponding to the 1st parameter in the + # group. class Optimizer will take care of saving/loading state. + state = self.state[p] + p_stacked = torch.stack(batch) + grad = torch.stack( + [torch.zeros_like(p) if p.grad is None else p.grad for p in batch] + ) + p_stacked.grad = grad + stacked_params_dict[key] = p_stacked + tuples.append((p_stacked, state, batch_names)) + + yield tuples # <-- calling code will do the actual optimization here! + + for ((stacked_params, _state, _names), batch) in zip(tuples, batches): + for i, p in enumerate(batch): # batch is list of Parameter + p.copy_(stacked_params[i]) + + +class ScaledAdam(BatchedOptimizer): + """ + Implements 'Scaled Adam', a variant of Adam where we scale each parameter's update + proportional to the norm of that parameter; and also learn the scale of the parameter, + in log space, subject to upper and lower limits (as if we had factored each parameter as + param = underlying_param * log_scale.exp()) + + + Args: + params: The parameters or param_groups to optimize (like other Optimizer subclasses) + lr: The learning rate. We will typically use a learning rate schedule that starts + at 0.03 and decreases over time, i.e. much higher than other common + optimizers. + clipping_scale: (e.g. 2.0) + A scale for gradient-clipping: if specified, the normalized gradients + over the whole model will be clipped to have 2-norm equal to + `clipping_scale` times the median 2-norm over the most recent period + of `clipping_update_period` minibatches. By "normalized gradients", + we mean after multiplying by the rms parameter value for this tensor + [for non-scalars]; this is appropriate because our update is scaled + by this quantity. + betas: beta1,beta2 are momentum constants for regular momentum, and moving sum-sq grad. + Must satisfy 0 < beta <= beta2 < 1. + scalar_lr_scale: A scaling factor on the learning rate, that we use to update the + scale of each parameter tensor and scalar parameters of the mode.. + If each parameter were decomposed + as p * p_scale.exp(), where (p**2).mean().sqrt() == 1.0, scalar_lr_scale + would be a the scaling factor on the learning rate of p_scale. + eps: A general-purpose epsilon to prevent division by zero + param_min_rms: Minimum root-mean-square value of parameter tensor, for purposes of + learning the scale on the parameters (we'll constrain the rms of each non-scalar + parameter tensor to be >= this value) + param_max_rms: Maximum root-mean-square value of parameter tensor, for purposes of + learning the scale on the parameters (we'll constrain the rms of each non-scalar + parameter tensor to be <= this value) + scalar_max: Maximum absolute value for scalar parameters (applicable if your + model has any parameters with numel() == 1). + size_update_period: The periodicity, in steps, with which we update the size (scale) + of the parameter tensor. This is provided to save a little time + in the update. + clipping_update_period: if clipping_scale is specified, this is the period + """ + + def __init__( + self, + params, + lr=3e-02, + clipping_scale=None, + betas=(0.9, 0.98), + scalar_lr_scale=0.1, + eps=1.0e-08, + param_min_rms=1.0e-05, + param_max_rms=3.0, + scalar_max=10.0, + size_update_period=4, + clipping_update_period=100, + parameters_names=None, + show_dominant_parameters=True, + ): + + assert parameters_names is not None, ( + "Please prepare parameters_names," + "which is a List[List[str]]. Each List[str] is for a group" + "and each str is for a parameter" + ) + defaults = dict( + lr=lr, + clipping_scale=clipping_scale, + betas=betas, + scalar_lr_scale=scalar_lr_scale, + eps=eps, + param_min_rms=param_min_rms, + param_max_rms=param_max_rms, + scalar_max=scalar_max, + size_update_period=size_update_period, + clipping_update_period=clipping_update_period, + ) + + super(ScaledAdam, self).__init__(params, defaults) + assert len(self.param_groups) == len(parameters_names) + self.parameters_names = parameters_names + self.show_dominant_parameters = show_dominant_parameters + + def __setstate__(self, state): + super(ScaledAdam, self).__setstate__(state) + + @torch.no_grad() + def step(self, closure=None): + """Performs a single optimization step. + + Arguments: + closure (callable, optional): A closure that reevaluates the model + and returns the loss. + """ + loss = None + if closure is not None: + with torch.enable_grad(): + loss = closure() + + batch = True + + for group, group_params_names in zip(self.param_groups, self.parameters_names): + + with self.batched_params(group["params"], group_params_names) as batches: + + # batches is list of pairs (stacked_param, state). stacked_param is like + # a regular parameter, and will have a .grad, but the 1st dim corresponds to + # a stacking dim, it is not a real dim. + + if ( + len(batches[0][1]) == 0 + ): # if len(first state) == 0: not yet initialized + clipping_scale = 1 + else: + clipping_scale = self._get_clipping_scale(group, batches) + + for p, state, _ in batches: + # Perform optimization step. + # grad is not going to be None, we handled that when creating the batches. + grad = p.grad + if grad.is_sparse: + raise RuntimeError( + "ScaledAdam optimizer does not support sparse gradients" + ) + # State initialization + if len(state) == 0: + self._init_state(group, p, state) + + self._step_one_batch(group, p, state, clipping_scale) + + return loss + + def _init_state(self, group: dict, p: Tensor, state: dict): + """ + Initializes state dict for parameter 'p'. Assumes that dim 0 of tensor p + is actually the batch dimension, corresponding to batched-together + parameters of a given shape. + + + Args: + group: Dict to look up configuration values. + p: The parameter that we are initializing the state for + state: Dict from string to whatever state we are initializing + """ + size_update_period = group["size_update_period"] + + state["step"] = 0 + + kwargs = {"device": p.device, "dtype": p.dtype} + + # 'delta' implements conventional momentum. There are + # several different kinds of update going on, so rather than + # compute "exp_avg" like in Adam, we store and decay a + # parameter-change "delta", which combines all forms of + # update. this is equivalent to how it's done in Adam, + # except for the first few steps. + state["delta"] = torch.zeros_like(p, memory_format=torch.preserve_format) + + batch_size = p.shape[0] + numel = p.numel() // batch_size + numel = p.numel() + + if numel > 1: + # "param_rms" just periodically records the scalar root-mean-square value of + # the parameter tensor. + # it has a shape like (batch_size, 1, 1, 1, 1) + param_rms = (p**2).mean(dim=list(range(1, p.ndim)), keepdim=True).sqrt() + state["param_rms"] = param_rms + + state["scale_exp_avg_sq"] = torch.zeros_like(param_rms) + state["scale_grads"] = torch.zeros( + size_update_period, *param_rms.shape, **kwargs + ) + + # exp_avg_sq is the weighted sum of scaled gradients. as in Adam. + state["exp_avg_sq"] = torch.zeros_like(p, memory_format=torch.preserve_format) + + def _get_clipping_scale( + self, group: dict, tuples: List[Tuple[Tensor, dict, List[str]]] + ) -> float: + """ + Returns a scalar factor <= 1.0 that dictates gradient clipping, i.e. we will scale the gradients + by this amount before applying the rest of the update. + + Args: + group: the parameter group, an item in self.param_groups + tuples: a list of tuples of (param, state, param_names) + where param is a batched set of parameters, + with a .grad (1st dim is batch dim) + and state is the state-dict where optimization parameters are kept. + param_names is a List[str] while each str is name for a parameter + in batched set of parameters "param". + """ + assert len(tuples) >= 1 + clipping_scale = group["clipping_scale"] + (first_p, first_state, _) = tuples[0] + step = first_state["step"] + if clipping_scale is None or step == 0: + # no clipping. return early on step == 0 because the other + # parameters' state won't have been initialized yet. + return 1.0 + clipping_update_period = group["clipping_update_period"] + + tot_sumsq = torch.tensor(0.0, device=first_p.device) + for (p, state, param_names) in tuples: + grad = p.grad + if grad.is_sparse: + raise RuntimeError( + "ScaledAdam optimizer does not support sparse gradients" + ) + if p.numel() == p.shape[0]: # a batch of scalars + tot_sumsq += (grad**2).sum() # sum() to change shape [1] to [] + else: + tot_sumsq += ((grad * state["param_rms"]) ** 2).sum() + + tot_norm = tot_sumsq.sqrt() + if "model_norms" not in first_state: + first_state["model_norms"] = torch.zeros( + clipping_update_period, device=p.device + ) + first_state["model_norms"][step % clipping_update_period] = tot_norm + + if step % clipping_update_period == 0: + # Print some stats. + # We don't reach here if step == 0 because we would have returned + # above. + sorted_norms = first_state["model_norms"].sort()[0].to("cpu") + quartiles = [] + for n in range(0, 5): + index = min( + clipping_update_period - 1, (clipping_update_period // 4) * n + ) + quartiles.append(sorted_norms[index].item()) + + median = quartiles[2] + threshold = clipping_scale * median + first_state["model_norm_threshold"] = threshold + percent_clipped = ( + first_state["num_clipped"] * 100.0 / clipping_update_period + if "num_clipped" in first_state + else 0.0 + ) + first_state["num_clipped"] = 0 + quartiles = " ".join(["%.3e" % x for x in quartiles]) + logging.info( + f"Clipping_scale={clipping_scale}, grad-norm quartiles {quartiles}, " + f"threshold={threshold:.3e}, percent-clipped={percent_clipped:.1f}" + ) + + if step < clipping_update_period: + return 1.0 # We have not yet estimated a norm to clip to. + else: + try: + model_norm_threshold = first_state["model_norm_threshold"] + except KeyError: + logging.info( + "Warning: model_norm_threshold not in state: possibly " + "you changed config when restarting, adding clipping_scale option?" + ) + return 1.0 + ans = min(1.0, (model_norm_threshold / (tot_norm + 1.0e-20)).item()) + if ans < 1.0: + first_state["num_clipped"] += 1 + if ans < 0.1: + logging.warn( + f"Scaling gradients by {ans}, model_norm_threshold={model_norm_threshold}" + ) + if self.show_dominant_parameters: + assert p.shape[0] == len(param_names) + self._show_gradient_dominating_parameter(tuples, tot_sumsq) + return ans + + def _show_gradient_dominating_parameter( + self, tuples: List[Tuple[Tensor, dict, List[str]]], tot_sumsq: Tensor + ): + """ + Show information of parameter wihch dominanting tot_sumsq. + + Args: + tuples: a list of tuples of (param, state, param_names) + where param is a batched set of parameters, + with a .grad (1st dim is batch dim) + and state is the state-dict where optimization parameters are kept. + param_names is a List[str] while each str is name for a parameter + in batched set of parameters "param". + tot_sumsq: sumsq of all parameters. Though it's could be calculated + from tuples, we still pass it to save some time. + """ + all_sumsq_orig = {} + for (p, state, batch_param_names) in tuples: + # p is a stacked batch parameters. + batch_grad = p.grad + if p.numel() == p.shape[0]: # a batch of scalars + batch_sumsq_orig = batch_grad**2 + # Dummpy values used by following `zip` statement. + batch_rms_orig = torch.ones(p.shape[0]) + else: + batch_rms_orig = state["param_rms"] + batch_sumsq_orig = ((batch_grad * batch_rms_orig) ** 2).sum( + dim=list(range(1, batch_grad.ndim)) + ) + + for name, sumsq_orig, rms, grad in zip( + batch_param_names, batch_sumsq_orig, batch_rms_orig, batch_grad + ): + + proportion_orig = sumsq_orig / tot_sumsq + all_sumsq_orig[name] = (proportion_orig, sumsq_orig, rms, grad) + + assert torch.isclose( + sum([value[0] for value in all_sumsq_orig.values()]).cpu(), + torch.tensor(1.0), + ) + sorted_by_proportion = { + k: v + for k, v in sorted( + all_sumsq_orig.items(), key=lambda item: item[1][0], reverse=True + ) + } + dominant_param_name = next(iter(sorted_by_proportion)) + ( + dominant_proportion, + dominant_sumsq, + dominant_rms, + dominant_grad, + ) = sorted_by_proportion[dominant_param_name] + logging.info( + f"Parameter Dominanting tot_sumsq {dominant_param_name}" + f" with proportion {dominant_proportion:.2f}," + f" where dominant_sumsq=(grad_sumsq*orig_rms_sq)" + f"={dominant_sumsq:.3e}," + f" grad_sumsq = {(dominant_grad**2).sum():.3e}," + f" orig_rms_sq={(dominant_rms**2).item():.3e}" + ) + + def _step_one_batch( + self, group: dict, p: Tensor, state: dict, clipping_scale: float + ): + """ + Do the step for one parameter, which is actually going to be a batch of + `real` parameters, with dim 0 as the batch dim. + Args: + group: dict to look up configuration values + p: parameter to update (actually multiple parameters stacked together + as a batch) + state: state-dict for p, to look up the optimizer state + """ + lr = group["lr"] + size_update_period = group["size_update_period"] + beta1 = group["betas"][0] + + grad = p.grad + if clipping_scale != 1.0: + grad = grad * clipping_scale + step = state["step"] + delta = state["delta"] + + delta.mul_(beta1) + batch_size = p.shape[0] + numel = p.numel() // batch_size + if numel > 1: + # Update the size/scale of p, and set param_rms + scale_grads = state["scale_grads"] + scale_grads[step % size_update_period] = (p * grad).sum( + dim=list(range(1, p.ndim)), keepdim=True + ) + if step % size_update_period == size_update_period - 1: + param_rms = state["param_rms"] # shape: (batch_size, 1, 1, ..) + param_rms.copy_( + (p**2).mean(dim=list(range(1, p.ndim)), keepdim=True).sqrt() + ) + if step > 0: + # self._size_update() learns the overall scale on the + # parameter, by shrinking or expanding it. + self._size_update(group, scale_grads, p, state) + + if numel == 1: + # For parameters with 1 element we just use regular Adam. + # Updates delta. + self._step_scalar(group, p, state) + else: + self._step(group, p, state) + + state["step"] = step + 1 + + def _size_update( + self, group: dict, scale_grads: Tensor, p: Tensor, state: dict + ) -> None: + """ + Called only where p.numel() > 1, this updates the scale of the parameter. + If we imagine: p = underlying_param * scale.exp(), and we are doing + gradient descent on underlying param and on scale, this function does the update + on `scale`. + + Args: + group: dict to look up configuration values + scale_grads: a tensor of shape (size_update_period, batch_size, 1, 1,...) containing + grads w.r.t. the scales. + p: The parameter to update + state: The state-dict of p + """ + + param_rms = state["param_rms"] + beta1, beta2 = group["betas"] + size_lr = group["lr"] * group["scalar_lr_scale"] + param_min_rms = group["param_min_rms"] + param_max_rms = group["param_max_rms"] + eps = group["eps"] + step = state["step"] + batch_size = p.shape[0] + + size_update_period = scale_grads.shape[0] + # correct beta2 for the size update period: we will have + # faster decay at this level. + beta2_corr = beta2**size_update_period + + scale_exp_avg_sq = state["scale_exp_avg_sq"] # shape: (batch_size, 1, 1, ..) + scale_exp_avg_sq.mul_(beta2_corr).add_( + (scale_grads**2).mean(dim=0), # mean over dim `size_update_period` + alpha=1 - beta2_corr, + ) # shape is (batch_size, 1, 1, ...) + + # The 1st time we reach here is when size_step == 1. + size_step = (step + 1) // size_update_period + bias_correction2 = 1 - beta2_corr**size_step + # we don't bother with bias_correction1; this will help prevent divergence + # at the start of training. + + denom = scale_exp_avg_sq.sqrt() + eps + + scale_step = ( + -size_lr * (bias_correction2**0.5) * scale_grads.sum(dim=0) / denom + ) + + is_too_small = param_rms < param_min_rms + is_too_large = param_rms > param_max_rms + + # when the param gets too small, just don't shrink it any further. + scale_step.masked_fill_(is_too_small, 0.0) + # when it gets too large, stop it from getting any larger. + scale_step.masked_fill_(is_too_large, -size_lr * size_update_period) + delta = state["delta"] + # the factor of (1-beta1) relates to momentum. + delta.add_(p * scale_step, alpha=(1 - beta1)) + + def _step(self, group: dict, p: Tensor, state: dict): + """ + This function does the core update of self.step(), in the case where the members of + the batch have more than 1 element. + + Args: + group: A dict which will be used to look up configuration values + p: The parameter to be updated + grad: The grad of p + state: The state-dict corresponding to parameter p + + This function modifies p. + """ + grad = p.grad + lr = group["lr"] + beta1, beta2 = group["betas"] + eps = group["eps"] + param_min_rms = group["param_min_rms"] + step = state["step"] + + exp_avg_sq = state["exp_avg_sq"] + exp_avg_sq.mul_(beta2).addcmul_(grad, grad, value=(1 - beta2)) + + this_step = state["step"] - (state["zero_step"] if "zero_step" in state else 0) + bias_correction2 = 1 - beta2 ** (this_step + 1) + if bias_correction2 < 0.99: + # note: not in-place. + exp_avg_sq = exp_avg_sq * (1.0 / bias_correction2) + + denom = exp_avg_sq.sqrt() + denom += eps + grad = grad / denom + + alpha = -lr * (1 - beta1) * state["param_rms"].clamp(min=param_min_rms) + + delta = state["delta"] + delta.add_(grad * alpha) + p.add_(delta) + + def _step_scalar(self, group: dict, p: Tensor, state: dict): + """ + A simplified form of the core update for scalar tensors, where we cannot get a good + estimate of the parameter rms. + """ + beta1, beta2 = group["betas"] + scalar_max = group["scalar_max"] + eps = group["eps"] + lr = group["lr"] * group["scalar_lr_scale"] + grad = p.grad + + exp_avg_sq = state["exp_avg_sq"] # shape: (batch_size,) + exp_avg_sq.mul_(beta2).addcmul_(grad, grad, value=1 - beta2) + + # bias_correction2 is like in Adam. Don't bother with bias_correction1; + # slower update at the start will help stability anyway. + bias_correction2 = 1 - beta2 ** (state["step"] + 1) + denom = (exp_avg_sq / bias_correction2).sqrt() + eps + + delta = state["delta"] + delta.add_(grad / denom, alpha=-lr * (1 - beta1)) + p.clamp_(min=-scalar_max, max=scalar_max) + p.add_(delta) + + +class LRScheduler(object): + """ + Base-class for learning rate schedulers where the learning-rate depends on both the + batch and the epoch. + """ + + def __init__(self, optimizer: Optimizer, verbose: bool = False): + # Attach optimizer + if not isinstance(optimizer, Optimizer): + raise TypeError("{} is not an Optimizer".format(type(optimizer).__name__)) + self.optimizer = optimizer + self.verbose = verbose + + for group in optimizer.param_groups: + group.setdefault("base_lr", group["lr"]) + + self.base_lrs = [group["base_lr"] for group in optimizer.param_groups] + + self.epoch = 0 + self.batch = 0 + + def state_dict(self): + """Returns the state of the scheduler as a :class:`dict`. + + It contains an entry for every variable in self.__dict__ which + is not the optimizer. + """ + return { + "base_lrs": self.base_lrs, + "epoch": self.epoch, + "batch": self.batch, + } + + def load_state_dict(self, state_dict): + """Loads the schedulers state. + + Args: + state_dict (dict): scheduler state. Should be an object returned + from a call to :meth:`state_dict`. + """ + self.__dict__.update(state_dict) + + def get_last_lr(self) -> List[float]: + """Return last computed learning rate by current scheduler. Will be a list of float.""" + return self._last_lr + + def get_lr(self): + # Compute list of learning rates from self.epoch and self.batch and + # self.base_lrs; this must be overloaded by the user. + # e.g. return [some_formula(self.batch, self.epoch, base_lr) for base_lr in self.base_lrs ] + raise NotImplementedError + + def step_batch(self, batch: Optional[int] = None) -> None: + # Step the batch index, or just set it. If `batch` is specified, it + # must be the batch index from the start of training, i.e. summed over + # all epochs. + # You can call this in any order; if you don't provide 'batch', it should + # of course be called once per batch. + if batch is not None: + self.batch = batch + else: + self.batch = self.batch + 1 + self._set_lrs() + + def step_epoch(self, epoch: Optional[int] = None): + # Step the epoch index, or just set it. If you provide the 'epoch' arg, + # you should call this at the start of the epoch; if you don't provide the 'epoch' + # arg, you should call it at the end of the epoch. + if epoch is not None: + self.epoch = epoch + else: + self.epoch = self.epoch + 1 + self._set_lrs() + + def _set_lrs(self): + values = self.get_lr() + assert len(values) == len(self.optimizer.param_groups) + + for i, data in enumerate(zip(self.optimizer.param_groups, values)): + param_group, lr = data + param_group["lr"] = lr + self.print_lr(self.verbose, i, lr) + self._last_lr = [group["lr"] for group in self.optimizer.param_groups] + + def print_lr(self, is_verbose, group, lr): + """Display the current learning rate.""" + if is_verbose: + logging.info( + f"Epoch={self.epoch}, batch={self.batch}: adjusting learning rate" + f" of group {group} to {lr:.4e}." + ) + + +class Eden(LRScheduler): + """ + Eden scheduler. + The basic formula (before warmup) is: + lr = base_lr * (((batch**2 + lr_batches**2) / lr_batches**2) ** -0.25 * + (((epoch**2 + lr_epochs**2) / lr_epochs**2) ** -0.25)) * warmup + where `warmup` increases from linearly 0.5 to 1 over `warmup_batches` batches + and then stays constant at 1. + + + E.g. suggest base_lr = 0.04 (passed to optimizer) if used with ScaledAdam + + Args: + optimizer: the optimizer to change the learning rates on + lr_batches: the number of batches after which we start significantly + decreasing the learning rate, suggest 5000. + lr_epochs: the number of epochs after which we start significantly + decreasing the learning rate, suggest 6 if you plan to do e.g. + 20 to 40 epochs, but may need smaller number if dataset is huge + and you will do few epochs. + """ + + def __init__( + self, + optimizer: Optimizer, + lr_batches: Union[int, float], + lr_epochs: Union[int, float], + warmup_batches: Union[int, float] = 500.0, + verbose: bool = False, + ): + super(Eden, self).__init__(optimizer, verbose) + self.lr_batches = lr_batches + self.lr_epochs = lr_epochs + self.warmup_batches = warmup_batches + + def get_lr(self): + factor = ( + (self.batch**2 + self.lr_batches**2) / self.lr_batches**2 + ) ** -0.25 * ( + ((self.epoch**2 + self.lr_epochs**2) / self.lr_epochs**2) ** -0.25 + ) + warmup_factor = ( + 1.0 + if self.batch >= self.warmup_batches + else 0.5 + 0.5 * (self.batch / self.warmup_batches) + ) + + return [x * factor * warmup_factor for x in self.base_lrs] + + +def _test_eden(): + m = torch.nn.Linear(100, 100) + optim = ScaledAdam(m.parameters(), lr=0.03) + + scheduler = Eden(optim, lr_batches=100, lr_epochs=2, verbose=True) + + for epoch in range(10): + scheduler.step_epoch(epoch) # sets epoch to `epoch` + + for step in range(20): + x = torch.randn(200, 100).detach() + x.requires_grad = True + y = m(x) + dy = torch.randn(200, 100).detach() + f = (y * dy).sum() + f.backward() + + optim.step() + scheduler.step_batch() + optim.zero_grad() + + logging.info(f"last lr = {scheduler.get_last_lr()}") + logging.info(f"state dict = {scheduler.state_dict()}") + + +# This is included mostly as a baseline for ScaledAdam. +class Eve(Optimizer): + """ + Implements Eve algorithm. This is a modified version of AdamW with a special + way of setting the weight-decay / shrinkage-factor, which is designed to make the + rms of the parameters approach a particular target_rms (default: 0.1). This is + for use with networks with 'scaled' versions of modules (see scaling.py), which + will be close to invariant to the absolute scale on the parameter matrix. + + The original Adam algorithm was proposed in `Adam: A Method for Stochastic Optimization`_. + The AdamW variant was proposed in `Decoupled Weight Decay Regularization`_. + Eve is unpublished so far. + + Arguments: + params (iterable): iterable of parameters to optimize or dicts defining + parameter groups + lr (float, optional): learning rate (default: 1e-3) + betas (Tuple[float, float], optional): coefficients used for computing + running averages of gradient and its square (default: (0.9, 0.999)) + eps (float, optional): term added to the denominator to improve + numerical stability (default: 1e-8) + weight_decay (float, optional): weight decay coefficient (default: 3e-4; + this value means that the weight would decay significantly after + about 3k minibatches. Is not multiplied by learning rate, but + is conditional on RMS-value of parameter being > target_rms. + target_rms (float, optional): target root-mean-square value of + parameters, if they fall below this we will stop applying weight decay. + + + .. _Adam: A Method for Stochastic Optimization: + https://arxiv.org/abs/1412.6980 + .. _Decoupled Weight Decay Regularization: + https://arxiv.org/abs/1711.05101 + .. _On the Convergence of Adam and Beyond: + https://openreview.net/forum?id=ryQu7f-RZ + """ + + def __init__( + self, + params, + lr=1e-3, + betas=(0.9, 0.98), + eps=1e-8, + weight_decay=1e-3, + target_rms=0.1, + ): + if not 0.0 <= lr: + raise ValueError("Invalid learning rate: {}".format(lr)) + if not 0.0 <= eps: + raise ValueError("Invalid epsilon value: {}".format(eps)) + if not 0.0 <= betas[0] < 1.0: + raise ValueError("Invalid beta parameter at index 0: {}".format(betas[0])) + if not 0.0 <= betas[1] < 1.0: + raise ValueError("Invalid beta parameter at index 1: {}".format(betas[1])) + if not 0 <= weight_decay <= 0.1: + raise ValueError("Invalid weight_decay value: {}".format(weight_decay)) + if not 0 < target_rms <= 10.0: + raise ValueError("Invalid target_rms value: {}".format(target_rms)) + defaults = dict( + lr=lr, + betas=betas, + eps=eps, + weight_decay=weight_decay, + target_rms=target_rms, + ) + super(Eve, self).__init__(params, defaults) + + def __setstate__(self, state): + super(Eve, self).