diff --git a/egs/ami/SURT/README.md b/egs/ami/SURT/README.md new file mode 100644 index 000000000..74a8ba014 --- /dev/null +++ b/egs/ami/SURT/README.md @@ -0,0 +1,156 @@ +# Introduction + +This is a multi-talker ASR recipe for the AMI and ICSI datasets. We train a Streaming +Unmixing and Recognition Transducer (SURT) model for the task. + +Please refer to the `egs/libricss/SURT` recipe README for details about the task and the +model. + +## Description of the recipe + +### Pre-requisites + +The recipes in this directory need the following packages to be installed: + +- [meeteval](https://github.com/fgnt/meeteval) +- [einops](https://github.com/arogozhnikov/einops) + +Additionally, we initialize the model with the pre-trained model from the LibriCSS recipe. +Please download this checkpoint (see below) or train the LibriCSS recipe first. + +### Training + +To train the model, run the following from within `egs/ami/SURT`: + +```bash +export CUDA_VISIBLE_DEVICES="0,1,2,3" + +python dprnn_zipformer/train.py \ + --use-fp16 True \ + --exp-dir dprnn_zipformer/exp/surt_base \ + --world-size 4 \ + --max-duration 500 \ + --max-duration-valid 250 \ + --max-cuts 200 \ + --num-buckets 50 \ + --num-epochs 30 \ + --enable-spec-aug True \ + --enable-musan False \ + --ctc-loss-scale 0.2 \ + --heat-loss-scale 0.2 \ + --base-lr 0.004 \ + --model-init-ckpt exp/libricss_base.pt \ + --chunk-width-randomization True \ + --num-mask-encoder-layers 4 \ + --num-encoder-layers 2,2,2,2,2 +``` + +The above is for SURT-base (~26M). For SURT-large (~38M), use: + +```bash + --model-init-ckpt exp/libricss_large.pt \ + --num-mask-encoder-layers 6 \ + --num-encoder-layers 2,4,3,2,4 \ + --model-init-ckpt exp/zipformer_large.pt \ +``` + +**NOTE:** You may need to decrease the `--max-duration` for SURT-large to avoid OOM. + +### Adaptation + +The training step above only trains on simulated mixtures. For best results, we also +adapt the final model on the AMI+ICSI train set. For this, run the following from within +`egs/ami/SURT`: + +```bash +export CUDA_VISIBLE_DEVICES="0" + +python dprnn_zipformer/train_adapt.py \ + --use-fp16 True \ + --exp-dir dprnn_zipformer/exp/surt_base_adapt \ + --world-size 4 \ + --max-duration 500 \ + --max-duration-valid 250 \ + --max-cuts 200 \ + --num-buckets 50 \ + --num-epochs 8 \ + --lr-epochs 2 \ + --enable-spec-aug True \ + --enable-musan False \ + --ctc-loss-scale 0.2 \ + --base-lr 0.0004 \ + --model-init-ckpt dprnn_zipformer/exp/surt_base/epoch-30.pt \ + --chunk-width-randomization True \ + --num-mask-encoder-layers 4 \ + --num-encoder-layers 2,2,2,2,2 +``` + +For SURT-large, use the following config: + +```bash + --num-mask-encoder-layers 6 \ + --num-encoder-layers 2,4,3,2,4 \ + --model-init-ckpt dprnn_zipformer/exp/surt_large/epoch-30.pt \ + --num-epochs 15 \ + --lr-epochs 4 \ +``` + + +### Decoding + +To decode the model, run the following from within `egs/ami/SURT`: + +#### Greedy search + +```bash +export CUDA_VISIBLE_DEVICES="0" + +python dprnn_zipformer/decode.py \ + --epoch 20 --avg 1 --use-averaged-model False \ + --exp-dir dprnn_zipformer/exp/surt_base_adapt \ + --max-duration 250 \ + --decoding-method greedy_search +``` + +#### Beam search + +```bash +python dprnn_zipformer/decode.py \ + --epoch 20 --avg 1 --use-averaged-model False \ + --exp-dir dprnn_zipformer/exp/surt_base_adapt \ + --max-duration 250 \ + --decoding-method modified_beam_search \ + --beam-size 4 +``` + +## Results (using beam search) + +**AMI** + +| Model | IHM-Mix | SDM | MDM | +|------------|:-------:|:----:|:----:| +| SURT-base | 39.8 | 65.4 | 46.6 | +| + adapt | 37.4 | 46.9 | 43.7 | +| SURT-large | 36.8 | 62.5 | 44.4 | +| + adapt | **35.1** | **44.6** | **41.4** | + +**ICSI** + +| Model | IHM-Mix | SDM | +|------------|:-------:|:----:| +| SURT-base | 28.3 | 60.0 | +| + adapt | 26.3 | 33.9 | +| SURT-large | 27.8 | 59.7 | +| + adapt | **24.4** | **32.3** | + +## Pre-trained models and logs + +* LibriCSS pre-trained model (for initialization): [base](https://huggingface.co/desh2608/icefall-surt-libricss-dprnn-zipformer/tree/main/exp/surt_base) [large](https://huggingface.co/desh2608/icefall-surt-libricss-dprnn-zipformer/tree/main/exp/surt_large) + +* Pre-trained models: + +* Training logs: + - surt_base: + - surt_base_adapt: + - surt_large: + - surt_large_adapt: diff --git a/egs/ami/SURT/dprnn_zipformer/asr_datamodule.py b/egs/ami/SURT/dprnn_zipformer/asr_datamodule.py new file mode 100644 index 000000000..ec8106bc3 --- /dev/null +++ b/egs/ami/SURT/dprnn_zipformer/asr_datamodule.py @@ -0,0 +1,399 @@ +# Copyright 2021 Piotr Żelasko +# Copyright 2022 Xiaomi Corporation (Author: Mingshuang Luo) +# +# See ../../../../LICENSE for clarification regarding multiple authors +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + + +import argparse +import inspect +import logging +from functools import lru_cache +from pathlib import Path +from typing import Any, Callable, Dict, List, Optional + +import torch +from lhotse import CutSet, Fbank, FbankConfig, load_manifest, load_manifest_lazy +from lhotse.dataset import ( # noqa F401 for PrecomputedFeatures + CutMix, + DynamicBucketingSampler, + K2SurtDataset, + PrecomputedFeatures, + SimpleCutSampler, + 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 AmiAsrDataModule: + """ + DataModule for k2 SURT 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, + - augmentation, + - on-the-fly feature extraction + + This class should be derived for specific corpora used in ASR tasks. + """ + + def __init__(self, args: argparse.Namespace): + self.args = args + + @classmethod + def add_arguments(cls, parser: argparse.ArgumentParser): + group = parser.add_argument_group( + title="ASR data related options", + description="These options are used for the preparation of " + "PyTorch DataLoaders from Lhotse CutSet's -- they control the " + "effective batch sizes, sampling strategies, applied data " + "augmentations, etc.", + ) + group.add_argument( + "--manifest-dir", + type=Path, + default=Path("data/manifests"), + 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( + "--max-duration-valid", + 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( + "--max-cuts", + type=int, + default=100, + help="Maximum number of cuts 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( + "--on-the-fly-feats", + type=str2bool, + default=False, + help=( + "When enabled, use on-the-fly cut mixing and feature " + "extraction. Will drop existing precomputed feature manifests " + "if available." + ), + ) + group.add_argument( + "--shuffle", + type=str2bool, + default=True, + help="When enabled (=default), the examples will be " + "shuffled for each epoch.", + ) + group.add_argument( + "--drop-last", + type=str2bool, + default=True, + help="Whether to drop last batch. Used by sampler.", + ) + group.add_argument( + "--return-cuts", + type=str2bool, + default=True, + help="When enabled, each batch will have the " + "field: batch['supervisions']['cut'] with the cuts that " + "were used to construct it.", + ) + + group.add_argument( + "--num-workers", + type=int, + default=2, + help="The number of training dataloader workers that " + "collect the batches.", + ) + + group.add_argument( + "--enable-spec-aug", + type=str2bool, + default=True, + help="When enabled, use SpecAugment for training dataset.", + ) + + group.add_argument( + "--spec-aug-time-warp-factor", + type=int, + default=80, + help="Used only when --enable-spec-aug is True. " + "It specifies the factor for time warping in SpecAugment. " + "Larger values mean more warping. " + "A value less than 1 means to disable time warp.", + ) + + group.add_argument( + "--enable-musan", + type=str2bool, + default=True, + help="When enabled, select noise from MUSAN and mix it" + "with training dataset. ", + ) + + def train_dataloaders( + self, + cuts_train: CutSet, + sampler_state_dict: Optional[Dict[str, Any]] = None, + sources: bool = False, + ) -> DataLoader: + """ + Args: + cuts_train: + CutSet for training. + sampler_state_dict: + The state dict for the training sampler. + """ + transforms = [] + if self.args.enable_musan: + logging.info("Enable MUSAN") + logging.info("About to get Musan cuts") + cuts_musan = load_manifest(self.args.manifest_dir / "musan_cuts.jsonl.gz") + transforms.append( + CutMix(cuts=cuts_musan, prob=0.5, snr=(10, 20), preserve_id=True) + ) + else: + logging.info("Disable MUSAN") + + 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 = K2SurtDataset( + input_strategy=OnTheFlyFeatures(Fbank(FbankConfig(num_mel_bins=80))) + if self.args.on_the_fly_feats + else PrecomputedFeatures(), + cut_transforms=transforms, + input_transforms=input_transforms, + return_cuts=self.args.return_cuts, + return_sources=sources, + strict=False, + ) + + if self.args.bucketing_sampler: + logging.info("Using DynamicBucketingSampler.") + train_sampler = DynamicBucketingSampler( + cuts_train, + max_duration=self.args.max_duration, + quadratic_duration=30.0, + max_cuts=self.args.max_cuts, + shuffle=self.args.shuffle, + num_buckets=self.args.num_buckets, + drop_last=self.args.drop_last, + ) + else: + logging.info("Using SingleCutSampler.") + train_sampler = SimpleCutSampler( + cuts_train, + max_duration=self.args.max_duration, + max_cuts=self.args.max_cuts, + shuffle=self.args.shuffle, + ) + logging.info("About to create train dataloader") + + if sampler_state_dict is not None: + logging.info("Loading sampler state dict") + train_sampler.load_state_dict(sampler_state_dict) + + # 'seed' is derived from