diff --git a/egs/libriheavy/ASR/RESULTS.md b/egs/libriheavy/ASR/RESULTS.md new file mode 100644 index 000000000..4fbedad98 --- /dev/null +++ b/egs/libriheavy/ASR/RESULTS.md @@ -0,0 +1,205 @@ +## Results + +### Zipformer PromptASR (zipformer + PromptASR + BERT text encoder) + +#### [zipformer_prompt_asr](./zipformer_prompt_asr) + +See for commit history and +our paper for more details. + + + +##### Training on the medium subset, with content & style prompt, **no** context list + +You can find a pre-trained model, training logs, decoding logs, and decoding results at: + +The training command is: + +```bash +causal=0 +subset=medium +memory_dropout_rate=0.05 +text_encoder_type=BERT + +python ./zipformer_prompt_asr/train_bert_encoder.py \ + --world-size 4 \ + --start-epoch 1 \ + --num-epochs 60 \ + --exp-dir ./zipformer_prompt_asr/exp \ + --use-fp16 True \ + --memory-dropout-rate $memory_dropout_rate \ + --causal $causal \ + --subset $subset \ + --manifest-dir data/fbank \ + --bpe-model data/lang_bpe_500_fallback_coverage_0.99/bpe.model \ + --max-duration 1000 \ + --text-encoder-type $text_encoder_type \ + --text-encoder-dim 768 \ + --use-context-list 0 \ + --top-k $top_k \ + --use-style-prompt 1 +``` + +The decoding results using utterance-level context (epoch-60-avg-10): + +| decoding method | lh-test-clean | lh-test-other | comment | +|----------------------|---------------|---------------|---------------------| +| modified_beam_search | 3.13 | 6.78 | --use-pre-text False --use-style-prompt False | +| modified_beam_search | 2.86 | 5.93 | --pre-text-transform upper-no-punc --style-text-transform upper-no-punc | +| modified_beam_search | 2.6 | 5.5 | --pre-text-transform mixed-punc --style-text-transform mixed-punc | + + +The decoding command is: + +```bash +for style in mixed-punc upper-no-punc; do + python ./zipformer_prompt_asr/decode_bert.py \ + --epoch 60 \ + --avg 10 \ + --use-averaged-model True \ + --post-normalization True \ + --causal False \ + --exp-dir ./zipformer_prompt_asr/exp \ + --manifest-dir data/fbank \ + --bpe-model data/lang_bpe_500_fallback_coverage_0.99/bpe.model \ + --max-duration 1000 \ + --decoding-method modified_beam_search \ + --beam-size 4 \ + --text-encoder-type BERT \ + --text-encoder-dim 768 \ + --memory-layer 0 \ + --use-ls-test-set False \ + --use-ls-context-list False \ + --max-prompt-lens 1000 \ + --use-pre-text True \ + --use-style-prompt True \ + --style-text-transform $style \ + --pre-text-transform $style \ + --compute-CER 0 +done +``` + +##### Training on the medium subset, with content & style prompt, **with** context list + +You can find a pre-trained model, training logs, decoding logs, and decoding results at: + +This model is trained with an extra type of content prompt (context words), thus it does better +on **word-level** context biasing. Note that to train this model, please first run `prepare_prompt_asr.sh` +to prepare a manifest containing context words. + +The training command is: + +```bash + +causal=0 +subset=medium +memory_dropout_rate=0.05 +text_encoder_type=BERT +use_context_list=True + +# prepare the required data for context biasing +./prepare_prompt_asr.sh --stage 0 --stop_stage 1 + +python ./zipformer_prompt_asr/train_bert_encoder.py \ + --world-size 4 \ + --start-epoch 1 \ + --num-epochs 50 \ + --exp-dir ./zipformer_prompt_asr/exp \ + --use-fp16 True \ + --memory-dropout-rate $memory_dropout_rate \ + --causal $causal \ + --subset $subset \ + --manifest-dir data/fbank \ + --bpe-model data/lang_bpe_500_fallback_coverage_0.99/bpe.model \ + --max-duration 1000 \ + --text-encoder-type $text_encoder_type \ + --text-encoder-dim 768 \ + --use-context-list $use_context_list \ + --top-k 10000 \ + --use-style-prompt 1 +``` + +*Utterance-level biasing:* + +| decoding method | lh-test-clean | lh-test-other | comment | +|----------------------|---------------|---------------|---------------------| +| modified_beam_search | 3.17 | 6.72 | --use-pre-text 0 --use-style-prompt 0 | +| modified_beam_search | 2.91 | 6.24 | --pre-text-transform upper-no-punc --style-text-transform upper-no-punc | +| modified_beam_search | 2.72 | 5.72 | --pre-text-transform mixed-punc --style-text-transform mixed-punc | + + +The decoding command for the table above is: + +```bash +for style in mixed-punc upper-no-punc; do + python ./zipformer_prompt_asr/decode_bert.py \ + --epoch 50 \ + --avg 10 \ + --use-averaged-model True \ + --post-normalization True \ + --causal False \ + --exp-dir ./zipformer_prompt_asr/exp \ + --manifest-dir data/fbank \ + --bpe-model data/lang_bpe_500_fallback_coverage_0.99/bpe.model \ + --max-duration 1000 \ + --decoding-method modified_beam_search \ + --beam-size 4 \ + --text-encoder-type BERT \ + --text-encoder-dim 768 \ + --memory-layer 0 \ + --use-ls-test-set False \ + --use-ls-context-list False \ + --max-prompt-lens 1000 \ + --use-pre-text True \ + --use-style-prompt True \ + --style-text-transform $style \ + --pre-text-transform $style \ + --compute-CER 0 +done +``` + +*Word-level biasing:* + +The results are reported on LibriSpeech test-sets using the biasing list provided from . +You need to set `--use-ls-test-set True` so that the LibriSpeech test sets are used. + +| decoding method | ls-test-clean | ls-test-other | comment | +|----------------------|---------------|---------------|---------------------| +| modified_beam_search | 2.4 | 5.08 | --use-pre-text 0 --use-style-prompt 0 | +| modified_beam_search | 2.14 | 4.62 | --use-ls-context-list 1 --pre-text-transform mixed-punc --style-text-transform mixed-punc --ls-distractors 0 | +| modified_beam_search | 2.14 | 4.64 | --use-ls-context-list 1 --pre-text-transform mixed-punc --style-text-transform mixed-punc --ls-distractors 100 | + +The decoding command is for the table above is: + +```bash +use_ls_test_set=1 +use_ls_context_list=1 + +for ls_distractors in 0 100; do + python ./zipformer_prompt_asr/decode_bert.py \ + --epoch 50 \ + --avg 10 \ + --use-averaged-model True \ + --post-normalization True \ + --causal False \ + --exp-dir ./zipformer_prompt_asr/exp \ + --manifest-dir data/fbank \ + --bpe-model data/lang_bpe_500_fallback_coverage_0.99/bpe.model \ + --max-duration 1000 \ + --decoding-method modified_beam_search \ + --beam-size 4 \ + --text-encoder-type BERT \ + --text-encoder-dim 768 \ + --memory-layer 0 \ + --use-ls-test-set $use_ls_test_setse \ + --use-ls-context-list $use_ls_context_list \ + --ls-distractors $ls_distractors \ + --max-prompt-lens 1000 \ + --use-pre-text True \ + --use-style-prompt True \ + --style-text-transform mixed-punc \ + --pre-text-transform mixed-punc \ + --compute-CER 0 +done + +``` diff --git a/egs/libriheavy/ASR/prepare_prompt_asr.sh b/egs/libriheavy/ASR/prepare_prompt_asr.sh new file mode 100755 index 000000000..b931cea26 --- /dev/null +++ b/egs/libriheavy/ASR/prepare_prompt_asr.sh @@ -0,0 +1,36 @@ +#!/usr/bin/env bash + +set -eou pipefail + +# This is the preparation recipe for PromptASR: https://arxiv.org/pdf/2309.07414 + +log() { + # This function is from espnet + local fname=${BASH_SOURCE[1]##*/} + echo -e "$(date '+%Y-%m-%d %H:%M:%S') (${fname}:${BASH_LINENO[0]}:${FUNCNAME[1]}) $*" +} + +stage=-1 +stop_stage=100 +manifest_dir=data/fbank +subset=medium +topk=10000 + +. shared/parse_options.sh || exit 1 + +if [ $stage -le 0 ] && [ $stop_stage -ge 0 ]; then + log "Stage 0: Download the meta biasing list for LibriSpeech" + mkdir -p data/context_biasing + cd data/context_biasing + git clone https://github.com/facebookresearch/fbai-speech.git + cd ../.. +fi + +if [ $stage -le 1 ] && [ $stop_stage -ge 1 ]; then + log "Stage 1: Add rare-words for context biasing to the manifest" + python zipformer_prompt_asr/utils.py \ + --manifest-dir $manifest_dir \ + --subset $subset \ + --top-k $topk + +fi diff --git a/egs/libriheavy/ASR/shared b/egs/libriheavy/ASR/shared new file mode 120000 index 000000000..4cbd91a7e --- /dev/null +++ b/egs/libriheavy/ASR/shared @@ -0,0 +1 @@ +../../../icefall/shared \ No newline at end of file diff --git a/egs/libriheavy/ASR/zipformer_prompt_asr/__init__.py b/egs/libriheavy/ASR/zipformer_prompt_asr/__init__.py new file mode 100644 index 000000000..e69de29bb diff --git a/egs/libriheavy/ASR/zipformer_prompt_asr/asr_datamodule.py b/egs/libriheavy/ASR/zipformer_prompt_asr/asr_datamodule.py new file mode 100644 index 000000000..690003377 --- /dev/null +++ b/egs/libriheavy/ASR/zipformer_prompt_asr/asr_datamodule.py @@ -0,0 +1,520 @@ +# Copyright 2021 Piotr Żelasko +# +# 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, Callable, Dict, List, Optional + +import torch +from dataset import PromptASRDataset +from lhotse import CutSet, Fbank, FbankConfig, load_manifest, load_manifest_lazy +from lhotse.dataset import ( # SingleCutSampler, + CutConcatenate, + CutMix, + DynamicBucketingSampler, + ExtraPadding, + K2SpeechRecognitionDataset, + PrecomputedFeatures, + SpecAugment, +) +from lhotse.dataset.input_strategies import 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 LibriHeavyAsrDataModule: + """ + 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 + + if args.use_context_list: + assert args.rare_word_file is not None + with open(args.rare_word_file, "r") as f: + self.rare_word_list = ( + f.read().lower().split() + ) # Use lower-cased for easier style transform + else: + self.rare_word_list = None + + @classmethod + def add_arguments(cls, parser: argparse.ArgumentParser): + group = parser.add_argument_group( + title="ASR data related options", + description="These options are used for the preparation of " + "PyTorch DataLoaders from Lhotse CutSet's -- they control the " + "effective batch sizes, sampling strategies, applied data " + "augmentations, etc.", + ) + group.add_argument( + "--manifest-dir", + type=Path, + default=Path("data/fbank"), + help="Path to directory with train/valid/test cuts.", + ) + group.add_argument( + "--max-duration", + type=int, + default=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( + "--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. ", + ) + + # Libriheavy specific arguments + group.add_argument( + "--subset", + type=str, + default="small", + help="Select the Libriheavy subset (small|medium|large)", + ) + + group.add_argument( + "--use-context-list", + type=str2bool, + default=False, + help="Use the context list of libri heavy", + ) + + group.add_argument( + "--top-k", + type=int, + default=10000, + help="""The top-k words are identified as common words, + the rest as rare words""", + ) + + group.add_argument( + "--with-decoding", + type=str2bool, + default=False, + help="If the texts field contain decoding", + ) + + group.add_argument( + "--random-left-padding", + type=str2bool, + ) + + group.add_argument( + "--rare-word-file", + type=str, + ) + + group.add_argument( + "--long-audio-cuts", + type=str, + default="data/manifest_npr/npr1_cuts_all_guids_0.jsonl.gz", + ) + + def train_dataloaders( + self, + cuts_train: CutSet, + sampler_state_dict: Optional[Dict[str, Any]] = None, + text_sampling_func: Callable[[List[str]], str] = 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, p=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 = PromptASRDataset( + cut_transforms=transforms, + input_transforms=input_transforms, + return_cuts=self.args.return_cuts, + text_sampling_func=text_sampling_func, + rare_word_list=self.rare_word_list, + ) + + 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 = PromptASRDataset( + cut_transforms=transforms, + input_strategy=OnTheFlyFeatures(Fbank(FbankConfig(num_mel_bins=80))), + input_transforms=input_transforms, + return_cuts=self.args.return_cuts, + text_sampling_func=text_sampling_func, + rare_word_list=self.rare_word_list, + ) + + 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=True, + ) + else: + raise NotImplementedError( + "SingleCutSampler is no longer supported by lhotse" + ) + 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, + text_sampling_func: Callable[[List[str]], str] = None, + ) -> DataLoader: + transforms = [] + if self.args.random_left_padding: + logging.info("Enable random left padding") + transforms.append( + ExtraPadding(extra_frames=16, randomized=True, direction="left") + ) + + 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 = PromptASRDataset( + cut_transforms=transforms, + input_strategy=OnTheFlyFeatures(Fbank(FbankConfig(num_mel_bins=80))), + return_cuts=self.args.return_cuts, + text_sampling_func=text_sampling_func, + rare_word_list=self.rare_word_list, + ) + else: + validate = PromptASRDataset( + cut_transforms=transforms, + return_cuts=self.args.return_cuts, + text_sampling_func=text_sampling_func, + rare_word_list=self.rare_word_list, + ) + 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 PrecomputedFeatures(), + 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_cuts(self) -> CutSet: + logging.info(f"About to get {self.args.subset} cuts") + + if self.args.use_context_list: + path = ( + self.args.manifest_dir + / f"libriheavy_cuts_{self.args.subset}_with_context_list_topk_{self.args.top_k}.jsonl.gz" + ) + elif self.args.with_decoding: + path = ( + self.args.manifest_dir + / f"libriheavy_cuts_{self.args.subset}_with_decoding.jsonl.gz" + ) + else: + path = ( + self.args.manifest_dir / f"libriheavy_cuts_{self.args.subset}.jsonl.gz" + ) + + logging.info(f"Loading manifest from {path}.") + cuts_train = CutSet.from_jsonl_lazy(path) + return cuts_train + + @lru_cache() + def dev_cuts(self) -> CutSet: + logging.info("About to get dev cuts") + cuts_valid = load_manifest_lazy( + self.args.manifest_dir / "libriheavy_cuts_dev.jsonl.gz" + ) + return cuts_valid + + @lru_cache() + def test_clean_cuts(self) -> CutSet: + logging.info("About to get test-clean cuts") + cuts_valid = load_manifest_lazy( + self.args.manifest_dir / "libriheavy_cuts_test-clean_official.jsonl.gz" + ) + return cuts_valid + + @lru_cache() + def test_other_cuts(self) -> CutSet: + logging.info("About to get test-other cuts") + cuts_valid = load_manifest_lazy( + self.args.manifest_dir / "libriheavy_cuts_test-other_official.jsonl.gz" + ) + return cuts_valid + + @lru_cache() + def librispeech_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 librispeech_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" + ) + + @lru_cache() + def long_audio_cuts(self) -> CutSet: + logging.info("About to get long audio cuts") + cuts = load_manifest_lazy( + self.args.long_audio_cuts, + ) + return cuts + + @lru_cache() + def test_dev_cuts(self) -> CutSet: + logging.info("About to get test dev cuts") + cuts = load_manifest_lazy( + self.args.manifest_dir / "libriheavy_cuts_test_dev.jsonl.gz" + ) + return cuts diff --git a/egs/libriheavy/ASR/zipformer_prompt_asr/beam_search.py b/egs/libriheavy/ASR/zipformer_prompt_asr/beam_search.py new file mode 120000 index 000000000..e24eca39f --- /dev/null +++ b/egs/libriheavy/ASR/zipformer_prompt_asr/beam_search.py @@ -0,0 +1 @@ +../../../librispeech/ASR/pruned_transducer_stateless2/beam_search.py \ No newline at end of file diff --git a/egs/libriheavy/ASR/zipformer_prompt_asr/dataset.py b/egs/libriheavy/ASR/zipformer_prompt_asr/dataset.py new file mode 100644 index 000000000..e0bf8f73d --- /dev/null +++ b/egs/libriheavy/ASR/zipformer_prompt_asr/dataset.py @@ -0,0 +1,586 @@ +# Copyright 2023 Xiaomi Corp. (authors: Xiaoyu Yang) +# +# See ../LICENSE for clarification regarding multiple authors +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import random +from typing import Callable, Dict, List, Optional, Union + +import numpy as np +import torch +from lhotse import validate +from lhotse.cut import CutSet +from lhotse.dataset import K2SpeechRecognitionDataset +from lhotse.dataset.input_strategies import BatchIO, PrecomputedFeatures +from lhotse.utils import compute_num_frames, ifnone +from text_normalization import ( + lower_all_char, + lower_only_alpha, + remove_non_alphabetic, + train_text_normalization, + upper_all_char, + upper_only_alpha, +) +from torch.utils.data.dataloader import DataLoader, default_collate + + +class PromptASRDataset(torch.utils.data.Dataset): + """This is a dataset for Prompt ASR. It supports the following features: + 1. Select a tuple of (text, pre_text, style_text) randomly from a + list of texts as supervisions. + + """ + + def __init__( + self, + return_cuts: bool = False, + cut_transforms: List[Callable[[CutSet], CutSet]] = None, + input_transforms: List[Callable[[torch.Tensor], torch.Tensor]] = None, + input_strategy: BatchIO = PrecomputedFeatures(), + text_sampling_func: Optional[Callable[[List[str]], str]] = None, + rare_word_list: Optional[List[str]] = None, + ): + """ + Icefall ASR IterableDataset constructor. See https://github.com/lhotse-speech/lhotse/blob/master/lhotse/dataset/speech_recognition.py + for more details. + + :param return_cuts: When ``True``, will additionally return a "cut" field in each batch with the Cut + objects used to create that batch. + :param cut_transforms: A list of transforms to be applied on each sampled batch, + before converting cuts to an input representation (audio/features). + Examples: cut concatenation, noise cuts mixing, etc. + :param input_transforms: A list of transforms to be applied on each sampled batch, + after the cuts are converted to audio/features. + Examples: normalization, SpecAugment, etc. + :param input_strategy: Converts cuts into a collated batch of audio/features. + By default, reads pre-computed features from disk. + :param text_sampling_func: Sampling a text as transcription from a list of texts. + """ + super().__init__() + # Initialize the fields + self.return_cuts = return_cuts + self.cut_transforms = ifnone(cut_transforms, []) + self.input_transforms = ifnone(input_transforms, []) + self.input_strategy = input_strategy + + # a text sampling function + self.text_sampling_func = text_sampling_func + self.rare_word_list = rare_word_list + + def __getitem__(self, cuts: CutSet) -> Dict[str, Union[torch.Tensor, List[str]]]: + """ + Return a new batch, with the batch size automatically determined using the constraints + of max_frames and max_cuts. + """ + validate_for_asr(cuts) + + # Sort the cuts by duration so that the first one determines the batch time dimensions. + cuts = cuts.sort_by_duration(ascending=False) + + # Optional CutSet transforms - e.g. padding, or speed perturbation that adjusts + # the supervision boundaries. + for tnfm in self.cut_transforms: + cuts = tnfm(cuts) + + # Sort the cuts again after transforms + cuts = cuts.sort_by_duration(ascending=False) + + # Get a tensor with batched feature matrices, shape (B, T, F) + # Collation performs auto-padding, if necessary. + input_tpl = self.input_strategy(cuts) + if len(input_tpl) == 3: + # An input strategy with fault tolerant audio reading mode. + # "cuts" may be a subset of the original "cuts" variable, + # that only has cuts for which we succesfully read the audio. + inputs, _, cuts = input_tpl + else: + inputs, _ = input_tpl + + # Get a dict of tensors that encode the positional information about supervisions + # in the batch of feature matrices. The tensors are named "sequence_idx", + # "start_frame/sample" and "num_frames/samples". + supervision_intervals = self.input_strategy.supervision_intervals(cuts) + + # Apply all available transforms on the inputs, i.e. either audio or features. + # This could be feature extraction, global MVN, SpecAugment, etc. + segments = torch.stack(list(supervision_intervals.values()), dim=1) + for tnfm in self.input_transforms: + inputs = tnfm(inputs, supervision_segments=segments) + + batch = { + "inputs": inputs, + "supervisions": default_collate( + [ + self.text_sampling_func( + texts=supervision.texts, + pre_texts=supervision.pre_texts, + context_list=supervision.context_list + if "context_list" in supervision.custom + else None, + rare_word_list=self.rare_word_list, + ) + if self.text_sampling_func is not None + else { + "text": train_text_normalization(supervision.texts[0]), + "pre_text": train_text_normalization(supervision.pre_texts[0]), + "style_text": train_text_normalization( + supervision.pre_texts[0] + ), + "transform_ids": 0, + } + for sequence_idx, cut in enumerate(cuts) + for supervision in cut.supervisions + ] + ), + } + # Update the 'supervisions' field with sequence_idx and start/num frames/samples + batch["supervisions"].update(supervision_intervals) + if self.return_cuts: + batch["supervisions"]["cut"] = [ + cut for cut in cuts for sup in cut.supervisions + ] + + has_word_alignments = all( + s.alignment is not None and "word" in s.alignment + for c in cuts + for s in c.supervisions + ) + + return batch + + +def validate_for_asr(cuts: CutSet) -> None: + validate(cuts) + tol = 2e-3 # 1ms + for cut in cuts: + for supervision in cut.supervisions: + assert supervision.start >= -tol, ( + f"Supervisions starting before the cut are not supported for ASR" + f" (sup id: {supervision.id}, cut id: {cut.id})" + ) + + # Supervision start time is relative to Cut ... + # https://lhotse.readthedocs.io/en/v0.10_e/cuts.html + # + # 'supervision.end' is end of supervision inside the Cut + assert supervision.end <= cut.duration + tol, ( + f"Supervisions ending after the cut " + f"are not supported for ASR" + f" (sup id: {supervision.id}, cut id: {cut.id})" + ) + + +def get_substring(s: str, min_len: int = 40, max_len: int = 250) -> str: + """A helper function that generates a random substring from a given string + + Args: + s (str): Input string + + Returns: + str: Returned substring + """ + min_len = min(len(s), min_len) + + start = random.randint(0, len(s) - min_len) + end = min(start + max_len, random.randint(start + min_len, len(s))) + + return s[start:end] + + +def triplet_text_sampling( + texts: List[str], + pre_texts: List[str], + context_list: Optional[str] = None, + rare_word_list: Optional[List[str]] = None, + transforms: Optional[List[Callable[[str], str]]] = None, + min_len_style: Optional[int] = 80, +) -> Dict[str, str]: + """This function generates a triplet of + (pre_text, style_text, ref_text). The style of style_text and ref_text + should **always** match, whereas the style of pre_text is arbitrary. + Suppose we have 2 different transforms A,B, and the preceding text is + referred to as pre_text. The following three tuples are all valid: + + (A(pre_text), A(style_text), A(ref_text)) + (A(pre_text), B(style_text), B(ref_text)) + (A(pre_text), A(style_text), A(ref_text)) + (B(pre_text), B(style_text), B(ref_text)) + + If transforms is not given, the following pre-defined transforms + are available: + 0: original (mixed-cased, with punc) + 1: upper_only_alpha (upper-cased, no punc) + + When the transform of text and pre_text match, we can use the whole + pre_text as the prompt text. + + Args: + texts (List[str]): + A list of ref_texts whose first item is the ground truth + text from books. + pre_texts (List[str]): + A list of pre_texts, whose first item is the groundtruth + pre_text from books. + context_list: Optional[str] = None, + A list of biasing words separated by space + rare_word_list: Optional[str] = None, + A list of rare-words separated by space (used as distractors) + transforms (List[Callable[[str], str]]): A list of possible transforms to be applied + + Returns: + A dictionary of ref_text, pre_text, style_text + """ + assert len(texts) == len(pre_texts) + assert len(texts) == 2 + + # we assume the first item to be ground truth + gt_text = texts[0] + gt_pre_text = pre_texts[0] + + if transforms is None: + transforms = [ + lambda x: x, # return it self + upper_only_alpha, + lower_only_alpha, + lower_all_char, + ] + + sampling_weight = [ + 0.7, + 0.3, + 0.0, + 0.0, + ] # Mixed-punc should have the largest sampling prob + + total_transforms = len(transforms) # do not use the recognized trans + + # Randomly sample transforms + i_text, i_pre_text = np.random.choice(total_transforms, 2, p=sampling_weight) + + # get the normalized text and pre_text + text = transforms[i_text](gt_text) + pre_text = transforms[i_pre_text](gt_pre_text) + + if i_text == i_pre_text: + style_text = get_substring(pre_text, min_len=min_len_style, max_len=150) + else: + # get the pre_text of same style as text + # For now, **don't** do transform to the style text, because we do it after the dataloader + style_text = gt_pre_text + # style_text = pre_texts[i_text] if i_text <= 1 else transforms[i_text-2](gt_pre_text) + style_text = get_substring(style_text, min_len=min_len_style, max_len=150) + + return { + "text": train_text_normalization(text), + "pre_text": train_text_normalization(pre_text), + "style_text": train_text_normalization(style_text), + "transform_ids": i_text, + } + + +def triplet_text_sampling_with_context_list( + texts: List[str], + pre_texts: List[str], + context_list: str, + rare_word_list: List[str], + transforms: Optional[List[Callable[[str], str]]] = None, + min_len_style: Optional[int] = 80, +) -> Dict[str, str]: + """This function generates a triplet of + (pre_text, style_text, ref_text). The pre_text is either the preceding text + or a list of words (context words + distractors). + The style of style_text and ref_text should **always** match, whereas + the style of pre_text is arbitrary. + Suppose we have 2 different transforms A,B, and the preceding text is + referred to as pre_text. The following three tuples are all valid: + + (A(pre_text), A(style_text), A(ref_text)) + (A(pre_text), B(style_text), B(ref_text)) + (A(pre_text), A(style_text), A(ref_text)) + (B(pre_text), B(style_text), B(ref_text)) + + If transforms is not given, the following pre-defined transforms + are available: + 0: original (mixed-cased, with punc) + 1: upper_only_alpha (upper-cased, no punc) + + When the transform of text and pre_text match, we can use the whole + pre_text as the prompt text. + + Args: + texts (List[str]): + A list of ref_texts whose first item is the ground truth + text from books. + pre_texts (List[str]): + A list of pre_texts, whose first item is the groundtruth + pre_text from books. + context_list: Optional[str] = None, + A list of biasing words separated by space + rare_word_list: Optional[str] = None, + A list of rare-words separated by space (used as distractors) + transforms (List[Callable[[str], str]]): A list of possible transforms to be applied + + Returns: + A dictionary of ref_text, pre_text, style_text + Returns: + str: A dictionary + """ + # import pdb; pdb.set_trace() + assert len(texts) == len(pre_texts) + assert len(texts) == 2 + + if context_list is not None: + context_list = context_list.lower() + + # we assume the first item to be ground truth + gt_text = texts[0] + gt_pre_text = pre_texts[0] + + if transforms is None: + transforms = [ + lambda x: x, # return it self + upper_only_alpha, + lower_only_alpha, + lower_all_char, + ] + + sampling_weight = [ + 0.7, + 0.3, + 0.0, + 0.0, + ] # Mixed-punc should have the largest sampling prob + + total_transforms = len(transforms) # do not use the recognized trans + + # Select a transformation randomly + i_text, i_pre_text = np.random.choice(total_transforms, 2, p=sampling_weight) + + # get the normalized text and pre_text + text = transforms[i_text](gt_text) + pre_text = get_pre_text_with_context_list2( + text=gt_text, + pre_text=gt_pre_text, + context_list=context_list, + rare_words_list=rare_word_list, + ) + pre_text = transforms[i_pre_text](pre_text) + + if i_text == i_pre_text: + style_text = get_substring(pre_text, min_len=min_len_style, max_len=150) + else: + # get the pre_text of same style as text + # For now, **don't** do transform to the style text + style_text = gt_pre_text + # style_text = pre_texts[i_text] if i_text <= 1 else transforms[i_text-2](gt_pre_text) + style_text = get_substring(style_text, min_len=min_len_style, max_len=150) + + return { + "text": train_text_normalization(text), + "pre_text": train_text_normalization(pre_text), + "style_text": train_text_normalization(style_text), + "transform_ids": i_text, + } + + +def get_pre_text_with_context_list( + text: str, + pre_text: str, + context_list: str, + rare_words_list: List[str] = None, +) -> str: + # Always get the first one, which is the gt (mixed-cased trans), but with upper_only_alpha + # By a small proportion of time, use the substring of ref_text as pre_text + + if context_list != "" and context_list is not None: + v = random.random() + if v < 0.5: + # correct + distractors + # sample distractors + num_distractors = random.randint(0, 50) + distractors = random.sample(rare_words_list, num_distractors) + # sample correct + correct = context_list.split() + i = random.randint(1, len(correct)) + correct = random.sample(correct, i) + # combine correct and distractors + pre_text = distractors + correct + random.shuffle(pre_text) + pre_text = " ".join(pre_text) + elif v < 0.7: + splitted = text.split() + sampling_weights = [len(w) ** 1.2 for w in splitted] + sampling_weights = [p / sum(sampling_weights) for p in sampling_weights] + i = random.randint(1, min(len(splitted), 20)) + splitted = list(np.random.choice(splitted, i, p=sampling_weights)) + num_distractors = random.randint(0, 70) + distractors = random.sample(rare_words_list, num_distractors) + splitted += distractors + random.shuffle(splitted) # shuffle the list + pre_text = " ".join(splitted) + else: + pre_text = pre_text + else: + v = random.random() + if v < 0.1: + splitted = text.split() + sampling_weights = [len(w) ** 1.2 for w in splitted] + sampling_weights = [p / sum(sampling_weights) for p in sampling_weights] + i = random.randint(1, min(len(splitted), 20)) + splitted = list(np.random.choice(splitted, i, p=sampling_weights)) + pre_text = " ".join(splitted) + num_distractors = random.randint(0, 70) + distractors = random.sample(rare_words_list, num_distractors) + splitted += distractors + random.shuffle(splitted) # shuffle the list + elif v < 0.2: + # full distractors + num_distractors = random.randint(5, 100) + distractors = random.sample(rare_words_list, num_distractors) + pre_text = " ".join(distractors) + + elif v < 0.3: + pre_text = get_substring(text, min_len=15, max_len=150) + else: + pre_text = pre_text + + return pre_text + + +def get_pre_text_with_context_list2( + text: str, + pre_text: str, + context_list: str, + rare_words_list: List[str] = None, +) -> str: + # Get the pre_text, either the ground truth preceding text or + # a list of words consisting of biasing words and distrators + # By a small proportion of time, use the substring of ref_text as pre_text + + if context_list != "" and context_list is not None: + v = random.random() + if v < 0.4: + # sample distractors + num_distractors = random.randint(50, 100) + distractors = random.sample(rare_words_list, num_distractors) + # sample correct + correct = context_list.split() + i = random.randint(1, len(correct)) + correct = random.sample(correct, i) + # combine correct and distractors + pre_text = distractors + correct + random.shuffle(pre_text) + pre_text = " ".join(pre_text) + elif v < 0.55: + splitted = text.split() + sampling_weights = [ + len(w) ** 1.2 for w in splitted + ] # longer words with higher weights + sampling_weights = [p / sum(sampling_weights) for p in sampling_weights] + i = random.randint(1, min(len(splitted), 20)) + splitted = list(np.random.choice(splitted, i, p=sampling_weights)) + num_distractors = random.randint(50, 100) + distractors = random.sample(rare_words_list, num_distractors) + splitted += distractors + random.shuffle(splitted) # shuffle the list + pre_text = " ".join(splitted) + else: + pre_text = pre_text + else: + v = random.random() + if v < 0.3: + splitted = text.split() + sampling_weights = [len(w) ** 1.2 for w in splitted] + sampling_weights = [p / sum(sampling_weights) for p in sampling_weights] + i = random.randint(1, min(len(splitted), 20)) + splitted = list(np.random.choice(splitted, i, p=sampling_weights)) + pre_text = " ".join(splitted) + num_distractors = random.randint(50, 100) + distractors = random.sample(rare_words_list, num_distractors) + splitted += distractors + random.shuffle(splitted) # shuffle the list + elif v < 0.4: + # full distractors + num_distractors = random.randint(5, 100) + distractors = random.sample(rare_words_list, num_distractors) + pre_text = " ".join(distractors) + elif v < 0.6: + pre_text = get_substring(text, min_len=15, max_len=150) + else: + pre_text = pre_text + + return pre_text + + +def naive_triplet_text_sampling( + texts: List[str], + pre_texts: List[str], + context_list: str = None, + rare_word_list: List[str] = None, + min_len_style: Optional[int] = 120, +): + # The most simplest text sampling function, used only for + # evaluation, use a fixed sentence as the style text + + return { + "text": train_text_normalization(texts[0]), + "pre_text": train_text_normalization(pre_texts[0]), + "style_text": "Mixed-case English transcription, with punctuation. Actually, it is fully not related. What do you think?", + "transform_ids": 0, + } + + +def random_shuffle_subset( + data: List[str], + p: float = 0.2, + p_mask: float = 0.05, +) -> List[str]: + """ + Randomly shuffle the subset by probability `p`, which means that p% of the samples + in the original batch are shuffled, the others are kept in the original order. + + With a probability of `p_mask`, replace the original string with an empty string. + + """ + + num_to_shuffle = int(len(data) * p) + id_to_shuffle = np.random.choice(len(data), num_to_shuffle, replace=False) + item_to_shuffle = [data[id] for id in id_to_shuffle] + random.shuffle(item_to_shuffle) + + for id, item in zip(id_to_shuffle, item_to_shuffle): + data[id] = item + + # Randomly mask a proportion of the data to empty string + if p_mask > 0: + for i in range(len(data)): + if random.random() < p_mask: + data[i] = "" + + return data + + +if __name__ == "__main__": + texts = [ + "AA, BB, cC, dD!", + "AA BB CC DD", + ] + + pre_texts = [ + "EE, Ff, Gg? EE, Ff, Gg? EE, Ff, Gg? EE, Ff, Gg?", + "EE FF GG EE FF GG EE FF GG EE FF GG EE FF GG", + ] + for i in range(10): + print(f"Run: {i}") + print(triplet_text_sampling(texts, pre_texts)) diff --git a/egs/libriheavy/ASR/zipformer_prompt_asr/decode_baseline.py b/egs/libriheavy/ASR/zipformer_prompt_asr/decode_baseline.py new file mode 100644 index 000000000..6a3bab3c8 --- /dev/null +++ b/egs/libriheavy/ASR/zipformer_prompt_asr/decode_baseline.py @@ -0,0 +1,791 @@ +#!/usr/bin/env python3 +# +# Copyright 2021-2022 Xiaomi Corporation (Author: Fangjun Kuang, +# Zengwei Yao, +# Xiaoyu Yang) +# +# See ../../../../LICENSE for clarification regarding multiple authors +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +""" +Usage: +(1) greedy search +./pruned_transducer_stateless7/decode.py \ + --epoch 28 \ + --avg 15 \ + --exp-dir ./pruned_transducer_stateless7/exp \ + --max-duration 600 \ + --decoding-method greedy_search + +(2) modified beam search +./pruned_transducer_stateless7/decode.py \ + --epoch 28 \ + --avg 15 \ + --exp-dir ./pruned_transducer_stateless7/exp \ + --max-duration 600 \ + --decoding-method modified_beam_search \ + --beam-size 4 + +""" + + +import argparse +import logging +import math +import warnings +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 LibriHeavyAsrDataModule +from beam_search import greedy_search, greedy_search_batch, modified_beam_search +from ls_text_normalization import word_normalization +from text_normalization import ( + ref_text_normalization, + remove_non_alphabetic, + upper_only_alpha, +) +from train_baseline import add_model_arguments, get_params, get_transducer_model +from utils import write_error_stats + +from icefall.checkpoint import ( + average_checkpoints, + average_checkpoints_with_averaged_model, + find_checkpoints, + load_checkpoint, +) +from icefall.lexicon import Lexicon +from icefall.utils import AttributeDict, setup_logger, store_transcripts, str2bool + +LOG_EPS = math.log(1e-10) + + +def get_parser(): + parser = argparse.ArgumentParser( + formatter_class=argparse.ArgumentDefaultsHelpFormatter + ) + + parser.add_argument( + "--epoch", + type=int, + default=30, + help="""It specifies the checkpoint to use for decoding. + Note: Epoch counts from 1. + You can specify --avg to use more checkpoints for model averaging.""", + ) + + parser.add_argument( + "--iter", + type=int, + default=0, + help="""If positive, --epoch is ignored and it + will use the checkpoint exp_dir/checkpoint-iter.pt. + You can specify --avg to use more checkpoints for model averaging. + """, + ) + + parser.add_argument( + "--avg", + type=int, + default=9, + help="Number of checkpoints to average. Automatically select " + "consecutive checkpoints before the checkpoint specified by " + "'--epoch' and '--iter'", + ) + + parser.add_argument( + "--use-averaged-model", + type=str2bool, + default=True, + help="Whether to load averaged model. Currently it only supports " + "using --epoch. If True, it would decode with the averaged model " + "over the epoch range from `epoch-avg` (excluded) to `epoch`." + "Actually only the models with epoch number of `epoch-avg` and " + "`epoch` are loaded for averaging. ", + ) + + parser.add_argument( + "--exp-dir", + type=str, + default="pruned_transducer_stateless7/exp", + help="The experiment dir", + ) + + parser.add_argument( + "--bpe-model", + type=str, + default="data/lang_bpe_500/bpe.model", + help="Path to the BPE model", + ) + + parser.add_argument( + "--lang-dir", + type=Path, + default="data/lang_bpe_500", + help="The lang dir containing word table and LG graph", + ) + + parser.add_argument( + "--decoding-method", + type=str, + default="greedy_search", + help="""Possible values are: + - greedy_search + - beam_search + - modified_beam_search + - fast_beam_search + - fast_beam_search_nbest + - fast_beam_search_nbest_oracle + - fast_beam_search_nbest_LG + - modified_beam_search_lm_shallow_fusion # for rnn lm shallow fusion + - modified_beam_search_LODR + If you use fast_beam_search_nbest_LG, you have to specify + `--lang-dir`, which should contain `LG.pt`. + """, + ) + + parser.add_argument( + "--beam-size", + type=int, + default=4, + help="""An integer indicating how many candidates we will keep for each + frame. Used only when --decoding-method is beam_search or + modified_beam_search.""", + ) + + parser.add_argument( + "--beam", + type=float, + default=20.0, + help="""A floating point value to calculate the cutoff score during beam + search (i.e., `cutoff = max-score - beam`), which is the same as the + `beam` in Kaldi. + Used only when --decoding-method is fast_beam_search, + fast_beam_search_nbest, fast_beam_search_nbest_LG, + and fast_beam_search_nbest_oracle + """, + ) + + parser.add_argument( + "--ngram-lm-scale", + type=float, + default=0.01, + help=""" + Used only when --decoding_method is fast_beam_search_nbest_LG. + It specifies the scale for n-gram LM scores. + """, + ) + + parser.add_argument( + "--max-contexts", + type=int, + default=8, + help="""Used only when --decoding-method is + fast_beam_search, fast_beam_search_nbest, fast_beam_search_nbest_LG, + and fast_beam_search_nbest_oracle""", + ) + + parser.add_argument( + "--max-states", + type=int, + default=64, + help="""Used only when --decoding-method is + fast_beam_search, fast_beam_search_nbest, fast_beam_search_nbest_LG, + and fast_beam_search_nbest_oracle""", + ) + + parser.add_argument( + "--max-sym-per-frame", + type=int, + default=1, + help="""Maximum number of symbols per frame. + Used only when --decoding_method is greedy_search""", + ) + + parser.add_argument( + "--num-paths", + type=int, + default=200, + help="""Number of paths for nbest decoding. + Used only when the decoding method is fast_beam_search_nbest, + fast_beam_search_nbest_LG, and fast_beam_search_nbest_oracle""", + ) + + parser.add_argument( + "--nbest-scale", + type=float, + default=0.5, + help="""Scale applied to lattice scores when computing nbest paths. + Used only when the decoding method is fast_beam_search_nbest, + fast_beam_search_nbest_LG, and fast_beam_search_nbest_oracle""", + ) + + parser.add_argument( + "--post-normalization", + type=str2bool, + default=True, + help="Normalized the recognition results by uppercasing and removing non-alphabetic symbols. ", + ) + + parser.add_argument( + "--long-audio-recog", + type=str2bool, + default=False, + ) + + parser.add_argument( + "--use-ls-test-set", + type=str2bool, + default=False, + help="Use librispeech test set for evaluation.", + ) + + parser.add_argument( + "--compute-CER", + type=str2bool, + default=True, + help="Reports CER. By default, only reports WER", + ) + + add_model_arguments(parser) + + return parser + + +def decode_one_batch( + params: AttributeDict, + model: nn.Module, + sp: spm.SentencePieceProcessor, + batch: dict, + word_table: Optional[k2.SymbolTable] = None, + decoding_graph: Optional[k2.Fsa] = None, +) -> Dict[str, List[List[str]]]: + """Decode one batch and return the result in a dict. The dict has the + following format: + + - key: It indicates the setting used for decoding. For example, + if greedy_search is used, it would be "greedy_search" + If beam search with a beam size of 7 is used, it would be + "beam_7" + - value: It contains the decoding result. `len(value)` equals to + batch size. `value[i]` is the decoding result for the i-th + utterance in the given batch. + Args: + params: + It's the return value of :func:`get_params`. + model: + The neural model. + sp: + The BPE model. + batch: + It is the return value from iterating + `lhotse.dataset.K2SpeechRecognitionDataset`. See its documentation + for the format of the `batch`. + word_table: + The word symbol table. + decoding_graph: + The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used + only when --decoding_method is fast_beam_search, fast_beam_search_nbest, + fast_beam_search_nbest_oracle, and fast_beam_search_nbest_LG. + LM: + A neural net LM for shallow fusion. Only used when `--use-shallow-fusion` + set to true. + ngram_lm: + A ngram lm. Used in LODR decoding. + ngram_lm_scale: + The scale of the ngram language model. + Returns: + Return the decoding result. See above description for the format of + the returned dict. + """ + device = next(model.parameters()).device + feature = batch["inputs"] + texts = batch["supervisions"]["text"] + batch_size = feature.size(0) + + # Get the transducer encoder output + 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) + + with warnings.catch_warnings(): + warnings.simplefilter("ignore") + encoder_out, encoder_out_lens = model.encode_audio( + feature=feature, + feature_lens=feature_lens, + ) + + hyps = [] + + if params.decoding_method == "greedy_search" and params.max_sym_per_frame == 1: + hyp_tokens = greedy_search_batch( + model=model, + encoder_out=encoder_out, + encoder_out_lens=encoder_out_lens, + ) + for hyp in sp.decode(hyp_tokens): + hyps.append(hyp.split()) + elif params.decoding_method == "modified_beam_search": + hyp_tokens = modified_beam_search( + model=model, + encoder_out=encoder_out, + encoder_out_lens=encoder_out_lens, + beam=params.beam_size, + ) + for hyp in sp.decode(hyp_tokens): + hyps.append(hyp.split()) + else: + batch_size = encoder_out.size(0) + + for i in range(batch_size): + # fmt: off + encoder_out_i = encoder_out[i:i+1, :encoder_out_lens[i]] + # fmt: on + if params.decoding_method == "greedy_search": + hyp = greedy_search( + model=model, + encoder_out=encoder_out_i, + max_sym_per_frame=params.max_sym_per_frame, + ) + else: + raise ValueError( + f"Unsupported decoding method: {params.decoding_method}" + ) + hyps.append(sp.decode(hyp).split()) + + if params.decoding_method == "greedy_search": + return {"greedy_search": hyps} + else: + return {f"beam_size_{params.beam_size}": hyps} + + +def decode_dataset( + dl: torch.utils.data.DataLoader, + params: AttributeDict, + model: nn.Module, + sp: spm.SentencePieceProcessor, + word_table: Optional[k2.SymbolTable] = None, + decoding_graph: Optional[k2.Fsa] = None, +) -> Dict[str, List[Tuple[str, List[str], List[str]]]]: + """Decode dataset. + + Args: + dl: + PyTorch's dataloader containing the dataset to decode. + params: + It is returned by :func:`get_params`. + model: + The neural model. + sp: + The BPE model. + word_table: + The word symbol table. + decoding_graph: + The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used + only when --decoding_method is fast_beam_search, fast_beam_search_nbest, + fast_beam_search_nbest_oracle, and fast_beam_search_nbest_LG. + LM: + A neural network LM, used during shallow fusion + Returns: + Return a dict, whose key may be "greedy_search" if greedy search + is used, or it may be "beam_7" if beam size of 7 is used. + Its value is a list of tuples. Each tuple contains two elements: + The first is the reference transcript, and the second is the + predicted result. + """ + num_cuts = 0 + + try: + num_batches = len(dl) + except TypeError: + num_batches = "?" + + if params.decoding_method == "greedy_search": + log_interval = 50 + else: + log_interval = 20 + + results = defaultdict(list) + for batch_idx, batch in enumerate(dl): + texts = batch["supervisions"]["text"] + cut_ids = [cut.id for cut in batch["supervisions"]["cut"]] + if not params.use_ls_test_set: + book_names = [ + cut.text_path.split("/")[-2] for cut in batch["supervisions"]["cut"] + ] + else: + book_names = ["" for _ in cut_ids] + + hyps_dict = decode_one_batch( + params=params, + model=model, + sp=sp, + decoding_graph=decoding_graph, + word_table=word_table, + batch=batch, + ) + + for name, hyps in hyps_dict.items(): + this_batch = [] + assert len(hyps) == len(texts) + for cut_id, book_name, hyp_words, ref_text in zip( + cut_ids, book_names, hyps, texts + ): + ref_text = ref_text_normalization(ref_text) + ref_words = ref_text.split() + this_batch.append((cut_id, ref_words, hyp_words)) + # if not params.use_ls_test_set: + # results[name + " " + book_name].extend(this_batch) + results[name].extend(this_batch) + + num_cuts += len(texts) + + if batch_idx % log_interval == 0: + batch_str = f"{batch_idx}/{num_batches}" + + logging.info(f"batch {batch_str}, cuts processed until now is {num_cuts}") + return results + + +def save_results( + params: AttributeDict, + test_set_name: str, + results_dict: Dict[str, List[Tuple[str, List[str], List[str]]]], + biasing_words: List[str] = None, +): + test_set_wers = dict() + test_set_cers = dict() + for key, results in results_dict.items(): + recog_path = params.res_dir / f"recogs-{test_set_name}-{params.suffix}.txt" + results = sorted(results) + store_transcripts(filename=recog_path, texts=results) + logging.info(f"The transcripts are stored in {recog_path}") + + # The following prints out WERs, per-word error statistics and aligned + # ref/hyp pairs. + errs_filename = params.res_dir / f"errs-{test_set_name}-{params.suffix}.txt" + with open(errs_filename, "w") as f: + wer = write_error_stats( + f, + f"{test_set_name}-{key}", + results, + enable_log=True, + biasing_words=biasing_words, + ) + test_set_wers[key] = wer + + logging.info("Wrote detailed error stats to {}".format(errs_filename)) + + if params.compute_CER: + # Write CER statistics + recog_path = ( + params.res_dir / f"recogs-{test_set_name}-char-{params.suffix}.txt" + ) + store_transcripts(filename=recog_path, texts=results, char_level=True) + errs_filename = ( + params.res_dir / f"errs-CER-{test_set_name}-{params.suffix}.txt" + ) + with open(errs_filename, "w") as f: + cer = write_error_stats( + f, + f"{test_set_name}-{key}", + results, + enable_log=True, + compute_CER=params.compute_CER, + ) + test_set_cers[key] = cer + + logging.info("Wrote detailed CER stats to {}".format(errs_filename)) + + test_set_wers = sorted(test_set_wers.items(), key=lambda x: x[1]) + errs_info = params.res_dir / f"wer-summary-{test_set_name}-{params.suffix}.txt" + with open(errs_info, "w") as f: + print("settings\tWER", file=f) + for key, val in test_set_wers: + print("{}\t{}".format(key, val), file=f) + + s = "\nFor {}, WER of different settings are:\n".format(test_set_name) + note = "\tbest for {}".format(test_set_name) + for