__setstate__(state) + + @torch.no_grad() + def step(self, closure=None): + """Performs a single optimization step. + + Arguments: + closure (callable, optional): A closure that reevaluates the model + and returns the loss. + """ + loss = None + if closure is not None: + with torch.enable_grad(): + loss = closure() + + for group in self.param_groups: + for p in group["params"]: + if p.grad is None: + continue + + # Perform optimization step + grad = p.grad + if grad.is_sparse: + raise RuntimeError("AdamW does not support sparse gradients") + + state = self.state[p] + + # State initialization + if len(state) == 0: + state["step"] = 0 + # Exponential moving average of gradient values + state["exp_avg"] = torch.zeros_like( + p, memory_format=torch.preserve_format + ) + # Exponential moving average of squared gradient values + state["exp_avg_sq"] = torch.zeros_like( + p, memory_format=torch.preserve_format + ) + + exp_avg, exp_avg_sq = state["exp_avg"], state["exp_avg_sq"] + + beta1, beta2 = group["betas"] + + state["step"] += 1 + bias_correction1 = 1 - beta1 ** state["step"] + bias_correction2 = 1 - beta2 ** state["step"] + + # Decay the first and second moment running average coefficient + exp_avg.mul_(beta1).add_(grad, alpha=1 - beta1) + exp_avg_sq.mul_(beta2).addcmul_(grad, grad, value=1 - beta2) + denom = (exp_avg_sq.sqrt() * (bias_correction2**-0.5)).add_( + group["eps"] + ) + + step_size = group["lr"] / bias_correction1 + target_rms = group["target_rms"] + weight_decay = group["weight_decay"] + + if p.numel() > 1: + # avoid applying this weight-decay on "scaling factors" + # (which are scalar). + is_above_target_rms = p.norm() > (target_rms * (p.numel() ** 0.5)) + p.mul_(1 - (weight_decay * is_above_target_rms)) + + p.addcdiv_(exp_avg, denom, value=-step_size) + + if random.random() < 0.0005: + step = (exp_avg / denom) * step_size + logging.info( + f"Delta rms = {(step**2).mean().item()}, shape = {step.shape}" + ) + + return loss + + +def _test_scaled_adam(hidden_dim: int): + import timeit + + from scaling import ScaledLinear + + E = 100 + B = 4 + T = 2 + logging.info("in test_eve_cain") + # device = torch.device('cuda') + device = torch.device("cpu") + dtype = torch.float32 + + fix_random_seed(42) + # these input_magnitudes and output_magnitudes are to test that + # Abel is working as we expect and is able to adjust scales of + # different dims differently. + input_magnitudes = (1.0 * torch.randn(E, dtype=dtype, device=device)).exp() + output_magnitudes = (1.0 * torch.randn(E, dtype=dtype, device=device)).exp() + + for iter in [1, 0]: + fix_random_seed(42) + Linear = torch.nn.Linear if iter == 0 else ScaledLinear + + m = torch.nn.Sequential( + Linear(E, hidden_dim), + torch.nn.PReLU(), + Linear(hidden_dim, hidden_dim), + torch.nn.PReLU(), + Linear(hidden_dim, E), + ).to(device) + + train_pairs = [ + ( + 100.0 + * torch.randn(B, T, E, device=device, dtype=dtype) + * input_magnitudes, + torch.randn(B, T, E, device=device, dtype=dtype) * output_magnitudes, + ) + for _ in range(20) + ] + + if iter == 0: + optim = Eve(m.parameters(), lr=0.003) + elif iter == 1: + optim = ScaledAdam(m.parameters(), lr=0.03, clipping_scale=2.0) + scheduler = Eden(optim, lr_batches=200, lr_epochs=5, verbose=False) + + start = timeit.default_timer() + avg_loss = 0.0 + for epoch in range(180): + scheduler.step_epoch() + # if epoch == 100 and iter in [2,3]: + # optim.reset_speedup() # check it doesn't crash. + + # if epoch == 130: + # opts = diagnostics.TensorDiagnosticOptions( + # 2 ** 22 + # ) # allow 4 megabytes per sub-module + # diagnostic = diagnostics.attach_diagnostics(m, opts) + + for n, (x, y) in enumerate(train_pairs): + y_out = m(x) + loss = ((y_out - y) ** 2).mean() * 100.0 + if epoch == 0 and n == 0: + avg_loss = loss.item() + else: + avg_loss = 0.98 * avg_loss + 0.02 * loss.item() + if n == 0 and epoch % 5 == 0: + # norm1 = '%.2e' % (m[0].weight**2).mean().sqrt().item() + # norm1b = '%.2e' % (m[0].bias**2).mean().sqrt().item() + # norm2 = '%.2e' % (m[2].weight**2).mean().sqrt().item() + # norm2b = '%.2e' % (m[2].bias**2).mean().sqrt().item() + # scale1 = '%.2e' % (m[0].weight_scale.exp().item()) + # scale1b = '%.2e' % (m[0].bias_scale.exp().item()) + # scale2 = '%.2e' % (m[2].weight_scale.exp().item()) + # scale2b = '%.2e' % (m[2].bias_scale.exp().item()) + lr = scheduler.get_last_lr()[0] + logging.info( + f"Iter {iter}, epoch {epoch}, batch {n}, avg_loss {avg_loss:.4g}, lr={lr:.4e}" + ) # , norms={norm1,norm1b,norm2,norm2b}") # scales={scale1,scale1b,scale2,scale2b} + loss.log().backward() + optim.step() + optim.zero_grad() + scheduler.step_batch() + + # diagnostic.print_diagnostics() + + stop = timeit.default_timer() + logging.info(f"Iter={iter}, Time taken: {stop - start}") + + logging.info(f"last lr = {scheduler.get_last_lr()}") + # logging.info("state dict = ", scheduler.state_dict()) + # logging.info("optim state_dict = ", optim.state_dict()) + logging.info(f"input_magnitudes = {input_magnitudes}") + logging.info(f"output_magnitudes = {output_magnitudes}") + + +if __name__ == "__main__": + torch.set_num_threads(1) + torch.set_num_interop_threads(1) + logging.getLogger().setLevel(logging.INFO) + import subprocess + + s = subprocess.check_output( + "git status -uno .; git log -1; git diff HEAD .", shell=True + ) + logging.info(s) + import sys + + if len(sys.argv) > 1: + hidden_dim = int(sys.argv[1]) + else: + hidden_dim = 200 + + _test_scaled_adam(hidden_dim) + _test_eden() diff --git a/egs/librispeech/ASR/zipformer_ctc_streaming/scaling.py b/egs/librispeech/ASR/zipformer_ctc_streaming/scaling.py new file mode 100644 index 000000000..156b91f09 --- /dev/null +++ b/egs/librispeech/ASR/zipformer_ctc_streaming/scaling.py @@ -0,0 +1,1179 @@ +# Copyright 2022 Xiaomi Corp. (authors: Daniel Povey) +# +# See ../../../../LICENSE for clarification regarding multiple authors +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + + +import collections +import logging +import random +from functools import reduce +from itertools import repeat +from typing import Optional, Tuple, Union + +import torch +import torch.nn as nn +import torch.nn.functional as F +from torch import Tensor +from torch.nn import Embedding as ScaledEmbedding + + +class ActivationBalancerFunction(torch.autograd.Function): + @staticmethod + def forward( + ctx, + x: Tensor, + scale_factor: Tensor, + sign_factor: Optional[Tensor], + channel_dim: int, + ) -> Tensor: + if channel_dim < 0: + channel_dim += x.ndim + ctx.channel_dim = channel_dim + xgt0 = x > 0 + if sign_factor is None: + ctx.save_for_backward(xgt0, scale_factor) + else: + ctx.save_for_backward(xgt0, scale_factor, sign_factor) + return x + + @staticmethod + def backward(ctx, x_grad: Tensor) -> Tuple[Tensor, None, None, None]: + if len(ctx.saved_tensors) == 3: + xgt0, scale_factor, sign_factor = ctx.saved_tensors + for _ in range(ctx.channel_dim, x_grad.ndim - 1): + scale_factor = scale_factor.unsqueeze(-1) + sign_factor = sign_factor.unsqueeze(-1) + factor = sign_factor + scale_factor * (xgt0.to(x_grad.dtype) - 0.5) + else: + xgt0, scale_factor = ctx.saved_tensors + for _ in range(ctx.channel_dim, x_grad.ndim - 1): + scale_factor = scale_factor.unsqueeze(-1) + factor = scale_factor * (xgt0.to(x_grad.dtype) - 0.5) + neg_delta_grad = x_grad.abs() * factor + return ( + x_grad - neg_delta_grad, + None, + None, + None, + ) + + +def _compute_scale_factor( + x: Tensor, + channel_dim: int, + min_abs: float, + max_abs: float, + gain_factor: float, + max_factor: float, +) -> Tensor: + if channel_dim < 0: + channel_dim += x.ndim + sum_dims = [d for d in range(x.ndim) if d != channel_dim] + x_abs_mean = torch.mean(x.abs(), dim=sum_dims).to(torch.float32) + + if min_abs == 0.0: + below_threshold = 0.0 + else: + # below_threshold is 0 if x_abs_mean > min_abs, can be at most max_factor if + # x_abs)_mean , min_abs. + below_threshold = ((min_abs - x_abs_mean) * (gain_factor / min_abs)).clamp( + min=0, max=max_factor + ) + + above_threshold = ((x_abs_mean - max_abs) * (gain_factor / max_abs)).clamp( + min=0, max=max_factor + ) + + return below_threshold - above_threshold + + +def _compute_sign_factor( + x: Tensor, + channel_dim: int, + min_positive: float, + max_positive: float, + gain_factor: float, + max_factor: float, +) -> Tensor: + if channel_dim < 0: + channel_dim += x.ndim + sum_dims = [d for d in range(x.ndim) if d != channel_dim] + proportion_positive = torch.mean((x > 0).to(torch.float32), dim=sum_dims) + if min_positive == 0.0: + factor1 = 0.0 + else: + # 0 if proportion_positive >= min_positive, else can be + # as large as max_factor. + factor1 = ( + (min_positive - proportion_positive) * (gain_factor / min_positive) + ).clamp_(min=0, max=max_factor) + + if max_positive == 1.0: + factor2 = 0.0 + else: + # 0 if self.proportion_positive <= max_positive, else can be + # as large as -max_factor. + factor2 = ( + (proportion_positive - max_positive) * (gain_factor / (1.0 - max_positive)) + ).clamp_(min=0, max=max_factor) + sign_factor = factor1 - factor2 + # require min_positive != 0 or max_positive != 1: + assert not isinstance(sign_factor, float) + return sign_factor + + +class ActivationScaleBalancerFunction(torch.autograd.Function): + """ + This object is used in class ActivationBalancer when the user specified + min_positive=0, max_positive=1, so there are no constraints on the signs + of the activations and only the absolute value has a constraint. + """ + + @staticmethod + def forward( + ctx, + x: Tensor, + sign_factor: Tensor, + scale_factor: Tensor, + channel_dim: int, + ) -> Tensor: + if channel_dim < 0: + channel_dim += x.ndim + ctx.channel_dim = channel_dim + xgt0 = x > 0 + ctx.save_for_backward(xgt0, sign_factor, scale_factor) + return x + + @staticmethod + def backward(ctx, x_grad: Tensor) -> Tuple[Tensor, None, None, None]: + xgt0, sign_factor, scale_factor = ctx.saved_tensors + for _ in range(ctx.channel_dim, x_grad.ndim - 1): + sign_factor = sign_factor.unsqueeze(-1) + scale_factor = scale_factor.unsqueeze(-1) + + factor = sign_factor + scale_factor * (xgt0.to(x_grad.dtype) - 0.5) + neg_delta_grad = x_grad.abs() * factor + return ( + x_grad - neg_delta_grad, + None, + None, + None, + ) + + +class RandomClampFunction(torch.autograd.Function): + @staticmethod + def forward( + ctx, + x: Tensor, + min: Optional[float], + max: Optional[float], + prob: float, + reflect: float, + ) -> Tensor: + x_clamped = torch.clamp(x, min=min, max=max) + mask = torch.rand_like(x) < prob + ans = torch.where(mask, x_clamped, x) + if x.requires_grad: + ctx.save_for_backward(ans == x) + ctx.reflect = reflect + if reflect != 0.0: + ans = ans * (1.0 + reflect) - (x * reflect) + return ans + + @staticmethod + def backward(ctx, ans_grad: Tensor) -> Tuple[Tensor, None, None, None, None]: + (is_same,) = ctx.saved_tensors + x_grad = ans_grad * is_same.to(ans_grad.dtype) + reflect = ctx.reflect + if reflect != 0.0: + x_grad = x_grad * (1.0 + reflect) - (ans_grad * reflect) + return x_grad, None, None, None, None + + +def random_clamp( + x: Tensor, + min: Optional[float] = None, + max: Optional[float] = None, + prob: float = 0.5, + reflect: float = 0.0, +): + return RandomClampFunction.apply(x, min, max, prob, reflect) + + +def random_cast_to_half(x: Tensor, min_abs: float = 5.0e-06) -> Tensor: + """ + A randomized way of casting a floating point value to half precision. + """ + if x.dtype == torch.float16: + return x + x_abs = x.abs() + is_too_small = x_abs < min_abs + # for elements where is_too_small is true, random_val will contain +-min_abs with + # probability (x.abs() / min_abs), and 0.0 otherwise. [so this preserves expectations, + # for those elements]. + random_val = min_abs * x.sign() * (torch.rand_like(x) * min_abs < x_abs) + return torch.where(is_too_small, random_val, x).to(torch.float16) + + +class RandomGradFunction(torch.autograd.Function): + """ + Does nothing in forward pass; in backward pass, gets rid of very small grads using + randomized approach that preserves expectations (intended to reduce roundoff). + """ + + @staticmethod + def forward(ctx, x: Tensor, min_abs: float) -> Tensor: + ctx.min_abs = min_abs + return x + + @staticmethod + def backward(ctx, ans_grad: Tensor) -> Tuple[Tensor, None]: + if ans_grad.dtype == torch.float16: + return ( + random_cast_to_half(ans_grad.to(torch.float32), min_abs=ctx.min_abs), + None, + ) + else: + return ans_grad, None + + +class RandomGrad(torch.nn.Module): + """ + Gets rid of very small gradients using an expectation-preserving method, intended to increase + accuracy of training when using amp (automatic mixed precision) + """ + + def __init__(self, min_abs: float = 5.0e-06): + super(RandomGrad, self).__init__() + self.min_abs = min_abs + + def forward(self, x: Tensor): + if torch.jit.is_scripting() or not self.training or torch.jit.is_tracing(): + return x + else: + return RandomGradFunction.apply(x, self.min_abs) + + +class SoftmaxFunction(torch.autograd.Function): + """ + Tries to handle half-precision derivatives in a randomized way that should + be more accurate for training than the default behavior. + """ + + @staticmethod + def forward(ctx, x: Tensor, dim: int): + ans = x.softmax(dim=dim) + # if x dtype is float16, x.softmax() returns a float32 because + # (presumably) that op does not support float16, and autocast + # is enabled. + if torch.is_autocast_enabled(): + ans = ans.to(torch.float16) + ctx.save_for_backward(ans) + ctx.x_dtype = x.dtype + ctx.dim = dim + return ans + + @staticmethod + def backward(ctx, ans_grad: Tensor): + (ans,) = ctx.saved_tensors + with torch.cuda.amp.autocast(enabled=False): + ans_grad = ans_grad.to(torch.float32) + ans = ans.to(torch.float32) + x_grad = ans_grad * ans + x_grad = x_grad - ans * x_grad.sum(dim=ctx.dim, keepdim=True) + return x_grad, None + + +def softmax(x: Tensor, dim: int): + if torch.jit.is_scripting() or torch.jit.is_tracing(): + return x.softmax(dim) + + return SoftmaxFunction.apply(x, dim) + + +class MaxEigLimiterFunction(torch.autograd.Function): + @staticmethod + def forward( + ctx, + x: Tensor, + coeffs: Tensor, + direction: Tensor, + channel_dim: int, + grad_scale: float, + ) -> Tensor: + ctx.channel_dim = channel_dim + ctx.grad_scale = grad_scale + ctx.save_for_backward(x.detach(), coeffs.detach(), direction.detach()) + return x + + @staticmethod + def backward(ctx, x_grad, *args): + with torch.enable_grad(): + (x_orig, coeffs, new_direction) = ctx.saved_tensors + x_orig.requires_grad = True + num_channels = x_orig.shape[ctx.channel_dim] + x = x_orig.transpose(ctx.channel_dim, -1).reshape(-1, num_channels) + new_direction.requires_grad = False + x = x - x.mean(dim=0) + x_var = (x**2).mean() + x_residual = x - coeffs * new_direction + x_residual_var = (x_residual**2).mean() + # `variance_proportion` is the proportion of the variance accounted for + # by the top eigen-direction. This is to be minimized. + variance_proportion = (x_var - x_residual_var) / (x_var + 1.0e-20) + variance_proportion.backward() + x_orig_grad = x_orig.grad + x_extra_grad = ( + x_orig.grad + * ctx.grad_scale + * x_grad.norm() + / (x_orig_grad.norm() + 1.0e-20) + ) + return x_grad + x_extra_grad.detach(), None, None, None, None + + +class BasicNorm(torch.nn.Module): + """ + This is intended to be a simpler, and hopefully cheaper, replacement for + LayerNorm. The observation this is based on, is that Transformer-type + networks, especially with pre-norm, sometimes seem to set one of the + feature dimensions to a large constant value (e.g. 50), which "defeats" + the LayerNorm because the output magnitude is then not strongly dependent + on the other (useful) features. Presumably the weight and bias of the + LayerNorm are required to allow it to do this. + + So the idea is to introduce this large constant value as an explicit + parameter, that takes the role of the "eps" in LayerNorm, so the network + doesn't have to do this trick. We make the "eps" learnable. + + Args: + num_channels: the number of channels, e.g. 512. + channel_dim: the axis/dimension corresponding to the channel, + interprted as an offset from the input's ndim if negative. + shis is NOT the num_channels; it should typically be one of + {-2, -1, 0, 1, 2, 3}. + eps: the initial "epsilon" that we add as ballast in: + scale = ((input_vec**2).mean() + epsilon)**-0.5 + Note: our epsilon is actually large, but we keep the name + to indicate the connection with conventional LayerNorm. + learn_eps: if true, we learn epsilon; if false, we keep it + at the initial value. + eps_min: float + eps_max: float + """ + + def __init__( + self, + num_channels: int, + channel_dim: int = -1, # CAUTION: see documentation. + eps: float = 0.25, + learn_eps: bool = True, + eps_min: float = -3.0, + eps_max: float = 3.0, + ) -> None: + super(BasicNorm, self).__init__() + self.num_channels = num_channels + self.channel_dim = channel_dim + if learn_eps: + self.eps = nn.Parameter(torch.tensor(eps).log().detach()) + else: + self.register_buffer("eps", torch.tensor(eps).log().detach()) + self.eps_min = eps_min + self.eps_max = eps_max + + def forward(self, x: Tensor) -> Tensor: + assert x.shape[self.channel_dim] == self.num_channels + eps = self.eps + if self.training and random.random() < 0.25: + # with probability 0.25, in training mode, clamp eps between the min + # and max; this will encourage it to learn parameters within the + # allowed range by making parameters that are outside the allowed + # range noisy. + + # gradients to allow the parameter to get back into the allowed + # region if it happens to exit it. + eps = eps.clamp(min=self.eps_min, max=self.eps_max) + scales = ( + torch.mean(x**2, dim=self.channel_dim, keepdim=True) + eps.exp() + ) ** -0.5 + return x * scales + + +def ScaledLinear(*args, initial_scale: float = 1.0, **kwargs) -> nn.Linear: + """ + Behaves like a constructor of a modified version of nn.Linear + that gives an easy way to set the default initial parameter scale. + + Args: + Accepts the standard args and kwargs that nn.Linear accepts + e.g. in_features, out_features, bias=False. + + initial_scale: you can override this if you want to increase + or decrease the initial magnitude of the module's output + (affects the initialization of weight_scale and bias_scale). + Another option, if you want to do something like this, is + to re-initialize the parameters. + """ + ans = nn.Linear(*args, **kwargs) + with torch.no_grad(): + ans.weight[:] *= initial_scale + if ans.bias is not None: + torch.nn.init.uniform_(ans.bias, -0.1 * initial_scale, 0.1 * initial_scale) + return ans + + +def ScaledConv1d(*args, initial_scale: float = 1.0, **kwargs) -> nn.Conv1d: + """ + Behaves like a constructor of a modified version of nn.Conv1d + that gives an easy way to set the default initial parameter scale. + + Args: + Accepts the standard args and kwargs that nn.Linear accepts + e.g. in_features, out_features, bias=False. + + initial_scale: you can override this if you want to increase + or decrease the initial magnitude of the module's output + (affects the initialization of weight_scale and bias_scale). + Another option, if you want to do something like this, is + to re-initialize the parameters. + """ + ans = nn.Conv1d(*args, **kwargs) + with torch.no_grad(): + ans.weight[:] *= initial_scale + if ans.bias is not None: + torch.nn.init.uniform_(ans.bias, -0.1 * initial_scale, 0.1 * initial_scale) + return ans + + +class ActivationBalancer(torch.nn.Module): + """ + Modifies the backpropped derivatives of a function to try to encourage, for + each channel, that it is positive at least a proportion `threshold` of the + time. It does this by multiplying negative derivative values by up to + (1+max_factor), and positive derivative values by up to (1-max_factor), + interpolated from 1 at the threshold to those extremal values when none + of the inputs are positive. + + Args: + num_channels: the number of channels + channel_dim: the dimension/axis corresponding to the channel, e.g. + -1, 0, 1, 2; will be interpreted as an offset from x.ndim if negative. + min_positive: the minimum, per channel, of the proportion of the time + that (x > 0), below which we start to modify the derivatives. + max_positive: the maximum, per channel, of the proportion of the time + that (x > 0), above which we start to modify the derivatives. + max_factor: the maximum factor by which we modify the derivatives for + either the sign constraint or the magnitude constraint; + e.g. with max_factor=0.02, the the derivatives would be multiplied by + values in the range [0.98..1.02]. + sign_gain_factor: determines the 'gain' with which we increase the + change in gradient once the constraints on min_positive and max_positive + are violated. + scale_gain_factor: determines the 'gain' with which we increase the + change in gradient once the constraints on min_abs and max_abs + are violated. + min_abs: the minimum average-absolute-value difference from the mean + value per channel, which we allow, before we start to modify + the derivatives to prevent this. + max_abs: the maximum average-absolute-value difference from the mean + value per channel, which we allow, before we start to modify + the derivatives to prevent this. + min_prob: determines the minimum probability with which we modify the + gradients for the {min,max}_positive and {min,max}_abs constraints, + on each forward(). This is done randomly to prevent all layers + from doing it at the same time. Early in training we may use + higher probabilities than this; it will decay to this value. + """ + + def __init__( + self, + num_channels: int, + channel_dim: int, + min_positive: float = 0.05, + max_positive: float = 0.95, + max_factor: float = 0.04, + sign_gain_factor: float = 0.01, + scale_gain_factor: float = 0.02, + min_abs: float = 0.2, + max_abs: float = 100.0, + min_prob: float = 0.1, + ): + super(ActivationBalancer, self).__init__() + self.num_channels = num_channels + self.channel_dim = channel_dim + self.min_positive = min_positive + self.max_positive = max_positive + self.max_factor = max_factor + self.min_abs = min_abs + self.max_abs = max_abs + self.min_prob = min_prob + self.sign_gain_factor = sign_gain_factor + self.scale_gain_factor = scale_gain_factor + + # count measures how many times the forward() function has been called. + # We occasionally sync this to a tensor called `count`, that exists to + # make sure it is synced to disk when we load and save the model. + self.cpu_count = 0 + self.register_buffer("count", torch.tensor(0, dtype=torch.int64)) + + def forward(self, x: Tensor) -> Tensor: + if torch.jit.is_scripting() or not x.requires_grad or torch.jit.is_tracing(): + return _no_op(x) + + count = self.cpu_count + self.cpu_count += 1 + + if random.random() < 0.01: + # Occasionally sync self.cpu_count with self.count. + # count affects the decay of 'prob'. don't do this on every iter, + # because syncing with the GPU is slow. + self.cpu_count = max(self.cpu_count, self.count.item()) + self.count.fill_(self.cpu_count) + + # the prob of doing some work exponentially decreases from 0.5 till it hits + # a floor at min_prob (==0.1, by default) + prob = max(self.min_prob, 0.5 ** (1 + (count / 4000.0))) + + if random.random() < prob: + sign_gain_factor = 0.5 + if self.min_positive != 0.0 or self.max_positive != 1.0: + sign_factor = _compute_sign_factor( + x, + self.channel_dim, + self.min_positive, + self.max_positive, + gain_factor=self.sign_gain_factor / prob, + max_factor=self.max_factor, + ) + else: + sign_factor = None + + scale_factor = _compute_scale_factor( + x.detach(), + self.channel_dim, + min_abs=self.min_abs, + max_abs=self.max_abs, + gain_factor=self.scale_gain_factor / prob, + max_factor=self.max_factor, + ) + return ActivationBalancerFunction.apply( + x, + scale_factor, + sign_factor, + self.channel_dim, + ) + else: + return _no_op(x) + + +def penalize_abs_values_gt(x: Tensor, limit: float, penalty: float) -> Tensor: + """ + Returns x unmodified, but in backprop will put a penalty for the excess of + the absolute values of elements of x over the limit "limit". E.g. if + limit == 10.0, then if x has any values over 10 it will get a penalty. + + Caution: the value of this penalty will be affected by grad scaling used + in automatic mixed precision training. For this reasons we use this, + it shouldn't really matter, or may even be helpful; we just use this + to disallow really implausible values of scores to be given to softmax. + """ + x_sign = x.sign() + over_limit = (x.abs() - limit) > 0 + # The following is a memory efficient way to penalize the absolute values of + # x that's over the limit. (The memory efficiency comes when you think + # about which items torch needs to cache for the autograd, and which ones it + # can throw away). The numerical value of aux_loss as computed here will + # actually be larger than it should be, by limit * over_limit.sum(), but it + # has the same derivative as the real aux_loss which is penalty * (x.abs() - + # limit).relu(). + aux_loss = penalty * ((x_sign * over_limit).to(torch.int8) * x) + # note: we don't do sum() here on aux)_loss, but it's as if we had done + # sum() due to how with_loss() works. + x = with_loss(x, aux_loss) + # you must use x for something, or this will be ineffective. + return x + + +def _diag(x: Tensor): # like .diag(), but works for tensors with 3 dims. + if x.ndim == 2: + return x.diag() + else: + (batch, dim, dim) = x.shape + x = x.reshape(batch, dim * dim) + x = x[:, :: dim + 1] + assert x.shape == (batch, dim) + return x + + +def _whitening_metric(x: Tensor, num_groups: int): + """ + Computes the "whitening metric", a value which will be 1.0 if all the eigenvalues of + of the centered feature covariance are the same within each group's covariance matrix + and also between groups. + Args: + x: a Tensor of shape (*, num_channels) + num_groups: the number of groups of channels, a number >=1 that divides num_channels + Returns: + Returns a scalar Tensor that will be 1.0 if the data is "perfectly white" and + greater than 1.0 otherwise. + """ + assert x.dtype != torch.float16 + x = x.reshape(-1, x.shape[-1]) + (num_frames, num_channels) = x.shape + assert num_channels % num_groups == 0 + channels_per_group = num_channels // num_groups + x = x.reshape(num_frames, num_groups, channels_per_group).transpose(0, 1) + # x now has shape (num_groups, num_frames, channels_per_group) + # subtract the mean so we use the centered, not uncentered, covariance. + # My experience has been that when we "mess with the gradients" like this, + # it's better not do anything that tries to move the mean around, because + # that can easily cause instability. + x = x - x.mean(dim=1, keepdim=True) + # x_covar: (num_groups, channels_per_group, channels_per_group) + x_covar = torch.matmul(x.transpose(1, 2), x) + x_covar_mean_diag = _diag(x_covar).mean() + # the following expression is what we'd get if we took the matrix product + # of each covariance and measured the mean of its trace, i.e. + # the same as _diag(torch.matmul(x_covar, x_covar)).mean(). + x_covarsq_mean_diag = (x_covar**2).sum() / (num_groups * channels_per_group) + # this metric will be >= 1.0; the larger it is, the less 'white' the data was. + metric = x_covarsq_mean_diag / (x_covar_mean_diag**2 + 1.0e-20) + return metric + + +class WhiteningPenaltyFunction(torch.autograd.Function): + @staticmethod + def forward( + ctx, x: Tensor, num_groups: int, whitening_limit: float, grad_scale: float + ) -> Tensor: + ctx.save_for_backward(x) + ctx.num_groups = num_groups + ctx.whitening_limit = whitening_limit + ctx.grad_scale = grad_scale + return x + + @staticmethod + def backward(ctx, x_grad: Tensor): + (x_orig,) = ctx.saved_tensors + with torch.enable_grad(): + with torch.cuda.amp.autocast(enabled=False): + x_detached = x_orig.to(torch.float32).detach() + x_detached.requires_grad = True + + metric = _whitening_metric(x_detached, ctx.num_groups) + + if random.random() < 0.005 or __name__ == "__main__": + logging.info( + f"Whitening: num_groups={ctx.num_groups}, num_channels={x_orig.shape[-1]}, " + f"metric={metric.item():.2f} vs. limit={ctx.whitening_limit}" + ) + + (metric - ctx.whitening_limit).relu().backward() + penalty_grad = x_detached.grad + scale = ctx.grad_scale * ( + x_grad.to(torch.float32).norm() / (penalty_grad.norm() + 1.0e-20) + ) + penalty_grad = penalty_grad * scale + return x_grad + penalty_grad.to(x_grad.dtype), None, None, None + + +class Whiten(nn.Module): + def __init__( + self, + num_groups: int, + whitening_limit: float, + prob: Union[float, Tuple[float, float]], + grad_scale: float, + ): + """ + Args: + num_groups: the number of groups to divide the channel dim into before + whitening. We will attempt to make the feature covariance + within each group, after mean subtraction, as "white" as possible, + while having the same trace across all groups. + whitening_limit: a value greater than 1.0, that dictates how much + freedom we have to violate the constraints. 