the current random state, which will have + # previously been set in the main process. + seed = torch.randint(0, 100000, ()).item() + worker_init_fn = _SeedWorkers(seed) + + train_dl = DataLoader( + train, + sampler=train_sampler, + batch_size=None, + num_workers=self.args.num_workers, + persistent_workers=False, + worker_init_fn=worker_init_fn, + ) + + return train_dl + + def valid_dataloaders(self, cuts_valid: CutSet) -> DataLoader: + transforms = [] + + logging.info("About to create dev dataset") + validate = K2SurtDataset( + input_strategy=OnTheFlyFeatures( + OnTheFlyFeatures(Fbank(FbankConfig(num_mel_bins=80))) + ) + if self.args.on_the_fly_feats + else PrecomputedFeatures(), + cut_transforms=transforms, + return_cuts=self.args.return_cuts, + return_sources=False, + strict=False, + ) + valid_sampler = DynamicBucketingSampler( + cuts_valid, + max_duration=self.args.max_duration_valid, + quadratic_duration=30.0, + max_cuts=self.args.max_cuts, + shuffle=False, + ) + logging.info("About to create dev dataloader") + + # '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) + + valid_dl = DataLoader( + validate, + sampler=valid_sampler, + batch_size=None, + num_workers=self.args.num_workers, + persistent_workers=False, + worker_init_fn=worker_init_fn, + ) + + return valid_dl + + def test_dataloaders(self, cuts: CutSet) -> DataLoader: + logging.debug("About to create test dataset") + test = K2SurtDataset( + input_strategy=OnTheFlyFeatures( + OnTheFlyFeatures(Fbank(FbankConfig(num_mel_bins=80))) + ) + if self.args.on_the_fly_feats + else PrecomputedFeatures(), + return_cuts=self.args.return_cuts, + return_sources=False, + strict=False, + ) + sampler = DynamicBucketingSampler( + cuts, + max_duration=self.args.max_duration_valid, + max_cuts=self.args.max_cuts, + shuffle=False, + ) + + # '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) + + logging.debug("About to create test dataloader") + test_dl = DataLoader( + test, + batch_size=None, + sampler=sampler, + num_workers=self.args.num_workers, + persistent_workers=False, + worker_init_fn=worker_init_fn, + ) + return test_dl + + @lru_cache() + def aimix_train_cuts( + self, + rvb_affix: str = "clean", + sources: bool = True, + ) -> CutSet: + logging.info("About to get train cuts") + source_affix = "_sources" if sources else "" + cs = load_manifest_lazy( + self.args.manifest_dir / f"cuts_train_{rvb_affix}{source_affix}.jsonl.gz" + ) + cs = cs.filter(lambda c: c.duration >= 1.0 and c.duration <= 30.0) + return cs + + @lru_cache() + def train_cuts( + self, + ) -> CutSet: + logging.info("About to get train cuts") + return load_manifest_lazy( + self.args.manifest_dir / "cuts_train_ami_icsi.jsonl.gz" + ) + + @lru_cache() + def ami_cuts(self, split: str = "dev", type: str = "sdm") -> CutSet: + logging.info(f"About to get AMI {split} {type} cuts") + return load_manifest_lazy( + self.args.manifest_dir / f"cuts_ami-{type}_{split}.jsonl.gz" + ) + + @lru_cache() + def icsi_cuts(self, split: str = "dev", type: str = "sdm") -> CutSet: + logging.info(f"About to get ICSI {split} {type} cuts") + return load_manifest_lazy( + self.args.manifest_dir / f"cuts_icsi-{type}_{split}.jsonl.gz" + ) diff --git a/egs/ami/SURT/dprnn_zipformer/beam_search.py b/egs/ami/SURT/dprnn_zipformer/beam_search.py new file mode 120000 index 000000000..581b29833 --- /dev/null +++ b/egs/ami/SURT/dprnn_zipformer/beam_search.py @@ -0,0 +1 @@ +../../../libricss/SURT/dprnn_zipformer/beam_search.py \ No newline at end of file diff --git a/egs/ami/SURT/dprnn_zipformer/decode.py b/egs/ami/SURT/dprnn_zipformer/decode.py new file mode 100755 index 000000000..d1a1eddc9 --- /dev/null +++ b/egs/ami/SURT/dprnn_zipformer/decode.py @@ -0,0 +1,622 @@ +#!/usr/bin/env python3 +# +# Copyright 2021-2022 Xiaomi Corporation (Author: Fangjun Kuang, +# Zengwei Yao) +# +# See ../../../../LICENSE for clarification regarding multiple authors +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +""" +Usage: +(1) greedy search +./dprnn_zipformer/decode.py \ + --epoch 20 \ + --avg 1 \ + --use-averaged-model false \ + --exp-dir ./dprnn_zipformer/exp_adapt \ + --max-duration 600 \ + --decoding-method greedy_search + +(2) beam search (not recommended) +./dprnn_zipformer/decode.py \ + --epoch 20 \ + --avg 1 \ + --use-averaged-model false \ + --exp-dir ./dprnn_zipformer/exp_adapt \ + --max-duration 600 \ + --decoding-method beam_search \ + --beam-size 4 + +(3) modified beam search +./dprnn_zipformer/decode.py \ + --epoch 20 \ + --avg 1 \ + --use-averaged-model false \ + --exp-dir ./dprnn_zipformer/exp_adapt \ + --max-duration 600 \ + --decoding-method modified_beam_search \ + --beam-size 4 +""" + + +import argparse +import logging +from collections import defaultdict +from pathlib import Path +from typing import Dict, List, Optional, Tuple + +import k2 +import sentencepiece as spm +import torch +import torch.nn as nn +from asr_datamodule import AmiAsrDataModule +from beam_search import ( + beam_search, + greedy_search, + greedy_search_batch, + modified_beam_search, +) +from lhotse.utils import EPSILON +from train import add_model_arguments, get_params, get_surt_model + +from icefall import LmScorer, NgramLm +from icefall.checkpoint import ( + average_checkpoints, + average_checkpoints_with_averaged_model, + find_checkpoints, + load_checkpoint, +) +from icefall.lexicon import Lexicon +from icefall.utils import ( + AttributeDict, + setup_logger, + store_transcripts, + str2bool, + write_surt_error_stats, +) + + +def get_parser(): + parser = argparse.ArgumentParser( + formatter_class=argparse.ArgumentDefaultsHelpFormatter + ) + + parser.add_argument( + "--epoch", + type=int, + default=20, + help="""It specifies the checkpoint to use for decoding. + Note: Epoch counts from 1. + You can specify --avg to use more checkpoints for model averaging.""", + ) + + parser.add_argument( + "--iter", + type=int, + default=0, + help="""If positive, --epoch is ignored and it + will use the checkpoint exp_dir/checkpoint-iter.pt. + You can specify --avg to use more checkpoints for model averaging. + """, + ) + + parser.add_argument( + "--avg", + type=int, + default=1, + help="Number of checkpoints to average. Automatically select " + "consecutive checkpoints before the checkpoint specified by " + "'--epoch' and '--iter'", + ) + + parser.add_argument( + "--use-averaged-model", + type=str2bool, + default=True, + help="Whether to load averaged model. Currently it only supports " + "using --epoch. If True, it would decode with the averaged model " + "over the epoch range from `epoch-avg` (excluded) to `epoch`." + "Actually only the models with epoch number of `epoch-avg` and " + "`epoch` are loaded for averaging. ", + ) + + parser.add_argument( + "--exp-dir", + type=str, + default="dprnn_zipformer/exp", + help="The experiment dir", + ) + + parser.add_argument( + "--bpe-model", + type=str, + default="data/lang_bpe_500/bpe.model", + help="Path to the BPE model", + ) + + parser.add_argument( + "--decoding-method", + type=str, + default="greedy_search", + help="""Possible values are: + - greedy_search + - beam_search + - modified_beam_search + """, + ) + + 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( + "--context-size", + type=int, + default=2, + help="The context size in the decoder. 1 means bigram; 2 means tri-gram", + ) + parser.add_argument( + "--max-sym-per-frame", + type=int, + default=1, + help="""Maximum number of symbols per frame. + Used only when --decoding_method is greedy_search""", + ) + + add_model_arguments(parser) + + return parser + + +def decode_one_batch( + params: AttributeDict, + model: nn.Module, + sp: spm.SentencePieceProcessor, + batch: dict, +) -> 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`. + Returns: + Return the decoding result. See above description for the format of + the returned dict. + """ + device = next(model.parameters()).device + feature = batch["inputs"] + assert feature.ndim == 3 + + feature = feature.to(device) + feature_lens = batch["input_lens"].to(device) + + # Apply the mask encoder + B, T, F = feature.shape + processed = model.mask_encoder(feature) # B,T,F*num_channels + masks = processed.view(B, T, F, params.num_channels).unbind(dim=-1) + x_masked = [feature * m for m in masks] + + # Recognition + # Stack the inputs along the batch axis + h = torch.cat(x_masked, dim=0) + h_lens = torch.cat([feature_lens for _ in range(params.num_channels)], dim=0) + encoder_out, encoder_out_lens = model.encoder(x=h, x_lens=h_lens) + + if model.joint_encoder_layer is not None: + encoder_out = model.joint_encoder_layer(encoder_out) + + def _group_channels(hyps: List[str]) -> List[List[str]]: + """ + Currently we have a batch of size M*B, where M is the number of + channels and B is the batch size. We need to group the hypotheses + into B groups, each of which contains M hypotheses. + + Example: + hyps = ['a1', 'b1', 'c1', 'a2', 'b2', 'c2'] + _group_channels(hyps) = [['a1', 'a2'], ['b1', 'b2'], ['c1', 'c2']] + """ + assert len(hyps) == B * params.num_channels + out_hyps = [] + for i in range(B): + out_hyps.append(hyps[i::B]) + return out_hyps + + 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) + 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) + else: + batch_size = encoder_out.size(0) + + for i in range(batch_size): + # fmt: off + encoder_out_i = encoder_out[i:i+1, :encoder_out_lens[i]] + # fmt: on + if params.decoding_method == "greedy_search": + hyp = greedy_search( + model=model, + encoder_out=encoder_out_i, + max_sym_per_frame=params.max_sym_per_frame, + ) + elif