key, val in test_set_wers: + s += "{}\t{}{}\n".format(key, val, note) + note = "" + logging.info(s) + + if params.compute_CER: + test_set_cers = sorted(test_set_cers.items(), key=lambda x: x[1]) + errs_info = params.res_dir / f"cer-summary-{test_set_name}-{params.suffix}.txt" + with open(errs_info, "w") as f: + print("settings\tcER", file=f) + for key, val in test_set_cers: + print("{}\t{}".format(key, val), file=f) + + s = "\nFor {}, CER of different settings are:\n".format(test_set_name) + note = "\tbest for {}".format(test_set_name) + for key, val in test_set_cers: + s += "{} CER\t{}{}\n".format(key, val, note) + note = "" + logging.info(s) + + +@torch.no_grad() +def main(): + parser = get_parser() + LibriHeavyAsrDataModule.add_arguments(parser) + args = parser.parse_args() + args.exp_dir = Path(args.exp_dir) + + params = get_params() + params.update(vars(args)) + + assert params.decoding_method in ( + "greedy_search", + "modified_beam_search", + ) + + if params.long_audio_recog: + params.res_dir = params.exp_dir / (params.decoding_method + "long_audio") + else: + params.res_dir = params.exp_dir / params.decoding_method + + if params.iter > 0: + params.suffix = f"iter-{params.iter}-avg-{params.avg}" + else: + params.suffix = f"epoch-{params.epoch}-avg-{params.avg}" + + if "beam_search" in params.decoding_method: + params.suffix += f"-{params.decoding_method}-beam-size-{params.beam_size}" + else: + params.suffix += f"-context-{params.context_size}" + params.suffix += f"-max-sym-per-frame-{params.max_sym_per_frame}" + + if "ngram" in params.decoding_method: + params.suffix += f"-ngram-lm-scale-{params.ngram_lm_scale}" + + if params.use_averaged_model: + params.suffix += "-use-averaged-model" + + setup_logger(f"{params.res_dir}/log-decode-{params.suffix}") + logging.info("Decoding started") + + device = torch.device("cpu") + if torch.cuda.is_available(): + device = torch.device("cuda", 0) + + logging.info(f"Device: {device}") + + sp = spm.SentencePieceProcessor() + sp.load(params.bpe_model) + + # and are defined in local/train_bpe_model.py + params.blank_id = sp.piece_to_id("") + params.unk_id = sp.piece_to_id("") + params.vocab_size = sp.get_piece_size() + + logging.info(params) + + logging.info("About to create model") + model = get_transducer_model(params) + + if not params.use_averaged_model: + if params.iter > 0: + filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[ + : params.avg + ] + if len(filenames) == 0: + raise ValueError( + f"No checkpoints found for" + f" --iter {params.iter}, --avg {params.avg}" + ) + elif len(filenames) < params.avg: + raise ValueError( + f"Not enough checkpoints ({len(filenames)}) found for" + f" --iter {params.iter}, --avg {params.avg}" + ) + logging.info(f"averaging {filenames}") + model.to(device) + model.load_state_dict(average_checkpoints(filenames, device=device)) + elif params.avg == 1: + load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model) + else: + start = params.epoch - params.avg + 1 + filenames = [] + for i in range(start, params.epoch + 1): + if i >= 1: + filenames.append(f"{params.exp_dir}/epoch-{i}.pt") + logging.info(f"averaging {filenames}") + model.to(device) + model.load_state_dict(average_checkpoints(filenames, device=device)) + else: + if params.iter > 0: + filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[ + : params.avg + 1 + ] + if len(filenames) == 0: + raise ValueError( + f"No checkpoints found for" + f" --iter {params.iter}, --avg {params.avg}" + ) + elif len(filenames) < params.avg + 1: + raise ValueError( + f"Not enough checkpoints ({len(filenames)}) found for" + f" --iter {params.iter}, --avg {params.avg}" + ) + filename_start = filenames[-1] + filename_end = filenames[0] + logging.info( + "Calculating the averaged model over iteration checkpoints" + f" from {filename_start} (excluded) to {filename_end}" + ) + model.to(device) + model.load_state_dict( + average_checkpoints_with_averaged_model( + filename_start=filename_start, + filename_end=filename_end, + device=device, + ), + strict=False, + ) + 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, + ), + strict=False, + ) + + model.to(device) + model.eval() + + LM = None + + decoding_graph = None + word_table = None + + num_param = sum([p.numel() for p in model.parameters()]) + logging.info(f"Number of model parameters: {num_param}") + + # we need cut ids to display recognition results. + args.return_cuts = True + libriheavy = LibriHeavyAsrDataModule(args) + + test_clean_cuts = libriheavy.test_clean_cuts() + test_other_cuts = libriheavy.test_other_cuts() + ls_test_clean_cuts = libriheavy.librispeech_test_clean_cuts() + ls_test_other_cuts = libriheavy.librispeech_test_other_cuts() + long_audio_cuts = libriheavy.long_audio_cuts() + + test_clean_dl = libriheavy.valid_dataloaders( + test_clean_cuts, + ) + test_other_dl = libriheavy.valid_dataloaders( + test_other_cuts, + ) + ls_test_clean_dl = libriheavy.test_dataloaders(ls_test_clean_cuts) + ls_test_other_dl = libriheavy.test_dataloaders(ls_test_other_cuts) + long_audio_dl = libriheavy.valid_dataloaders( + long_audio_cuts, + ) + + if params.use_ls_test_set: + test_sets = ["ls-test-clean", "ls-test-other"] + test_dl = [ls_test_clean_dl, ls_test_other_dl] + else: + test_sets = ["test-clean", "test-other"] + test_dl = [test_clean_dl, test_other_dl] + + if params.long_audio_recog: + test_sets = ["long-audio"] + test_dl = [long_audio_dl] + + for test_set, test_dl in zip(test_sets, test_dl): + if params.use_ls_test_set: + f = open( + "data/context_biasing/LibriSpeechBiasingLists/all_rare_words.txt", "r" + ) + biasing_words = f.read().strip().split() + f.close() + else: + biasing_words = None + results_dict = decode_dataset( + dl=test_dl, + params=params, + model=model, + sp=sp, + word_table=word_table, + decoding_graph=decoding_graph, + ) + + save_results( + params=params, + test_set_name=test_set, + results_dict=results_dict, + ) + + if params.post_normalization: + if "-post-normalization" not in params.suffix: + params.suffix += "-post-normalization" + + new_res = {} + for k in results_dict: + new_ans = [] + for item in results_dict[k]: + id, ref, hyp = item + if params.use_ls_test_set: + hyp = ( + " ".join(hyp).replace("-", " ").split() + ) # handle the hypens + hyp = upper_only_alpha(" ".join(hyp)).split() + hyp = [word_normalization(w.upper()) for w in hyp] + hyp = " ".join(hyp).split() + hyp = [w for w in hyp if w != ""] + ref = upper_only_alpha(" ".join(ref)).split() + else: + hyp = upper_only_alpha(" ".join(hyp)).split() + ref = upper_only_alpha(" ".join(ref)).split() + new_ans.append((id, ref, hyp)) + new_res[k] = new_ans + + save_results( + params=params, + test_set_name=test_set, + results_dict=new_res, + biasing_words=biasing_words, + ) + + if params.suffix.endswith("-post-normalization"): + params.suffix = params.suffix.replace("-post-normalization", "") + + logging.info("Done!") + + +if __name__ == "__main__": + main() diff --git a/egs/libriheavy/ASR/zipformer_prompt_asr/decode_bert.py b/egs/libriheavy/ASR/zipformer_prompt_asr/decode_bert.py new file mode 100755 index 000000000..e71999b0a --- /dev/null +++ b/egs/libriheavy/ASR/zipformer_prompt_asr/decode_bert.py @@ -0,0 +1,1025 @@ +#!/usr/bin/env python3 +# +# Copyright 2021-2022 Xiaomi Corporation (Author: Fangjun Kuang, +# Zengwei Yao, +# Xiaoyu Yang) +# +# See ../../../../LICENSE for clarification regarding multiple authors +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +""" +Usage: +(1) greedy search +./zipformer_prompt_asr/decode_bert.py \ + --epoch 28 \ + --avg 15 \ + --exp-dir ./zipformer_prompt_asr/exp \ + --max-duration 1000 \ + --decoding-method greedy_search \ + --text-encoder-type BERT \ + --memory-layer 0 \ + --use-pre-text True \ + --use-style-prompt True \ + --max-prompt-lens 1000 \ + --style-text-transform mixed-punc \ + --pre-text-transform mixed-punc \ + --compute-CER 0 + + +(2) modified beam search +./zipformer_prompt_asr/decode_bert.py \ + --epoch 28 \ + --avg 15 \ + --exp-dir ./zipformer_prompt_asr/exp \ + --max-duration 1000 \ + --decoding-method modified_beam_search \ + --beam-size 4 \ + --text-encoder-type BERT \ + --memory-layer 0 \ + --use-pre-text True \ + --use-style-prompt True \ + --max-prompt-lens 1000 \ + --style-text-transform mixed-punc \ + --pre-text-transform mixed-punc \ + --compute-CER 0 + +(3) Decode LibriSpeech + +./zipformer_prompt_asr/decode_bert.py \ + --epoch 28 \ + --avg 15 \ + --exp-dir ./zipformer_prompt_asr/exp \ + --max-duration 1000 \ + --decoding-method modified_beam_search \ + --use-ls-test-set True \ + --beam-size 4 \ + --text-encoder-type BERT \ + --memory-layer 0 \ + --use-pre-text True \ + --use-style-prompt True \ + --max-prompt-lens 1000 \ + --style-text-transform mixed-punc \ + --pre-text-transform mixed-punc \ + --compute-CER 0 + +(4) Decode LibriSpeech + biasing list + +biasing_list=100 # could also be 0 + +./zipformer_prompt_asr/decode_bert.py \ + --epoch 28 \ + --avg 15 \ + --exp-dir ./zipformer_prompt_asr/exp \ + --max-duration 1000 \ + --decoding-method modified_beam_search \ + --beam-size 4 \ + --use-ls-test-set True \ + --use-ls-context-list True \ + --biasing-level utterance \ + --ls-distractors $biasing_list \ + --post-normalization True \ + --text-encoder-type BERT \ + --max-prompt-lens 1000 \ + --style-text-transform mixed-punc \ + --pre-text-transform mixed-punc + + +""" + + +import argparse +import logging +import math +import warnings +from collections import defaultdict +from pathlib import Path +from typing import Callable, Dict, List, Optional, Tuple + +import k2 +import sentencepiece as spm +import torch +import torch.nn as nn +from asr_datamodule import LibriHeavyAsrDataModule +from beam_search import greedy_search, greedy_search_batch, modified_beam_search +from dataset import naive_triplet_text_sampling, random_shuffle_subset +from ls_text_normalization import word_normalization +from text_normalization import ( + _apply_style_transform, + lower_all_char, + lower_only_alpha, + ref_text_normalization, + remove_non_alphabetic, + train_text_normalization, + upper_all_char, + upper_only_alpha, +) +from train_bert_encoder import ( + _encode_texts_as_bytes_with_tokenizer, + add_model_arguments, + get_params, + get_tokenizer, + get_transducer_model, +) +from transformers import BertModel, BertTokenizer +from utils import brian_biasing_list, get_facebook_biasing_list, write_error_stats + +from icefall.checkpoint import ( + average_checkpoints, + average_checkpoints_with_averaged_model, + find_checkpoints, + load_checkpoint, +) +from icefall.lexicon import Lexicon +from icefall.utils import AttributeDict, setup_logger, store_transcripts, str2bool + +LOG_EPS = math.log(1e-10) + + +def get_parser(): + parser = argparse.ArgumentParser( + formatter_class=argparse.ArgumentDefaultsHelpFormatter + ) + + parser.add_argument( + "--epoch", + type=int, + default=30, + help="""It specifies the checkpoint to use for decoding. + Note: Epoch counts from 1. + You can specify --avg to use more checkpoints for model averaging.""", + ) + + parser.add_argument( + "--iter", + type=int, + default=0, + help="""If positive, --epoch is ignored and it + will use the checkpoint exp_dir/checkpoint-iter.pt. + You can specify --avg to use more checkpoints for model averaging. + """, + ) + + parser.add_argument( + "--avg", + type=int, + default=9, + help="Number of checkpoints to average. Automatically select " + "consecutive checkpoints before the checkpoint specified by " + "'--epoch' and '--iter'", + ) + + parser.add_argument( + "--use-averaged-model", + type=str2bool, + default=True, + help="Whether to load averaged model. Currently it only supports " + "using --epoch. If True, it would decode with the averaged model " + "over the epoch range from `epoch-avg` (excluded) to `epoch`." + "Actually only the models with epoch number of `epoch-avg` and " + "`epoch` are loaded for averaging. ", + ) + + parser.add_argument( + "--exp-dir", + type=str, + default="pruned_transducer_stateless7/exp", + help="The experiment dir", + ) + + parser.add_argument( + "--bpe-model", + type=str, + default="data/lang_bpe_500/bpe.model", + help="Path to the BPE model", + ) + + parser.add_argument( + "--lang-dir", + type=Path, + default="data/lang_bpe_500", + help="The lang dir containing word table and LG graph", + ) + + parser.add_argument( + "--decoding-method", + type=str, + default="greedy_search", + help="""Possible values are: + - greedy_search + - beam_search + - modified_beam_search + - fast_beam_search + - fast_beam_search_nbest + - fast_beam_search_nbest_oracle + - fast_beam_search_nbest_LG + - modified_beam_search_lm_shallow_fusion # for rnn lm shallow fusion + - modified_beam_search_LODR + If you use fast_beam_search_nbest_LG, you have to specify + `--lang-dir`, which should contain `LG.pt`. + """, + ) + + parser.add_argument( + "--beam-size", + type=int, + default=4, + help="""An integer indicating how many candidates we will keep for each + frame. Used only when --decoding-method is beam_search or + modified_beam_search.""", + ) + + parser.add_argument( + "--beam", + type=float, + default=20.0, + help="""A floating point value to calculate the cutoff score during beam + search (i.e., `cutoff = max-score - beam`), which is the same as the + `beam` in Kaldi. + Used only when --decoding-method is fast_beam_search, + fast_beam_search_nbest, fast_beam_search_nbest_LG, + and fast_beam_search_nbest_oracle + """, + ) + + parser.add_argument( + "--ngram-lm-scale", + type=float, + default=0.01, + help=""" + Used only when --decoding_method is fast_beam_search_nbest_LG. + It specifies the scale for n-gram LM scores. + """, + ) + + parser.add_argument( + "--max-contexts", + type=int, + default=8, + help="""Used only when --decoding-method is + fast_beam_search, fast_beam_search_nbest, fast_beam_search_nbest_LG, + and fast_beam_search_nbest_oracle""", + ) + + parser.add_argument( + "--max-states", + type=int, + default=64, + help="""Used only when --decoding-method is + fast_beam_search, fast_beam_search_nbest, fast_beam_search_nbest_LG, + and fast_beam_search_nbest_oracle""", + ) + + parser.add_argument( + "--max-sym-per-frame", + type=int, + default=1, + help="""Maximum number of symbols per frame. + Used only when --decoding_method is greedy_search""", + ) + + parser.add_argument( + "--num-paths", + type=int, + default=200, + help="""Number of paths for nbest decoding. + Used only when the decoding method is fast_beam_search_nbest, + fast_beam_search_nbest_LG, and fast_beam_search_nbest_oracle""", + ) + + parser.add_argument( + "--nbest-scale", + type=float, + default=0.5, + help="""Scale applied to lattice scores when computing nbest paths. + Used only when the decoding method is fast_beam_search_nbest, + fast_beam_search_nbest_LG, and fast_beam_search_nbest_oracle""", + ) + + parser.add_argument( + "--use-pre-text", + type=str2bool, + default=True, + help="Use pre-text is available during decoding", + ) + + parser.add_argument( + "--use-style-prompt", + type=str2bool, + default=True, + help="Use style prompt when evaluation", + ) + + parser.add_argument( + "--max-prompt-lens", + type=int, + default=1000, + ) + + parser.add_argument( + "--post-normalization", + type=str2bool, + default=True, + help="Normalized the recognition results by uppercasing and removing non-alphabetic symbols. ", + ) + + parser.add_argument( + "--compute-CER", + type=str2bool, + default=False, + help="Reports CER. By default, only reports WER", + ) + + parser.add_argument( + "--style-text-transform", + type=str, + choices=["mixed-punc", "upper-no-punc", "lower-no-punc", "lower-punc"], + default="mixed-punc", + help="The style of style prompt, i.e style_text", + ) + + parser.add_argument( + "--pre-text-transform", + type=str, + choices=["mixed-punc", "upper-no-punc", "lower-no-punc", "lower-punc"], + default="mixed-punc", + help="The style of content prompt, i.e pre_text", + ) + + parser.add_argument( + "--use-ls-test-set", + type=str2bool, + default=False, + help="Use librispeech test set for evaluation.", + ) + + parser.add_argument( + "--use-ls-context-list", + type=str2bool, + default=False, + help="If use a fixed context list for LibriSpeech decoding", + ) + + parser.add_argument( + "--biasing-level", + type=str, + default="utterance", + choices=["utterance", "Book", "Chapter"], + ) + + parser.add_argument( + "--ls-distractors", + type=int, + default=0, + help="The number of distractors into context list for LibriSpeech decoding", + ) + + add_model_arguments(parser) + + return parser + + +def decode_one_batch( + params: AttributeDict, + model: nn.Module, + sp: spm.SentencePieceProcessor, + tokenizer: spm.SentencePieceProcessor, + batch: dict, + biasing_dict: dict = None, + word_table: Optional[k2.SymbolTable] = None, + decoding_graph: Optional[k2.Fsa] = None, +) -> Dict[str, List[List[str]]]: + """Decode one batch and return the result in a dict. The dict has the + following format: + + - key: It indicates the setting used for decoding. For example, + if greedy_search is used, it would be "greedy_search" + If beam search with a beam size of 7 is used, it would be + "beam_7" + - value: It contains the decoding result. `len(value)` equals to + batch size. `value[i]` is the decoding result for the i-th + utterance in the given batch. + Args: + params: + It's the return value of :func:`get_params`. + model: + The neural model. + sp: + The BPE model. + tokenizer: + Tokenizer for the text encoder + batch: + It is the return value from iterating + `lhotse.dataset.K2SpeechRecognitionDataset`. See its documentation + for the format of the `batch`. + biasing_dict: + A dictionary in the form `{cut_id: :w1 w2"}` that contains a list + of biasing words (separated with space) + word_table: + The word symbol table. + decoding_graph: + The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used + only when --decoding_method is fast_beam_search, fast_beam_search_nbest, + fast_beam_search_nbest_oracle, and fast_beam_search_nbest_LG. + LM: + A neural net LM for shallow fusion. Only used when `--use-shallow-fusion` + set to true. + ngram_lm: + A ngram lm. Used in LODR decoding. + ngram_lm_scale: + The scale of the ngram language model. + Returns: + Return the decoding result. See above description for the format of + the returned dict. + """ + device = next(model.parameters()).device + feature = batch["inputs"] + cuts = batch["supervisions"]["cut"] + cut_ids = [c.supervisions[0].id for c in cuts] + batch_size = feature.size(0) + + if "pre_text" in batch["supervisions"] and params.use_pre_text: + pre_texts = batch["supervisions"]["pre_text"] + pre_texts = [train_text_normalization(t) for t in pre_texts] + else: + pre_texts = ["" for _ in range(batch_size)] + + # get the librispeech biasing data + if params.use_pre_text and (params.use_ls_context_list and params.use_ls_test_set): + if params.biasing_level == "utterance": + pre_texts = [biasing_dict[id] for id in cut_ids] + elif params.biasing_level == "Chapter": + chapter_ids = [c.split("-")[1] for c in cut_ids] + pre_texts = [biasing_dict[id] for id in chapter_ids] + elif params.biasing_level == "Book": + chapter_ids = [c.split("-")[1] for c in cut_ids] + pre_texts = [biasing_dict[id] for id in chapter_ids] + else: + raise ValueError(f"Unseen biasing level: {params.biasing_level}") + if params.pre_text_transform == "mixed-punc": + pre_texts = [t.lower() for t in pre_texts] + + # get style_text + if params.use_style_prompt: + fixed_sentence = "Mixed-case English transcription, with punctuation. Actually, it's fully not related." + style_texts = batch["supervisions"].get( + "style_text", [fixed_sentence for _ in range(batch_size)] + ) + style_texts = [train_text_normalization(t) for t in style_texts] + else: + style_texts = ["" for _ in range(batch_size)] # use empty string + + # Get the text embedding + if params.use_pre_text or params.use_style_prompt: + # apply style transform to the pre_text and style_text + pre_texts = _apply_style_transform(pre_texts, params.pre_text_transform) + if not params.use_ls_context_list: + pre_texts = [t[-params.max_prompt_lens :] for t in pre_texts] + + if params.use_style_prompt: + style_texts = _apply_style_transform( + style_texts, params.style_text_transform + ) + + with warnings.catch_warnings(): + warnings.simplefilter("ignore") + + # Use tokenizer to prepare input for text encoder + encoded_inputs, style_lens = _encode_texts_as_bytes_with_tokenizer( + pre_texts=pre_texts, + style_texts=style_texts, + tokenizer=tokenizer, + device=device, + no_limit=True, + ) + logging.info( + f"Shape of the encoded prompts: {encoded_inputs['input_ids'].shape}" + ) + + memory, memory_key_padding_mask = model.encode_text( + encoded_inputs=encoded_inputs, + style_lens=style_lens, + ) # (T,B,C) + else: + memory = None + memory_key_padding_mask = None + + # Get the transducer encoder output + 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) + + with warnings.catch_warnings(): + warnings.simplefilter("ignore") + encoder_out, encoder_out_lens = model.encode_audio( + feature=feature, + feature_lens=feature_lens, + memory=memory, + memory_key_padding_mask=memory_key_padding_mask, + ) + + hyps = [] + + if params.decoding_method == "greedy_search" and params.max_sym_per_frame == 1: + hyp_tokens = greedy_search_batch( + model=model, + encoder_out=encoder_out, + encoder_out_lens=encoder_out_lens, + ) + for hyp in sp.decode(hyp_tokens): + hyps.append(hyp.split()) + elif params.decoding_method == "modified_beam_search": + hyp_tokens = modified_beam_search( + model=model, + encoder_out=encoder_out, + encoder_out_lens=encoder_out_lens, + beam=params.beam_size, + ) + for hyp in sp.decode(hyp_tokens): + hyps.append(hyp.split()) + else: + batch_size = encoder_out.size(0) + + for i in range(batch_size): + # fmt: off + encoder_out_i = encoder_out[i:i+1, :encoder_out_lens[i]] + # fmt: on + if params.decoding_method == "greedy_search": + hyp = greedy_search( + model=model, + encoder_out=encoder_out_i, + max_sym_per_frame=params.max_sym_per_frame, + ) + else: + raise ValueError( + f"Unsupported decoding method: {params.decoding_method}" + ) + hyps.append(sp.decode(hyp).split()) + + if params.decoding_method == "greedy_search": + return {"greedy_search": hyps} + else: + return {f"beam_size_{params.beam_size}": hyps} + + +def decode_dataset( + dl: torch.utils.data.DataLoader, + params: AttributeDict, + model: nn.Module, + sp: spm.SentencePieceProcessor, + tokenizer: spm.SentencePieceProcessor, + biasing_dict: Dict = None, + word_table: Optional[k2.SymbolTable] = None, + decoding_graph: Optional[k2.Fsa] = None, +) -> Dict[str, List[Tuple[str, List[str], List[str]]]]: + """Decode dataset. + + Args: + dl: + PyTorch's dataloader containing the dataset to decode. + params: + It is returned by :func:`get_params`. + model: + The neural model. + sp: + The BPE model. + tokenizer: + Tokenizer for the text encoder + biasing_dict: + A dictionary in the form `{cut_id: :w1 w2"}` that contains a list + of biasing words (separated with space) + word_table: + The word symbol table. + decoding_graph: + The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used + only when --decoding_method is fast_beam_search, fast_beam_search_nbest, + fast_beam_search_nbest_oracle, and fast_beam_search_nbest_LG. + LM: + A neural network LM, used during shallow fusion + Returns: + Return a dict, whose key may be "greedy_search" if greedy search + is used, or it may be "beam_7" if beam size of 7 is used. + Its value is a list of tuples. Each tuple contains two elements: + The first is the reference transcript, and the second is the + predicted result. + """ + num_cuts = 0 + + try: + num_batches = len(dl) + except TypeError: + num_batches = "?" + + if params.decoding_method == "greedy_search": + log_interval = 50 + else: + log_interval = 20 + + results = defaultdict(list) + for batch_idx, batch in enumerate(dl): + texts = batch["supervisions"][ + "text" + ] # By default, this should be in mixed-punc format + + # the style of ref_text should match style_text + texts = _apply_style_transform(texts, params.style_text_transform) + if params.use_style_prompt: + texts = _apply_style_transform(texts, params.style_text_transform) + + cut_ids = [cut.id for cut in batch["supervisions"]["cut"]] + if not params.use_ls_test_set: + try: + book_names = [ + cut.text_path.split("/")[-2] for cut in batch["supervisions"]["cut"] + ] + except AttributeError: + book_names = [ + cut.id.split("/")[0] for cut in batch["supervisions"]["cut"] + ] + else: + book_names = ["" for _ in cut_ids] + + hyps_dict = decode_one_batch( + params=params, + model=model, + sp=sp, + tokenizer=tokenizer, + biasing_dict=biasing_dict, + decoding_graph=decoding_graph, + word_table=word_table, + batch=batch, + ) + + for name, hyps in hyps_dict.items(): + this_batch = [] + assert len(hyps) == len(texts) + for cut_id, book_name, hyp_words, ref_text in zip( + cut_ids, book_names, hyps, texts + ): + ref_text = ref_text_normalization( + ref_text + ) # remove full-width symbols & some book marks + ref_words = ref_text.split() + this_batch.append((cut_id, ref_words, hyp_words)) + results[name].extend(this_batch) + + num_cuts += len(texts) + + if batch_idx % log_interval == 0: + batch_str = f"{batch_idx}/{num_batches}" + + logging.info(f"batch {batch_str}, cuts processed until now is {num_cuts}") + return results + + +def save_results( + params: AttributeDict, + test_set_name: str, + results_dict: Dict[str, List[Tuple[str, List[str], List[str]]]], + biasing_words: List[str] = None, +): + test_set_wers = dict() + test_set_cers = dict() + for key, results in results_dict.items(): + recog_path = params.res_dir / f"recogs-{test_set_name}-{params.suffix}.txt" + results = sorted(results) + store_transcripts(filename=recog_path, texts=results) + logging.info(f"The transcripts are stored in {recog_path}") + + # The following prints out WERs, per-word error statistics and aligned + # ref/hyp pairs. + errs_filename = params.res_dir / f"errs-{test_set_name}-{params.suffix}.txt" + with open(errs_filename, "w") as f: + wer = write_error_stats( + f, f"{test_set_name}-{key}", results, enable_log=True + ) + test_set_wers[key] = wer + + logging.info("Wrote detailed error stats to {}".format(errs_filename)) + + if params.compute_CER: + # Write CER statistics + recog_path = ( + params.res_dir / f"recogs-{test_set_name}-char-{params.suffix}.txt" + ) + store_transcripts(filename=recog_path, texts=results, char_level=True) + errs_filename = ( + params.res_dir / f"errs-CER-{test_set_name}-{params.suffix}.txt" + ) + with open(errs_filename, "w") as f: + cer = write_error_stats( + f, + f"{test_set_name}-{key}", + results, + enable_log=True, + compute_CER=params.compute_CER, + ) + test_set_cers[key] = cer + + logging.info("Wrote detailed CER stats to {}".format(errs_filename)) + + test_set_wers = sorted(test_set_wers.items(), key=lambda x: x[1]) + errs_info = params.res_dir / f"wer-summary-{test_set_name}-{params.suffix}.txt" + with open(errs_info, "w") as f: + print("settings\tWER", file=f) + for key, val in test_set_wers: + print("{}\t{}".format(key, val), file=f) + + s = "\nFor {}, WER of different settings are:\n".format(test_set_name) + note = "\tbest for {}".format(test_set_name) + for key, val in test_set_wers: + s += "{}\t{}{}\n".format(key, val, note) + note = "" + logging.info(s) + + if params.compute_CER: + test_set_cers = sorted(test_set_cers.items(), key=lambda x: x[1]) + errs_info = params.res_dir / f"cer-summary-{test_set_name}-{params.suffix}.txt" + with open(errs_info, "w") as f: + print("settings\tCER", file=f) + for key, val in test_set_cers: + print("{}\t{}".format(key, val), file=f) + + s = "\nFor {}, CER of different settings are:\n".format(test_set_name) + note = "\tbest for {}".format(test_set_name) + for key, val in test_set_cers: + s += "{} CER\t{}{}\n".format(key, val, note) + note = "" + logging.info(s) + + +@torch.no_grad() +def main(): + parser = get_parser() + LibriHeavyAsrDataModule.add_arguments(parser) + args = parser.parse_args() + args.exp_dir = Path(args.exp_dir) + + params = get_params() + params.update(vars(args)) + + assert params.decoding_method in ( + "greedy_search", + "modified_beam_search", + ) + + params.res_dir = params.exp_dir / params.decoding_method + + if params.iter > 0: + params.suffix = f"iter-{params.iter}-avg-{params.avg}" + else: + params.suffix = f"epoch-{params.epoch}-avg-{params.avg}" + + if params.causal: + assert ( + "," not in params.chunk_size + ), "chunk_size should be one value in decoding." + assert ( + "," not in params.left_context_frames + ), "left_context_frames should be one value in decoding." + params.suffix += f"-chunk-{params.chunk_size}" + params.suffix += f"-left-context-{params.left_context_frames}" + + if "beam_search" in params.decoding_method: + params.suffix += f"-{params.decoding_method}-beam-size-{params.beam_size}" + else: + params.suffix += f"-context-{params.context_size}" + params.suffix += f"-max-sym-per-frame-{params.max_sym_per_frame}" + + if params.use_pre_text: + params.suffix += ( + f"-pre-text-{params.pre_text_transform}-len-{params.max_prompt_lens}" + ) + + if params.use_style_prompt: + params.suffix += f"-style-prompt-{params.style_text_transform}" + + if params.use_ls_context_list: + assert ( + params.use_pre_text + ), "Must set --use-pre-text to True if using context list" + params.suffix += f"-use-{params.biasing_level}-level-ls-context-list" + if params.biasing_level == "utterance" and params.ls_distractors: + params.suffix += f"-ls-context-distractors-{params.ls_distractors}" + + if params.use_averaged_model: + params.suffix += "-use-averaged-model" + + setup_logger(f"{params.res_dir}/log-decode-{params.suffix}") + logging.info("Decoding started") + + device = torch.device("cpu") + if torch.cuda.is_available(): + device = torch.device("cuda", 0) + + logging.info(f"Device: {device}") + + sp = spm.SentencePieceProcessor() + sp.load(params.bpe_model) + + # and are defined in local/train_bpe_model.py + params.blank_id = sp.piece_to_id("") + params.unk_id = sp.piece_to_id("") + params.vocab_size = sp.get_piece_size() + + logging.info(params) + + logging.info("About to create model") + model = get_transducer_model(params) + tokenizer = get_tokenizer(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, + ), + strict=False, + ) + + model.to(device) + model.eval() + + LM = None + + decoding_graph = None + word_table = None + + num_param = sum([p.numel() for p in model.parameters()]) + logging.info(f"Number of model parameters: {num_param}") + + # we need cut ids to display recognition results. + args.return_cuts = True + libriheavy = LibriHeavyAsrDataModule(args) + + test_clean_cuts = libriheavy.test_clean_cuts() + test_other_cuts = libriheavy.test_other_cuts() + ls_test_clean_cuts = libriheavy.librispeech_test_clean_cuts() + ls_test_other_cuts = libriheavy.librispeech_test_other_cuts() + + test_clean_dl = libriheavy.valid_dataloaders( + test_clean_cuts, text_sampling_func=naive_triplet_text_sampling + ) + test_other_dl = libriheavy.valid_dataloaders( + test_other_cuts, text_sampling_func=naive_triplet_text_sampling + ) + ls_test_clean_dl = libriheavy.test_dataloaders(ls_test_clean_cuts) + ls_test_other_dl = libriheavy.test_dataloaders(ls_test_other_cuts) + + if params.use_ls_test_set: + test_sets = ["ls-test-clean", "ls-test-other"] + test_dl = [ls_test_clean_dl, ls_test_other_dl] + else: + 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): + biasing_dict = None + if params.use_ls_context_list: + if test_set == "ls-test-clean": + biasing_dict = get_facebook_biasing_list( + test_set="test-clean", + num_distractors=params.ls_distractors, + ) + elif test_set == "ls-test-other": + biasing_dict = get_facebook_biasing_list( + test_set="test-other", + num_distractors=params.ls_distractors, + ) + + results_dict = decode_dataset( + dl=test_dl, + params=params, + model=model, + sp=sp, + tokenizer=tokenizer, + biasing_dict=biasing_dict, + word_table=word_table, + decoding_graph=decoding_graph, + ) + + save_results( + params=params, + test_set_name=test_set, + results_dict=results_dict, + ) + + if params.post_normalization: + if "-post-normalization" not in params.suffix: + params.suffix += "-post-normalization" + + new_res = {} + for k in results_dict: + new_ans = [] + for item in results_dict[k]: + id, ref, hyp = item + if params.use_ls_test_set: + hyp = ( + " ".join(hyp).replace("-", " ").split() + ) # handle the hypens + hyp = upper_only_alpha(" ".join(hyp)).split() + hyp = [word_normalization(w.upper()) for w in hyp] + hyp = " ".join(hyp).split() + hyp = [w for w in hyp if w != ""] + ref = upper_only_alpha(" ".join(ref)).split() + else: + hyp = upper_only_alpha(" ".join(hyp)).split() + ref = upper_only_alpha(" ".join(ref)).split() + new_ans.append((id, ref, hyp)) + new_res[k] = new_ans + + save_results( + params=params, + test_set_name=test_set, + results_dict=new_res, + ) + + if params.suffix.endswith("-post-normalization"): + params.suffix = params.suffix.replace("-post-normalization", "") + + logging.info("Done!") + + +if __name__ == "__main__": + main() diff --git a/egs/libriheavy/ASR/zipformer_prompt_asr/decode_bert_with_style_save_decoding_mp.py b/egs/libriheavy/ASR/zipformer_prompt_asr/decode_bert_with_style_save_decoding_mp.py new file mode 100755 index 000000000..4559ebb6d --- /dev/null +++ b/egs/libriheavy/ASR/zipformer_prompt_asr/decode_bert_with_style_save_decoding_mp.py @@ -0,0 +1,963 @@ +#!/usr/bin/env python3 +# +# Copyright 2021-2022 Xiaomi Corporation (Author: Fangjun Kuang, +# Zengwei Yao, +# Xiaoyu Yang) +# +# See ../../../../LICENSE for clarification regarding multiple authors +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +""" +Usage: +(1) greedy search +./pruned_transducer_stateless7/decode.py \ + --epoch 28 \ + --avg 15 \ + --exp-dir ./pruned_transducer_stateless7/exp \ + --max-duration 600 \ + --decoding-method greedy_search + +(2) modified beam search +./pruned_transducer_stateless7/decode.py \ + --epoch 28 \ + --avg 15 \ + --exp-dir ./pruned_transducer_stateless7/exp \ + --max-duration 600 \ + --decoding-method modified_beam_search \ + --beam-size 4 + +""" + + +import argparse +import logging +import math +import warnings +from collections import defaultdict +from pathlib import Path +from typing import Callable, Dict, List, Optional, Tuple + +import k2 +import sentencepiece as spm +import torch +import torch.multiprocessing as mp +import torch.nn as nn +from asr_datamodule import LibriHeavyAsrDataModule +from beam_search import ( + greedy_search, + greedy_search_batch, + greedy_search_batch_with_context, + greedy_search_with_context, + modified_beam_search, +) +from dataset import naive_triplet_text_sampling, random_shuffle_subset +from lhotse import load_manifest_lazy +from text_normalization import ( + lower_all_char, + lower_only_alpha, + ref_text_normalization, + remove_non_alphabetic, + train_text_normalization, + upper_all_char, + upper_only_alpha, +) +from train_bert_encoder_with_style import ( + _encode_texts_as_bytes_with_tokenizer, + add_model_arguments, + get_params, + get_tokenizer, + get_transducer_model, +) +from transformers import BertModel, BertTokenizer +from utils import get_facebook_biasing_list + +from icefall.checkpoint import ( + average_checkpoints, + average_checkpoints_with_averaged_model, + find_checkpoints, + load_checkpoint, +) +from icefall.lexicon import Lexicon +from icefall.utils import ( + AttributeDict, + setup_logger, + store_transcripts, + str2bool, + write_error_stats, +) + +LOG_EPS = math.log(1e-10) + + +def get_parser(): + parser = argparse.ArgumentParser( + formatter_class=argparse.ArgumentDefaultsHelpFormatter + ) + + parser.add_argument( + "--world-size", + type=int, + ) + + parser.add_argument( + "--epoch", + type=int, + default=30, + help="""It specifies the checkpoint to use for decoding. + Note: Epoch counts from 1. + You can specify --avg to use more checkpoints for model averaging.""", + ) + + parser.add_argument( + "--iter", + type=int, + default=0, + help="""If positive, --epoch is ignored and it + will use the checkpoint exp_dir/checkpoint-iter.pt. + You can specify --avg to use more checkpoints for model averaging. + """, + ) + + parser.add_argument( + "--avg", + type=int, + default=9, + help="Number of checkpoints to average. Automatically select " + "consecutive checkpoints before the checkpoint specified by " + "'--epoch' and '--iter'", + ) + + parser.add_argument( + "--use-averaged-model", + type=str2bool, + default=True, + help="Whether to load averaged model. Currently it only supports " + "using --epoch. If True, it would decode with the averaged model " + "over the epoch range from `epoch-avg` (excluded) to `epoch`." + "Actually only the models with epoch number of `epoch-avg` and " + "`epoch` are loaded for averaging. ", + ) + + parser.add_argument( + "--exp-dir", + type=str, + default="pruned_transducer_stateless7/exp", + help="The experiment dir", + ) + + parser.add_argument( + "--log-dir", + type=str, + required=True, + help="Where to store the logs", + ) + + parser.add_argument( + "--bpe-model", + type=str, + default="data/lang_bpe_500/bpe.model", + help="Path to the BPE model", + ) + + parser.add_argument( + "--lang-dir", + type=Path, + default="data/lang_bpe_500", + help="The lang dir containing word table and LG graph", + ) + + parser.add_argument( + "--decoding-method", + type=str, + default="greedy_search", + help="""Possible values are: + - greedy_search + - beam_search + - modified_beam_search + - fast_beam_search + - fast_beam_search_nbest + - fast_beam_search_nbest_oracle + - fast_beam_search_nbest_LG + - modified_beam_search_lm_shallow_fusion # for rnn lm shallow fusion + - modified_beam_search_LODR + If you use fast_beam_search_nbest_LG, you have to specify + `--lang-dir`, which should contain `LG.pt`. + """, + ) + + parser.add_argument( + "--beam-size", + type=int, + default=4, + help="""An integer indicating how many candidates we will keep for each + frame. Used only when --decoding-method is beam_search or + modified_beam_search.""", + ) + + parser.add_argument( + "--beam", + type=float, + default=20.0, + help="""A floating point value to calculate the cutoff score during beam + search (i.e., `cutoff = max-score - beam`), which is the same as the + `beam` in Kaldi. + Used only when --decoding-method is fast_beam_search, + fast_beam_search_nbest, fast_beam_search_nbest_LG, + and fast_beam_search_nbest_oracle + """, + ) + + parser.add_argument( + "--ngram-lm-scale", + type=float, + default=0.01, + help=""" + Used only when --decoding_method is fast_beam_search_nbest_LG. + It specifies the scale for n-gram LM scores. + """, + ) + + parser.add_argument( + "--max-contexts", + type=int, + default=8, + help="""Used only when --decoding-method is + fast_beam_search, fast_beam_search_nbest, fast_beam_search_nbest_LG, + and fast_beam_search_nbest_oracle""", + ) + + parser.add_argument( + "--max-states", + type=int, + default=64, + help="""Used only when --decoding-method is + fast_beam_search, fast_beam_search_nbest, fast_beam_search_nbest_LG, + and fast_beam_search_nbest_oracle""", + ) + + parser.add_argument( + "--max-sym-per-frame", + type=int, + default=1, + help="""Maximum number of symbols per frame. + Used only when --decoding_method is greedy_search""", + ) + + parser.add_argument( + "--num-paths", + type=int, + default=200, + help="""Number of paths for nbest decoding. + Used only when the decoding method is fast_beam_search_nbest, + fast_beam_search_nbest_LG, and fast_beam_search_nbest_oracle""", + ) + + parser.add_argument( + "--nbest-scale", + type=float, + default=0.5, + help="""Scale applied to lattice scores when computing nbest paths. + Used only when the decoding method is fast_beam_search_nbest, + fast_beam_search_nbest_LG, and fast_beam_search_nbest_oracle""", + ) + + parser.add_argument( + "--input-manifest", + type=str, + required=True, + help="The input manifest to be decoded", + ) + + parser.add_argument( + "--output-manifest", + type=str, + required=True, + help="Where to store the output manifest (directory)", + ) + + parser.add_argument( + "--use-pre-text", + type=str2bool, + default=True, + help="Use pre-text is available during decoding", + ) + + parser.add_argument( + "--use-style-prompt", + type=str2bool, + default=True, + help="Use style prompt when evaluation", + ) + + parser.add_argument( + "--use-context-embedding", + type=str2bool, + default=False, + help="Use context fuser when evaluation", + ) + + parser.add_argument( + "--post-normalization", + type=str2bool, + default=True, + help="Normalized the recognition results by uppercasing and removing non-alphabetic symbols. ", + ) + + parser.add_argument( + "--compute-CER", + type=str2bool, + default=True, + help="Reports CER. By default, only reports WER", + ) + + parser.add_argument( + "--style-text-transform", + type=str, + choices=["mixed-punc", "upper-no-punc", "lower-no-punc", "lower-punc"], + default="mixed-punc", + help="The style of style prompt, i.e style_text", + ) + + parser.add_argument( + "--pre-text-transform", + type=str, + choices=["mixed-punc", "upper-no-punc", "lower-no-punc", "lower-punc"], + default="mixed-punc", + help="The style of content prompt, i.e pre_text", + ) + + parser.add_argument( + "--use-ls-test-set", + type=str2bool, + default=False, + help="Use librispeech test set for evaluation.", + ) + + parser.add_argument( + "--use-ls-context-list", + type=str2bool, + default=False, + help="If use a fixed context list for LibriSpeech decoding", + ) + + add_model_arguments(parser) + + return parser + + +def _apply_style_transform(text: List[str], transform: str) -> List[str]: + """Apply transform to a list of text. By default, the text are in + ground truth format, i.e mixed-punc. + + Args: + text (List[str]): Input text string + transform (str): Transform to be applied + + Returns: + List[str]: _description_ + """ + if transform == "mixed-punc": + return text + elif transform == "upper-no-punc": + return [upper_only_alpha(s) for s in text] + elif transform == "lower-no-punc": + return [lower_only_alpha(s) for s in text] + elif transform == "lower-punc": + return [lower_all_char(s) for s in text] + else: + raise NotImplementedError(f"Unseen transform: {transform}") + + +def decode_one_batch( + params: AttributeDict, + model: nn.Module, + sp: spm.SentencePieceProcessor, + tokenizer, + batch: dict, + biasing_dict: dict = None, + word_table: Optional[k2.SymbolTable] = None, + decoding_graph: Optional[k2.Fsa] = None, +) -> Dict[str, List[List[str]]]: + """Decode one batch and return the result in a dict. The dict has the + following format: + + - key: It indicates the setting used for decoding. For example, + if greedy_search is used, it would be "greedy_search" + If beam search with a beam size of 7 is used, it would be + "beam_7" + - value: It contains the decoding result. `len(value)` equals to + batch size. `value[i]` is the decoding result for the i-th + utterance in the given batch. + Args: + params: + It's the return value of :func:`get_params`. + model: + The neural model. + sp: + The BPE model. + batch: + It is the return value from iterating + `lhotse.dataset.K2SpeechRecognitionDataset`. See its documentation + for the format of the `batch`. + word_table: + The word symbol table. + decoding_graph: + The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used + only when --decoding_method is fast_beam_search, fast_beam_search_nbest, + fast_beam_search_nbest_oracle, and fast_beam_search_nbest_LG. + LM: + A neural net LM for shallow fusion. Only used when `--use-shallow-fusion` + set to true. + ngram_lm: + A ngram lm. Used in LODR decoding. + ngram_lm_scale: + The scale of the ngram language model. + Returns: + Return the decoding result. See above description for the format of + the returned dict. + """ + device = next(model.parameters()).device + feature = batch["inputs"] + cuts = batch["supervisions"]["cut"] + cut_ids = [c.supervisions[0].id for c in cuts] + batch_size = feature.size(0) + + # get pre_text + if "pre_text" in batch["supervisions"] and params.use_pre_text: + pre_texts = batch["supervisions"][ + "text" + ] # use the ground truth ref text as pre_text + pre_texts = [train_text_normalization(t) for t in pre_texts] + else: + pre_texts = ["" for _ in range(batch_size)] + + if params.use_ls_context_list: + pre_texts = [biasing_dict[id] for id in cut_ids] + + # get style_text + if params.use_style_prompt: + fixed_sentence = "Mixed-case English transcription, with punctuation. Actually, it's fully not related." + style_texts = batch["supervisions"].get( + "style_text", [fixed_sentence for _ in range(batch_size)] + ) + style_texts = [train_text_normalization(t) for t in style_texts] + else: + style_texts = ["" for _ in range(batch_size)] # use empty string + + # Get the text embedding input + if params.use_pre_text or params.use_style_prompt: + + # apply style transform to the pre_text and style_text + pre_texts = _apply_style_transform(pre_texts, params.pre_text_transform) + # pre_texts = random_shuffle_subset(pre_texts, p=1.0, p_mask=0.0) + if params.use_style_prompt: + style_texts = _apply_style_transform( + style_texts, params.style_text_transform + ) + + with warnings.catch_warnings(): + warnings.simplefilter("ignore") + + # Use tokenizer to prepare input for text encoder + encoded_inputs, style_lens = _encode_texts_as_bytes_with_tokenizer( + pre_texts=pre_texts, + style_texts=style_texts, + tokenizer=tokenizer, + device=device, + ) + + memory, memory_key_padding_mask = model.encode_text( + encoded_inputs=encoded_inputs, + style_lens=style_lens, + ) # (T,B,C) + else: + memory = None + memory_key_padding_mask = None + + # Get the transducer encoder output + 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) + + with warnings.catch_warnings(): + warnings.simplefilter("ignore") + encoder_out, encoder_out_lens = model.encode_audio( + feature=feature, + feature_lens=feature_lens, + memory=memory, + memory_key_padding_mask=memory_key_padding_mask, + ) + + hyps = [] + + if params.decoding_method == "greedy_search" and params.max_sym_per_frame == 1: + if memory is None or not params.use_context_embedding: + hyp_tokens = greedy_search_batch( + model=model, + encoder_out=encoder_out, + encoder_out_lens=encoder_out_lens, + ) + else: + memory = memory.permute(1, 0, 2) # (T,N,C) -> (N,T,C) + context = model.context_fuser( + memory, padding_mask=memory_key_padding_mask + ) # (N,C) + context = model.joiner.context_proj(context) # (N,C) + hyp_tokens = greedy_search_batch_with_context( + model=model, + encoder_out=encoder_out, + encoder_out_lens=encoder_out_lens, + context=context, + ) + for hyp in sp.decode(hyp_tokens): + hyps.append(hyp.split()) + elif params.decoding_method == "modified_beam_search": + hyp_tokens = modified_beam_search( + model=model, + encoder_out=encoder_out, + encoder_out_lens=encoder_out_lens, + beam=params.beam_size, + ) + for hyp in sp.decode(hyp_tokens): + hyps.append(hyp.split()) + else: + batch_size = encoder_out.size(0) + + for i in range(batch_size): + # fmt: off + encoder_out_i = encoder_out[i:i+1, :encoder_out_lens[i]] + # fmt: on + if params.decoding_method == "greedy_search": + if memory is None or not params.use_context_embedding: + hyp = greedy_search( + model=model, + encoder_out=encoder_out_i, + max_sym_per_frame=params.max_sym_per_frame, + ) + else: + cur_context = context[i : i + 1, :] + hyp = greedy_search_with_context( + model=model, + encoder_out=encoder_out_i, + context=cur_context, + max_sym_per_frame=params.max_sym_per_frame, + ) + else: + raise ValueError( + f"Unsupported decoding method: {params.decoding_method}" + ) + hyps.append(sp.decode(hyp).split()) + + if params.decoding_method == "greedy_search": + return {"greedy_search": hyps} + else: + return {f"beam_size_{params.beam_size}": hyps} + + +def