1.0 would mean perfectly + white, with exactly the same trace across groups; larger values + give more freedom. E.g. 2.0. + prob: the probability with which we apply the gradient modification + (also affects the grad scale). May be supplied as a float, + or as a pair (min_prob, max_prob) + + grad_scale: determines the scale on the gradient term from this object, + relative to the rest of the gradient on the attention weights. + E.g. 0.02 (you may want to use smaller values than this if prob is large) + """ + super(Whiten, self).__init__() + assert num_groups >= 1 + assert whitening_limit >= 1 + assert grad_scale >= 0 + self.num_groups = num_groups + self.whitening_limit = whitening_limit + if isinstance(prob, float): + assert 0 < prob <= 1 + self.prob = prob + else: + (self.min_prob, self.max_prob) = prob + assert 0 < self.min_prob < self.max_prob <= 1 + self.prob = self.max_prob + + self.grad_scale = grad_scale + + def forward(self, x: Tensor) -> Tensor: + """ + In the forward pass, this function just returns the input unmodified. + In the backward pass, it will modify the gradients to ensure that the + distribution in each group has close to (lambda times I) as the covariance + after mean subtraction, with the same lambda across groups. + For whitening_limit > 1, there will be more freedom to violate this + constraint. + + Args: + x: the input of shape (*, num_channels) + + Returns: + x, unmodified. You should make sure + you use the returned value, or the graph will be freed + and nothing will happen in backprop. + """ + if not x.requires_grad or random.random() > self.prob or self.grad_scale == 0: + return _no_op(x) + else: + if hasattr(self, "min_prob") and random.random() < 0.25: + # occasionally switch between min_prob and max_prob, based on whether + # we are above or below the threshold. + if ( + _whitening_metric(x.to(torch.float32), self.num_groups) + > self.whitening_limit + ): + # there would be a change to the grad. + self.prob = self.max_prob + else: + self.prob = self.min_prob + + return WhiteningPenaltyFunction.apply( + x, self.num_groups, self.whitening_limit, self.grad_scale + ) + + +class WithLoss(torch.autograd.Function): + @staticmethod + def forward(ctx, x: Tensor, y: Tensor): + ctx.y_shape = y.shape + return x + + @staticmethod + def backward(ctx, ans_grad: Tensor): + return ans_grad, torch.ones( + ctx.y_shape, dtype=ans_grad.dtype, device=ans_grad.device + ) + + +def with_loss(x, y): + if torch.jit.is_scripting() or torch.jit.is_tracing(): + return x + # returns x but adds y.sum() to the loss function. + return WithLoss.apply(x, y) + + +def _no_op(x: Tensor) -> Tensor: + if torch.jit.is_scripting() or torch.jit.is_tracing(): + return x + else: + # a no-op function that will have a node in the autograd graph, + # to avoid certain bugs relating to backward hooks + return x.chunk(1, dim=-1)[0] + + +class Identity(torch.nn.Module): + def __init__(self): + super(Identity, self).__init__() + + def forward(self, x): + return _no_op(x) + + +class MaxEig(torch.nn.Module): + """ + Modifies the backpropped derivatives of a function to try to discourage + that any given direction in activation space accounts for more than + a specified proportion of the covariance (e.g. 0.2). + + + Args: + num_channels: the number of channels + channel_dim: the dimension/axis corresponding to the channel, e.g. + -1, 0, 1, 2; will be interpreted as an offset from x.ndim if negative. + max_var_per_eig: the maximum proportion of the variance of the + features/channels, after mean subtraction, that can come from + any given eigenvalue. + min_prob: the minimum probability with which we apply this during any invocation + of forward(), assuming last time we applied the constraint it was + not active; supplied for speed. + scale: determines the scale with which we modify the gradients, relative + to the existing / unmodified gradients + """ + + def __init__( + self, + num_channels: int, + channel_dim: int, + max_var_per_eig: float = 0.2, + min_prob: float = 0.01, + scale: float = 0.01, + ): + super(MaxEig, self).__init__() + self.num_channels = num_channels + self.channel_dim = channel_dim + self.scale = scale + assert max_var_per_eig == 0.0 or max_var_per_eig > 1.0 / num_channels + self.max_var_per_eig = max_var_per_eig + + # we figure out the dominant direction using the power method: starting with + # a random vector, keep multiplying by the covariance and renormalizing. + with torch.no_grad(): + # arbitrary.. would use randn() but want to leave the rest of the model's + # random parameters unchanged for comparison + direction = torch.arange(num_channels).to(torch.float) + direction = direction / direction.norm() + self.register_buffer("max_eig_direction", direction) + + self.min_prob = min_prob + # cur_prob is the current probability we'll use to apply the ActivationBalancer. + # We'll regress this towards prob, each tiem we try to apply it and it is not + # active. + self.cur_prob = 1.0 + + def forward(self, x: Tensor) -> Tensor: + if ( + torch.jit.is_scripting() + or self.max_var_per_eig <= 0 + or random.random() > self.cur_prob + or torch.jit.is_tracing() + ): + return _no_op(x) + + with torch.cuda.amp.autocast(enabled=False): + eps = 1.0e-20 + orig_x = x + x = x.to(torch.float32) + with torch.no_grad(): + x = x.transpose(self.channel_dim, -1).reshape(-1, self.num_channels) + x = x - x.mean(dim=0) + new_direction, coeffs = self._find_direction_coeffs( + x, self.max_eig_direction + ) + x_var = (x**2).mean() + x_residual = x - coeffs * new_direction + x_residual_var = (x_residual**2).mean() + + # `variance_proportion` is the proportion of the variance accounted for + # by the top eigen-direction. + variance_proportion = (x_var - x_residual_var) / (x_var + 1.0e-20) + + # ensure new direction is nonzero even if x == 0, by including `direction`. + self._set_direction(0.1 * self.max_eig_direction + new_direction) + + if random.random() < 0.01 or __name__ == "__main__": + logging.info( + f"variance_proportion = {variance_proportion.item()}, shape={tuple(orig_x.shape)}, cur_prob={self.cur_prob}" + ) + + if variance_proportion >= self.max_var_per_eig: + # The constraint is active. Note, we should quite rarely + # reach here, only near the beginning of training if we are + # starting to diverge, should this constraint be active. + cur_prob = self.cur_prob + self.cur_prob = 1.0 # next time, do the update with probability 1.0. + return MaxEigLimiterFunction.apply( + orig_x, coeffs, new_direction, self.channel_dim, self.scale + ) + else: + # let self.cur_prob exponentially approach self.min_prob, as + # long as the constraint is inactive. + self.cur_prob = 0.75 * self.cur_prob + 0.25 * self.min_prob + return orig_x + + def _set_direction(self, direction: Tensor): + """ + Sets self.max_eig_direction to a normalized version of `direction` + """ + direction = direction.detach() + direction = direction / direction.norm() + direction_sum = direction.sum().item() + if direction_sum - direction_sum == 0: # no inf/nan + self.max_eig_direction[:] = direction + else: + logging.info( + f"Warning: sum of direction in MaxEig is {direction_sum}, " + "num_channels={self.num_channels}, channel_dim={self.channel_dim}" + ) + + def _find_direction_coeffs( + self, x: Tensor, prev_direction: Tensor + ) -> Tuple[Tensor, Tensor, Tensor]: + """ + Figure out (an approximation to) the proportion of the variance of a set of + feature vectors that can be attributed to the top eigen-direction. + Args: + x: a Tensor of shape (num_frames, num_channels), with num_frames > 1. + prev_direction: a Tensor of shape (num_channels,), that is our previous estimate + of the top eigen-direction, or a random direction if this is the first + iteration. Does not have to be normalized, but should be nonzero. + + Returns: (cur_direction, coeffs), where: + cur_direction: a Tensor of shape (num_channels,) that is the current + estimate of the top eigen-direction. + coeffs: a Tensor of shape (num_frames, 1) that minimizes, or + approximately minimizes, (x - coeffs * cur_direction).norm() + """ + (num_frames, num_channels) = x.shape + assert num_channels > 1 and num_frames > 1 + assert prev_direction.shape == (num_channels,) + # `coeffs` are the coefficients of `prev_direction` in x. + # actually represent the coeffs up to a constant positive factor. + coeffs = (x * prev_direction).sum(dim=1, keepdim=True) + 1.0e-10 + cur_direction = (x * coeffs).sum(dim=0) / ((coeffs**2).sum() + 1.0e-20) + return cur_direction, coeffs + + +class DoubleSwishFunction(torch.autograd.Function): + """ + double_swish(x) = x * torch.sigmoid(x-1) + This is a definition, originally motivated by its close numerical + similarity to swish(swish(x)), where swish(x) = x * sigmoid(x). + + Memory-efficient derivative computation: + double_swish(x) = x * s, where s(x) = torch.sigmoid(x-1) + double_swish'(x) = d/dx double_swish(x) = x * s'(x) + x' * s(x) = x * s'(x) + s(x). + Now, s'(x) = s(x) * (1-s(x)). + double_swish'(x) = x * s'(x) + s(x). + = x * s(x) * (1-s(x)) + s(x). + = double_swish(x) * (1-s(x)) + s(x) + ... so we just need to remember s(x) but not x itself. + """ + + @staticmethod + def forward(ctx, x: Tensor) -> Tensor: + requires_grad = x.requires_grad + x_dtype = x.dtype + if x.dtype == torch.float16: + x = x.to(torch.float32) + + s = torch.sigmoid(x - 1.0) + y = x * s + + if requires_grad: + deriv = y * (1 - s) + s + # notes on derivative of x * sigmoid(x - 1): + # https://www.wolframalpha.com/input?i=d%2Fdx+%28x+*+sigmoid%28x-1%29%29 + # min \simeq -0.043638. Take floor as -0.043637 so it's a lower bund + # max \simeq 1.1990. Take ceil to be 1.2 so it's an upper bound. + # the combination of "+ torch.rand_like(deriv)" and casting to torch.uint8 (which + # floors), should be expectation-preserving. + floor = -0.043637 + ceil = 1.2 + d_scaled = (deriv - floor) * (255.0 / (ceil - floor)) + torch.rand_like( + deriv + ) + if __name__ == "__main__": + # for self-testing only. + assert d_scaled.min() >= 0.0 + assert d_scaled.max() < 256.0 + d_int = d_scaled.to(torch.uint8) + ctx.save_for_backward(d_int) + if x.dtype == torch.float16 or torch.is_autocast_enabled(): + y = y.to(torch.float16) + return y + + @staticmethod + def backward(ctx, y_grad: Tensor) -> Tensor: + (d,) = ctx.saved_tensors + # the same constants as used in forward pass. + floor = -0.043637 + ceil = 1.2 + d = d * ((ceil - floor) / 255.0) + floor + return y_grad * d + + +class DoubleSwish(torch.nn.Module): + def forward(self, x: Tensor) -> Tensor: + """Return double-swish activation function which is an approximation to Swish(Swish(x)), + that we approximate closely with x * sigmoid(x-1). + """ + if torch.jit.is_scripting() or torch.jit.is_tracing(): + return x * torch.sigmoid(x - 1.0) + return DoubleSwishFunction.apply(x) + + +def _test_max_eig(): + for proportion in [0.1, 0.5, 10.0]: + logging.info(f"proportion = {proportion}") + x = torch.randn(100, 128) + direction = torch.randn(128) + coeffs = torch.randn(100, 1) + x += proportion * direction * coeffs + + x.requires_grad = True + + num_channels = 128 + m = MaxEig( + num_channels, 1, 0.5, scale=0.1 # channel_dim # max_var_per_eig + ) # grad_scale + + for _ in range(4): + y = m(x) + + y_grad = torch.randn_like(x) + y.backward(gradient=y_grad) + + if proportion < 0.2: + assert torch.allclose(x.grad, y_grad, atol=1.0e-02) + elif proportion > 1.0: + assert not torch.allclose(x.grad, y_grad) + + +def _test_whiten(): + for proportion in [0.1, 0.5, 10.0]: + logging.info(f"_test_whiten(): proportion = {proportion}") + x = torch.randn(100, 128) + direction = torch.randn(128) + coeffs = torch.randn(100, 1) + x += proportion * direction * coeffs + + x.requires_grad = True + + num_channels = 128 + m = Whiten( + 1, 5.0, prob=1.0, grad_scale=0.1 # num_groups # whitening_limit, + ) # grad_scale + + for _ in range(4): + y = m(x) + + y_grad = torch.randn_like(x) + y.backward(gradient=y_grad) + + if proportion < 0.2: + assert torch.allclose(x.grad, y_grad) + elif proportion > 1.0: + assert not torch.allclose(x.grad, y_grad) + + +def _test_activation_balancer_sign(): + probs = torch.arange(0, 1, 0.01) + N = 1000 + x = 1.0 * ((2.0 * (torch.rand(probs.numel(), N) < probs.unsqueeze(-1))) - 1.0) + x = x.detach() + x.requires_grad = True + m = ActivationBalancer( + probs.numel(), + channel_dim=0, + min_positive=0.05, + max_positive=0.95, + max_factor=0.2, + min_abs=0.0, + ) + + y_grad = torch.sign(torch.randn(probs.numel(), N)) + + y = m(x) + y.backward(gradient=y_grad) + print("_test_activation_balancer_sign: x = ", x) + print("_test_activation_balancer_sign: y grad = ", y_grad) + print("_test_activation_balancer_sign: x grad = ", x.grad) + + +def _test_activation_balancer_magnitude(): + magnitudes = torch.arange(0, 1, 0.01) + N = 1000 + x = torch.sign(torch.randn(magnitudes.numel(), N)) * magnitudes.unsqueeze(-1) + x = x.detach() + x.requires_grad = True + m = ActivationBalancer( + magnitudes.numel(), + channel_dim=0, + min_positive=0.0, + max_positive=1.0, + max_factor=0.2, + min_abs=0.2, + max_abs=0.8, + min_prob=1.0, + ) + + y_grad = torch.sign(torch.randn(magnitudes.numel(), N)) + + y = m(x) + y.backward(gradient=y_grad) + print("_test_activation_balancer_magnitude: x = ", x) + print("_test_activation_balancer_magnitude: y grad = ", y_grad) + print("_test_activation_balancer_magnitude: x grad = ", x.grad) + + +def _test_basic_norm(): + num_channels = 128 + m = BasicNorm(num_channels=num_channels, channel_dim=1) + + x = torch.randn(500, num_channels) + + y = m(x) + + assert y.shape == x.shape + x_rms = (x**2).mean().sqrt() + y_rms = (y**2).mean().sqrt() + print("x rms = ", x_rms) + print("y rms = ", y_rms) + assert y_rms < x_rms + assert y_rms > 0.5 * x_rms + + +def _test_double_swish_deriv(): + x = torch.randn(10, 12, dtype=torch.double) * 3.0 + x.requires_grad = True + m = DoubleSwish() + + tol = (1.2 - (-0.043637)) / 255.0 + torch.autograd.gradcheck(m, x, atol=tol) + + # for self-test. + x = torch.randn(1000, 1000, dtype=torch.double) * 3.0 + x.requires_grad = True + y = m(x) + + +def _test_softmax(): + a = torch.randn(2, 10, dtype=torch.float64) + b = a.clone() + a.requires_grad = True + b.requires_grad = True + a.softmax(dim=1)[:, 0].sum().backward() + print("a grad = ", a.grad) + softmax(b, dim=1)[:, 0].sum().backward() + print("b grad = ", b.grad) + assert torch.allclose(a.grad, b.grad) + + +if __name__ == "__main__": + logging.getLogger().setLevel(logging.INFO) + torch.set_num_threads(1) + torch.set_num_interop_threads(1) + _test_softmax() + _test_whiten() + _test_max_eig() + _test_activation_balancer_sign() + _test_activation_balancer_magnitude() + _test_basic_norm() + _test_double_swish_deriv() diff --git a/egs/librispeech/ASR/zipformer_ctc_streaming/scaling_converter.py b/egs/librispeech/ASR/zipformer_ctc_streaming/scaling_converter.py new file mode 100644 index 000000000..56165d1f9 --- /dev/null +++ b/egs/librispeech/ASR/zipformer_ctc_streaming/scaling_converter.py @@ -0,0 +1,114 @@ +# Copyright 2022 Xiaomi Corp. (authors: Fangjun Kuang) +# +# See ../../../../LICENSE for clarification regarding multiple authors +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +""" +This file replaces various modules in a model. +Specifically, ActivationBalancer is replaced with an identity operator; +Whiten is also replaced with an identity operator; +BasicNorm is replaced by a module with `exp` removed. +""" + +import copy +from typing import List + +import torch +import torch.nn as nn +from scaling import ActivationBalancer, BasicNorm, Whiten + + +class NonScaledNorm(nn.Module): + """See BasicNorm for doc""" + + def __init__( + self, + num_channels: int, + eps_exp: float, + channel_dim: int = -1, # CAUTION: see documentation. + ): + super().__init__() + self.num_channels = num_channels + self.channel_dim = channel_dim + self.eps_exp = eps_exp + + def forward(self, x: torch.Tensor) -> torch.Tensor: + if not torch.jit.is_tracing(): + assert x.shape[self.channel_dim] == self.num_channels + scales = ( + torch.mean(x * x, dim=self.channel_dim, keepdim=True) + self.eps_exp + ).pow(-0.5) + return x * scales + + +def convert_basic_norm(basic_norm: BasicNorm) -> NonScaledNorm: + assert isinstance(basic_norm, BasicNorm), type(BasicNorm) + norm = NonScaledNorm( + num_channels=basic_norm.num_channels, + eps_exp=basic_norm.eps.data.exp().item(), + channel_dim=basic_norm.channel_dim, + ) + return norm + + +# Copied from https://pytorch.org/docs/1.9.0/_modules/torch/nn/modules/module.html#Module.get_submodule # noqa +# get_submodule was added to nn.Module at v1.9.0 +def get_submodule(model, target): + if target == "": + return model + atoms: List[str] = target.split(".") + mod: torch.nn.Module = model + for item in atoms: + if not hasattr(mod, item): + raise AttributeError( + mod._get_name() + " has no " "attribute `" + item + "`" + ) + mod = getattr(mod, item) + if not isinstance(mod, torch.nn.Module): + raise AttributeError("`" + item + "` is not " "an nn.Module") + return mod + + +def convert_scaled_to_non_scaled( + model: nn.Module, + inplace: bool = False, +): + """ + Args: + model: + The model to be converted. + inplace: + If True, the input model is modified inplace. + If False, the input model is copied and we modify the copied version. + Return: + Return a model without scaled layers. + """ + if not inplace: + model = copy.deepcopy(model) + + d = {} + for name, m in model.named_modules(): + if isinstance(m, BasicNorm): + d[name] = convert_basic_norm(m) + elif isinstance(m, (ActivationBalancer, Whiten)): + d[name] = nn.Identity() + + for k, v in d.items(): + if "." in k: + parent, child = k.rsplit(".", maxsplit=1) + setattr(get_submodule(model, parent), child, v) + else: + setattr(model, k, v) + + return model diff --git a/egs/librispeech/ASR/zipformer_ctc_streaming/subsampling.py b/egs/librispeech/ASR/zipformer_ctc_streaming/subsampling.py new file mode 100644 index 000000000..8e0f73d05 --- /dev/null +++ b/egs/librispeech/ASR/zipformer_ctc_streaming/subsampling.py @@ -0,0 +1,153 @@ +# 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 torch +import torch.nn as nn + + +class Conv2dSubsampling(nn.Module): + """Convolutional 2D subsampling (to 1/4 length). + + Convert an input of shape (N, T, idim) to an output + with shape (N, T', odim), where + T' = ((T-1)//2 - 1)//2, which approximates T' == T//4 + + It is based on + https://github.com/espnet/espnet/blob/master/espnet/nets/pytorch_backend/transformer/subsampling.py # noqa + """ + + def __init__(self, idim: int, odim: int) -> None: + """ + Args: + idim: + Input dim. The input shape is (N, T, idim). + Caution: It requires: T >=7, idim >=7 + odim: + Output dim. The output shape is (N, ((T-1)//2 - 1)//2, odim) + """ + assert idim >= 7 + super().__init__() + self.conv = nn.Sequential( + nn.Conv2d(in_channels=1, out_channels=odim, kernel_size=3, stride=2), + nn.ReLU(), + nn.Conv2d(in_channels=odim, out_channels=odim, kernel_size=3, stride=2), + nn.ReLU(), + ) + self.out = nn.Linear(odim * (((idim - 1) // 2 - 1) // 2), odim) + + def forward(self, x: torch.Tensor) -> torch.Tensor: + """Subsample x. + + Args: + x: + Its shape is (N, T, idim). + + Returns: + Return a tensor of shape (N, ((T-1)//2 - 1)//2, odim) + """ + # On entry, x is (N, T, idim) + x = x.unsqueeze(1) # (N, T, idim) -> (N, 1, T, idim) i.e., (N, C, H, W) + x = self.conv(x) + # Now x is of shape (N, odim, ((T-1)//2 - 1)//2, ((idim-1)//2 - 1)//2) + b, c, t, f = x.size() + x = self.out(x.transpose(1, 2).contiguous().view(b, t, c * f)) + # Now x is of shape (N, ((T-1)//2 - 1))//2, odim) + return x + + +class VggSubsampling(nn.Module): + """Trying to follow the setup described in the following paper: + https://arxiv.org/pdf/1910.09799.pdf + + This paper is not 100% explicit so I am guessing to some extent, + and trying to compare with other VGG implementations. + + Convert an input of shape (N, T, idim) to an output + with shape (N, T', odim), where + T' = ((T-1)//2 - 1)//2, which approximates T' = T//4 + """ + + def __init__(self, idim: int, odim: int) -> None: + """Construct a VggSubsampling object. + + This uses 2 VGG blocks with 2 Conv2d layers each, + subsampling its input by a factor of 4 in the time dimensions. + + Args: + idim: + Input dim. The input shape is (N, T, idim). + Caution: It requires: T >=7, idim >=7 + odim: + Output dim. The output shape is (N, ((T-1)//2 - 1)//2, odim) + """ + super().__init__() + + cur_channels = 1 + layers = [] + block_dims = [32, 64] + + # The decision to use padding=1 for the 1st convolution, then padding=0 + # for the 2nd and for the max-pooling, and ceil_mode=True, was driven by + # a back-compatibility concern so that the number of frames at the + # output would be equal to: + # (((T-1)//2)-1)//2. + # We can consider changing this by using padding=1 on the + # 2nd convolution, so the num-frames at the output would be T//4. + for block_dim in block_dims: + layers.append( + torch.nn.Conv2d( + in_channels=cur_channels, + out_channels=block_dim, + kernel_size=3, + padding=1, + stride=1, + ) + ) + layers.append(torch.nn.ReLU()) + layers.append( + torch.nn.Conv2d( + in_channels=block_dim, + out_channels=block_dim, + kernel_size=3, + padding=0, + stride=1, + ) + ) + layers.append( + torch.nn.MaxPool2d(kernel_size=2, stride=2, padding=0, ceil_mode=True) + ) + cur_channels = block_dim + + self.layers = nn.Sequential(*layers) + + self.out = nn.Linear(block_dims[-1] * (((idim - 1) // 2 - 1) // 2), odim) + + def forward(self, x: torch.Tensor) -> torch.Tensor: + """Subsample x. + + Args: + x: + Its shape is (N, T, idim). + + Returns: + Return a tensor of shape (N, ((T-1)//2 - 1)//2, odim) + """ + x = x.unsqueeze(1) + x = self.layers(x) + b, c, t, f = x.size() + x = self.out(x.transpose(1, 2).contiguous().view(b, t, c * f)) + return x diff --git a/egs/librispeech/ASR/zipformer_ctc_streaming/train.py b/egs/librispeech/ASR/zipformer_ctc_streaming/train.py new file mode 100755 index 000000000..4019ee87c --- /dev/null +++ b/egs/librispeech/ASR/zipformer_ctc_streaming/train.py @@ -0,0 +1,1167 @@ +#!/usr/bin/env python3 +# Copyright 2021 Xiaomi Corp. (authors: Fangjun Kuang, +# Wei Kang +# Mingshuang Luo) +# +# See ../../../../LICENSE for clarification regarding multiple authors +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +""" +Usage: + export CUDA_VISIBLE_DEVICES="0,1,2,3" + ./zipformer_ctc/train.py \ + --exp-dir ./zipformer_ctc/exp \ + --world-size 4 \ + --max-duration 500 \ + --num-epochs 30 +""" + +import argparse +import copy +import logging +from pathlib import Path +from shutil import copyfile +from typing import Any, Dict, Optional, Tuple, Union + +import torch +import torch.multiprocessing as mp +import torch.nn as nn +from asr_datamodule import LibriSpeechAsrDataModule +from decoder import Decoder +from lhotse.cut import Cut +from lhotse.dataset.sampling.base import CutSampler +from lhotse.utils import fix_random_seed +from model import CTCModel +from optim import Eden, LRScheduler, ScaledAdam +from torch import Tensor +from torch.cuda.amp import GradScaler +from torch.nn.parallel import DistributedDataParallel as DDP +from torch.nn.utils import clip_grad_norm_ +from torch.utils.tensorboard import SummaryWriter +from zipformer import Zipformer + +from icefall import diagnostics +from icefall.bpe_graph_compiler import BpeCtcTrainingGraphCompiler +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.lexicon import Lexicon +from icefall.utils import AttributeDict, MetricsTracker, setup_logger, str2bool + +LRSchedulerType = Union[torch.optim.lr_scheduler._LRScheduler, LRScheduler] + + +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 module in model.modules(): + if hasattr(module, "batch_count"): + module.batch_count = batch_count + + +def add_model_arguments(parser: argparse.ArgumentParser): + parser.add_argument( + "--num-encoder-layers", + type=str, + default="2,4,3,2,4", + help="Number of zipformer encoder layers, comma separated.", + ) + + parser.add_argument( + "--feedforward-dims", + type=str, + default="1024,1024,2048,2048,1024", + help="Feedforward dimension of the zipformer encoder layers, comma separated.", + ) + + parser.add_argument( + "--nhead", + type=str, + default="8,8,8,8,8", + help="Number of attention heads in the zipformer encoder layers.", + ) + + parser.add_argument( + "--encoder-dims", + type=str, + default="384,384,384,384,384", + help="Embedding dimension in the 2 blocks of zipformer encoder layers, comma separated", + ) + + parser.add_argument( + "--attention-dims", + type=str, + default="192,192,192,192,192", + help="""Attention dimension in the 2 blocks of zipformer encoder layers, comma separated; + not the same as embedding dimension.""", + ) + + parser.add_argument( + "--encoder-unmasked-dims", + type=str, + default="256,256,256,256,256", + help="Unmasked dimensions in the encoders, relates to augmentation during training. " + "Must be <= each of encoder_dims. Empirically, less than 256 seems to make performance " + " worse.", + ) + + parser.add_argument( + "--zipformer-downsampling-factors", + type=str, + default="1,2,4,8,2", + help="Downsampling factor for each stack of encoder layers.", + ) + + parser.add_argument( + "--cnn-module-kernels", + type=str, + default="31,31,31,31,31", + help="Sizes of kernels in convolution modules", + ) + + parser.add_argument( + "--num-decoder-layers", + type=int, + default=6, + help="""Number of decoder layer of transformer decoder. + Setting this to 0 will not create the decoder at all (pure CTC model) + """, + ) + + parser.add_argument( + "--short-chunk-size", + type=int, + default=50, + help="""Chunk length of dynamic training, the chunk size would be either + max sequence length of current batch or uniformly sampled from (1, short_chunk_size). + """, + ) + + parser.add_argument( + "--num-left-chunks", + type=int, + default=4, + help="How many left context can be seen in chunks when calculating attention.", + ) + + parser.add_argument( + "--decode-chunk-len", + type=int, + default=32, + help="The chunk size for decoding (in frames before subsampling)", + ) + + +def get_parser(): + parser = argparse.ArgumentParser( + formatter_class=argparse.ArgumentDefaultsHelpFormatter + ) + + parser.add_argument( + "--world-size", + type=int, + default=1, + help="Number of GPUs for DDP training.", + ) + + parser.add_argument( + "--master-port", + type=int, + default=12354, + help="Master port to use for DDP training.", + ) + + parser.add_argument( + "--tensorboard", + type=str2bool, + default=True, + help="Should various information be logged in tensorboard.", + ) + + parser.add_argument( + "--num-epochs", + type=int, + default=30, + help="Number of epochs to train.", + ) + + parser.add_argument( + "--start-epoch", + type=int, + default=1, + help="""Resume training from this epoch. It should be positive. + If larger than 1, it will load checkpoint from + exp-dir/epoch-{start_epoch-1}.pt + """, + ) + + parser.add_argument( + "--start-batch", + type=int, + default=0, + help="""If positive, --start-epoch is ignored and + it loads the checkpoint from exp-dir/checkpoint-{start_batch}.pt + """, + ) + + parser.add_argument( + "--exp-dir", + type=str, + default="conformer_ctc/exp", + help="""The experiment dir. + It specifies the directory where all training related + files, e.g., checkpoints, log, etc, are saved + """, + ) + + parser.add_argument( + "--lang-dir", + type=str, + default="data/lang_bpe_500", + help="""The lang dir + It contains language related input files such as + "lexicon.txt" + """, + ) + + parser.add_argument( + "--base-lr", type=float, default=0.05, help="The base learning rate." + ) + + parser.add_argument( + "--lr-batches", + type=float, + default=5000, + help="""Number of steps that affects how rapidly the learning rate + decreases. We suggest not to change this.""", + ) + + parser.add_argument( + "--lr-epochs", + type=float, + default=3.5, + help="""Number of epochs that affects how rapidly the learning rate decreases. + """, + ) + + parser.add_argument( + "--att-rate", + type=float, + default=0.8, + help="""The attention rate. + The total loss is (1 - att_rate) * ctc_loss + att_rate * att_loss + """, + ) + + parser.add_argument( + "--seed", + type=int, + default=42, + help="The seed for random generators intended for reproducibility", + ) + + parser.add_argument( + "--print-diagnostics", + type=str2bool, + default=False, + help="Accumulate stats on activations, print them and exit.", + ) + + parser.add_argument( + "--inf-check", + type=str2bool, + default=False, + help="Add hooks to check for infinite module outputs and gradients.", + ) + + parser.add_argument( + "--save-every-n", + type=int, + default=2000, + help="""Save checkpoint after processing this number of batches" + periodically. We save checkpoint to exp-dir/ whenever + params.batch_idx_train % save_every_n == 0. The checkpoint filename + has the form: f'exp-dir/checkpoint-{params.batch_idx_train}.pt' + Note: It also saves checkpoint to `exp-dir/epoch-xxx.pt` at the + end of each epoch where `xxx` is the epoch number counting from 0. + """, + ) + + parser.add_argument( + "--keep-last-k", + type=int, + default=30, + help="""Only keep this number of checkpoints on disk. + For instance, if it is 3, there are only 3 checkpoints + in the exp-dir with filenames `checkpoint-xxx.pt`. + It does not affect checkpoints with name `epoch-xxx.pt`. + """, + ) + + parser.add_argument( + "--average-period", + type=int, + default=200, + help="""Update the averaged model, namely `model_avg`, after processing + this number of batches. `model_avg` is a separate version of model, + in which each floating-point parameter is the average of all the + parameters from the start of training. Each time we take the average, + we do: `model_avg = model * (average_period / batch_idx_train) + + model_avg * ((batch_idx_train - average_period) / batch_idx_train)`. + """, + ) + + parser.add_argument( + "--use-fp16", + type=str2bool, + default=False, + help="Whether to use half precision training.", + ) + + add_model_arguments(parser) + + return parser + + +def get_params() -> AttributeDict: + """Return a dict containing training parameters. + + All training related parameters that are not passed from the commandline + are saved in the variable `params`. + + Commandline options are merged into `params` after they are parsed, so + you can also access them via `params`. + + Explanation of options saved in `params`: + + - best_train_loss: Best training loss so far. It is used to select + the model that has the lowest training loss. It is + updated during the training. + + - best_valid_loss: Best validation loss so far. It is used to select + the model that has the lowest validation loss. It is + updated during the training. + + - best_train_epoch: It is the epoch that has the best training loss. + + - best_valid_epoch: It is the epoch that has the best validation loss. + + - batch_idx_train: Used to writing statistics to tensorboard. It + contains number of batches trained so far across + epochs. + + - log_interval: Print training loss if batch_idx % log_interval` is 0 + + - reset_interval: Reset statistics if batch_idx % reset_interval is 0 + + - valid_interval: Run validation if batch_idx % valid_interval is 0 + + - feature_dim: The model input dim. It has to match the one used + in computing features. + + - subsampling_factor: The subsampling factor for the model. + + - use_feat_batchnorm: Normalization for the input features, can be a + boolean indicating whether to do batch + normalization, or a float which means just scaling + the input features with this float value. + If given a float value, we will remove batchnorm + layer in `ConvolutionModule` as well. + + - attention_dim: Hidden dim for multi-head attention model. + + - head: Number of heads of multi-head attention model. + + - num_decoder_layers: Number of decoder layer of transformer decoder. + + - beam_size: It is used in k2.ctc_loss + + - reduction: It is used in k2.ctc_loss + + - use_double_scores: It is used in k2.ctc_loss + + - weight_decay: The weight_decay for the optimizer. + + - warm_step: The warm_step for Noam optimizer. + """ + params = AttributeDict( + { + "frame_shift_ms": 10.0, + "allowed_excess_duration_ratio": 0.1, + "best_train_loss": float("inf"), + "best_valid_loss": float("inf"), + "best_train_epoch": -1, + "best_valid_epoch": -1, + "batch_idx_train": 0, + "log_interval": 50, + "reset_interval": 200, + "valid_interval": 3000, # For the 100h subset, use 800 + # parameters for zipformer + "feature_dim": 80, + "subsampling_factor": 4, # not passed in, this is fixed. + "warm_step": 2000, + "env_info": get_env_info(), + # parameters for loss + "beam_size": 10, + "reduction": "sum", + "use_double_scores": True, + # parameters for decoding + "search_beam": 20, + "output_beam": 8, + "min_active_states": 30, + "max_active_states": 10000, + } + ) + + return params + + +def get_encoder_model(params: AttributeDict) -> nn.Module: + # TODO: We can add an option to switch between Zipformer and Transformer + def to_int_tuple(s: str): + return tuple(map(int, s.split(","))) + + encoder = Zipformer( + num_features=params.feature_dim, + output_downsampling_factor=2, + zipformer_downsampling_factors=to_int_tuple( + params.zipformer_downsampling_factors + ), + encoder_dims=to_int_tuple(params.encoder_dims), + attention_dim=to_int_tuple(params.attention_dims), + encoder_unmasked_dims=to_int_tuple(params.encoder_unmasked_dims), + nhead=to_int_tuple(params.nhead), + feedforward_dim=to_int_tuple(params.feedforward_dims), + cnn_module_kernels=to_int_tuple(params.cnn_module_kernels), + num_encoder_layers=to_int_tuple(params.num_encoder_layers), + num_left_chunks=params.num_left_chunks, + short_chunk_size=params.short_chunk_size, + decode_chunk_size=params.decode_chunk_len // 2, + ) + return encoder + + +def get_decoder_model(params: AttributeDict) -> nn.Module: + decoder = Decoder( + num_layers=params.num_decoder_layers, + num_classes=params.vocab_size, + d_model=int(params.encoder_dims.split(",")[-1]), + ) + return decoder + + +def get_ctc_model(params: AttributeDict) -> nn.Module: + encoder = get_encoder_model(params) + decoder = get_decoder_model(params) + + model = CTCModel( + encoder=encoder, + decoder=decoder, + encoder_dim=int(params.encoder_dims.split(",")[-1]), + vocab_size=params.vocab_size, + ) + return model + + +def load_checkpoint_if_available( + params: AttributeDict, + model: nn.Module, + model_avg: nn.Module = None, + optimizer: Optional[torch.optim.Optimizer] = None, + scheduler: Optional[LRSchedulerType] = None, +) -> Optional[Dict[str, Any]]: + """Load checkpoint from file. + + If params.start_batch is positive, it will load the checkpoint from + `params.exp_dir/checkpoint-{params.start_batch}.pt`. Otherwise, if + params.start_epoch is larger than 1, it will load the checkpoint from + `params.start_epoch - 1`. + + Apart from loading state dict for `model` and `optimizer` it also updates + `best_train_epoch`, `best_train_loss`, `best_valid_epoch`, + and `best_valid_loss` in `params`. + + Args: + params: + The return value of :func:`get_params`. + model: + The training model. + model_avg: + The stored model averaged from the start of training. + optimizer: + The optimizer that we are using. + scheduler: + The scheduler that we are using. + Returns: + Return a dict containing previously saved training info. + """ + if params.start_batch > 0: + filename = params.exp_dir / f"checkpoint-{params.start_batch}.pt" + elif params.start_epoch > 1: + filename = params.exp_dir / f"epoch-{params.start_epoch-1}.pt" + else: + return None + + assert filename.is_file(), f"{filename} does not exist!" + + saved_params = load_checkpoint( + filename, + model=model, + model_avg=model_avg, + optimizer=optimizer, + scheduler=scheduler, + ) + + keys = [ + "best_train_epoch", + "best_valid_epoch", + "batch_idx_train", + "best_train_loss", + "best_valid_loss", + ] + for k in keys: + params[k] = saved_params[k] + + if params.start_batch > 0: + if "cur_epoch" in saved_params: + params["start_epoch"] = saved_params["cur_epoch"] + + if "cur_batch_idx" in saved_params: + params["cur_batch_idx"] = saved_params["cur_batch_idx"] + + 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: nn.Module, + batch: dict, + graph_compiler: BpeCtcTrainingGraphCompiler, + is_training: bool, +) -> Tuple[Tensor, MetricsTracker]: + """ + Compute CTC loss given the model and its inputs. + + Args: + params: + Parameters for training. See :func:`get_params`. + model: + The model for training. It is an instance of Conformer in our case. + batch: + A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()` + for the content in it. + graph_compiler: + It is used to build a decoding graph from a ctc topo and training + transcript. The training transcript is contained in the given `batch`, + while the ctc topo is built when this compiler is instantiated. + 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 = graph_compiler.device + feature = batch["inputs"] + # at entry, feature is (N, T, C) + assert feature.ndim == 3 + feature = feature.to(device) + + supervisions = batch["supervisions"] + feature_lens = supervisions["num_frames"].to(device) + with torch.set_grad_enabled(is_training): + ctc_loss, att_loss, tot_frames = model( + feature, + feature_lens, + supervisions, + graph_compiler, + subsampling_factor=params.subsampling_factor, + beam_size=params.beam_size, + reduction=params.reduction, + use_double_scores=params.use_double_scores, + ) + + info = MetricsTracker() + info["frames"] = tot_frames + info["ctc_loss"] = ctc_loss.detach().cpu().item() + if params.att_rate != 0.0: + info["att_loss"] = att_loss.detach().cpu().item() + + loss = (1.0 - params.att_rate) * ctc_loss + params.att_rate * att_loss + assert loss.requires_grad == is_training, f"{loss.requires_grad} != {is_training}" + info["loss"] = loss.detach().cpu().item() + + # `utt_duration` and `utt_pad_proportion` would be normalized by `utterances` # noqa + info["utterances"] = feature.size(0) + # averaged input duration in frames over utterances + info["utt_duration"] = supervisions["num_frames"].sum().item() + # averaged padding proportion over utterances + info["utt_pad_proportion"] = ( + ((feature.size(1) - supervisions["num_frames"]) / feature.size(1)).sum().item() + ) + + return loss, info + + +def compute_validation_loss( + params: AttributeDict, + model: nn.Module, + graph_compiler: BpeCtcTrainingGraphCompiler, + 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, + graph_compiler=graph_compiler, + 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, + graph_compiler: BpeCtcTrainingGraphCompiler, + 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. + graph_compiler: + It is used to convert transcripts to FSAs. + train_dl: + Dataloader for the training dataset. + valid_dl: + Dataloader for the validation dataset. + tb_writer: + Writer to write log messages to tensorboard. + world_size: + Number of nodes in DDP training. If it is 1, DDP is disabled. + """ + model.train() + + tot_loss = MetricsTracker() + + for batch_idx, batch in enumerate(train_dl): + params.batch_idx_train += 1 + batch_size = len(batch["supervisions"]["text"]) + + try: + with torch.cuda.amp.autocast(enabled=params.use_fp16): + loss, loss_info = compute_loss( + params=params, + model=model, + batch=batch, + graph_compiler=graph_compiler, + is_training=True, + ) + # summary stats + tot_loss = (tot_loss * (1 - 1 / params.reset_interval)) + loss_info + + # NOTE: We use reduction==sum and loss is computed over utterances + # in the batch and there is no normalization to it so far. + scaler.scale(loss).backward() + set_batch_count(model, params.batch_idx_train) + scheduler.step_batch(params.batch_idx_train) + + scaler.step(optimizer) + scaler.update() + optimizer.zero_grad() + except: # noqa + display_and_save_batch(batch, params=params, graph_compiler=graph_compiler) + 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 + ): + params.cur_batch_idx = batch_idx + 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, + ) + del params.cur_batch_idx + 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 < 1.0 or (cur_grad_scale < 8.0 and batch_idx % 400 == 0): + scaler.update(cur_grad_scale * 2.0) + if cur_grad_scale < 0.01: + logging.warning(f"Grad scale is small: {cur_grad_scale}") + if cur_grad_scale < 1.0e-05: + raise RuntimeError( + f"grad_scale is too small, exiting: {cur_grad_scale}" + ) + + if batch_idx % params.log_interval == 0: + cur_lr = scheduler.get_last_lr()[0] + 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, + graph_compiler=graph_compiler, + valid_dl=valid_dl, + world_size=world_size, + ) + model.train() + logging.info(f"Epoch {params.cur_epoch}, validation: {valid_info}") + logging.info( + f"Maximum memory allocated so far is {torch.cuda.max_memory_allocated()//1000000}MB" + ) + if tb_writer is not None: + valid_info.write_summary( + tb_writer, "train/valid_", params.batch_idx_train + ) + + loss_value = tot_loss["loss"] / tot_loss["frames"] + params.train_loss = loss_value + if params.train_loss < params.best_train_loss: + params.best_train_epoch = params.cur_epoch + params.best_train_loss = params.train_loss + + +def run(rank, world_size, args): + """ + Args: + rank: + It is a value between 0 and `world_size-1`, which is + passed automatically by `mp.spawn()` in :func:`main`. + The node with rank 0 is responsible for saving checkpoint. + world_size: + Number of GPUs for DDP training. + args: + The return value of get_parser().parse_args() + """ + params = get_params() + params.update(vars(args)) + + fix_random_seed(params.seed) + if world_size > 1: + setup_dist(rank, world_size, params.master_port) + + setup_logger(f"{params.exp_dir}/log/log-train") + logging.info("Training started") + logging.info(params) + + if args.tensorboard and rank == 0: + tb_writer = SummaryWriter(log_dir=f"{params.exp_dir}/tensorboard") + else: + tb_writer = None + + lexicon = Lexicon(params.lang_dir) + max_token_id = max(lexicon.tokens) + num_classes = max_token_id + 1 # +1 for the blank + params.vocab_size = num_classes + + device = torch.device("cpu") + if torch.cuda.is_available(): + device = torch.device("cuda", rank) + + graph_compiler = BpeCtcTrainingGraphCompiler( + params.lang_dir, + device=device, + sos_token="", + eos_token="", + ) + logging.info("About to create model") + + model = get_ctc_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) + + parameters_names = [] + parameters_names.append( + [name_param_pair[0] for name_param_pair in model.named_parameters()] + ) + optimizer = ScaledAdam( + model.parameters(), + lr=params.base_lr, + clipping_scale=2.0, + parameters_names=parameters_names, + ) + + scheduler = Eden(optimizer, params.lr_batches, params.lr_epochs) + + if checkpoints and "optimizer" in checkpoints: + logging.info("Loading optimizer state dict") + optimizer.load_state_dict(checkpoints["optimizer"]) + + if ( + checkpoints + and "scheduler" in checkpoints + and checkpoints["scheduler"] is not None + ): + logging.info("Loading scheduler state dict") + scheduler.load_state_dict(checkpoints["scheduler"]) + + if params.print_diagnostics: + opts = diagnostics.TensorDiagnosticOptions( + 2**22 + ) # allow 4 megabytes per sub-module + diagnostic = diagnostics.attach_diagnostics(model, opts) + + if params.inf_check: + register_inf_check_hooks(model) + + librispeech = LibriSpeechAsrDataModule(args) + + if params.mini_libri: + train_cuts = librispeech.train_clean_5_cuts() + else: + train_cuts = librispeech.train_clean_100_cuts() + if params.full_libri: + train_cuts += librispeech.train_clean_360_cuts() + train_cuts += librispeech.train_other_500_cuts() + + def remove_short_and_long_utt(c: Cut): + # Keep only utterances with duration between 1 second and 20 seconds + # + # Caution: There is a reason to select 20.0 here. Please see + # ../local/display_manifest_statistics.py + # + # You should use ../local/display_manifest_statistics.py to get + # an utterance duration distribution for your dataset to select + # the threshold + return 1.0 <= c.duration <= 25.0 + + train_cuts = train_cuts.filter(remove_short_and_long_utt) + + train_dl = librispeech.train_dataloaders(train_cuts) + + + if params.mini_libri: + valid_cuts = librispeech.dev_clean_2_cuts() + else: + valid_cuts = librispeech.dev_clean_cuts() + valid_cuts += librispeech.dev_other_cuts() + valid_dl = librispeech.valid_dataloaders(valid_cuts) + + scan_pessimistic_batches_for_oom( + model=model, + train_dl=train_dl, + optimizer=optimizer, + graph_compiler=graph_compiler, + 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, + graph_compiler=graph_compiler, + 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, + graph_compiler: BpeCtcTrainingGraphCompiler, +) -> None: + """Display the batch statistics and save the batch into disk. + + Args: + batch: + A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()` + for the content in it. + params: + Parameters for training. See :func:`get_params`. + graph_compiler: + It is used to build a decoding graph from a ctc topo and training + transcript. The training transcript is contained in the given `batch`, + while the ctc topo is built when this compiler is instantiated. + """ + from lhotse.utils import uuid4 + + filename = f"{params.exp_dir}/batch-{uuid4()}.pt" + logging.info(f"Saving batch to {filename}") + torch.save(batch, filename) + + supervisions = batch["supervisions"] + features = batch["inputs"] + + logging.info(f"features shape: {features.shape}") + + y = graph_compiler.texts_to_ids(supervisions["text"]) + num_tokens = sum(len(i) for i in y) + logging.info(f"num tokens: {num_tokens}") + + +def scan_pessimistic_batches_for_oom( + model: nn.Module, + train_dl: torch.utils.data.DataLoader, + optimizer: torch.optim.Optimizer, + graph_compiler: BpeCtcTrainingGraphCompiler, + params: AttributeDict, +): + from lhotse.dataset import find_pessimistic_batches + + logging.info( + "Sanity check -- see if any of the batches in epoch 0 would cause OOM." + ) + batches, crit_values = find_pessimistic_batches(train_dl.sampler) + for criterion, cuts in batches.items(): + batch = train_dl.dataset[cuts] + try: + optimizer.zero_grad() + loss, _ = compute_loss( + params=params, + model=model, + batch=batch, + graph_compiler=graph_compiler, + is_training=True, + ) + loss.backward() + clip_grad_norm_(model.parameters(), 5.0, 2.0) + optimizer.step() + except RuntimeError as e: + if "CUDA out of memory" in str(e): + logging.error( + "Your GPU ran out of memory with the current " + "max_duration setting. We recommend decreasing " + "max_duration and trying again.\n" + f"Failing criterion: {criterion} " + f"(={crit_values[criterion]}) ..." + ) + raise + + +def main(): + parser = get_parser() + LibriSpeechAsrDataModule.add_arguments(parser) + args = parser.parse_args() + args.exp_dir = Path(args.exp_dir) + args.lang_dir = Path(args.lang_dir) + + world_size = args.world_size + assert world_size >= 1 + if world_size > 1: + mp.spawn(run, args=(world_size, args), nprocs=world_size, join=True) + else: + run(rank=0, world_size=1, args=args) + + +torch.set_num_threads(1) +torch.set_num_interop_threads(1) + +if __name__ == "__main__": + main() diff --git a/egs/librispeech/ASR/zipformer_ctc_streaming/transformer.py b/egs/librispeech/ASR/zipformer_ctc_streaming/transformer.py new file mode 100644 index 000000000..0566cfc81 --- /dev/null +++ b/egs/librispeech/ASR/zipformer_ctc_streaming/transformer.py @@ -0,0 +1,928 @@ +# Copyright 2021 University of Chinese Academy of Sciences (author: Han Zhu) +# +# 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 math +from typing import Dict, List, Optional, Tuple, Union + +import torch +import torch.nn as nn +from label_smoothing import LabelSmoothingLoss +from subsampling import Conv2dSubsampling, VggSubsampling +from torch.nn.utils.rnn import pad_sequence + +# Note: TorchScript requires Dict/List/etc. to be fully typed. +Supervisions = Dict[str, torch.Tensor] + + +class Transformer(nn.Module): + def __init__( + self, + num_features: int, + num_classes: int, + subsampling_factor: int = 4, + d_model: int = 256, + nhead: int = 4, + dim_feedforward: int = 2048, + num_encoder_layers: int = 12, + num_decoder_layers: int = 6, + dropout: float = 0.1, + normalize_before: bool = True, + vgg_frontend: bool = False, + use_feat_batchnorm: Union[float, bool] = 0.1, + ) -> None: + """ + Args: + num_features: + The input dimension of the model. + num_classes: + The output dimension of the model. + subsampling_factor: + Number of output frames is num_in_frames // subsampling_factor. + Currently, subsampling_factor MUST be 4. + d_model: + Attention dimension. + nhead: + Number of heads in multi-head attention. + Must satisfy d_model // nhead == 0. + dim_feedforward: + The output dimension of the feedforward layers in encoder/decoder. + num_encoder_layers: + Number of encoder layers. + num_decoder_layers: + Number of decoder layers. + dropout: + Dropout in encoder/decoder. + normalize_before: + If True, use pre-layer norm; False to use post-layer norm. + vgg_frontend: + True to use vgg style frontend for subsampling. + use_feat_batchnorm: + True to use batchnorm for the input layer. + Float value to scale the input layer. + False to do nothing. + """ + super().