params.decoding_method == "beam_search": + hyp = beam_search( + model=model, + encoder_out=encoder_out_i, + beam=params.beam_size, + ) + else: + raise ValueError( + f"Unsupported decoding method: {params.decoding_method}" + ) + hyps.append(sp.decode(hyp)) + + if params.decoding_method == "greedy_search": + return {"greedy_search": _group_channels(hyps)} + elif "fast_beam_search" in params.decoding_method: + key = f"beam_{params.beam}_" + key += f"max_contexts_{params.max_contexts}_" + key += f"max_states_{params.max_states}" + if "nbest" in params.decoding_method: + key += f"_num_paths_{params.num_paths}_" + key += f"nbest_scale_{params.nbest_scale}" + if "LG" in params.decoding_method: + key += f"_ngram_lm_scale_{params.ngram_lm_scale}" + + return {key: _group_channels(hyps)} + else: + return {f"beam_size_{params.beam_size}": _group_channels(hyps)} + + +def decode_dataset( + dl: torch.utils.data.DataLoader, + params: AttributeDict, + model: nn.Module, + sp: spm.SentencePieceProcessor, +) -> 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. + 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): + cut_ids = [cut.id for cut in batch["cuts"]] + cuts_batch = batch["cuts"] + + hyps_dict = decode_one_batch( + params=params, + model=model, + sp=sp, + ) + + for name, hyps in hyps_dict.items(): + this_batch = [] + for cut_id, hyp_words in zip(cut_ids, hyps): + # Reference is a list of supervision texts sorted by start time. + ref_words = [ + s.text.strip() + for s in sorted( + cuts_batch[cut_id].supervisions, key=lambda s: s.start + ) + ] + this_batch.append((cut_id, ref_words, hyp_words)) + + results[name].extend(this_batch) + + num_cuts += len(cut_ids) + + if batch_idx % log_interval == 0: + batch_str = f"{batch_idx}/{num_batches}" + + logging.info(f"batch {batch_str}, cuts processed until now is {num_cuts}") + return results + + +def save_results( + params: AttributeDict, + test_set_name: str, + results_dict: Dict[str, List[Tuple[str, List[str], List[str]]]], +): + test_set_wers = dict() + for key, results in results_dict.items(): + recog_path = ( + params.res_dir / f"recogs-{test_set_name}-{key}-{params.suffix}.txt" + ) + results = sorted(results) + store_transcripts(filename=recog_path, texts=results) + logging.info(f"The transcripts are stored in {recog_path}") + + # The following prints out WERs, per-word error statistics and aligned + # ref/hyp pairs. + errs_filename = ( + params.res_dir / f"errs-{test_set_name}-{key}-{params.suffix}.txt" + ) + with open(errs_filename, "w") as f: + wer = write_surt_error_stats( + f, + f"{test_set_name}-{key}", + results, + enable_log=True, + num_channels=params.num_channels, + ) + test_set_wers[key] = wer + + logging.info("Wrote detailed error stats to {}".format(errs_filename)) + + test_set_wers = sorted(test_set_wers.items(), key=lambda x: x[1]) + errs_info = ( + params.res_dir / f"wer-summary-{test_set_name}-{key}-{params.suffix}.txt" + ) + with open(errs_info, "w") as f: + print("settings\tWER", file=f) + for key, val in test_set_wers: + print("{}\t{}".format(key, val), file=f) + + s = "\nFor {}, WER of different settings are:\n".format(test_set_name) + note = "\tbest for {}".format(test_set_name) + for key, val in test_set_wers: + s += "{}\t{}{}\n".format(key, val, note) + note = "" + logging.info(s) + + +@torch.no_grad() +def main(): + parser = get_parser() + LmScorer.add_arguments(parser) + AmiAsrDataModule.add_arguments(parser) + args = parser.parse_args() + args.exp_dir = Path(args.exp_dir) + args.lang_dir = Path(args.lang_dir) + + params = get_params() + params.update(vars(args)) + + assert params.decoding_method in ( + "greedy_search", + "beam_search", + "modified_beam_search", + ), f"Decoding method {params.decoding_method} is not supported." + 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 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_surt_model(params) + assert model.encoder.decode_chunk_size == params.decode_chunk_len // 2, ( + model.encoder.decode_chunk_size, + params.decode_chunk_len, + ) + + if not params.use_averaged_model: + if params.iter > 0: + filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[ + : params.avg + ] + if len(filenames) == 0: + raise ValueError( + f"No checkpoints found for" + f" --iter {params.iter}, --avg {params.avg}" + ) + elif len(filenames) < params.avg: + raise ValueError( + f"Not enough checkpoints ({len(filenames)}) found for" + f" --iter {params.iter}, --avg {params.avg}" + ) + logging.info(f"averaging {filenames}") + model.to(device) + model.load_state_dict(average_checkpoints(filenames, device=device)) + elif params.avg == 1: + load_checkpoint(f"{params.exp_dir}/epoch-{params.epoch}.pt", model) + else: + start = params.epoch - params.avg + 1 + filenames = [] + for i in range(start, params.epoch + 1): + if i >= 1: + filenames.append(f"{params.exp_dir}/epoch-{i}.pt") + logging.info(f"averaging {filenames}") + model.to(device) + model.load_state_dict(average_checkpoints(filenames, device=device)) + else: + if params.iter > 0: + filenames = find_checkpoints(params.exp_dir, iteration=-params.iter)[ + : params.avg + 1 + ] + if len(filenames) == 0: + raise ValueError( + f"No checkpoints found for" + f" --iter {params.iter}, --avg {params.avg}" + ) + elif len(filenames) < params.avg + 1: + raise ValueError( + f"Not enough checkpoints ({len(filenames)}) found for" + f" --iter {params.iter}, --avg {params.avg}" + ) + filename_start = filenames[-1] + filename_end = filenames[0] + logging.info( + "Calculating the averaged model over iteration checkpoints" + f" from {filename_start} (excluded) to {filename_end}" + ) + model.to(device) + model.load_state_dict( + average_checkpoints_with_averaged_model( + filename_start=filename_start, + filename_end=filename_end, + device=device, + ) + ) + else: + assert params.avg > 0, params.avg + start = params.epoch - params.avg + assert start >= 1, start + filename_start = f"{params.exp_dir}/epoch-{start}.pt" + filename_end = f"{params.exp_dir}/epoch-{params.epoch}.pt" + logging.info( + f"Calculating the averaged model over epoch range from " + f"{start} (excluded) to {params.epoch}" + ) + model.to(device) + model.load_state_dict( + average_checkpoints_with_averaged_model( + filename_start=filename_start, + filename_end=filename_end, + device=device, + ) + ) + + model.to(device) + model.eval() + + num_param = sum([p.numel() for p in model.parameters()]) + logging.info(f"Number of model parameters: {num_param}") + + # we need cut ids to display recognition results. + args.return_cuts = True + ami = AmiAsrDataModule(args) + + # NOTE(@desh2608): we filter segments longer than 120s to avoid OOM errors in decoding. + # However, 99.9% of the segments are shorter than 120s, so this should not + # substantially affect the results. In future, we will implement an overlapped + # inference method to avoid OOM errors. + + test_sets = {} + for split in ["dev", "test"]: + for type in ["ihm-mix", "sdm", "mdm8-bf"]: + test_sets[f"ami-{split}_{type}"] = ( + ami.ami_cuts(split=split, type=type) + .trim_to_supervision_groups(max_pause=0.0) + .filter(lambda c: 0.1 < c.duration < 120.0) + .to_eager() + ) + + for split in ["dev", "test"]: + for type in ["ihm-mix", "sdm"]: + test_sets[f"icsi-{split}_{type}"] = ( + ami.icsi_cuts(split=split, type=type) + .trim_to_supervision_groups(max_pause=0.0) + .filter(lambda c: 0.1 < c.duration < 120.0) + .to_eager() + ) + + for test_set, test_cuts in test_sets.items(): + test_dl = ami.test_dataloaders(test_cuts) + results_dict = decode_dataset( + dl=test_dl, + params=params, + model=model, + sp=sp, + ) + + save_results( + params=params, + test_set_name=test_set, + results_dict=results_dict, + ) + + logging.info("Done!") + + +if __name__ == "__main__": + main() diff --git a/egs/ami/SURT/dprnn_zipformer/decoder.py b/egs/ami/SURT/dprnn_zipformer/decoder.py new file mode 120000 index 000000000..c34865c25 --- /dev/null +++ b/egs/ami/SURT/dprnn_zipformer/decoder.py @@ -0,0 +1 @@ +../../../libricss/SURT/dprnn_zipformer/decoder.py \ No newline at end of file diff --git a/egs/ami/SURT/dprnn_zipformer/dprnn.py b/egs/ami/SURT/dprnn_zipformer/dprnn.py new file mode 120000 index 000000000..8918beb32 --- /dev/null +++ b/egs/ami/SURT/dprnn_zipformer/dprnn.py @@ -0,0 +1 @@ +../../../libricss/SURT/dprnn_zipformer/dprnn.py \ No newline at end of file diff --git a/egs/ami/SURT/dprnn_zipformer/encoder_interface.py b/egs/ami/SURT/dprnn_zipformer/encoder_interface.py new file mode 120000 index 000000000..0ba945d0f --- /dev/null +++ b/egs/ami/SURT/dprnn_zipformer/encoder_interface.py @@ -0,0 +1 @@ +../../../libricss/SURT/dprnn_zipformer/encoder_interface.py \ No newline at end of file diff --git a/egs/ami/SURT/dprnn_zipformer/export.py b/egs/ami/SURT/dprnn_zipformer/export.py new file mode 120000 index 000000000..3deae4471 --- /dev/null +++ b/egs/ami/SURT/dprnn_zipformer/export.py @@ -0,0 +1 @@ +../../../libricss/SURT/dprnn_zipformer/export.py \ No newline at end of file diff --git a/egs/ami/SURT/dprnn_zipformer/joiner.py b/egs/ami/SURT/dprnn_zipformer/joiner.py new file mode 120000 index 000000000..79fbe8769 --- /dev/null +++ b/egs/ami/SURT/dprnn_zipformer/joiner.py @@ -0,0 +1 @@ +../../../libricss/SURT/dprnn_zipformer/joiner.py \ No newline at end of file diff --git a/egs/ami/SURT/dprnn_zipformer/model.py b/egs/ami/SURT/dprnn_zipformer/model.py new file mode 120000 index 000000000..ae8c65c99 --- /dev/null +++ b/egs/ami/SURT/dprnn_zipformer/model.py @@ -0,0 +1 @@ +../../../libricss/SURT/dprnn_zipformer/model.py \ No newline at end of file diff --git a/egs/ami/SURT/dprnn_zipformer/optim.py b/egs/ami/SURT/dprnn_zipformer/optim.py new file mode 120000 index 000000000..366d0f7a2 --- /dev/null +++ b/egs/ami/SURT/dprnn_zipformer/optim.py @@ -0,0 +1 @@ +../../../libricss/SURT/dprnn_zipformer/optim.py \ No newline at end of file diff --git a/egs/ami/SURT/dprnn_zipformer/scaling.py b/egs/ami/SURT/dprnn_zipformer/scaling.py new file mode 120000 index 000000000..f11d49d77 --- /dev/null +++ b/egs/ami/SURT/dprnn_zipformer/scaling.py @@ -0,0 +1 @@ +../../../libricss/SURT/dprnn_zipformer/scaling.py \ No newline at end of file diff --git a/egs/ami/SURT/dprnn_zipformer/scaling_converter.py b/egs/ami/SURT/dprnn_zipformer/scaling_converter.py new file mode 120000 index 000000000..1533cbe0e --- /dev/null +++ b/egs/ami/SURT/dprnn_zipformer/scaling_converter.py @@ -0,0 +1 @@ +../../../libricss/SURT/dprnn_zipformer/scaling_converter.py \ No newline at end of file diff --git a/egs/ami/SURT/dprnn_zipformer/test_model.py b/egs/ami/SURT/dprnn_zipformer/test_model.py new file mode 120000 index 000000000..1259849e0 --- /dev/null +++ b/egs/ami/SURT/dprnn_zipformer/test_model.py @@ -0,0 +1 @@ +../../../librispeech/ASR/pruned_transducer_stateless7_streaming/test_model.py \ No newline at end of file diff --git a/egs/ami/SURT/dprnn_zipformer/train.py b/egs/ami/SURT/dprnn_zipformer/train.py new file mode 100755 index 000000000..cd5fafc34 --- /dev/null +++ b/egs/ami/SURT/dprnn_zipformer/train.py @@ -0,0 +1,1420 @@ +#!/usr/bin/env python3 +# Copyright 2021 Xiaomi Corp. (authors: Fangjun Kuang, +# Wei Kang, +# Mingshuang Luo,) +# Zengwei Yao) +# +# See ../../../../LICENSE for clarification regarding multiple authors +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +""" +Usage: + +export CUDA_VISIBLE_DEVICES="0,1,2,3" + +cd egs/ami/SURT/ +./prepare.sh + +./dprnn_zipformer/train.py \ + --world-size 4 \ + --num-epochs 30 \ + --start-epoch 1 \ + --use-fp16 1 \ + --exp-dir dprnn_zipformer/exp \ + --max-duration 650 +""" + +import argparse +import copy +import logging +import warnings +from itertools import chain +from pathlib import Path +from shutil import copyfile +from typing import Any, Dict, Optional, Tuple, Union + +import k2 +import optim +import sentencepiece as spm +import torch +import torch.multiprocessing as mp +import torch.nn as nn +from asr_datamodule import AmiAsrDataModule +from decoder import Decoder +from dprnn import DPRNN +from einops.layers.torch import Rearrange +from joiner import Joiner +from lhotse.cut import Cut +from lhotse.dataset.sampling.base import CutSampler +from lhotse.utils import LOG_EPSILON, fix_random_seed +from model import SURT +from optim import Eden, ScaledAdam +from scaling import ScaledLinear, ScaledLSTM +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 Zipformer + +from icefall import diagnostics +from icefall.checkpoint import load_checkpoint, remove_checkpoints +from icefall.checkpoint import save_checkpoint as save_checkpoint_impl +from icefall.checkpoint import ( + save_checkpoint_with_global_batch_idx, + update_averaged_model, +) +from icefall.dist import cleanup_dist, setup_dist +from icefall.env import get_env_info +from icefall.utils import AttributeDict, MetricsTracker, setup_logger, str2bool + +LRSchedulerType = Union[torch.optim.lr_scheduler._LRScheduler, optim.LRScheduler] + + +def set_batch_count(model: Union[nn.Module, DDP], batch_count: float) -> None: + if isinstance(model, DDP): + # get underlying nn.Module + model = model.module + for module in model.modules(): + if hasattr(module, "batch_count"): + module.batch_count = batch_count + + +def add_model_arguments(parser: argparse.ArgumentParser): + parser.add_argument( + "--num-mask-encoder-layers", + type=int, + default=4, + help="Number of layers in the DPRNN based mask encoder.", + ) + + parser.add_argument( + "--mask-encoder-dim", + type=int, + default=256, + help="Hidden dimension of the LSTM blocks in DPRNN.", + ) + + parser.add_argument( + "--mask-encoder-segment-size", + type=int, + default=32, + help="Segment size of the SegLSTM in DPRNN. Ideally, this should be equal to the " + "decode-chunk-length of the zipformer encoder.", + ) + + parser.add_argument( + "--chunk-width-randomization", + type=bool, + default=False, + help="Whether to randomize the chunk width in DPRNN.", + ) + + # Zipformer config is based on: + # https://github.com/k2-fsa/icefall/pull/745#issuecomment-1405282740 + parser.add_argument( + "--num-encoder-layers", + type=str, + default="2,2,2,2,2", + help="Number of zipformer encoder layers, comma separated.", + ) + + parser.add_argument( + "--feedforward-dims", + type=str, + default="768,768,768,768,768", + help="Feedforward dimension of the zipformer encoder layers, comma separated.", + ) + + parser.add_argument( + "--nhead", + type=str, + default="8,8,8,8,8", + help="Number of attention heads in the zipformer encoder layers.", + ) + + parser.add_argument( + "--encoder-dims", + type=str, + default="256,256,256,256,256", + help="Embedding dimension in the 2 blocks of zipformer encoder layers, comma separated", + ) + + parser.add_argument( + "--attention-dims", + type=str, + default="192,192,192,192,192", + help="""Attention dimension in the 2 blocks of zipformer encoder layers, comma separated; + not the same as embedding dimension.""", + ) + + parser.add_argument( + "--encoder-unmasked-dims", + type=str, + default="192,192,192,192,192", + help="Unmasked dimensions in the encoders, relates to augmentation during training. " + "Must be <= each of encoder_dims. Empirically, less than 256 seems to make performance " + " worse.", + ) + + parser.add_argument( + "--zipformer-downsampling-factors", + type=str, + default="1,2,4,8,2", + help="Downsampling factor for each stack of encoder layers.", + ) + + parser.add_argument( + "--cnn-module-kernels", + type=str, + default="31,31,31,31,31", + help="Sizes of kernels in convolution modules", + ) + + parser.add_argument( + "--use-joint-encoder-layer", + type=str, + default="lstm", + choices=["linear", "lstm", "none"], + help="Whether to use a joint layer to combine all branches.", + ) + + 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( + "--short-chunk-size", + type=int, + default=50, + help="""Chunk length of dynamic training, the chunk size would be either + max sequence length of current batch or uniformly sampled from (1, short_chunk_size). + """, + ) + + parser.add_argument( + "--num-left-chunks", + type=int, + default=4, + help="How many left context can be seen in chunks when calculating attention.", + ) + + parser.add_argument( + "--decode-chunk-len", + type=int, + default=32, + help="The chunk size for decoding (in frames before subsampling)", + ) + + +def get_parser(): + parser = argparse.ArgumentParser( + formatter_class=argparse.ArgumentDefaultsHelpFormatter + ) + + parser.add_argument( + "--world-size", + type=int, + default=1, + help="Number of GPUs for DDP training.", + ) + + parser.add_argument( + "--master-port", + type=int, + default=12354, + help="Master port to use for DDP training.", + ) + + parser.add_argument( + "--tensorboard", + type=str2bool, + default=True, + help="Should various information be logged in tensorboard.", + ) + + parser.add_argument( + "--num-epochs", + type=int, + default=30, + help="Number of epochs to train.", + ) + + parser.add_argument( + "--start-epoch", + type=int, + default=1, + help="""Resume training from this epoch. It should be positive. + If larger than 1, it will load checkpoint from + exp-dir/epoch-{start_epoch-1}.pt + """, + ) + + parser.add_argument( + "--start-batch", + type=int, + default=0, + help="""If positive, --start-epoch is ignored and + it loads the checkpoint from exp-dir/checkpoint-{start_batch}.pt + """, + ) + + parser.add_argument( + "--exp-dir", + type=str, + default="conv_lstm_transducer_stateless_ctc/exp", + help="""The experiment dir. + It specifies the directory where all training related + files, e.g., checkpoints, log, etc, are saved + """, + ) + + parser.add_argument( + "--model-init-ckpt", + type=str, + default=None, + help="""The model checkpoint to initialize the model (either full or part). + If not specified, the model is randomly initialized. + """, + ) + + 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=0.004, help="The base learning rate." + ) + + parser.add_argument( + "--lr-batches", + type=float, + default=5000, + help="""Number of steps that affects how rapidly the learning rate + decreases. We suggest not to change this.""", + ) + + parser.add_argument( + "--lr-epochs", + type=float, + default=5, + help="""Number of epochs that affects how rapidly the learning rate decreases. + """, + ) + + parser.add_argument( + "--context-size", + type=int, + default=2, + help="The context size in the decoder. 