decode_dataset( + dl: torch.utils.data.DataLoader, + params: AttributeDict, + model: nn.Module, + sp: spm.SentencePieceProcessor, + tokenizer, + biasing_dict: Dict = None, + word_table: Optional[k2.SymbolTable] = None, + decoding_graph: Optional[k2.Fsa] = None, +) -> Dict[str, List[Tuple[str, List[str], List[str]]]]: + """Decode dataset. + + Args: + dl: + PyTorch's dataloader containing the dataset to decode. + params: + It is returned by :func:`get_params`. + model: + The neural model. + sp: + The BPE model. + word_table: + The word symbol table. + decoding_graph: + The decoding graph. Can be either a `k2.trivial_graph` or HLG, Used + only when --decoding_method is fast_beam_search, fast_beam_search_nbest, + fast_beam_search_nbest_oracle, and fast_beam_search_nbest_LG. + LM: + A neural network LM, used during shallow fusion + Returns: + Return a dict, whose key may be "greedy_search" if greedy search + is used, or it may be "beam_7" if beam size of 7 is used. + Its value is a list of tuples. Each tuple contains two elements: + The first is the reference transcript, and the second is the + predicted result. + """ + num_cuts = 0 + + try: + num_batches = len(dl) + except TypeError: + num_batches = "?" + + if params.decoding_method == "greedy_search": + log_interval = 40 + else: + log_interval = 20 + + results = defaultdict(list) + for batch_idx, batch in enumerate(dl): + texts = batch["supervisions"][ + "text" + ] # By default, this should be in mixed-punc format + + # the style of ref_text should match style_text + texts = _apply_style_transform(texts, params.style_text_transform) + if params.use_style_prompt: + texts = _apply_style_transform(texts, params.style_text_transform) + + cut_ids = [cut.id for cut in batch["supervisions"]["cut"]] + + hyps_dict = decode_one_batch( + params=params, + model=model, + sp=sp, + tokenizer=tokenizer, + biasing_dict=biasing_dict, + decoding_graph=decoding_graph, + word_table=word_table, + batch=batch, + ) + + for name, hyps in hyps_dict.items(): + this_batch = [] + assert len(hyps) == len(texts) + for cut_id, hyp_words, ref_text in zip(cut_ids, hyps, texts): + ref_text = ref_text_normalization( + ref_text + ) # remove full-width symbols & some book marks + ref_words = ref_text.split() + this_batch.append((cut_id, ref_words, hyp_words)) + + results[name].extend(this_batch) + + num_cuts += len(texts) + + if batch_idx % log_interval == 0: + batch_str = f"{batch_idx}/{num_batches}" + + logging.info(f"batch {batch_str}, cuts processed until now is {num_cuts}") + return results + + +def save_results( + params: AttributeDict, + test_set_name: str, + results_dict: Dict[str, List[Tuple[str, List[str], List[str]]]], +): + test_set_wers = dict() + test_set_cers = dict() + for key, results in results_dict.items(): + recog_path = params.res_dir / f"recogs-{test_set_name}-{params.suffix}.txt" + results = sorted(results) + store_transcripts(filename=recog_path, texts=results) + logging.info(f"The transcripts are stored in {recog_path}") + + # The following prints out WERs, per-word error statistics and aligned + # ref/hyp pairs. + errs_filename = params.res_dir / f"errs-{test_set_name}-{params.suffix}.txt" + with open(errs_filename, "w") as f: + wer = write_error_stats( + f, f"{test_set_name}-{key}", results, enable_log=True + ) + test_set_wers[key] = wer + + logging.info("Wrote detailed error stats to {}".format(errs_filename)) + + if params.compute_CER: + # Write CER statistics + recog_path = ( + params.res_dir / f"recogs-{test_set_name}-char-{params.suffix}.txt" + ) + store_transcripts(filename=recog_path, texts=results, char_level=True) + errs_filename = ( + params.res_dir / f"errs-CER-{test_set_name}-{params.suffix}.txt" + ) + with open(errs_filename, "w") as f: + cer = write_error_stats( + f, + f"{test_set_name}-{key}", + results, + enable_log=True, + compute_CER=params.compute_CER, + ) + test_set_cers[key] = cer + + logging.info("Wrote detailed CER stats to {}".format(errs_filename)) + + test_set_wers = sorted(test_set_wers.items(), key=lambda x: x[1]) + errs_info = params.res_dir / f"wer-summary-{test_set_name}-{params.suffix}.txt" + with open(errs_info, "w") as f: + print("settings\tWER", file=f) + for key, val in test_set_wers: + print("{}\t{}".format(key, val), file=f) + + s = "\nFor {}, WER of different settings are:\n".format(test_set_name) + note = "\tbest for {}".format(test_set_name) + for key, val in test_set_wers: + s += "{}\t{}{}\n".format(key, val, note) + note = "" + logging.info(s) + + if params.compute_CER: + test_set_cers = sorted(test_set_cers.items(), key=lambda x: x[1]) + errs_info = params.res_dir / f"cer-summary-{test_set_name}-{params.suffix}.txt" + with open(errs_info, "w") as f: + print("settings\tCER", file=f) + for key, val in test_set_cers: + print("{}\t{}".format(key, val), file=f) + + s = "\nFor {}, CER of different settings are:\n".format(test_set_name) + note = "\tbest for {}".format(test_set_name) + for key, val in test_set_cers: + s += "{} CER\t{}{}\n".format(key, val, note) + note = "" + logging.info(s) + + +def add_decoding_result_to_manifest( + in_manifest, + out_manifest: str, + results_dict: Dict, +): + # write the decoding results with prompt to the manifest as an + # extra ref text + new_ans = {} + for key, value in results_dict.items(): + for items in value: + id, ref, hyp = items + new_ans[id] = " ".join(hyp) + + def _add_decoding(c): + key = c.supervisions[0].id + c.supervisions[0].texts.append(new_ans[key]) + return c + + in_manifest = in_manifest.map(_add_decoding) + logging.info(f"Saving manifest to {out_manifest}") + in_manifest.to_file(out_manifest) + + +def main(): + parser = get_parser() + LibriHeavyAsrDataModule.add_arguments(parser) + args = parser.parse_args() + args.exp_dir = Path(args.exp_dir) + + cuts = load_manifest_lazy(args.input_manifest) + + world_size = args.world_size + assert world_size >= 1 + if world_size > 1: + splitted_cuts = cuts.split(num_splits=world_size) + mp.spawn( + run, args=(world_size, args, splitted_cuts), nprocs=world_size, join=True + ) + else: + run(rank=0, world_size=1, args=args, cuts=cuts) + + +@torch.no_grad() +def run(rank, world_size, args, cuts): + params = get_params() + params.update(vars(args)) + params.res_dir = params.exp_dir / params.decoding_method + + if params.iter > 0: + params.suffix = f"iter-{params.iter}-avg-{params.avg}" + else: + params.suffix = f"epoch-{params.epoch}-avg-{params.avg}" + + if params.use_pre_text: + params.suffix += f"-pre-text-{params.pre_text_transform}" + + if params.use_style_prompt: + params.suffix += f"-style-prompt-{params.style_text_transform}" + + params.suffix += f"-{rank}" + + world_size = params.world_size + + params.output_manifest = Path(params.output_manifest) + if world_size > 1: + cuts = cuts[rank] + out_name = params.output_manifest / f"with_decoding_job_{rank}.jsonl.gz" + else: + out_name = params.output_manifest / "with_decoding.jsonl.gz" + + device = torch.device("cpu") + if torch.cuda.is_available(): + device = torch.device("cuda", rank) + + setup_logger(f"{params.log_dir}/log-get-manifest-with-decoding-{rank}") + logging.info("Decoding started") + + logging.info(f"Device: {device}") + + sp = spm.SentencePieceProcessor() + sp.load(params.bpe_model) + + # and are defined in local/train_bpe_model.py + params.blank_id = sp.piece_to_id("") + params.unk_id = sp.piece_to_id("") + params.vocab_size = sp.get_piece_size() + + logging.info(params) + + logging.info("About to create model") + model = get_transducer_model(params) + tokenizer = get_tokenizer(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() + + LM = None + + decoding_graph = None + word_table = None + + num_param = sum([p.numel() for p in model.parameters()]) + logging.info(f"Number of model parameters: {num_param}") + + # we need cut ids to display recognition results. + args.return_cuts = True + libriheavy = LibriHeavyAsrDataModule(args) + + dl = libriheavy.valid_dataloaders( + cuts, text_sampling_func=naive_triplet_text_sampling + ) + + test_sets = ["test"] + test_dl = [dl] + + for test_set, test_dl in zip(test_sets, test_dl): + biasing_dict = None + + results_dict = decode_dataset( + dl=test_dl, + params=params, + model=model, + sp=sp, + tokenizer=tokenizer, + biasing_dict=biasing_dict, + word_table=word_table, + decoding_graph=decoding_graph, + ) + + # save_results( + # params=params, + # test_set_name=test_set, + # results_dict=results_dict, + # ) + + add_decoding_result_to_manifest( + in_manifest=cuts, + out_manifest=out_name, + results_dict=results_dict, + ) + + logging.info("Done!") + + +# torch.set_num_threads(1) +# torch.set_num_interop_threads(1) + +if __name__ == "__main__": + main() diff --git a/egs/libriheavy/ASR/zipformer_prompt_asr/decoder.py b/egs/libriheavy/ASR/zipformer_prompt_asr/decoder.py new file mode 100644 index 000000000..93e0f9f7e --- /dev/null +++ b/egs/libriheavy/ASR/zipformer_prompt_asr/decoder.py @@ -0,0 +1,130 @@ +# 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 +import torch.nn.functional as F +from scaling import Balancer + + +class Decoder(nn.Module): + """This class modifies the stateless decoder from the following paper: + + RNN-transducer with stateless prediction network + https://ieeexplore.ieee.org/stamp/stamp.jsp?arnumber=9054419 + + It removes the recurrent connection from the decoder, i.e., the prediction + network. Different from the above paper, it adds an extra Conv1d + right after the embedding layer. + + TODO: Implement https://arxiv.org/pdf/2109.07513.pdf + """ + + def __init__( + self, + vocab_size: int, + decoder_dim: int, + blank_id: int, + context_size: int, + ): + """ + Args: + vocab_size: + Number of tokens of the modeling unit including blank. + decoder_dim: + Dimension of the input embedding, and of the decoder output. + blank_id: + The ID of the blank symbol. + context_size: + Number of previous words to use to predict the next word. + 1 means bigram; 2 means trigram. n means (n+1)-gram. + """ + super().__init__() + + self.embedding = nn.Embedding( + num_embeddings=vocab_size, + embedding_dim=decoder_dim, + padding_idx=blank_id, + ) + # the balancers are to avoid any drift in the magnitude of the + # embeddings, which would interact badly with parameter averaging. + self.balancer = Balancer( + decoder_dim, + channel_dim=-1, + min_positive=0.0, + max_positive=1.0, + min_abs=0.5, + max_abs=1.0, + prob=0.05, + ) + + self.blank_id = blank_id + + assert context_size >= 1, context_size + self.context_size = context_size + self.vocab_size = vocab_size + + if context_size > 1: + self.conv = nn.Conv1d( + in_channels=decoder_dim, + out_channels=decoder_dim, + kernel_size=context_size, + padding=0, + groups=decoder_dim // 4, # group size == 4 + bias=False, + ) + self.balancer2 = Balancer( + decoder_dim, + channel_dim=-1, + min_positive=0.0, + max_positive=1.0, + min_abs=0.5, + max_abs=1.0, + prob=0.05, + ) + + def forward(self, y: torch.Tensor, need_pad: bool = True) -> torch.Tensor: + """ + Args: + y: + A 2-D tensor of shape (N, U). + need_pad: + True to left pad the input. Should be True during training. + False to not pad the input. Should be False during inference. + Returns: + Return a tensor of shape (N, U, decoder_dim). + """ + y = y.to(torch.int64) + # this stuff about clamp() is a temporary fix for a mismatch + # at utterance start, we use negative ids in beam_search.py + embedding_out = self.embedding(y.clamp(min=0)) * (y >= 0).unsqueeze(-1) + + embedding_out = self.balancer(embedding_out) + + if self.context_size > 1: + embedding_out = embedding_out.permute(0, 2, 1) + if need_pad is True: + embedding_out = F.pad(embedding_out, pad=(self.context_size - 1, 0)) + else: + # During inference time, there is no need to do extra padding + # as we only need one output + assert embedding_out.size(-1) == self.context_size + embedding_out = self.conv(embedding_out) + embedding_out = embedding_out.permute(0, 2, 1) + embedding_out = F.relu(embedding_out) + embedding_out = self.balancer2(embedding_out) + + return embedding_out diff --git a/egs/libriheavy/ASR/zipformer_prompt_asr/encoder_interface.py b/egs/libriheavy/ASR/zipformer_prompt_asr/encoder_interface.py new file mode 100644 index 000000000..257facce4 --- /dev/null +++ b/egs/libriheavy/ASR/zipformer_prompt_asr/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/libriheavy/ASR/zipformer_prompt_asr/export_PromptASR.py b/egs/libriheavy/ASR/zipformer_prompt_asr/export_PromptASR.py new file mode 100644 index 000000000..e0bc556a8 --- /dev/null +++ b/egs/libriheavy/ASR/zipformer_prompt_asr/export_PromptASR.py @@ -0,0 +1,255 @@ +#!/usr/bin/env python3 +# +# Copyright 2021-2023 Xiaomi Corporation (Author: Xiaoyu Yang) +# +# See ../../../../LICENSE for clarification regarding multiple authors +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +# This script converts several saved checkpoints +# to a single one using model averaging. + +""" +Export `model.state_dict()` + +- For non-streaming model: + +./zipformer_prompt_asr/export_PromptASR.py \ + --exp-dir ./zipformer_prompt_asr/exp \ + --tokens data/lang_bpe_500_fallback_coverage_0.99/tokens.txt \ + --epoch 50 \ + --avg 10 + +- For streaming model: + +./zipformer_prompt_asr/export_PromptASR.py \ + --exp-dir ./zipformer_prompt_asr/exp \ + --causal 1 \ + --tokens data/lang_bpe_500_fallback_coverage_0.99/tokens.txt \ + --epoch 50 \ + --avg 10 + +It will generate a file `pretrained.pt` in the given `exp_dir`. You can later +load it by `icefall.checkpoint.load_checkpoint()`. + +""" + +import argparse +import logging +from pathlib import Path +from typing import List, Tuple + +import k2 +import torch +from torch import Tensor, nn +from train_bert_encoder import add_model_arguments, get_params, get_transducer_model + +from icefall.checkpoint import ( + average_checkpoints, + average_checkpoints_with_averaged_model, + find_checkpoints, + load_checkpoint, +) +from icefall.utils import make_pad_mask, num_tokens, 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 1. + You can specify --avg to use more checkpoints for model averaging.""", + ) + + parser.add_argument( + "--iter", + type=int, + default=0, + help="""If positive, --epoch is ignored and it + will use the checkpoint exp_dir/checkpoint-iter.pt. + You can specify --avg to use more checkpoints for model averaging. + """, + ) + + parser.add_argument( + "--avg", + type=int, + default=9, + help="Number of checkpoints to average. Automatically select " + "consecutive checkpoints before the checkpoint specified by " + "'--epoch' and '--iter'", + ) + + parser.add_argument( + "--use-averaged-model", + type=str2bool, + default=True, + help="Whether to load averaged model. Currently it only supports " + "using --epoch. If True, it would decode with the averaged model " + "over the epoch range from `epoch-avg` (excluded) to `epoch`." + "Actually only the models with epoch number of `epoch-avg` and " + "`epoch` are loaded for averaging. ", + ) + + parser.add_argument( + "--exp-dir", + type=str, + default="zipformer/exp", + help="""It specifies the directory where all training related + files, e.g., checkpoints, log, etc, are saved + """, + ) + + parser.add_argument( + "--tokens", + type=str, + default="data/lang_bpe_500/tokens.txt", + help="Path to the tokens.txt", + ) + + parser.add_argument( + "--jit", + type=str2bool, + default=False, + help="""True to save a model after applying torch.jit.script. + It will generate a file named jit_script.pt. + Check ./jit_pretrained.py for how to use it. + """, + ) + + add_model_arguments(parser) + + return parser + + +@torch.no_grad() +def main(): + args = get_parser().parse_args() + args.exp_dir = Path(args.exp_dir) + + params = get_params() + params.update(vars(args)) + + device = torch.device("cpu") + # if torch.cuda.is_available(): + # device = torch.device("cuda", 0) + + logging.info(f"device: {device}") + + token_table = k2.SymbolTable.from_file(params.tokens) + params.blank_id = token_table[""] + params.vocab_size = num_tokens(token_table) + 1 + + logging.info(params) + + logging.info("About to create model") + model = get_transducer_model(params) + + if not params.use_averaged_model: + if params.iter > 0: + filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[ + : params.avg + ] + if len(filenames) == 0: + raise ValueError( + f"No checkpoints found for" + f" --iter {params.iter}, --avg {params.avg}" + ) + elif len(filenames) < params.avg: + raise ValueError( + f"Not enough checkpoints ({len(filenames)}) found for" + f" --iter {params.iter}, --avg {params.avg}" + ) + logging.info(f"averaging {filenames}") + model.load_state_dict(average_checkpoints(filenames, device=device)) + elif params.avg == 1: + load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model) + else: + start = params.epoch - params.avg + 1 + filenames = [] + for i in range(start, params.epoch + 1): + if i >= 1: + filenames.append(f"{params.exp_dir}/epoch-{i}.pt") + logging.info(f"averaging {filenames}") + model.load_state_dict(average_checkpoints(filenames, device=device)) + else: + if params.iter > 0: + filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[ + : params.avg + 1 + ] + if len(filenames) == 0: + raise ValueError( + f"No checkpoints found for" + f" --iter {params.iter}, --avg {params.avg}" + ) + elif len(filenames) < params.avg + 1: + raise ValueError( + f"Not enough checkpoints ({len(filenames)}) found for" + f" --iter {params.iter}, --avg {params.avg}" + ) + filename_start = filenames[-1] + filename_end = filenames[0] + logging.info( + "Calculating the averaged model over iteration checkpoints" + f" from {filename_start} (excluded) to {filename_end}" + ) + model.load_state_dict( + average_checkpoints_with_averaged_model( + filename_start=filename_start, + filename_end=filename_end, + device=device, + ) + ) + elif params.avg == 1: + load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model) + else: + assert params.avg > 0, params.avg + start = params.epoch - params.avg + assert start >= 1, start + filename_start = f"{params.exp_dir}/epoch-{start}.pt" + filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt" + logging.info( + f"Calculating the averaged model over epoch range from " + f"{start} (excluded) to {params.epoch}" + ) + model.load_state_dict( + average_checkpoints_with_averaged_model( + filename_start=filename_start, + filename_end=filename_end, + device=device, + ) + ) + + model.eval() + + assert params.jit is False, "Jit is not supported yet" + + logging.info("Not using torchscript. Export model.state_dict()") + # Save it using a format so that it can be loaded + # by :func:`load_checkpoint` + filename = params.exp_dir / "pretrained.pt" + torch.save({"model": model.state_dict()}, str(filename)) + logging.info(f"Saved to {filename}") + + +if __name__ == "__main__": + formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s" + + logging.basicConfig(format=formatter, level=logging.INFO) + main() diff --git a/egs/libriheavy/ASR/zipformer_prompt_asr/joiner.py b/egs/libriheavy/ASR/zipformer_prompt_asr/joiner.py new file mode 100644 index 000000000..59f822748 --- /dev/null +++ b/egs/libriheavy/ASR/zipformer_prompt_asr/joiner.py @@ -0,0 +1,86 @@ +# Copyright 2023 Xiaomi Corp. (authors: Xiaoyu Yang) +# +# See ../../../../LICENSE for clarification regarding multiple authors +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import torch +import torch.nn as nn +from scaling import ScaledLinear + + +class Joiner(nn.Module): + def __init__( + self, + encoder_dim: int, + decoder_dim: int, + joiner_dim: int, + vocab_size: int, + context_dim: int = 512, + context_injection: bool = False, + ): + super().__init__() + + self.encoder_proj = ScaledLinear(encoder_dim, joiner_dim, initial_scale=0.25) + self.decoder_proj = ScaledLinear(decoder_dim, joiner_dim, initial_scale=0.25) + self.output_linear = nn.Linear(joiner_dim, vocab_size) + if context_injection: + self.context_proj = ScaledLinear( + context_dim, joiner_dim, initial_scale=0.25 + ) + else: + self.context_proj = None + + def forward( + self, + encoder_out: torch.Tensor, + decoder_out: torch.Tensor, + context: torch.Tensor = None, + project_input: bool = True, + ) -> torch.Tensor: + """ + Args: + encoder_out: + Output from the encoder. Its shape is (N, T, s_range, C). + decoder_out: + Output from the decoder. Its shape is (N, T, s_range, C). + context: + An embedding vector representing the previous context information + project_input: + If true, apply input projections encoder_proj and decoder_proj. + If this is false, it is the user's responsibility to do this + manually. + Returns: + Return a tensor of shape (N, T, s_range, C). + """ + assert encoder_out.ndim == decoder_out.ndim == 4 + assert encoder_out.shape[:-1] == decoder_out.shape[:-1] + + if project_input: + if context: + logit = ( + self.encoder_proj(encoder_out) + + self.decoder_proj(decoder_out) + + self.context_proj(context) + ) + else: + logit = self.encoder_proj(encoder_out) + self.decoder_proj(decoder_out) + else: + if context is not None: + logit = encoder_out + decoder_out + context.unsqueeze(1).unsqueeze(1) + else: + logit = encoder_out + decoder_out + + logit = self.output_linear(torch.tanh(logit)) + + return logit diff --git a/egs/libriheavy/ASR/zipformer_prompt_asr/ls_text_normalization.py b/egs/libriheavy/ASR/zipformer_prompt_asr/ls_text_normalization.py new file mode 100644 index 000000000..9a693ca4f --- /dev/null +++ b/egs/libriheavy/ASR/zipformer_prompt_asr/ls_text_normalization.py @@ -0,0 +1,153 @@ +import re + +words = { + 0: "zero", + 1: "one", + 2: "two", + 3: "three", + 4: "four", + 5: "five", + 6: "six", + 7: "seven", + 8: "eight", + 9: "nine", + 10: "ten", + 11: "eleven", + 12: "twelve", + 13: "thirteen", + 14: "fourteen", + 15: "fifteen", + 16: "sixteen", + 17: "seventeen", + 18: "eighteen", + 19: "nineteen", + 20: "twenty", + 30: "thirty", + 40: "forty", + 50: "fifty", + 60: "sixty", + 70: "seventy", + 80: "eighty", + 90: "ninety", +} +ordinal_nums = [ + "zeroth", + "first", + "second", + "third", + "fourth", + "fifth", + "sixth", + "seventh", + "eighth", + "ninth", + "tenth", + "eleventh", + "twelfth", + "thirteenth", + "fourteenth", + "fifteenth", + "sixteenth", + "seventeenth", + "eighteenth", + "nineteenth", + "twentieth", +] + +num_ordinal_dict = {num: ordinal_nums[num] for num in range(21)} + + +def year_to_words(num: int): + assert isinstance(num, int), num + # check if a num is representing a year + if num > 1500 and num < 2000: + return words[num // 100] + " " + num_to_words(num % 100) + elif num == 2000: + return "TWO THOUSAND" + elif num > 2000: + return "TWO THOUSAND AND " + num_to_words(num % 100) + else: + return num_to_words(num) + + +def num_to_words(num: int): + # Return the English words of a integer number + + # If this is a year number + if num > 1500 and num < 2030: + return year_to_words(num) + + if num < 20: + return words[num] + if num < 100: + if num % 10 == 0: + return words[num // 10 * 10] + else: + return words[num // 10 * 10] + " " + words[num % 10] + if num < 1000: + return words[num // 100] + " hundred and " + num_to_words(num % 100) + if num < 1000000: + return num_to_words(num // 1000) + " thousand " + num_to_words(num % 1000) + return num + + +def num_to_ordinal_word(num: int): + + return num_ordinal_dict.get(num, num_to_words(num)).upper() + + +def replace_full_width_symbol(s: str) -> str: + # replace full-width symbol with theri half width counterpart + s = s.replace("“", '"') + s = s.replace("”", '"') + s = s.replace("‘", "'") + s = s.replace("’", "'") + + return s + + +def decoding_normalization(text: str) -> str: + text = replace_full_width_symbol(text) + + # Only keep all alpha-numeric characters, hypen and apostrophe + text = text.replace("-", " ") + text = re.sub(r"[^a-zA-Z0-9\s']+", "", text) + return text + + +def word_normalization(word: str) -> str: + # 1 .Use full word for some abbreviation + # 2. Convert digits to english words + # 3. Convert ordinal number to english words + if word == "MRS": + return "MISSUS" + if word == "MR": + return "MISTER" + if word == "ST": + return "SAINT" + if word == "ECT": + return "ET CETERA" + if word.isnumeric(): + word = num_to_words(int(word)) + return str(word).upper() + # e.g 9TH, 6TH + if word[-2:] == "TH" and word[0].isnumeric(): + return num_to_ordinal_word(int(word[:-2])).upper() + if word[0] == "'": + return word[1:] + + return word + + +def simple_normalization(text: str) -> str: + text = replace_full_width_symbol(text) + text = text.replace("--", " ") + + return text + + +if __name__ == "__main__": + + s = str(1830) + out = word_normalization(s) + print(s, out) diff --git a/egs/libriheavy/ASR/zipformer_prompt_asr/model_baseline.py b/egs/libriheavy/ASR/zipformer_prompt_asr/model_baseline.py new file mode 100644 index 000000000..77b4057c4 --- /dev/null +++ b/egs/libriheavy/ASR/zipformer_prompt_asr/model_baseline.py @@ -0,0 +1,262 @@ +# 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. + + +import random +import warnings +from typing import Optional, Tuple + +import k2 +import torch +import torch.nn as nn +from encoder_interface import EncoderInterface +from scaling import ScaledLinear, penalize_abs_values_gt +from torch import Tensor + +from icefall.utils import add_sos, make_pad_mask + + +class Transducer(nn.Module): + """It implements https://arxiv.org/pdf/1211.3711.pdf + "Sequence Transduction with Recurrent Neural Networks" + """ + + def __init__( + self, + encoder_embed: nn.Module, + encoder: EncoderInterface, + decoder: nn.Module, + joiner: nn.Module, + encoder_dim: int, + decoder_dim: int, + joiner_dim: int, + vocab_size: int, + ): + """ + Args: + encoder_embed: + It is a Convolutional 2D subsampling module. It converts + an input of shape (N, T, idim) to an output of of shape + (N, T', odim), where T' = (T-3)//2-2 = (T-7)//2. + encoder: + It is the transcription network in the paper. Its accepts + two inputs: `x` of (N, T, encoder_dim) and `x_lens` of shape (N,). + It returns two tensors: `logits` of shape (N, T, encoder_dm) and + `logit_lens` of shape (N,). + decoder: + It is the prediction network in the paper. Its input shape + is (N, U) and its output shape is (N, U, decoder_dim). + It should contain one attribute: `blank_id`. + joiner: + It has two inputs with shapes: (N, T, encoder_dim) and (N, U, decoder_dim). + Its output shape is (N, T, U, vocab_size). Note that its output contains + unnormalized probs, i.e., not processed by log-softmax. + """ + super().__init__() + assert isinstance(encoder, EncoderInterface), type(encoder) + assert hasattr(decoder, "blank_id") + + self.encoder_embed = encoder_embed + self.encoder = encoder + self.decoder = decoder + self.joiner = joiner + + self.simple_am_proj = ScaledLinear( + encoder_dim, + vocab_size, + initial_scale=0.25, + ) + self.simple_lm_proj = ScaledLinear( + decoder_dim, + vocab_size, + initial_scale=0.25, + ) + + def forward( + self, + x: torch.Tensor, + x_lens: torch.Tensor, + y: k2.RaggedTensor, + prune_range: int = 5, + am_scale: float = 0.0, + lm_scale: float = 0.0, + ) -> torch.Tensor: + """ + Args: + x: + A 3-D tensor of shape (N, T, C). + x_lens: + A 1-D tensor of shape (N,). It contains the number of frames in `x` + before padding. + x_lens: + A 1-D tensor of shape (N,). It contains the number of frames in `x` + before padding. + text: + A 2-D tensor of integer dtype containing prompt text, of shape (N, T). + It is exptected to contain the style prompt (first) and then the content + prompt. + text_lens: + A 1-D tensor of shape (N,). It contains the number of elements (bytes) + in `text` before padding, which will include the lengths of the + style plus the content prompt. + style_lens: + A 1-D tensor of shape (N,), containing the number of elements (bytes) + within each row of `text` that correspond to the style prompt (these + are expected to come first). + y: + A ragged tensor with 2 axes [utt][label]. It contains labels of each + utterance. + prune_range: + The prune range for rnnt loss, it means how many symbols(context) + we are considering for each frame to compute the loss. + am_scale: + The scale to smooth the loss with am (output of encoder network) + part + lm_scale: + The scale to smooth the loss with lm (output of predictor network) + part + Returns: + Return the transducer loss. + + Note: + Regarding am_scale & lm_scale, it will make the loss-function one of + the form: + lm_scale * lm_probs + am_scale * am_probs + + (1-lm_scale-am_scale) * combined_probs + """ + assert x.ndim == 3, x.shape + assert x_lens.ndim == 1, x_lens.shape + assert y.num_axes == 2, y.num_axes + + assert x.size(0) == x_lens.size(0) == y.dim0 + + x, x_lens = self.encoder_embed(x, x_lens) + + src_key_padding_mask = make_pad_mask(x_lens) + x = x.permute(1, 0, 2) # (N, T, C) -> (T, N, C) + + encoder_out, x_lens = self.encoder( + x, + x_lens, + src_key_padding_mask, + ) + encoder_out = encoder_out.permute(1, 0, 2) # (T, N, C) ->(N, T, C) + + assert torch.all(x_lens > 0) + + # Now for the decoder, i.e., the prediction network + row_splits = y.shape.row_splits(1) + y_lens = row_splits[1:] - row_splits[:-1] + + blank_id = self.decoder.blank_id + sos_y = add_sos(y, sos_id=blank_id) + + # sos_y_padded: [B, S + 1], start with SOS. + sos_y_padded = sos_y.pad(mode="constant", padding_value=blank_id) + + # decoder_out: [B, S + 1, decoder_dim] + decoder_out = self.decoder(sos_y_padded) + + # Note: y does not start with SOS + # y_padded : [B, S] + y_padded = y.pad(mode="constant", padding_value=0) + + y_padded = y_padded.to(torch.int64) + boundary = torch.zeros( + (encoder_out.size(0), 4), + dtype=torch.int64, + device=encoder_out.device, + ) + boundary[:, 2] = y_lens + boundary[:, 3] = x_lens + + lm = self.simple_lm_proj(decoder_out) + am = self.simple_am_proj(encoder_out) + + with torch.cuda.amp.autocast(enabled=False): + simple_loss, (px_grad, py_grad) = k2.rnnt_loss_smoothed( + lm=lm.float(), + am=am.float(), + symbols=y_padded, + termination_symbol=blank_id, + lm_only_scale=lm_scale, + am_only_scale=am_scale, + boundary=boundary, + reduction="sum", + return_grad=True, + ) + + # ranges : [B, T, prune_range] + ranges = k2.get_rnnt_prune_ranges( + px_grad=px_grad, + py_grad=py_grad, + boundary=boundary, + s_range=prune_range, + ) + + # am_pruned : [B, T, prune_range, encoder_dim] + # lm_pruned : [B, T, prune_range, decoder_dim] + am_pruned, lm_pruned = k2.do_rnnt_pruning( + am=self.joiner.encoder_proj(encoder_out), + lm=self.joiner.decoder_proj(decoder_out), + ranges=ranges, + ) + + # logits : [B, T, prune_range, vocab_size] + + # project_input=False since we applied the decoder's input projections + # prior to do_rnnt_pruning (this is an optimization for speed). + logits = self.joiner(am_pruned, lm_pruned, project_input=False) + + with torch.cuda.amp.autocast(enabled=False): + pruned_loss = k2.rnnt_loss_pruned( + logits=logits.float(), + symbols=y_padded, + ranges=ranges, + termination_symbol=blank_id, + boundary=boundary, + reduction="sum", + ) + + return (simple_loss, pruned_loss) + + def encode_audio( + self, + feature: Tensor, + feature_lens: Tensor, + memory: Optional[Tensor] = None, + memory_key_padding_mask: Optional[Tensor] = None, + ) -> Tuple[Tensor, Tensor]: + """Encode the input audio features + + Args: + feature (Tensor): Input audio (N,T,C) + feature_lens (Tensor): Length of input audio (N,) + Returns: + Tuple[Tensor, Tensor]: Encoded acoustic features and length + """ + x, x_lens = self.encoder_embed(feature, feature_lens) + src_key_padding_mask = make_pad_mask(x_lens) + x = x.permute(1, 0, 2) # (N, T, C) -> (T, N, C) + + encoder_out, encoder_out_lens = self.encoder( + x=x, + x_lens=x_lens, + src_key_padding_mask=src_key_padding_mask, + ) + encoder_out = encoder_out.permute(1, 0, 2) # (T, N, C) ->(N, T, C) + + return encoder_out, encoder_out_lens diff --git a/egs/libriheavy/ASR/zipformer_prompt_asr/model_with_BERT.py b/egs/libriheavy/ASR/zipformer_prompt_asr/model_with_BERT.py new file mode 100644 index 000000000..21c7b4fac --- /dev/null +++ b/egs/libriheavy/ASR/zipformer_prompt_asr/model_with_BERT.py @@ -0,0 +1,392 @@ +# 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. + + +import random +import warnings +from typing import Dict, Optional, Tuple + +import k2 +import torch +import torch.nn as nn +from encoder_interface import EncoderInterface +from scaling import ScaledLinear, penalize_abs_values_gt +from torch import Tensor + +from icefall.utils import add_sos, make_pad_mask + + +class PromptedTransducer(nn.Module): + """It implements https://arxiv.org/pdf/1211.3711.pdf + "Sequence Transduction with Recurrent Neural Networks" + """ + + def __init__( + self, + encoder_embed: nn.Module, + encoder: EncoderInterface, + text_encoder: EncoderInterface, + decoder: nn.Module, + joiner: nn.Module, + encoder_dim: int, + decoder_dim: int, + joiner_dim: int, + vocab_size: int, + use_BERT: bool = True, + text_encoder_type: str = "BERT", + text_encoder_adapter: bool = False, + freeze_text_encoder: bool = True, + context_fuser: nn.Module = None, + ): + """ + Args: + encoder_embed: + It is a Convolutional 2D subsampling module. It converts + an input of shape (N, T, idim) to an output of of shape + (N, T', odim), where T' = (T-3)//2-2 = (T-7)//2. + encoder: + It is the transcription network in the paper. Its accepts + two inputs: `x` of (N, T, encoder_dim) and `x_lens` of shape (N,). + It returns two tensors: `logits` of shape (N, T, encoder_dm) and + `logit_lens` of shape (N,). + text_encoder: + This is a encoder that processes text information (e.g content prompt + and style prompt). The input is `x` of (N,T) and `x_lens` of shape (N,). + decoder: + It is the prediction network in the paper. Its input shape + is (N, U) and its output shape is (N, U, decoder_dim). + It should contain one attribute: `blank_id`. + joiner: + It has two inputs with shapes: (N, T, encoder_dim) and (N, U, decoder_dim). + Its output shape is (N, T, U, vocab_size). Note that its output contains + unnormalized probs, i.e., not processed by log-softmax. + text_encoder_type: + The type of the text_encoder. Supported are (BERT, DistilBERT) + context_fuser + A optional module that fuses the embeddings of text encoder. The fused embedding + will be added to the joiner. + """ + super().__init__() + assert isinstance(encoder, EncoderInterface), type(encoder) + assert hasattr(decoder, "blank_id") + + self.encoder_embed = encoder_embed + self.encoder = encoder + self.text_encoder = text_encoder + self.decoder = decoder + self.joiner = joiner + + self.simple_am_proj = ScaledLinear( + encoder_dim, + vocab_size, + initial_scale=0.25, + ) + self.simple_lm_proj = ScaledLinear( + decoder_dim, + vocab_size, + initial_scale=0.25, + ) + + self.use_BERT = use_BERT # if the text encoder is a pre-trained BERT + self.context_fuser = context_fuser + + assert text_encoder_type in ( + "BERT", + "DistilBERT", + "BERT-UNCASED", + ), f"Unseen text_encoder type {text_encoder_type}" + self.text_encoder_dim = ( + self.text_encoder.config.hidden_size + if text_encoder_type in ("BERT", "BERT-UNCASED") + else self.text_encoder.config.dim + ) + self.freeze_text_encoder = freeze_text_encoder + + if text_encoder_adapter: + self.text_encoder_adapter = nn.Sequential( + nn.Linear(self.text_encoder_dim, self.text_encoder_dim, bias=False), + nn.Tanh(), + ) + else: + self.text_encoder_adapter = None + + self.style_prompt_embedding = nn.Parameter( + torch.full((self.text_encoder_dim,), 0.5) + ) + + def forward( + self, + x: torch.Tensor, + x_lens: torch.Tensor, + encoded_inputs: Dict, + style_lens: torch.Tensor, + y: k2.RaggedTensor, + prune_range: int = 5, + am_scale: float = 0.0, + lm_scale: float = 0.0, + use_pre_text: bool = True, + ) -> torch.Tensor: + """ + Args: + x: + A 3-D tensor of shape (N, T, C). + x_lens: + A 1-D tensor of shape (N,). It contains the number of frames in `x` + before padding. + x_lens: + A 1-D tensor of shape (N,). It contains the number of frames in `x` + before padding. + text: + A 2-D tensor of integer dtype containing prompt text, of shape (N, T). + It is exptected to contain the style prompt (first) and then the content + prompt. + text_lens: + A 1-D tensor of shape (N,). It contains the number of elements (bytes) + in `text` before padding, which will include the lengths of the + style plus the content prompt. + style_lens: + A 1-D tensor of shape (N,), containing the number of elements (bytes) + within each row of `text` that correspond to the style prompt (these + are expected to come first). + y: + A ragged tensor with 2 axes [utt][label]. It contains labels of each + utterance. + prune_range: + The prune range for rnnt loss, it means how many symbols(context) + we are considering for each frame to compute the loss. + am_scale: + The scale to smooth the loss with am (output of encoder network) + part + lm_scale: + The scale to smooth the loss with lm (output of predictor network) + part + Returns: + Return the transducer loss. + + Note: + Regarding am_scale & lm_scale, it will make the loss-function one of + the form: + lm_scale * lm_probs + am_scale * am_probs + + (1-lm_scale-am_scale) * combined_probs + """ + if self.freeze_text_encoder: + self.text_encoder.eval() + assert x.ndim == 3, x.shape + assert x_lens.ndim == 1, x_lens.shape + assert y.num_axes == 2, y.num_axes + + assert x.size(0) == x_lens.size(0) == y.dim0 + + x, x_lens = self.encoder_embed(x, x_lens) + + src_key_padding_mask = make_pad_mask(x_lens) + x = x.permute(1, 0, 2) # (N, T, C) -> (T, N, C) + + # freeze the BERT text encoder + + if use_pre_text: + memory, memory_key_padding_mask = self.encode_text( + encoded_inputs, style_lens=style_lens + ) + else: + memory = None + memory_key_padding_mask = None + + encoder_out, x_lens = self.encoder( + x, + x_lens, + src_key_padding_mask, + memory=memory, + memory_key_padding_mask=memory_key_padding_mask, + ) + encoder_out = encoder_out.permute(1, 0, 2) # (T, N, C) ->(N, T, C) + + assert torch.all(x_lens > 0) + + # Now for the decoder, i.e., the prediction network + row_splits = y.shape.row_splits(1) + y_lens = row_splits[1:] - row_splits[:-1] + + blank_id = self.decoder.blank_id + sos_y = add_sos(y, sos_id=blank_id) + + # sos_y_padded: [B, S + 1], start with SOS. + sos_y_padded = sos_y.pad(mode="constant", padding_value=blank_id) + + # decoder_out: [B, S + 1, decoder_dim] + decoder_out = self.decoder(sos_y_padded) + + # Note: y does not start with SOS + # y_padded : [B, S] + y_padded = y.pad(mode="constant", padding_value=0) + + y_padded = y_padded.to(torch.int64) + boundary = torch.zeros( + (encoder_out.size(0), 4), + dtype=torch.int64, + device=encoder_out.device, + ) + boundary[:, 2] = y_lens + boundary[:, 3] = x_lens + + lm = self.simple_lm_proj(decoder_out) + am = self.simple_am_proj(encoder_out) + + with torch.cuda.amp.autocast(enabled=False): + simple_loss, (px_grad, py_grad) = k2.rnnt_loss_smoothed( + lm=lm.float(), + am=am.float(), + symbols=y_padded, + termination_symbol=blank_id, + lm_only_scale=lm_scale, + am_only_scale=am_scale, + boundary=boundary, + reduction="sum", + return_grad=True, + ) + + # ranges : [B, T, prune_range] + ranges = k2.get_rnnt_prune_ranges( + px_grad=px_grad, + py_grad=py_grad, + boundary=boundary, + s_range=prune_range, + ) + + # am_pruned : [B, T, prune_range, encoder_dim] + # lm_pruned : [B, T, prune_range, decoder_dim] + am_pruned, lm_pruned = k2.do_rnnt_pruning( + am=self.joiner.encoder_proj(encoder_out), + lm=self.joiner.decoder_proj(decoder_out), + ranges=ranges, + ) + + # logits : [B, T, prune_range, vocab_size] + + # project_input=False since we applied the decoder's input projections + # prior to do_rnnt_pruning (this is an optimization for speed). + if self.context_fuser is not None and memory is not None: + memory = memory.permute(1, 0, 2) # (T,N,C) -> (N,T,C) + context = self.context_fuser(memory, padding_mask=memory_key_padding_mask) + context = self.joiner.context_proj(context) + else: + context = None + + logits = self.joiner(am_pruned, lm_pruned, context=context, project_input=False) + + with torch.cuda.amp.autocast(enabled=False): + pruned_loss = k2.rnnt_loss_pruned( + logits=logits.float(), + symbols=y_padded, + ranges=ranges, + termination_symbol=blank_id, + boundary=boundary, + reduction="sum", + ) + + return (simple_loss, pruned_loss) + + def _add_style_indicator(self, memory: Tensor, style_lens: Tensor): + """ + Adds to `memory` an indicator that is 1.0 for positions that correspond to + the `style prompt` and 0 elsewhere. The scale can be fixed because the + scale of the embedding vector can adjust to compensate. + + Args: + memory: (memory_len, batch_size, embed_dim) + style_lens: (batch_size,), a vector of lengths of the style prompt. + """ + + (memory_len, batch_size, embed_dim) = memory.shape + + indicator = ( + torch.arange(memory_len, device=memory.device).unsqueeze(-1) < style_lens + ) + indicator = indicator.to(memory.dtype) + + extra_term = torch.zeros_like(memory) + extra_term += indicator.unsqueeze(-1) * self.style_prompt_embedding.expand( + memory_len, batch_size, self.text_encoder_dim + ) + + return memory + extra_term + + def encode_text( + self, + encoded_inputs: Dict, + style_lens: Tensor, + ) -> Tuple[Tensor, Tensor]: + """Get the embeddings of text + + Args: + encoded_inputs: The encoded inputs generated by a tokenizer (Dict) + + Returns: + Tuple[Tensor, Tensor]: Returns the text embeddings encoded by the + text_encoder and the attention mask + """ + text_lens = encoded_inputs.pop("length") # need to use pop to remove this item + + # Freeze the pre-trained text encoder + with torch.no_grad(): + memory = self.text_encoder(**encoded_inputs)["last_hidden_state"] # (B,T,C) + memory = memory.permute(1, 0, 2) + + # Text encoder adapter + if self.text_encoder_adapter is not None: + memory = self.text_encoder_adapter(memory) + + memory = self._add_style_indicator(memory, style_lens) + + memory_key_padding_mask = make_pad_mask(text_lens) + + return memory, memory_key_padding_mask + + def encode_audio( + self, + feature: Tensor, + feature_lens: Tensor, + memory: Optional[Tensor], + memory_key_padding_mask: Optional[Tensor], + ) -> Tuple[Tensor, Tensor]: + """Encode the input audio features + + Args: + feature (Tensor): Input audio (N,T,C) + feature_lens (Tensor): Length of input audio (N,) + memory (Tensor): Embeddings from the text encoder + memory_key_padding_mask (Tensor): _description_ + + Returns: + Tuple[Tensor, Tensor]: _description_ + """ + x, x_lens = self.encoder_embed(feature, feature_lens) + src_key_padding_mask = make_pad_mask(x_lens) + x = x.permute(1, 0, 2) # (N, T, C) -> (T, N, C) + + encoder_out, encoder_out_lens = self.encoder( + x=x, + x_lens=x_lens, + src_key_padding_mask=src_key_padding_mask, + memory=memory, + memory_key_padding_mask=memory_key_padding_mask, + ) + encoder_out = encoder_out.permute(1, 0, 2) # (T, N, C) ->(N, T, C) + + return encoder_out, encoder_out_lens + + +Transducer = PromptedTransducer # for decoding diff --git a/egs/libriheavy/ASR/zipformer_prompt_asr/optim.py b/egs/libriheavy/ASR/zipformer_prompt_asr/optim.py new file mode 100644 index 000000000..a767761eb --- /dev/null +++ b/egs/libriheavy/ASR/zipformer_prompt_asr/optim.py @@ -0,0 +1,1168 @@ +# 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 Dict, List, Optional, Tuple, Union + +import torch +from lhotse.utils import fix_random_seed +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) + Unlike common optimizers, which accept model.parameters() or groups of parameters(), + this optimizer could accept model.named_parameters() or groups of named_parameters(). + See comments of function _get_names_of_parameters for its 4 possible cases. + 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, + ): + + 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, + ) + + # If params only contains parameters or group of parameters, + # i.e when parameter names are not given, + # this flag will be set to False in funciton _get_names_of_parameters. + self.show_dominant_parameters = True + param_groups, parameters_names = self._get_names_of_parameters(params) + super(ScaledAdam, self).__init__(param_groups, defaults) + assert len(self.param_groups) == len(parameters_names) + self.parameters_names = parameters_names + + def _get_names_of_parameters( + self, params_or_named_params + ) -> Tuple[List[Dict], List[List[str]]]: + """ + Args: + params_or_named_params: according to the way ScaledAdam is initialized in train.py, + this argument could be one of following 4 cases, + case 1, a generator of parameter, e.g.: + optimizer = ScaledAdam(model.parameters(), lr=params.base_lr, clipping_scale=3.0) + + case 2, a list of parameter groups with different config, e.g.: + model_param_groups = [ + {'params': model.encoder.parameters(), 'lr': 0.05}, + {'params': model.decoder.parameters(), 'lr': 0.01}, + {'params': model.joiner.parameters(), 'lr': 0.03}, + ] + optimizer = ScaledAdam(model_param_groups, lr=params.base_lr, clipping_scale=3.0) + + case 3, a generator of named_parameter, e.g.: + optimizer = ScaledAdam(model.named_parameters(), lr=params.base_lr, clipping_scale=3.0) + + case 4, a list of named_parameter groups with different config, e.g.: + model_named_param_groups = [ + {'named_params': model.encoder.named_parameters(), 'lr': 0.05}, + {'named_params': model.decoder.named_parameters(), 'lr': 0.01}, + {'named_params': model.joiner.named_parameters(), 'lr': 0.03}, + ] + optimizer = ScaledAdam(model_named_param_groups, lr=params.base_lr, clipping_scale=3.0) + + For case 1 and case 2, input params is used to initialize the underlying torch.optimizer. + For case 3 and case 4, firstly, names and params are extracted from input named_params, + then, these extracted params are used to initialize the underlying torch.optimizer, + and these extracted names are mainly used by function + `_show_gradient_dominating_parameter` + + Returns: + Returns a tuple containing 2 elements: + - `param_groups` with type List[Dict], each Dict element is a parameter group. + An example of `param_groups` could be: + [ + {'params': `one iterable of Parameter`, 'lr': 0.05}, + {'params': `another iterable of Parameter`, 'lr': 0.08}, + {'params': `a third iterable of Parameter`, 'lr': 0.1}, + ] + - `param_gruops_names` with type List[List[str]], + each `List[str]` is for a group['params'] in param_groups, + and each `str` is the name of a parameter. + A dummy name "foo" is related to each parameter, + if input are params without names, i.e. case 1 or case 2. + """ + # variable naming convention in this function: + # p is short for param. + # np is short for named_param. + # p_or_np is short for param_or_named_param. + # cur is short for current. + # group is a dict, e.g. {'params': iterable of parameter, 'lr': 0.05, other fields}. + # groups is a List[group] + + iterable_or_groups = list(params_or_named_params) + if len(iterable_or_groups) == 0: + raise ValueError("optimizer got an empty parameter list") + + # The first value of returned tuple. A list of dicts containing at + # least 'params' as a key. + param_groups = [] + + # The second value of returned tuple, + # a List[List[str]], each sub-List is for a group. + param_groups_names = [] + + if not isinstance(iterable_or_groups[0], dict): + # case 1 or case 3, + # the input is an iterable of parameter or named parameter. + param_iterable_cur_group = [] + param_names_cur_group = [] + for p_or_np in iterable_or_groups: + if isinstance(p_or_np, tuple): + # case 3 + name, param = p_or_np + else: + # case 1 + assert isinstance(p_or_np, torch.Tensor) + param = p_or_np + # Assign a dummy name as a placeholder + name = "foo" + self.show_dominant_parameters = False + param_iterable_cur_group.append(param) + param_names_cur_group.append(name) + param_groups.append({"params": param_iterable_cur_group}) + param_groups_names.append(param_names_cur_group) + else: + # case 2 or case 4 + # the input is groups of parameter or named parameter. + for cur_group in iterable_or_groups: + assert "named_params" in cur_group + name_list = [x[0] for x in cur_group["named_params"]] + p_list = [x[1] for x in cur_group["named_params"]] + del cur_group["named_params"] + cur_group["params"] = p_list + param_groups.append(cur_group) + param_groups_names.append(name_list) + + return param_groups, param_groups_names + + 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 which dominates 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 + # Dummy 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 dominating 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 + + # when the param gets too small, just don't shrink it any further. + scale_step.masked_fill_(is_too_small, 0.0) + + # and ensure the parameter rms after update never exceeds param_max_rms. + # We have to look at the trained model for parameters at or around the + # param_max_rms, because sometimes they can indicate a problem with the + # topology or settings. + scale_step = torch.minimum(scale_step, (param_max_rms - param_rms) / param_rms) + + 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[Union[int, float]] = 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/libriheavy/ASR/zipformer_prompt_asr/pretrained.py b/egs/libriheavy/ASR/zipformer_prompt_asr/pretrained.py new file mode 100644 index 000000000..48fd2612a --- /dev/null +++ b/egs/libriheavy/ASR/zipformer_prompt_asr/pretrained.py @@ -0,0 +1,359 @@ +#!