__init__() + self.use_feat_batchnorm = use_feat_batchnorm + assert isinstance(use_feat_batchnorm, (float, bool)) + if isinstance(use_feat_batchnorm, bool) and use_feat_batchnorm: + self.feat_batchnorm = nn.BatchNorm1d(num_features) + + self.num_features = num_features + self.num_classes = num_classes + self.subsampling_factor = subsampling_factor + if subsampling_factor != 4: + raise NotImplementedError("Support only 'subsampling_factor=4'.") + + # self.encoder_embed converts the input of shape (N, T, num_classes) + # to the shape (N, T//subsampling_factor, d_model). + # That is, it does two things simultaneously: + # (1) subsampling: T -> T//subsampling_factor + # (2) embedding: num_classes -> d_model + if vgg_frontend: + self.encoder_embed = VggSubsampling(num_features, d_model) + else: + self.encoder_embed = Conv2dSubsampling(num_features, d_model) + + self.encoder_pos = PositionalEncoding(d_model, dropout) + + encoder_layer = TransformerEncoderLayer( + d_model=d_model, + nhead=nhead, + dim_feedforward=dim_feedforward, + dropout=dropout, + normalize_before=normalize_before, + ) + + if normalize_before: + encoder_norm = nn.LayerNorm(d_model) + else: + encoder_norm = None + + self.encoder = nn.TransformerEncoder( + encoder_layer=encoder_layer, + num_layers=num_encoder_layers, + norm=encoder_norm, + ) + + # TODO(fangjun): remove dropout + self.encoder_output_layer = nn.Sequential( + nn.Dropout(p=dropout), nn.Linear(d_model, num_classes) + ) + + if num_decoder_layers > 0: + self.decoder_num_class = ( + self.num_classes + ) # bpe model already has sos/eos symbol + + self.decoder_embed = nn.Embedding( + num_embeddings=self.decoder_num_class, embedding_dim=d_model + ) + self.decoder_pos = PositionalEncoding(d_model, dropout) + + decoder_layer = TransformerDecoderLayer( + d_model=d_model, + nhead=nhead, + dim_feedforward=dim_feedforward, + dropout=dropout, + normalize_before=normalize_before, + ) + + if normalize_before: + decoder_norm = nn.LayerNorm(d_model) + else: + decoder_norm = None + + self.decoder = nn.TransformerDecoder( + decoder_layer=decoder_layer, + num_layers=num_decoder_layers, + norm=decoder_norm, + ) + + self.decoder_output_layer = torch.nn.Linear(d_model, self.decoder_num_class) + + self.decoder_criterion = LabelSmoothingLoss() + else: + self.decoder_criterion = None + + def forward( + self, x: torch.Tensor, supervision: Optional[Supervisions] = None + ) -> Tuple[torch.Tensor, torch.Tensor, Optional[torch.Tensor]]: + """ + Args: + x: + The input tensor. Its shape is (N, T, C). + supervision: + Supervision in lhotse format. + See https://github.com/lhotse-speech/lhotse/blob/master/lhotse/dataset/speech_recognition.py#L32 # noqa + (CAUTION: It contains length information, i.e., start and number of + frames, before subsampling) + + Returns: + Return a tuple containing 3 tensors: + - CTC output for ctc decoding. Its shape is (N, T, C) + - Encoder output with shape (T, N, C). It can be used as key and + value for the decoder. + - Encoder output padding mask. It can be used as + memory_key_padding_mask for the decoder. Its shape is (N, T). + It is None if `supervision` is None. + """ + if isinstance(self.use_feat_batchnorm, bool) and self.use_feat_batchnorm: + x = x.permute(0, 2, 1) # (N, T, C) -> (N, C, T) + x = self.feat_batchnorm(x) + x = x.permute(0, 2, 1) # (N, C, T) -> (N, T, C) + if isinstance(self.use_feat_batchnorm, float): + x *= self.use_feat_batchnorm + encoder_memory, memory_key_padding_mask = self.run_encoder(x, supervision) + x = self.ctc_output(encoder_memory) + return x, encoder_memory, memory_key_padding_mask + + def run_encoder( + self, x: torch.Tensor, supervisions: Optional[Supervisions] = None + ) -> Tuple[torch.Tensor, Optional[torch.Tensor]]: + """Run the transformer encoder. + + Args: + x: + The model input. Its shape is (N, T, C). + supervisions: + Supervision in lhotse format. + See https://github.com/lhotse-speech/lhotse/blob/master/lhotse/dataset/speech_recognition.py#L32 # noqa + CAUTION: It contains length information, i.e., start and number of + frames, before subsampling + It is read directly from the batch, without any sorting. It is used + to compute the encoder padding mask, which is used as memory key + padding mask for the decoder. + Returns: + Return a tuple with two tensors: + - The encoder output, with shape (T, N, C) + - encoder padding mask, with shape (N, T). + The mask is None if `supervisions` is None. + It is used as memory key padding mask in the decoder. + """ + x = self.encoder_embed(x) + x = self.encoder_pos(x) + x = x.permute(1, 0, 2) # (N, T, C) -> (T, N, C) + mask = encoder_padding_mask(x.size(0), supervisions) + mask = mask.to(x.device) if mask is not None else None + x = self.encoder(x, src_key_padding_mask=mask) # (T, N, C) + + return x, mask + + def ctc_output(self, x: torch.Tensor) -> torch.Tensor: + """ + Args: + x: + The output tensor from the transformer encoder. + Its shape is (T, N, C) + + Returns: + Return a tensor that can be used for CTC decoding. + Its shape is (N, T, C) + """ + x = self.encoder_output_layer(x) + x = x.permute(1, 0, 2) # (T, N, C) ->(N, T, C) + x = nn.functional.log_softmax(x, dim=-1) # (N, T, C) + return x + + @torch.jit.export + def decoder_forward( + self, + memory: torch.Tensor, + memory_key_padding_mask: torch.Tensor, + token_ids: List[List[int]], + sos_id: int, + eos_id: int, + ) -> torch.Tensor: + """ + Args: + memory: + It's the output of the encoder with shape (T, N, C) + memory_key_padding_mask: + The padding mask from the encoder. + token_ids: + A list-of-list IDs. Each sublist contains IDs for an utterance. + The IDs can be either phone IDs or word piece IDs. + sos_id: + sos token id + eos_id: + eos token id + + Returns: + A scalar, the **sum** of label smoothing loss over utterances + in the batch without any normalization. + """ + ys_in = add_sos(token_ids, sos_id=sos_id) + ys_in = [torch.tensor(y) for y in ys_in] + ys_in_pad = pad_sequence(ys_in, batch_first=True, padding_value=float(eos_id)) + + ys_out = add_eos(token_ids, eos_id=eos_id) + ys_out = [torch.tensor(y) for y in ys_out] + ys_out_pad = pad_sequence(ys_out, batch_first=True, padding_value=float(-1)) + + device = memory.device + ys_in_pad = ys_in_pad.to(device) + ys_out_pad = ys_out_pad.to(device) + + tgt_mask = generate_square_subsequent_mask(ys_in_pad.shape[-1]).to(device) + + tgt_key_padding_mask = decoder_padding_mask(ys_in_pad, ignore_id=eos_id) + # TODO: Use length information to create the decoder padding mask + # We set the first column to False since the first column in ys_in_pad + # contains sos_id, which is the same as eos_id in our current setting. + tgt_key_padding_mask[:, 0] = False + + tgt = self.decoder_embed(ys_in_pad) # (N, T) -> (N, T, C) + tgt = self.decoder_pos(tgt) + tgt = tgt.permute(1, 0, 2) # (N, T, C) -> (T, N, C) + pred_pad = self.decoder( + tgt=tgt, + memory=memory, + tgt_mask=tgt_mask, + tgt_key_padding_mask=tgt_key_padding_mask, + memory_key_padding_mask=memory_key_padding_mask, + ) # (T, N, C) + pred_pad = pred_pad.permute(1, 0, 2) # (T, N, C) -> (N, T, C) + pred_pad = self.decoder_output_layer(pred_pad) # (N, T, C) + + decoder_loss = self.decoder_criterion(pred_pad, ys_out_pad) + + return decoder_loss + + @torch.jit.export + def decoder_nll( + self, + memory: torch.Tensor, + memory_key_padding_mask: torch.Tensor, + token_ids: List[torch.Tensor], + sos_id: int, + eos_id: int, + ) -> torch.Tensor: + """ + Args: + memory: + It's the output of the encoder with shape (T, N, C) + memory_key_padding_mask: + The padding mask from the encoder. + token_ids: + A list-of-list IDs (e.g., word piece IDs). + Each sublist represents an utterance. + sos_id: + The token ID for SOS. + eos_id: + The token ID for EOS. + Returns: + A 2-D tensor of shape (len(token_ids), max_token_length) + representing the cross entropy loss (i.e., negative log-likelihood). + """ + # The common part between this function and decoder_forward could be + # extracted as a separate function. + if isinstance(token_ids[0], torch.Tensor): + # This branch is executed by torchscript in C++. + # See https://github.com/k2-fsa/k2/pull/870 + # https://github.com/k2-fsa/k2/blob/3c1c18400060415b141ccea0115fd4bf0ad6234e/k2/torch/bin/attention_rescore.cu#L286 + token_ids = [tolist(t) for t in token_ids] + + ys_in = add_sos(token_ids, sos_id=sos_id) + ys_in = [torch.tensor(y) for y in ys_in] + ys_in_pad = pad_sequence(ys_in, batch_first=True, padding_value=float(eos_id)) + + ys_out = add_eos(token_ids, eos_id=eos_id) + ys_out = [torch.tensor(y) for y in ys_out] + ys_out_pad = pad_sequence(ys_out, batch_first=True, padding_value=float(-1)) + + device = memory.device + ys_in_pad = ys_in_pad.to(device, dtype=torch.int64) + ys_out_pad = ys_out_pad.to(device, dtype=torch.int64) + + tgt_mask = generate_square_subsequent_mask(ys_in_pad.shape[-1]).to(device) + + tgt_key_padding_mask = decoder_padding_mask(ys_in_pad, ignore_id=eos_id) + # TODO: Use length information to create the decoder padding mask + # We set the first column to False since the first column in ys_in_pad + # contains sos_id, which is the same as eos_id in our current setting. + tgt_key_padding_mask[:, 0] = False + + tgt = self.decoder_embed(ys_in_pad) # (B, T) -> (B, T, F) + tgt = self.decoder_pos(tgt) + tgt = tgt.permute(1, 0, 2) # (B, T, F) -> (T, B, F) + pred_pad = self.decoder( + tgt=tgt, + memory=memory, + tgt_mask=tgt_mask, + tgt_key_padding_mask=tgt_key_padding_mask, + memory_key_padding_mask=memory_key_padding_mask, + ) # (T, B, F) + pred_pad = pred_pad.permute(1, 0, 2) # (T, B, F) -> (B, T, F) + pred_pad = self.decoder_output_layer(pred_pad) # (B, T, F) + # nll: negative log-likelihood + nll = torch.nn.functional.cross_entropy( + pred_pad.view(-1, self.decoder_num_class), + ys_out_pad.view(-1), + ignore_index=-1, + reduction="none", + ) + + nll = nll.view(pred_pad.shape[0], -1) + + return nll + + +class TransformerEncoderLayer(nn.Module): + """ + Modified from torch.nn.TransformerEncoderLayer. + Add support of normalize_before, + i.e., use layer_norm before the first block. + + Args: + d_model: + the number of expected features in the input (required). + nhead: + the number of heads in the multiheadattention models (required). + dim_feedforward: + the dimension of the feedforward network model (default=2048). + dropout: + the dropout value (default=0.1). + activation: + the activation function of intermediate layer, relu or + gelu (default=relu). + normalize_before: + whether to use layer_norm before the first block. + + Examples:: + >>> encoder_layer = TransformerEncoderLayer(d_model=512, nhead=8) + >>> src = torch.rand(10, 32, 512) + >>> out = encoder_layer(src) + """ + + def __init__( + self, + d_model: int, + nhead: int, + dim_feedforward: int = 2048, + dropout: float = 0.1, + activation: str = "relu", + normalize_before: bool = True, + ) -> None: + super(TransformerEncoderLayer, self).__init__() + self.self_attn = nn.MultiheadAttention(d_model, nhead, dropout=0.0) + # Implementation of Feedforward model + self.linear1 = nn.Linear(d_model, dim_feedforward) + self.dropout = nn.Dropout(dropout) + self.linear2 = nn.Linear(dim_feedforward, d_model) + + self.norm1 = nn.LayerNorm(d_model) + self.norm2 = nn.LayerNorm(d_model) + self.dropout1 = nn.Dropout(dropout) + self.dropout2 = nn.Dropout(dropout) + + self.activation = _get_activation_fn(activation) + + self.normalize_before = normalize_before + + def __setstate__(self, state): + if "activation" not in state: + state["activation"] = nn.functional.relu + super(TransformerEncoderLayer, self).__setstate__(state) + + def forward( + self, + src: torch.Tensor, + src_mask: Optional[torch.Tensor] = None, + src_key_padding_mask: Optional[torch.Tensor] = None, + ) -> torch.Tensor: + """ + Pass the input through the encoder layer. + + Args: + src: the sequence to the encoder layer (required). + src_mask: the mask for the src sequence (optional). + src_key_padding_mask: the mask for the src keys per batch (optional) + + Shape: + src: (S, N, E). + src_mask: (S, S). + src_key_padding_mask: (N, S). + S is the source sequence length, T is the target sequence length, + N is the batch size, E is the feature number + """ + residual = src + if self.normalize_before: + src = self.norm1(src) + src2 = self.self_attn( + src, + src, + src, + attn_mask=src_mask, + key_padding_mask=src_key_padding_mask, + )[0] + src = residual + self.dropout1(src2) + if not self.normalize_before: + src = self.norm1(src) + + residual = src + if self.normalize_before: + src = self.norm2(src) + src2 = self.linear2(self.dropout(self.activation(self.linear1(src)))) + src = residual + self.dropout2(src2) + if not self.normalize_before: + src = self.norm2(src) + return src + + +class TransformerDecoderLayer(nn.Module): + """ + Modified from torch.nn.TransformerDecoderLayer. + Add support of normalize_before, + i.e., use layer_norm before the first block. + + Args: + d_model: + the number of expected features in the input (required). + nhead: + the number of heads in the multiheadattention models (required). + dim_feedforward: + the dimension of the feedforward network model (default=2048). + dropout: + the dropout value (default=0.1). + activation: + the activation function of intermediate layer, relu or + gelu (default=relu). + + Examples:: + >>> decoder_layer = nn.TransformerDecoderLayer(d_model=512, nhead=8) + >>> memory = torch.rand(10, 32, 512) + >>> tgt = torch.rand(20, 32, 512) + >>> out = decoder_layer(tgt, memory) + """ + + def __init__( + self, + d_model: int, + nhead: int, + dim_feedforward: int = 2048, + dropout: float = 0.1, + activation: str = "relu", + normalize_before: bool = True, + ) -> None: + super(TransformerDecoderLayer, self).__init__() + self.self_attn = nn.MultiheadAttention(d_model, nhead, dropout=0.0) + self.src_attn = nn.MultiheadAttention(d_model, nhead, dropout=0.0) + # Implementation of Feedforward model + self.linear1 = nn.Linear(d_model, dim_feedforward) + self.dropout = nn.Dropout(dropout) + self.linear2 = nn.Linear(dim_feedforward, d_model) + + self.norm1 = nn.LayerNorm(d_model) + self.norm2 = nn.LayerNorm(d_model) + self.norm3 = nn.LayerNorm(d_model) + self.dropout1 = nn.Dropout(dropout) + self.dropout2 = nn.Dropout(dropout) + self.dropout3 = nn.Dropout(dropout) + + self.activation = _get_activation_fn(activation) + + self.normalize_before = normalize_before + + def __setstate__(self, state): + if "activation" not in state: + state["activation"] = nn.functional.relu + super(TransformerDecoderLayer, self).__setstate__(state) + + def forward( + self, + tgt: torch.Tensor, + memory: torch.Tensor, + tgt_mask: Optional[torch.Tensor] = None, + memory_mask: Optional[torch.Tensor] = None, + tgt_key_padding_mask: Optional[torch.Tensor] = None, + memory_key_padding_mask: Optional[torch.Tensor] = None, + ) -> torch.Tensor: + """Pass the inputs (and mask) through the decoder layer. + + Args: + tgt: + the sequence to the decoder layer (required). + memory: + the sequence from the last layer of the encoder (required). + tgt_mask: + the mask for the tgt sequence (optional). + memory_mask: + the mask for the memory sequence (optional). + tgt_key_padding_mask: + the mask for the tgt keys per batch (optional). + memory_key_padding_mask: + the mask for the memory keys per batch (optional). + + Shape: + tgt: (T, N, E). + memory: (S, N, E). + tgt_mask: (T, T). + memory_mask: (T, S). + tgt_key_padding_mask: (N, T). + memory_key_padding_mask: (N, S). + S is the source sequence length, T is the target sequence length, + N is the batch size, E is the feature number + """ + residual = tgt + if self.normalize_before: + tgt = self.norm1(tgt) + tgt2 = self.self_attn( + tgt, + tgt, + tgt, + attn_mask=tgt_mask, + key_padding_mask=tgt_key_padding_mask, + )[0] + tgt = residual + self.dropout1(tgt2) + if not self.normalize_before: + tgt = self.norm1(tgt) + + residual = tgt + if self.normalize_before: + tgt = self.norm2(tgt) + tgt2 = self.src_attn( + tgt, + memory, + memory, + attn_mask=memory_mask, + key_padding_mask=memory_key_padding_mask, + )[0] + tgt = residual + self.dropout2(tgt2) + if not self.normalize_before: + tgt = self.norm2(tgt) + + residual = tgt + if self.normalize_before: + tgt = self.norm3(tgt) + tgt2 = self.linear2(self.dropout(self.activation(self.linear1(tgt)))) + tgt = residual + self.dropout3(tgt2) + if not self.normalize_before: + tgt = self.norm3(tgt) + return tgt + + +def _get_activation_fn(activation: str): + if activation == "relu": + return nn.functional.relu + elif activation == "gelu": + return nn.functional.gelu + + raise RuntimeError("activation should be relu/gelu, not {}".format(activation)) + + +class PositionalEncoding(nn.Module): + """This class implements the positional encoding + proposed in the following paper: + + - Attention Is All You Need: https://arxiv.org/pdf/1706.03762.pdf + + PE(pos, 2i) = sin(pos / (10000^(2i/d_modle)) + PE(pos, 2i+1) = cos(pos / (10000^(2i/d_modle)) + + Note:: + + 1 / (10000^(2i/d_model)) = exp(-log(10000^(2i/d_model))) + = exp(-1* 2i / d_model * log(100000)) + = exp(2i * -(log(10000) / d_model)) + """ + + def __init__(self, d_model: int, dropout: float = 0.1) -> None: + """ + Args: + d_model: + Embedding dimension. + dropout: + Dropout probability to be applied to the output of this module. + """ + super().__init__() + self.d_model = d_model + self.xscale = math.sqrt(self.d_model) + self.dropout = nn.Dropout(p=dropout) + # not doing: self.pe = None because of errors thrown by torchscript + self.pe = torch.zeros(1, 0, self.d_model, dtype=torch.float32) + + def extend_pe(self, x: torch.Tensor) -> None: + """Extend the time t in the positional encoding if required. + + The shape of `self.pe` is (1, T1, d_model). The shape of the input x + is (N, T, d_model). If T > T1, then we change the shape of self.pe + to (N, T, d_model). Otherwise, nothing is done. + + Args: + x: + It is a tensor of shape (N, T, C). + Returns: + Return None. + """ + if self.pe is not None: + if self.pe.size(1) >= x.size(1): + self.pe = self.pe.to(dtype=x.dtype, device=x.device) + return + pe = torch.zeros(x.size(1), self.d_model, dtype=torch.float32) + position = torch.arange(0, x.size(1), dtype=torch.float32).unsqueeze(1) + div_term = torch.exp( + torch.arange(0, self.d_model, 2, dtype=torch.float32) + * -(math.log(10000.0) / self.d_model) + ) + pe[:, 0::2] = torch.sin(position * div_term) + pe[:, 1::2] = torch.cos(position * div_term) + pe = pe.unsqueeze(0) + # Now pe is of shape (1, T, d_model), where T is x.size(1) + self.pe = pe.to(device=x.device, dtype=x.dtype) + + def forward(self, x: torch.Tensor) -> torch.Tensor: + """ + Add positional encoding. + + Args: + x: + Its shape is (N, T, C) + + Returns: + Return a tensor of shape (N, T, C) + """ + self.extend_pe(x) + x = x * self.xscale + self.pe[:, : x.size(1), :] + return self.dropout(x) + + +class Noam(object): + """ + Implements Noam optimizer. + + Proposed in + "Attention Is All You Need", https://arxiv.org/pdf/1706.03762.pdf + + Modified from + https://github.com/espnet/espnet/blob/master/espnet/nets/pytorch_backend/transformer/optimizer.py # noqa + + Args: + params: + iterable of parameters to optimize or dicts defining parameter groups + model_size: + attention dimension of the transformer model + factor: + learning rate factor + warm_step: + warmup steps + """ + + def __init__( + self, + params, + model_size: int = 256, + factor: float = 10.0, + warm_step: int = 25000, + weight_decay=0, + ) -> None: + """Construct an Noam object.""" + self.optimizer = torch.optim.Adam( + params, lr=0, betas=(0.9, 0.98), eps=1e-9, weight_decay=weight_decay + ) + self._step = 0 + self.warmup = warm_step + self.factor = factor + self.model_size = model_size + self._rate = 0 + + @property + def param_groups(self): + """Return param_groups.""" + return self.optimizer.param_groups + + def step(self): + """Update parameters and rate.""" + self._step += 1 + rate = self.rate() + for p in self.optimizer.param_groups: + p["lr"] = rate + self._rate = rate + self.optimizer.step() + + def rate(self, step=None): + """Implement `lrate` above.""" + if step is None: + step = self._step + return ( + self.factor + * self.model_size ** (-0.5) + * min(step ** (-0.5), step * self.warmup ** (-1.5)) + ) + + def zero_grad(self): + """Reset gradient.""" + self.optimizer.zero_grad() + + def state_dict(self): + """Return state_dict.""" + return { + "_step": self._step, + "warmup": self.warmup, + "factor": self.factor, + "model_size": self.model_size, + "_rate": self._rate, + "optimizer": self.optimizer.state_dict(), + } + + def load_state_dict(self, state_dict): + """Load state_dict.""" + for key, value in state_dict.items(): + if key == "optimizer": + self.optimizer.load_state_dict(state_dict["optimizer"]) + else: + setattr(self, key, value) + + +def encoder_padding_mask( + max_len: int, supervisions: Optional[Supervisions] = None +) -> Optional[torch.Tensor]: + """Make mask tensor containing indexes of padded part. + + TODO:: + This function **assumes** that the model uses + a subsampling factor of 4. We should remove that + assumption later. + + Args: + max_len: + Maximum length of input features. + CAUTION: It is the length after subsampling. + supervisions: + Supervision in lhotse format. + See https://github.com/lhotse-speech/lhotse/blob/master/lhotse/dataset/speech_recognition.py#L32 # noqa + (CAUTION: It contains length information, i.e., start and number of + frames, before subsampling) + + Returns: + Tensor: Mask tensor of dimension (batch_size, input_length), + True denote the masked indices. + """ + if supervisions is None: + return None + + supervision_segments = torch.stack( + ( + supervisions["sequence_idx"], + supervisions["start_frame"], + supervisions["num_frames"], + ), + 1, + ).to(torch.int32) + + lengths = [0 for _ in range(int(supervision_segments[:, 0].max().item()) + 1)] + for idx in range(supervision_segments.size(0)): + # Note: TorchScript doesn't allow to unpack tensors as tuples + sequence_idx = supervision_segments[idx, 0].item() + start_frame = supervision_segments[idx, 1].item() + num_frames = supervision_segments[idx, 2].item() + lengths[sequence_idx] = start_frame + num_frames + + lengths = [((i - 1) // 2 - 1) // 2 for i in lengths] + bs = int(len(lengths)) + seq_range = torch.arange(0, max_len, dtype=torch.int64) + seq_range_expand = seq_range.unsqueeze(0).expand(bs, max_len) + # Note: TorchScript doesn't implement Tensor.new() + seq_length_expand = torch.tensor( + lengths, device=seq_range_expand.device, dtype=seq_range_expand.dtype + ).unsqueeze(-1) + mask = seq_range_expand >= seq_length_expand + + return mask + + +def decoder_padding_mask(ys_pad: torch.Tensor, ignore_id: int = -1) -> torch.Tensor: + """Generate a length mask for input. + + The masked position are filled with True, + Unmasked positions are filled with False. + + Args: + ys_pad: + padded tensor of dimension (batch_size, input_length). + ignore_id: + the ignored number (the padding number) in ys_pad + + Returns: + Tensor: + a bool tensor of the same shape as the input tensor. + """ + ys_mask = ys_pad == ignore_id + return ys_mask + + +def generate_square_subsequent_mask(sz: int) -> torch.Tensor: + """Generate a square mask for the sequence. The masked positions are + filled with float('-inf'). Unmasked positions are filled with float(0.0). + The mask can be used for masked self-attention. + + For instance, if sz is 3, it returns:: + + tensor([[0., -inf, -inf], + [0., 0., -inf], + [0., 0., 0]]) + + Args: + sz: mask size + + Returns: + A square mask of dimension (sz, sz) + """ + mask = (torch.triu(torch.ones(sz, sz)) == 1).transpose(0, 1) + mask = ( + mask.float() + .masked_fill(mask == 0, float("-inf")) + .masked_fill(mask == 1, float(0.0)) + ) + return mask + + +def add_sos(token_ids: List[List[int]], sos_id: int) -> List[List[int]]: + """Prepend sos_id to each utterance. + + Args: + token_ids: + A list-of-list of token IDs. Each sublist contains + token IDs (e.g., word piece IDs) of an utterance. + sos_id: + The ID of the SOS token. + + Return: + Return a new list-of-list, where each sublist starts + with SOS ID. + """ + return [[sos_id] + utt for utt in token_ids] + + +def add_eos(token_ids: List[List[int]], eos_id: int) -> List[List[int]]: + """Append eos_id to each utterance. + + Args: + token_ids: + A list-of-list of token IDs. Each sublist contains + token IDs (e.g., word piece IDs) of an utterance. + eos_id: + The ID of the EOS token. + + Return: + Return a new list-of-list, where each sublist ends + with EOS ID. + """ + return [utt + [eos_id] for utt in token_ids] + + +def tolist(t: torch.Tensor) -> List[int]: + """Used by jit""" + return torch.jit.annotate(List[int], t.tolist()) diff --git a/egs/librispeech/ASR/zipformer_ctc_streaming/zipformer.py b/egs/librispeech/ASR/zipformer_ctc_streaming/zipformer.py new file mode 100644 index 000000000..88beb38c1 --- /dev/null +++ b/egs/librispeech/ASR/zipformer_ctc_streaming/zipformer.py @@ -0,0 +1,2881 @@ +#!/usr/bin/env python3 +# Copyright 2022 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 itertools +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 ( + ScaledLinear, # not as in other dirs.. just scales down initial parameter values. +) +from scaling import ( + ActivationBalancer, + BasicNorm, + DoubleSwish, + Identity, + MaxEig, + ScaledConv1d, + Whiten, + _diag, + penalize_abs_values_gt, + random_clamp, + softmax, +) +from torch import Tensor, nn + +from icefall.dist import get_rank +from icefall.utils import make_pad_mask, subsequent_chunk_mask + + +def stack_states(state_list: List[List[Tensor]]) -> List[Tensor]: + """Stack list of zipformer states that correspond to separate utterances + into a single emformer state, so that it can be used as an input for + zipformer when those utterances are formed into a batch. + + Note: + It is the inverse of :func:`unstack_states`. + + Args: + state_list: + Each element in state_list corresponding to the internal state + of the zipformer model for a single utterance. + ``states[i]`` is a list of 7 * num_encoders elements of i-th utterance. + ``states[i][0:num_encoders]`` is the cached numbers of past frames. + ``states[i][num_encoders:2*num_encoders]`` is the cached average tensors. + ``states[i][2*num_encoders:3*num_encoders]`` is the cached key tensors of the first attention modules. + ``states[i][3*num_encoders:4*num_encoders]`` is the cached value tensors of the first attention modules. + ``states[i][4*num_encoders:5*num_encoders]`` is the cached value tensors of the second attention modules. + ``states[i][5*num_encoders:6*num_encoders]`` is the cached left contexts of the first convolution modules. + ``states[i][6*num_encoders:7*num_encoders]`` is the cached left contexts of the second convolution modules. + + Returns: + A new state corresponding to a batch of utterances. + See the input argument of :func:`unstack_states` for the meaning + of the returned tensor. + """ + batch_size = len(state_list) + assert len(state_list[0]) % 7 == 0, len(state_list[0]) + num_encoders = len(state_list[0]) // 7 + + cached_len = [] + cached_avg = [] + cached_key = [] + cached_val = [] + cached_val2 = [] + cached_conv1 = [] + cached_conv2 = [] + + # For cached_len + len_list = [state_list[n][0:num_encoders] for n in range(batch_size)] + for i in range(num_encoders): + # len_avg: (num_layers, batch_size) + len_avg = torch.cat([len_list[n][i] for n in range(batch_size)], dim=1) + cached_len.append(len_avg) + + # For cached_avg + avg_list = [ + state_list[n][num_encoders : 2 * num_encoders] for n in range(batch_size) + ] + for i in range(num_encoders): + # avg: (num_layers, batch_size, D) + avg = torch.cat([avg_list[n][i] for n in range(batch_size)], dim=1) + cached_avg.append(avg) + + # For cached_key + key_list = [ + state_list[n][2 * num_encoders : 3 * num_encoders] for n in range(batch_size) + ] + for i in range(num_encoders): + # key: (num_layers, left_context_size, batch_size, D) + key = torch.cat([key_list[n][i] for n in range(batch_size)], dim=2) + cached_key.append(key) + + # For cached_val + val_list = [ + state_list[n][3 * num_encoders : 4 * num_encoders] for n in range(batch_size) + ] + for i in range(num_encoders): + # val: (num_layers, left_context_size, batch_size, D) + val = torch.cat([val_list[n][i] for n in range(batch_size)], dim=2) + cached_val.append(val) + + # For cached_val2 + val2_list = [ + state_list[n][4 * num_encoders : 5 * num_encoders] for n in range(batch_size) + ] + for i in range(num_encoders): + # val2: (num_layers, left_context_size, batch_size, D) + val2 = torch.cat([val2_list[n][i] for n in range(batch_size)], dim=2) + cached_val2.append(val2) + + # For cached_conv1 + conv1_list = [ + state_list[n][5 * num_encoders : 6 * num_encoders] for n in range(batch_size) + ] + for i in range(num_encoders): + # conv1: (num_layers, batch_size, D, kernel-1) + conv1 = torch.cat([conv1_list[n][i] for n in range(batch_size)], dim=1) + cached_conv1.append(conv1) + + # For cached_conv2 + conv2_list = [ + state_list[n][6 * num_encoders : 7 * num_encoders] for n in range(batch_size) + ] + for i in range(num_encoders): + # conv2: (num_layers, batch_size, D, kernel-1) + conv2 = torch.cat([conv2_list[n][i] for n in range(batch_size)], dim=1) + cached_conv2.append(conv2) + + states = ( + cached_len + + cached_avg + + cached_key + + cached_val + + cached_val2 + + cached_conv1 + + cached_conv2 + ) + return states + + +def unstack_states(states: List[Tensor]) -> List[List[Tensor]]: + """Unstack the zipformer state corresponding to a batch of utterances + into a list of states, where the i-th entry is the state from the i-th + utterance in the batch. + + Note: + It is the inverse of :func:`stack_states`. + + Args: + states: + A list of 7 * num_encoders elements: + ``states[0:num_encoders]`` is the cached numbers of past frames. + ``states[num_encoders:2*num_encoders]`` is the cached average tensors. + ``states[2*num_encoders:3*num_encoders]`` is the cached key tensors of the first attention modules. + ``states[3*num_encoders:4*num_encoders]`` is the cached value tensors of the first attention modules. + ``states[4*num_encoders:5*num_encoders]`` is the cached value tensors of the second attention modules. + ``states[5*num_encoders:6*num_encoders]`` is the cached left contexts of the first convolution modules. + ``states[6*num_encoders:7*num_encoders]`` is the cached left contexts of the second convolution modules. + + Returns: + A list of states. + ``states[i]`` is a list of 7 * num_encoders elements of i-th utterance. + """ + assert len(states) % 7 == 0, len(states) + num_encoders = len(states) // 7 + ( + cached_len, + cached_avg, + cached_key, + cached_val, + cached_val2, + cached_conv1, + cached_conv2, + ) = (states[i * num_encoders : (i + 1) * num_encoders] for i in range(7)) + + batch_size = cached_len[0].shape[1] + + len_list = [[] for _ in range(batch_size)] + for i in range(num_encoders): + # cached_len[i]: (num_layers, batch_size) + len_avg = cached_len[i].chunk(chunks=batch_size, dim=1) + for n in range(batch_size): + len_list[n].append(len_avg[n]) + + avg_list = [[] for _ in range(batch_size)] + for i in range(num_encoders): + # cached_avg[i]: (num_layers, batch_size, D) + avg = cached_avg[i].chunk(chunks=batch_size, dim=1) + for n in range(batch_size): + avg_list[n].append(avg[n]) + + key_list = [[] for _ in range(batch_size)] + for i in range(num_encoders): + # cached_key[i]: (num_layers, left_context, batch_size, D) + key = cached_key[i].chunk(chunks=batch_size, dim=2) + for n in range(batch_size): + key_list[n].append(key[n]) + + val_list = [[] for _ in range(batch_size)] + for i in range(num_encoders): + # cached_val[i]: (num_layers, left_context, batch_size, D) + val = cached_val[i].chunk(chunks=batch_size, dim=2) + for n in range(batch_size): + val_list[n].append(val[n]) + + val2_list = [[] for _ in range(batch_size)] + for i in range(num_encoders): + # cached_val2[i]: (num_layers, left_context, batch_size, D) + val2 = cached_val2[i].chunk(chunks=batch_size, dim=2) + for n in range(batch_size): + val2_list[n].append(val2[n]) + + conv1_list = [[] for _ in range(batch_size)] + for i in range(num_encoders): + # cached_conv1[i]: (num_layers, batch_size, D, kernel-1) + conv1 = cached_conv1[i].chunk(chunks=batch_size, dim=1) + for n in range(batch_size): + conv1_list[n].append(conv1[n]) + + conv2_list = [[] for _ in range(batch_size)] + for i in range(num_encoders): + # cached_conv2[i]: (num_layers, batch_size, D, kernel-1) + conv2 = cached_conv2[i].chunk(chunks=batch_size, dim=1) + for n in range(batch_size): + conv2_list[n].append(conv2[n]) + + state_list = [ + ( + len_list[i] + + avg_list[i] + + key_list[i] + + val_list[i] + + val2_list[i] + + conv1_list[i] + + conv2_list[i] + ) + for i in range(batch_size) + ] + return state_list + + +class Zipformer(EncoderInterface): + """ + Args: + num_features (int): Number of input features + d_model: (int,int): embedding dimension of 2 encoder stacks + attention_dim: (int,int): attention dimension of 2 encoder stacks + nhead (int, int): number of heads + dim_feedforward (int, int): feedforward dimension in 2 encoder stacks + num_encoder_layers (int): number of encoder layers + dropout (float): dropout rate + cnn_module_kernel (int): Kernel size of convolution module + vgg_frontend (bool): whether to use vgg frontend. + warmup_batches (float): number of batches to warm up over + """ + + def __init__( + self, + num_features: int, + output_downsampling_factor: int = 2, + encoder_dims: Tuple[int] = (384, 384), + attention_dim: Tuple[int] = (256, 256), + encoder_unmasked_dims: Tuple[int] = (256, 256), + zipformer_downsampling_factors: Tuple[int] = (2, 4), + nhead: Tuple[int] = (8, 8), + feedforward_dim: Tuple[int] = (1536, 2048), + num_encoder_layers: Tuple[int] = (12, 12), + dropout: float = 0.1, + cnn_module_kernels: Tuple[int] = (31, 31), + pos_dim: int = 4, + num_left_chunks: int = 4, + short_chunk_threshold: float = 0.75, + short_chunk_size: int = 50, + decode_chunk_size: int = 16, + warmup_batches: float = 4000.0, + ) -> None: + super(Zipformer, self).