1 means bigram; 2 means tri-gram", + ) + + parser.add_argument( + "--prune-range", + type=int, + default=5, + help="The prune range for rnnt loss, it means how many symbols(context)" + "we are using to compute the loss", + ) + + parser.add_argument( + "--lm-scale", + type=float, + default=0.25, + help="The scale to smooth the loss with lm " + "(output of prediction network) part.", + ) + + parser.add_argument( + "--am-scale", + type=float, + default=0.0, + help="The scale to smooth the loss with am (output of encoder network) part.", + ) + + parser.add_argument( + "--simple-loss-scale", + type=float, + default=0.5, + help="To get pruning ranges, we will calculate a simple version" + "loss(joiner is just addition), this simple loss also uses for" + "training (as a regularization item). We will scale the simple loss" + "with this parameter before adding to the final loss.", + ) + + parser.add_argument( + "--ctc-loss-scale", + type=float, + default=0.2, + help="Scale for CTC loss.", + ) + + parser.add_argument( + "--heat-loss-scale", + type=float, + default=0.2, + help="Scale for HEAT loss on separated sources.", + ) + + parser.add_argument( + "--seed", + type=int, + default=42, + help="The seed for random generators intended for reproducibility", + ) + + parser.add_argument( + "--print-diagnostics", + type=str2bool, + default=False, + help="Accumulate stats on activations, print them and exit.", + ) + + parser.add_argument( + "--save-every-n", + type=int, + default=2000, + help="""Save checkpoint after processing this number of batches" + periodically. We save checkpoint to exp-dir/ whenever + params.batch_idx_train % save_every_n == 0. The checkpoint filename + has the form: f'exp-dir/checkpoint-{params.batch_idx_train}.pt' + Note: It also saves checkpoint to `exp-dir/epoch-xxx.pt` at the + end of each epoch where `xxx` is the epoch number counting from 0. + """, + ) + + parser.add_argument( + "--keep-last-k", + type=int, + default=1, + help="""Only keep this number of checkpoints on disk. + For instance, if it is 3, there are only 3 checkpoints + in the exp-dir with filenames `checkpoint-xxx.pt`. + It does not affect checkpoints with name `epoch-xxx.pt`. + """, + ) + + parser.add_argument( + "--average-period", + type=int, + default=100, + help="""Update the averaged model, namely `model_avg`, after processing + this number of batches. `model_avg` is a separate version of model, + in which each floating-point parameter is the average of all the + parameters from the start of training. Each time we take the average, + we do: `model_avg = model * (average_period / batch_idx_train) + + model_avg * ((batch_idx_train - average_period) / batch_idx_train)`. + """, + ) + + parser.add_argument( + "--use-fp16", + type=str2bool, + default=False, + help="Whether to use half precision training.", + ) + + add_model_arguments(parser) + + return parser + + +def get_params() -> AttributeDict: + """Return a dict containing training parameters. + + All training related parameters that are not passed from the commandline + are saved in the variable `params`. + + Commandline options are merged into `params` after they are parsed, so + you can also access them via `params`. + + Explanation of options saved in `params`: + + - best_train_loss: Best training loss so far. It is used to select + the model that has the lowest training loss. It is + updated during the training. + + - best_valid_loss: Best validation loss so far. It is used to select + the model that has the lowest validation loss. It is + updated during the training. + + - best_train_epoch: It is the epoch that has the best training loss. + + - best_valid_epoch: It is the epoch that has the best validation loss. + + - batch_idx_train: Used to writing statistics to tensorboard. It + contains number of batches trained so far across + epochs. + + - log_interval: Print training loss if batch_idx % log_interval` is 0 + + - reset_interval: Reset statistics if batch_idx % reset_interval is 0 + + - valid_interval: Run validation if batch_idx % valid_interval is 0 + + - feature_dim: The model input dim. It has to match the one used + in computing features. + + - subsampling_factor: The subsampling factor for the model. + + - num_decoder_layers: Number of decoder layer of transformer decoder. + + - warm_step: The warm_step for Noam optimizer. + """ + params = AttributeDict( + { + "best_train_loss": float("inf"), + "best_valid_loss": float("inf"), + "best_train_epoch": -1, + "best_valid_epoch": -1, + "batch_idx_train": 0, + "log_interval": 50, + "reset_interval": 200, + "valid_interval": 2000, + # parameters for SURT + "num_channels": 2, + "feature_dim": 80, + "subsampling_factor": 4, # not passed in, this is fixed + # parameters for Noam + "model_warm_step": 5000, # arg given to model, not for lrate + # parameters for ctc loss + "beam_size": 10, + "use_double_scores": True, + "env_info": get_env_info(), + } + ) + + return params + + +def get_mask_encoder_model(params: AttributeDict) -> nn.Module: + mask_encoder = DPRNN( + feature_dim=params.feature_dim, + input_size=params.mask_encoder_dim, + hidden_size=params.mask_encoder_dim, + output_size=params.feature_dim * params.num_channels, + segment_size=params.mask_encoder_segment_size, + num_blocks=params.num_mask_encoder_layers, + chunk_width_randomization=params.chunk_width_randomization, + ) + return mask_encoder + + +def get_encoder_model(params: AttributeDict) -> nn.Module: + # TODO: We can add an option to switch between Zipformer and Transformer + def to_int_tuple(s: str): + return tuple(map(int, s.split(","))) + + encoder = Zipformer( + num_features=params.feature_dim, + output_downsampling_factor=2, + zipformer_downsampling_factors=to_int_tuple( + params.zipformer_downsampling_factors + ), + encoder_dims=to_int_tuple(params.encoder_dims), + attention_dim=to_int_tuple(params.attention_dims), + encoder_unmasked_dims=to_int_tuple(params.encoder_unmasked_dims), + nhead=to_int_tuple(params.nhead), + feedforward_dim=to_int_tuple(params.feedforward_dims), + cnn_module_kernels=to_int_tuple(params.cnn_module_kernels), + num_encoder_layers=to_int_tuple(params.num_encoder_layers), + num_left_chunks=params.num_left_chunks, + short_chunk_size=params.short_chunk_size, + decode_chunk_size=params.decode_chunk_len // 2, + ) + return encoder + + +def get_joint_encoder_layer(params: AttributeDict) -> nn.Module: + class TakeFirst(nn.Module): + def forward(self, x): + return x[0] + + if params.use_joint_encoder_layer == "linear": + encoder_dim = int(params.encoder_dims.split(",")[-1]) + joint_layer = nn.Sequential( + Rearrange("(c b) t d -> b t (c d)", c=params.num_channels), + nn.Linear( + params.num_channels * encoder_dim, params.num_channels * encoder_dim + ), + nn.ReLU(), + Rearrange("b t (c d) -> (c b) t d", c=params.num_channels), + ) + elif params.use_joint_encoder_layer == "lstm": + encoder_dim = int(params.encoder_dims.split(",")[-1]) + joint_layer = nn.Sequential( + Rearrange("(c b) t d -> b t (c d)", c=params.num_channels), + ScaledLSTM( + input_size=params.num_channels * encoder_dim, + hidden_size=params.num_channels * encoder_dim, + num_layers=1, + bias=True, + batch_first=True, + dropout=0.0, + bidirectional=False, + ), + TakeFirst(), + nn.ReLU(), + Rearrange("b t (c d) -> (c b) t d", c=params.num_channels), + ) + elif params.use_joint_encoder_layer == "none": + joint_layer = None + else: + raise ValueError( + f"Unknown joint encoder layer type: {params.use_joint_encoder_layer}" + ) + return joint_layer + + +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=int(params.encoder_dims.split(",")[-1]), + decoder_dim=params.decoder_dim, + joiner_dim=params.joiner_dim, + vocab_size=params.vocab_size, + ) + return joiner + + +def get_surt_model( + params: AttributeDict, +) -> nn.Module: + mask_encoder = get_mask_encoder_model(params) + encoder = get_encoder_model(params) + joint_layer = get_joint_encoder_layer(params) + decoder = get_decoder_model(params) + joiner = get_joiner_model(params) + + model = SURT( + mask_encoder=mask_encoder, + encoder=encoder, + joint_encoder_layer=joint_layer, + decoder=decoder, + joiner=joiner, + num_channels=params.num_channels, + encoder_dim=int(params.encoder_dims.split(",")[-1]), + decoder_dim=params.decoder_dim, + joiner_dim=params.joiner_dim, + vocab_size=params.vocab_size, + ) + return model + + +def load_checkpoint_if_available( + params: AttributeDict, + model: nn.Module, + model_avg: nn.Module = None, + optimizer: Optional[torch.optim.Optimizer] = None, + scheduler: Optional[LRSchedulerType] = None, +) -> Optional[Dict[str, Any]]: + """Load checkpoint from file. + + If params.start_batch is positive, it will load the checkpoint from + `params.exp_dir/checkpoint-{params.start_batch}.pt`. Otherwise, if + params.start_epoch is larger than 1, it will load the checkpoint from + `params.start_epoch - 1`. + + Apart from loading state dict for `model` and `optimizer` it also updates + `best_train_epoch`, `best_train_loss`, `best_valid_epoch`, + and `best_valid_loss` in `params`. + + Args: + params: + The return value of :func:`get_params`. + model: + The training model. + model_avg: + The stored model averaged from the start of training. + optimizer: + The optimizer that we are using. + scheduler: + The scheduler that we are using. + Returns: + Return a dict containing previously saved training info. + """ + if params.start_batch > 0: + filename = params.exp_dir / f"checkpoint-{params.start_batch}.pt" + elif params.start_epoch > 1: + filename = params.exp_dir / f"epoch-{params.start_epoch-1}.pt" + else: + return None + + assert filename.is_file(), f"{filename} does not exist!" + + saved_params = load_checkpoint( + filename, + model=model, + model_avg=model_avg, + optimizer=optimizer, + scheduler=scheduler, + ) + + keys = [ + "best_train_epoch", + "best_valid_epoch", + "batch_idx_train", + "best_train_loss", + "best_valid_loss", + ] + for k in keys: + params[k] = saved_params[k] + + if params.start_batch > 0: + if "cur_epoch" in saved_params: + params["start_epoch"] = saved_params["cur_epoch"] + + return saved_params + + +def save_checkpoint( + params: AttributeDict, + model: Union[nn.Module, DDP], + model_avg: Optional[nn.Module] = None, + optimizer: Optional[torch.optim.Optimizer] = None, + scheduler: Optional[LRSchedulerType] = None, + sampler: Optional[CutSampler] = None, + scaler: Optional[GradScaler] = None, + rank: int = 0, +) -> None: + """Save model, optimizer, scheduler and training stats to file. + + Args: + params: + It is returned by :func:`get_params`. + model: + The training model. + model_avg: + The stored model averaged from the start of training. + optimizer: + The optimizer used in the training. + sampler: + The sampler for the training dataset. + scaler: + The scaler used for mix precision training. + """ + if rank != 0: + return + filename = params.exp_dir / f"epoch-{params.cur_epoch}.pt" + save_checkpoint_impl( + filename=filename, + model=model, + model_avg=model_avg, + params=params, + optimizer=optimizer, + scheduler=scheduler, + sampler=sampler, + scaler=scaler, + rank=rank, + ) + + if params.best_train_epoch == params.cur_epoch: + best_train_filename = params.exp_dir / "best-train-loss.pt" + copyfile(src=filename, dst=best_train_filename) + + if params.best_valid_epoch == params.cur_epoch: + best_valid_filename = params.exp_dir / "best-valid-loss.pt" + copyfile(src=filename, dst=best_valid_filename) + + +def compute_heat_loss(x_masked, batch, num_channels=2) -> Tensor: + """ + Compute HEAT loss for separated sources using the output of mask encoder. + Args: + x_masked: + The output of mask encoder. It is a tensor of shape (B, T, C). + batch: + A batch of data. See `lhotse.dataset.K2SurtDatasetWithSources()` + for the content in it. + num_channels: + The number of output branches in the SURT model. + """ + B, T, D = x_masked[0].shape + device = x_masked[0].device + + # Create training targets for each channel. + targets = [] + for i in range(num_channels): + target = torch.ones_like(x_masked[i]) * LOG_EPSILON + targets.append(target) + + source_feats = batch["source_feats"] + source_boundaries = batch["source_boundaries"] + input_lens = batch["input_lens"].to(device) + # Assign sources to channels based on the HEAT criteria + for b in range(B): + cut_source_feats = source_feats[b] + cut_source_boundaries = source_boundaries[b] + last_seg_end = [0 for _ in range(num_channels)] + for source_feat, (start, end) in zip(cut_source_feats, cut_source_boundaries): + assigned = False + end = min(end, T) + source_feat = source_feat[: end - start, :] + for i in range(num_channels): + if start >= last_seg_end[i]: + targets[i][b, start:end, :] += source_feat.to(device) + last_seg_end[i] = max(end, last_seg_end[i]) + assigned = True + break + if not assigned: + min_end_channel = last_seg_end.index(min(last_seg_end)) + targets[min_end_channel][b, start:end, :] += source_feat.to(device) + last_seg_end[min_end_channel] = max(end, last_seg_end[min_end_channel]) + + # Get padding mask based on input lengths + pad_mask = torch.arange(T, device=device).expand(B, T) > input_lens.unsqueeze(1) + pad_mask = pad_mask.unsqueeze(-1) + + # Compute masked loss for each channel + losses = torch.zeros((num_channels, B, T, D), device=device) + for i in range(num_channels): + loss = nn.functional.mse_loss(x_masked[i], targets[i], reduction="none") + # Apply padding mask to loss + loss.masked_fill_(pad_mask, 0) + losses[i] = loss + + # loss: C x B x T x D. pad_mask: B x T x 1 + # We want to compute loss for each item in the batch. Each item has loss given + # by the sum over C, and average over T and D. For T, we need to use the padding. + loss = losses.sum(0).mean(-1).sum(-1) / batch["input_lens"].to(device) + return loss + + +def compute_loss( + params: AttributeDict, + model: Union[nn.Module, DDP], + sp: spm.SentencePieceProcessor, + batch: dict, + is_training: bool, +) -> Tuple[Tensor, MetricsTracker]: + """ + Compute RNN-T loss given the model and its inputs. + + Args: + params: + Parameters for training. See :func:`get_params`. + model: + The model for training. It is an instance of Conformer in our case. + batch: + A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()` + for the content in it. + is_training: + True for training. False for validation. When it is True, this + function enables autograd during computation; when it is False, it + disables autograd. + """ + device = model.device if isinstance(model, DDP) else next(model.parameters()).device + feature = batch["inputs"].to(device) + feature_lens = batch["input_lens"].to(device) + + # at entry, feature is (N, T, C) + assert feature.ndim == 3 + + # The dataloader returns text as a list of cuts, each of which is a list of channel + # text. We flatten this to a list where all channels are together, i.e., it looks like + # [utt1_ch1, utt2_ch1, ..., uttN_ch1, utt1_ch2, ...., uttN,ch2]. + text = [val for tup in zip(*batch["text"]) for val in tup] + assert len(text) == len(feature) * params.num_channels + + # Convert all channel texts to token IDs and create a ragged tensor. + y = sp.encode(text, out_type=int) + y = k2.RaggedTensor(y).to(device) + + batch_idx_train = params.batch_idx_train + warm_step = params.model_warm_step + + with torch.set_grad_enabled(is_training): + (simple_loss, pruned_loss, ctc_loss, x_masked) = model( + x=feature, + x_lens=feature_lens, + y=y, + prune_range=params.prune_range, + am_scale=params.am_scale, + lm_scale=params.lm_scale, + reduction="none", + subsampling_factor=params.subsampling_factor, + ) + simple_loss_is_finite = torch.isfinite(simple_loss) + pruned_loss_is_finite = torch.isfinite(pruned_loss) + ctc_loss_is_finite = torch.isfinite(ctc_loss) + + # Compute HEAT loss + if is_training and params.heat_loss_scale > 0.0: + heat_loss = compute_heat_loss( + x_masked, batch, num_channels=params.num_channels + ) + else: + heat_loss = torch.tensor(0.0, device=device) + + heat_loss_is_finite = torch.isfinite(heat_loss) + is_finite = ( + simple_loss_is_finite + & pruned_loss_is_finite + & ctc_loss_is_finite + & heat_loss_is_finite + ) + if not torch.all(is_finite): + logging.info( + "Not all losses are finite!\n" + f"simple_losses: {simple_loss}\n" + f"pruned_losses: {pruned_loss}\n" + f"ctc_losses: {ctc_loss}\n" + f"heat_losses: {heat_loss}\n" + ) + display_and_save_batch(batch, params=params, sp=sp) + simple_loss = simple_loss[simple_loss_is_finite] + pruned_loss = pruned_loss[pruned_loss_is_finite] + ctc_loss = ctc_loss[ctc_loss_is_finite] + heat_loss = heat_loss[heat_loss_is_finite] + + # If either all simple_loss or pruned_loss is inf or nan, + # we stop the training process by raising an exception + if ( + torch.all(~simple_loss_is_finite) + or torch.all(~pruned_loss_is_finite) + or torch.all(~ctc_loss_is_finite) + or torch.all(~heat_loss_is_finite) + ): + raise ValueError( + "There are too many utterances in this batch " + "leading to inf or nan losses." + ) + + simple_loss_sum = simple_loss.sum() + pruned_loss_sum = pruned_loss.sum() + ctc_loss_sum = ctc_loss.sum() + heat_loss_sum = heat_loss.sum() + + 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_sum + + pruned_loss_scale * pruned_loss_sum + + params.ctc_loss_scale * ctc_loss_sum + + params.heat_loss_scale * heat_loss_sum + ) + + assert loss.requires_grad == is_training + + info = MetricsTracker() + with warnings.catch_warnings(): + warnings.simplefilter("ignore") + # info["frames"] is an approximate number for two reasons: + # (1) The acutal subsampling factor is ((lens - 1) // 2 - 1) // 2 + # (2) If some utterances in the batch lead to inf/nan loss, they + # are filtered out. + info["frames"] = (feature_lens // params.subsampling_factor).sum().item() + + # `utt_duration` and `utt_pad_proportion` would be normalized by `utterances` # noqa + info["utterances"] = feature.size(0) + # averaged input duration in frames over utterances + info["utt_duration"] = feature_lens.sum().item() + # averaged padding proportion over utterances + info["utt_pad_proportion"] = ( + ((feature.size(1) - feature_lens) / feature.size(1)).sum().item() + ) + + # Note: We use reduction=sum while computing the loss. + info["loss"] = loss.detach().cpu().item() + info["simple_loss"] = simple_loss_sum.detach().cpu().item() + info["pruned_loss"] = pruned_loss_sum.detach().cpu().item() + if params.ctc_loss_scale > 0.0: + info["ctc_loss"] = ctc_loss_sum.detach().cpu().item() + if params.heat_loss_scale > 0.0: + info["heat_loss"] = heat_loss_sum.