/usr/bin/env python3 +# Copyright 2021-2023 Xiaomi Corp. (authors: Fangjun Kuang, Zengwei Yao) +# +# See ../../../../LICENSE for clarification regarding multiple authors +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +""" +This script loads a checkpoint (`pretrained.pt`) and uses it to decode waves. +You can generate the checkpoint with the following command: + +./zipformer/export_PromptASR.py \ + --exp-dir ./zipformer/exp \ + --tokens data/lang_bpe_500_fallback_coverage_0.99/tokens.txt \ + --epoch 50 \ + --avg 10 + +Utterance level context biasing: + +./zipformer/pretrained.py \ + --checkpoint ./zipformer/exp/pretrained.pt \ + --tokens data/lang_bpe_500_fallback_coverage_0.99/tokens.txt \ + --method modified_beam_search \ + --use-pre-text True \ + --content-prompt "bessy random words hello k2 ASR" \ + --use-style-prompt True \ + librispeech.flac + + +Word level context biasing: + +./zipformer/pretrained.py \ + --checkpoint ./zipformer/exp/pretrained.pt \ + --tokens data/lang_bpe_500_fallback_coverage_0.99/tokens.txt \ + --method modified_beam_search \ + --use-pre-text True \ + --content-prompt "The topic is about horses." \ + --use-style-prompt True \ + test.wav + + +""" + +import argparse +import logging +import math +import warnings +from typing import List + +import k2 +import kaldifeat +import sentencepiece as spm +import torch +import torchaudio +from beam_search import greedy_search_batch, modified_beam_search +from text_normalization import _apply_style_transform, train_text_normalization +from torch.nn.utils.rnn import pad_sequence +from train_bert_encoder import ( + _encode_texts_as_bytes_with_tokenizer, + add_model_arguments, + get_params, + get_tokenizer, + get_transducer_model, +) + +from icefall.utils import make_pad_mask, num_tokens, str2bool + + +def get_parser(): + parser = argparse.ArgumentParser( + formatter_class=argparse.ArgumentDefaultsHelpFormatter + ) + + parser.add_argument( + "--checkpoint", + type=str, + required=True, + help="Path to the checkpoint. " + "The checkpoint is assumed to be saved by " + "icefall.checkpoint.save_checkpoint().", + ) + + parser.add_argument( + "--bpe-model", + type=str, + default="data/lang_bpe_500_fallback_coverage_0.99/bpe.model", + help="""Path to tokens.txt.""", + ) + + parser.add_argument( + "--method", + type=str, + default="greedy_search", + help="""Possible values are: + - greedy_search + - modified_beam_search + - fast_beam_search + """, + ) + + parser.add_argument( + "sound_files", + type=str, + nargs="+", + help="The input sound file(s) to transcribe. " + "Supported formats are those supported by torchaudio.load(). " + "For example, wav and flac are supported. " + "The sample rate has to be 16kHz.", + ) + + parser.add_argument( + "--sample-rate", + type=int, + default=16000, + help="The sample rate of the input sound file", + ) + + parser.add_argument( + "--beam-size", + type=int, + default=4, + help="""An integer indicating how many candidates we will keep for each + frame. Used only when --method is beam_search or + modified_beam_search.""", + ) + + parser.add_argument( + "--max-sym-per-frame", + type=int, + default=1, + help="""Maximum number of symbols per frame. Used only when + --method is greedy_search. + """, + ) + + parser.add_argument( + "--use-pre-text", + type=str2bool, + default=True, + help="Use content prompt during decoding", + ) + + parser.add_argument( + "--use-style-prompt", + type=str2bool, + default=True, + help="Use style prompt during decoding", + ) + + parser.add_argument( + "--pre-text-transform", + type=str, + choices=["mixed-punc", "upper-no-punc", "lower-no-punc", "lower-punc"], + default="mixed-punc", + help="The style of content prompt, i.e pre_text", + ) + + parser.add_argument( + "--style-text-transform", + type=str, + choices=["mixed-punc", "upper-no-punc", "lower-no-punc", "lower-punc"], + default="mixed-punc", + help="The style of style prompt, i.e style_text", + ) + + parser.add_argument( + "--content-prompt", type=str, default="", help="The content prompt for decoding" + ) + + parser.add_argument( + "--style-prompt", + type=str, + default="Mixed-cased English text with punctuations, feel free to change it.", + help="The style prompt for decoding", + ) + + add_model_arguments(parser) + + return parser + + +def read_sound_files( + filenames: List[str], expected_sample_rate: float +) -> List[torch.Tensor]: + """Read a list of sound files into a list 1-D float32 torch tensors. + Args: + filenames: + A list of sound filenames. + expected_sample_rate: + The expected sample rate of the sound files. + Returns: + Return a list of 1-D float32 torch tensors. + """ + ans = [] + for f in filenames: + wave, sample_rate = torchaudio.load(f) + assert ( + sample_rate == expected_sample_rate + ), f"expected sample rate: {expected_sample_rate}. Given: {sample_rate}" + # We use only the first channel + ans.append(wave[0].contiguous()) + return ans + + +@torch.no_grad() +def main(): + parser = get_parser() + args = parser.parse_args() + + params = get_params() + + params.update(vars(args)) + + sp = spm.SentencePieceProcessor() + sp.load(params.bpe_model) + + # is defined in local/train_bpe_model.py + params.blank_id = sp.piece_to_id("") + params.unk_id = sp.piece_to_id("") + params.vocab_size = sp.get_piece_size() + + logging.info(f"{params}") + + device = torch.device("cpu") + if torch.cuda.is_available(): + device = torch.device("cuda", 0) + + logging.info(f"device: {device}") + + if params.causal: + assert ( + "," not in params.chunk_size + ), "chunk_size should be one value in decoding." + assert ( + "," not in params.left_context_frames + ), "left_context_frames should be one value in decoding." + + logging.info("Creating model") + model = get_transducer_model(params) + tokenizer = get_tokenizer(params) # for text encoder + + num_param = sum([p.numel() for p in model.parameters()]) + logging.info(f"Number of model parameters: {num_param}") + + checkpoint = torch.load(args.checkpoint, map_location="cpu") + model.load_state_dict(checkpoint["model"], strict=False) + model.to(device) + model.eval() + + logging.info("Constructing Fbank computer") + opts = kaldifeat.FbankOptions() + opts.device = device + opts.frame_opts.dither = 0 + opts.frame_opts.snip_edges = False + opts.frame_opts.samp_freq = params.sample_rate + opts.mel_opts.num_bins = params.feature_dim + + fbank = kaldifeat.Fbank(opts) + + assert ( + len(params.sound_files) == 1 + ), "Only support decoding one audio at this moment" + logging.info(f"Reading sound files: {params.sound_files}") + waves = read_sound_files( + filenames=params.sound_files, expected_sample_rate=params.sample_rate + ) + waves = [w.to(device) for w in waves] + + logging.info("Decoding started") + features = fbank(waves) + feature_lengths = [f.size(0) for f in features] + + features = pad_sequence(features, batch_first=True, padding_value=math.log(1e-10)) + feature_lengths = torch.tensor(feature_lengths, device=device) + + # encode prompts + if params.use_pre_text: + pre_text = [train_text_normalization(params.content_prompt)] + pre_text = _apply_style_transform(pre_text, params.pre_text_transform) + else: + pre_text = [""] + + if params.use_style_prompt: + style_text = [params.style_prompt] + style_text = _apply_style_transform(style_text, params.style_text_transform) + else: + style_text = [""] + + if params.use_pre_text or params.use_style_prompt: + encoded_inputs, style_lens = _encode_texts_as_bytes_with_tokenizer( + pre_texts=pre_text, + style_texts=style_text, + tokenizer=tokenizer, + device=device, + no_limit=True, + ) + + memory, memory_key_padding_mask = model.encode_text( + encoded_inputs=encoded_inputs, + style_lens=style_lens, + ) # (T,B,C) + else: + memory = None + memory_key_padding_mask = None + + with warnings.catch_warnings(): + warnings.simplefilter("ignore") + encoder_out, encoder_out_lens = model.encode_audio( + feature=features, + feature_lens=feature_lengths, + memory=memory, + memory_key_padding_mask=memory_key_padding_mask, + ) + + hyps = [] + msg = f"Using {params.method}" + logging.info(msg) + + if params.method == "modified_beam_search": + hyp_tokens = modified_beam_search( + model=model, + encoder_out=encoder_out, + encoder_out_lens=encoder_out_lens, + beam=params.beam_size, + ) + hyps.append(sp.decode(hyp_tokens)[0]) + elif params.method == "greedy_search" and params.max_sym_per_frame == 1: + hyp_tokens = greedy_search_batch( + model=model, + encoder_out=encoder_out, + encoder_out_lens=encoder_out_lens, + ) + hyps.append(sp.decode(hyp_tokens)[0]) + else: + raise ValueError(f"Unsupported method: {params.method}") + + s = "\n" + for filename, hyp in zip(params.sound_files, hyps): + s += f"{filename}:\n{hyp}\n\n" + logging.info(s) + + logging.info("Decoding Done") + + +if __name__ == "__main__": + formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s" + + logging.basicConfig(format=formatter, level=logging.INFO) + main() diff --git a/egs/libriheavy/ASR/zipformer_prompt_asr/scaling.py b/egs/libriheavy/ASR/zipformer_prompt_asr/scaling.py new file mode 100644 index 000000000..0e6764ba0 --- /dev/null +++ b/egs/libriheavy/ASR/zipformer_prompt_asr/scaling.py @@ -0,0 +1,1872 @@ +# 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 math +import random +from functools import reduce +from itertools import repeat +from typing import Optional, Tuple, Union + +import k2 +import torch +import torch.nn as nn +import torch.nn.functional as F +from torch import Tensor +from torch.cuda.amp import custom_bwd, custom_fwd +from torch.nn import Embedding as ScaledEmbedding + + +class PiecewiseLinear(object): + """ + Piecewise linear function, from float to float, specified as nonempty list of (x,y) pairs with + the x values in order. x values <[initial x] or >[final x] are map to [initial y], [final y] + respectively. + """ + + def __init__(self, *args): + assert len(args) >= 1 + if len(args) == 1 and isinstance(args[0], PiecewiseLinear): + self.pairs = list(args[0].pairs) + else: + self.pairs = [(float(x), float(y)) for x, y in args] + for (x, y) in self.pairs: + assert isinstance(x, float) or isinstance(x, int) + assert isinstance(y, float) or isinstance(y, int) + + for i in range(len(self.pairs) - 1): + assert self.pairs[i + 1][0] > self.pairs[i][0], self.pairs + + def __str__(self): + # e.g. 'PiecewiseLinear((0., 10.), (100., 0.))' + return f"PiecewiseLinear({str(self.pairs)[1:-1]})" + + def __call__(self, x): + if x <= self.pairs[0][0]: + return self.pairs[0][1] + elif x >= self.pairs[-1][0]: + return self.pairs[-1][1] + else: + cur_x, cur_y = self.pairs[0] + for i in range(1, len(self.pairs)): + next_x, next_y = self.pairs[i] + if x >= cur_x and x <= next_x: + return cur_y + (next_y - cur_y) * (x - cur_x) / (next_x - cur_x) + cur_x, cur_y = next_x, next_y + assert False + + def __mul__(self, alpha): + return PiecewiseLinear(*[(x, y * alpha) for x, y in self.pairs]) + + def __add__(self, x): + if isinstance(x, float) or isinstance(x, int): + return PiecewiseLinear(*[(p[0], p[1] + x) for p in self.pairs]) + s, x = self.get_common_basis(x) + return PiecewiseLinear( + *[(sp[0], sp[1] + xp[1]) for sp, xp in zip(s.pairs, x.pairs)] + ) + + def max(self, x): + if isinstance(x, float) or isinstance(x, int): + x = PiecewiseLinear((0, x)) + s, x = self.get_common_basis(x, include_crossings=True) + return PiecewiseLinear( + *[(sp[0], max(sp[1], xp[1])) for sp, xp in zip(s.pairs, x.pairs)] + ) + + def min(self, x): + if isinstance(x, float) or isinstance(x, int): + x = PiecewiseLinear((0, x)) + s, x = self.get_common_basis(x, include_crossings=True) + return PiecewiseLinear( + *[(sp[0], min(sp[1], xp[1])) for sp, xp in zip(s.pairs, x.pairs)] + ) + + def __eq__(self, other): + return self.pairs == other.pairs + + def get_common_basis(self, p: "PiecewiseLinear", include_crossings: bool = False): + """ + Returns (self_mod, p_mod) which are equivalent piecewise lienar + functions to self and p, but with the same x values. + + p: the other piecewise linear function + include_crossings: if true, include in the x values positions + where the functions indicate by this and p crosss. + """ + assert isinstance(p, PiecewiseLinear) + + # get sorted x-values without repetition. + x_vals = sorted(set([x for x, y in self.pairs] + [x for x, y in p.pairs])) + y_vals1 = [self(x) for x in x_vals] + y_vals2 = [p(x) for x in x_vals] + + if include_crossings: + extra_x_vals = [] + for i in range(len(x_vals) - 1): + if (y_vals1[i] > y_vals2[i]) != (y_vals1[i + 1] > y_vals2[i + 1]): + # if the two lines in this subsegment potentially cross each other.. + diff_cur = abs(y_vals1[i] - y_vals2[i]) + diff_next = abs(y_vals1[i + 1] - y_vals2[i + 1]) + # `pos`, between 0 and 1, gives the relative x position, + # with 0 being x_vals[i] and 1 being x_vals[i+1]. + pos = diff_cur / (diff_cur + diff_next) + extra_x_val = x_vals[i] + pos * (x_vals[i + 1] - x_vals[i]) + extra_x_vals.append(extra_x_val) + if len(extra_x_vals) > 0: + x_vals = sorted(set(x_vals + extra_x_vals)) + y_vals1 = [self(x) for x in x_vals] + y_vals2 = [p(x) for x in x_vals] + return ( + PiecewiseLinear(*zip(x_vals, y_vals1)), + PiecewiseLinear(*zip(x_vals, y_vals2)), + ) + + +class ScheduledFloat(torch.nn.Module): + """ + This object is a torch.nn.Module only because we want it to show up in [top_level module].modules(); + it does not have a working forward() function. You are supposed to cast it to float, as + in, float(parent_module.whatever), and use it as something like a dropout prob. + + It is a floating point value whose value changes depending on the batch count of the + training loop. It is a piecewise linear function where you specifiy the (x,y) pairs + in sorted order on x; x corresponds to the batch index. For batch-index values before the + first x or after the last x, we just use the first or last y value. + + Example: + self.dropout = ScheduledFloat((0.0, 0.2), (4000.0, 0.0), default=0.0) + + `default` is used when self.batch_count is not set or in training or mode or in + torch.jit scripting mode. + """ + + def __init__(self, *args, default: float = 0.0): + super().__init__() + # self.batch_count and self.name will be written to in the training loop. + self.batch_count = None + self.name = None + self.default = default + self.schedule = PiecewiseLinear(*args) + + def extra_repr(self) -> str: + return ( + f"batch_count={self.batch_count}, schedule={str(self.schedule.pairs[1:-1])}" + ) + + def __float__(self): + batch_count = self.batch_count + if batch_count is None or not self.training or torch.jit.is_scripting(): + return float(self.default) + else: + ans = self.schedule(self.batch_count) + if random.random() < 0.0002: + logging.info( + f"ScheduledFloat: name={self.name}, batch_count={self.batch_count}, ans={ans}" + ) + return ans + + def __add__(self, x): + if isinstance(x, float) or isinstance(x, int): + return ScheduledFloat(self.schedule + x, default=self.default) + else: + return ScheduledFloat( + self.schedule + x.schedule, default=self.default + x.default + ) + + def max(self, x): + if isinstance(x, float) or isinstance(x, int): + return ScheduledFloat(self.schedule.max(x), default=self.default) + else: + return ScheduledFloat( + self.schedule.max(x.schedule), default=max(self.default, x.default) + ) + + +FloatLike = Union[float, ScheduledFloat] + + +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 CutoffEstimator: + """ + Estimates cutoffs of an arbitrary numerical quantity such that a specified + proportion of items will be above the cutoff on average. + + p is the proportion of items that should be above the cutoff. + """ + + def __init__(self, p: float): + self.p = p + # total count of items + self.count = 0 + # total count of items that were above the cutoff + self.count_above = 0 + # initial cutoff value + self.cutoff = 0 + + def __call__(self, x: float) -> bool: + """ + Returns true if x is above the cutoff. + """ + ans = x > self.cutoff + self.count += 1 + if ans: + self.count_above += 1 + cur_p = self.count_above / self.count + delta_p = cur_p - self.p + if (delta_p > 0) == ans: + q = abs(delta_p) + self.cutoff = x * q + self.cutoff * (1 - q) + return ans + + +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): + 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 BiasNormFunction(torch.autograd.Function): + # This computes: + # scales = (torch.mean((x - bias) ** 2, keepdim=True)) ** -0.5 * log_scale.exp() + # return (x - bias) * scales + # (after unsqueezing the bias), but it does it in a memory-efficient way so that + # it can just store the returned value (chances are, this will also be needed for + # some other reason, related to the next operation, so we can save memory). + @staticmethod + def forward( + ctx, + x: Tensor, + bias: Tensor, + log_scale: Tensor, + channel_dim: int, + store_output_for_backprop: bool, + ) -> Tensor: + assert bias.ndim == 1 + if channel_dim < 0: + channel_dim = channel_dim + x.ndim + ctx.store_output_for_backprop = store_output_for_backprop + ctx.channel_dim = channel_dim + for _ in range(channel_dim + 1, x.ndim): + bias = bias.unsqueeze(-1) + scales = ( + torch.mean((x - bias) ** 2, dim=channel_dim, keepdim=True) ** -0.5 + ) * log_scale.exp() + ans = x * scales + ctx.save_for_backward( + ans.detach() if store_output_for_backprop else x, + scales.detach(), + bias.detach(), + log_scale.detach(), + ) + return ans + + @staticmethod + def backward(ctx, ans_grad: Tensor) -> Tensor: + ans_or_x, scales, bias, log_scale = ctx.saved_tensors + if ctx.store_output_for_backprop: + x = ans_or_x / scales + else: + x = ans_or_x + x = x.detach() + x.requires_grad = True + bias.requires_grad = True + log_scale.requires_grad = True + with torch.enable_grad(): + # recompute scales from x, bias and log_scale. + scales = ( + torch.mean((x - bias) ** 2, dim=ctx.channel_dim, keepdim=True) ** -0.5 + ) * log_scale.exp() + ans = x * scales + ans.backward(gradient=ans_grad) + return x.grad, bias.grad.flatten(), log_scale.grad, None, None + + +class BiasNorm(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. + + Instead, we give the BiasNorm a trainable bias that it can use when + computing the scale for normalization. We also give it a (scalar) + trainable scale on the output. + + + 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}. + log_scale: the initial log-scale that we multiply the output by; this + is learnable. + log_scale_min: FloatLike, minimum allowed value of log_scale + log_scale_max: FloatLike, maximum allowed value of log_scale + store_output_for_backprop: only possibly affects memory use; recommend + to set to True if you think the output of this module is more likely + than the input of this module to be required to be stored for the + backprop. + """ + + def __init__( + self, + num_channels: int, + channel_dim: int = -1, # CAUTION: see documentation. + log_scale: float = 1.0, + log_scale_min: float = -1.5, + log_scale_max: float = 1.5, + store_output_for_backprop: bool = False, + ) -> None: + super(BiasNorm, self).__init__() + self.num_channels = num_channels + self.channel_dim = channel_dim + self.log_scale = nn.Parameter(torch.tensor(log_scale)) + self.bias = nn.Parameter(torch.zeros(num_channels)) + + self.log_scale_min = log_scale_min + self.log_scale_max = log_scale_max + + self.store_output_for_backprop = store_output_for_backprop + + def forward(self, x: Tensor) -> Tensor: + assert x.shape[self.channel_dim] == self.num_channels + + if torch.jit.is_scripting(): + channel_dim = self.channel_dim + if channel_dim < 0: + channel_dim += x.ndim + bias = self.bias + for _ in range(channel_dim + 1, x.ndim): + bias = bias.unsqueeze(-1) + scales = ( + torch.mean((x - bias) ** 2, dim=channel_dim, keepdim=True) ** -0.5 + ) * self.log_scale.exp() + return x * scales + + log_scale = limit_param_value( + self.log_scale, + min=float(self.log_scale_min), + max=float(self.log_scale_max), + training=self.training, + ) + + return BiasNormFunction.apply( + x, self.bias, log_scale, self.channel_dim, self.store_output_for_backprop + ) + + +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 + + +def ScaledConv2d(*args, initial_scale: float = 1.0, **kwargs) -> nn.Conv2d: + """ + Behaves like a constructor of a modified version of nn.Conv2d + 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, but: + NO PADDING-RELATED ARGS. + + 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.Conv2d(*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 ChunkCausalDepthwiseConv1d(torch.nn.Module): + """ + Behaves like a depthwise 1d convolution, except that it is causal in + a chunkwise way, as if we had a block-triangular attention mask. + The chunk size is provided at test time (it should probably be + kept in sync with the attention mask). + + This has a little more than twice the parameters of a conventional + depthwise conv1d module: we implement it by having one + depthwise convolution, of half the width, that is causal (via + right-padding); and one depthwise convolution that is applied only + within chunks, that we multiply by a scaling factor which depends + on the position within the chunk. + + 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. + """ + + def __init__( + self, + channels: int, + kernel_size: int, + initial_scale: float = 1.0, + bias: bool = True, + ): + super().__init__() + assert kernel_size % 2 == 1 + + half_kernel_size = (kernel_size + 1) // 2 + # will pad manually, on one side. + self.causal_conv = nn.Conv1d( + in_channels=channels, + out_channels=channels, + groups=channels, + kernel_size=half_kernel_size, + padding=0, + bias=True, + ) + + self.chunkwise_conv = nn.Conv1d( + in_channels=channels, + out_channels=channels, + groups=channels, + kernel_size=kernel_size, + padding=kernel_size // 2, + bias=bias, + ) + + # first row is correction factors added to the scale near the left edge of the chunk, + # second row is correction factors added to the scale near the right edge of the chunk, + # both of these are added to a default scale of 1.0. + self.chunkwise_conv_scale = nn.Parameter(torch.zeros(2, channels, kernel_size)) + self.kernel_size = kernel_size + + with torch.no_grad(): + self.causal_conv.weight[:] *= initial_scale + self.chunkwise_conv.weight[:] *= initial_scale + if bias: + torch.nn.init.uniform_( + self.causal_conv.bias, -0.1 * initial_scale, 0.1 * initial_scale + ) + + def forward(self, x: Tensor, chunk_size: int = -1) -> Tensor: + """ + Forward function. Args: + x: a Tensor of shape (batch_size, channels, seq_len) + chunk_size: the chunk size, in frames; does not have to divide seq_len exactly. + """ + (batch_size, num_channels, seq_len) = x.shape + + half_kernel_size = self.kernel_size + 1 // 2 + # left_pad is half_kernel_size - 1 where half_kernel_size is the size used + # in the causal conv. It's the amount by which we must pad on the left, + # to make the convolution causal. + left_pad = self.kernel_size // 2 + + if chunk_size < 0 or chunk_size > seq_len: + chunk_size = seq_len + right_pad = -seq_len % chunk_size + + x = torch.nn.functional.pad(x, (left_pad, right_pad)) + + x_causal = self.causal_conv(x[..., : left_pad + seq_len]) + assert x_causal.shape == (batch_size, num_channels, seq_len) + + x_chunk = x[..., left_pad:] + num_chunks = x_chunk.shape[2] // chunk_size + x_chunk = x_chunk.reshape(batch_size, num_channels, num_chunks, chunk_size) + x_chunk = x_chunk.permute(0, 2, 1, 3).reshape( + batch_size * num_chunks, num_channels, chunk_size + ) + x_chunk = self.chunkwise_conv(x_chunk) # does not change shape + + chunk_scale = self._get_chunk_scale(chunk_size) + + x_chunk = x_chunk * chunk_scale + x_chunk = x_chunk.reshape( + batch_size, num_chunks, num_channels, chunk_size + ).permute(0, 2, 1, 3) + x_chunk = x_chunk.reshape(batch_size, num_channels, num_chunks * chunk_size)[ + ..., :seq_len + ] + + return x_chunk + x_causal + + def _get_chunk_scale(self, chunk_size: int): + """Returns tensor of shape (num_channels, chunk_size) that will be used to + scale the output of self.chunkwise_conv.""" + left_edge = self.chunkwise_conv_scale[0] + right_edge = self.chunkwise_conv_scale[1] + if chunk_size < self.kernel_size: + left_edge = left_edge[:, :chunk_size] + right_edge = right_edge[:, -chunk_size:] + else: + t = chunk_size - self.kernel_size + channels = left_edge.shape[0] + pad = torch.zeros( + channels, t, device=left_edge.device, dtype=left_edge.dtype + ) + left_edge = torch.cat((left_edge, pad), dim=-1) + right_edge = torch.cat((pad, right_edge), dim=-1) + return 1.0 + (left_edge + right_edge) + + +class BalancerFunction(torch.autograd.Function): + @staticmethod + def forward( + ctx, + x: Tensor, + min_mean: float, + max_mean: float, + min_rms: float, + max_rms: float, + grad_scale: float, + channel_dim: int, + ) -> Tensor: + if channel_dim < 0: + channel_dim += x.ndim + ctx.channel_dim = channel_dim + ctx.save_for_backward(x) + ctx.config = (min_mean, max_mean, min_rms, max_rms, grad_scale, channel_dim) + return x + + @staticmethod + def backward(ctx, x_grad: Tensor) -> Tuple[Tensor, None, None, None, None, None]: + (x,) = ctx.saved_tensors + (min_mean, max_mean, min_rms, max_rms, grad_scale, channel_dim) = ctx.config + + try: + with torch.enable_grad(): + with torch.cuda.amp.autocast(enabled=False): + x = x.to(torch.float32) + x = x.detach() + x.requires_grad = True + mean_dims = [i for i in range(x.ndim) if i != channel_dim] + uncentered_var = (x**2).mean(dim=mean_dims, keepdim=True) + mean = x.mean(dim=mean_dims, keepdim=True) + stddev = (uncentered_var - (mean * mean)).clamp(min=1.0e-20).sqrt() + rms = uncentered_var.clamp(min=1.0e-20).sqrt() + + m = mean / stddev + # part of loss that relates to mean / stddev + m_loss = (m - m.clamp(min=min_mean, max=max_mean)).abs() + + # put a much larger scale on the RMS-max-limit loss, so that if both it and the + # m_loss are violated we fix the RMS loss first. + rms_clamped = rms.clamp(min=min_rms, max=max_rms) + r_loss = (rms_clamped / rms).log().abs() + + loss = m_loss + r_loss + + loss.backward(gradient=torch.ones_like(loss)) + loss_grad = x.grad + loss_grad_rms = ( + (loss_grad**2) + .mean(dim=mean_dims, keepdim=True) + .sqrt() + .clamp(min=1.0e-20) + ) + + loss_grad = loss_grad * (grad_scale / loss_grad_rms) + + x_grad_float = x_grad.to(torch.float32) + # scale each element of loss_grad by the absolute value of the corresponding + # element of x_grad, which we view as a noisy estimate of its magnitude for that + # (frame and dimension). later we can consider factored versions. + x_grad_mod = x_grad_float + (x_grad_float.abs() * loss_grad) + x_grad = x_grad_mod.to(x_grad.dtype) + except Exception as e: + logging.info( + f"Caught exception in Balancer backward: {e}, size={list(x_grad.shape)}, will continue." + ) + + return x_grad, None, None, None, None, None, None + + +class Balancer(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. + 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. + 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. + """ + + def __init__( + self, + num_channels: int, + channel_dim: int, + min_positive: FloatLike = 0.05, + max_positive: FloatLike = 0.95, + min_abs: FloatLike = 0.2, + max_abs: FloatLike = 100.0, + grad_scale: FloatLike = 0.04, + prob: Optional[FloatLike] = None, + ): + super().__init__() + + if prob is None: + prob = ScheduledFloat((0.0, 0.5), (8000.0, 0.125), default=0.4) + self.prob = prob + # 5% of the time we will return and do nothing because memory usage is + # too high. + self.mem_cutoff = CutoffEstimator(0.05) + + # actually self.num_channels is no longer needed except for an assertion. + self.num_channels = num_channels + self.channel_dim = channel_dim + self.min_positive = min_positive + self.max_positive = max_positive + self.min_abs = min_abs + self.max_abs = max_abs + self.grad_scale = grad_scale + + def forward(self, x: Tensor) -> Tensor: + if ( + torch.jit.is_scripting() + or not x.requires_grad + or (x.is_cuda and self.mem_cutoff(torch.cuda.memory_allocated())) + ): + return _no_op(x) + + prob = float(self.prob) + if random.random() < prob: + # The following inner-functions convert from the way we historically specified + # these limitations, as limits on the absolute value and the proportion of positive + # values, to limits on the RMS value and the (mean / stddev). + def _abs_to_rms(x): + # for normally distributed data, if the expected absolute value is x, the + # expected rms value will be sqrt(pi/2) * x. + return 1.25331413732 * x + + def _proportion_positive_to_mean(x): + def _atanh(x): + eps = 1.0e-10 + # eps is to prevent crashes if x is exactly 0 or 1. + # we'll just end up returning a fairly large value. + return (math.log(1 + x + eps) - math.log(1 - x + eps)) / 2.0 + + def _approx_inverse_erf(x): + # 1 / (sqrt(pi) * ln(2)), + # see https://math.stackexchange.com/questions/321569/approximating-the-error-function-erf-by-analytical-functions + # this approximation is extremely crude and gets progressively worse for + # x very close to -1 or +1, but we mostly care about the "middle" region + # e.g. _approx_inverse_erf(0.05) = 0.0407316414078772, + # and math.erf(0.0407316414078772) = 0.045935330944660666, + # which is pretty close to 0.05. + return 0.8139535143 * _atanh(x) + + # first convert x from the range 0..1 to the range -1..1 which the error + # function returns + x = -1 + (2 * x) + return _approx_inverse_erf(x) + + min_mean = _proportion_positive_to_mean(float(self.min_positive)) + max_mean = _proportion_positive_to_mean(float(self.max_positive)) + min_rms = _abs_to_rms(float(self.min_abs)) + max_rms = _abs_to_rms(float(self.max_abs)) + grad_scale = float(self.grad_scale) + + assert x.shape[self.channel_dim] == self.num_channels + + return BalancerFunction.apply( + x, min_mean, max_mean, min_rms, max_rms, grad_scale, self.channel_dim + ) + else: + return _no_op(x) + + +def penalize_abs_values_gt( + x: Tensor, limit: float, penalty: float, name: str = None +) -> 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. + + The name is for randomly printed debug info. + """ + 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, name) + # 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, module: nn.Module) -> Tensor: + ctx.save_for_backward(x) + ctx.module = module + return x + + @staticmethod + def backward(ctx, x_grad: Tensor): + (x_orig,) = ctx.saved_tensors + w = ctx.module + + try: + 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, w.num_groups) + + if random.random() < 0.005 or __name__ == "__main__": + logging.info( + f"Whitening: name={w.name}, num_groups={w.num_groups}, num_channels={x_orig.shape[-1]}, " + f"metric={metric.item():.2f} vs. limit={float(w.whitening_limit)}" + ) + + if metric < float(w.whitening_limit): + w.prob = w.min_prob + return x_grad, None + else: + w.prob = w.max_prob + metric.backward() + penalty_grad = x_detached.grad + scale = w.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 + except Exception as e: + logging.info( + f"Caught exception in Whiten backward: {e}, size={list(x_grad.shape)}, will continue." + ) + return x_grad, None + + +class Whiten(nn.Module): + def __init__( + self, + num_groups: int, + whitening_limit: FloatLike, + prob: Union[float, Tuple[float, float]], + grad_scale: FloatLike, + ): + """ + 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 float(whitening_limit) >= 1 + assert grad_scale >= 0 + self.num_groups = num_groups + self.whitening_limit = whitening_limit + self.grad_scale = grad_scale + + if isinstance(prob, float): + prob = (prob, prob) + (self.min_prob, self.max_prob) = prob + assert 0 < self.min_prob <= self.max_prob <= 1 + self.prob = self.max_prob + self.name = None # will be set in training loop + + 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. + """ + grad_scale = float(self.grad_scale) + if not x.requires_grad or random.random() > self.prob or grad_scale == 0: + return _no_op(x) + else: + return WhiteningPenaltyFunction.apply(x, self) + + +class WithLoss(torch.autograd.Function): + @staticmethod + def forward(ctx, x: Tensor, y: Tensor, name: str): + ctx.y_shape = y.shape + if random.random() < 0.002 and name is not None: + loss_sum = y.sum().item() + logging.info(f"WithLoss: name={name}, loss-sum={loss_sum:.3e}") + 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), + None, + ) + + +def with_loss(x, y, name): + # returns x but adds y.sum() to the loss function. + return WithLoss.apply(x, y, name) + + +class ScaleGradFunction(torch.autograd.Function): + @staticmethod + def forward(ctx, x: Tensor, alpha: float) -> Tensor: + ctx.alpha = alpha + return x + + @staticmethod + def backward(ctx, grad: Tensor): + return grad * ctx.alpha, None + + +def scale_grad(x: Tensor, alpha: float): + return ScaleGradFunction.apply(x, alpha) + + +class ScaleGrad(nn.Module): + def __init__(self, alpha: float): + super().__init__() + self.alpha = alpha + + def forward(self, x: Tensor) -> Tensor: + return scale_grad(x, self.alpha) + + +class LimitParamValue(torch.autograd.Function): + @staticmethod + def forward(ctx, x: Tensor, min: float, max: float): + ctx.save_for_backward(x) + assert max >= min + ctx.min = min + ctx.max = max + return x + + @staticmethod + def backward(ctx, x_grad: Tensor): + (x,) = ctx.saved_tensors + # where x < ctx.min, ensure all grads are negative (this will tend to make + # x more positive). + x_grad = x_grad * torch.where( + torch.logical_and(x_grad > 0, x < ctx.min), -1.0, 1.0 + ) + # where x > ctx.max, ensure all grads are positive (this will tend to make + # x more negative). + x_grad *= torch.where(torch.logical_and(x_grad < 0, x > ctx.max), -1.0, 1.0) + return x_grad, None, None + + +def limit_param_value( + x: Tensor, min: float, max: float, prob: float = 0.6, training: bool = True +): + # You apply this to (typically) an nn.Parameter during training to ensure that its + # (elements mostly) stays within a supplied range. This is done by modifying the + # gradients in backprop. + # It's not necessary to do this on every batch: do it only some of the time, + # to save a little time. + if training and random.random() < prob: + return LimitParamValue.apply(x, min, max) + else: + return x + + +def _no_op(x: Tensor) -> Tensor: + if torch.jit.is_scripting(): + 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 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 -0.044 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 = -0.044 + 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 __init__(self): + super().__init__() + + 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(): + return x * torch.sigmoid(x - 1.0) + return DoubleSwishFunction.apply(x) + + +# Dropout2 is just like normal dropout, except it supports schedules on the dropout rates. +class Dropout2(nn.Module): + def __init__(self, p: FloatLike): + super().__init__() + self.p = p + + def forward(self, x: Tensor) -> Tensor: + return torch.nn.functional.dropout(x, p=float(self.p), training=self.training) + + +class MulForDropout3(torch.autograd.Function): + # returns (x * y * alpha) where alpha is a float and y doesn't require + # grad and is zero-or-one. + @staticmethod + @custom_fwd + def forward(ctx, x, y, alpha): + assert not y.requires_grad + ans = x * y * alpha + ctx.save_for_backward(ans) + ctx.alpha = alpha + return ans + + @staticmethod + @custom_bwd + def backward(ctx, ans_grad): + (ans,) = ctx.saved_tensors + x_grad = ctx.alpha * ans_grad * (ans != 0) + return x_grad, None, None + + +# Dropout3 is just like normal dropout, except it supports schedules on the dropout rates, +# and it lets you choose one dimension to share the dropout mask over +class Dropout3(nn.Module): + def __init__(self, p: FloatLike, shared_dim: int): + super().__init__() + self.p = p + self.shared_dim = shared_dim + + def forward(self, x: Tensor) -> Tensor: + p = float(self.p) + if not self.training or p == 0: + return _no_op(x) + scale = 1.0 / (1 - p) + rand_shape = list(x.shape) + rand_shape[self.shared_dim] = 1 + mask = torch.rand(*rand_shape, device=x.device) > p + ans = MulForDropout3.apply(x, mask, scale) + return ans + + +class SwooshLFunction(torch.autograd.Function): + """ + swoosh(x) = log(1 + exp(x-4)) - 0.08*x - 0.035 + """ + + @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) + + zero = torch.tensor(0.0, dtype=x.dtype, device=x.device) + + coeff = -0.08 + + with torch.cuda.amp.autocast(enabled=False): + with torch.enable_grad(): + x = x.detach() + x.requires_grad = True + y = torch.logaddexp(zero, x - 4.0) + coeff * x - 0.035 + + if not requires_grad: + return y + y.backward(gradient=torch.ones_like(y)) + + grad = x.grad + floor = coeff + ceil = 1.0 + coeff + 0.005 + + d_scaled = (grad - floor) * (255.0 / (ceil - floor)) + torch.rand_like( + grad + ) + 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. + + coeff = -0.08 + floor = coeff + ceil = 1.0 + coeff + 0.005 + d = d * ((ceil - floor) / 255.0) + floor + return y_grad * d + + +class SwooshL(torch.nn.Module): + def forward(self, x: Tensor) -> Tensor: + """Return Swoosh-L activation.""" + if torch.jit.is_scripting(): + zero = torch.tensor(0.0, dtype=x.dtype, device=x.device) + return torch.logaddexp(zero, x - 4.0) - 0.08 * x - 0.035 + if not x.requires_grad: + return k2.swoosh_l_forward(x) + else: + return k2.swoosh_l(x) + # return SwooshLFunction.apply(x) + + +class SwooshRFunction(torch.autograd.Function): + """ + swoosh(x) = log(1 + exp(x-1)) - 0.08*x - 0.313261687 + + derivatives are between -0.08 and 0.92. + + """ + + @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) + + zero = torch.tensor(0.0, dtype=x.dtype, device=x.device) + + with torch.cuda.amp.autocast(enabled=False): + with torch.enable_grad(): + x = x.detach() + x.requires_grad = True + y = torch.logaddexp(zero, x - 1.0) - 0.08 * x - 0.313261687 + + if not requires_grad: + return y + y.backward(gradient=torch.ones_like(y)) + + grad = x.grad + floor = -0.08 + ceil = 0.925 + + d_scaled = (grad - floor) * (255.0 / (ceil - floor)) + torch.rand_like( + grad + ) + 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 = -0.08 + ceil = 0.925 + d = d * ((ceil - floor) / 255.0) + floor + return y_grad * d + + +class SwooshR(torch.nn.Module): + def forward(self, x: Tensor) -> Tensor: + """Return Swoosh-R activation.""" + if torch.jit.is_scripting(): + zero = torch.tensor(0.0, dtype=x.dtype, device=x.device) + return torch.logaddexp(zero, x - 1.0) - 0.08 * x - 0.313261687 + if not x.requires_grad: + return k2.swoosh_r_forward(x) + else: + return k2.swoosh_r(x) + # return SwooshRFunction.apply(x) + + +# simple version of SwooshL that does not redefine the backprop, used in +# ActivationDropoutAndLinearFunction. +def SwooshLForward(x: Tensor): + x_offset = x - 4.0 + log_sum = (1.0 + x_offset.exp()).log().to(x.dtype) + log_sum = torch.where(log_sum == float("inf"), x_offset, log_sum) + return log_sum - 0.08 * x - 0.035 + + +# simple version of SwooshR that does not redefine the backprop, used in +# ActivationDropoutAndLinearFunction. +def SwooshRForward(x: Tensor): + x_offset = x - 1.0 + log_sum = (1.0 + x_offset.exp()).log().to(x.dtype) + log_sum = torch.where(log_sum == float("inf"), x_offset, log_sum) + return log_sum - 0.08 * x - 0.313261687 + + +class ActivationDropoutAndLinearFunction(torch.autograd.Function): + @staticmethod + @custom_fwd + def forward( + ctx, + x: Tensor, + weight: Tensor, + bias: Optional[Tensor], + activation: str, + dropout_p: float, + dropout_shared_dim: Optional[int], + ): + if dropout_p != 0.0: + dropout_shape = list(x.shape) + if dropout_shared_dim is not None: + dropout_shape[dropout_shared_dim] = 1 + # else it won't be very memory efficient. + dropout_mask = (1.0 / (1.0 - dropout_p)) * ( + torch.rand(*dropout_shape, device=x.device, dtype=x.dtype) > dropout_p + ) + else: + dropout_mask = None + + ctx.save_for_backward(x, weight, bias, dropout_mask) + + ctx.activation = activation + + forward_activation_dict = { + "SwooshL": k2.swoosh_l_forward, + "SwooshR": k2.swoosh_r_forward, + } + # it will raise a KeyError if this fails. This will be an error. We let it + # propagate to the user. + activation_func = forward_activation_dict[activation] + x = activation_func(x) + if dropout_mask is not None: + x = x * dropout_mask + x = torch.nn.functional.linear(x, weight, bias) + return x + + @staticmethod + @custom_bwd + def backward(ctx, ans_grad: Tensor): + saved = ctx.saved_tensors + (x, weight, bias, dropout_mask) = saved + + forward_and_deriv_activation_dict = { + "SwooshL": k2.swoosh_l_forward_and_deriv, + "SwooshR": k2.swoosh_r_forward_and_deriv, + } + # the following lines a KeyError if the activation is unrecognized. + # This will be an error. We let it propagate to the user. + func = forward_and_deriv_activation_dict[ctx.activation] + + y, func_deriv = func(x) + if dropout_mask is not None: + y = y * dropout_mask + # now compute derivative of y w.r.t. weight and bias.. + # y: (..., in_channels), ans_grad: (..., out_channels), + (out_channels, in_channels) = weight.shape + + in_channels = y.shape[-1] + g = ans_grad.reshape(-1, out_channels) + weight_deriv = torch.matmul(g.t(), y.reshape(-1, in_channels)) + y_deriv = torch.matmul(ans_grad, weight) + bias_deriv = None if bias is None else g.sum(dim=0) + x_deriv = y_deriv * func_deriv + if dropout_mask is not None: + # order versus func_deriv does not matter + x_deriv = x_deriv * dropout_mask + + return x_deriv, weight_deriv, bias_deriv, None, None, None + + +class ActivationDropoutAndLinear(torch.nn.Module): + """ + This merges an activation function followed by dropout and then a nn.Linear module; + it does so in a memory efficient way so that it only stores the input to the whole + module. If activation == SwooshL and dropout_shared_dim != None, this will be + equivalent to: + nn.Sequential(SwooshL(), + Dropout3(dropout_p, shared_dim=dropout_shared_dim), + ScaledLinear(in_channels, out_channels, bias=bias, + initial_scale=initial_scale)) + If dropout_shared_dim is None, the dropout would be equivalent to + Dropout2(dropout_p). Note: Dropout3 will be more memory efficient as the dropout + mask is smaller. + + Args: + in_channels: number of input channels, e.g. 256 + out_channels: number of output channels, e.g. 256 + bias: if true, have a bias + activation: the activation function, for now just support SwooshL. + dropout_p: the dropout probability or schedule (happens after nonlinearity). + dropout_shared_dim: the dimension, if any, across which the dropout mask is + shared (e.g. the time dimension). If None, this may be less memory + efficient if there are modules before this one that cache the input + for their backprop (e.g. Balancer or Whiten). + """ + + def __init__( + self, + in_channels: int, + out_channels: int, + bias: bool = True, + activation: str = "SwooshL", + dropout_p: FloatLike = 0.0, + dropout_shared_dim: Optional[int] = -1, + initial_scale: float = 1.0, + ): + super().__init__() + # create a temporary module of nn.Linear that we'll steal the + # weights and bias from + layer = ScaledLinear( + in_channels, out_channels, bias=bias, initial_scale=initial_scale + ) + + self.weight = layer.weight + # register_parameter properly handles making it a parameter when l.bias + # is None. I think there is some reason for doing it this way rather + # than just setting it to None but I don't know what it is, maybe + # something to do with exporting the module.. + self.register_parameter("bias", layer.bias) + + self.activation = activation + self.dropout_p = dropout_p + self.dropout_shared_dim = dropout_shared_dim + + def forward(self, x: Tensor): + if torch.jit.is_scripting(): + if self.activation == "SwooshL": + x = SwooshLForward(x) + elif self.activation == "SwooshR": + x = SwooshRForward(x) + else: + assert False, self.activation + return torch.nn.functional.linear(x, self.weight, self.bias) + + return ActivationDropoutAndLinearFunction.apply( + x, + self.weight, + self.bias, + self.activation, + float(self.dropout_p), + self.dropout_shared_dim, + ) + + +class ClipGradFunction(torch.autograd.Function): + @staticmethod + def forward(ctx, x: Tensor, limit: float): + ctx.limit = limit + return x + + @staticmethod + def backward(ctx, x_grad, *args): + return x_grad.clamp(-ctx.limit, ctx.limit), None + + +def clip_grad(x: Tensor, limit: float): + return ClipGradFunction.apply(x, limit) + + +class AbsValuePenalizer(nn.Module): + """ + This module adds a penalty to the loss function when ever the absolute value of + any element of the input tensor exceeds a certain limit. + """ + + def __init__(self, limit: float, prob: float = 0.1, penalty: float = 1.0e-04): + super().