__init__() + + self.num_features = num_features + assert 0 < encoder_dims[0] <= encoder_dims[1] + self.encoder_dims = encoder_dims + self.encoder_unmasked_dims = encoder_unmasked_dims + self.zipformer_downsampling_factors = zipformer_downsampling_factors + self.output_downsampling_factor = output_downsampling_factor + + self.num_left_chunks = num_left_chunks + self.short_chunk_threshold = short_chunk_threshold + self.short_chunk_size = short_chunk_size + + # Used in decoding + self.decode_chunk_size = decode_chunk_size + + # will be written to, see set_batch_count() + self.batch_count = 0 + self.warmup_end = warmup_batches + + for u, d in zip(encoder_unmasked_dims, encoder_dims): + assert u <= d, (u, d) + + # self.encoder_embed converts the input of shape (N, T, num_features) + # to the shape (N, (T - 7)//2, encoder_dims). + # That is, it does two things simultaneously: + # (1) subsampling: T -> (T - 7)//2 + # (2) embedding: num_features -> encoder_dims + self.encoder_embed = Conv2dSubsampling( + num_features, encoder_dims[0], dropout=dropout + ) + + # each one will be ZipformerEncoder or DownsampledZipformerEncoder + encoders = [] + + self.num_encoders = len(encoder_dims) + for i in range(self.num_encoders): + encoder_layer = ZipformerEncoderLayer( + encoder_dims[i], + attention_dim[i], + nhead[i], + feedforward_dim[i], + dropout, + cnn_module_kernels[i], + pos_dim, + ) + + # For the segment of the warmup period, we let the Conv2dSubsampling + # layer learn something. Then we start to warm up the other encoders. + encoder = ZipformerEncoder( + encoder_layer, + num_encoder_layers[i], + dropout, + warmup_begin=warmup_batches * (i + 1) / (self.num_encoders + 1), + warmup_end=warmup_batches * (i + 2) / (self.num_encoders + 1), + ) + + if zipformer_downsampling_factors[i] != 1: + encoder = DownsampledZipformerEncoder( + encoder, + input_dim=encoder_dims[i - 1] if i > 0 else encoder_dims[0], + output_dim=encoder_dims[i], + downsample=zipformer_downsampling_factors[i], + ) + encoders.append(encoder) + self.encoders = nn.ModuleList(encoders) + + # initializes self.skip_layers and self.skip_modules + self._init_skip_modules() + + self.downsample_output = AttentionDownsample( + encoder_dims[-1], encoder_dims[-1], downsample=output_downsampling_factor + ) + + def _get_layer_skip_dropout_prob(self): + if not self.training: + return 0.0 + batch_count = self.batch_count + min_dropout_prob = 0.025 + + if batch_count > self.warmup_end: + return min_dropout_prob + else: + return 0.5 - (batch_count / self.warmup_end) * (0.5 - min_dropout_prob) + + def _init_skip_modules(self): + """ + If self.zipformer_downampling_factors = (1, 2, 4, 8, 4, 2), then at the input of layer + indexed 4 (in zero indexing), with has subsapling_factor=4, we combine the output of + layers 2 and 3; and at the input of layer indexed 5, which which has subsampling_factor=2, + we combine the outputs of layers 1 and 5. + """ + skip_layers = [] + skip_modules = [] + z = self.zipformer_downsampling_factors + for i in range(len(z)): + if i <= 1 or z[i - 1] <= z[i]: + skip_layers.append(None) + skip_modules.append(SimpleCombinerIdentity()) + else: + # TEMP + for j in range(i - 2, -1, -1): + if z[j] <= z[i] or j == 0: + # TEMP logging statement. + logging.info( + f"At encoder stack {i}, which has downsampling_factor={z[i]}, we will " + f"combine the outputs of layers {j} and {i-1}, with downsampling_factors={z[j]} and {z[i-1]}." + ) + skip_layers.append(j) + skip_modules.append( + SimpleCombiner( + self.encoder_dims[j], + self.encoder_dims[i - 1], + min_weight=(0.0, 0.25), + ) + ) + break + self.skip_layers = skip_layers + self.skip_modules = nn.ModuleList(skip_modules) + + def get_feature_masks(self, x: torch.Tensor) -> List[float]: + # Note: The actual return type is Union[List[float], List[Tensor]], + # but to make torch.jit.script() work, we use List[float] + """ + 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 + (num_frames, batch_size, encoder_dims0) + """ + num_encoders = len(self.encoder_dims) + if torch.jit.is_scripting() or not self.training: + return [1.0] * num_encoders + + (num_frames0, batch_size, _encoder_dims0) = x.shape + + assert self.encoder_dims[0] == _encoder_dims0, ( + self.encoder_dims, + _encoder_dims0, + ) + + max_downsampling_factor = max(self.zipformer_downsampling_factors) + + num_frames_max = num_frames0 + max_downsampling_factor - 1 + + feature_mask_dropout_prob = 0.15 + + # frame_mask_max shape: (num_frames_max, batch_size, 1) + frame_mask_max = ( + torch.rand(num_frames_max, batch_size, 1, device=x.device) + > feature_mask_dropout_prob + ).to(x.dtype) + + feature_masks = [] + for i in range(num_encoders): + ds = self.zipformer_downsampling_factors[i] + upsample_factor = max_downsampling_factor // ds + + frame_mask = ( + frame_mask_max.unsqueeze(1) + .expand(num_frames_max, upsample_factor, batch_size, 1) + .reshape(num_frames_max * upsample_factor, batch_size, 1) + ) + num_frames = (num_frames0 + ds - 1) // ds + frame_mask = frame_mask[:num_frames] + feature_mask = torch.ones( + num_frames, + batch_size, + self.encoder_dims[i], + dtype=x.dtype, + device=x.device, + ) + u = self.encoder_unmasked_dims[i] + feature_mask[:, :, u:] *= frame_mask + feature_masks.append(feature_mask) + + return feature_masks + + def forward( + self, + x: torch.Tensor, + x_lens: torch.Tensor, + ) -> Tuple[torch.Tensor, torch.Tensor]: + """ + Args: + x: + The input tensor. Its shape is (batch_size, seq_len, feature_dim). + x_lens: + A tensor of shape (batch_size,) containing the number of frames in + `x` before padding. + chunk_size: + The chunk size used in evaluation mode. + Returns: + Return a tuple containing 2 tensors: + - embeddings: its shape is (batch_size, output_seq_len, encoder_dims[-1]) + - lengths, a tensor of shape (batch_size,) containing the number + of frames in `embeddings` before padding. + """ + x = self.encoder_embed(x) + + x = x.permute(1, 0, 2) # (N, T, C) -> (T, N, C) + + lengths = (x_lens - 7) >> 1 + assert x.size(0) == lengths.max().item(), (x.shape, lengths, lengths.max()) + mask = make_pad_mask(lengths) + + outputs = [] + feature_masks = self.get_feature_masks(x) + + if self.training: + # Training mode + max_ds = max(self.zipformer_downsampling_factors) + # Generate dynamic chunk-wise attention mask during training + max_len = x.size(0) // max_ds + short_chunk_size = self.short_chunk_size // max_ds + chunk_size = torch.randint(1, max_len, (1,)).item() + if chunk_size > (max_len * self.short_chunk_threshold): + # Full attention + chunk_size = x.size(0) + else: + # Chunk-wise attention + chunk_size = chunk_size % short_chunk_size + 1 + chunk_size *= max_ds + else: + chunk_size = self.decode_chunk_size + # Evaluation mode + for ds in self.zipformer_downsampling_factors: + assert chunk_size % ds == 0, (chunk_size, ds) + + attn_mask = ~subsequent_chunk_mask( + size=x.size(0), + chunk_size=chunk_size, + num_left_chunks=self.num_left_chunks, + device=x.device, + ) + + for i, (module, skip_module) in enumerate( + zip(self.encoders, self.skip_modules) + ): + ds = self.zipformer_downsampling_factors[i] + k = self.skip_layers[i] + if isinstance(k, int): + layer_skip_dropout_prob = self._get_layer_skip_dropout_prob() + if torch.jit.is_scripting(): + x = skip_module(outputs[k], x) + elif (not self.training) or random.random() > layer_skip_dropout_prob: + x = skip_module(outputs[k], x) + x = module( + x, + feature_mask=feature_masks[i], + src_key_padding_mask=None if mask is None else mask[..., ::ds], + attn_mask=attn_mask[::ds, ::ds], + ) + outputs.append(x) + + x = self.downsample_output(x) + # class Downsample has this rounding behavior.. + assert self.output_downsampling_factor == 2, self.output_downsampling_factor + lengths = (lengths + 1) >> 1 + + x = x.permute(1, 0, 2) # (T, N, C) ->(N, T, C) + + return x, lengths + + def streaming_forward( + self, + x: torch.Tensor, + x_lens: torch.Tensor, + states: List[Tensor], + ) -> Tuple[Tensor, Tensor, List[Tensor]]: + """ + Args: + x: + The input tensor. Its shape is (batch_size, seq_len, feature_dim). + seq_len is the input chunk length. + x_lens: + A tensor of shape (batch_size,) containing the number of frames in + `x` before padding. + states: + A list of 7 * num_encoders elements: + ``states[0:num_encoders]`` is the cached numbers of past frames. + ``states[num_encoders:2*num_encoders]`` is the cached average tensors. + ``states[2*num_encoders:3*num_encoders]`` is the cached key tensors of the first attention modules. + ``states[3*num_encoders:4*num_encoders]`` is the cached value tensors of the first attention modules. + ``states[4*num_encoders:5*num_encoders]`` is the cached value tensors of the second attention modules. + ``states[5*num_encoders:6*num_encoders]`` is the cached left contexts of the first convolution modules. + ``states[6*num_encoders:7*num_encoders]`` is the cached left contexts of the second convolution modules. + + Returns: + Return a tuple containing 3 tensors: + - embeddings: its shape is (batch_size, output_seq_len, encoder_dims[-1]) + - lengths, a tensor of shape (batch_size,) containing the number + of frames in `embeddings` before padding. + - updated states. + """ + assert len(states) == 7 * self.num_encoders, (len(states), self.num_encoders) + + cached_len = states[: self.num_encoders] + cached_avg = states[self.num_encoders : 2 * self.num_encoders] + cached_key = states[2 * self.num_encoders : 3 * self.num_encoders] + cached_val = states[3 * self.num_encoders : 4 * self.num_encoders] + cached_val2 = states[4 * self.num_encoders : 5 * self.num_encoders] + cached_conv1 = states[5 * self.num_encoders : 6 * self.num_encoders] + cached_conv2 = states[6 * self.num_encoders : 7 * self.num_encoders] + + x = self.encoder_embed(x) + x = x.permute(1, 0, 2) # (N, T, C) -> (T, N, C) + lengths = (x_lens - 7) >> 1 + assert x.size(0) == lengths.max().item(), (x.shape, lengths, lengths.max()) + + outputs = [] + new_cached_len = [] + new_cached_avg = [] + new_cached_key = [] + new_cached_val = [] + new_cached_val2 = [] + new_cached_conv1 = [] + new_cached_conv2 = [] + + for i, (module, skip_module) in enumerate( + zip(self.encoders, self.skip_modules) + ): + k = self.skip_layers[i] + if isinstance(k, int): + x = skip_module(outputs[k], x) + x, len_avg, avg, key, val, val2, conv1, conv2 = module.streaming_forward( + x, + cached_len=cached_len[i], + cached_avg=cached_avg[i], + cached_key=cached_key[i], + cached_val=cached_val[i], + cached_val2=cached_val2[i], + cached_conv1=cached_conv1[i], + cached_conv2=cached_conv2[i], + ) + outputs.append(x) + # Update caches + new_cached_len.append(len_avg) + new_cached_avg.append(avg) + new_cached_key.append(key) + new_cached_val.append(val) + new_cached_val2.append(val2) + new_cached_conv1.append(conv1) + new_cached_conv2.append(conv2) + + x = self.downsample_output(x) + # class Downsample has this rounding behavior.. + assert self.output_downsampling_factor == 2, self.output_downsampling_factor + lengths = (lengths + 1) >> 1 + + x = x.permute(1, 0, 2) # (T, N, C) ->(N, T, C) + + new_states = ( + new_cached_len + + new_cached_avg + + new_cached_key + + new_cached_val + + new_cached_val2 + + new_cached_conv1 + + new_cached_conv2 + ) + return x, lengths, new_states + + @torch.jit.export + def get_init_state( + self, + device: torch.device = torch.device("cpu"), + ) -> List[Tensor]: + """Get initial states. + A list of 7 * num_encoders elements: + ``states[0:num_encoders]`` is the cached numbers of past frames. + ``states[num_encoders:2*num_encoders]`` is the cached average tensors. + ``states[2*num_encoders:3*num_encoders]`` is the cached key tensors of the first attention modules. + ``states[3*num_encoders:4*num_encoders]`` is the cached value tensors of the first attention modules. + ``states[4*num_encoders:5*num_encoders]`` is the cached value tensors of the second attention modules. + ``states[5*num_encoders:6*num_encoders]`` is the cached left contexts of the first convolution modules. + ``states[6*num_encoders:7*num_encoders]`` is the cached left contexts of the second convolution modules. + """ + cached_len = [] + cached_avg = [] + cached_key = [] + cached_val = [] + cached_val2 = [] + cached_conv1 = [] + cached_conv2 = [] + + left_context_len = self.decode_chunk_size * self.num_left_chunks + + for i, encoder in enumerate(self.encoders): + num_layers = encoder.num_layers + ds = self.zipformer_downsampling_factors[i] + + len_avg = torch.zeros(num_layers, 1, dtype=torch.int32, device=device) + cached_len.append(len_avg) + + avg = torch.zeros(num_layers, 1, encoder.d_model, device=device) + cached_avg.append(avg) + + key = torch.zeros( + num_layers, + left_context_len // ds, + 1, + encoder.attention_dim, + device=device, + ) + cached_key.append(key) + + val = torch.zeros( + num_layers, + left_context_len // ds, + 1, + encoder.attention_dim // 2, + device=device, + ) + cached_val.append(val) + + val2 = torch.zeros( + num_layers, + left_context_len // ds, + 1, + encoder.attention_dim // 2, + device=device, + ) + cached_val2.append(val2) + + conv1 = torch.zeros( + num_layers, + 1, + encoder.d_model, + encoder.cnn_module_kernel - 1, + device=device, + ) + cached_conv1.append(conv1) + + conv2 = torch.zeros( + num_layers, + 1, + encoder.d_model, + encoder.cnn_module_kernel - 1, + device=device, + ) + cached_conv2.append(conv2) + + states = ( + cached_len + + cached_avg + + cached_key + + cached_val + + cached_val2 + + cached_conv1 + + cached_conv2 + ) + return states + + +class ZipformerEncoderLayer(nn.Module): + """ + ZipformerEncoderLayer is made up of self-attn, feedforward and convolution networks. + + Args: + d_model: 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 = ZipformerEncoderLayer(d_model=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, + d_model: int, + attention_dim: int, + nhead: int, + feedforward_dim: int = 2048, + dropout: float = 0.1, + cnn_module_kernel: int = 31, + pos_dim: int = 4, + ) -> None: + super(ZipformerEncoderLayer, self).__init__() + + self.d_model = d_model + self.attention_dim = attention_dim + self.cnn_module_kernel = cnn_module_kernel + + # will be written to, see set_batch_count() + self.batch_count = 0 + + self.self_attn = RelPositionMultiheadAttention( + d_model, + attention_dim, + nhead, + pos_dim, + dropout=0.0, + ) + + self.pooling = PoolingModule(d_model) + + self.feed_forward1 = FeedforwardModule(d_model, feedforward_dim, dropout) + + self.feed_forward2 = FeedforwardModule(d_model, feedforward_dim, dropout) + + self.feed_forward3 = FeedforwardModule(d_model, feedforward_dim, dropout) + + self.conv_module1 = ConvolutionModule(d_model, cnn_module_kernel) + + self.conv_module2 = ConvolutionModule(d_model, cnn_module_kernel) + + self.norm_final = BasicNorm(d_model) + + self.bypass_scale = nn.Parameter(torch.tensor(0.5)) + + # try to ensure the output is close to zero-mean (or at least, zero-median). + self.balancer = ActivationBalancer( + d_model, + channel_dim=-1, + min_positive=0.45, + max_positive=0.55, + max_abs=6.0, + ) + self.whiten = Whiten( + num_groups=1, whitening_limit=5.0, prob=(0.025, 0.25), grad_scale=0.01 + ) + + def get_bypass_scale(self): + if torch.jit.is_scripting() or not self.training: + return self.bypass_scale + if random.random() < 0.1: + # ensure we get grads if self.bypass_scale becomes out of range + return self.bypass_scale + # hardcode warmup period for bypass scale + warmup_period = 20000.0 + initial_clamp_min = 0.75 + final_clamp_min = 0.25 + if self.batch_count > warmup_period: + clamp_min = final_clamp_min + else: + clamp_min = initial_clamp_min - (self.batch_count / warmup_period) * ( + initial_clamp_min - final_clamp_min + ) + return self.bypass_scale.clamp(min=clamp_min, max=1.0) + + def get_dynamic_dropout_rate(self): + # return dropout rate for the dynamic modules (self_attn, pooling, convolution); this + # starts at 0.2 and rapidly decreases to 0. Its purpose is to keep the training stable + # at the beginning, by making the network focus on the feedforward modules. + if torch.jit.is_scripting() or not self.training: + return 0.0 + warmup_period = 2000.0 + initial_dropout_rate = 0.2 + final_dropout_rate = 0.0 + if self.batch_count > warmup_period: + return final_dropout_rate + else: + return initial_dropout_rate - ( + initial_dropout_rate * final_dropout_rate + ) * (self.batch_count / warmup_period) + + def forward( + self, + src: Tensor, + pos_emb: Tensor, + 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 layer (required). + pos_emb: Positional embedding tensor (required). + src_mask: the mask for the src sequence (optional). + src_key_padding_mask: the mask for the src keys per batch (optional). + batch_split: if not None, this layer will only be applied to + + Shape: + src: (S, N, E). + pos_emb: (N, 2*S-1, E) + src_mask: (S, S). + src_key_padding_mask: (N, S). + S is the source sequence length, N is the batch size, E is the feature number + """ + src_orig = src + + # macaron style feed forward module + src = src + self.feed_forward1(src) + + # dropout rate for submodules that interact with time. + dynamic_dropout = self.get_dynamic_dropout_rate() + + # pooling module + if torch.jit.is_scripting(): + src = src + self.pooling(src, src_key_padding_mask=src_key_padding_mask) + elif random.random() >= dynamic_dropout: + src = src + self.pooling(src, src_key_padding_mask=src_key_padding_mask) + + if torch.jit.is_scripting(): + src_att, attn_weights = self.self_attn( + src, + pos_emb=pos_emb, + attn_mask=attn_mask, + key_padding_mask=src_key_padding_mask, + ) + src = src + src_att + + src = src + self.conv_module1( + src, src_key_padding_mask=src_key_padding_mask + ) + + src = src + self.feed_forward2(src) + + src = src + self.self_attn.forward2(src, attn_weights) + + src = src + self.conv_module2( + src, src_key_padding_mask=src_key_padding_mask + ) + else: + use_self_attn = random.random() >= dynamic_dropout + if use_self_attn: + src_att, attn_weights = self.self_attn( + src, + pos_emb=pos_emb, + attn_mask=attn_mask, + key_padding_mask=src_key_padding_mask, + ) + src = src + src_att + + if random.random() >= dynamic_dropout: + src = src + self.conv_module1( + src, src_key_padding_mask=src_key_padding_mask + ) + + src = src + self.feed_forward2(src) + + if use_self_attn: + src = src + self.self_attn.forward2(src, attn_weights) + + if random.random() >= dynamic_dropout: + src = src + self.conv_module2( + src, src_key_padding_mask=src_key_padding_mask + ) + + src = src + self.feed_forward3(src) + + src = self.norm_final(self.balancer(src)) + + delta = src - src_orig + + src = src_orig + delta * self.get_bypass_scale() + + return self.whiten(src) + + def streaming_forward( + self, + src: Tensor, + pos_emb: Tensor, + cached_len: Tensor, + cached_avg: Tensor, + cached_key: Tensor, + cached_val: Tensor, + cached_val2: Tensor, + cached_conv1: Tensor, + cached_conv2: Tensor, + ) -> Tuple[Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor]: + """ + Pass the input through the encoder layer. + + Args: + src: the sequence to the encoder layer (required). + pos_emb: Positional embedding tensor (required). + cached_len: processed number of past frames. + cached_avg: cached average of past frames. + cached_key: cached key tensor of left context for the first attention module. + cached_val: cached value tensor of left context for the first attention module. + cached_val2: cached value tensor of left context for the second attention module. + cached_conv1: cached left context for the first convolution module. + cached_conv2: cached left context for the second convolution module. + + Shape: + src: (S, N, E). + pos_emb: (N, left_context_len+2*S-1, E) + cached_len: (N,) + N is the batch size. + cached_avg: (N, C). + N is the batch size, C is the feature dimension. + cached_key: (left_context_len, N, K). + N is the batch size, K is the key dimension. + cached_val: (left_context_len, N, V). + N is the batch size, V is the key dimension. + cached_val2: (left_context_len, N, V). + N is the batch size, V is the key dimension. + cached_conv1: (N, C, kernel_size-1). + N is the batch size, C is the convolution channels. + cached_conv2: (N, C, kernel_size-1). + N is the batch size, C is the convolution channels. + """ + src_orig = src + + # macaron style feed forward module + src = src + self.feed_forward1(src) + + src_pool, cached_len, cached_avg = self.pooling.streaming_forward( + src, + cached_len=cached_len, + cached_avg=cached_avg, + ) + src = src + src_pool + + ( + src_attn, + attn_weights, + cached_key, + cached_val, + ) = self.self_attn.streaming_forward( + src, + pos_emb=pos_emb, + cached_key=cached_key, + cached_val=cached_val, + ) + src = src + src_attn + + src_conv, cached_conv1 = self.conv_module1.streaming_forward( + src, + cache=cached_conv1, + ) + src = src + src_conv + + src = src + self.feed_forward2(src) + + src_attn, cached_val2 = self.self_attn.streaming_forward2( + src, + attn_weights, + cached_val=cached_val2, + ) + src = src + src_attn + + src_conv, cached_conv2 = self.conv_module2.streaming_forward( + src, + cache=cached_conv2, + ) + src = src + src_conv + + src = src + self.feed_forward3(src) + + src = self.norm_final(self.balancer(src)) + + delta = src - src_orig + + src = src_orig + delta * self.bypass_scale + + return ( + src, + cached_len, + cached_avg, + cached_key, + cached_val, + cached_val2, + cached_conv1, + cached_conv2, + ) + + +class ZipformerEncoder(nn.Module): + r"""ZipformerEncoder is a stack of N encoder layers + + Args: + encoder_layer: an instance of the ZipformerEncoderLayer() class (required). + num_layers: the number of sub-encoder-layers in the encoder (required). + + Examples:: + >>> encoder_layer = ZipformerEncoderLayer(d_model=512, nhead=8) + >>> zipformer_encoder = ZipformerEncoder(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, + dropout: float, + warmup_begin: float, + warmup_end: float, + ) -> None: + super().