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. + """ + torch.cuda.empty_cache() + model.train() + + tot_loss = MetricsTracker() + + cur_batch_idx = params.get("cur_batch_idx", 0) + + for batch_idx, batch in enumerate(train_dl): + if batch_idx < cur_batch_idx: + continue + cur_batch_idx = batch_idx + + params.batch_idx_train += 1 + batch_size = batch["inputs"].shape[0] + + 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() + set_batch_count(model, params.batch_idx_train) + scheduler.step_batch(params.batch_idx_train) + + scaler.step(optimizer) + scaler.update() + optimizer.zero_grad() + except: # noqa + display_and_save_batch(batch, params=params, 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 < 1.0 or (cur_grad_scale < 8.0 and batch_idx % 400 == 0): + scaler.update(cur_grad_scale * 2.0) + if cur_grad_scale < 0.01: + logging.warning(f"Grad scale is small: {cur_grad_scale}") + if cur_grad_scale < 1.0e-05: + raise RuntimeError( + f"grad_scale is too small, exiting: {cur_grad_scale}" + ) + + if batch_idx % params.log_interval == 0: + cur_lr = scheduler.get_last_lr()[0] + cur_grad_scale = scaler._scale.item() if params.use_fp16 else 1.0 + + logging.info( + f"Epoch {params.cur_epoch}, " + f"batch {batch_idx}, loss[{loss_info}], " + f"tot_loss[{tot_loss}], batch size: {batch_size}, " + f"lr: {cur_lr:.2e}, " + + (f"grad_scale: {scaler._scale.item()}" if params.use_fp16 else "") + ) + + if tb_writer is not None: + tb_writer.add_scalar( + "train/learning_rate", cur_lr, params.batch_idx_train + ) + + loss_info.write_summary( + tb_writer, "train/current_", params.batch_idx_train + ) + tot_loss.write_summary(tb_writer, "train/tot_", params.batch_idx_train) + if params.use_fp16: + tb_writer.add_scalar( + "train/grad_scale", cur_grad_scale, params.batch_idx_train + ) + + if batch_idx % params.valid_interval == 0 and not params.print_diagnostics: + logging.info("Computing validation loss") + valid_info = compute_validation_loss( + params=params, + model=model, + 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_surt_model(params) + + num_param = sum([p.numel() for p in model.parameters()]) + logging.info(f"Number of model parameters: {num_param}") + + assert params.save_every_n >= params.average_period + model_avg: Optional[nn.Module] = None + if rank == 0: + # model_avg is only used with rank 0 + model_avg = copy.deepcopy(model) + + assert params.start_epoch > 0, params.start_epoch + checkpoints = load_checkpoint_if_available( + params=params, model=model, model_avg=model_avg + ) + + model.to(device) + + if checkpoints is None and params.model_init_ckpt is not None: + logging.info( + f"Initializing model with checkpoint from {params.model_init_ckpt}" + ) + init_ckpt = torch.load(params.model_init_ckpt, map_location=device) + model.load_state_dict(init_ckpt["model"], strict=False) + + if world_size > 1: + logging.info("Using DDP") + model = DDP(model, device_ids=[rank], find_unused_parameters=True) + + parameters_names = [] + parameters_names.append( + [name_param_pair[0] for name_param_pair in model.named_parameters()] + ) + optimizer = ScaledAdam( + model.parameters(), + lr=params.base_lr, + clipping_scale=2.0, + parameters_names=parameters_names, + ) + + scheduler = Eden(optimizer, params.lr_batches, params.lr_epochs) + + if checkpoints and "optimizer" in checkpoints: + logging.info("Loading optimizer state dict") + optimizer.load_state_dict(checkpoints["optimizer"]) + + if ( + checkpoints + and "scheduler" in checkpoints + and checkpoints["scheduler"] is not None + ): + logging.info("Loading scheduler state dict") + scheduler.load_state_dict(checkpoints["scheduler"]) + + if params.print_diagnostics: + diagnostic = diagnostics.attach_diagnostics(model) + + ami = AmiAsrDataModule(args) + + train_cuts = ami.aimix_train_cuts(rvb_affix="comb", sources=True) + dev_cuts = ami.ami_cuts(split="dev", type="ihm-mix") + dev_cuts = dev_cuts.trim_to_supervision_groups(max_pause=0.0).filter( + lambda c: 0.2 <= c.duration <= 60.0 + ) + + if params.start_batch > 0 and checkpoints and "sampler" in checkpoints: + # We only load the sampler's state dict when it loads a checkpoint + # saved in the middle of an epoch + sampler_state_dict = checkpoints["sampler"] + else: + sampler_state_dict = None + + train_dl = ami.train_dataloaders( + train_cuts, + sampler_state_dict=sampler_state_dict, + sources=True, + ) + valid_dl = ami.valid_dataloaders(dev_cuts) + + 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) + + features = batch["inputs"] + + logging.info(f"features shape: {features.shape}") + + y = [sp.encode(text_ch) for text_ch in batch["text"]] + num_tokens = [sum(len(yi) for yi in y_ch) for y_ch in y] + logging.info(f"num tokens: {num_tokens}") + + +def main(): + parser = get_parser() + AmiAsrDataModule.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) +torch.multiprocessing.set_sharing_strategy("file_system") + +if __name__ == "__main__": + main() diff --git a/egs/ami/SURT/dprnn_zipformer/train_adapt.py b/egs/ami/SURT/dprnn_zipformer/train_adapt.py new file mode 100755 index 000000000..9f3b4425f --- /dev/null +++ b/egs/ami/SURT/dprnn_zipformer/train_adapt.py @@ -0,0 +1,1411 @@ +#!/usr/bin/env python3 +# Copyright 2021 Xiaomi Corp. (authors: Fangjun Kuang, +# Wei Kang, +# Mingshuang Luo,) +# Zengwei Yao) +# +# See ../../../../LICENSE for clarification regarding multiple authors +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +""" +Usage: + +# ./dprnn_zipformer/train.py should be run before this script. + +export CUDA_VISIBLE_DEVICES="0,1,2,3" + +./dprnn_zipformer/train.py \ + --world-size 4 \ + --num-epochs 30 \ + --start-epoch 1 \ + --use-fp16 1 \ + --exp-dir dprnn_zipformer/exp_adapt \ + --model-init-ckpt dprnn_zipformer/exp/epoch-30.pt \ + --max-duration 550 +""" + +import argparse +import copy +import logging +import warnings +from itertools import chain +from pathlib import Path +from shutil import copyfile +from typing import Any, Dict, Optional, Tuple, Union + +import k2 +import optim +import sentencepiece as spm +import torch +import torch.multiprocessing as mp +import torch.nn as nn +from asr_datamodule import AmiAsrDataModule +from decoder import Decoder +from dprnn import DPRNN +from einops.layers.torch import Rearrange +from joiner import Joiner +from lhotse.cut import Cut +from lhotse.dataset.sampling.base import CutSampler +from lhotse.utils import LOG_EPSILON, fix_random_seed +from model import SURT +from optim import Eden, ScaledAdam +from scaling import ScaledLinear, ScaledLSTM +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 Zipformer + +from icefall import diagnostics +from icefall.checkpoint import load_checkpoint, remove_checkpoints +from icefall.checkpoint import save_checkpoint as save_checkpoint_impl +from icefall.checkpoint import ( + save_checkpoint_with_global_batch_idx, + update_averaged_model, +) +from icefall.dist import cleanup_dist, setup_dist +from icefall.env import get_env_info +from icefall.utils import AttributeDict, MetricsTracker, setup_logger, str2bool + +LRSchedulerType = Union[torch.optim.lr_scheduler._LRScheduler, optim.LRScheduler] + + +def set_batch_count(model: Union[nn.Module, DDP], batch_count: float) -> None: + if isinstance(model, DDP): + # get underlying nn.Module + model = model.module + for module in model.modules(): + if hasattr(module, "batch_count"): + module.batch_count = batch_count + + +def add_model_arguments(parser: argparse.ArgumentParser): + parser.add_argument( + "--num-mask-encoder-layers", + type=int, + default=4, + help="Number of layers in the DPRNN based mask encoder.", + ) + + parser.add_argument( + "--mask-encoder-dim", + type=int, + default=256, + help="Hidden dimension of the LSTM blocks in DPRNN.", + ) + + parser.add_argument( + "--mask-encoder-segment-size", + type=int, + default=32, + help="Segment size of the SegLSTM in DPRNN. Ideally, this should be equal to the " + "decode-chunk-length of the zipformer encoder.", + ) + + parser.add_argument( + "--chunk-width-randomization", + type=bool, + default=False, + help="Whether to randomize the chunk width in DPRNN.", + ) + + # Zipformer config is based on: + # https://github.com/k2-fsa/icefall/pull/745#issuecomment-1405282740 + parser.add_argument( + "--num-encoder-layers", + type=str, + default="2,2,2,2,2", + help="Number of zipformer encoder layers, comma separated.", + ) + + parser.add_argument( + "--feedforward-dims", + type=str, + default="768,768,768,768,768", + help="Feedforward dimension of the zipformer encoder layers, comma separated.", + ) + + parser.add_argument( + "--nhead", + type=str, + default="8,8,8,8,8", + help="Number of attention heads in the zipformer encoder layers.", + ) + + parser.add_argument( + "--encoder-dims", + type=str, + default="256,256,256,256,256", + help="Embedding dimension in the 2 blocks of zipformer encoder layers, comma separated", + ) + + parser.add_argument( + "--attention-dims", + type=str, + default="192,192,192,192,192", + help="""Attention dimension in the 2 blocks of zipformer encoder layers, comma separated; + not the same as embedding dimension.""", + ) + + parser.add_argument( + "--encoder-unmasked-dims", + type=str, + default="192,192,192,192,192", + help="Unmasked dimensions in the encoders, relates to augmentation during training. " + "Must be <= each of encoder_dims. Empirically, less than 256 seems to make performance " + " worse.", + ) + + parser.add_argument( + "--zipformer-downsampling-factors", + type=str, + default="1,2,4,8,2", + help="Downsampling factor for each stack of encoder layers.", + ) + + parser.add_argument( + "--cnn-module-kernels", + type=str, + default="31,31,31,31,31", + help="Sizes of kernels in convolution modules", + ) + + parser.add_argument( + "--use-joint-encoder-layer", + type=str, + default="linear", + choices=["linear", "lstm", "none"], + help="Whether to use a joint layer to combine all branches.", + ) + + 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( + "--short-chunk-size", + type=int, + default=50, + help="""Chunk length of dynamic training, the chunk size would be either + max sequence length of current batch or uniformly sampled from (1, short_chunk_size). + """, + ) + + parser.add_argument( + "--num-left-chunks", + type=int, + default=4, + help="How many left context can be seen in chunks when calculating attention.", + ) + + parser.add_argument( + "--decode-chunk-len", + type=int, + default=32, + help="The chunk size for decoding (in frames before subsampling)", + ) + + +def get_parser(): + parser = argparse.ArgumentParser( + formatter_class=argparse.ArgumentDefaultsHelpFormatter + ) + + parser.add_argument( + "--world-size", + type=int, + default=1, + help="Number of GPUs for DDP training.", + ) + + parser.add_argument( + "--master-port", + type=int, + default=12354, + help="Master port to use for DDP training.", + ) + + parser.add_argument( + "--tensorboard", + type=str2bool, + default=True, + help="Should various information be logged in tensorboard.", + ) + + parser.add_argument( + "--num-epochs", + type=int, + default=20, + 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="conv_lstm_transducer_stateless_ctc/exp", + help="""The experiment dir. + It specifies the directory where all training related + files, e.g., checkpoints, log, etc, are saved + """, + ) + + parser.add_argument( + "--model-init-ckpt", + type=str, + default=None, + help="""The model checkpoint to initialize the model (either full or part). + If not specified, the model is randomly initialized. + """, + ) + + 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=0.0001, help="The base learning rate." + ) + + parser.add_argument( + "--lr-batches", + type=float, + default=5000, + help="""Number of steps that affects how rapidly the learning rate + decreases. We suggest not to change this.""", + ) + + parser.add_argument( + "--lr-epochs", + type=float, + default=2, + help="""Number of epochs that affects how rapidly the learning rate decreases. + """, + ) + + parser.add_argument( + "--context-size", + type=int, + default=2, + help="The context size in the decoder. 