__init__() + self.limit = limit + self.penalty = penalty + + self.prob = prob + self.name = None # will be set in training loop + + # 20% of the time we will return and do nothing because memory usage is + # too high. + self.mem_cutoff = CutoffEstimator(0.2) + + def forward(self, x: Tensor) -> Tensor: + if ( + torch.jit.is_scripting() + or not x.requires_grad + or not self.training + or random.random() > self.prob + ): + # or (x.is_cuda and self.mem_cutoff(torch.cuda.memory_allocated())) + return _no_op(x) # the _no_op op is to make our diagnostics code work. + + x = penalize_abs_values_gt( + x, limit=self.limit, penalty=self.penalty, name=self.name + ) + return x + + +def convert_num_channels(x: Tensor, num_channels: int) -> Tensor: + if num_channels <= x.shape[-1]: + return x[..., :num_channels] + else: + shape = list(x.shape) + shape[-1] = num_channels - shape[-1] + zeros = torch.zeros(*shape, dtype=x.dtype, device=x.device) + return torch.cat((x, zeros), dim=-1) + + +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_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 = Balancer( + probs.numel(), + channel_dim=0, + min_positive=0.05, + max_positive=0.95, + min_abs=0.0, + prob=1.0, + ) + + y_grad = torch.sign(torch.randn(probs.numel(), N)) + + y = m(x) + y.backward(gradient=y_grad) + print("_test_balancer_sign: x = ", x) + print("_test_balancer_sign: y grad = ", y_grad) + print("_test_balancer_sign: x grad = ", x.grad) + + +def _test_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 = Balancer( + magnitudes.numel(), + channel_dim=0, + min_positive=0.0, + max_positive=1.0, + min_abs=0.2, + max_abs=0.7, + prob=1.0, + ) + + y_grad = torch.sign(torch.randn(magnitudes.numel(), N)) + + y = m(x) + y.backward(gradient=y_grad) + print("_test_balancer_magnitude: x = ", x) + print("_test_balancer_magnitude: y grad = ", y_grad) + print("_test_balancer_magnitude: x grad = ", x.grad) + + +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_swooshl_deriv(): + x = torch.randn(10, 12, dtype=torch.double) * 3.0 + x.requires_grad = True + m = SwooshL() + + tol = 1.0 / 255.0 + torch.autograd.gradcheck(m, x, atol=tol, eps=0.01) + + # for self-test. + x = torch.randn(1000, 1000, dtype=torch.double) * 3.0 + x.requires_grad = True + y = m(x) + + +def _test_swooshr_deriv(): + x = torch.randn(10, 12, dtype=torch.double) * 3.0 + x.requires_grad = True + m = SwooshR() + + tol = 1.0 / 255.0 + torch.autograd.gradcheck(m, x, atol=tol, eps=0.01) + + # 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) + + +def _test_piecewise_linear(): + p = PiecewiseLinear((0, 10.0)) + for x in [-100, 0, 100]: + assert p(x) == 10.0 + p = PiecewiseLinear((0, 10.0), (1, 0.0)) + for x, y in [(-100, 10.0), (0, 10.0), (0.5, 5.0), (1, 0.0), (2, 0.0)]: + print("x, y = ", x, y) + assert p(x) == y, (x, p(x), y) + + q = PiecewiseLinear((0.5, 15.0), (0.6, 1.0)) + x_vals = [-1.0, 0.0, 0.1, 0.2, 0.5, 0.6, 0.7, 0.9, 1.0, 2.0] + pq = p.max(q) + for x in x_vals: + y1 = max(p(x), q(x)) + y2 = pq(x) + assert abs(y1 - y2) < 0.001 + pq = p.min(q) + for x in x_vals: + y1 = min(p(x), q(x)) + y2 = pq(x) + assert abs(y1 - y2) < 0.001 + pq = p + q + for x in x_vals: + y1 = p(x) + q(x) + y2 = pq(x) + assert abs(y1 - y2) < 0.001 + + +def _test_activation_dropout_and_linear(): + in_channels = 20 + out_channels = 30 + + for bias in [True, False]: + # actually we don't test for dropout_p != 0.0 because forward functions will give + # different answers. This is because we are using the k2 implementation of + # swoosh_l an swoosh_r inside SwooshL() and SwooshR(), and they call randn() + # internally, messing up the random state. + for dropout_p in [0.0]: + for activation in ["SwooshL", "SwooshR"]: + m1 = nn.Sequential( + SwooshL() if activation == "SwooshL" else SwooshR(), + Dropout3(p=dropout_p, shared_dim=-1), + ScaledLinear( + in_channels, out_channels, bias=bias, initial_scale=0.5 + ), + ) + m2 = ActivationDropoutAndLinear( + in_channels, + out_channels, + bias=bias, + initial_scale=0.5, + activation=activation, + dropout_p=dropout_p, + ) + with torch.no_grad(): + m2.weight[:] = m1[2].weight + if bias: + m2.bias[:] = m1[2].bias + # make sure forward gives same result. + x1 = torch.randn(10, in_channels) + x1.requires_grad = True + + # TEMP. + assert torch.allclose( + SwooshRFunction.apply(x1), SwooshRForward(x1), atol=1.0e-03 + ) + + x2 = x1.clone().detach() + x2.requires_grad = True + seed = 10 + torch.manual_seed(seed) + y1 = m1(x1) + y_grad = torch.randn_like(y1) + y1.backward(gradient=y_grad) + torch.manual_seed(seed) + y2 = m2(x2) + y2.backward(gradient=y_grad) + + print( + f"bias = {bias}, dropout_p = {dropout_p}, activation = {activation}" + ) + print("y1 = ", y1) + print("y2 = ", y2) + assert torch.allclose(y1, y2, atol=0.02) + assert torch.allclose(m1[2].weight.grad, m2.weight.grad, atol=1.0e-05) + if bias: + assert torch.allclose(m1[2].bias.grad, m2.bias.grad, atol=1.0e-05) + print("x1.grad = ", x1.grad) + print("x2.grad = ", x2.grad) + + def isclose(a, b): + # return true if cosine similarity is > 0.9. + return (a * b).sum() > 0.9 * ( + (a**2).sum() * (b**2).sum() + ).sqrt() + + # the SwooshL() implementation has a noisy gradient due to 1-byte + # storage of it. + assert isclose(x1.grad, x2.grad) + + +if __name__ == "__main__": + logging.getLogger().setLevel(logging.INFO) + torch.set_num_threads(1) + torch.set_num_interop_threads(1) + _test_piecewise_linear() + _test_softmax() + _test_whiten() + _test_balancer_sign() + _test_balancer_magnitude() + _test_swooshl_deriv() + _test_swooshr_deriv() + _test_activation_dropout_and_linear() + _test_double_swish_deriv() diff --git a/egs/libriheavy/ASR/zipformer_prompt_asr/subsampling.py b/egs/libriheavy/ASR/zipformer_prompt_asr/subsampling.py new file mode 100644 index 000000000..7acbc1808 --- /dev/null +++ b/egs/libriheavy/ASR/zipformer_prompt_asr/subsampling.py @@ -0,0 +1,276 @@ +#!/usr/bin/env python3 +# Copyright 2023 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 warnings +from typing import Tuple + +import torch +from scaling import ( + Balancer, + BiasNorm, + Dropout3, + FloatLike, + Optional, + ScaledConv2d, + ScaleGrad, + ScheduledFloat, + SwooshL, + SwooshR, + Whiten, +) +from torch import Tensor, nn + + +class ConvNeXt(nn.Module): + """ + Our interpretation of the ConvNeXt module as used in https://arxiv.org/pdf/2206.14747.pdf + """ + + def __init__( + self, + channels: int, + hidden_ratio: int = 3, + kernel_size: Tuple[int, int] = (7, 7), + layerdrop_rate: FloatLike = None, + ): + super().__init__() + padding = ((kernel_size[0] - 1) // 2, (kernel_size[1] - 1) // 2) + hidden_channels = channels * hidden_ratio + if layerdrop_rate is None: + layerdrop_rate = ScheduledFloat((0.0, 0.2), (20000.0, 0.015)) + self.layerdrop_rate = layerdrop_rate + + self.depthwise_conv = nn.Conv2d( + in_channels=channels, + out_channels=channels, + groups=channels, + kernel_size=kernel_size, + padding=padding, + ) + + self.pointwise_conv1 = nn.Conv2d( + in_channels=channels, out_channels=hidden_channels, kernel_size=1 + ) + + self.hidden_balancer = Balancer( + hidden_channels, + channel_dim=1, + min_positive=0.3, + max_positive=1.0, + min_abs=0.75, + max_abs=5.0, + ) + + self.activation = SwooshL() + self.pointwise_conv2 = ScaledConv2d( + in_channels=hidden_channels, + out_channels=channels, + kernel_size=1, + initial_scale=0.01, + ) + + self.out_balancer = Balancer( + channels, + channel_dim=1, + min_positive=0.4, + max_positive=0.6, + min_abs=1.0, + max_abs=6.0, + ) + self.out_whiten = Whiten( + num_groups=1, + whitening_limit=5.0, + prob=(0.025, 0.25), + grad_scale=0.01, + ) + + def forward(self, x: Tensor) -> Tensor: + if torch.jit.is_scripting() or not self.training: + return self.forward_internal(x) + layerdrop_rate = float(self.layerdrop_rate) + + if layerdrop_rate != 0.0: + batch_size = x.shape[0] + mask = ( + torch.rand((batch_size, 1, 1, 1), dtype=x.dtype, device=x.device) + > layerdrop_rate + ) + else: + mask = None + # turns out this caching idea does not work with --world-size > 1 + # return caching_eval(self.forward_internal, x, mask) + return self.forward_internal(x, mask) + + def forward_internal( + self, x: Tensor, layer_skip_mask: Optional[Tensor] = None + ) -> Tensor: + """ + x layout: (N, C, H, W), i.e. (batch_size, num_channels, num_frames, num_freqs) + + The returned value has the same shape as x. + """ + bypass = x + x = self.depthwise_conv(x) + x = self.pointwise_conv1(x) + x = self.hidden_balancer(x) + x = self.activation(x) + x = self.pointwise_conv2(x) + + if layer_skip_mask is not None: + x = x * layer_skip_mask + + x = bypass + x + x = self.out_balancer(x) + x = x.transpose(1, 3) # (N, W, H, C); need channel dim to be last + x = self.out_whiten(x) + x = x.transpose(1, 3) # (N, C, H, W) + + return x + + +class Conv2dSubsampling(nn.Module): + """Convolutional 2D subsampling (to 1/2 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: FloatLike = 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-3)//2, out_channels) + layer1_channels: + Number of channels in layer1 + layer1_channels: + Number of channels in layer2 + bottleneck: + bottleneck dimension for 1d squeeze-excite + """ + assert in_channels >= 7 + super().__init__() + + # The ScaleGrad module is there to prevent the gradients + # w.r.t. the weight or bias of the first Conv2d module in self.conv from + # exceeding the range of fp16 when using automatic mixed precision (amp) + # training. (The second one is necessary to stop its bias from getting + # a too-large gradient). + + self.conv = nn.Sequential( + nn.Conv2d( + in_channels=1, + out_channels=layer1_channels, + kernel_size=3, + padding=(0, 1), # (time, freq) + ), + ScaleGrad(0.2), + Balancer(layer1_channels, channel_dim=1, max_abs=1.0), + SwooshR(), + nn.Conv2d( + in_channels=layer1_channels, + out_channels=layer2_channels, + kernel_size=3, + stride=2, + padding=0, + ), + Balancer(layer2_channels, channel_dim=1, max_abs=4.0), + SwooshR(), + nn.Conv2d( + in_channels=layer2_channels, + out_channels=layer3_channels, + kernel_size=3, + stride=(1, 2), # (time, freq) + ), + Balancer(layer3_channels, channel_dim=1, max_abs=4.0), + SwooshR(), + ) + + # just one convnext layer + self.convnext = ConvNeXt(layer3_channels, kernel_size=(7, 7)) + + out_width = (((in_channels - 1) // 2) - 1) // 2 + + self.out = nn.Linear(out_width * layer3_channels, out_channels) + # use a larger than normal grad_scale on this whitening module; there is + # only one such module, so there is not a concern about adding together + # many copies of this extra gradient term. + self.out_whiten = Whiten( + num_groups=1, + whitening_limit=ScheduledFloat((0.0, 4.0), (20000.0, 8.0), default=4.0), + prob=(0.025, 0.25), + grad_scale=0.02, + ) + + # max_log_eps=0.0 is to prevent both eps and the output of self.out from + # getting large, there is an unnecessary degree of freedom. + self.out_norm = BiasNorm(out_channels) + self.dropout = Dropout3(dropout, shared_dim=1) + + def forward(self, x: torch.Tensor, x_lens: torch.Tensor) -> torch.Tensor: + """Subsample x. + + Args: + x: + Its shape is (N, T, idim). + x_lens: + A tensor of shape (batch_size,) containing the number of frames in + + Returns: + - a tensor of shape (N, ((T-1)//2 - 1)//2, odim) + - output lengths, of shape (batch_size,) + """ + # On entry, x is (N, T, idim) + x = x.unsqueeze(1) # (N, T, idim) -> (N, 1, T, idim) i.e., (N, C, H, W) + # scaling x by 0.1 allows us to use a larger grad-scale in fp16 "amp" (automatic mixed precision) + # training, since the weights in the first convolution are otherwise the limiting factor for getting infinite + # gradients. + x = self.conv(x) + x = self.convnext(x) + + # Now x is of shape (N, odim, ((T-3)//2 - 1)//2, ((idim-1)//2 - 1)//2) + b, c, t, f = x.size() + + x = x.transpose(1, 2).reshape(b, t, c * f) + # now x: (N, ((T-1)//2 - 1))//2, out_width * layer3_channels)) + + x = self.out(x) + # Now x is of shape (N, ((T-1)//2 - 1))//2, odim) + x = self.out_whiten(x) + x = self.out_norm(x) + x = self.dropout(x) + + with warnings.catch_warnings(): + warnings.simplefilter("ignore") + x_lens = (x_lens - 7) // 2 + assert x.size(1) == x_lens.max().item() + + return x, x_lens diff --git a/egs/libriheavy/ASR/zipformer_prompt_asr/test_model.py b/egs/libriheavy/ASR/zipformer_prompt_asr/test_model.py new file mode 100755 index 000000000..13483637d --- /dev/null +++ b/egs/libriheavy/ASR/zipformer_prompt_asr/test_model.py @@ -0,0 +1,119 @@ +#!/usr/bin/env python3 +# Copyright 2022 Xiaomi Corp. (authors: Fangjun Kuang) +# +# See ../../../../LICENSE for clarification regarding multiple authors +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + + +""" +To run this file, do: + + cd icefall/egs/librispeech/ASR + python ./pruned_transducer_stateless4/test_model.py +""" + +from scaling import ScheduledFloat +from train_subformer import get_params, get_text_encoder, get_transducer_model +from zipformer import Zipformer2 + + +def test_model_1(): + params = get_params() + params.vocab_size = 500 + params.blank_id = 0 + params.context_size = 2 + params.num_encoder_layers = 24 + params.dim_feedforward = 1536 # 384 * 4 + params.encoder_dim = 384 + model = get_transducer_model(params) + num_param = sum([p.numel() for p in model.parameters()]) + print(f"Number of model parameters: {num_param}") + + +# See Table 1 from https://arxiv.org/pdf/2005.08100.pdf +def test_model_M(): + params = get_params() + params.vocab_size = 500 + params.blank_id = 0 + params.context_size = 2 + params.num_encoder_layers = "2,4,3,2,4" + params.feedforward_dims = "1024,1024,2048,2048,1024" + params.nhead = "8,8,8,8,8" + params.encoder_dims = "384,384,384,384,384" + params.attention_dims = "192,192,192,192,192" + params.encoder_unmasked_dims = "256,256,256,256,256" + params.zipformer_downsampling_factors = "1,2,4,8,2" + params.cnn_module_kernels = "31,31,15,15" + + params.text_encoder_dim = (192, 192, 256, 384) + params.decoder_dim = 512 + params.joiner_dim = 512 + model = Zipformer2( + output_downsampling_factor=8, + downsampling_factor=(1, 2, 4, 8), + num_encoder_layers=(2, 4, 4, 4), + encoder_dim=(192, 192, 256, 384), + encoder_unmasked_dim=(192, 192, 256, 256), + query_head_dim=(32, 32, 32, 32), + pos_head_dim=(4, 4, 4, 4), + value_head_dim=(12, 12, 12, 12), + pos_dim=48, + num_heads=(4, 4, 4, 8), + feedforward_dim=( + 384, + 512, + 768, + 1024, + ), # could increase this if there is nough data + cnn_module_kernel=(31, 31, 15, 15), + dropout=ScheduledFloat((0.0, 0.3), (20000.0, 0.1)), + warmup_batches=4000.0, + causal=False, + ) + num_param = sum([p.numel() for p in model.parameters()]) + print(f"Number of model parameters: {num_param}") + + model = Zipformer2( + output_downsampling_factor=8, + downsampling_factor=(1, 2, 4, 8), + num_encoder_layers=(2, 4, 6, 6), + encoder_dim=(256, 256, 384, 512), + encoder_unmasked_dim=(196, 196, 256, 256), + query_head_dim=(32, 32, 32, 32), + pos_head_dim=(4, 4, 4, 4), + value_head_dim=(12, 12, 12, 12), + pos_dim=48, + num_heads=(4, 4, 4, 8), + feedforward_dim=( + 384, + 512, + 768, + 1024, + ), # could increase this if there is nough data + cnn_module_kernel=(31, 31, 15, 15), + dropout=ScheduledFloat((0.0, 0.3), (20000.0, 0.1)), + warmup_batches=4000.0, + causal=False, + ) + num_param = sum([p.numel() for p in model.parameters()]) + print(f"Number of model parameters: {num_param}") + + +def main(): + # test_model_1() + test_model_M() + + +if __name__ == "__main__": + main() diff --git a/egs/libriheavy/ASR/zipformer_prompt_asr/text_normalization.py b/egs/libriheavy/ASR/zipformer_prompt_asr/text_normalization.py new file mode 100644 index 000000000..efb4acc3c --- /dev/null +++ b/egs/libriheavy/ASR/zipformer_prompt_asr/text_normalization.py @@ -0,0 +1,101 @@ +# Copyright 2023 Xiaomi Corp. (authors: Xiaoyu Yang) +# +# See ../LICENSE for clarification regarding multiple authors +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import re +from typing import List + + +def train_text_normalization(s: str) -> str: + # replace full-width with half-width + s = s.replace("“", '"') + s = s.replace("”", '"') + s = s.replace("‘", "'") + s = s.replace("’", "'") + if s[:2] == '" ': # remove the starting double quote + s = s[2:] + + return s + + +def ref_text_normalization(ref_text: str) -> str: + # Rule 1: Remove the [FN#[]] + p = r"[FN#[0-9]*]" + pattern = re.compile(p) + + res = pattern.findall(ref_text) + ref_text = re.sub(p, "", ref_text) + + ref_text = train_text_normalization(ref_text) + + return ref_text + + +def remove_non_alphabetic(text: str, strict: bool = True) -> str: + # Recommend to set strict to False + if not strict: + # Note, this also keeps space, single quote(') and hypen (-) + text = text.replace("-", " ") + text = text.replace("—", " ") + return re.sub(r"[^a-zA-Z0-9\s']+", "", text) + else: + # only keeps space + return re.sub(r"[^a-zA-Z\s]+", "", text) + + +def upper_only_alpha(text: str) -> str: + return remove_non_alphabetic(text.upper(), strict=False) + + +def lower_only_alpha(text: str) -> str: + return remove_non_alphabetic(text.lower(), strict=False) + + +def lower_all_char(text: str) -> str: + return text.lower() + + +def upper_all_char(text: str) -> str: + return text.upper() + + +def _apply_style_transform(text: List[str], transform: str) -> List[str]: + """Apply transform to a list of text. By default, the text are in + ground truth format, i.e mixed-punc. + + Args: + text (List[str]): Input text string + transform (str): Transform to be applied + + Returns: + List[str]: _description_ + """ + if transform == "mixed-punc": + return text + elif transform == "upper-no-punc": + return [upper_only_alpha(s) for s in text] + elif transform == "lower-no-punc": + return [lower_only_alpha(s) for s in text] + elif transform == "lower-punc": + return [lower_all_char(s) for s in text] + else: + raise NotImplementedError(f"Unseen transform: {transform}") + + +if __name__ == "__main__": + ref_text = "Mixed-case English transcription, with punctuation. Actually, it is fully not related." + print(ref_text) + res = upper_only_alpha(ref_text) + print(res) diff --git a/egs/libriheavy/ASR/zipformer_prompt_asr/train_baseline.py b/egs/libriheavy/ASR/zipformer_prompt_asr/train_baseline.py new file mode 100644 index 000000000..7075c9154 --- /dev/null +++ b/egs/libriheavy/ASR/zipformer_prompt_asr/train_baseline.py @@ -0,0 +1,1390 @@ +#!/usr/bin/env python3 +# Copyright 2021-2022 Xiaomi Corp. (authors: Fangjun Kuang, +# Wei Kang, +# Mingshuang Luo,) +# Zengwei Yao) +# +# See ../../../../LICENSE for clarification regarding multiple authors +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +""" +Usage: + + +# For mix precision training: + +export CUDA_VISIBLE_DEVICES="0,1,2,3" + +./zipformer/train.py \ + --world-size 4 \ + --num-epochs 30 \ + --start-epoch 1 \ + --use-fp16 1 \ + --exp-dir zipformer/exp \ + --max-duration 1000 + +# To train a streaming model + +./zipformer/train.py \ + --world-size 4 \ + --num-epochs 30 \ + --start-epoch 1 \ + --use-fp16 1 \ + --causal 1 + --exp-dir zipformer/exp \ + --max-duration 1000 + +""" + + +import argparse +import copy +import logging +import random +import warnings +from pathlib import Path +from shutil import copyfile +from typing import Any, Dict, List, Optional, Tuple, Union + +import k2 +import optim +import sentencepiece as spm +import torch +import torch.multiprocessing as mp +import torch.nn as nn +from asr_datamodule import LibriHeavyAsrDataModule +from decoder import Decoder +from joiner import Joiner +from lhotse.cut import Cut +from lhotse.dataset.sampling.base import CutSampler +from lhotse.utils import fix_random_seed +from model_baseline import Transducer +from optim import Eden, ScaledAdam +from scaling import ScheduledFloat +from subsampling import Conv2dSubsampling +from text_normalization import train_text_normalization, upper_only_alpha +from torch import Tensor +from torch.cuda.amp import GradScaler +from torch.nn.parallel import DistributedDataParallel as DDP +from torch.utils.tensorboard import SummaryWriter +from zipformer import Zipformer2 + +from icefall import diagnostics +from icefall.checkpoint import load_checkpoint, remove_checkpoints +from icefall.checkpoint import save_checkpoint as save_checkpoint_impl +from icefall.checkpoint import ( + save_checkpoint_with_global_batch_idx, + update_averaged_model, +) +from icefall.dist import cleanup_dist, setup_dist +from icefall.env import get_env_info +from icefall.hooks import register_inf_check_hooks +from icefall.utils import ( + AttributeDict, + MetricsTracker, + get_parameter_groups_with_lrs, + setup_logger, + str2bool, +) + +LRSchedulerType = Union[torch.optim.lr_scheduler._LRScheduler, optim.LRScheduler] + + +def get_first( + texts: List[str], + pre_texts: List[str], + context_list: Optional[str] = None, + rare_word_list: Optional[List[str]] = None, +) -> str: + # Always get the first one, which is the mixed-cased text with punc + out = {"text": texts[0], "pre_text": pre_texts[0]} + return out + + +def get_upper_only_alpha( + texts: List[str], + pre_texts: List[str], + context_list: Optional[str] = None, + rare_word_list: Optional[List[str]] = None, +) -> str: + # Always get the first one, which is the mixed-cased text with punc, + # but with upper case it and remove punctuation + out = { + "text": upper_only_alpha(texts[0]), + "pre_text": upper_only_alpha(pre_texts[0]), + } + return out + + +def get_adjusted_batch_count(params: AttributeDict) -> float: + # returns the number of batches we would have used so far if we had used the reference + # duration. This is for purposes of set_batch_count(). + return ( + params.batch_idx_train + * (params.max_duration * params.world_size) + / params.ref_duration + ) + + +def set_batch_count(model: Union[nn.Module, DDP], batch_count: float) -> None: + if isinstance(model, DDP): + # get underlying nn.Module + model = model.module + for name, module in model.named_modules(): + if hasattr(module, "batch_count"): + module.batch_count = batch_count + if hasattr(module, "name"): + module.name = name + + +def add_model_arguments(parser: argparse.ArgumentParser): + parser.add_argument( + "--num-encoder-layers", + type=str, + default="2,2,3,4,3,2", + help="Number of zipformer encoder layers per stack, comma separated.", + ) + + parser.add_argument( + "--downsampling-factor", + type=str, + default="1,2,4,8,4,2", + help="Downsampling factor for each stack of encoder layers.", + ) + + parser.add_argument( + "--feedforward-dim", + type=str, + default="512,768,1024,1536,1024,768", + help="Feedforward dimension of the zipformer encoder layers, per stack, comma separated.", + ) + + parser.add_argument( + "--num-heads", + type=str, + default="4,4,4,8,4,4", + help="Number of attention heads in the zipformer encoder layers: a single int or comma-separated list.", + ) + + parser.add_argument( + "--encoder-dim", + type=str, + default="192,256,384,512,384,256", + help="Embedding dimension in encoder stacks: a single int or comma-separated list.", + ) + + parser.add_argument( + "--text-encoder-dim", + type=str, + default="256,256,384,512", + help="Embedding dimension in text encoder stacks: a comma-separated list of 4 elements, " + "or you should change other configs in the code.", + ) + + parser.add_argument( + "--query-head-dim", + type=str, + default="32", + help="Query/key dimension per head in encoder stacks: a single int or comma-separated list.", + ) + + parser.add_argument( + "--value-head-dim", + type=str, + default="12", + help="Value dimension per head in encoder stacks: a single int or comma-separated list.", + ) + + parser.add_argument( + "--pos-head-dim", + type=str, + default="4", + help="Positional-encoding dimension per head in encoder stacks: a single int or comma-separated list.", + ) + + parser.add_argument( + "--pos-dim", + type=int, + default="48", + help="Positional-encoding embedding dimension", + ) + + parser.add_argument( + "--encoder-unmasked-dim", + type=str, + default="192,192,256,256,256,192", + help="Unmasked dimensions in the encoders, relates to augmentation during training. " + "A single int or comma-separated list. Must be <= each corresponding encoder_dim.", + ) + + parser.add_argument( + "--cnn-module-kernel", + type=str, + default="31,31,15,15,15,31", + help="Sizes of convolutional kernels in convolution modules in each encoder stack: " + "a single int or comma-separated list.", + ) + + parser.add_argument( + "--decoder-dim", + type=int, + default=512, + help="Embedding dimension in the decoder model.", + ) + + parser.add_argument( + "--context-size", + type=int, + default=2, + help="The context size in the decoder. 1 means bigram; " "2 means tri-gram", + ) + + parser.add_argument( + "--joiner-dim", + type=int, + default=512, + help="""Dimension used in the joiner model. + Outputs from the encoder and decoder model are projected + to this dimension before adding. + """, + ) + + parser.add_argument( + "--causal", + type=str2bool, + default=False, + help="If True, use causal version of model.", + ) + + parser.add_argument( + "--chunk-size", + type=str, + default="16,32,64,-1", + help="Chunk sizes (at 50Hz frame rate) will be chosen randomly from this list during training. " + " Must be just -1 if --causal=False", + ) + + parser.add_argument( + "--left-context-frames", + type=str, + default="64,128,256,-1", + help="Maximum left-contexts for causal training, measured in frames which will " + "be converted to a number of chunks. If splitting into chunks, " + "chunk left-context frames will be chosen randomly from this list; else not relevant.", + ) + + +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="pruned_transducer_stateless7/exp", + help="""The experiment dir. + It specifies the directory where all training related + files, e.g., checkpoints, log, etc, are saved + """, + ) + + parser.add_argument( + "--bpe-model", + type=str, + default="data/lang_bpe_500/bpe.model", + help="Path to the BPE model", + ) + + parser.add_argument( + "--base-lr", type=float, default=0.045, help="The base learning rate." + ) + + parser.add_argument( + "--lr-batches", + type=float, + default=7500, + help="""Number of steps that affects how rapidly the learning rate + decreases. We suggest not to change this.""", + ) + + parser.add_argument( + "--lr-epochs", + type=float, + default=3.5, + help="""Number of epochs that affects how rapidly the learning rate decreases. + """, + ) + + parser.add_argument( + "--ref-duration", + type=float, + default=600, + help="Reference batch duration for purposes of adjusting batch counts for setting various " + "schedules inside the model", + ) + + parser.add_argument( + "--prune-range", + type=int, + default=5, + help="The prune range for rnnt loss, it means how many symbols(context)" + "we are using to compute the loss", + ) + + parser.add_argument( + "--lm-scale", + type=float, + default=0.25, + help="The scale to smooth the loss with lm " + "(output of prediction network) part.", + ) + + parser.add_argument( + "--am-scale", + type=float, + default=0.0, + help="The scale to smooth the loss with am (output of encoder network)" "part.", + ) + + parser.add_argument( + "--simple-loss-scale", + type=float, + default=0.5, + help="To get pruning ranges, we will calculate a simple version" + "loss(joiner is just addition), this simple loss also uses for" + "training (as a regularization item). We will scale the simple loss" + "with this parameter before adding to the final loss.", + ) + + parser.add_argument( + "--seed", + type=int, + default=42, + help="The seed for random generators intended for reproducibility", + ) + + parser.add_argument( + "--print-diagnostics", + type=str2bool, + default=False, + help="Accumulate stats on activations, print them and exit.", + ) + + parser.add_argument( + "--inf-check", + type=str2bool, + default=False, + help="Add hooks to check for infinite module outputs and gradients.", + ) + + parser.add_argument( + "--save-every-n", + type=int, + default=4000, + help="""Save checkpoint after processing this number of batches" + periodically. We save checkpoint to exp-dir/ whenever + params.batch_idx_train % save_every_n == 0. The checkpoint filename + has the form: f'exp-dir/checkpoint-{params.batch_idx_train}.pt' + Note: It also saves checkpoint to `exp-dir/epoch-xxx.pt` at the + end of each epoch where `xxx` is the epoch number counting from 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. + + - encoder_dim: Hidden dim for multi-head attention model. + + - num_decoder_layers: Number of decoder layer of transformer decoder. + + - warm_step: The warmup period that dictates the decay of the + scale on "simple" (un-pruned) loss. + """ + params = AttributeDict( + { + "best_train_loss": float("inf"), + "best_valid_loss": float("inf"), + "best_train_epoch": -1, + "best_valid_epoch": -1, + "batch_idx_train": 0, + "log_interval": 50, + "reset_interval": 200, + "valid_interval": 3000, # For the 100h subset, use 800 + # parameters for zipformer + "feature_dim": 80, + "subsampling_factor": 4, # not passed in, this is fixed. + "warm_step": 2000, + "env_info": get_env_info(), + } + ) + + return params + + +def _to_int_tuple(s: str): + return tuple(map(int, s.split(","))) + + +def get_encoder_embed(params: AttributeDict) -> nn.Module: + # encoder_embed converts the input of shape (N, T, num_features) + # to the shape (N, (T - 7) // 2, encoder_dims). + # That is, it does two things simultaneously: + # (1) subsampling: T -> (T - 7) // 2 + # (2) embedding: num_features -> encoder_dims + # In the normal configuration, we will downsample once more at the end + # by a factor of 2, and most of the encoder stacks will run at a lower + # sampling rate. + encoder_embed = Conv2dSubsampling( + in_channels=params.feature_dim, + out_channels=_to_int_tuple(params.encoder_dim)[0], + dropout=ScheduledFloat((0.0, 0.3), (20000.0, 0.1)), + ) + return encoder_embed + + +def get_encoder_model(params: AttributeDict) -> nn.Module: + encoder = Zipformer2( + output_downsampling_factor=2, + downsampling_factor=_to_int_tuple(params.downsampling_factor), + num_encoder_layers=_to_int_tuple(params.num_encoder_layers), + encoder_dim=_to_int_tuple(params.encoder_dim), + encoder_unmasked_dim=_to_int_tuple(params.encoder_unmasked_dim), + query_head_dim=_to_int_tuple(params.query_head_dim), + pos_head_dim=_to_int_tuple(params.pos_head_dim), + value_head_dim=_to_int_tuple(params.value_head_dim), + pos_dim=params.pos_dim, + num_heads=_to_int_tuple(params.num_heads), + feedforward_dim=_to_int_tuple(params.feedforward_dim), + cnn_module_kernel=_to_int_tuple(params.cnn_module_kernel), + dropout=ScheduledFloat((0.0, 0.3), (20000.0, 0.1)), + warmup_batches=4000.0, + causal=params.causal, + chunk_size=_to_int_tuple(params.chunk_size), + left_context_frames=_to_int_tuple(params.left_context_frames), + ) + return encoder + + +def get_decoder_model(params: AttributeDict) -> nn.Module: + decoder = Decoder( + vocab_size=params.vocab_size, + decoder_dim=params.decoder_dim, + blank_id=params.blank_id, + context_size=params.context_size, + ) + return decoder + + +def get_joiner_model(params: AttributeDict) -> nn.Module: + joiner = Joiner( + encoder_dim=max(_to_int_tuple(params.encoder_dim)), + decoder_dim=params.decoder_dim, + joiner_dim=params.joiner_dim, + vocab_size=params.vocab_size, + ) + return joiner + + +def get_transducer_model(params: AttributeDict) -> nn.Module: + encoder_embed = get_encoder_embed(params) + encoder = get_encoder_model(params) + decoder = get_decoder_model(params) + joiner = get_joiner_model(params) + + model = Transducer( + encoder_embed=encoder_embed, + encoder=encoder, + decoder=decoder, + joiner=joiner, + encoder_dim=int(max(params.encoder_dim.split(","))), + decoder_dim=params.decoder_dim, + joiner_dim=params.joiner_dim, + vocab_size=params.vocab_size, + ) + return model + + +def load_checkpoint_if_available( + params: AttributeDict, + model: nn.Module, + model_avg: nn.Module = None, + optimizer: Optional[torch.optim.Optimizer] = None, + scheduler: Optional[LRSchedulerType] = None, +) -> Optional[Dict[str, Any]]: + """Load checkpoint from file. + + If params.start_batch is positive, it will load the checkpoint from + `params.exp_dir/checkpoint-{params.start_batch}.pt`. Otherwise, if + params.start_epoch is larger than 1, it will load the checkpoint from + `params.start_epoch - 1`. + + Apart from loading state dict for `model` and `optimizer` it also updates + `best_train_epoch`, `best_train_loss`, `best_valid_epoch`, + and `best_valid_loss` in `params`. + + Args: + params: + The return value of :func:`get_params`. + model: + The training model. + model_avg: + The stored model averaged from the start of training. + optimizer: + The optimizer that we are using. + scheduler: + The scheduler that we are using. + Returns: + Return a dict containing previously saved training info. + """ + if params.start_batch > 0: + filename = params.exp_dir / f"checkpoint-{params.start_batch}.pt" + elif params.start_epoch > 1: + filename = params.exp_dir / f"epoch-{params.start_epoch-1}.pt" + else: + return None + + assert filename.is_file(), f"{filename} does not exist!" + + saved_params = load_checkpoint( + filename, + model=model, + model_avg=model_avg, + optimizer=optimizer, + scheduler=scheduler, + ) + + keys = [ + "best_train_epoch", + "best_valid_epoch", + "batch_idx_train", + "best_train_loss", + "best_valid_loss", + ] + for k in keys: + params[k] = saved_params[k] + + if params.start_batch > 0: + if "cur_epoch" in saved_params: + params["start_epoch"] = saved_params["cur_epoch"] + + 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: Union[nn.Module, DDP], + sp: spm.SentencePieceProcessor, + batch: dict, + 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 Zipformer in our case. + batch: + A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()` + for the content in it. + is_training: + True for training. False for validation. When it is True, this + function enables autograd during computation; when it is False, it + disables autograd. + warmup: a floating point value which increases throughout training; + values >= 1.0 are fully warmed up and have all modules present. + """ + device = model.device if isinstance(model, DDP) else next(model.parameters()).device + feature = batch["inputs"] + # at entry, feature is (N, T, C) + assert feature.ndim == 3 + feature = feature.to(device) + + supervisions = batch["supervisions"] + feature_lens = supervisions["num_frames"].to(device) + + batch_idx_train = params.batch_idx_train + warm_step = params.warm_step + + texts = batch["supervisions"]["text"] + texts = [train_text_normalization(t) for t in texts] + y = sp.encode(texts, out_type=int) + y = k2.RaggedTensor(y).to(device) + + if random.random() < 0.02: + logging.info(f"Ref texts: {texts[0]}") + + with torch.set_grad_enabled(is_training): + simple_loss, pruned_loss = model( + x=feature, + x_lens=feature_lens, + y=y, + prune_range=params.prune_range, + am_scale=params.am_scale, + lm_scale=params.lm_scale, + ) + + s = params.simple_loss_scale + # take down the scale on the simple loss from 1.0 at the start + # to params.simple_loss scale by warm_step. + simple_loss_scale = ( + s + if batch_idx_train >= warm_step + else 1.0 - (batch_idx_train / warm_step) * (1.0 - s) + ) + pruned_loss_scale = ( + 1.0 + if batch_idx_train >= warm_step + else 0.1 + 0.9 * (batch_idx_train / warm_step) + ) + + loss = simple_loss_scale * simple_loss + pruned_loss_scale * pruned_loss + + assert loss.requires_grad == is_training + + info = MetricsTracker() + with warnings.catch_warnings(): + warnings.simplefilter("ignore") + info["frames"] = (feature_lens // params.subsampling_factor).sum().item() + + # Note: We use reduction=sum while computing the loss. + info["loss"] = loss.detach().cpu().item() + info["simple_loss"] = simple_loss.detach().cpu().item() + info["pruned_loss"] = pruned_loss.detach().cpu().item() + + return loss, info + + +def compute_validation_loss( + params: AttributeDict, + model: Union[nn.Module, DDP], + sp: spm.SentencePieceProcessor, + valid_dl: torch.utils.data.DataLoader, + world_size: int = 1, +) -> MetricsTracker: + """Run the validation process.""" + model.eval() + + tot_loss = MetricsTracker() + + for batch_idx, batch in enumerate(valid_dl): + loss, loss_info = compute_loss( + params=params, + model=model, + sp=sp, + batch=batch, + is_training=False, + ) + assert loss.requires_grad is False + tot_loss = tot_loss + loss_info + + if world_size > 1: + tot_loss.reduce(loss.device) + + loss_value = tot_loss["loss"] / tot_loss["frames"] + if loss_value < params.best_valid_loss: + params.best_valid_epoch = params.cur_epoch + params.best_valid_loss = loss_value + + return tot_loss + + +def train_one_epoch( + params: AttributeDict, + model: Union[nn.Module, DDP], + optimizer: torch.optim.Optimizer, + scheduler: LRSchedulerType, + sp: spm.SentencePieceProcessor, + train_dl: torch.utils.data.DataLoader, + valid_dl: torch.utils.data.DataLoader, + scaler: GradScaler, + model_avg: Optional[nn.Module] = None, + tb_writer: Optional[SummaryWriter] = None, + world_size: int = 1, + rank: int = 0, +) -> None: + """Train the model for one epoch. + + The training loss from the mean of all frames is saved in + `params.train_loss`. It runs the validation process every + `params.valid_interval` batches. + + Args: + params: + It is returned by :func:`get_params`. + model: + The model for training. + optimizer: + The optimizer we are using. + scheduler: + The learning rate scheduler, we call step() every step. + train_dl: + Dataloader for the training dataset. + valid_dl: + Dataloader for the validation dataset. + scaler: + The scaler used for mix precision training. + model_avg: + The stored model averaged from the start of training. + tb_writer: + Writer to write log messages to tensorboard. + world_size: + Number of nodes in DDP training. If it is 1, DDP is disabled. + rank: + The rank of the node in DDP training. If no DDP is used, it should + be set to 0. + """ + model.train() + + tot_loss = MetricsTracker() + + cur_batch_idx = params.get("cur_batch_idx", 0) + + saved_bad_model = False + + def save_bad_model(suffix: str = ""): + save_checkpoint_impl( + filename=params.exp_dir / f"bad-model{suffix}-{rank}.pt", + model=model, + model_avg=model_avg, + params=params, + optimizer=optimizer, + scheduler=scheduler, + sampler=train_dl.sampler, + scaler=scaler, + rank=0, + ) + + for batch_idx, batch in enumerate(train_dl): + if batch_idx % 10 == 0: + set_batch_count(model, get_adjusted_batch_count(params)) + if batch_idx < cur_batch_idx: + continue + cur_batch_idx = batch_idx + + params.batch_idx_train += 1 + batch_size = len(batch["supervisions"]["text"]) + + try: + with torch.cuda.amp.autocast(enabled=params.use_fp16): + loss, loss_info = compute_loss( + params=params, + model=model, + sp=sp, + batch=batch, + is_training=True, + ) + # summary stats + tot_loss = (tot_loss * (1 - 1 / params.reset_interval)) + loss_info + + # NOTE: We use reduction==sum and loss is computed over utterances + # in the batch and there is no normalization to it so far. + scaler.scale(loss).backward() + scheduler.step_batch(params.batch_idx_train) + + scaler.step(optimizer) + scaler.update() + optimizer.zero_grad() + except: # noqa + save_bad_model() + display_and_save_batch(batch, params=params, sp=sp) + raise + + if params.print_diagnostics and batch_idx == 5: + return + + if ( + rank == 0 + and params.batch_idx_train > 0 + and params.batch_idx_train % params.average_period == 0 + ): + update_averaged_model( + params=params, + model_cur=model, + model_avg=model_avg, + ) + + if ( + params.batch_idx_train > 0 + and params.batch_idx_train % params.save_every_n == 0 + ): + 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 < 8.0 or (cur_grad_scale < 32.0 and batch_idx % 400 == 0): + scaler.update(cur_grad_scale * 2.0) + if cur_grad_scale < 0.01: + if not saved_bad_model: + save_bad_model(suffix="-first-warning") + saved_bad_model = True + logging.warning(f"Grad scale is small: {cur_grad_scale}") + if cur_grad_scale < 1.0e-05: + save_bad_model() + raise RuntimeError( + f"grad_scale is too small, exiting: {cur_grad_scale}" + ) + + if batch_idx % params.log_interval == 0: + cur_lr = max(scheduler.get_last_lr()) + cur_grad_scale = scaler._scale.item() if params.use_fp16 else 1.0 + + logging.info( + f"Epoch {params.cur_epoch}, " + f"batch {batch_idx}, loss[{loss_info}], " + f"tot_loss[{tot_loss}], batch size: {batch_size}, " + f"lr: {cur_lr:.2e}, " + + (f"grad_scale: {scaler._scale.item()}" if params.use_fp16 else "") + ) + + if tb_writer is not None: + tb_writer.add_scalar( + "train/learning_rate", cur_lr, params.batch_idx_train + ) + + loss_info.write_summary( + tb_writer, "train/current_", params.batch_idx_train + ) + tot_loss.write_summary(tb_writer, "train/tot_", params.batch_idx_train) + if params.use_fp16: + tb_writer.add_scalar( + "train/grad_scale", + cur_grad_scale, + params.batch_idx_train, + ) + + if batch_idx % params.valid_interval == 0 and not params.print_diagnostics: + logging.info("Computing validation loss") + valid_info = compute_validation_loss( + params=params, + model=model, + sp=sp, + valid_dl=valid_dl, + world_size=world_size, + ) + model.train() + logging.info(f"Epoch {params.cur_epoch}, validation: {valid_info}") + logging.info( + f"Maximum memory allocated so far is {torch.cuda.max_memory_allocated()//1000000}MB" + ) + if tb_writer is not None: + valid_info.write_summary( + tb_writer, "train/valid_", params.batch_idx_train + ) + + loss_value = tot_loss["loss"] / tot_loss["frames"] + params.train_loss = loss_value + if params.train_loss < params.best_train_loss: + params.best_train_epoch = params.cur_epoch + params.best_train_loss = params.train_loss + + +def run(rank, world_size, args): + """ + Args: + rank: + It is a value between 0 and `world_size-1`, which is + passed automatically by `mp.spawn()` in :func:`main`. + The node with rank 0 is responsible for saving checkpoint. + world_size: + Number of GPUs for DDP training. + args: + The return value of get_parser().parse_args() + """ + params = get_params() + params.update(vars(args)) + + fix_random_seed(params.seed) + if world_size > 1: + setup_dist(rank, world_size, params.master_port) + + setup_logger(f"{params.exp_dir}/log/log-train") + logging.info("Training started") + + if args.tensorboard and rank == 0: + tb_writer = SummaryWriter(log_dir=f"{params.exp_dir}/tensorboard") + else: + tb_writer = None + + device = torch.device("cpu") + if torch.cuda.is_available(): + device = torch.device("cuda", rank) + logging.info(f"Device: {device}") + + sp = spm.SentencePieceProcessor() + sp.load(params.bpe_model) + + # is defined in local/train_bpe_model.py + params.blank_id = sp.piece_to_id("") + params.vocab_size = sp.get_piece_size() + + logging.info(params) + + logging.info("About to create model") + model = get_transducer_model(params) + + num_param = sum([p.numel() for p in model.parameters()]) + logging.info(f"Number of model parameters: {num_param}") + + assert params.save_every_n >= params.average_period + model_avg: Optional[nn.Module] = None + if rank == 0: + # model_avg is only used with rank 0 + model_avg = copy.deepcopy(model).to(torch.float64) + + assert params.start_epoch > 0, params.start_epoch + checkpoints = load_checkpoint_if_available( + params=params, model=model, model_avg=model_avg + ) + + model.to(device) + if world_size > 1: + logging.info("Using DDP") + model = DDP(model, device_ids=[rank], find_unused_parameters=True) + + optimizer = ScaledAdam( + get_parameter_groups_with_lrs(model, lr=params.base_lr, include_names=True), + lr=params.base_lr, # should have no effect + clipping_scale=2.0, + ) + + scheduler = Eden(optimizer, params.lr_batches, params.lr_epochs) + + if checkpoints and "optimizer" in checkpoints: + logging.info("Loading optimizer state dict") + optimizer.load_state_dict(checkpoints["optimizer"]) + + if ( + checkpoints + and "scheduler" in checkpoints + and checkpoints["scheduler"] is not None + ): + logging.info("Loading scheduler state dict") + scheduler.load_state_dict(checkpoints["scheduler"]) + + if params.print_diagnostics: + opts = diagnostics.TensorDiagnosticOptions( + 2**22 + ) # allow 4 megabytes per sub-module + diagnostic = diagnostics.attach_diagnostics(model, opts) + + if params.inf_check: + register_inf_check_hooks(model) + + libriheavy = LibriHeavyAsrDataModule(args) + + train_cuts = libriheavy.train_cuts() + + def remove_short_and_long_utt(c: Cut): + # Keep only utterances with duration between 1 second and 20 seconds + # + # Caution: There is a reason to select 20.0 here. Please see + # ../local/display_manifest_statistics.py + # + # You should use ../local/display_manifest_statistics.py to get + # an utterance duration distribution for your dataset to select + # the threshold + if c.duration < 1.0 or c.duration > 30.0: + return False + + # In pruned RNN-T, we require that T >= S + # where T is the number of feature frames after subsampling + # and S is the number of tokens in the utterance + + # In ./zipformer.py, the conv module uses the following expression + # for subsampling + T = ((c.num_frames - 7) // 2 + 1) // 2 + tokens = sp.encode(c.supervisions[0].texts[0], out_type=str) + + if T < len(tokens): + logging.warning( + f"Exclude cut with ID {c.id} from training. " + f"Number of frames (before subsampling): {c.num_frames}. " + f"Number of frames (after subsampling): {T}. " + f"Text: {c.supervisions[0].texts[0]}. " + f"Tokens: {tokens}. " + f"Number of tokens: {len(tokens)}" + ) + return False + + return True + + train_cuts = train_cuts.filter(remove_short_and_long_utt) + + if params.start_batch > 0 and checkpoints and "sampler" in checkpoints: + # We only load the sampler's state dict when it loads a checkpoint + # saved in the middle of an epoch + sampler_state_dict = checkpoints["sampler"] + else: + sampler_state_dict = None + + text_sampling_func = get_upper_only_alpha + logging.info(f"Text sampling func: {text_sampling_func}") + train_dl = libriheavy.train_dataloaders( + train_cuts, + sampler_state_dict=sampler_state_dict, + text_sampling_func=text_sampling_func, + ) + + valid_cuts = libriheavy.dev_cuts() + valid_dl = libriheavy.valid_dataloaders(valid_cuts) + + if not params.print_diagnostics: + scan_pessimistic_batches_for_oom( + model=model, + train_dl=train_dl, + optimizer=optimizer, + sp=sp, + params=params, + ) + + scaler = GradScaler(enabled=params.use_fp16, init_scale=1.0) + if checkpoints and "grad_scaler" in checkpoints: + logging.info("Loading grad scaler state dict") + scaler.load_state_dict(checkpoints["grad_scaler"]) + + for epoch in range(params.start_epoch, params.num_epochs + 1): + scheduler.step_epoch(epoch - 1) + fix_random_seed(params.seed + epoch - 1) + train_dl.sampler.set_epoch(epoch - 1) + + if tb_writer is not None: + tb_writer.add_scalar("train/epoch", epoch, params.batch_idx_train) + + params.cur_epoch = epoch + + train_one_epoch( + params=params, + model=model, + model_avg=model_avg, + optimizer=optimizer, + scheduler=scheduler, + sp=sp, + train_dl=train_dl, + valid_dl=valid_dl, + scaler=scaler, + tb_writer=tb_writer, + world_size=world_size, + rank=rank, + ) + + if params.print_diagnostics: + diagnostic.print_diagnostics() + break + + save_checkpoint( + params=params, + model=model, + model_avg=model_avg, + optimizer=optimizer, + scheduler=scheduler, + sampler=train_dl.sampler, + scaler=scaler, + rank=rank, + ) + + logging.info("Done!") + + if world_size > 1: + torch.distributed.barrier() + cleanup_dist() + + +def display_and_save_batch( + batch: dict, + params: AttributeDict, + sp: spm.SentencePieceProcessor, +) -> None: + """Display the batch statistics and save the batch into disk. + + Args: + batch: + A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()` + for the content in it. + params: + Parameters for training. See :func:`get_params`. + sp: + The BPE model. + """ + from lhotse.utils import uuid4 + + filename = f"{params.exp_dir}/batch-{uuid4()}.pt" + logging.info(f"Saving batch to {filename}") + torch.save(batch, filename) + + supervisions = batch["supervisions"] + features = batch["inputs"] + + logging.info(f"features shape: {features.shape}") + + y = sp.encode(supervisions["text"], out_type=int) + num_tokens = sum(len(i) for i in y) + logging.info(f"num tokens: {num_tokens}") + + +def scan_pessimistic_batches_for_oom( + model: Union[nn.Module, DDP], + train_dl: torch.utils.data.DataLoader, + optimizer: torch.optim.Optimizer, + sp: spm.SentencePieceProcessor, + params: AttributeDict, +): + from lhotse.dataset import find_pessimistic_batches + + logging.info( + "Sanity check -- see if any of the batches in epoch 1 would cause OOM." + ) + batches, crit_values = find_pessimistic_batches(train_dl.sampler) + for criterion, cuts in batches.items(): + batch = train_dl.dataset[cuts] + try: + with torch.cuda.amp.autocast(enabled=params.use_fp16): + loss, _ = compute_loss( + params=params, + model=model, + sp=sp, + batch=batch, + is_training=True, + ) + loss.backward() + optimizer.zero_grad() + except Exception as e: + if "CUDA out of memory" in str(e): + logging.error( + "Your GPU ran out of memory with the current " + "max_duration setting. We recommend decreasing " + "max_duration and trying again.\n" + f"Failing criterion: {criterion} " + f"(={crit_values[criterion]}) ..." + ) + display_and_save_batch(batch, params=params, sp=sp) + raise + logging.info( + f"Maximum memory allocated so far is {torch.cuda.max_memory_allocated()//1000000}MB" + ) + + +def main(): + parser = get_parser() + LibriHeavyAsrDataModule.add_arguments(parser) + args = parser.parse_args() + args.exp_dir = Path(args.exp_dir) + + world_size = args.world_size + assert world_size >= 1 + if world_size > 1: + mp.spawn(run, args=(world_size, args), nprocs=world_size, join=True) + else: + run(rank=0, world_size=1, args=args) + + +torch.set_num_threads(1) +torch.set_num_interop_threads(1) + +if __name__ == "__main__": + main() diff --git a/egs/libriheavy/ASR/zipformer_prompt_asr/train_bert_encoder.py b/egs/libriheavy/ASR/zipformer_prompt_asr/train_bert_encoder.py new file mode 100755 index 000000000..e253d1118 --- /dev/null +++ b/egs/libriheavy/ASR/zipformer_prompt_asr/train_bert_encoder.py @@ -0,0 +1,1798 @@ +#!