__init__() + # will be written to, see set_batch_count() Note: in inference time this + # may be zero but should be treated as large, we can check if + # self.training is true. + self.batch_count = 0 + self.warmup_begin = warmup_begin + self.warmup_end = warmup_end + # module_seed is for when we need a random number that is unique to the module but + # shared across jobs. It's used to randomly select how many layers to drop, + # so that we can keep this consistent across worker tasks (for efficiency). + self.module_seed = torch.randint(0, 1000, ()).item() + + self.encoder_pos = RelPositionalEncoding(encoder_layer.d_model, dropout) + + self.layers = nn.ModuleList( + [copy.deepcopy(encoder_layer) for i in range(num_layers)] + ) + self.num_layers = num_layers + + self.d_model = encoder_layer.d_model + self.attention_dim = encoder_layer.attention_dim + self.cnn_module_kernel = encoder_layer.cnn_module_kernel + + assert 0 <= warmup_begin <= warmup_end, (warmup_begin, warmup_end) + + delta = (1.0 / num_layers) * (warmup_end - warmup_begin) + cur_begin = warmup_begin + for i in range(num_layers): + self.layers[i].warmup_begin = cur_begin + cur_begin += delta + self.layers[i].warmup_end = cur_begin + + def get_layers_to_drop(self, rnd_seed: int): + ans = set() + if not self.training: + return ans + + batch_count = self.batch_count + num_layers = len(self.layers) + + def get_layerdrop_prob(layer: int) -> float: + layer_warmup_begin = self.layers[layer].warmup_begin + layer_warmup_end = self.layers[layer].warmup_end + + initial_layerdrop_prob = 0.5 + final_layerdrop_prob = 0.05 + + if batch_count == 0: + # As a special case, if batch_count == 0, return 0 (drop no + # layers). This is rather ugly, I'm afraid; it is intended to + # enable our scan_pessimistic_batches_for_oom() code to work correctly + # so if we are going to get OOM it will happen early. + # also search for 'batch_count' with quotes in this file to see + # how we initialize the warmup count to a random number between + # 0 and 10. + return 0.0 + elif batch_count < layer_warmup_begin: + return initial_layerdrop_prob + elif batch_count > layer_warmup_end: + return final_layerdrop_prob + else: + # linearly interpolate + t = (batch_count - layer_warmup_begin) / layer_warmup_end + assert 0.0 <= t < 1.001, t + return initial_layerdrop_prob + t * ( + final_layerdrop_prob - initial_layerdrop_prob + ) + + shared_rng = random.Random(batch_count + self.module_seed) + independent_rng = random.Random(rnd_seed) + + layerdrop_probs = [get_layerdrop_prob(i) for i in range(num_layers)] + tot = sum(layerdrop_probs) + # Instead of drawing the samples independently, we first randomly decide + # how many layers to drop out, using the same random number generator between + # jobs so that all jobs drop out the same number (this is for speed). + # Then we use an approximate approach to drop out the individual layers + # with their specified probs while reaching this exact target. + num_to_drop = int(tot) + int(shared_rng.random() < (tot - int(tot))) + + layers = list(range(num_layers)) + independent_rng.shuffle(layers) + + # go through the shuffled layers until we get the required number of samples. + if num_to_drop > 0: + for layer in itertools.cycle(layers): + if independent_rng.random() < layerdrop_probs[layer]: + ans.add(layer) + if len(ans) == num_to_drop: + break + if shared_rng.random() < 0.005 or __name__ == "__main__": + logging.info( + f"warmup_begin={self.warmup_begin:.1f}, warmup_end={self.warmup_end:.1f}, " + f"batch_count={batch_count:.1f}, num_to_drop={num_to_drop}, layers_to_drop={ans}" + ) + return ans + + def forward( + self, + src: Tensor, + # Note: The type of feature_mask should be Union[float, Tensor], + # but to make torch.jit.script() work, we use `float` here + feature_mask: 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). + feature_mask: something that broadcasts with src, that we'll multiply `src` + by at every layer. + mask: the mask for the src sequence (optional). + src_key_padding_mask: the mask for the src keys per batch (optional). + + Shape: + src: (S, N, E). + pos_emb: (N, 2*S-1, E) + mask: (S, S). + src_key_padding_mask: (N, S). + S is the source sequence length, T is the target sequence length, N is the batch size, E is the feature number + + Returns: (x, x_no_combine), both of shape (S, N, E) + """ + pos_emb = self.encoder_pos(src) + output = src + + if torch.jit.is_scripting(): + layers_to_drop = [] + else: + rnd_seed = src.numel() + random.randint(0, 1000) + layers_to_drop = self.get_layers_to_drop(rnd_seed) + + output = output * feature_mask + + for i, mod in enumerate(self.layers): + if not torch.jit.is_scripting(): + if i in layers_to_drop: + continue + output = mod( + output, + pos_emb, + attn_mask=attn_mask, + src_key_padding_mask=src_key_padding_mask, + ) + + output = output * feature_mask + + return output + + @torch.jit.export + def streaming_forward( + self, + src: Tensor, + cached_len: Tensor, + cached_avg: Tensor, + cached_key: Tensor, + cached_val: Tensor, + cached_val2: Tensor, + cached_conv1: Tensor, + cached_conv2: Tensor, + ) -> Tuple[Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor]: + r"""Pass the input through the encoder layers in turn. + + Args: + src: the sequence to the encoder (required). + cached_len: number of past frames. + cached_avg: cached average of past frames. + cached_key: cached key tensor for first attention module. + cached_val: cached value tensor for first attention module. + cached_val2: cached value tensor for second attention module. + cached_conv1: cached left contexts for the first convolution module. + cached_conv2: cached left contexts for the second convolution module. + + Shape: + src: (S, N, E). + cached_len: (N,) + N is the batch size. + cached_avg: (num_layers, N, C). + N is the batch size, C is the feature dimension. + cached_key: (num_layers, left_context_len, N, K). + N is the batch size, K is the key dimension. + cached_val: (num_layers, left_context_len, N, V). + N is the batch size, V is the key dimension. + cached_val2: (num_layers, left_context_len, N, V). + N is the batch size, V is the key dimension. + cached_conv1: (num_layers, N, C, kernel_size-1). + N is the batch size, C is the convolution channels. + cached_conv2: (num_layers, N, C, kernel_size-1). + N is the batch size, C is the convolution channels. + + Returns: A tuple of 8 tensors: + - output tensor + - updated cached number of past frmaes. + - updated cached average of past frmaes. + - updated cached key tensor of of the first attention module. + - updated cached value tensor of of the first attention module. + - updated cached value tensor of of the second attention module. + - updated cached left contexts of the first convolution module. + - updated cached left contexts of the second convolution module. + """ + assert cached_len.size(0) == self.num_layers, ( + cached_len.size(0), + self.num_layers, + ) + assert cached_avg.size(0) == self.num_layers, ( + cached_avg.size(0), + self.num_layers, + ) + assert cached_key.size(0) == self.num_layers, ( + cached_key.size(0), + self.num_layers, + ) + assert cached_val.size(0) == self.num_layers, ( + cached_val.size(0), + self.num_layers, + ) + assert cached_val2.size(0) == self.num_layers, ( + cached_val2.size(0), + self.num_layers, + ) + assert cached_conv1.size(0) == self.num_layers, ( + cached_conv1.size(0), + self.num_layers, + ) + assert cached_conv2.size(0) == self.num_layers, ( + cached_conv2.size(0), + self.num_layers, + ) + + left_context_len = cached_key.shape[1] + pos_emb = self.encoder_pos(src, left_context_len) + output = src + + new_cached_len = [] + new_cached_avg = [] + new_cached_key = [] + new_cached_val = [] + new_cached_val2 = [] + new_cached_conv1 = [] + new_cached_conv2 = [] + for i, mod in enumerate(self.layers): + output, len_avg, avg, key, val, val2, conv1, conv2 = mod.streaming_forward( + output, + pos_emb, + cached_len=cached_len[i], + cached_avg=cached_avg[i], + cached_key=cached_key[i], + cached_val=cached_val[i], + cached_val2=cached_val2[i], + cached_conv1=cached_conv1[i], + cached_conv2=cached_conv2[i], + ) + # Update caches + new_cached_len.append(len_avg) + new_cached_avg.append(avg) + new_cached_key.append(key) + new_cached_val.append(val) + new_cached_val2.append(val2) + new_cached_conv1.append(conv1) + new_cached_conv2.append(conv2) + + return ( + output, + torch.stack(new_cached_len, dim=0), + torch.stack(new_cached_avg, dim=0), + torch.stack(new_cached_key, dim=0), + torch.stack(new_cached_val, dim=0), + torch.stack(new_cached_val2, dim=0), + torch.stack(new_cached_conv1, dim=0), + torch.stack(new_cached_conv2, dim=0), + ) + + +class DownsampledZipformerEncoder(nn.Module): + r""" + DownsampledZipformerEncoder 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, input_dim: int, output_dim: int, downsample: int + ): + super(DownsampledZipformerEncoder, self).__init__() + self.downsample_factor = downsample + self.downsample = AttentionDownsample(input_dim, output_dim, downsample) + self.encoder = encoder + self.num_layers = encoder.num_layers + self.d_model = encoder.d_model + self.attention_dim = encoder.attention_dim + self.cnn_module_kernel = encoder.cnn_module_kernel + self.upsample = SimpleUpsample(output_dim, downsample) + self.out_combiner = SimpleCombiner( + input_dim, output_dim, min_weight=(0.0, 0.25) + ) + + def forward( + self, + src: Tensor, + # Note: the type of feature_mask should be Unino[float, Tensor], + # but to make torch.jit.script() happ, we use float here + feature_mask: 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). + feature_mask: something that broadcasts with src, that we'll multiply `src` + by at every layer. feature_mask is expected to be already downsampled by + self.downsample_factor. + attn_mask: attention mask (optional). Should be downsampled already. + src_key_padding_mask: the mask for the src keys per batch (optional). Should be downsampled already. + + Shape: + src: (S, N, E). + attn_mask: (S, S). + src_key_padding_mask: (N, S). + S is the source sequence length, T is the target sequence length, N is the batch size, E is the feature number + + Returns: output of shape (S, N, F) where F is the number of output features + (output_dim to constructor) + """ + src_orig = src + src = self.downsample(src) + + src = self.encoder( + src, + 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, + cached_len: Tensor, + cached_avg: Tensor, + cached_key: Tensor, + cached_val: Tensor, + cached_val2: Tensor, + cached_conv1: Tensor, + cached_conv2: Tensor, + ) -> Tuple[Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor]: + r"""Downsample, go through encoder, upsample. + + Args: + src: the sequence to the encoder (required). + cached_avg: cached average value of past frames. + cached_len: length of past frames. + cached_key: cached key tensor for the first attention module. + cached_val: cached value tensor for the first attention module. + cached_val2: cached value tensor for the second attention module. + cached_conv1: cached left context for the first convolution module. + cached_conv2: cached left context for the second convolution module. + + Shape: + src: (S, N, E). + cached_len: (N,) + N is the batch size. + cached_avg: (num_layers, N, C). + N is the batch size, C is the feature dimension. + cached_key: (num_layers, left_context_len, N, K). + N is the batch size, K is the key dimension. + cached_val: (num_layers, left_context_len, N, V). + N is the batch size, V is the key dimension. + cached_val2: (num_layers, left_context_len, N, V). + N is the batch size, V is the key dimension. + cached_conv1: (num_layers, N, C, kernel_size-1). + N is the batch size, C is the convolution channels. + cached_conv2: (num_layers, N, C, kernel_size-1). + N is the batch size, C is the convolution channels. + Returns: output of shape (S, N, F) where F is the number of output features + (output_dim to constructor) + """ + src_orig = src + src = self.downsample(src) + + ( + src, + cached_len, + cached_avg, + cached_key, + cached_val, + cached_val2, + cached_conv1, + cached_conv2, + ) = self.encoder.streaming_forward( + src, + cached_len=cached_len, + cached_avg=cached_avg, + cached_key=cached_key, + cached_val=cached_val, + cached_val2=cached_val2, + cached_conv1=cached_conv1, + cached_conv2=cached_conv2, + ) + 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), + cached_len, + cached_avg, + cached_key, + cached_val, + cached_val2, + cached_conv1, + cached_conv2, + ) + + +class AttentionDownsample(torch.nn.Module): + """ + Does downsampling with attention, by weighted sum, and a projection.. + """ + + def __init__(self, in_channels: int, out_channels: int, downsample: int): + """ + Require out_channels > in_channels. + """ + super(AttentionDownsample, self).__init__() + self.query = nn.Parameter(torch.randn(in_channels) * (in_channels**-0.5)) + + # fill in the extra dimensions with a projection of the input + if out_channels > in_channels: + self.extra_proj = nn.Linear( + in_channels * downsample, out_channels - in_channels, bias=False + ) + else: + self.extra_proj = None + self.downsample = downsample + + def forward(self, src: Tensor) -> Tensor: + """ + x: (seq_len, 1, in_channels) + Returns a tensor of shape + ( (seq_len+downsample-1)//downsample, batch_size, out_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 + if seq_len != d_seq_len * ds: + # 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.shape[0], d_seq_len, ds) + + src = src.reshape(d_seq_len, ds, batch_size, in_channels) + scores = (src * self.query).sum(dim=-1, keepdim=True) + + if not torch.jit.is_scripting() and not torch.jit.is_tracing(): + scores = penalize_abs_values_gt(scores, limit=10.0, penalty=1.0e-04) + + weights = scores.softmax(dim=1) + + # ans1 is the first `in_channels` channels of the output + ans = (src * weights).sum(dim=1) + src = src.permute(0, 2, 1, 3).reshape(d_seq_len, batch_size, ds * in_channels) + + if self.extra_proj is not None: + ans2 = self.extra_proj(src) + ans = torch.cat((ans, ans2), dim=2) + 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.bias = nn.Parameter(torch.randn(upsample, num_channels) * 0.01) + + 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.bias.shape[0] + (seq_len, batch_size, num_channels) = src.shape + src = src.unsqueeze(1).expand(seq_len, upsample, batch_size, num_channels) + src = src + self.bias.unsqueeze(1) + src = src.reshape(seq_len * upsample, batch_size, num_channels) + return src + + +class SimpleCombinerIdentity(nn.Module): + def __init__(self, *args, **kwargs): + super().__init__() + + def forward(self, src1: Tensor, src2: Tensor) -> Tensor: + return src1 + + +class SimpleCombiner(torch.nn.Module): + """ + A very simple way of combining 2 vectors of 2 different dims, via a + learned weighted combination in the shared part of the dim. + Args: + dim1: the dimension of the first input, e.g. 256 + dim2: the dimension of the second input, e.g. 384. + The output will have the same dimension as dim2. + """ + + def __init__(self, dim1: int, dim2: int, min_weight: Tuple[float] = (0.0, 0.0)): + super(SimpleCombiner, self).__init__() + assert dim2 >= dim1, (dim2, dim1) + self.weight1 = nn.Parameter(torch.zeros(())) + self.min_weight = min_weight + + def forward(self, src1: Tensor, src2: Tensor) -> Tensor: + """ + src1: (*, dim1) + src2: (*, dim2) + + Returns: a tensor of shape (*, dim2) + """ + assert src1.shape[:-1] == src2.shape[:-1], (src1.shape, src2.shape) + + weight1 = self.weight1 + if not torch.jit.is_scripting(): + if ( + self.training + and random.random() < 0.25 + and self.min_weight != (0.0, 0.0) + ): + weight1 = weight1.clamp( + min=self.min_weight[0], max=1.0 - self.min_weight[1] + ) + + src1 = src1 * weight1 + src2 = src2 * (1.0 - weight1) + + src1_dim = src1.shape[-1] + src2_dim = src2.shape[-1] + if src1_dim != src2_dim: + if src1_dim < src2_dim: + src1 = torch.nn.functional.pad(src1, (0, src2_dim - src1_dim)) + else: + src1 = src1[:src2_dim] + + return src1 + src2 + + +class RelPositionalEncoding(torch.nn.Module): + """Relative positional encoding module. + + See : Appendix B in "Transformer-XL: Attentive Language Models Beyond a Fixed-Length Context" + Modified from https://github.com/espnet/espnet/blob/master/espnet/nets/pytorch_backend/transformer/embedding.py + + Args: + d_model: Embedding dimension. + dropout_rate: Dropout rate. + max_len: Maximum input length. + + """ + + def __init__( + self, + d_model: int, + dropout_rate: float, + max_len: int = 5000, + ) -> None: + """Construct a PositionalEncoding object.""" + super(RelPositionalEncoding, self).__init__() + self.d_model = d_model + self.dropout = torch.nn.Dropout(dropout_rate) + self.pe = None + 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.""" + x_size_left = 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(1) >= x_size_left * 2 - 1: + # Note: TorchScript doesn't implement operator== for torch.Device + if self.pe.dtype != x.dtype or str(self.pe.device) != str(x.device): + self.pe = self.pe.to(dtype=x.dtype, device=x.device) + return + # Suppose `i` means to the position of query vecotr and `j` means the + # position of key vector. We use position relative positions when keys + # are to the left (i>j) and negative relative positions otherwise (i Tensor: + """Add positional encoding. + + Args: + x (torch.Tensor): Input tensor (time, batch, `*`). + left_context_len: (int): Length of cached left context. + + Returns: + torch.Tensor: Encoded tensor (batch, left_context_len + 2*time-1, `*`). + + """ + self.extend_pe(x, left_context_len) + x_size_left = x.size(0) + left_context_len + pos_emb = self.pe[ + :, + self.pe.size(1) // 2 + - x_size_left + + 1 : self.pe.size(1) // 2 # noqa E203 + + x.size(0), + ] + return self.dropout(pos_emb) + + +class RelPositionMultiheadAttention(nn.Module): + r"""Multi-Head Attention layer with relative position encoding + + 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: total dimension of the model. + attention_dim: dimension in the attention module, may be less or more than embed_dim + but must be a multiple of num_heads. + num_heads: parallel attention heads. + dropout: a Dropout layer on attn_output_weights. Default: 0.0. + + Examples:: + + >>> rel_pos_multihead_attn = RelPositionMultiheadAttention(embed_dim, num_heads) + >>> attn_output, attn_output_weights = multihead_attn(query, key, value, pos_emb) + """ + + def __init__( + self, + embed_dim: int, + attention_dim: int, + num_heads: int, + pos_dim: int, + dropout: float = 0.0, + ) -> None: + super(RelPositionMultiheadAttention, self).__init__() + self.embed_dim = embed_dim + self.attention_dim = attention_dim + self.num_heads = num_heads + self.dropout = dropout + self.head_dim = attention_dim // num_heads + self.pos_dim = pos_dim + assert self.head_dim % 2 == 0, self.head_dim + assert self.head_dim * num_heads == attention_dim, ( + self.head_dim, + num_heads, + attention_dim, + ) + + # the initial_scale is supposed to take over the "scaling" factor of + # head_dim ** -0.5, dividing it between the query and key. + in_proj_dim = ( + 2 * attention_dim + + attention_dim // 2 # query, key + + pos_dim * num_heads # value + ) # positional encoding query + + self.in_proj = ScaledLinear( + embed_dim, in_proj_dim, bias=True, initial_scale=self.head_dim**-0.25 + ) + + # self.whiten_values is applied on the values in forward(); + # it just copies the keys but prevents low-rank distribution by modifying grads. + self.whiten_values = Whiten( + num_groups=num_heads, + whitening_limit=2.0, + prob=(0.025, 0.25), + grad_scale=0.025, + ) + self.whiten_keys = Whiten( + num_groups=num_heads, + whitening_limit=2.0, + prob=(0.025, 0.25), + grad_scale=0.025, + ) + + # linear transformation for positional encoding. + self.linear_pos = ScaledLinear( + embed_dim, num_heads * pos_dim, bias=False, initial_scale=0.05 + ) + + # the following are for diagnosics only, see --print-diagnostics option. + # they only copy their inputs. + self.copy_pos_query = Identity() + self.copy_query = Identity() + + self.out_proj = ScaledLinear( + attention_dim // 2, embed_dim, bias=True, initial_scale=0.05 + ) + + self.in_proj2 = nn.Linear(embed_dim, attention_dim // 2, bias=False) + self.out_proj2 = ScaledLinear( + attention_dim // 2, embed_dim, bias=True, initial_scale=0.05 + ) + # self.whiten_values2 is applied on the values in forward2() + self.whiten_values2 = Whiten( + num_groups=num_heads, + whitening_limit=2.0, + prob=(0.025, 0.25), + grad_scale=0.025, + ) + + def forward( + self, + x: Tensor, + pos_emb: Tensor, + key_padding_mask: Optional[Tensor] = None, + attn_mask: Optional[Tensor] = None, + ) -> Tuple[Tensor, Tensor]: + r""" + Args: + x: input to be projected to query, key, value + pos_emb: Positional embedding tensor + key_padding_mask: if provided, specified padding elements in the key will + be ignored by the attention. When given a binary mask and a value is True, + the corresponding value on the attention layer will be ignored. When given + a byte mask and a value is non-zero, the corresponding value on the attention + layer will be ignored + attn_mask: 2D or 3D mask that prevents attention to certain positions. A 2D mask will be broadcasted for all + the batches while a 3D mask allows to specify a different mask for the entries of each batch. + + Shape: + - Inputs: + - x: :math:`(L, N, E)` where L is the target sequence length, N is the batch size, E is + the embedding dimension. + - pos_emb: :math:`(N, 2*L-1, E)` where L is the target sequence length, N is the batch size, E is + the embedding dimension. + - key_padding_mask: :math:`(N, S)` where N is the batch size, S is the source sequence length. + If a ByteTensor is provided, the non-zero positions will be ignored while the position + with the zero positions will be unchanged. If a BoolTensor is provided, the positions with the + value of ``True`` will be ignored while the position with the value of ``False`` will be unchanged. + - attn_mask: 2D mask :math:`(L, S)` where L is the target sequence length, S is the source sequence length. + 3D mask :math:`(N*num_heads, L, S)` where N is the batch size, L is the target sequence length, + S is the source sequence length. attn_mask ensure that position i is allowed to attend the unmasked + positions. If a ByteTensor is provided, the non-zero positions are not allowed to attend + while the zero positions will be unchanged. If a BoolTensor is provided, positions with ``True`` + is not allowed to attend while ``False`` values will be unchanged. If a FloatTensor + is provided, it will be added to the attention weight. + + - Returns: (attn_output, attn_weights) + + - attn_output: :math:`(S, N, E)` where S is the sequence length, N is the batch size, + E is the embedding dimension. + - attn_weights: :math:`(N * N, S, S)` where N is the batch size, H is the num-heads + and S is the sequence length. + """ + x, weights = self.multi_head_attention_forward( + self.in_proj(x), + self.linear_pos(pos_emb), + self.attention_dim, + self.num_heads, + self.dropout, + self.out_proj.weight, + self.out_proj.bias, + training=self.training, + key_padding_mask=key_padding_mask, + attn_mask=attn_mask, + ) + return x, weights + + def streaming_forward( + self, + x: Tensor, + pos_emb: Tensor, + cached_key: Tensor, + cached_val: Tensor, + ) -> Tuple[Tensor, Tensor, Tensor, Tensor]: + r""" + Args: + x: input to be projected to query, key, value + pos_emb: Positional embedding tensor + attn_mask: 2D or 3D mask that prevents attention to certain positions. A 2D mask will be broadcasted for all + the batches while a 3D mask allows to specify a different mask for the entries of each batch. + + Shape: + - Inputs: + - x: :math:`(L, N, E)` where L is the target sequence length, N is the batch size, E is + the embedding dimension. + - pos_emb: :math:`(N, 2*L-1, E)` where L is the target sequence length, N is the batch size, E is + the embedding dimension. + - attn_mask: 2D mask :math:`(L, S)` where L is the target sequence length, S is the source sequence length. + 3D mask :math:`(N*num_heads, L, S)` where N is the batch size, L is the target sequence length, + S is the source sequence length. attn_mask ensure that position i is allowed to attend the unmasked + positions. If a ByteTensor is provided, the non-zero positions are not allowed to attend + while the zero positions will be unchanged. If a BoolTensor is provided, positions with ``True`` + is not allowed to attend while ``False`` values will be unchanged. If a FloatTensor + is provided, it will be added to the attention weight. + - cached_key: :math:`(left_context_len, N, K)`, where N is the batch size, K is the key dimension. + - cached_val: :math:`(left_context_len, N, V)`, where N is the batch size, V is the value dimension. + + - Returns: (attn_output, attn_weights, cached_key, cached_val) + + - attn_output: :math:`(S, N, E)` where S is the sequence length, N is the batch size, + E is the embedding dimension. + - attn_weights: :math:`(N * N, S, S)` where N is the batch size, H is the num-heads + and S is the sequence length. + - cached_key: :math:`(left_context_len, N, K)`, updated cached attention key tensor of + left context + - cached_val: :math:`(left_context_len, N, K)`, updated cached attention value tensor of + """ + ( + x, + weights, + cached_key, + cached_val, + ) = self.streaming_multi_head_attention_forward( + self.in_proj(x), + self.linear_pos(pos_emb), + self.attention_dim, + self.num_heads, + self.out_proj.weight, + self.out_proj.bias, + cached_key=cached_key, + cached_val=cached_val, + ) + return x, weights, cached_key, cached_val + + def multi_head_attention_forward( + self, + x_proj: Tensor, + pos: Tensor, + attention_dim: int, + num_heads: int, + dropout_p: float, + out_proj_weight: Tensor, + out_proj_bias: Tensor, + training: bool = True, + key_padding_mask: Optional[Tensor] = None, + attn_mask: Optional[Tensor] = None, + ) -> Tuple[Tensor, Tensor]: + r""" + Args: + x_proj: the projected input, to be split into query, key, value. + pos: head-specific biases arising from the positional embeddings. + attention_dim: dimension inside attention mechanism + num_heads: parallel attention heads. + dropout_p: probability of an element to be zeroed. + out_proj_weight, out_proj_bias: the output projection weight and bias. + training: apply dropout if is ``True``. + key_padding_mask: if provided, specified padding elements in the key will + be ignored by the attention. This is an binary mask. When the value is True, + the corresponding value on the attention layer will be filled with -inf. + attn_mask: 2D or 3D mask that prevents attention to certain positions. A 2D mask will be broadcasted for all + the batches while a 3D mask allows to specify a different mask for the entries of each batch. + + Shape: + Inputs: + - x: :math:`(L, N, 7 * A // 2)` where L is the target sequence length, N is the batch size, A is + the attention dimension. Will be split into (query, key, value, pos). + - pos: :math:`(N, 2*L-1, A//2)` or :math:`(1, 2*L-1, A//2)` where L is the sequence + length, N is the batch size, and A is the attention dim. + - key_padding_mask: :math:`(N, S)` where N is the batch size, S is the source sequence length. + If a ByteTensor is provided, the non-zero positions will be ignored while the zero positions + will be unchanged. If a BoolTensor is provided, the positions with the + value of ``True`` will be ignored while the position with the value of ``False`` will be unchanged. + - attn_mask: 2D mask :math:`(L, S)` where L is the target sequence length, S is the source sequence length. + 3D mask :math:`(N*num_heads, L, S)` where N is the batch size, L is the target sequence length, + S is the source sequence length. attn_mask ensures that position i is allowed to attend the unmasked + positions. If a ByteTensor is provided, the non-zero positions are not allowed to attend + while the zero positions will be unchanged. If a BoolTensor is provided, positions with ``True`` + are not allowed to attend while ``False`` values will be unchanged. If a FloatTensor + is provided, it will be added to the attention weight. + + Outputs: + - attn_output: :math:`(L, N, E)` where L is the target sequence length, N is the batch size, + E is the embedding dimension. + - attn_weights: :math:`(N * H, S, S)` where N is the batch size, + H is the num-heads, S is the sequence length. + """ + + seq_len, bsz, _ = x_proj.size() + + head_dim = attention_dim // num_heads + pos_dim = self.pos_dim # positional-encoding dim per head + assert ( + head_dim * num_heads == attention_dim + ), f"attention_dim must be divisible by num_heads: {head_dim}, {num_heads}, {attention_dim}" + + # self-attention + q = x_proj[..., 0:attention_dim] + k = x_proj[..., attention_dim : 2 * attention_dim] + value_dim = attention_dim // 2 + v = x_proj[..., 2 * attention_dim : 2 * attention_dim + value_dim] + # p is the position-encoding query, its dimension is num_heads*pos_dim.. + p = x_proj[..., 2 * attention_dim + value_dim :] + + k = self.whiten_keys(k) # does nothing in the forward pass. + v = self.whiten_values(v) # does nothing in the forward pass. + q = self.copy_query(q) # for diagnostics only, does nothing. + p = self.copy_pos_query(p) # for diagnostics only, does nothing. + + if attn_mask is not None: + assert ( + attn_mask.dtype == torch.float32 + or attn_mask.dtype == torch.float64 + or attn_mask.dtype == torch.float16 + or attn_mask.dtype == torch.uint8 + or attn_mask.dtype == torch.bool + ), "Only float, byte, and bool types are supported for attn_mask, not {}".format( + attn_mask.dtype + ) + if attn_mask.dtype == torch.uint8: + warnings.warn( + "Byte tensor for attn_mask is deprecated. Use bool tensor instead." + ) + attn_mask = attn_mask.to(torch.bool) + + if attn_mask.dim() == 2: + attn_mask = attn_mask.unsqueeze(0) + if list(attn_mask.size()) != [1, seq_len, seq_len]: + raise RuntimeError("The size of the 2D attn_mask is not correct.") + elif attn_mask.dim() == 3: + if list(attn_mask.size()) != [ + bsz * num_heads, + seq_len, + seq_len, + ]: + raise RuntimeError("The size of the 3D attn_mask is not correct.") + else: + raise RuntimeError( + "attn_mask's dimension {} is not supported".format(attn_mask.dim()) + ) + # attn_mask's dim is 3 now. + + # convert ByteTensor key_padding_mask to bool + if key_padding_mask is not None and key_padding_mask.dtype == torch.uint8: + warnings.warn( + "Byte tensor for key_padding_mask is deprecated. Use bool tensor instead." + ) + key_padding_mask = key_padding_mask.to(torch.bool) + + q = q.reshape(seq_len, bsz, num_heads, head_dim) + p = p.reshape(seq_len, bsz, num_heads, pos_dim) + k = k.reshape(seq_len, bsz, num_heads, head_dim) + v = v.reshape(seq_len, bsz * num_heads, head_dim // 2).transpose(0, 1) + + if key_padding_mask is not None: + assert key_padding_mask.size(0) == bsz, "{} == {}".format( + key_padding_mask.size(0), bsz + ) + assert key_padding_mask.size(1) == seq_len, "{} == {}".format( + key_padding_mask.size(1), seq_len + ) + + q = q.permute(1, 2, 0, 3) # (batch, head, time1, head_dim) + p = p.permute(1, 2, 0, 3) # (batch, head, time1, pos_dim) + k = k.permute(1, 2, 3, 0) # (batch, head, d_k, time2) + + seq_len2 = 2 * seq_len - 1 + pos = pos.reshape(1, seq_len2, num_heads, pos_dim).permute(0, 2, 3, 1) + # pos shape now: (batch, head, pos_dim, seq_len2) + + # (batch, head, time1, pos_dim) x (1, head, pos_dim, seq_len2) -> (batch, head, time1, seq_len2) + # [where seq_len2 represents relative position.] + pos_weights = torch.matmul(p, pos) + # the following .as_strided() expression converts the last axis of pos_weights 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. + pos_weights = pos_weights.as_strided( + (bsz, num_heads, seq_len, seq_len), + ( + pos_weights.stride(0), + pos_weights.stride(1), + pos_weights.stride(2) - pos_weights.stride(3), + pos_weights.stride(3), + ), + storage_offset=pos_weights.stride(3) * (seq_len - 1), + ) + + # caution: they are really scores at this point. + attn_output_weights = torch.matmul(q, k) + pos_weights + + if not torch.jit.is_scripting(): + if 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, a limit on entropy, because once the entropy + # gets very small gradients through the softmax can become very small, and + # some mechanisms like that become ineffective. + attn_output_weights = penalize_abs_values_gt( + attn_output_weights, limit=25.0, penalty=1.0e-04 + ) + + # attn_output_weights: (batch, head, time1, time2) + attn_output_weights = attn_output_weights.view( + bsz * num_heads, seq_len, seq_len + ) + + if attn_mask is not None: + if attn_mask.dtype == torch.bool: + attn_output_weights = attn_output_weights.masked_fill( + attn_mask, float("-inf") + ) + else: + attn_output_weights = attn_output_weights + attn_mask + + if key_padding_mask is not None: + attn_output_weights = attn_output_weights.view( + bsz, num_heads, seq_len, seq_len + ) + attn_output_weights = attn_output_weights.masked_fill( + key_padding_mask.unsqueeze(1).unsqueeze(2), + float("-inf"), + ) + attn_output_weights = attn_output_weights.view( + bsz * num_heads, seq_len, seq_len + ) + + # Using this version of softmax, defined in scaling.py, + # should save a little of the memory used in backprop by, if + # we are in automatic mixed precision mode (amp) == autocast, + # only storing the half-precision output for backprop purposes. + attn_output_weights = softmax(attn_output_weights, dim=-1) + + # If we are using chunk-wise attention mask and setting a limited + # num_left_chunks, the attention may only see the padding values which + # will also be masked out by `key_padding_mask`. At this circumstances, + # the whole column of `attn_output_weights` will be `-inf` + # (i.e. be `nan` after softmax). So we fill `0.0` at the masking + # positions to avoid invalid loss value below. + if ( + attn_mask is not None + and attn_mask.dtype == torch.bool + and key_padding_mask is not None + ): + if attn_mask.size(0) != 1: + attn_mask = attn_mask.view(bsz, num_heads, seq_len, seq_len) + combined_mask = attn_mask | key_padding_mask.unsqueeze(1).unsqueeze(2) + else: + # attn_mask.shape == (1, tgt_len, src_len) + combined_mask = attn_mask.unsqueeze(0) | key_padding_mask.unsqueeze( + 1 + ).unsqueeze(2) + + attn_output_weights = attn_output_weights.view( + bsz, num_heads, seq_len, seq_len + ) + attn_output_weights = attn_output_weights.masked_fill(combined_mask, 0.0) + attn_output_weights = attn_output_weights.view( + bsz * num_heads, seq_len, seq_len + ) + + attn_output_weights = nn.functional.dropout( + attn_output_weights, p=dropout_p, training=training + ) + + attn_output = torch.bmm(attn_output_weights, v) + assert list(attn_output.size()) == [bsz * num_heads, seq_len, head_dim // 2] + attn_output = ( + attn_output.transpose(0, 1) + .contiguous() + .view(seq_len, bsz, attention_dim // 2) + ) + attn_output = nn.functional.linear(attn_output, out_proj_weight, out_proj_bias) + + return attn_output, attn_output_weights + + def streaming_multi_head_attention_forward( + self, + x_proj: Tensor, + pos: Tensor, + attention_dim: int, + num_heads: int, + out_proj_weight: Tensor, + out_proj_bias: Tensor, + cached_key: Tensor, + cached_val: Tensor, + ) -> Tuple[Tensor, Tensor, Tensor, Tensor]: + r""" + Args: + x_proj: the projected input, to be split into query, key, value. + pos: head-specific biases arising from the positional embeddings. + attention_dim: dimension inside attention mechanism + num_heads: parallel attention heads. + out_proj_weight, out_proj_bias: the output projection weight and bias. + cached_key: cached attention key tensor of left context. + cached_val: cached attention value tensor of left context. + + Shape: + Inputs: + - x: :math:`(L, N, 7 * A // 2)` where L is the target sequence length, N is the batch size, A is + the attention dimension. Will be split into (query, key, value, pos). + - pos: :math:`(N, 2*L-1, A//2)` or :math:`(1, 2*L-1, A//2)` where L is the sequence + length, N is the batch size, and A is the attention dim. + If a ByteTensor is provided, the non-zero positions will be ignored while the zero positions + will be unchanged. If a BoolTensor is provided, the positions with the + value of ``True`` will be ignored while the position with the value of ``False`` will be unchanged. + + Outputs: + - attn_output: :math:`(L, N, E)` where L is the target sequence length, N is the batch size, + E is the embedding dimension. + - attn_weights: :math:`(N * H, S, S)` where N is the batch size, + H is the num-heads, S is the sequence length. + - cached_key: :math:`(left_context_len, N, K)`, updated cached attention key tensor of left context. + - cached_val: :math:`(left_context_len, N, K)`, updated cached attention value tensor of left context. + """ + + seq_len, bsz, _ = x_proj.size() + + head_dim = attention_dim // num_heads + pos_dim = self.pos_dim # positional-encoding dim per head + assert ( + head_dim * num_heads == attention_dim + ), f"attention_dim must be divisible by num_heads: {head_dim}, {num_heads}, {attention_dim}" + + # self-attention + q = x_proj[..., 0:attention_dim] + k = x_proj[..., attention_dim : 2 * attention_dim] + value_dim = attention_dim // 2 + v = x_proj[..., 2 * attention_dim : 2 * attention_dim + value_dim] + # p is the position-encoding query, its dimension is num_heads*pos_dim.. + p = x_proj[..., 2 * attention_dim + value_dim :] + + left_context_len = cached_key.shape[0] + assert left_context_len > 0, left_context_len + assert cached_key.shape[0] == cached_val.shape[0], ( + cached_key.shape, + cached_val.shape, + ) + # Pad cached left contexts + k = torch.cat([cached_key, k], dim=0) + v = torch.cat([cached_val, v], dim=0) + # Update cached left contexts + cached_key = k[-left_context_len:, ...] + cached_val = v[-left_context_len:, ...] + + # The length of key and value + kv_len = k.shape[0] + + q = q.reshape(seq_len, bsz, num_heads, head_dim) + p = p.reshape(seq_len, bsz, num_heads, pos_dim) + k = k.reshape(kv_len, bsz, num_heads, head_dim) + v = v.reshape(kv_len, bsz * num_heads, head_dim // 2).transpose(0, 1) + + q = q.permute(1, 2, 0, 3) # (batch, head, time1, head_dim) + p = p.permute(1, 2, 0, 3) # (batch, head, time1, pos_dim) + k = k.permute(1, 2, 3, 0) # (batch, head, d_k, time2) + + seq_len2 = 2 * seq_len - 1 + left_context_len + pos = pos.reshape(1, seq_len2, num_heads, pos_dim).permute(0, 2, 3, 1) + # pos shape now: (batch, head, pos_dim, seq_len2) + + # (batch, head, time1, pos_dim) x (1, head, pos_dim, seq_len2) -> (batch, head, time1, seq_len2) + # [where seq_len2 represents relative position.] + pos_weights = torch.matmul(p, pos) + # the following .as_strided() expression converts the last axis of pos_weights 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. + pos_weights = pos_weights.as_strided( + (bsz, num_heads, seq_len, kv_len), + ( + pos_weights.stride(0), + pos_weights.stride(1), + pos_weights.stride(2) - pos_weights.stride(3), + pos_weights.stride(3), + ), + storage_offset=pos_weights.stride(3) * (seq_len - 1), + ) + + # caution: they are really scores at this point. + attn_output_weights = torch.matmul(q, k) + pos_weights + + # attn_output_weights: (batch, head, time1, time2) + attn_output_weights = attn_output_weights.view(bsz * num_heads, seq_len, kv_len) + + # Using this version of softmax, defined in scaling.py, + # should save a little of the memory used in backprop by, if + # we are in automatic mixed precision mode (amp) == autocast, + # only storing the half-precision output for backprop purposes. + attn_output_weights = softmax(attn_output_weights, dim=-1) + + attn_output = torch.bmm(attn_output_weights, v) + assert list(attn_output.size()) == [bsz * num_heads, seq_len, head_dim // 2] + attn_output = ( + attn_output.transpose(0, 1) + .contiguous() + .view(seq_len, bsz, attention_dim // 2) + ) + attn_output = nn.functional.linear(attn_output, out_proj_weight, out_proj_bias) + + return attn_output, attn_output_weights, cached_key, cached_val + + def forward2( + self, + x: Tensor, + attn_weights: Tensor, + ) -> Tensor: + """ + Second forward function, where we re-use the attn_weights returned by the first forward function + but with different input. + Args: + x: input, of shape (seq_len, batch_size, embed_dim) + attn_weights: attention weights returned by forward(), of shape (batch_size * num_heads, seq_len, seq_len) + Returns: + output of the same shape as x, i.e. (seq_len, batch_size, embed_dim) + """ + num_heads = self.num_heads + (seq_len, bsz, embed_dim) = x.shape + head_dim = self.attention_dim // num_heads + # v: (tgt_len, bsz, embed_dim // 2) + v = self.in_proj2(x) + v = self.whiten_values2(v) # does nothing in the forward pass. + v = v.reshape(seq_len, bsz * num_heads, head_dim // 2).transpose(0, 1) + + # now v: (bsz * num_heads, seq_len, head_dim // 2) + attn_output = torch.bmm(attn_weights, v) + + if not torch.jit.is_scripting(): + if random.random() < 0.001 or __name__ == "__main__": + self._print_attn_stats(attn_weights, attn_output) + + # attn_output: (bsz * num_heads, seq_len, head_dim) + attn_output = ( + attn_output.transpose(0, 1) + .contiguous() + .view(seq_len, bsz, self.attention_dim // 2) + ) + # returned value is of shape (seq_len, bsz, embed_dim), like x. + return self.out_proj2(attn_output) + + def streaming_forward2( + self, + x: Tensor, + attn_weights: Tensor, + cached_val: Tensor, + ) -> Tuple[Tensor, Tensor]: + """ + Second forward function, where we re-use the attn_weights returned by the first forward function + but with different input. + Args: + x: input, of shape (seq_len, batch_size, embed_dim) + attn_weights: attention weights returned by forward(), of shape (batch_size * num_heads, seq_len, seq_len) + cached_val: cached attention value tensor of left context. + Returns: + - output of the same shape as x, i.e. (seq_len, batch_size, embed_dim) + - updated cached attention value tensor of left context. + """ + num_heads = self.num_heads + (seq_len, bsz, embed_dim) = x.shape + head_dim = self.attention_dim // num_heads + # v: (tgt_len, bsz, embed_dim // 2) + v = self.in_proj2(x) + + left_context_len = cached_val.shape[0] + assert left_context_len > 0, left_context_len + v = torch.cat([cached_val, v], dim=0) + cached_val = v[-left_context_len:] + + seq_len2 = left_context_len + seq_len + v = v.reshape(seq_len2, bsz * num_heads, head_dim // 2).transpose(0, 1) + + # now v: (bsz * num_heads, seq_len, head_dim // 2) + attn_output = torch.bmm(attn_weights, v) + + # attn_output: (bsz * num_heads, seq_len, head_dim) + attn_output = ( + attn_output.transpose(0, 1) + .contiguous() + .view(seq_len, bsz, self.attention_dim // 2) + ) + # returned value is of shape (seq_len, bsz, embed_dim), like x. + return self.out_proj2(attn_output), cached_val + + def _print_attn_stats(self, attn_weights: Tensor, attn_output: Tensor): + # attn_weights: (batch_size * num_heads, seq_len, seq_len) + # attn_output: (bsz * num_heads, seq_len, head_dim) + (n, seq_len, head_dim) = attn_output.shape + num_heads = self.num_heads + bsz = n // num_heads + + with torch.no_grad(): + with torch.cuda.amp.autocast(enabled=False): + attn_weights = attn_weights.to(torch.float32) + attn_output = attn_output.to(torch.float32) + attn_weights_entropy = ( + -((attn_weights + 1.0e-20).log() * attn_weights) + .sum(dim=-1) + .reshape(bsz, num_heads, seq_len) + .mean(dim=(0, 2)) + ) + attn_output = attn_output.reshape(bsz, num_heads, seq_len, head_dim) + attn_output = attn_output.permute(1, 0, 2, 3).reshape( + num_heads, bsz * seq_len, head_dim + ) + attn_output_mean = attn_output.mean(dim=1, keepdim=True) + attn_output = attn_output - attn_output_mean + attn_covar = torch.matmul(attn_output.transpose(1, 2), attn_output) / ( + bsz * seq_len + ) + # attn_covar: (num_heads, head_dim, head_dim) + # eigs, _ = torch.symeig(attn_covar) + # logging.info(f"attn_weights_entropy = {attn_weights_entropy}, output_eigs = {eigs}") + + attn_covar = _diag(attn_covar).mean(dim=1) # (num_heads,) + embed_dim = self.in_proj2.weight.shape[1] + in_proj_covar = ( + self.in_proj2.weight.reshape(num_heads, head_dim, embed_dim) ** 2 + ).mean(dim=(1, 2)) + out_proj_covar = ( + self.out_proj2.weight.reshape(embed_dim, num_heads, head_dim) ** 2 + ).mean(dim=(0, 2)) + logging.info( + f"attn_weights_entropy = {attn_weights_entropy}, covar={attn_covar}, in_proj_covar={in_proj_covar}, out_proj_covar={out_proj_covar}" + ) + + +class PoolingModule(nn.Module): + """ + Averages the input over the time dimension and project with a square matrix. + """ + + def __init__(self, d_model: int): + super().__init__() + self.proj = ScaledLinear(d_model, d_model, initial_scale=0.1, bias=False) + + def forward( + self, + x: Tensor, + src_key_padding_mask: Optional[Tensor] = None, + ) -> Tensor: + """ + Args: + x: a Tensor of shape (T, N, C) + src_key_padding_mask: a Tensor of bool, of shape (N, T), with True in masked + positions. + + Returns: + - output, a Tensor of shape (T, N, C). + """ + if src_key_padding_mask is not None: + # False in padding positions + padding_mask = src_key_padding_mask.logical_not().to(x.dtype) # (N, T) + # Cumulated numbers of frames from start + cum_mask = padding_mask.cumsum(dim=1) # (N, T) + x = x.cumsum(dim=0) # (T, N, C) + pooling_mask = padding_mask / cum_mask + pooling_mask = pooling_mask.transpose(0, 1).contiguous().unsqueeze(-1) + # now pooling_mask: (T, N, 1) + x = x * pooling_mask # (T, N, C) + else: + num_frames = x.shape[0] + cum_mask = torch.arange(1, num_frames + 1).unsqueeze(1) # (T, 1) + x = x.cumsum(dim=0) # (T, N, C) + pooling_mask = (1.0 / cum_mask).unsqueeze(2) + # now pooling_mask: (T, N, 1) + x = x * pooling_mask + + x = self.proj(x) + return x + + def streaming_forward( + self, + x: Tensor, + cached_len: Tensor, + cached_avg: Tensor, + ) -> Tuple[Tensor, Tensor, Tensor]: + """ + Args: + x: a Tensor of shape (T, N, C) + cached_len: a Tensor of int, of shape (N,), containing the number of + past frames in batch. + cached_avg: a Tensor of shape (N, C), the average over all past frames + in batch. + + Returns: + A tuple of 2 tensors: + - output, a Tensor of shape (T, N, C). + - updated cached_avg, a Tensor of shape (N, C). + """ + x = x.cumsum(dim=0) # (T, N, C) + x = x + (cached_avg * cached_len.unsqueeze(1)).unsqueeze(0) + # Cumulated numbers of frames from start + cum_mask = torch.arange(1, x.size(0) + 1, device=x.device) + cum_mask = cum_mask.unsqueeze(1) + cached_len.unsqueeze(0) # (T, N) + pooling_mask = (1.0 / cum_mask).unsqueeze(2) + # now pooling_mask: (T, N, 1) + x = x * pooling_mask # (T, N, C) + + cached_len = cached_len + x.size(0) + cached_avg = x[-1] + + x = self.proj(x) + return x, cached_len, cached_avg + + +class FeedforwardModule(nn.Module): + """Feedforward module in Zipformer model.""" + + def __init__(self, d_model: int, feedforward_dim: int, dropout: float): + super(FeedforwardModule, self).__init__() + self.in_proj = nn.Linear(d_model, feedforward_dim) + self.balancer = ActivationBalancer( + feedforward_dim, channel_dim=-1, max_abs=10.0, min_prob=0.25 + ) + self.activation = DoubleSwish() + self.dropout = nn.Dropout(dropout) + self.out_proj = ScaledLinear(feedforward_dim, d_model, initial_scale=0.01) + + def forward(self, x: Tensor): + x = self.in_proj(x) + x = self.balancer(x) + x = self.activation(x) + x = self.dropout(x) + x = self.out_proj(x) + return x + + +class ConvolutionModule(nn.Module): + """ConvolutionModule in Zipformer 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, bias: bool = True) -> None: + """Construct an ConvolutionModule object.""" + super(ConvolutionModule, self).__init__() + # kernerl_size should be a odd number for 'SAME' padding + assert (kernel_size - 1) % 2 == 0, kernel_size + + self.pointwise_conv1 = nn.Conv1d( + channels, + 2 * channels, + kernel_size=1, + stride=1, + padding=0, + bias=bias, + ) + + # after pointwise_conv1 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.deriv_balancer1 = ActivationBalancer( + 2 * channels, + channel_dim=1, + max_abs=10.0, + min_positive=0.05, + max_positive=1.0, + ) + + # Will pad cached left context + self.lorder = kernel_size - 1 + self.depthwise_conv = nn.Conv1d( + channels, + channels, + kernel_size, + stride=1, + padding=0, + groups=channels, + bias=bias, + ) + + self.deriv_balancer2 = ActivationBalancer( + channels, + channel_dim=1, + min_positive=0.05, + max_positive=1.0, + max_abs=20.0, + ) + + self.activation = DoubleSwish() + + self.pointwise_conv2 = ScaledConv1d( + channels, + channels, + kernel_size=1, + stride=1, + padding=0, + bias=bias, + initial_scale=0.05, + ) + + def forward( + self, + x: Tensor, + src_key_padding_mask: Optional[Tensor] = None, + ) -> 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 bool in masked positions. + + Returns: + - Output tensor (#time, batch, channels). + """ + # exchange the temporal dimension and the feature dimension + x = x.permute(1, 2, 0) # (#batch, channels, time). + + # GLU mechanism + x = self.pointwise_conv1(x) # (batch, 2*channels, time) + + x = self.deriv_balancer1(x) + x = nn.functional.glu(x, dim=1) # (batch, channels, time) + + if src_key_padding_mask is not None: + x.masked_fill_(src_key_padding_mask.unsqueeze(1).expand_as(x), 0.0) + + # 1D Depthwise Conv + # Make depthwise_conv causal by + # manualy padding self.lorder zeros to the left + x = nn.functional.pad(x, (self.lorder, 0), "constant", 0.0) + x = self.depthwise_conv(x) + + x = self.deriv_balancer2(x) + x = self.activation(x) + + x = self.pointwise_conv2(x) # (batch, channel, time) + + return x.permute(2, 0, 1) + + def streaming_forward( + self, + x: Tensor, + cache: Tensor, + ) -> Tuple[Tensor, Tensor]: + """Compute convolution module. + + Args: + x: Input tensor (#time, batch, channels). + src_key_padding_mask: the mask for the src keys per batch: + (batch, #time), contains bool in masked positions. + cache: Cached left context for depthwise_conv, with shape of + (batch, channels, #kernel_size-1). Only used in real streaming decoding. + + Returns: + A tuple of 2 tensors: + - Output tensor (#time, batch, channels). + - New cached left context, with shape of (batch, channels, #kernel_size-1). + """ + # exchange the temporal dimension and the feature dimension + x = x.permute(1, 2, 0) # (#batch, channels, time). + + # GLU mechanism + x = self.pointwise_conv1(x) # (batch, 2*channels, time) + + x = self.deriv_balancer1(x) + x = nn.functional.glu(x, dim=1) # (batch, channels, time) + + # 1D Depthwise Conv + assert cache.shape == (x.size(0), x.size(1), self.lorder), ( + cache.shape, + (x.size(0), x.size(1), self.lorder), + ) + x = torch.cat([cache, x], dim=2) + # Update cache + cache = x[:, :, -self.lorder :] + x = self.depthwise_conv(x) + + x = self.deriv_balancer2(x) + x = self.activation(x) + + x = self.pointwise_conv2(x) # (batch, channel, time) + + return x.permute(2, 0, 1), cache + + +class Conv2dSubsampling(nn.Module): + """Convolutional 2D subsampling (to 1/4 length). + + Convert an input of shape (N, T, idim) to an output + with shape (N, T', odim), where + T' = (T-3)//2 - 2 == (T-7)//2 + + It is based on + https://github.com/espnet/espnet/blob/master/espnet/nets/pytorch_backend/transformer/subsampling.py # noqa + """ + + def __init__( + self, + in_channels: int, + out_channels: int, + layer1_channels: int = 8, + layer2_channels: int = 32, + layer3_channels: int = 128, + dropout: float = 0.1, + ) -> None: + """ + Args: + in_channels: + Number of channels in. The input shape is (N, T, in_channels). + Caution: It requires: T >=7, in_channels >=7 + out_channels + Output dim. The output shape is (N, (T-7)//2, out_channels) + layer1_channels: + Number of channels in layer1 + layer2_channels: + Number of channels in layer2 + layer3_channels: + Number of channels in layer3 + """ + assert in_channels >= 7, in_channels + super().__init__() + + self.conv = nn.Sequential( + nn.Conv2d( + in_channels=1, + out_channels=layer1_channels, + kernel_size=3, + padding=(0, 1), # (time, freq) + ), + ActivationBalancer(layer1_channels, channel_dim=1), + DoubleSwish(), + nn.Conv2d( + in_channels=layer1_channels, + out_channels=layer2_channels, + kernel_size=3, + stride=2, + padding=0, + ), + ActivationBalancer(layer2_channels, channel_dim=1), + DoubleSwish(), + nn.Conv2d( + in_channels=layer2_channels, + out_channels=layer3_channels, + kernel_size=3, + stride=(1, 2), # (time, freq) + ), + ActivationBalancer(layer3_channels, channel_dim=1), + DoubleSwish(), + ) + out_height = (((in_channels - 1) // 2) - 1) // 2 + self.out = ScaledLinear(out_height * layer3_channels, out_channels) + self.dropout = nn.Dropout(dropout) + + def forward(self, x: torch.Tensor) -> torch.Tensor: + """Subsample x. + + Args: + x: + Its shape is (N, T, idim). + + Returns: + Return a tensor of shape (N, (T-7)//2, odim) + """ + # On entry, x is (N, T, idim) + x = x.unsqueeze(1) # (N, T, idim) -> (N, 1, T, idim) i.e., (N, C, H, W) + x = self.conv(x) + # Now x is of shape (N, odim, (T-7)//2, ((idim-1)//2 - 1)//2) + b, c, t, f = x.size() + x = self.out(x.transpose(1, 2).reshape(b, t, c * f)) + # Now x is of shape (N, (T-7)//2, odim) + x = self.dropout(x) + return x + + +def _test_zipformer_main(): + feature_dim = 50 + batch_size = 5 + seq_len = 47 + feature_dim = 50 + # Just make sure the forward pass runs. + + c = Zipformer( + num_features=feature_dim, + encoder_dims=(64, 96), + encoder_unmasked_dims=(48, 64), + nhead=(4, 4), + decode_chunk_size=4, + ) + # Just make sure the forward pass runs. + f = c( + torch.randn(batch_size, seq_len, feature_dim), + torch.full((batch_size,), seq_len, dtype=torch.int64), + ) + assert ((seq_len - 7) // 2 + 1) // 2 == f[0].shape[1], (seq_len, f.shape[1]) + f[0].sum().backward() + c.eval() + f = c( + torch.randn(batch_size, seq_len, feature_dim), + torch.full((batch_size,), seq_len, dtype=torch.int64), + ) + f # to remove flake8 warnings + + +def _test_conv2d_subsampling(): + num_features = 80 + encoder_dims = 384 + dropout = 0.1 + encoder_embed = Conv2dSubsampling(num_features, encoder_dims, dropout=dropout) + for i in range(20, 40): + x = torch.rand(2, i, num_features) + y = encoder_embed(x) + assert (x.shape[1] - 7) // 2 == y.shape[1], (x.shape[1], y.shape[1]) + + +def _test_pooling_module(): + N, S, C = 2, 12, 32 + chunk_len = 4 + m = PoolingModule(d_model=C) + + # test chunk-wise forward with padding_mask + x = torch.randn(S, N, C) + y = m(x) + cached_len = torch.zeros(N, dtype=torch.int32) + cached_avg = torch.zeros(N, C) + for i in range(S // chunk_len): + start = i * chunk_len + end = start + chunk_len + x_chunk = x[start:end] + y_chunk, cached_len, cached_avg = m.streaming_forward( + x_chunk, + cached_len=cached_len, + cached_avg=cached_avg, + ) + assert torch.allclose(y_chunk, y[start:end]), (y_chunk, y[start:end]) + + +def _test_state_stack_unstack(): + m = Zipformer( + num_features=80, + encoder_dims=(64, 96), + encoder_unmasked_dims=(48, 64), + nhead=(4, 4), + zipformer_downsampling_factors=(4, 8), + num_left_chunks=2, + decode_chunk_size=8, + ) + s1 = m.get_init_state() + s2 = m.get_init_state() + states = stack_states([s1, s2]) + new_s1, new_s2 = unstack_states(states) + for i in range(m.num_encoders * 7): + for x, y in zip(s1[i], new_s1[i]): + assert torch.equal(x, y) + for x, y in zip(s2[i], new_s2[i]): + assert torch.equal(x, y) + + +if __name__ == "__main__": + logging.getLogger().setLevel(logging.INFO) + torch.set_num_threads(1) + torch.set_num_interop_threads(1) + _test_zipformer_main() + _test_conv2d_subsampling() + _test_pooling_module() + _test_state_stack_unstack()