/usr/bin/env python3 +# Copyright 2021-2022 Xiaomi Corp. (authors: Xiaoyu Yang, +# +# +# See ../../../../LICENSE for clarification regarding multiple authors +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +""" +Usage: + +export CUDA_VISIBLE_DEVICES="0,1,2,3" + +# For mix precision training: + +(1) Non-streaming model, **without** context list + +./zipformer_prompt_asr/train.py \ + --world-size 4 \ + --num-epochs 30 \ + --start-epoch 1 \ + --use-fp16 1 \ + --subset medium \ + --causal False \ + --exp-dir zipformer_prompt_asr/exp \ + --max-duration 1000 \ + --memory-layer 0 \ + --text-encoder-type BERT \ + --text-encoder-dim 768 \ + --use-style-prompt True \ + --use-context-list False + +(2) Non-streaming model, **with** context list + +./zipformer_prompt_asr/train.py \ + --world-size 4 \ + --num-epochs 30 \ + --start-epoch 1 \ + --use-fp16 1 \ + --subset medium \ + --causal False \ + --exp-dir zipformer_prompt_asr/exp \ + --max-duration 1000 \ + --memory-layer 0 \ + --text-encoder-type BERT \ + --text-encoder-dim 768 \ + --use-style-prompt True \ + --use-context-list True \ + --top-k 10000 \ + --rare-word-file data/context_biasing/small_rare_words_topk_10000.txt + + +""" + + +import argparse +import copy +import logging +import os +import random +import warnings +from pathlib import Path +from shutil import copyfile +from typing import Any, Dict, List, Optional, Tuple, Union + +import k2 +import numpy +import optim +import sentencepiece as spm +import torch +import torch.multiprocessing as mp +import torch.nn as nn +from asr_datamodule import LibriHeavyAsrDataModule +from dataset import ( + naive_triplet_text_sampling, + random_shuffle_subset, + triplet_text_sampling, + triplet_text_sampling_with_context_list, +) +from decoder import Decoder +from joiner import Joiner +from lhotse.cut import Cut +from lhotse.dataset.sampling.base import CutSampler +from lhotse.utils import fix_random_seed +from model_with_BERT import PromptedTransducer +from optim import Eden, ScaledAdam +from scaling import Balancer, BiasNorm, Dropout3, ScaleGrad, ScheduledFloat, SwooshR +from subsampling import Conv2dSubsampling +from text_normalization import ( + lower_all_char, + lower_only_alpha, + train_text_normalization, + upper_all_char, + upper_only_alpha, +) +from torch import Tensor +from torch.cuda.amp import GradScaler +from torch.nn.parallel import DistributedDataParallel as DDP +from torch.utils.tensorboard import SummaryWriter +from zipformer import Zipformer2 + +from icefall import diagnostics +from icefall.checkpoint import load_checkpoint, remove_checkpoints +from icefall.checkpoint import save_checkpoint as save_checkpoint_impl +from icefall.checkpoint import ( + save_checkpoint_with_global_batch_idx, + update_averaged_model, +) +from icefall.dist import cleanup_dist, setup_dist +from icefall.env import get_env_info +from icefall.hooks import register_inf_check_hooks +from icefall.utils import ( + AttributeDict, + MetricsTracker, + get_parameter_groups_with_lrs, + setup_logger, + str2bool, +) + +LRSchedulerType = Union[torch.optim.lr_scheduler._LRScheduler, optim.LRScheduler] + +style_transforms = [ + lambda x: x, # return it self + upper_only_alpha, + lower_only_alpha, + lower_all_char, +] + + +def get_first(texts: List[str], pre_texts: List[str]) -> str: + out = { + "text": texts[0], + "pre_text": pre_texts[0], + "style_text": "", + "transform_ids": 0, + } + return out + + +def get_upper_only_alpha(texts: List[str], pre_texts: List[str]) -> str: + # Always get the first one, which is the gt (mixed-cased trans), but with upper_only_alpha + out = { + "text": upper_only_alpha(texts[0]), + "pre_text": upper_only_alpha(pre_texts[0]), + "style_text": "", + "transform_ids": 0, + } + return out + + +def get_adjusted_batch_count(params: AttributeDict) -> float: + # returns the number of batches we would have used so far if we had used the reference + # duration. This is for purposes of set_batch_count(). + return ( + params.batch_idx_train + * (params.max_duration * params.world_size) + / params.ref_duration + ) + + +def set_batch_count(model: Union[nn.Module, DDP], batch_count: float) -> None: + if isinstance(model, DDP): + # get underlying nn.Module + model = model.module + for name, module in model.named_modules(): + if hasattr(module, "batch_count"): + module.batch_count = batch_count + if hasattr(module, "name"): + module.name = name + + +def add_model_arguments(parser: argparse.ArgumentParser): + parser.add_argument( + "--num-encoder-layers", + type=str, + default="2,2,3,4,3,2", + help="Number of zipformer encoder layers per stack, comma separated.", + ) + + parser.add_argument( + "--downsampling-factor", + type=str, + default="1,2,4,8,4,2", + help="Downsampling factor for each stack of encoder layers.", + ) + + parser.add_argument( + "--feedforward-dim", + type=str, + default="512,768,1024,1536,1024,768", + help="Feedforward dimension of the zipformer encoder layers, per stack, comma separated.", + ) + + parser.add_argument( + "--num-heads", + type=str, + default="4,4,4,8,4,4", + help="Number of attention heads in the zipformer encoder layers: a single int or comma-separated list.", + ) + + parser.add_argument( + "--encoder-dim", + type=str, + default="192,256,384,512,384,256", + help="Embedding dimension in encoder stacks: a single int or comma-separated list.", + ) + + parser.add_argument( + "--memory-dropout-rate", + type=float, + default=0.05, + help="By which probability, dropout the memory when doing cross-attention.", + ) + + parser.add_argument( + "--memory-layer", + type=int, + default=0, + help="Start doing cross-attention from which layer. Zero-indexed", + ) + + parser.add_argument( + "--query-head-dim", + type=str, + default="32", + help="Query/key dimension per head in encoder stacks: a single int or comma-separated list.", + ) + + parser.add_argument( + "--value-head-dim", + type=str, + default="12", + help="Value dimension per head in encoder stacks: a single int or comma-separated list.", + ) + + parser.add_argument( + "--pos-head-dim", + type=str, + default="4", + help="Positional-encoding dimension per head in encoder stacks: a single int or comma-separated list.", + ) + + parser.add_argument( + "--pos-dim", + type=int, + default="48", + help="Positional-encoding embedding dimension", + ) + + parser.add_argument( + "--encoder-unmasked-dim", + type=str, + default="192,192,256,256,256,192", + help="Unmasked dimensions in the encoders, relates to augmentation during training. " + "A single int or comma-separated list. Must be <= each corresponding encoder_dim.", + ) + + parser.add_argument( + "--cnn-module-kernel", + type=str, + default="31,31,15,15,15,31", + help="Sizes of convolutional kernels in convolution modules in each encoder stack: " + "a single int or comma-separated list.", + ) + + parser.add_argument( + "--decoder-dim", + type=int, + default=512, + help="Embedding dimension in the decoder model.", + ) + + parser.add_argument( + "--joiner-dim", + type=int, + default=512, + help="""Dimension used in the joiner model. + Outputs from the encoder and decoder model are projected + to this dimension before adding. + """, + ) + + parser.add_argument( + "--context-size", + type=int, + default=2, + help="The context size in the decoder. 1 means bigram; " "2 means tri-gram", + ) + + parser.add_argument( + "--causal", + type=str2bool, + default=False, + help="If True, use causal version of model.", + ) + + parser.add_argument( + "--chunk-size", + type=str, + default="16,32,64,-1", + help="Chunk sizes (at 50Hz frame rate) will be chosen randomly from this list during training. " + " Must be just -1 if --causal=False", + ) + + parser.add_argument( + "--left-context-frames", + type=str, + default="64,128,256,-1", + help="Maximum left-contexts for causal training, measured in frames which will " + "be converted to a number of chunks. If splitting into chunks, " + "chunk left-context frames will be chosen randomly from this list; else not relevant.", + ) + + parser.add_argument( + "--freeze-text-encoder", + type=str2bool, + default=True, + ) + + parser.add_argument( + "--text-encoder-type", + type=str, + default="BERT", + choices=["BERT", "DistilBERT"], + help="Type of the text encoder", + ) + + parser.add_argument( + "--text-encoder-dim", + type=int, + default=768, + help="Dimension of the text encoder", + ) + + parser.add_argument( + "--text-encoder-adapter", + type=str2bool, + default=False, + help="An adapter for pre-trained BERT", + ) + + parser.add_argument( + "--context-injection", + type=str2bool, + default=False, + help="Inject context embedding into the joiner", + ) + + parser.add_argument( + "--context-dropout-rate", + type=float, + default=0.05, + ) + + +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="pruned_transducer_stateless7/exp", + help="""The experiment dir. + It specifies the directory where all training related + files, e.g., checkpoints, log, etc, are saved + """, + ) + + parser.add_argument( + "--bpe-model", + type=str, + default="data/lang_bpe_500/bpe.model", + help="Path to the BPE model", + ) + + parser.add_argument( + "--base-lr", type=float, default=0.045, help="The base learning rate." + ) + + parser.add_argument( + "--lr-batches", + type=float, + default=7500, + help="""Number of steps that affects how rapidly the learning rate + decreases. We suggest not to change this.""", + ) + + parser.add_argument( + "--lr-epochs", + type=float, + default=3.5, + help="""Number of epochs that affects how rapidly the learning rate decreases. + """, + ) + + parser.add_argument( + "--ref-duration", + type=float, + default=600, + help="Reference batch duration for purposes of adjusting batch counts for setting various " + "schedules inside the model", + ) + + parser.add_argument( + "--prune-range", + type=int, + default=5, + help="The prune range for rnnt loss, it means how many symbols(context)" + "we are using to compute the loss", + ) + + parser.add_argument( + "--lm-scale", + type=float, + default=0.25, + help="The scale to smooth the loss with lm " + "(output of prediction network) part.", + ) + + parser.add_argument( + "--am-scale", + type=float, + default=0.0, + help="The scale to smooth the loss with am (output of encoder network)" "part.", + ) + + parser.add_argument( + "--simple-loss-scale", + type=float, + default=0.5, + help="To get pruning ranges, we will calculate a simple version" + "loss(joiner is just addition), this simple loss also uses for" + "training (as a regularization item). We will scale the simple loss" + "with this parameter before adding to the final loss.", + ) + + parser.add_argument( + "--seed", + type=int, + default=42, + help="The seed for random generators intended for reproducibility", + ) + + parser.add_argument( + "--print-diagnostics", + type=str2bool, + default=False, + help="Accumulate stats on activations, print them and exit.", + ) + + parser.add_argument( + "--inf-check", + type=str2bool, + default=False, + help="Add hooks to check for infinite module outputs and gradients.", + ) + + parser.add_argument( + "--save-every-n", + type=int, + default=4000, + help="""Save checkpoint after processing this number of batches" + periodically. We save checkpoint to exp-dir/ whenever + params.batch_idx_train % save_every_n == 0. The checkpoint filename + has the form: f'exp-dir/checkpoint-{params.batch_idx_train}.pt' + Note: It also saves checkpoint to `exp-dir/epoch-xxx.pt` at the + end of each epoch where `xxx` is the epoch number counting from 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.", + ) + + parser.add_argument( + "--use-style-prompt", + type=str2bool, + default=True, + help="Whether to use style prompt.", + ) + + # arguments for using prompt + parser.add_argument( + "--pre-text-shuffle-prob", + type=float, + default=0.05, + help="The proportion of pre_text to be shuffled with in a batch", + ) + + parser.add_argument( + "--style-text-shuffle-prob", + type=float, + default=0.2, + help="The proportion of style_text to be shuffled with in a batch", + ) + + parser.add_argument( + "--prompt-mask-prob", + type=float, + default=0.05, + help="The probability of masking prompts", + ) + + parser.add_argument( + "--forced-upper-pre-text", + type=str2bool, + default=False, + help="Forced format of pre-text", + ) + + add_model_arguments(parser) + + return parser + + +def get_params() -> AttributeDict: + """Return a dict containing training parameters. + + All training related parameters that are not passed from the commandline + are saved in the variable `params`. + + Commandline options are merged into `params` after they are parsed, so + you can also access them via `params`. + + Explanation of options saved in `params`: + + - best_train_loss: Best training loss so far. It is used to select + the model that has the lowest training loss. It is + updated during the training. + + - best_valid_loss: Best validation loss so far. It is used to select + the model that has the lowest validation loss. It is + updated during the training. + + - best_train_epoch: It is the epoch that has the best training loss. + + - best_valid_epoch: It is the epoch that has the best validation loss. + + - batch_idx_train: Used to writing statistics to tensorboard. It + contains number of batches trained so far across + epochs. + + - log_interval: Print training loss if batch_idx % log_interval` is 0 + + - reset_interval: Reset statistics if batch_idx % reset_interval is 0 + + - valid_interval: Run validation if batch_idx % valid_interval is 0 + + - feature_dim: The model input dim. It has to match the one used + in computing features. + + - subsampling_factor: The subsampling factor for the model. + + - encoder_dim: Hidden dim for multi-head attention model. + + - num_decoder_layers: Number of decoder layer of transformer decoder. + + - warm_step: The warmup period that dictates the decay of the + scale on "simple" (un-pruned) loss. + """ + params = AttributeDict( + { + "best_train_loss": float("inf"), + "best_valid_loss": float("inf"), + "best_train_epoch": -1, + "best_valid_epoch": -1, + "batch_idx_train": 0, + "log_interval": 50, + "reset_interval": 200, + "valid_interval": 3000, # For the 100h subset, use 800 + # parameters for zipformer + "feature_dim": 80, + "subsampling_factor": 4, # not passed in, this is fixed. + "warm_step": 2000, + "env_info": get_env_info(), + } + ) + + return params + + +def _to_int_tuple(s: str): + return tuple(map(int, s.split(","))) + + +def get_encoder_embed(params: AttributeDict) -> nn.Module: + # encoder_embed converts the input of shape (N, T, num_features) + # to the shape (N, (T - 7) // 2, encoder_dims). + # That is, it does two things simultaneously: + # (1) subsampling: T -> (T - 7) // 2 + # (2) embedding: num_features -> encoder_dims + # In the normal configuration, we will downsample once more at the end + # by a factor of 2, and most of the encoder stacks will run at a lower + # sampling rate. + encoder_embed = Conv2dSubsampling( + in_channels=params.feature_dim, + out_channels=_to_int_tuple(params.encoder_dim)[0], + dropout=ScheduledFloat((0.0, 0.3), (20000.0, 0.1)), + ) + return encoder_embed + + +class TextEmbedding(nn.Module): + def __init__( + self, + num_embeddings: int = 256, + embedding_dim: int = 256, + kernel_size: int = 3, + layer1_channels: int = 256, + layer2_channels: int = 256, + bias: bool = True, + dropout: float = 0.1, + ): + super().__init__() + self.embed = nn.Embedding( + num_embeddings=num_embeddings, # we encode the text as UTF-8 bytes + embedding_dim=embedding_dim, # + ) + + assert embedding_dim == layer1_channels # for depth wise convolution + self.conv = nn.Sequential( + nn.Conv1d( + embedding_dim, + layer1_channels, # depthwise convolution + kernel_size=kernel_size, + stride=1, + padding=(kernel_size - 1) // 2, + groups=layer1_channels, + bias=True, + ), + ScaleGrad(0.2), + Balancer(layer1_channels, channel_dim=1, min_positive=0.1, max_abs=1.0), + nn.ReLU(), + nn.Conv1d( + layer1_channels, + layer2_channels, + kernel_size=1, # pointwise convolution + stride=1, + padding=0, + bias=True, + ), + Balancer(layer2_channels, channel_dim=1, min_positive=0.1, max_abs=1.0), + nn.ReLU(), + ) + + self.out_norm = BiasNorm(layer2_channels) + self.dropout = Dropout3(dropout, shared_dim=1) + + def forward(self, text: torch.Tensor) -> torch.Tensor: + """Forward function of the text embedding + + Args: + text (torch.Tensor): Text in UTF-8 bytes (T,N) + Returns: + The embeddings of text (T,N,C) + """ + text = self.embed(text) # (T,N,C) + + # src = text + text = text.permute(1, 2, 0) # (T,N,C) -> (N,C,T) + text = self.conv(text) + text = text.permute(2, 0, 1) # (N,C,T) -> (T,N,C) + # src = src + text + + text = self.out_norm(text) + text = self.dropout(text) + + return text + + +def get_text_encoder(params: AttributeDict) -> nn.Module: + # Return a text encoder + if params.text_encoder_type == "BERT": # This is a BERT-base-cased + from transformers import BertModel + + logging.info("Loading pre-trained BERT-base-cased as text encoder") + if os.path.exists("data/models/bert-base-cased"): + model = BertModel.from_pretrained("data/models/bert-base-cased") + else: + model = BertModel.from_pretrained("bert-base-cased") + assert params.text_encoder_dim == 768 + elif params.text_encoder_type == "BERT-large": + from transformers import BertModel + + logging.info("Loading pre-trained BERT-large-uncased as text encoder") + if os.path.exists("data/models/bert-large-uncased"): + model = BertModel.from_pretrained("data/models/bert-large-uncased") + else: + model = BertModel.from_pretrained("bert-large-uncased") + assert params.text_encoder_dim == 1024 + elif params.text_encoder_type == "DistilBERT": + from transformers import DistilBertModel # This is a DistilBERT-base-cased + + logging.info("Loading pre-trained DistilBERT-base-cased as text encoder") + model = DistilBertModel.from_pretrained("distilbert-base-cased") + assert params.text_encoder_dim == 768 + else: + raise ValueError() + + return model + + +def get_tokenizer(params: AttributeDict): + + if params.text_encoder_type == "BERT": + from transformers import BertTokenizer + + # This is a BERT-base-cased + if os.path.exists("data/models/bert-base-cased"): + tokenizer = BertTokenizer.from_pretrained("data/models/bert-base-cased") + else: + tokenizer = BertTokenizer.from_pretrained("bert-base-cased") + elif params.text_encoder_type == "BERT-large": + from transformers import BertTokenizer + + # This is a BERT-large-uncased + if os.path.exists("data/models/bert-large-uncased"): + tokenizer = BertTokenizer.from_pretrained("data/models/bert-large-uncased") + else: + tokenizer = BertTokenizer.from_pretrained("bert-large-uncased") + elif params.text_encoder_type == "DistilBERT": + from transformers import DistilBertTokenizer + + tokenizer = DistilBertTokenizer.from_pretrained("distilbert-base-cased") + else: + raise ValueError() + + return tokenizer + + +def get_encoder_model(params: AttributeDict) -> nn.Module: + encoder = Zipformer2( + output_downsampling_factor=2, + downsampling_factor=_to_int_tuple(params.downsampling_factor), + num_encoder_layers=_to_int_tuple(params.num_encoder_layers), + encoder_dim=_to_int_tuple(params.encoder_dim), + encoder_unmasked_dim=_to_int_tuple(params.encoder_unmasked_dim), + query_head_dim=_to_int_tuple(params.query_head_dim), + pos_head_dim=_to_int_tuple(params.pos_head_dim), + value_head_dim=_to_int_tuple(params.value_head_dim), + pos_dim=params.pos_dim, + num_heads=_to_int_tuple(params.num_heads), + feedforward_dim=_to_int_tuple(params.feedforward_dim), + cnn_module_kernel=_to_int_tuple(params.cnn_module_kernel), + dropout=ScheduledFloat((0.0, 0.3), (20000.0, 0.1)), + warmup_batches=4000.0, + causal=params.causal, + chunk_size=_to_int_tuple(params.chunk_size), + left_context_frames=_to_int_tuple(params.left_context_frames), + memory_dim=params.text_encoder_dim, # This is fixed as the BERT base model is 768-D + memory_layer=params.memory_layer, + memory_dropout_rate=params.memory_dropout_rate, + ) + return encoder + + +def get_decoder_model(params: AttributeDict) -> nn.Module: + decoder = Decoder( + vocab_size=params.vocab_size, + decoder_dim=params.decoder_dim, + blank_id=params.blank_id, + context_size=params.context_size, + ) + return decoder + + +def get_joiner_model(params: AttributeDict) -> nn.Module: + joiner = Joiner( + encoder_dim=max(_to_int_tuple(params.encoder_dim)), + decoder_dim=params.decoder_dim, + joiner_dim=params.joiner_dim, + vocab_size=params.vocab_size, + context_dim=4 * 768 + if params.context_injection + else -1, # the output dim of text encoder + context_injection=params.context_injection, + ) + return joiner + + +def get_transducer_model(params: AttributeDict) -> nn.Module: + text_encoder = get_text_encoder(params) # This should be a cased BERT base model + num_param = sum([p.numel() for p in text_encoder.parameters()]) + logging.info(f"Num params in text encoder: {num_param}") + + encoder_embed = get_encoder_embed(params) + encoder = get_encoder_model(params) + + decoder = get_decoder_model(params) + joiner = get_joiner_model(params) + + model = PromptedTransducer( + encoder_embed=encoder_embed, + encoder=encoder, + text_encoder=text_encoder, + decoder=decoder, + joiner=joiner, + encoder_dim=int(max(params.encoder_dim.split(","))), + decoder_dim=params.decoder_dim, + joiner_dim=params.joiner_dim, + vocab_size=params.vocab_size, + text_encoder_type=params.text_encoder_type, + text_encoder_adapter=params.text_encoder_adapter, + freeze_text_encoder=params.freeze_text_encoder, + context_fuser=None, + ) + + 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 _encode_texts_as_bytes_with_tokenizer( + pre_texts: List[str], + style_texts: List[str], + tokenizer, + device: torch.device, + max_len: int = 500, + no_limit: bool = False, +) -> Tuple[Dict, Tensor]: + """ + Encode texts as bytes and then integer tensors. + Note that the style text will be added to the beginning of texts. + """ + batch_size = len(pre_texts) + max_len = min(max_len, 500) + + if no_limit: + allowed_lens = [5000 - len(s) for s in style_texts] + else: + allowed_lens = [1000 - len(s) for s in style_texts] + truncated_pre_texts = [pre_texts[i][-allowed_lens[i] :] for i in range(batch_size)] + combined_text = [ + style_texts[i] + " [SEP] " + truncated_pre_texts[i] for i in range(batch_size) + ] + + encoded_style_texts = tokenizer( + style_texts, + return_tensors="pt", + padding=True, + truncation=True, + return_length=True, + max_length=max_len, + ) + style_lens = encoded_style_texts["length"].to(device) + + # Use tokenizer to prepare input for text encoder + encoded_inputs = tokenizer( + combined_text, + return_tensors="pt", + padding=True, + truncation=True, + return_length=True, + max_length=max_len, + ).to(device) + + return encoded_inputs, style_lens + + +def compute_loss( + params: AttributeDict, + model: Union[nn.Module, DDP], + sp: spm.SentencePieceProcessor, + tokenizer, + batch: dict, + 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 Zipformer in our case. + batch: + A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()` + for the content in it. + is_training: + True for training. False for validation. When it is True, this + function enables autograd during computation; when it is False, it + disables autograd. + warmup: a floating point value which increases throughout training; + values >= 1.0 are fully warmed up and have all modules present. + """ + device = model.device if isinstance(model, DDP) else next(model.parameters()).device + feature = batch["inputs"] + # at entry, feature is (N, T, C) + assert feature.ndim == 3 + feature = feature.to(device) + batch_size = feature.size(0) + + supervisions = batch["supervisions"] + feature_lens = supervisions["num_frames"].to(device) + + batch_idx_train = params.batch_idx_train + warm_step = params.warm_step + + texts = batch["supervisions"]["text"] + pre_texts = batch["supervisions"]["pre_text"] + style_texts = batch["supervisions"][ + "style_text" + ] # the style texts are in gt format + transform_ids = batch["supervisions"]["transform_ids"] + + # This is to replace full-width symbols with half-width symbols + texts = [train_text_normalization(t) for t in texts] + pre_texts = [train_text_normalization(t) for t in pre_texts] + style_texts = [train_text_normalization(t) for t in style_texts] + + y = sp.encode( + texts, out_type=int + ) # sp.encode treats consecutive space as a single space + y = k2.RaggedTensor(y).to(device) + + if params.forced_upper_pre_text: + pre_texts = [upper_only_alpha(p) for p in pre_texts] + + # only shuffle the pre_text and style texts if during training, and use style prompt + if is_training: + # randomly shuffle&mask the pre_text + pre_texts = random_shuffle_subset( + pre_texts, + p=params.pre_text_shuffle_prob, + p_mask=params.prompt_mask_prob, + ) + + if params.use_style_prompt: + if random.random() < 0.5: + # randomly shuffle the style_text + # now the style_texts are all in gt format + style_texts = random_shuffle_subset( + style_texts, + p=params.style_text_shuffle_prob, + p_mask=params.prompt_mask_prob, + ) + + assert len(transform_ids) == len(style_texts) + + for i in range(len(style_texts)): + t = transform_ids[i] # get the transform id + style_texts[i] = style_transforms[t](style_texts[i]) + + if not params.use_style_prompt: + style_texts = [ + "" for _ in style_texts + ] # use empty string for style texts if don't use style prompt + + if random.random() < 0.05: + logging.info(f"Pre texts: {pre_texts[0]}") + logging.info(f"Ref texts: {texts[0]}") + logging.info(f"Style texts: {style_texts[0]}") + + encoded_inputs, style_lens = _encode_texts_as_bytes_with_tokenizer( + pre_texts=pre_texts, + style_texts=style_texts, + tokenizer=tokenizer, + device=device, + ) + + if random.random() < 0.02: + logging.info(f"Shape of encoded texts: {encoded_inputs['input_ids'].shape} ") + + with torch.set_grad_enabled(is_training): + simple_loss, pruned_loss = model( + x=feature, + x_lens=feature_lens, + encoded_inputs=encoded_inputs, + style_lens=style_lens, + y=y, + prune_range=params.prune_range, + am_scale=params.am_scale, + lm_scale=params.lm_scale, + ) + + s = params.simple_loss_scale + # take down the scale on the simple loss from 1.0 at the start + # to params.simple_loss scale by warm_step. + simple_loss_scale = ( + s + if batch_idx_train >= warm_step + else 1.0 - (batch_idx_train / warm_step) * (1.0 - s) + ) + pruned_loss_scale = ( + 1.0 + if batch_idx_train >= warm_step + else 0.1 + 0.9 * (batch_idx_train / warm_step) + ) + + loss = simple_loss_scale * simple_loss + pruned_loss_scale * pruned_loss + + assert loss.requires_grad == is_training + + info = MetricsTracker() + with warnings.catch_warnings(): + warnings.simplefilter("ignore") + info["frames"] = (feature_lens // params.subsampling_factor).sum().item() + + # Note: We use reduction=sum while computing the loss. + info["loss"] = loss.detach().cpu().item() + info["simple_loss"] = simple_loss.detach().cpu().item() + info["pruned_loss"] = pruned_loss.detach().cpu().item() + + return loss, info + + +def compute_validation_loss( + params: AttributeDict, + model: Union[nn.Module, DDP], + sp: spm.SentencePieceProcessor, + tokenizer, + valid_dl: torch.utils.data.DataLoader, + world_size: int = 1, +) -> MetricsTracker: + """Run the validation process.""" + model.eval() + + tot_loss = MetricsTracker() + + for batch_idx, batch in enumerate(valid_dl): + loss, loss_info = compute_loss( + params=params, + model=model, + sp=sp, + tokenizer=tokenizer, + batch=batch, + is_training=False, + ) + assert loss.requires_grad is False + tot_loss = tot_loss + loss_info + + if world_size > 1: + tot_loss.reduce(loss.device) + + loss_value = tot_loss["loss"] / tot_loss["frames"] + if loss_value < params.best_valid_loss: + params.best_valid_epoch = params.cur_epoch + params.best_valid_loss = loss_value + + return tot_loss + + +def train_one_epoch( + params: AttributeDict, + model: Union[nn.Module, DDP], + optimizer: torch.optim.Optimizer, + scheduler: LRSchedulerType, + sp: spm.SentencePieceProcessor, + tokenizer, + train_dl: torch.utils.data.DataLoader, + valid_dl: torch.utils.data.DataLoader, + scaler: GradScaler, + model_avg: Optional[nn.Module] = None, + tb_writer: Optional[SummaryWriter] = None, + world_size: int = 1, + rank: int = 0, +) -> None: + """Train the model for one epoch. + + The training loss from the mean of all frames is saved in + `params.train_loss`. It runs the validation process every + `params.valid_interval` batches. + + Args: + params: + It is returned by :func:`get_params`. + model: + The model for training. + optimizer: + The optimizer we are using. + scheduler: + The learning rate scheduler, we call step() every step. + train_dl: + Dataloader for the training dataset. + valid_dl: + Dataloader for the validation dataset. + scaler: + The scaler used for mix precision training. + model_avg: + The stored model averaged from the start of training. + tb_writer: + Writer to write log messages to tensorboard. + world_size: + Number of nodes in DDP training. If it is 1, DDP is disabled. + rank: + The rank of the node in DDP training. If no DDP is used, it should + be set to 0. + """ + model.train() + + tot_loss = MetricsTracker() + + cur_batch_idx = params.get("cur_batch_idx", 0) + + saved_bad_model = False + + def save_bad_model(suffix: str = ""): + save_checkpoint_impl( + filename=params.exp_dir / f"bad-model{suffix}-{rank}.pt", + model=model, + model_avg=model_avg, + params=params, + optimizer=optimizer, + scheduler=scheduler, + sampler=train_dl.sampler, + scaler=scaler, + rank=0, + ) + + for batch_idx, batch in enumerate(train_dl): + if batch_idx % 10 == 0: + set_batch_count(model, get_adjusted_batch_count(params)) + if batch_idx < cur_batch_idx: + continue + cur_batch_idx = batch_idx + + params.batch_idx_train += 1 + batch_size = len(batch["supervisions"]["text"]) + + try: + with torch.cuda.amp.autocast(enabled=params.use_fp16): + loss, loss_info = compute_loss( + params=params, + model=model, + sp=sp, + tokenizer=tokenizer, + batch=batch, + is_training=True, + ) + # summary stats + tot_loss = (tot_loss * (1 - 1 / params.reset_interval)) + loss_info + + # NOTE: We use reduction==sum and loss is computed over utterances + # in the batch and there is no normalization to it so far. + scaler.scale(loss).backward() + scheduler.step_batch(params.batch_idx_train) + + scaler.step(optimizer) + scaler.update() + optimizer.zero_grad() + except: # noqa + save_bad_model() + display_and_save_batch(batch, params=params, sp=sp) + raise + + if params.print_diagnostics and batch_idx == 5: + return + + if ( + rank == 0 + and params.batch_idx_train > 0 + and params.batch_idx_train % params.average_period == 0 + ): + update_averaged_model( + params=params, + model_cur=model, + model_avg=model_avg, + ) + + if ( + params.batch_idx_train > 0 + and params.batch_idx_train % params.save_every_n == 0 + ): + 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 < 8.0 or (cur_grad_scale < 32.0 and batch_idx % 400 == 0): + scaler.update(cur_grad_scale * 2.0) + if cur_grad_scale < 0.01: + if not saved_bad_model: + save_bad_model(suffix="-first-warning") + saved_bad_model = True + logging.warning(f"Grad scale is small: {cur_grad_scale}") + if cur_grad_scale < 1.0e-05: + save_bad_model() + raise RuntimeError( + f"grad_scale is too small, exiting: {cur_grad_scale}" + ) + + if batch_idx % params.log_interval == 0: + cur_lr = max(scheduler.get_last_lr()) + cur_grad_scale = scaler._scale.item() if params.use_fp16 else 1.0 + + logging.info( + f"Epoch {params.cur_epoch}, " + f"batch {batch_idx}, loss[{loss_info}], " + f"tot_loss[{tot_loss}], batch size: {batch_size}, " + f"lr: {cur_lr:.2e}, " + + (f"grad_scale: {scaler._scale.item()}" if params.use_fp16 else "") + ) + + if tb_writer is not None: + tb_writer.add_scalar( + "train/learning_rate", cur_lr, params.batch_idx_train + ) + + loss_info.write_summary( + tb_writer, "train/current_", params.batch_idx_train + ) + tot_loss.write_summary(tb_writer, "train/tot_", params.batch_idx_train) + if params.use_fp16: + tb_writer.add_scalar( + "train/grad_scale", + cur_grad_scale, + params.batch_idx_train, + ) + + if batch_idx % params.valid_interval == 0 and not params.print_diagnostics: + logging.info("Computing validation loss") + valid_info = compute_validation_loss( + params=params, + model=model, + sp=sp, + tokenizer=tokenizer, + valid_dl=valid_dl, + world_size=world_size, + ) + model.train() + logging.info(f"Epoch {params.cur_epoch}, validation: {valid_info}") + logging.info( + f"Maximum memory allocated so far is {torch.cuda.max_memory_allocated()//1000000}MB" + ) + if tb_writer is not None: + valid_info.write_summary( + tb_writer, "train/valid_", params.batch_idx_train + ) + + loss_value = tot_loss["loss"] / tot_loss["frames"] + params.train_loss = loss_value + if params.train_loss < params.best_train_loss: + params.best_train_epoch = params.cur_epoch + params.best_train_loss = params.train_loss + + +def run(rank, world_size, args): + """ + Args: + rank: + It is a value between 0 and `world_size-1`, which is + passed automatically by `mp.spawn()` in :func:`main`. + The node with rank 0 is responsible for saving checkpoint. + world_size: + Number of GPUs for DDP training. + args: + The return value of get_parser().parse_args() + """ + params = get_params() + params.update(vars(args)) + + fix_random_seed(params.seed) + if world_size > 1: + setup_dist(rank, world_size, params.master_port) + + setup_logger(f"{params.exp_dir}/log/log-train") + logging.info("Training started") + + if not params.use_style_prompt: + if params.pre_text_shuffle_prob == 0.0: + logging.info( + f"Pre_text shuffle prob is set to: {params.pre_text_shuffle_prob}" + ) + logging.info( + "If style prompt is not used, you should be careful when shuffling the pre_text within the same batch" + ) + logging.info("Hard set this probability to 0.0!") + params.pre_text_shuffle_prob = 0.0 + + if args.tensorboard and rank == 0: + tb_writer = SummaryWriter(log_dir=f"{params.exp_dir}/tensorboard") + else: + tb_writer = None + + device = torch.device("cpu") + if torch.cuda.is_available(): + device = torch.device("cuda", rank) + logging.info(f"Device: {device}") + + sp = spm.SentencePieceProcessor() + sp.load(params.bpe_model) + + # is defined in local/train_bpe_model.py + params.blank_id = sp.piece_to_id("") + params.vocab_size = sp.get_piece_size() + + logging.info(params) + + logging.info("About to create model") + model = get_transducer_model(params) + tokenizer = get_tokenizer(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) + + if params.freeze_text_encoder: + freeze_modules = ["text_encoder"] + logging.info( + "Freeze the parameters of text encoder and don't include them in the optimizer" + ) + else: + freeze_modules = [] + + optimizer = ScaledAdam( + get_parameter_groups_with_lrs( + model, lr=params.base_lr, include_names=True, freeze_modules=freeze_modules + ), + lr=params.base_lr, # should have no effect + clipping_scale=2.0, + ) + + scheduler = Eden(optimizer, params.lr_batches, params.lr_epochs) + + if checkpoints and "optimizer" in checkpoints: + logging.info("Loading optimizer state dict") + optimizer.load_state_dict(checkpoints["optimizer"]) + + if ( + checkpoints + and "scheduler" in checkpoints + and checkpoints["scheduler"] is not None + ): + logging.info("Loading scheduler state dict") + scheduler.load_state_dict(checkpoints["scheduler"]) + + if params.print_diagnostics: + args.max_duration = 100 + 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) + + libriheavy = LibriHeavyAsrDataModule(args) + + train_cuts = libriheavy.train_cuts() + + def remove_short_and_long_utt(c: Cut): + # Keep only utterances with duration between 1 second and 20 seconds + # + # Caution: There is a reason to select 20.0 here. Please see + # ../local/display_manifest_statistics.py + # + # You should use ../local/display_manifest_statistics.py to get + # an utterance duration distribution for your dataset to select + # the threshold + if c.duration < 1.0 or c.duration > 30.0: + return False + + # In pruned RNN-T, we require that T >= S + # where T is the number of feature frames after subsampling + # and S is the number of tokens in the utterance + + # In ./zipformer.py, the conv module uses the following expression + # for subsampling + T = ((c.num_frames - 7) // 2 + 1) // 2 + tokens = sp.encode(c.supervisions[0].texts[0], out_type=str) + + if T < len(tokens): + logging.warning( + f"Exclude cut with ID {c.id} from training. " + f"Number of frames (before subsampling): {c.num_frames}. " + f"Number of frames (after subsampling): {T}. " + f"Text: {c.supervisions[0].texts[0]}. " + f"Tokens: {tokens}. " + f"Number of tokens: {len(tokens)}" + ) + return False + + return True + + train_cuts = train_cuts.filter(remove_short_and_long_utt) + + if params.start_batch > 0 and checkpoints and "sampler" in checkpoints: + # We only load the sampler's state dict when it loads a checkpoint + # saved in the middle of an epoch + sampler_state_dict = checkpoints["sampler"] + else: + sampler_state_dict = None + + if params.use_context_list: + text_sampling_func = triplet_text_sampling_with_context_list + else: + text_sampling_func = triplet_text_sampling + + logging.info(f"Text sampling: {text_sampling_func}") + + train_dl = libriheavy.train_dataloaders( + train_cuts, + sampler_state_dict=sampler_state_dict, + text_sampling_func=text_sampling_func, + ) + + # For fair comparison, use fixed sampling in valid dataloaders + valid_cuts = libriheavy.dev_cuts() + valid_dl = libriheavy.valid_dataloaders( + valid_cuts, text_sampling_func=naive_triplet_text_sampling + ) + + if not params.print_diagnostics: + scan_pessimistic_batches_for_oom( + model=model, + train_dl=train_dl, + optimizer=optimizer, + sp=sp, + tokenizer=tokenizer, + params=params, + ) + + scaler = GradScaler(enabled=params.use_fp16, init_scale=1.0) + if checkpoints and "grad_scaler" in checkpoints: + logging.info("Loading grad scaler state dict") + scaler.load_state_dict(checkpoints["grad_scaler"]) + + for epoch in range(params.start_epoch, params.num_epochs + 1): + scheduler.step_epoch(epoch - 1) + fix_random_seed(params.seed + epoch - 1) + train_dl.sampler.set_epoch(epoch - 1) + + if tb_writer is not None: + tb_writer.add_scalar("train/epoch", epoch, params.batch_idx_train) + + params.cur_epoch = epoch + + train_one_epoch( + params=params, + model=model, + model_avg=model_avg, + optimizer=optimizer, + scheduler=scheduler, + sp=sp, + tokenizer=tokenizer, + train_dl=train_dl, + valid_dl=valid_dl, + scaler=scaler, + tb_writer=tb_writer, + world_size=world_size, + rank=rank, + ) + + if params.print_diagnostics: + diagnostic.print_diagnostics() + break + + save_checkpoint( + params=params, + model=model, + model_avg=model_avg, + optimizer=optimizer, + scheduler=scheduler, + sampler=train_dl.sampler, + scaler=scaler, + rank=rank, + ) + + logging.info("Done!") + + if world_size > 1: + torch.distributed.barrier() + cleanup_dist() + + +def display_and_save_batch( + batch: dict, + params: AttributeDict, + sp: spm.SentencePieceProcessor, +) -> None: + """Display the batch statistics and save the batch into disk. + + Args: + batch: + A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()` + for the content in it. + params: + Parameters for training. See :func:`get_params`. + sp: + The BPE model. + """ + from lhotse.utils import uuid4 + + filename = f"{params.exp_dir}/batch-{uuid4()}.pt" + logging.info(f"Saving batch to {filename}") + torch.save(batch, filename) + + supervisions = batch["supervisions"] + features = batch["inputs"] + + logging.info(f"features shape: {features.shape}") + + y = sp.encode(supervisions["text"], out_type=int) + num_tokens = sum(len(i) for i in y) + logging.info(f"num tokens: {num_tokens}") + + +def scan_pessimistic_batches_for_oom( + model: Union[nn.Module, DDP], + train_dl: torch.utils.data.DataLoader, + optimizer: torch.optim.Optimizer, + sp: spm.SentencePieceProcessor, + tokenizer: spm.SentencePieceProcessor, + params: AttributeDict, +): + from lhotse.dataset import find_pessimistic_batches + + logging.info( + "Sanity check -- see if any of the batches in epoch 1 would cause OOM." + ) + batches, crit_values = find_pessimistic_batches(train_dl.sampler) + for criterion, cuts in batches.items(): + batch = train_dl.dataset[cuts] + try: + with torch.cuda.amp.autocast(enabled=params.use_fp16): + loss, _ = compute_loss( + params=params, + model=model, + sp=sp, + tokenizer=tokenizer, + batch=batch, + is_training=True, + ) + loss.backward() + optimizer.zero_grad() + except Exception as e: + if "CUDA out of memory" in str(e): + logging.error( + "Your GPU ran out of memory with the current " + "max_duration setting. We recommend decreasing " + "max_duration and trying again.\n" + f"Failing criterion: {criterion} " + f"(={crit_values[criterion]}) ..." + ) + display_and_save_batch(batch, params=params, sp=sp) + raise + logging.info( + f"Maximum memory allocated so far is {torch.cuda.max_memory_allocated()//1000000}MB" + ) + + +def main(): + parser = get_parser() + LibriHeavyAsrDataModule.add_arguments(parser) + args = parser.parse_args() + args.exp_dir = Path(args.exp_dir) + + world_size = args.world_size + assert world_size >= 1 + if world_size > 1: + mp.spawn(run, args=(world_size, args), nprocs=world_size, join=True) + else: + run(rank=0, world_size=1, args=args) + + +torch.set_num_threads(1) +torch.set_num_interop_threads(1) + +if __name__ == "__main__": + main() diff --git a/egs/libriheavy/ASR/zipformer_prompt_asr/transcribe_bert.py b/egs/libriheavy/ASR/zipformer_prompt_asr/transcribe_bert.py new file mode 100644 index 000000000..ef0c48e8a --- /dev/null +++ b/egs/libriheavy/ASR/zipformer_prompt_asr/transcribe_bert.py @@ -0,0 +1,515 @@ +# Copyright 2023 Xiaomi Corp. (authors: Xiaoyu Yang) +# +# See ../../../../LICENSE for clarification regarding multiple authors +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +""" +Usage: + +python ./zipformer_prompt_asr/transcribe_bert.py \ + --epoch 50 \ + --avg 10 \ + --exp-dir ./zipformer_prompt_asr/exp \ + --manifest-dir data/long_audios/long_audio.jsonl.gz \ + --pre-text-transform mixed-punc \ + --style-text-transform mixed-punc \ + --num-history 5 \ + --use-pre-text True \ + --use-gt-pre-text False + + +""" + +import argparse +import logging +import math +import warnings +from pathlib import Path +from typing import List + +import k2 +import kaldifeat +import sentencepiece as spm +import torch +import torchaudio +from beam_search import ( + beam_search, + fast_beam_search_one_best, + greedy_search, + greedy_search_batch, + modified_beam_search, +) +from decode_bert import _apply_style_transform +from lhotse import Fbank, load_manifest +from text_normalization import ( + lower_all_char, + lower_only_alpha, + ref_text_normalization, + remove_non_alphabetic, + train_text_normalization, + upper_all_char, + upper_only_alpha, +) +from tqdm import tqdm +from train_bert_encoder import ( + _encode_texts_as_bytes_with_tokenizer, + add_model_arguments, + get_params, + get_tokenizer, + get_transducer_model, +) + +from icefall.checkpoint import ( + average_checkpoints, + average_checkpoints_with_averaged_model, + find_checkpoints, + load_checkpoint, +) +from icefall.utils import ( + AttributeDict, + setup_logger, + store_transcripts, + str2bool, + write_error_stats, +) + + +def get_parser(): + parser = argparse.ArgumentParser( + formatter_class=argparse.ArgumentDefaultsHelpFormatter + ) + + parser.add_argument( + "--epoch", + type=int, + default=30, + help="""It specifies the checkpoint to use for decoding. + Note: Epoch counts from 1. + You can specify --avg to use more checkpoints for model averaging.""", + ) + + parser.add_argument( + "--iter", + type=int, + default=0, + help="""If positive, --epoch is ignored and it + will use the checkpoint exp_dir/checkpoint-iter.pt. + You can specify --avg to use more checkpoints for model averaging. + """, + ) + + parser.add_argument( + "--avg", + type=int, + default=9, + help="Number of checkpoints to average. Automatically select " + "consecutive checkpoints before the checkpoint specified by " + "'--epoch' and '--iter'", + ) + + parser.add_argument( + "--exp-dir", + type=str, + default="pruned_transducer_stateless7/exp", + help="The experiment dir", + ) + + parser.add_argument( + "--bpe-model", + type=str, + default="data/lang_bpe_500/bpe.model", + help="""Path to bpe.model.""", + ) + + parser.add_argument( + "--method", + type=str, + default="greedy_search", + help="""Possible values are: + - greedy_search + - beam_search + - modified_beam_search + - fast_beam_search + """, + ) + + parser.add_argument( + "--beam-size", + type=int, + default=4, + ) + + parser.add_argument( + "--manifest-dir", + type=str, + default="data/long_audios/long_audio.jsonl.gz", + help="""This is the manfiest for long audio transcription. + The cust are intended to be sorted, i.e first sort by recording ID and + then sort by start timestamp""", + ) + + parser.add_argument( + "--use-pre-text", + type=str2bool, + default=False, + help="Whether use pre-text when decoding the current chunk", + ) + + parser.add_argument( + "--use-style-prompt", + type=str2bool, + default=True, + help="Use style prompt when evaluation", + ) + + parser.add_argument( + "--pre-text-transform", + type=str, + choices=["mixed-punc", "upper-no-punc", "lower-no-punc", "lower-punc"], + default="mixed-punc", + help="The style of content prompt, i.e pre_text", + ) + + parser.add_argument( + "--style-text-transform", + type=str, + choices=["mixed-punc", "upper-no-punc", "lower-no-punc", "lower-punc"], + default="mixed-punc", + help="The style of style prompt, i.e style_text", + ) + + parser.add_argument( + "--num-history", + type=int, + default=2, + help="How many previous chunks to look if using pre-text for decoding", + ) + + parser.add_argument( + "--use-gt-pre-text", + type=str2bool, + default=False, + help="Whether use gt pre text when using content prompt", + ) + + parser.add_argument( + "--post-normalization", + type=str2bool, + default=True, + ) + + add_model_arguments(parser) + + return parser + + +@torch.no_grad() +def main(): + parser = get_parser() + args = parser.parse_args() + args.exp_dir = Path(args.exp_dir) + + params = get_params() + + params.update(vars(args)) + + sp = spm.SentencePieceProcessor() + sp.load(params.bpe_model) + + # is defined in local/train_bpe_model.py + params.blank_id = sp.piece_to_id("") + params.unk_id = sp.piece_to_id("") + params.vocab_size = sp.get_piece_size() + + params.res_dir = params.exp_dir / "long_audio_transcribe" + params.res_dir.mkdir(exist_ok=True) + + if params.iter > 0: + params.suffix = f"iter-{params.iter}-avg-{params.avg}" + else: + params.suffix = f"epoch-{params.epoch}-avg-{params.avg}" + + if "beam_search" in params.method: + params.suffix += f"-{params.method}-beam-size-{params.beam_size}" + + if params.use_pre_text: + if params.use_gt_pre_text: + params.suffix += f"-use-gt-pre-text-{params.pre_text_transform}-history-{params.num_history}" + else: + params.suffix += ( + f"-pre-text-{params.pre_text_transform}-history-{params.num_history}" + ) + + book_name = params.manifest_dir.split("/")[-1].replace(".jsonl.gz", "") + setup_logger( + f"{params.res_dir}/log-decode-{book_name}-{params.suffix}", log_level="info" + ) + logging.info("Decoding started") + + device = torch.device("cpu") + if torch.cuda.is_available(): + device = torch.device("cuda", 0) + + logging.info(f"device: {device}") + + logging.info("Creating model") + model = get_transducer_model(params) + tokenizer = get_tokenizer(params) + + num_param = sum([p.numel() for p in model.parameters()]) + logging.info(f"Number of model parameters: {num_param}") + + 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() + model.device = device + + # load manifest + manifest = load_manifest(params.manifest_dir) + + results = [] + count = 0 + + last_recording = "" + last_end = -1 + history = [] + num_pre_texts = [] + + for cut in manifest: + if cut.has_features: + feat = cut.load_features() + feat_lens = cut.num_frames + else: + feat = cut.compute_features(extractor=Fbank()) + feat_lens = feat.shape[0] + + cur_recording = cut.recording.id + + if cur_recording != last_recording: + last_recording = cur_recording + history = [] # clean up the history + last_end = -1 + logging.info("Moving on to the next recording") + else: + if cut.start < last_end - 0.2: # overlap with the previous cuts + logging.warning("An overlap exists between current cut and last cut") + logging.warning("Skipping this cut!") + continue + if cut.start > last_end + 10: + logging.warning( + f"Large time gap between the current and previous utterance: {cut.start - last_end}." + ) + + # prepare input + x = torch.tensor(feat, device=device).unsqueeze(0) + x_lens = torch.tensor( + [ + feat_lens, + ], + device=device, + ) + + if params.use_pre_text: + if params.num_history > 0: + pre_texts = history[-params.num_history :] + else: + pre_texts = [] + num_pre_texts.append(len(pre_texts)) + pre_texts = [train_text_normalization(" ".join(pre_texts))] + fixed_sentence = "Mixed-case English transcription, with punctuation. Actually, it is fully not related." + style_texts = [fixed_sentence] + + pre_texts = _apply_style_transform(pre_texts, params.pre_text_transform) + if params.use_style_prompt: + style_texts = _apply_style_transform( + style_texts, params.style_text_transform + ) + + # encode prompts + with warnings.catch_warnings(): + warnings.simplefilter("ignore") + + encoded_inputs, style_lens = _encode_texts_as_bytes_with_tokenizer( + pre_texts=pre_texts, + style_texts=style_texts, + tokenizer=tokenizer, + device=device, + no_limit=True, + ) + if params.num_history > 5: + logging.info( + f"Shape of encoded texts: {encoded_inputs['input_ids'].shape} " + ) + + memory, memory_key_padding_mask = model.encode_text( + encoded_inputs=encoded_inputs, + style_lens=style_lens, + ) # (T,B,C) + else: + memory = None + memory_key_padding_mask = None + + with warnings.catch_warnings(): + warnings.simplefilter("ignore") + encoder_out, encoder_out_lens = model.encode_audio( + feature=x, + feature_lens=x_lens, + memory=memory, + memory_key_padding_mask=memory_key_padding_mask, + ) + + if params.method == "greedy_search": + hyp_tokens = greedy_search_batch( + model=model, + encoder_out=encoder_out, + encoder_out_lens=encoder_out_lens, + ) + elif params.method == "modified_beam_search": + hyp_tokens = modified_beam_search( + model=model, + encoder_out=encoder_out, + encoder_out_lens=encoder_out_lens, + beam=params.beam_size, + ) + + hyp = sp.decode(hyp_tokens)[0] # in string format + ref_text = ref_text_normalization( + cut.supervisions[0].texts[0] + ) # required to match the training + + # extend the history + if params.use_gt_pre_text: + history.append(ref_text) + else: + history.append(hyp) + last_end = cut.end # update the last end timestamp + + # append the current decoding result + hyp = hyp.split() + ref = ref_text.split() + results.append((cut.id, ref, hyp)) + + count += 1 + if count % 100 == 0: + logging.info(f"Cuts processed until now: {count}/{len(manifest)}") + logging.info( + f"Averaged context numbers of last 100 samples is: {sum(num_pre_texts[-100:])/100}" + ) + + logging.info(f"A total of {count} cuts") + logging.info( + f"Averaged context numbers of whole set is: {sum(num_pre_texts)/len(num_pre_texts)}" + ) + + results = sorted(results) + recog_path = ( + params.res_dir / f"recogs-long-audio-{params.method}-{params.suffix}.txt" + ) + store_transcripts(filename=recog_path, texts=results) + logging.info(f"The transcripts are stored in {recog_path}") + + errs_filename = ( + params.res_dir / f"errs-long-audio-{params.method}-{params.suffix}.txt" + ) + with open(errs_filename, "w") as f: + wer = write_error_stats( + f, + f"long-audio-{params.method}", + results, + enable_log=True, + compute_CER=False, + ) + + logging.info("Wrote detailed error stats to {}".format(errs_filename)) + + if params.post_normalization: + params.suffix += "-post-normalization" + + new_res = [] + for item in results: + id, ref, hyp = item + hyp = upper_only_alpha(" ".join(hyp)).split() + ref = upper_only_alpha(" ".join(ref)).split() + new_res.append((id, ref, hyp)) + + new_res = sorted(new_res) + recog_path = ( + params.res_dir + / f"recogs-long-audio-{params.method}-{params.suffix}-post-normalization.txt" + ) + store_transcripts(filename=recog_path, texts=new_res) + logging.info(f"The transcripts are stored in {recog_path}") + + errs_filename = ( + params.res_dir + / f"errs-long-audio-{params.method}-{params.suffix}-post-normalization.txt" + ) + with open(errs_filename, "w") as f: + wer = write_error_stats( + f, + f"long-audio-{params.method}", + new_res, + enable_log=True, + compute_CER=False, + ) + + logging.info("Wrote detailed error stats to {}".format(errs_filename)) + + +if __name__ == "__main__": + main() diff --git a/egs/libriheavy/ASR/zipformer_prompt_asr/utils.py b/egs/libriheavy/ASR/zipformer_prompt_asr/utils.py new file mode 100644 index 000000000..533982519 --- /dev/null +++ b/egs/libriheavy/ASR/zipformer_prompt_asr/utils.py @@ -0,0 +1,439 @@ +import argparse +import ast +import glob +import logging +import os +from collections import defaultdict +from typing import Dict, Iterable, List, TextIO, Tuple, Union + +import kaldialign +from lhotse import load_manifest, load_manifest_lazy +from lhotse.cut import Cut, CutSet +from text_normalization import remove_non_alphabetic +from tqdm import tqdm + + +def get_parser(): + parser = argparse.ArgumentParser( + formatter_class=argparse.ArgumentDefaultsHelpFormatter + ) + + parser.add_argument( + "--manifest-dir", + type=str, + default="data/fbank", + help="Where are the manifest stored", + ) + + parser.add_argument( + "--subset", type=str, default="medium", help="Which subset to work with" + ) + + parser.add_argument( + "--top-k", + type=int, + default=10000, + help="How many words to keep", + ) + + return parser + + +def get_facebook_biasing_list( + test_set: str, + num_distractors: int = 100, +) -> Dict: + # Get the biasing list from the meta paper: https://arxiv.org/pdf/2104.02194.pdf + assert num_distractors in (0, 100, 500, 1000, 2000), num_distractors + if num_distractors == 0: + if test_set == "test-clean": + biasing_file = "data/context_biasing/fbai-speech/is21_deep_bias/ref/test-clean.biasing_100.tsv" + elif test_set == "test-other": + biasing_file = "data/context_biasing/fbai-speech/is21_deep_bias/ref/test-other.biasing_100.tsv" + else: + raise ValueError(f"Unseen test set {test_set}") + else: + if test_set == "test-clean": + biasing_file = f"data/context_biasing/fbai-speech/is21_deep_bias/ref/test-clean.biasing_{num_distractors}.tsv" + elif test_set == "test-other": + biasing_file = f"data/context_biasing/fbai-speech/is21_deep_bias/ref/test-other.biasing_{num_distractors}.tsv" + else: + raise ValueError(f"Unseen test set {test_set}") + + f = open(biasing_file, "r") + data = f.readlines() + f.close() + + output = dict() + for line in data: + id, _, l1, l2 = line.split("\t") + if num_distractors > 0: # use distractors + biasing_list = ast.literal_eval(l2) + else: + biasing_list = ast.literal_eval(l1) + biasing_list = [w.strip().upper() for w in biasing_list] + output[id] = " ".join(biasing_list) + + return output + + +def brian_biasing_list(level: str): + # The biasing list from Brian's paper: https://arxiv.org/pdf/2109.00627.pdf + root_dir = f"data/context_biasing/LibriSpeechBiasingLists/{level}Level" + all_files = glob.glob(root_dir + "/*") + biasing_dict = {} + for f in all_files: + k = f.split("/")[-1] + fin = open(f, "r") + data = fin.read().strip().split() + biasing_dict[k] = " ".join(data) + fin.close() + + return biasing_dict + + +def get_rare_words( + subset: str = "medium", + top_k: int = 10000, + # min_count: int = 10000, +): + """Get a list of rare words appearing less than `min_count` times + + Args: + subset: The dataset + top_k (int): How many frequent words + """ + txt_path = f"data/tmp/transcript_words_{subset}.txt" + rare_word_file = f"data/context_biasing/{subset}_rare_words_topk_{top_k}.txt" + + if os.path.exists(rare_word_file): + print("File exists, do not proceed!") + return + + print("---Identifying rare words in the manifest---") + count_file = f"data/tmp/transcript_words_{subset}_count.txt" + if not os.path.exists(count_file): + with open(txt_path, "r") as file: + words = file.read().upper().split() + word_count = {} + for word in words: + word = remove_non_alphabetic(word, strict=False) + word = word.split() + for w in word: + if w not in word_count: + word_count[w] = 1 + else: + word_count[w] += 1 + + word_count = list(word_count.items()) # convert to a list of tuple + word_count = sorted(word_count, key=lambda w: int(w[1]), reverse=True) + with open(count_file, "w") as fout: + for w, count in word_count: + fout.write(f"{w}\t{count}\n") + + else: + word_count = {} + with open(count_file, "r") as fin: + word_count = fin.read().strip().split("\n") + word_count = [pair.split("\t") for pair in word_count] + word_count = sorted(word_count, key=lambda w: int(w[1]), reverse=True) + + print(f"A total of {len(word_count)} words appeared!") + rare_words = [] + for word, count in word_count[top_k:]: + rare_words.append(word + "\n") + print(f"A total of {len(rare_words)} are identified as rare words.") + + with open(rare_word_file, "w") as f: + f.writelines(rare_words) + + +def add_context_list_to_manifest( + manifest_dir: str, + subset: str = "medium", + top_k: int = 10000, +): + """Generate a context list of rare words for each utterance in the manifest + + Args: + manifest_dir: Where to store the manifest with context list + subset (str): Subset + top_k (int): How many frequent words + + """ + orig_manifest_dir = f"{manifest_dir}/libriheavy_cuts_{subset}.jsonl.gz" + target_manifest_dir = orig_manifest_dir.replace( + ".jsonl.gz", f"_with_context_list_topk_{top_k}.jsonl.gz" + ) + if os.path.exists(target_manifest_dir): + print(f"Target file exits at {target_manifest_dir}!") + return + + rare_words_file = f"data/context_biasing/{subset}_rare_words_topk_{top_k}.txt" + print(f"---Reading rare words from {rare_words_file}---") + with open(rare_words_file, "r") as f: + rare_words = f.read() + rare_words = rare_words.split("\n") + rare_words = set(rare_words) + print(f"A total of {len(rare_words)} rare words!") + + cuts = load_manifest_lazy(orig_manifest_dir) + print(f"Loaded manifest from {orig_manifest_dir}") + + def _add_context(c: Cut): + splits = ( + remove_non_alphabetic(c.supervisions[0].texts[0], strict=False) + .upper() + .split() + ) + found = [] + for w in splits: + if w in rare_words: + found.append(w) + c.supervisions[0].context_list = " ".join(found) + return c + + cuts = cuts.map(_add_context) + print(f"---Saving manifest with context list to {target_manifest_dir}---") + cuts.to_file(target_manifest_dir) + print("Finished") + + +def check( + manifest_dir: str, + subset: str = "medium", + top_k: int = 10000, +): + # Show how many samples in the training set have a context list + # and the average length of context list + print("--- Calculating the stats over the manifest ---") + + manifest_dir = f"{manifest_dir}/libriheavy_cuts_{subset}_with_context_list_topk_{top_k}.jsonl.gz" + cuts = load_manifest_lazy(manifest_dir) + total_cuts = len(cuts) + has_context_list = [c.supervisions[0].context_list != "" for c in cuts] + context_list_len = [len(c.supervisions[0].context_list.split()) for c in cuts] + print(f"{sum(has_context_list)}/{total_cuts} cuts have context list! ") + print( + f"Average length of non-empty context list is {sum(context_list_len)/sum(has_context_list)}" + ) + + +def write_error_stats( + f: TextIO, + test_set_name: str, + results: List[Tuple[str, str]], + enable_log: bool = True, + compute_CER: bool = False, + biasing_words: List[str] = None, +) -> float: + """Write statistics based on predicted results and reference transcripts. It also calculates the + biasing word error rate as described in https://arxiv.org/pdf/2104.02194.pdf + + It will write the following to the given file: + + - WER + - number of insertions, deletions, substitutions, corrects and total + reference words. For example:: + + Errors: 23 insertions, 57 deletions, 212 substitutions, over 2606 + reference words (2337 correct) + + - The difference between the reference transcript and predicted result. + An instance is given below:: + + THE ASSOCIATION OF (EDISON->ADDISON) ILLUMINATING COMPANIES + + The above example shows that the reference word is `EDISON`, + but it is predicted to `ADDISON` (a substitution error). + + Another example is:: + + FOR THE FIRST DAY (SIR->*) I THINK + + The reference word `SIR` is missing in the predicted + results (a deletion error). + results: + An iterable of tuples. The first element is the cut_id, the second is + the reference transcript and the third element is the predicted result. + enable_log: + If True, also print detailed WER to the console. + Otherwise, it is written only to the given file. + biasing_words: + All the words in the biasing list + Returns: + Return None. + """ + subs: Dict[Tuple[str, str], int] = defaultdict(int) + ins: Dict[str, int] = defaultdict(int) + dels: Dict[str, int] = defaultdict(int) + + # `words` stores counts per word, as follows: + # corr, ref_sub, hyp_sub, ins, dels + words: Dict[str, List[int]] = defaultdict(lambda: [0, 0, 0, 0, 0]) + num_corr = 0 + ERR = "*" + + if compute_CER: + for i, res in enumerate(results): + cut_id, ref, hyp = res + ref = list("".join(ref)) + hyp = list("".join(hyp)) + results[i] = (cut_id, ref, hyp) + + for cut_id, ref, hyp in results: + ali = kaldialign.align(ref, hyp, ERR) + for ref_word, hyp_word in ali: + if ref_word == ERR: # INSERTION + ins[hyp_word] += 1 + words[hyp_word][3] += 1 + elif hyp_word == ERR: # DELETION + dels[ref_word] += 1 + words[ref_word][4] += 1 + elif hyp_word != ref_word: # SUBSTITUTION + subs[(ref_word, hyp_word)] += 1 + words[ref_word][1] += 1 + words[hyp_word][2] += 1 + else: + words[ref_word][0] += 1 + num_corr += 1 + ref_len = sum([len(r) for _, r, _ in results]) + sub_errs = sum(subs.values()) + ins_errs = sum(ins.values()) + del_errs = sum(dels.values()) + tot_errs = sub_errs + ins_errs + del_errs + tot_err_rate = "%.2f" % (100.0 * tot_errs / ref_len) + + if enable_log: + logging.info( + f"[{test_set_name}] %WER {tot_errs / ref_len:.2%} " + f"[{tot_errs} / {ref_len}, {ins_errs} ins, " + f"{del_errs} del, {sub_errs} sub ]" + ) + + print(f"%WER = {tot_err_rate}", file=f) + print( + f"Errors: {ins_errs} insertions, {del_errs} deletions, " + f"{sub_errs} substitutions, over {ref_len} reference " + f"words ({num_corr} correct)", + file=f, + ) + print( + "Search below for sections starting with PER-UTT DETAILS:, " + "SUBSTITUTIONS:, DELETIONS:, INSERTIONS:, PER-WORD STATS:", + file=f, + ) + + print("", file=f) + print("PER-UTT DETAILS: corr or (ref->hyp) ", file=f) + for cut_id, ref, hyp in results: + ali = kaldialign.align(ref, hyp, ERR) + combine_successive_errors = True + if combine_successive_errors: + ali = [[[x], [y]] for x, y in ali] + for i in range(len(ali) - 1): + if ali[i][0] != ali[i][1] and ali[i + 1][0] != ali[i + 1][1]: + ali[i + 1][0] = ali[i][0] + ali[i + 1][0] + ali[i + 1][1] = ali[i][1] + ali[i + 1][1] + ali[i] = [[], []] + ali = [ + [ + list(filter(lambda a: a != ERR, x)), + list(filter(lambda a: a != ERR, y)), + ] + for x, y in ali + ] + ali = list(filter(lambda x: x != [[], []], ali)) + ali = [ + [ + ERR if x == [] else " ".join(x), + ERR if y == [] else " ".join(y), + ] + for x, y in ali + ] + + print( + f"{cut_id}:\t" + + " ".join( + ( + ref_word if ref_word == hyp_word else f"({ref_word}->{hyp_word})" + for ref_word, hyp_word in ali + ) + ), + file=f, + ) + + print("", file=f) + print("SUBSTITUTIONS: count ref -> hyp", file=f) + + for count, (ref, hyp) in sorted([(v, k) for k, v in subs.items()], reverse=True): + print(f"{count} {ref} -> {hyp}", file=f) + + print("", file=f) + print("DELETIONS: count ref", file=f) + for count, ref in sorted([(v, k) for k, v in dels.items()], reverse=True): + print(f"{count} {ref}", file=f) + + print("", file=f) + print("INSERTIONS: count hyp", file=f) + for count, hyp in sorted([(v, k) for k, v in ins.items()], reverse=True): + print(f"{count} {hyp}", file=f) + + unbiased_word_counts = 0 + unbiased_word_errs = 0 + biased_word_counts = 0 + biased_word_errs = 0 + + print("", file=f) + print("PER-WORD STATS: word corr tot_errs count_in_ref count_in_hyp", file=f) + + for _, word, counts in sorted( + [(sum(v[1:]), k, v) for k, v in words.items()], reverse=True + ): + (corr, ref_sub, hyp_sub, ins, dels) = counts + tot_errs = ref_sub + hyp_sub + ins + dels + # number of appearances of "word" in reference text + ref_count = ( + corr + ref_sub + dels + ) # correct + in ref but got substituted + deleted + # number of appearances of "word" in hyp text + hyp_count = corr + hyp_sub + ins + + if biasing_words is not None: + if word in biasing_words: + biased_word_counts += ref_count + biased_word_errs += ins + dels + ref_sub + else: + unbiased_word_counts += ref_count + unbiased_word_errs += ins + dels + hyp_sub + + print(f"{word} {corr} {tot_errs} {ref_count} {hyp_count}", file=f) + + if biasing_words is not None: + B_WER = "%.2f" % (100 * biased_word_errs / biased_word_counts) + U_WER = "%.2f" % (100 * unbiased_word_errs / unbiased_word_counts) + logging.info(f"Biased WER: {B_WER} [{biased_word_errs}/{biased_word_counts}] ") + logging.info( + f"Un-biased WER: {U_WER} [{unbiased_word_errs}/{unbiased_word_counts}]" + ) + + return float(tot_err_rate) + + +if __name__ == "__main__": + parser = get_parser() + args = parser.parse_args() + manifest_dir = args.manifest_dir + subset = args.subset + top_k = args.top_k + get_rare_words(subset=subset, top_k=top_k) + add_context_list_to_manifest( + manifest_dir=manifest_dir, + subset=subset, + top_k=top_k, + ) + check( + manifest_dir=manifest_dir, + subset=subset, + top_k=top_k, + ) diff --git a/egs/libriheavy/ASR/zipformer_prompt_asr/zipformer.py b/egs/libriheavy/ASR/zipformer_prompt_asr/zipformer.py new file mode 100644 index 000000000..d1cf90ffb --- /dev/null +++ b/egs/libriheavy/ASR/zipformer_prompt_asr/zipformer.py @@ -0,0 +1,2310 @@ +#!/usr/bin/env python3 +# Copyright 2022-2023 Xiaomi Corp. (authors: Daniel Povey, +# Zengwei Yao, +# Xiaoyu Yang) +# +# See ../../../../LICENSE for clarification regarding multiple authors +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import copy +import logging +import math +import random +import warnings +from typing import List, Optional, Tuple, Union + +import torch +from encoder_interface import EncoderInterface +from scaling import ( + Identity, # more friendly to backward hooks than nn.Identity(), for diagnostic reasons. +) +from scaling import ( + ScaledLinear, # not as in other dirs.. just scales down initial parameter values. +) +from scaling import ( + ActivationDropoutAndLinear, + Balancer, + BiasNorm, + ChunkCausalDepthwiseConv1d, + Dropout2, + FloatLike, + ScheduledFloat, + Whiten, + convert_num_channels, + limit_param_value, + penalize_abs_values_gt, + softmax, +) +from torch import Tensor, nn + + +class Zipformer2(EncoderInterface): + """ + Args: + + Note: all "int or Tuple[int]" arguments below will be treated as lists of the same length + as downsampling_factor if they are single ints or one-element tuples. The length of + downsampling_factor defines the number of stacks. + + output_downsampling_factor (int): how much to downsample at the output. Note: + we also downsample by a factor of 2 in the Conv2dSubsampling encoder. + You should probably leave this at 2. + downsampling_factor (Tuple[int]): downsampling factor for each encoder stack. + Note: this is in addition to the downsampling factor of 2 that is applied in + the frontend (self.encoder_embed). + encoder_dim (Tuple[int]): embedding dimension of each of the encoder stacks, one per + encoder stack. + num_encoder_layers (int or Tuple[int])): number of encoder layers for each stack + encoder_unmasked_dim (int or Tuple[int]): unmasked dimension in each of + the encoder stacks for purposes of per-frame dropout (recommend 256 for + now). + query_head_dim (int or Tuple[int]): dimension of query and key per attention + head: per stack, if a tuple.. + value_head_dim (int or Tuple[int]): dimension of value in each attention head + pos_head_dim (int or Tuple[int]): dimension of positional-encoding projection per + attention head + num_heads: (int or Tuple[int]): number of heads in the self-attention mechanism. + Must be at least 4. + feedforward_dim (int or Tuple[int]): hidden dimension in feedforward modules + cnn_module_kernel (int or Tuple[int])): Kernel size of convolution module + + pos_dim (int): the dimension of each positional-encoding vector prior to projection, + e.g. 128. + + dropout (float): dropout rate + warmup_batches (float): number of batches to warm up over; this controls + dropout of encoder layers. + causal (bool): if True, support chunkwise causal convolution. This should + not hurt WER as no modeling power is lost, but the convolution modules will be + slightly slower and use more memory. Enables use of the chunk_size and + left_context_chunks options in forward(), which simulates streaming + decoding. + chunk_size: (list of int): only set this to other than [-1] if causal; + the chunk size will be randomly chosen from this list. -1 means no chunking. + left_context_frames: (list of int): determines the number of left- + context chunks for causal training; will be rounded to a number of + chunks. Must not be less than cnn_module_kernel (after factoring in + rounding and downsampling); an error will be thrown if this is violated. + memory_dim: if supplied and >0, will be the dimension of the memory embeddings + passed into the zipformer (e.g. this might be the output of another + Zipformer used to create embedding vectors.) + memory_dropout_rate: By this probability, do not use the provided memory for + cross-attention. This should give robustness to the model when evaluated + without memory. + memory_layer: if supplied and >0, only add cross-attention module starting from + the specified layer. + """ + + def __init__( + self, + output_downsampling_factor: int = 2, + downsampling_factor: Tuple[int] = (2, 4), + encoder_dim: Union[int, Tuple[int]] = 384, + num_encoder_layers: Union[int, Tuple[int]] = 4, + encoder_unmasked_dim: Union[int, Tuple[int]] = 256, + query_head_dim: Union[int, Tuple[int]] = 24, + pos_head_dim: Union[int, Tuple[int]] = 4, + value_head_dim: Union[int, Tuple[int]] = 12, + num_heads: Union[int, Tuple[int]] = 8, + feedforward_dim: Union[int, Tuple[int]] = 1536, + cnn_module_kernel: Union[int, Tuple[int]] = 31, + pos_dim: int = 192, + dropout: FloatLike = None, # see code below for default + warmup_batches: float = 4000.0, + causal: bool = False, + chunk_size: Tuple[int] = [-1], + left_context_frames: Tuple[int] = [-1], + memory_dim: int = -1, + memory_dropout_rate: float = 0.05, + memory_layer: int = 0, + ) -> None: + super(Zipformer2, self).__init__() + + if dropout is None: + dropout = ScheduledFloat((0.0, 0.3), (20000.0, 0.1)) + + def _to_tuple(x): + """Converts a single int or a 1-tuple of an int to a tuple with the same length + as downsampling_factor""" + if isinstance(x, int): + x = (x,) + if len(x) == 1: + x = x * len(downsampling_factor) + else: + assert len(x) == len(downsampling_factor) and isinstance(x[0], int) + return x + + self.output_downsampling_factor = output_downsampling_factor # int + self.downsampling_factor = downsampling_factor # tuple + self.encoder_dim = encoder_dim = _to_tuple(encoder_dim) # tuple + self.encoder_unmasked_dim = encoder_unmasked_dim = _to_tuple( + encoder_unmasked_dim + ) # tuple + num_encoder_layers = _to_tuple(num_encoder_layers) + self.num_encoder_layers = num_encoder_layers + self.query_head_dim = query_head_dim = _to_tuple(query_head_dim) + self.value_head_dim = value_head_dim = _to_tuple(value_head_dim) + pos_head_dim = _to_tuple(pos_head_dim) + self.num_heads = num_heads = _to_tuple(num_heads) + feedforward_dim = _to_tuple(feedforward_dim) + self.cnn_module_kernel = cnn_module_kernel = _to_tuple(cnn_module_kernel) + + self.causal = causal + self.chunk_size = chunk_size + self.left_context_frames = left_context_frames + self.memory_dropout_rate = memory_dropout_rate + self.memory_layer = memory_layer + + for u, d in zip(encoder_unmasked_dim, encoder_dim): + assert u <= d + + # each one will be Zipformer2Encoder or DownsampledZipformer2Encoder + encoders = [] + + num_encoders = len(downsampling_factor) + for i in range(num_encoders): + encoder_layer = Zipformer2EncoderLayer( + embed_dim=encoder_dim[i], + pos_dim=pos_dim, + num_heads=num_heads[i], + query_head_dim=query_head_dim[i], + pos_head_dim=pos_head_dim[i], + value_head_dim=value_head_dim[i], + feedforward_dim=feedforward_dim[i], + memory_dim=memory_dim if i >= self.memory_layer else -1, + dropout=dropout, + cnn_module_kernel=cnn_module_kernel[i], + causal=causal, + ) + + # For the segment of the warmup period, we let the Conv2dSubsampling + # layer learn something. Then we start to warm up the other encoders. + encoder = Zipformer2Encoder( + encoder_layer, + num_encoder_layers[i], + pos_dim=pos_dim, + dropout=dropout, + warmup_begin=warmup_batches * (i + 1) / (num_encoders + 1), + warmup_end=warmup_batches * (i + 2) / (num_encoders + 1), + final_layerdrop_rate=0.035 * (downsampling_factor[i] ** 0.5), + ) + + if downsampling_factor[i] != 1: + encoder = DownsampledZipformer2Encoder( + encoder, + dim=encoder_dim[i], + downsample=downsampling_factor[i], + dropout=dropout, + ) + + encoders.append(encoder) + + self.encoders = nn.ModuleList(encoders) + + self.downsample_output = SimpleDownsample( + max(encoder_dim), downsample=output_downsampling_factor, dropout=dropout + ) + + def get_feature_masks(self, x: Tensor) -> Union[List[float], List[Tensor]]: + """ + In eval mode, returns [1.0] * num_encoders; in training mode, returns a number of + randomized feature masks, one per encoder. + On e.g. 15% of frames, these masks will zero out all enocder dims larger than + some supplied number, e.g. >256, so in effect on those frames we are using + a smaller encoer dim. + + We generate the random masks at this level because we want the 2 masks to 'agree' + all the way up the encoder stack. This will mean that the 1st mask will have + mask values repeated self.zipformer_subsampling_factor times. + + Args: + x: the embeddings (needed for the shape and dtype and device), of shape + (1, batch_size, encoder_dims0) + """ + num_encoders = len(self.encoder_dim) + if not self.training: + return [1.0] * num_encoders + + (num_frames0, batch_size, _encoder_dims0) = x.shape + + assert self.encoder_dim[0] == _encoder_dims0, ( + self.encoder_dim[0], + _encoder_dims0, + ) + + feature_mask_dropout_prob = 0.125 + + # mask1 shape: (1, batch_size, 1) + mask1 = ( + torch.rand(1, batch_size, 1, device=x.device) > feature_mask_dropout_prob + ).to(x.dtype) + + # mask2 has additional sequences masked, about twice the number. + mask2 = torch.logical_and( + mask1, + ( + torch.rand(1, batch_size, 1, device=x.device) + > feature_mask_dropout_prob + ).to(x.dtype), + ) + + # dim: (1, batch_size, 2) + mask = torch.cat((mask1, mask2), dim=-1) + + feature_masks = [] + for i in range(num_encoders): + channels = self.encoder_dim[i] + feature_mask = torch.ones( + 1, batch_size, channels, dtype=x.dtype, device=x.device + ) + u1 = self.encoder_unmasked_dim[i] + u2 = u1 + (channels - u1) // 2 + + feature_mask[:, :, u1:u2] *= mask[..., 0:1] + feature_mask[:, :, u2:] *= mask[..., 1:2] + + feature_masks.append(feature_mask) + + return feature_masks + + def get_chunk_info(self) -> Tuple[int, int]: + """ + Returns chunk_size and left_context_chunks. + """ + if not self.causal: + return -1, -1 + + if torch.jit.is_scripting() or torch.jit.is_tracing(): + assert len(self.chunk_size) == 1, self.chunk_size + chunk_size = self.chunk_size[0] + else: + chunk_size = random.choice(self.chunk_size) + if chunk_size == -1: + left_context_chunks = -1 + else: + if torch.jit.is_scripting() or torch.jit.is_tracing(): + assert len(self.left_context_frames) == 1, self.left_context_frames + left_context_frames = self.left_context_frames[0] + else: + left_context_frames = random.choice(self.left_context_frames) + # Note: in Python, -1 // n == -1 for n > 0 + left_context_chunks = left_context_frames // chunk_size + if left_context_chunks == 0: + left_context_chunks = 1 + + return chunk_size, left_context_chunks + + def forward( + self, + x: torch.Tensor, + x_lens: torch.Tensor, + src_key_padding_mask: Optional[torch.Tensor] = None, + memory: Optional[Tensor] = None, + memory_key_padding_mask: Optional[Tensor] = None, + ) -> 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. + src_key_padding_mask: + The mask for padding, of shape (batch_size, seq_len); True means + masked position. May be None. + memory: optionally, the memory embeddings of shape (memory_len, batch_size, memory_dim) + memory_key_padding_mask: optionally the mask for padding of memory input (for source- + attention), of shape (batch_size, memory_len); True means + masked position. May be None. + + Returns: + Return a tuple containing 2 tensors: + - embeddings: its shape is (batch_size, output_seq_len, max(encoder_dim)) + - lengths, a tensor of shape (batch_size,) containing the number + of frames in `embeddings` before padding. + """ + outputs = [] + if torch.jit.is_scripting() or torch.jit.is_tracing(): + feature_masks = [1.0] * len(self.encoder_dim) + else: + feature_masks = self.get_feature_masks(x) + + chunk_size, left_context_chunks = self.get_chunk_info() + + if torch.jit.is_scripting() or torch.jit.is_tracing(): + # Not support exporting a model for simulating streaming decoding + attn_mask = None + else: + attn_mask = self._get_attn_mask(x, chunk_size, left_context_chunks) + + if self.training and memory is not None: + batch_size = x.shape[1] + # setting memory to zero should be equivalent to not using the + # memory input at all, since the Attention module has no biases. + memory = memory * ( + torch.rand(batch_size, 1, device=memory.device) + > self.memory_dropout_rate + ) + + for i, module in enumerate(self.encoders): + ds = self.downsampling_factor[i] + x = convert_num_channels(x, self.encoder_dim[i]) + + x = module( + x, + chunk_size=chunk_size, + feature_mask=feature_masks[i], + src_key_padding_mask=( + None + if src_key_padding_mask is None + else src_key_padding_mask[..., ::ds] + ), + attn_mask=attn_mask, + memory=memory if i >= self.memory_layer else None, + memory_key_padding_mask=memory_key_padding_mask + if i >= self.memory_layer + else None, + ) + outputs.append(x) + + # if the last output has the largest dimension, x will be unchanged, + # it will be the same as outputs[-1]. Otherwise it will be concatenated + # from different pieces of 'outputs', taking each dimension from the + # most recent output that has it present. + x = self._get_full_dim_output(outputs) + x = self.downsample_output(x) + # class Downsample has this rounding behavior.. + assert self.output_downsampling_factor == 2, self.output_downsampling_factor + if torch.jit.is_scripting() or torch.jit.is_tracing(): + lengths = (x_lens + 1) // 2 + else: + with warnings.catch_warnings(): + warnings.simplefilter("ignore") + lengths = (x_lens + 1) // 2 + + return x, lengths + + def _get_attn_mask( + self, x: Tensor, chunk_size: int, left_context_chunks: int + ) -> Optional[Tensor]: + """ + Return None if chunk_size == -1, else return attention mask of shape + (seq_len, seq_len), interpreted as (tgt_seq_len, src_seq_len). True + means a masked position. + Args: + x: embeddings after self.encoder_embed(), of shape (seq_len, batch_size, embed_dim). + chunk_size: chunk size, must divide + """ + if chunk_size <= 0: + return None + assert all(chunk_size % d == 0 for d in self.downsampling_factor) + if left_context_chunks >= 0: + num_encoders = len(self.encoder_dim) + assert all( + chunk_size * left_context_chunks + >= (self.cnn_module_kernel[i] // 2) * self.downsampling_factor[i] + for i in range(num_encoders) + ) + else: + left_context_chunks = 1000000 + + seq_len = x.shape[0] + + # t is frame index, shape (seq_len,) + t = torch.arange(seq_len, dtype=torch.int32, device=x.device) + # c is chunk index for each frame, shape (seq_len,) + if torch.jit.is_scripting() or torch.jit.is_tracing(): + c = t // chunk_size + else: + with warnings.catch_warnings(): + warnings.simplefilter("ignore") + c = t // chunk_size + src_c = c + tgt_c = c.unsqueeze(-1) + + attn_mask = torch.logical_or(src_c > tgt_c, src_c < tgt_c - left_context_chunks) + if __name__ == "__main__": + logging.info(f"attn_mask = {attn_mask}") + return attn_mask + + def _get_full_dim_output(self, outputs: List[Tensor]): + num_encoders = len(self.encoder_dim) + assert len(outputs) == num_encoders + output_dim = max(self.encoder_dim) + output_pieces = [outputs[-1]] + cur_dim = self.encoder_dim[-1] + for i in range(num_encoders - 2, -1, -1): + d = self.encoder_dim[i] + if d > cur_dim: + this_output = outputs[i] + output_pieces.append(this_output[..., cur_dim:d]) + cur_dim = d + assert cur_dim == output_dim + return torch.cat(output_pieces, dim=-1) + + def streaming_forward( + self, + x: Tensor, + x_lens: Tensor, + states: List[Tensor], + src_key_padding_mask: Tensor, + ) -> Tuple[Tensor, Tensor, List[Tensor]]: + """ + Args: + x: + The input tensor. Its shape is (seq_len, batch_size, feature_dim). + x_lens: + A tensor of shape (batch_size,) containing the number of frames in + `x` before padding. + states: list of cached tensors of all encoder layers. For layer-i, + states[i*6:(i+1)*6] is (cached_key, cached_nonlin_attn, cached_val1, cached_val2, + cached_conv1, cached_conv2). + src_key_padding_mask: + The mask for padding, of shape (batch_size, seq_len); True means + masked position. May be None. + Returns: + Return a tuple containing 2 tensors: + - embeddings: its shape is (output_seq_len, batch_size, max(encoder_dim)) + - lengths, a tensor of shape (batch_size,) containing the number + of frames in `embeddings` before padding. + - updated states + """ + outputs = [] + new_states = [] + layer_offset = 0 + + for i, module in enumerate(self.encoders): + num_layers = module.num_layers + ds = self.downsampling_factor[i] + x = convert_num_channels(x, self.encoder_dim[i]) + + x, new_layer_states = module.streaming_forward( + x, + states=states[layer_offset * 6 : (layer_offset + num_layers) * 6], + left_context_len=self.left_context_frames[0] // ds, + src_key_padding_mask=src_key_padding_mask[..., ::ds], + ) + layer_offset += num_layers + outputs.append(x) + new_states += new_layer_states + + # if the last output has the largest dimension, x will be unchanged, + # it will be the same as outputs[-1]. Otherwise it will be concatenated + # from different pieces of 'outputs', taking each dimension from the + # most recent output that has it present. + x = self._get_full_dim_output(outputs) + x = self.downsample_output(x) + # class Downsample has this rounding behavior.. + assert self.output_downsampling_factor == 2 + if torch.jit.is_scripting() or torch.jit.is_tracing(): + lengths = (x_lens + 1) // 2 + else: + with warnings.catch_warnings(): + warnings.simplefilter("ignore") + lengths = (x_lens + 1) // 2 + + return x, lengths, new_states + + @torch.jit.export + def get_init_states( + self, + batch_size: int = 1, + device: torch.device = torch.device("cpu"), + ) -> List[Tensor]: + """Get initial states. + + A list of cached tensors of all encoder layers. For layer-i, states[i*6:(i+1)*6] + is (cached_key, cached_nonlin_attn, cached_val1, cached_val2, cached_conv1, cached_conv2). + """ + states = [] + for i, module in enumerate(self.encoders): + num_layers = module.num_layers + embed_dim = self.encoder_dim[i] + ds = self.downsampling_factor[i] + num_heads = self.num_heads[i] + key_dim = self.query_head_dim[i] * num_heads + value_dim = self.value_head_dim[i] * num_heads + downsample_left = self.left_context_frames[0] // ds + nonlin_attn_head_dim = 3 * embed_dim // 4 + conv_left_pad = self.cnn_module_kernel[i] // 2 + for layer in range(num_layers): + cached_key = torch.zeros(downsample_left, batch_size, key_dim).to( + device + ) + cached_nonlin_attn = torch.zeros( + 1, batch_size, downsample_left, nonlin_attn_head_dim + ).to(device) + cached_val1 = torch.zeros(downsample_left, batch_size, value_dim).to( + device + ) + cached_val2 = torch.zeros(downsample_left, batch_size, value_dim).to( + device + ) + cached_conv1 = torch.zeros(batch_size, embed_dim, conv_left_pad).to( + device + ) + cached_conv2 = torch.zeros(batch_size, embed_dim, conv_left_pad).to( + device + ) + states += [ + cached_key, + cached_nonlin_attn, + cached_val1, + cached_val2, + cached_conv1, + cached_conv2, + ] + + return states + + +def _whitening_schedule(x: float, ratio: float = 2.0) -> ScheduledFloat: + return ScheduledFloat((0.0, x), (20000.0, ratio * x), default=x) + + +def _balancer_schedule(min_prob: float): + return ScheduledFloat((0.0, 0.4), (8000.0, min_prob)) + + +class Zipformer2EncoderLayer(nn.Module): + """ + Args: + embed_dim: the number of expected features in the input (required). + nhead: the number of heads in the multiheadattention models (required). + feedforward_dim: the dimension of the feedforward network model (default=2048). + dropout: the dropout value (default=0.1). + cnn_module_kernel (int): Kernel size of convolution module. + + Examples:: + >>> encoder_layer = Zipformer2EncoderLayer(embed_dim=512, nhead=8) + >>> src = torch.rand(10, 32, 512) + >>> pos_emb = torch.rand(32, 19, 512) + >>> out = encoder_layer(src, pos_emb) + """ + + def __init__( + self, + embed_dim: int, + pos_dim: int, + num_heads: int, + query_head_dim: int, + pos_head_dim: int, + value_head_dim: int, + feedforward_dim: int, + dropout: FloatLike = 0.1, + cnn_module_kernel: int = 31, + causal: bool = False, + memory_dim: int = -1, + attention_skip_rate: FloatLike = ScheduledFloat( + (0.0, 0.2), (4000.0, 0.05), (16000, 0.0), default=0 + ), + conv_skip_rate: FloatLike = ScheduledFloat( + (0.0, 0.2), (4000.0, 0.05), (16000, 0.0), default=0 + ), + const_attention_rate: FloatLike = ScheduledFloat( + (0.0, 0.25), (4000.0, 0.025), default=0 + ), + ff2_skip_rate: FloatLike = ScheduledFloat( + (0.0, 0.1), (4000.0, 0.01), (50000.0, 0.0) + ), + ff3_skip_rate: FloatLike = ScheduledFloat( + (0.0, 0.1), (4000.0, 0.01), (50000.0, 0.0) + ), + bypass_skip_rate: FloatLike = ScheduledFloat( + (0.0, 0.5), (4000.0, 0.02), default=0 + ), + ) -> None: + super(Zipformer2EncoderLayer, self).__init__() + self.embed_dim = embed_dim + + # self.bypass implements layer skipping as well as bypass; see its default values. + self.bypass = BypassModule( + embed_dim, skip_rate=bypass_skip_rate, straight_through_rate=0 + ) + # bypass_mid is bypass used in the middle of the layer. + self.bypass_mid = BypassModule(embed_dim, straight_through_rate=0) + + # skip probability for dynamic modules (meaning: anything but feedforward). + self.attention_skip_rate = copy.deepcopy(attention_skip_rate) + # an additional skip probability that applies to ConvModule to stop it from + # contributing too much early on. + self.conv_skip_rate = copy.deepcopy(conv_skip_rate) + + # ff2_skip_rate is to prevent the ff2 module from having output that's too big + # compared to its residual. + self.ff2_skip_rate = copy.deepcopy(ff2_skip_rate) + self.ff3_skip_rate = copy.deepcopy(ff3_skip_rate) + + self.const_attention_rate = copy.deepcopy(const_attention_rate) + + self.self_attn_weights = RelPositionMultiheadAttentionWeights( + embed_dim, + pos_dim=pos_dim, + num_heads=num_heads, + query_head_dim=query_head_dim, + pos_head_dim=pos_head_dim, + dropout=0.0, + ) + + self.self_attn1 = Attention(embed_dim, embed_dim, num_heads, value_head_dim) + + self.self_attn2 = Attention(embed_dim, embed_dim, num_heads, value_head_dim) + + if memory_dim > 0: + self.attn_weights = MultiheadAttentionWeights( + memory_dim, + embed_dim, + num_heads=num_heads, + head_dim=query_head_dim, + dropout=0.0, + ) + self.src_attn1 = Attention(memory_dim, embed_dim, num_heads, value_head_dim) + self.src_attn2 = Attention(memory_dim, embed_dim, num_heads, value_head_dim) + self.memory_balancer = Balancer( + embed_dim, + channel_dim=-1, + min_abs=0.015, + ) + + self.feed_forward1 = FeedforwardModule( + embed_dim, (feedforward_dim * 3) // 4, dropout + ) + + self.feed_forward2 = FeedforwardModule(embed_dim, feedforward_dim, dropout) + + self.feed_forward3 = FeedforwardModule( + embed_dim, (feedforward_dim * 5) // 4, dropout + ) + + self.nonlin_attention = NonlinAttention( + embed_dim, hidden_channels=3 * embed_dim // 4 + ) + + self.conv_module1 = ConvolutionModule( + embed_dim, cnn_module_kernel, causal=causal + ) + + self.conv_module2 = ConvolutionModule( + embed_dim, cnn_module_kernel, causal=causal + ) + + # self.attention_squeeze = AttentionSqueeze(embed_dim, embed_dim // 2) + + self.norm = BiasNorm(embed_dim) + + self.bypass_scale = nn.Parameter(torch.full((embed_dim,), 0.5)) + + self.balancer1 = Balancer( + embed_dim, + channel_dim=-1, + min_positive=0.45, + max_positive=0.55, + min_abs=0.2, + max_abs=4.0, + ) + + # balancer for output of NonlinAttentionModule + self.balancer_na = Balancer( + embed_dim, + channel_dim=-1, + min_positive=0.3, + max_positive=0.7, + min_abs=ScheduledFloat((0.0, 0.004), (4000.0, 0.02)), + prob=0.05, # out of concern for memory usage + ) + + # balancer for output of feedforward2, prevent it from staying too + # small. give this a very small probability, even at the start of + # training, it's to fix a rare problem and it's OK to fix it slowly. + self.balancer_ff2 = Balancer( + embed_dim, + channel_dim=-1, + min_positive=0.3, + max_positive=0.7, + min_abs=ScheduledFloat((0.0, 0.0), (4000.0, 0.1), default=0.0), + max_abs=2.0, + prob=0.05, + ) + + self.balancer_ff3 = Balancer( + embed_dim, + channel_dim=-1, + min_positive=0.3, + max_positive=0.7, + min_abs=ScheduledFloat((0.0, 0.0), (4000.0, 0.2), default=0.0), + max_abs=4.0, + prob=0.05, + ) + + self.whiten = Whiten( + num_groups=1, + whitening_limit=_whitening_schedule(4.0, ratio=3.0), + prob=(0.025, 0.25), + grad_scale=0.01, + ) + + self.balancer2 = Balancer( + embed_dim, + channel_dim=-1, + min_positive=0.45, + max_positive=0.55, + min_abs=0.1, + max_abs=4.0, + ) + + def get_bypass_scale(self, batch_size: int): + # returns bypass-scale of shape (num_channels,), + # or (batch_size, num_channels,). This is actually the + # scale on the non-residual term, so 0 correponds to bypassing + # this module. + if torch.jit.is_scripting() or not self.training: + return self.bypass_scale + else: + ans = limit_param_value( + self.bypass_scale, + min=float(self.bypass_min), + max=float(self.bypass_max), + ) + layer_skip_rate = float(self.layer_skip_rate) + if layer_skip_rate != 0.0: + mask = torch.rand((batch_size, 1), device=ans.device) > layer_skip_rate + ans = ans * mask + # now ans is of shape (batch_size, num_channels), and is zero for sequences + # on which we have randomly chosen to do layer-skipping. + return ans + + def get_sequence_dropout_mask( + self, x: Tensor, dropout_rate: float + ) -> Optional[Tensor]: + if dropout_rate == 0.0 or not self.training or torch.jit.is_scripting(): + return None + batch_size = x.shape[1] + mask = (torch.rand(batch_size, 1, device=x.device) > dropout_rate).to(x.dtype) + return mask + + def sequence_dropout(self, x: Tensor, dropout_rate: float) -> Tensor: + """ + Apply sequence-level dropout to x. + x shape: (seq_len, batch_size, embed_dim) + """ + dropout_mask = self.get_sequence_dropout_mask(x, dropout_rate) + if dropout_mask is None: + return x + else: + return x * dropout_mask + + def forward( + self, + src: Tensor, + pos_emb: Tensor, + chunk_size: int = -1, + attn_mask: Optional[Tensor] = None, + src_key_padding_mask: Optional[Tensor] = None, + memory: Optional[Tensor] = None, + memory_key_padding_mask: Optional[Tensor] = None, + ) -> Tensor: + """ + Pass the input through the encoder layer. + Args: + src: the sequence to the encoder (required): shape (seq_len, batch_size, embedding_dim). + pos_emb: (1, 2*seq_len-1, pos_emb_dim) or (batch_size, 2*seq_len-1, pos_emb_dim) + chunk_size: the number of frames per chunk, of >= 0; if -1, no chunking. + feature_mask: something that broadcasts with src, that we'll multiply `src` + by at every layer: if a Tensor, likely of shape (seq_len, batch_size, embedding_dim) + attn_mask: the attention mask, of shape (batch_size, seq_len, seq_len) or (seq_len, seq_len), + interpreted as (batch_size, tgt_seq_len, src_seq_len) or (tgt_seq_len, src_seq_len). + True means masked position. May be None. + src_key_padding_mask: the mask for padding, of shape (batch_size, seq_len); True means + masked position. May be None. + + Returns: + A tensor which has the same shape as src + """ + src_orig = src + + # dropout rate for non-feedforward submodules + attention_skip_rate = float(self.attention_skip_rate) if self.training else 0.0 + + # attn_weights: (num_heads, batch_size, seq_len, seq_len) + attn_weights = self.self_attn_weights( + src, + pos_emb=pos_emb, + attn_mask=attn_mask, + key_padding_mask=src_key_padding_mask, + ) + + if memory is not None and hasattr(self, "attn_weights"): + src_attn_weights = self.attn_weights(memory, src, memory_key_padding_mask) + + src = src + self.feed_forward1(src) + + attn_dropout_mask = self.get_sequence_dropout_mask(src, attention_skip_rate) + + if True: + selected_attn_weights = attn_weights[0:2] + if random.random() < float(self.const_attention_rate): + # Make attention weights constant. The intention is to + # encourage these modules to do something similar to an + # averaging-over-time operation. + # only need the mask, can just use the 1st one and expand later + selected_attn_weights = selected_attn_weights[0:1] + selected_attn_weights = (selected_attn_weights > 0.0).to( + selected_attn_weights.dtype + ) + selected_attn_weights = selected_attn_weights * ( + 1.0 / selected_attn_weights.sum(dim=-1, keepdim=True) + ) + selected_attn_weights = selected_attn_weights.expand(2, -1, -1, -1) + + na = self.balancer_na(self.nonlin_attention(src, selected_attn_weights[0:1])) + + src = src + (na if attn_dropout_mask is None else na * attn_dropout_mask) + + self_attn = self.self_attn1(src, attn_weights) + + src = src + ( + self_attn if attn_dropout_mask is None else self_attn * attn_dropout_mask + ) + + if memory is not None and hasattr(self, "attn_weights"): + src = src + self.sequence_dropout( + self.memory_balancer(self.src_attn1(memory, src_attn_weights)), + attention_skip_rate, + ) + + src = src + self.sequence_dropout( + self.conv_module1( + src, chunk_size=chunk_size, src_key_padding_mask=src_key_padding_mask + ), + float(self.conv_skip_rate), + ) + + src = src + self.sequence_dropout( + self.balancer_ff2(self.feed_forward2(src)), float(self.ff2_skip_rate) + ) + + # bypass in the middle of the layer. + src = self.bypass_mid(src_orig, src) + + self_attn = self.self_attn2(src, attn_weights) + + src = src + ( + self_attn if attn_dropout_mask is None else self_attn * attn_dropout_mask + ) + + if memory is not None and hasattr(self, "attn_weights"): + src = src + self.sequence_dropout( + self.memory_balancer(self.src_attn2(memory, src_attn_weights)), + attention_skip_rate, + ) + + src = src + self.sequence_dropout( + self.conv_module2( + src, chunk_size=chunk_size, src_key_padding_mask=src_key_padding_mask + ), + float(self.conv_skip_rate), + ) + + src = src + self.sequence_dropout( + self.balancer_ff3(self.feed_forward3(src)), float(self.ff3_skip_rate) + ) + + src = self.balancer1(src) + src = self.norm(src) + + src = self.bypass(src_orig, src) + + src = self.balancer2(src) + src = self.whiten(src) + + return src + + +class Zipformer2Encoder(nn.Module): + r"""Zipformer2Encoder is a stack of N encoder layers + + Args: + encoder_layer: an instance of the Zipformer2EncoderLayer() class (required). + num_layers: the number of sub-encoder-layers in the encoder (required). + pos_dim: the dimension for the relative positional encoding + + Examples:: + >>> encoder_layer = Zipformer2EncoderLayer(embed_dim=512, nhead=8) + >>> zipformer_encoder = Zipformer2Encoder(encoder_layer, num_layers=6) + >>> src = torch.rand(10, 32, 512) + >>> out = zipformer_encoder(src) + """ + + def __init__( + self, + encoder_layer: nn.Module, + num_layers: int, + pos_dim: int, + dropout: float, + warmup_begin: float, + warmup_end: float, + initial_layerdrop_rate: float = 0.5, + final_layerdrop_rate: float = 0.05, + ) -> None: + super().__init__() + self.encoder_pos = CompactRelPositionalEncoding( + pos_dim, dropout_rate=0.15, length_factor=1.0 + ) + + self.layers = nn.ModuleList( + [copy.deepcopy(encoder_layer) for i in range(num_layers)] + ) + self.num_layers = num_layers + + assert 0 <= warmup_begin <= warmup_end + + delta = (1.0 / num_layers) * (warmup_end - warmup_begin) + cur_begin = warmup_begin # interpreted as a training batch index + for i in range(num_layers): + cur_end = cur_begin + delta + self.layers[i].bypass.skip_rate = ScheduledFloat( + (cur_begin, initial_layerdrop_rate), + (cur_end, final_layerdrop_rate), + default=0.0, + ) + cur_begin = cur_end + + def forward( + self, + src: Tensor, + chunk_size: int = -1, + feature_mask: Union[Tensor, float] = 1.0, + attn_mask: Optional[Tensor] = None, + src_key_padding_mask: Optional[Tensor] = None, + memory: Optional[Tensor] = None, + memory_key_padding_mask: Optional[Tensor] = None, + ) -> Tensor: + r"""Pass the input through the encoder layers in turn. + + Args: + src: the sequence to the encoder (required): shape (seq_len, batch_size, embedding_dim). + chunk_size: the number of frames per chunk, of >= 0; if -1, no chunking. + feature_mask: something that broadcasts with src, that we'll multiply `src` + by at every layer: if a Tensor, likely of shape (seq_len, batch_size, embedding_dim) + attn_mask: the attention mask, of shape (batch_size, seq_len, seq_len) or (seq_len, seq_len), + interpreted as (batch_size, tgt_seq_len, src_seq_len) or (tgt_seq_len, src_seq_len). + True means masked position. May be None. + src_key_padding_mask: the mask for padding, of shape (batch_size, seq_len); True means + masked position. May be None. + memory: optionally, the memory embeddings of shape (memory_len, batch_size, memory_dim) + memory_key_padding_mask: optionally the mask for padding of memory input (for source- + attention), of shape (batch_size, memory_len); True means + masked position. May be None. + + Returns: a Tensor with the same shape as src. + """ + pos_emb = self.encoder_pos(src) + output = src + + output = output * feature_mask + + for i, mod in enumerate(self.layers): + output = mod( + output, + pos_emb, + chunk_size=chunk_size, + attn_mask=attn_mask, + src_key_padding_mask=src_key_padding_mask, + memory=memory, + memory_key_padding_mask=memory_key_padding_mask, + ) + + output = output * feature_mask + + return output + + +class BypassModule(nn.Module): + """ + An nn.Module that implements a learnable bypass scale, and also randomized per-sequence + layer-skipping. The bypass is limited during early stages of training to be close to + "straight-through", i.e. to not do the bypass operation much initially, in order to + force all the modules to learn something. + """ + + def __init__( + self, + embed_dim: int, + skip_rate: FloatLike = 0.0, + straight_through_rate: FloatLike = 0.0, + scale_min: FloatLike = ScheduledFloat((0.0, 0.9), (20000.0, 0.2), default=0), + scale_max: FloatLike = 1.0, + ): + super().__init__() + self.bypass_scale = nn.Parameter(torch.full((embed_dim,), 0.5)) + self.skip_rate = copy.deepcopy(skip_rate) + self.straight_through_rate = copy.deepcopy(straight_through_rate) + self.scale_min = copy.deepcopy(scale_min) + self.scale_max = copy.deepcopy(scale_max) + + def _get_bypass_scale(self, batch_size: int): + # returns bypass-scale of shape (num_channels,), + # or (batch_size, num_channels,). This is actually the + # scale on the non-residual term, so 0 correponds to bypassing + # this module. + if torch.jit.is_scripting() or not self.training: + return self.bypass_scale + else: + ans = limit_param_value( + self.bypass_scale, min=float(self.scale_min), max=float(self.scale_max) + ) + skip_rate = float(self.skip_rate) + if skip_rate != 0.0: + mask = torch.rand((batch_size, 1), device=ans.device) > skip_rate + ans = ans * mask + # now ans is of shape (batch_size, num_channels), and is zero for sequences + # on which we have randomly chosen to do layer-skipping. + straight_through_rate = float(self.straight_through_rate) + if straight_through_rate != 0.0: + mask = ( + torch.rand((batch_size, 1), device=ans.device) + < straight_through_rate + ) + ans = torch.maximum(ans, mask.to(ans.dtype)) + + return ans + + def forward(self, src_orig: Tensor, src: Tensor): + """ + Args: src_orig and src are both of shape (seq_len, batch_size, num_channels) + Returns: something with the same shape as src and src_orig + """ + bypass_scale = self._get_bypass_scale(src.shape[1]) + return src_orig + (src - src_orig) * bypass_scale + + +class DownsampledZipformer2Encoder(nn.Module): + r""" + DownsampledZipformer2Encoder is a zipformer encoder evaluated at a reduced frame rate, + after convolutional downsampling, and then upsampled again at the output, and combined + with the origin input, so that the output has the same shape as the input. + """ + + def __init__( + self, encoder: nn.Module, dim: int, downsample: int, dropout: FloatLike + ): + super(DownsampledZipformer2Encoder, self).__init__() + self.downsample_factor = downsample + self.downsample = SimpleDownsample(dim, downsample, dropout) + self.encoder = encoder + self.upsample = SimpleUpsample(dim, downsample) + self.out_combiner = BypassModule(dim, straight_through_rate=0.025) + + def forward( + self, + src: Tensor, + chunk_size: int = -1, + feature_mask: Union[Tensor, float] = 1.0, + attn_mask: Optional[Tensor] = None, + src_key_padding_mask: Optional[Tensor] = None, + memory: Optional[Tensor] = None, + memory_key_padding_mask: Optional[Tensor] = None, + ) -> Tuple[Tensor, Tensor]: + r"""Downsample, go through encoder, upsample. + + Args: + src: the sequence to the encoder (required): shape (seq_len, batch_size, embedding_dim). + feature_mask: something that broadcasts with src, that we'll multiply `src` + by at every layer: if a Tensor, likely of shape (seq_len, batch_size, embedding_dim) + attn_mask: the attention mask, of shape (batch_size, seq_len, seq_len) or (seq_len, seq_len), + interpreted as (batch_size, tgt_seq_len, src_seq_len) or (tgt_seq_len, src_seq_len). + True means masked position. May be None. + src_key_padding_mask: the mask for padding, of shape (batch_size, seq_len); True means + masked position. May be None. + memory: optionally, the memory embeddings of shape (memory_len, batch_size, memory_dim) + memory_key_padding_mask: optionally the mask for padding of memory input (for source- + attention), of shape (batch_size, memory_len); True means + masked position. May be None. + + Returns: a Tensor with the same shape as src. + """ + src_orig = src + src = self.downsample(src) + ds = self.downsample_factor + if attn_mask is not None: + attn_mask = attn_mask[::ds, ::ds] + + src = self.encoder( + src, + chunk_size=chunk_size // ds, + feature_mask=feature_mask, + attn_mask=attn_mask, + src_key_padding_mask=src_key_padding_mask, + memory=memory, + memory_key_padding_mask=memory_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) + + +class SimpleDownsample(torch.nn.Module): + """ + Does downsampling with attention, by weighted sum, and a projection.. + """ + + def __init__(self, channels: int, downsample: int, dropout: FloatLike): + super(SimpleDownsample, self).__init__() + + self.bias = nn.Parameter(torch.zeros(downsample)) + + self.name = None # will be set from training code + self.dropout = copy.deepcopy(dropout) + + self.downsample = downsample + + def forward(self, src: Tensor) -> Tensor: + """ + x: (seq_len, batch_size, in_channels) + Returns a tensor of shape + ( (seq_len+downsample-1)//downsample, batch_size, channels) + """ + (seq_len, batch_size, in_channels) = src.shape + ds = self.downsample + d_seq_len = (seq_len + ds - 1) // ds + + # Pad to an exact multiple of self.downsample + 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 = src.reshape(d_seq_len, ds, batch_size, in_channels) + + weights = self.bias.softmax(dim=0) + # weights: (downsample, 1, 1) + weights = weights.unsqueeze(-1).unsqueeze(-1) + + # ans1 is the first `in_channels` channels of the output + ans = (src * weights).sum(dim=1) + + return ans + + +class SimpleUpsample(torch.nn.Module): + """ + A very simple form of upsampling that mostly just repeats the input, but + also adds a position-specific bias. + """ + + def __init__(self, num_channels: int, upsample: int): + super(SimpleUpsample, self).__init__() + self.upsample = upsample + + def forward(self, src: Tensor) -> Tensor: + """ + x: (seq_len, batch_size, num_channels) + Returns a tensor of shape + ( (seq_len*upsample), batch_size, num_channels) + """ + upsample = self.upsample + (seq_len, batch_size, num_channels) = src.shape + src = src.unsqueeze(1).expand(seq_len, upsample, batch_size, num_channels) + src = src.reshape(seq_len * upsample, batch_size, num_channels) + return src + + +class CompactRelPositionalEncoding(torch.nn.Module): + """ + Relative positional encoding module. This version is "compact" meaning it is able to encode + the important information about the relative position in a relatively small number of dimensions. + The goal is to make it so that small differences between large relative offsets (e.g. 1000 vs. 1001) + make very little difference to the embedding. Such differences were potentially important + when encoding absolute position, but not important when encoding relative position because there + is now no need to compare two large offsets with each other. + + Our embedding works done by projecting the interval [-infinity,infinity] to a finite interval + using the atan() function, before doing the fourier transform of that fixed interval. The + atan() function would compress the "long tails" too small, + making it hard to distinguish between different magnitudes of large offsets, so we use a logarithmic + function to compress large offsets to a smaller range before applying atan(). + Scalings are chosen in such a way that the embedding can clearly distinguish invidual offsets as long + as they are quite close to the origin, e.g. abs(offset) <= about sqrt(embedding_dim) + + + Args: + embed_dim: Embedding dimension. + dropout_rate: Dropout rate. + max_len: Maximum input length: just a heuristic for initialization. + length_factor: a heuristic scale (should be >= 1.0) which, if larger, gives + less weight to small differences of offset near the origin. + """ + + def __init__( + self, + embed_dim: int, + dropout_rate: FloatLike, + max_len: int = 1000, + length_factor: float = 1.0, + ) -> None: + """Construct a CompactRelPositionalEncoding object.""" + super(CompactRelPositionalEncoding, self).__init__() + self.embed_dim = embed_dim + assert embed_dim % 2 == 0 + self.dropout = Dropout2(dropout_rate) + self.pe = None + assert length_factor >= 1.0 + self.length_factor = length_factor + self.extend_pe(torch.tensor(0.0).expand(max_len)) + + def extend_pe(self, x: Tensor) -> None: + """Reset the positional encodings.""" + if self.pe is not None: + # self.pe contains both positive and negative parts + # the length of self.pe is 2 * input_len - 1 + if self.pe.size(0) >= x.size(0) * 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 + + T = x.size(0) + # if T == 4, x would contain [ -3, -2, 1, 0, 1, 2, 3 ] + x = torch.arange(-(T - 1), T, device=x.device).to(torch.float32).unsqueeze(1) + + freqs = 1 + torch.arange(self.embed_dim // 2, device=x.device) + + # `compression_length` this is arbitrary/heuristic, if it is larger we have more resolution + # for small time offsets but less resolution for large time offsets. + compression_length = self.embed_dim**0.5 + # x_compressed, like X, goes from -infinity to infinity as T goes from -infinity to infinity; + # but it does so more slowly than T for large absolute values of T. + # The formula is chosen so that d(x_compressed )/dx is 1 around x == 0, which + # is important. + x_compressed = ( + compression_length + * x.sign() + * ((x.abs() + compression_length).log() - math.log(compression_length)) + ) + + # if self.length_factor == 1.0, then length_scale is chosen so that the + # FFT can exactly separate points close to the origin (T == 0). So this + # part of the formulation is not really heuristic. + # But empirically, for ASR at least, length_factor > 1.0 seems to work better. + length_scale = self.length_factor * self.embed_dim / (2.0 * math.pi) + + # note for machine implementations: if atan is not available, we can use: + # x.sign() * ((1 / (x.abs() + 1)) - 1) * (-math.pi/2) + # check on wolframalpha.com: plot(sign(x) * (1 / ( abs(x) + 1) - 1 ) * -pi/2 , atan(x)) + x_atan = (x_compressed / length_scale).atan() # results between -pi and pi + + cosines = (x_atan * freqs).cos() + sines = (x_atan * freqs).sin() + + pe = torch.zeros(x.shape[0], self.embed_dim, device=x.device) + pe[:, 0::2] = cosines + pe[:, 1::2] = sines + pe[:, -1] = 1.0 # for bias. + + self.pe = pe.to(dtype=x.dtype) + + def forward(self, x: torch.Tensor) -> Tensor: + """Create positional encoding. + + Args: + x (torch.Tensor): Input tensor (time, batch, `*`). + + Returns: + positional embedding, of shape (1, 2*time-1, `*`). + + """ + self.extend_pe(x) + pos_emb = self.pe[ + self.pe.size(0) // 2 + - x.size(0) + + 1 : self.pe.size(0) // 2 # noqa E203 + + x.size(0), + :, + ] + pos_emb = pos_emb.unsqueeze(0) + return self.dropout(pos_emb) + + +class RelPositionMultiheadAttentionWeights(nn.Module): + r"""Module that computes multi-head attention weights with relative position encoding. + Various other modules consume the resulting attention weights: see, for example, the + SimpleAttention module which allows you to compute conventional attention. + + This is a quite heavily modified from: "Transformer-XL: Attentive Language Models Beyond a Fixed-Length Context", + we have to write up the differences. + + + Args: + embed_dim: number of channels at the input to this module, e.g. 256 + pos_dim: dimension of the positional encoding vectors, e.g. 128. + num_heads: number of heads to compute weights for, e.g. 8 + query_head_dim: dimension of the query (and key), per head. e.g. 24. + pos_head_dim: dimension of the projected positional encoding per head, e.g. 4. + dropout: dropout probability for attn_output_weights. Default: 0.0. + pos_emb_skip_rate: probability for skipping the pos_emb part of the scores on + any given call to forward(), in training time. + """ + + def __init__( + self, + embed_dim: int, + pos_dim: int, + num_heads: int, + query_head_dim: int, + pos_head_dim: int, + dropout: float = 0.0, + pos_emb_skip_rate: FloatLike = ScheduledFloat((0.0, 0.5), (4000.0, 0.0)), + ) -> None: + super().__init__() + self.embed_dim = embed_dim + self.num_heads = num_heads + self.query_head_dim = query_head_dim + self.pos_head_dim = pos_head_dim + self.dropout = dropout + self.pos_emb_skip_rate = copy.deepcopy(pos_emb_skip_rate) + self.name = None # will be overwritten in training code; for diagnostics. + + key_head_dim = query_head_dim + in_proj_dim = (query_head_dim + key_head_dim + pos_head_dim) * num_heads + + # the initial_scale is supposed to take over the "scaling" factor of + # head_dim ** -0.5 that has been used in previous forms of attention, + # dividing it between the query and key. Note: this module is intended + # to be used with the ScaledAdam optimizer; with most other optimizers, + # it would be necessary to apply the scaling factor in the forward function. + self.in_proj = ScaledLinear( + embed_dim, in_proj_dim, bias=True, initial_scale=query_head_dim**-0.25 + ) + + self.whiten_keys = Whiten( + num_groups=num_heads, + whitening_limit=_whitening_schedule(3.0), + prob=(0.025, 0.25), + grad_scale=0.025, + ) + + # add a balancer for the keys that runs with very small probability, and + # tries to enforce that all dimensions have mean around zero. The + # weights produced by this module are invariant to adding a constant to + # the keys, so the derivative of the bias is mathematically zero; but + # due to how Adam/ScaledAdam work, it can learn a fairly large nonzero + # bias because the small numerical roundoff tends to have a non-random + # sign. This module is intended to prevent that. Use a very small + # probability; that should be suffixient to fix the problem. + self.balance_keys = Balancer( + key_head_dim * num_heads, + channel_dim=-1, + min_positive=0.4, + max_positive=0.6, + min_abs=0.0, + max_abs=100.0, + prob=0.025, + ) + + # linear transformation for positional encoding. + self.linear_pos = ScaledLinear( + pos_dim, num_heads * pos_head_dim, bias=False, initial_scale=0.05 + ) + + # the following are for diagnosics only, see --print-diagnostics option + self.copy_pos_query = Identity() + self.copy_query = Identity() + + def forward( + self, + x: Tensor, + pos_emb: Tensor, + chunk_size: int = -1, + key_padding_mask: Optional[Tensor] = None, + attn_mask: Optional[Tensor] = None, + ) -> Tensor: + r""" + Args: + x: input of shape (seq_len, batch_size, embed_dim) + pos_emb: Positional embedding tensor, of shape (1, 2*seq_len - 2, pos_dim) + chunk_size + key_padding_mask: a bool tensor of shape (batch_size, seq_len). Positions that + are True in this mask will be ignored as sources in the attention weighting. + attn_mask: mask of shape (seq_len, seq_len) or (batch_size, seq_len, seq_len), + interpreted as ([batch_size,] tgt_seq_len, src_seq_len) + saying which positions are allowed to attend to which other positions. + Returns: + a tensor of attention weights, of shape (hum_heads, batch_size, seq_len, seq_len) + interpreted as (hum_heads, batch_size, tgt_seq_len, src_seq_len). + """ + x = self.in_proj(x) + query_head_dim = self.query_head_dim + pos_head_dim = self.pos_head_dim + num_heads = self.num_heads + + seq_len, batch_size, _ = x.shape + + query_dim = query_head_dim * num_heads + + q = x[..., 0:query_dim] + k = x[..., query_dim : 2 * query_dim] + # p is the position-encoding query + p = x[..., 2 * query_dim :] + assert p.shape[-1] == num_heads * pos_head_dim + + q = self.copy_query(q) # for diagnostics only, does nothing. + k = self.whiten_keys(self.balance_keys(k)) # does nothing in the forward pass. + p = self.copy_pos_query(p) # for diagnostics only, does nothing. + + q = q.reshape(seq_len, batch_size, num_heads, query_head_dim) + p = p.reshape(seq_len, batch_size, num_heads, pos_head_dim) + k = k.reshape(seq_len, batch_size, num_heads, query_head_dim) + + # time1 refers to target, time2 refers to source. + q = q.permute(2, 1, 0, 3) # (head, batch, time1, query_head_dim) + p = p.permute(2, 1, 0, 3) # (head, batch, time1, pos_head_dim) + k = k.permute(2, 1, 3, 0) # (head, batch, d_k, time2) + + attn_scores = torch.matmul(q, k) + + if not self.training or random.random() >= float(self.pos_emb_skip_rate): + pos_emb = self.linear_pos(pos_emb) + seq_len2 = 2 * seq_len - 1 + pos_emb = pos_emb.reshape(-1, seq_len2, num_heads, pos_head_dim).permute( + 2, 0, 3, 1 + ) + # pos shape now: (head, {1 or batch_size}, pos_dim, seq_len2) + + # (head, batch, time1, pos_dim) x (head, 1, pos_dim, seq_len2) -> (head, batch, time1, seq_len2) + # [where seq_len2 represents relative position.] + pos_scores = torch.matmul(p, pos_emb) + # the following .as_strided() expression converts the last axis of pos_scores from relative + # to absolute position. I don't know whether I might have got the time-offsets backwards or + # not, but let this code define which way round it is supposed to be. + pos_scores = pos_scores.as_strided( + (num_heads, batch_size, seq_len, seq_len), + ( + pos_scores.stride(0), + pos_scores.stride(1), + pos_scores.stride(2) - pos_scores.stride(3), + pos_scores.stride(3), + ), + storage_offset=pos_scores.stride(3) * (seq_len - 1), + ) + + attn_scores = attn_scores + pos_scores + + if self.training and random.random() < 0.1: + # This is away of limiting the attention scores to not be + # too large. It incurs a penalty if any of them has an absolute + # value greater than 25.0. this should be outside the normal range + # of the attention scores. We use this mechanism instead of, say, + # something added to the loss function involving the entropy, + # because once the entropy gets very small gradients through the + # softmax can become very small, and we'd get zero derivatives. The + # choices of 1.0e-04 as the scale on the penalty makes this + # mechanism vulnerable to the absolute scale of the loss function, + # but we view this as a failsafe to avoid "implausible" parameter + # values rather than a regularization method that should be active + # under normal circumstances. + attn_scores = penalize_abs_values_gt( + attn_scores, limit=25.0, penalty=1.0e-04, name=self.name + ) + + assert attn_scores.shape == (num_heads, batch_size, seq_len, seq_len) + + if attn_mask is not None: + assert attn_mask.dtype == torch.bool + # use -1000 to avoid nan's where attn_mask and key_padding_mask make + # all scores zero. It's important that this be large enough that exp(-1000) + # is exactly zero, for reasons related to const_attention_rate, it + # compares the final weights with zero. + attn_scores = attn_scores.masked_fill(attn_mask, -1000) + + if key_padding_mask is not None: + assert key_padding_mask.shape == ( + batch_size, + seq_len, + ), key_padding_mask.shape + attn_scores = attn_scores.masked_fill( + key_padding_mask.unsqueeze(1), + -1000, + ) + + # We use our own version of softmax, defined in scaling.py, which should + # save a little of the memory used in backprop by, if we are in + # automatic mixed precision mode (amp / autocast), by only storing the + # half-precision output for backprop purposes. + attn_weights = softmax(attn_scores, dim=-1) + + if random.random() < 0.001: + self._print_attn_entropy(attn_weights) + + attn_weights = nn.functional.dropout( + attn_weights, p=self.dropout, training=self.training + ) + + return attn_weights + + def _print_attn_entropy(self, attn_weights: Tensor): + # attn_weights: (num_heads, batch_size, seq_len, seq_len) + (num_heads, batch_size, seq_len, seq_len) = attn_weights.shape + + with torch.no_grad(): + with torch.cuda.amp.autocast(enabled=False): + attn_weights = attn_weights.to(torch.float32) + attn_weights_entropy = ( + -((attn_weights + 1.0e-20).log() * attn_weights) + .sum(dim=-1) + .mean(dim=(1, 2)) + ) + logging.info( + f"name={self.name}, attn_weights_entropy = {attn_weights_entropy}" + ) + + +class Attention(nn.Module): + """ + The simplest possible attention module. This one works with already-computed attention + weights, e.g. as computed by RelPositionMultiheadAttentionWeights. + + Args: + embed_dim_in: the input embedding dimension + embed_dim_out: the output embedding dimension (normally the same as input) + num_heads: the number of attention heads + value_head_dim: the value dimension per head + """ + + def __init__( + self, + embed_dim_in: int, + embed_dim_out: int, + num_heads: int, + value_head_dim: int, + ) -> None: + super().__init__() + self.in_proj = nn.Linear(embed_dim_in, num_heads * value_head_dim, bias=False) + + # Note we set bias to False so that input of 0 will have no effect + self.out_proj = ScaledLinear( + num_heads * value_head_dim, embed_dim_out, bias=False, initial_scale=0.05 + ) + + self.whiten = Whiten( + num_groups=1, + whitening_limit=_whitening_schedule(7.5, ratio=3.0), + prob=(0.025, 0.25), + grad_scale=0.01, + ) + + def forward( + self, + x: Tensor, + attn_weights: Tensor, + ) -> Tensor: + """ + Args: + x: input tensor, of shape (seq_len, batch_size, embed_dim) + attn_weights: a tensor of shape (num_heads, batch_size, query_len, key_len), + Expect attn_weights.sum(dim=-1) == 1. The input here is the value in the + original attention mechanism. + Returns: + a tensor with the same shape as x. + """ + (num_heads, batch_size, query_len, key_len) = attn_weights.shape + + x = self.in_proj(x) # (key_len, batch_size, num_heads * value_head_dim) + x = x.reshape(key_len, batch_size, num_heads, -1).permute(2, 1, 0, 3) + # now x: (num_heads, batch_size, key_len, value_head_dim) + value_head_dim = x.shape[-1] + + # todo: see whether there is benefit in overriding matmul + x = torch.matmul(attn_weights, x) + # v: (num_heads, batch_size, query_len, value_head_dim) + + x = ( + x.permute(2, 1, 0, 3) + .contiguous() + .view(query_len, batch_size, num_heads * value_head_dim) + ) + + # returned value is of shape (query_len, batch_size, embed_dim), like the input. + x = self.out_proj(x) + x = self.whiten(x) + + return x + + def streaming_forward( + self, + x: Tensor, + attn_weights: Tensor, + cached_val: Tensor, + left_context_len: int, + ) -> Tuple[Tensor, Tensor]: + """ + Args: + x: input tensor, of shape (seq_len, batch_size, embed_dim) + attn_weights: a tensor of shape (num_heads, batch_size, seq_len, seq_len), + with seq_len being interpreted as (tgt_seq_len, src_seq_len). Expect + attn_weights.sum(dim=-1) == 1. + cached_val: cached attention value tensor of left context, + of shape (left_context_len, batch_size, value_dim) + left_context_len: number of left context frames. + + Returns: + - attention weighted output, a tensor with the same shape as x. + - updated cached attention value tensor of left context. + """ + (seq_len, batch_size, embed_dim) = x.shape + num_heads = attn_weights.shape[0] + seq_len2 = seq_len + left_context_len + assert attn_weights.shape == (num_heads, batch_size, seq_len, seq_len2) + + x = self.in_proj(x) # (seq_len, batch_size, num_heads * value_head_dim) + + # Pad cached left contexts + assert cached_val.shape[0] == left_context_len, ( + cached_val.shape[0], + left_context_len, + ) + x = torch.cat([cached_val, x], dim=0) + # Update cached left contexts + cached_val = x[-left_context_len:, ...] + + x = x.reshape(seq_len2, batch_size, num_heads, -1).permute(2, 1, 0, 3) + # now x: (num_heads, batch_size, seq_len, value_head_dim) + value_head_dim = x.shape[-1] + + # todo: see whether there is benefit in overriding matmul + x = torch.matmul(attn_weights, x) + # v: (num_heads, batch_size, seq_len, value_head_dim) + + x = ( + x.permute(2, 1, 0, 3) + .contiguous() + .view(seq_len, batch_size, num_heads * value_head_dim) + ) + + # returned value is of shape (seq_len, batch_size, embed_dim), like the input. + x = self.out_proj(x) + + return x, cached_val + + +class MultiheadAttentionWeights(nn.Module): + r"""Module that computes multi-head cross-attention weights. Allows src and target + to have different dims. + + Args: + key_embed_dim: number of channels of the thing that we'll project to + make the query (corresponds to source). e.g. 256 + query_embed_dim: number of channels of the thing that we'll project to + make the query (corresponds to target). e.g. 256 + num_heads: number of heads to compute weights for, e.g. 8 + head_dim: dimension of the query and key, per head. e.g. 24. + dropout: dropout probability for attn_output_weights. Default: 0.0. + """ + + def __init__( + self, + key_embed_dim: int, + query_embed_dim: int, + num_heads: int, + head_dim: int, + dropout: float = 0.0, + ) -> None: + super().__init__() + self.key_embed_dim = key_embed_dim + self.query_embed_dim = query_embed_dim + self.num_heads = num_heads + self.head_dim = head_dim + self.dropout = dropout + self.name = None # will be overwritten in training code; for diagnostics. + + # the initial_scale is supposed to take over the "scaling" factor of + # head_dim ** -0.5 that has been used in previous forms of attention, + # dividing it between the query and key. Note: this module is intended + # to be used with the ScaledAdam optimizer; with most other optimizers, + # it would be necessary to apply the scaling factor in the forward function. + self.query_in_proj = ScaledLinear( + query_embed_dim, + head_dim * num_heads, + bias=True, + initial_scale=head_dim**-0.25, + ) + + # weights produced by this module are invariant to adding a constant to + # the keys, so we don't need a bias for the keys. + self.key_in_proj = ScaledLinear( + key_embed_dim, + head_dim * num_heads, + bias=False, + initial_scale=head_dim**-0.25, + ) + + self.whiten_keys = Whiten( + num_groups=num_heads, + whitening_limit=_whitening_schedule(3.0), + prob=(0.025, 0.25), + grad_scale=0.025, + ) + + def forward( + self, + key: Tensor, + query: Tensor, + key_padding_mask: Optional[Tensor] = None, + ) -> Tensor: + r""" + Args: + key: input of shape (key_len, batch_size, key_embed_dim) + query: input of shape (query_len, batch_size, query_embed_dim) + key_padding_mask: an optional bool tensor of shape (batch_size, key_len). Positions that + are True in this mask will be ignored as sources in the attention weighting. + Returns: + a tensor of attention weights, of shape (hum_heads, batch_size, query_len, key_len) + """ + q = self.query_in_proj(query) + k = self.key_in_proj(key) + + head_dim = self.head_dim + num_heads = self.num_heads + + query_len, batch_size, _ = q.shape + key_len, _batch_size, _ = k.shape + assert _batch_size == batch_size + + k = self.whiten_keys(k) # does nothing in the forward pass. + + q = q.reshape(query_len, batch_size, num_heads, head_dim) + k = k.reshape(key_len, batch_size, num_heads, head_dim) + + # time1 refers to target, time2 refers to source. + q = q.permute(2, 1, 0, 3) # (head, batch, time1, query_head_dim) + k = k.permute(2, 1, 3, 0) # (head, batch, d_k, time2) + + attn_scores = torch.matmul(q, k) + + if self.training and random.random() < 0.1: + # This is a 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 25.0. this should be outside the normal range + # of the attention scores. We use this mechanism instead of, say, + # something added to the loss function involving the entropy, + # because once the entropy gets very small gradients through the + # softmax can become very small, and we'd get zero derivatives. The + # choices of 1.0e-04 as the scale on the penalty makes this + # mechanism vulnerable to the absolute scale of the loss function, + # but we view this as a failsafe to avoid "implausible" parameter + # values rather than a regularization method that should be active + # under normal circumstances. + attn_scores = penalize_abs_values_gt( + attn_scores, limit=25.0, penalty=1.0e-04, name=self.name + ) + + assert attn_scores.shape == (num_heads, batch_size, query_len, key_len) + + if key_padding_mask is not None: + assert key_padding_mask.shape == ( + batch_size, + key_len, + ), key_padding_mask.shape + attn_scores = attn_scores.masked_fill( + key_padding_mask.unsqueeze(1), + -1000, + ) + + # We use our own version of softmax, defined in scaling.py, which should + # save a little of the memory used in backprop by, if we are in + # automatic mixed precision mode (amp / autocast), by only storing the + # half-precision output for backprop purposes. + attn_weights = softmax(attn_scores, dim=-1) + + if random.random() < 0.001: + self._print_attn_entropy(attn_weights) + + attn_weights = nn.functional.dropout( + attn_weights, p=self.dropout, training=self.training + ) + + return attn_weights + + def _print_attn_entropy(self, attn_weights: Tensor): + # attn_weights: (num_heads, batch_size, seq_len, seq_len) + (num_heads, batch_size, seq_len, seq_len) = attn_weights.shape + + with torch.no_grad(): + with torch.cuda.amp.autocast(enabled=False): + attn_weights = attn_weights.to(torch.float32) + attn_weights_entropy = ( + -((attn_weights + 1.0e-20).log() * attn_weights) + .sum(dim=-1) + .mean(dim=(1, 2)) + ) + logging.info( + f"name={self.name}, attn_weights_entropy = {attn_weights_entropy}" + ) + + +class FeedforwardModule(nn.Module): + """Feedforward module in Zipformer2 model.""" + + def __init__(self, embed_dim: int, feedforward_dim: int, dropout: FloatLike): + super(FeedforwardModule, self).__init__() + self.in_proj = nn.Linear(embed_dim, feedforward_dim) + + self.hidden_balancer = Balancer( + feedforward_dim, + channel_dim=-1, + min_positive=0.3, + max_positive=1.0, + min_abs=0.75, + max_abs=5.0, + ) + + # shared_dim=0 means we share the dropout mask along the time axis + self.out_proj = ActivationDropoutAndLinear( + feedforward_dim, + embed_dim, + activation="SwooshL", + dropout_p=dropout, + dropout_shared_dim=0, + bias=True, + initial_scale=0.1, + ) + + self.out_whiten = Whiten( + num_groups=1, + whitening_limit=_whitening_schedule(7.5), + prob=(0.025, 0.25), + grad_scale=0.01, + ) + + def forward(self, x: Tensor): + x = self.in_proj(x) + x = self.hidden_balancer(x) + # out_proj contains SwooshL activation, then dropout, then linear. + x = self.out_proj(x) + x = self.out_whiten(x) + return x + + +class NonlinAttention(nn.Module): + """This is like the ConvolutionModule, but refactored so that we use multiplication by attention weights (borrowed + from the attention module) in place of actual convolution. We also took out the second nonlinearity, the + one after the attention mechanism. + + Args: + channels (int): The number of channels of conv layers. + """ + + def __init__( + self, + channels: int, + hidden_channels: int, + ) -> None: + super().__init__() + + self.hidden_channels = hidden_channels + + self.in_proj = nn.Linear(channels, hidden_channels * 3, bias=True) + + # balancer that goes before the sigmoid. Have quite a large min_abs value, at 2.0, + # because we noticed that well-trained instances of this module have abs-value before the sigmoid + # starting from about 3, and poorly-trained instances of the module have smaller abs values + # before the sigmoid. + self.balancer = Balancer( + hidden_channels, + channel_dim=-1, + min_positive=ScheduledFloat((0.0, 0.25), (20000.0, 0.05)), + max_positive=ScheduledFloat((0.0, 0.75), (20000.0, 0.95)), + min_abs=0.5, + max_abs=5.0, + ) + self.tanh = nn.Tanh() + + self.identity1 = Identity() # for diagnostics. + self.identity2 = Identity() # for diagnostics. + self.identity3 = Identity() # for diagnostics. + + self.out_proj = ScaledLinear( + hidden_channels, channels, bias=True, initial_scale=0.05 + ) + + self.whiten1 = Whiten( + num_groups=1, + whitening_limit=_whitening_schedule(5.0), + prob=(0.025, 0.25), + grad_scale=0.01, + ) + + self.whiten2 = Whiten( + num_groups=1, + whitening_limit=_whitening_schedule(5.0, ratio=3.0), + prob=(0.025, 0.25), + grad_scale=0.01, + ) + + def forward( + self, + x: Tensor, + attn_weights: Tensor, + ) -> Tensor: + """. + Args: + x: a Tensor of shape (seq_len, batch_size, num_channels) + attn_weights: a Tensor of shape (num_heads, batch_size, seq_len, seq_len) + Returns: + a Tensor with the same shape as x + """ + x = self.in_proj(x) + + (seq_len, batch_size, _) = x.shape + hidden_channels = self.hidden_channels + + s, x, y = x.chunk(3, dim=-1) + + # s will go through tanh. + + s = self.balancer(s) + s = self.tanh(s) + + s = s.unsqueeze(-1).reshape(seq_len, batch_size, hidden_channels) + x = self.whiten1(x) + x = x * s + x = self.identity1(x) # diagnostics only, it's the identity. + + (seq_len, batch_size, embed_dim) = x.shape + num_heads = attn_weights.shape[0] + assert attn_weights.shape == (num_heads, batch_size, seq_len, seq_len) + + x = x.reshape(seq_len, batch_size, num_heads, -1).permute(2, 1, 0, 3) + # now x: (num_heads, batch_size, seq_len, head_dim) + x = torch.matmul(attn_weights, x) + # now x: (num_heads, batch_size, seq_len, head_dim) + x = x.permute(2, 1, 0, 3).reshape(seq_len, batch_size, -1) + + y = self.identity2(y) + x = x * y + x = self.identity3(x) + + x = self.out_proj(x) + x = self.whiten2(x) + return x + + def streaming_forward( + self, + x: Tensor, + attn_weights: Tensor, + cached_x: Tensor, + left_context_len: int, + ) -> Tuple[Tensor, Tensor]: + """. + Args: + x: a Tensor of shape (seq_len, batch_size, num_channels) + attn_weights: a Tensor of shape (num_heads, batch_size, seq_len, seq_len) + cached_x: left context, a Tensor of shape + (num_heads, batch_size, left_context_len, head_dim) + left_context_len: number of left context frames. + Returns: + - a Tensor with the same shape as x + - updated left context with same shape as cached_x + """ + x = self.in_proj(x) + + (seq_len, batch_size, _) = x.shape + hidden_channels = self.hidden_channels + + s, x, y = x.chunk(3, dim=-1) + + # s will go through tanh. + s = self.tanh(s) + + s = s.unsqueeze(-1).reshape(seq_len, batch_size, hidden_channels) + x = x * s + + (seq_len, batch_size, embed_dim) = x.shape + num_heads = attn_weights.shape[0] + assert attn_weights.shape == ( + num_heads, + batch_size, + seq_len, + left_context_len + seq_len, + ) + + x = x.reshape(seq_len, batch_size, num_heads, -1).permute(2, 1, 0, 3) + # now x: (num_heads, batch_size, seq_len, head_dim) + + # Pad cached tensor + assert cached_x.shape[2] == left_context_len, ( + cached_x.shape[2], + left_context_len, + ) + x_pad = torch.cat([cached_x, x], dim=2) + # Update cached tensor + cached_x = x_pad[:, :, -left_context_len:, :] + + x = torch.matmul(attn_weights, x_pad) + # now x: (num_heads, batch_size, seq_len, head_dim) + x = x.permute(2, 1, 0, 3).reshape(seq_len, batch_size, -1) + + x = x * y + + x = self.out_proj(x) + return x, cached_x + + +class ConvolutionModule(nn.Module): + """ConvolutionModule in Zipformer2 model. + Modified from https://github.com/espnet/espnet/blob/master/espnet/nets/pytorch_backend/zipformer/convolution.py + + Args: + channels (int): The number of channels of conv layers. + kernel_size (int): Kernerl size of conv layers. + bias (bool): Whether to use bias in conv layers (default=True). + + """ + + def __init__( + self, + channels: int, + kernel_size: int, + causal: bool, + ) -> None: + """Construct a ConvolutionModule object.""" + super(ConvolutionModule, self).__init__() + # kernerl_size should be a odd number for 'SAME' padding + assert (kernel_size - 1) % 2 == 0 + + bottleneck_dim = channels + self.causal = causal + + self.in_proj = nn.Linear( + channels, + 2 * bottleneck_dim, + ) + # the gradients on in_proj are a little noisy, likely to do with the + # sigmoid in glu. + + # after in_proj we put x through a gated linear unit (nn.functional.glu). + # For most layers the normal rms value of channels of x seems to be in + # the range 1 to 4, but sometimes, for some reason, for layer 0 the + # rms ends up being very large, between 50 and 100 for different channels. + # This will cause very peaky and + # sparse derivatives for the sigmoid gating function, which will tend to make + # the loss function not learn effectively. (for most layers the average absolute values + # are in the range 0.5..9.0, and the average p(x>0), i.e. positive proportion, + # at the output of pointwise_conv1.output is around 0.35 to 0.45 for different + # layers, which likely breaks down as 0.5 for the "linear" half and + # 0.2 to 0.3 for the part that goes into the sigmoid. The idea is that if we + # constrain the rms values to a reasonable range via a constraint of max_abs=10.0, + # it will be in a better position to start learning something, i.e. to latch onto + # the correct range. + self.balancer1 = Balancer( + bottleneck_dim, + channel_dim=-1, + min_positive=ScheduledFloat((0.0, 0.05), (8000.0, 0.025)), + max_positive=1.0, + min_abs=1.5, + max_abs=ScheduledFloat((0.0, 5.0), (8000.0, 10.0), default=1.0), + ) + + self.activation1 = Identity() # for diagnostics + + self.sigmoid = nn.Sigmoid() + + self.activation2 = Identity() # for diagnostics + + assert kernel_size % 2 == 1 + + self.depthwise_conv = ( + ChunkCausalDepthwiseConv1d(channels=bottleneck_dim, kernel_size=kernel_size) + if causal + else nn.Conv1d( + in_channels=bottleneck_dim, + out_channels=bottleneck_dim, + groups=bottleneck_dim, + kernel_size=kernel_size, + padding=kernel_size // 2, + ) + ) + + self.balancer2 = Balancer( + bottleneck_dim, + channel_dim=1, + min_positive=ScheduledFloat((0.0, 0.1), (8000.0, 0.05)), + max_positive=1.0, + min_abs=ScheduledFloat((0.0, 0.2), (20000.0, 0.5)), + max_abs=10.0, + ) + + self.whiten = Whiten( + num_groups=1, + whitening_limit=_whitening_schedule(7.5), + prob=(0.025, 0.25), + grad_scale=0.01, + ) + + self.out_proj = ActivationDropoutAndLinear( + bottleneck_dim, + channels, + activation="SwooshR", + dropout_p=0.0, + initial_scale=0.05, + ) + + def forward( + self, + x: Tensor, + src_key_padding_mask: Optional[Tensor] = None, + chunk_size: int = -1, + ) -> Tensor: + """Compute convolution module. + + Args: + x: Input tensor (#time, batch, channels). + src_key_padding_mask: the mask for the src keys per batch (optional): + (batch, #time), contains True in masked positions. + + Returns: + Tensor: Output tensor (#time, batch, channels). + + """ + + x = self.in_proj(x) # (time, batch, 2*channels) + + x, s = x.chunk(2, dim=-1) + s = self.balancer1(s) + s = self.sigmoid(s) + x = self.activation1(x) # identity. + x = x * s + x = self.activation2(x) # identity + + # (time, batch, channels) + + # exchange the temporal dimension and the feature dimension + x = x.permute(1, 2, 0) # (#batch, channels, time). + + if src_key_padding_mask is not None: + x = x.masked_fill(src_key_padding_mask.unsqueeze(1).expand_as(x), 0.0) + + if ( + not torch.jit.is_scripting() + and not torch.jit.is_tracing() + and chunk_size >= 0 + ): + # Not support exporting a model for simulated streaming decoding + assert ( + self.causal + ), "Must initialize model with causal=True if you use chunk_size" + x = self.depthwise_conv(x, chunk_size=chunk_size) + else: + x = self.depthwise_conv(x) + + x = self.balancer2(x) + x = x.permute(2, 0, 1) # (time, batch, channels) + + x = self.whiten(x) # (time, batch, channels) + x = self.out_proj(x) # (time, batch, channels) + + return x + + def streaming_forward( + self, + x: Tensor, + cache: Tensor, + src_key_padding_mask: Tensor, + ) -> Tuple[Tensor, Tensor]: + """Compute convolution module in streaming forward mode. + + Args: + x: Input tensor (#time, batch, channels). + cache: cached left context for depthwise_conv of shape + (#batch, channels, left_pad) + src_key_padding_mask: the mask for the src keys per batch (optional): + (batch, #time), contains True in masked positions. + + Returns: + - Output tensor (#time, batch, channels). + - Updated cache (#batch, channels, left_pad) + """ + + x = self.in_proj(x) # (time, batch, 2*channels) + + x, s = x.chunk(2, dim=2) + s = self.sigmoid(s) + x = x * s + # (time, batch, channels) + + # exchange the temporal dimension and the feature dimension + x = x.permute(1, 2, 0) # (#batch, channels, time). + + if src_key_padding_mask is not None: + x = x.masked_fill(src_key_padding_mask.unsqueeze(1).expand_as(x), 0.0) + + x, cache = self.depthwise_conv.streaming_forward(x, cache=cache) + + x = x.permute(2, 0, 1) # (time, batch, channels) + + x = self.out_proj(x) # (time, batch, channels) + + return x, cache + + +class ScalarMultiply(nn.Module): + def __init__(self, scale: float): + super().__init__() + self.scale = scale + + def forward(self, x): + return x * self.scale + + +def _test_zipformer_main(causal: bool = False): + batch_size = 5 + seq_len = 20 + # Just make sure the forward pass runs. + memory_dim = 100 + + c = Zipformer2( + encoder_dim=(64, 96), + encoder_unmasked_dim=(48, 64), + num_heads=(4, 4), + causal=causal, + chunk_size=(4,) if causal else (-1,), + left_context_frames=(64,), + memory_dim=memory_dim, + ) + batch_size = 5 + seq_len = 20 + # Just make sure the forward pass runs. + f = c( + torch.randn(seq_len, batch_size, 64), + torch.full((batch_size,), seq_len, dtype=torch.int64), + memory=torch.randn(101, batch_size, memory_dim), + ) + f[0].sum().backward() + c.eval() + f = c( + torch.randn(seq_len, batch_size, 64), + torch.full((batch_size,), seq_len, dtype=torch.int64), + ) + f # to remove flake8 warnings + + +if __name__ == "__main__": + logging.getLogger().setLevel(logging.INFO) + torch.set_num_threads(1) + torch.set_num_interop_threads(1) + _test_zipformer_main(False) + _test_zipformer_main(True) diff --git a/icefall/utils.py b/icefall/utils.py index 8fda3a4ca..410340d9d 100644 --- a/icefall/utils.py +++ b/icefall/utils.py @@ -483,7 +483,7 @@ def load_alignments(filename: str) -> Tuple[int, Dict[str, List[int]]]: def store_transcripts( - filename: Pathlike, texts: Iterable[Tuple[str, str, str]] + filename: Pathlike, texts: Iterable[Tuple[str, str, str]], char_level: bool = False ) -> None: """Save predicted results and reference transcripts to a file. @@ -500,6 +500,9 @@ def store_transcripts( """ with open(filename, "w") as f: for cut_id, ref, hyp in texts: + if char_level: + ref = list("".join(ref)) + hyp = list("".join(hyp)) print(f"{cut_id}:\tref={ref}", file=f) print(f"{cut_id}:\thyp={hyp}", file=f) @@ -557,6 +560,7 @@ def write_error_stats( test_set_name: str, results: List[Tuple[str, str]], enable_log: bool = True, + compute_CER: bool = False, sclite_mode: bool = False, ) -> float: """Write statistics based on predicted results and reference transcripts. @@ -585,7 +589,7 @@ def write_error_stats( The reference word `SIR` is missing in the predicted results (a deletion error). results: - An iterable of tuples. The first element is the cur_id, the second is + An iterable of tuples. The first element is the cut_id, the second is the reference transcript and the third element is the predicted result. enable_log: If True, also print detailed WER to the console. @@ -602,6 +606,14 @@ def write_error_stats( words: Dict[str, List[int]] = defaultdict(lambda: [0, 0, 0, 0, 0]) num_corr = 0 ERR = "*" + + if compute_CER: + for i, res in enumerate(results): + cut_id, ref, hyp = res + ref = list("".join(ref)) + hyp = list("".join(hyp)) + results[i] = (cut_id, ref, hyp) + for cut_id, ref, hyp in results: ali = kaldialign.align(ref, hyp, ERR, sclite_mode=sclite_mode) for ref_word, hyp_word in ali: @@ -1426,7 +1438,10 @@ def measure_gradient_norms(model: nn.Module, norm: str = "l1") -> Dict[str, floa def get_parameter_groups_with_lrs( - model: nn.Module, lr: float, include_names: bool = False + model: nn.Module, + lr: float, + include_names: bool = False, + freeze_modules: List[str] = [], ) -> List[dict]: """ This is for use with the ScaledAdam optimizers (more recent versions that accept lists of @@ -1450,6 +1465,8 @@ def get_parameter_groups_with_lrs( ... ] """ + named_modules = list(model.named_modules()) + # flat_lr_scale just contains the lr_scale explicitly specified # for each prefix of the name, e.g. 'encoder.layers.3', these need # to be multiplied for all prefix of the name of any given parameter. @@ -1469,6 +1486,15 @@ def get_parameter_groups_with_lrs( split_name = name.split(".") # caution: as a special case, if the name is '', split_name will be [ '' ]. prefix = split_name[0] + if prefix == "module": # DDP + module_name = split_name[1] + if module_name in freeze_modules: + logging.info(f"Remove {name} from parameters") + continue + else: + if prefix in freeze_modules: + logging.info(f"Remove {name} from parameters") + continue cur_lr = lr * flat_lr_scale[prefix] if prefix != "": cur_lr *= flat_lr_scale[""]