1 means bigram; 2 means tri-gram", + ) + + parser.add_argument( + "--prune-range", + type=int, + default=5, + help="The prune range for rnnt loss, it means how many symbols(context)" + "we are using to compute the loss", + ) + + parser.add_argument( + "--lm-scale", + type=float, + default=0.25, + help="The scale to smooth the loss with lm " + "(output of prediction network) part.", + ) + + parser.add_argument( + "--am-scale", + type=float, + default=0.0, + help="The scale to smooth the loss with am (output of encoder network) part.", + ) + + parser.add_argument( + "--simple-loss-scale", + type=float, + default=0.5, + help="To get pruning ranges, we will calculate a simple version" + "loss(joiner is just addition), this simple loss also uses for" + "training (as a regularization item). We will scale the simple loss" + "with this parameter before adding to the final loss.", + ) + + parser.add_argument( + "--ctc-loss-scale", + type=float, + default=0.2, + help="Scale for CTC loss.", + ) + + parser.add_argument( + "--seed", + type=int, + default=42, + help="The seed for random generators intended for reproducibility", + ) + + parser.add_argument( + "--print-diagnostics", + type=str2bool, + default=False, + help="Accumulate stats on activations, print them and exit.", + ) + + parser.add_argument( + "--save-every-n", + type=int, + default=2000, + help="""Save checkpoint after processing this number of batches" + periodically. We save checkpoint to exp-dir/ whenever + params.batch_idx_train % save_every_n == 0. The checkpoint filename + has the form: f'exp-dir/checkpoint-{params.batch_idx_train}.pt' + Note: It also saves checkpoint to `exp-dir/epoch-xxx.pt` at the + end of each epoch where `xxx` is the epoch number counting from 0. + """, + ) + + parser.add_argument( + "--keep-last-k", + type=int, + default=1, + help="""Only keep this number of checkpoints on disk. + For instance, if it is 3, there are only 3 checkpoints + in the exp-dir with filenames `checkpoint-xxx.pt`. + It does not affect checkpoints with name `epoch-xxx.pt`. + """, + ) + + parser.add_argument( + "--average-period", + type=int, + default=100, + help="""Update the averaged model, namely `model_avg`, after processing + this number of batches. `model_avg` is a separate version of model, + in which each floating-point parameter is the average of all the + parameters from the start of training. Each time we take the average, + we do: `model_avg = model * (average_period / batch_idx_train) + + model_avg * ((batch_idx_train - average_period) / batch_idx_train)`. + """, + ) + + parser.add_argument( + "--use-fp16", + type=str2bool, + default=False, + help="Whether to use half precision training.", + ) + + add_model_arguments(parser) + + return parser + + +def get_params() -> AttributeDict: + """Return a dict containing training parameters. + + All training related parameters that are not passed from the commandline + are saved in the variable `params`. + + Commandline options are merged into `params` after they are parsed, so + you can also access them via `params`. + + Explanation of options saved in `params`: + + - best_train_loss: Best training loss so far. It is used to select + the model that has the lowest training loss. It is + updated during the training. + + - best_valid_loss: Best validation loss so far. It is used to select + the model that has the lowest validation loss. It is + updated during the training. + + - best_train_epoch: It is the epoch that has the best training loss. + + - best_valid_epoch: It is the epoch that has the best validation loss. + + - batch_idx_train: Used to writing statistics to tensorboard. It + contains number of batches trained so far across + epochs. + + - log_interval: Print training loss if batch_idx % log_interval` is 0 + + - reset_interval: Reset statistics if batch_idx % reset_interval is 0 + + - valid_interval: Run validation if batch_idx % valid_interval is 0 + + - feature_dim: The model input dim. It has to match the one used + in computing features. + + - subsampling_factor: The subsampling factor for the model. + + - num_decoder_layers: Number of decoder layer of transformer decoder. + + - warm_step: The warm_step for Noam optimizer. + """ + params = AttributeDict( + { + "best_train_loss": float("inf"), + "best_valid_loss": float("inf"), + "best_train_epoch": -1, + "best_valid_epoch": -1, + "batch_idx_train": 0, + "log_interval": 50, + "reset_interval": 200, + "valid_interval": 2000, + # parameters for SURT + "num_channels": 2, + "feature_dim": 80, + "subsampling_factor": 4, # not passed in, this is fixed + # parameters for Noam + "model_warm_step": 5000, # arg given to model, not for lrate + # parameters for ctc loss + "beam_size": 10, + "use_double_scores": True, + "env_info": get_env_info(), + } + ) + + return params + + +def get_mask_encoder_model(params: AttributeDict) -> nn.Module: + mask_encoder = DPRNN( + feature_dim=params.feature_dim, + input_size=params.mask_encoder_dim, + hidden_size=params.mask_encoder_dim, + output_size=params.feature_dim * params.num_channels, + segment_size=params.mask_encoder_segment_size, + num_blocks=params.num_mask_encoder_layers, + chunk_width_randomization=params.chunk_width_randomization, + ) + return mask_encoder + + +def get_encoder_model(params: AttributeDict) -> nn.Module: + # TODO: We can add an option to switch between Zipformer and Transformer + def to_int_tuple(s: str): + return tuple(map(int, s.split(","))) + + encoder = Zipformer( + num_features=params.feature_dim, + output_downsampling_factor=2, + zipformer_downsampling_factors=to_int_tuple( + params.zipformer_downsampling_factors + ), + encoder_dims=to_int_tuple(params.encoder_dims), + attention_dim=to_int_tuple(params.attention_dims), + encoder_unmasked_dims=to_int_tuple(params.encoder_unmasked_dims), + nhead=to_int_tuple(params.nhead), + feedforward_dim=to_int_tuple(params.feedforward_dims), + cnn_module_kernels=to_int_tuple(params.cnn_module_kernels), + num_encoder_layers=to_int_tuple(params.num_encoder_layers), + num_left_chunks=params.num_left_chunks, + short_chunk_size=params.short_chunk_size, + decode_chunk_size=params.decode_chunk_len // 2, + ) + return encoder + + +def get_joint_encoder_layer(params: AttributeDict) -> nn.Module: + class TakeFirst(nn.Module): + def forward(self, x): + return x[0] + + if params.use_joint_encoder_layer == "linear": + encoder_dim = int(params.encoder_dims.split(",")[-1]) + joint_layer = nn.Sequential( + Rearrange("(c b) t d -> b t (c d)", c=params.num_channels), + nn.Linear( + params.num_channels * encoder_dim, params.num_channels * encoder_dim + ), + nn.ReLU(), + Rearrange("b t (c d) -> (c b) t d", c=params.num_channels), + ) + elif params.use_joint_encoder_layer == "lstm": + encoder_dim = int(params.encoder_dims.split(",")[-1]) + joint_layer = nn.Sequential( + Rearrange("(c b) t d -> b t (c d)", c=params.num_channels), + ScaledLSTM( + input_size=params.num_channels * encoder_dim, + hidden_size=params.num_channels * encoder_dim, + num_layers=1, + bias=True, + batch_first=True, + dropout=0.0, + bidirectional=False, + ), + TakeFirst(), + nn.ReLU(), + Rearrange("b t (c d) -> (c b) t d", c=params.num_channels), + ) + elif params.use_joint_encoder_layer == "none": + joint_layer = None + else: + raise ValueError( + f"Unknown joint encoder layer type: {params.use_joint_encoder_layer}" + ) + return joint_layer + + +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=int(params.encoder_dims.split(",")[-1]), + decoder_dim=params.decoder_dim, + joiner_dim=params.joiner_dim, + vocab_size=params.vocab_size, + ) + return joiner + + +def get_surt_model( + params: AttributeDict, +) -> nn.Module: + mask_encoder = get_mask_encoder_model(params) + encoder = get_encoder_model(params) + joint_layer = get_joint_encoder_layer(params) + decoder = get_decoder_model(params) + joiner = get_joiner_model(params) + + model = SURT( + mask_encoder=mask_encoder, + encoder=encoder, + joint_encoder_layer=joint_layer, + decoder=decoder, + joiner=joiner, + num_channels=params.num_channels, + encoder_dim=int(params.encoder_dims.split(",")[-1]), + decoder_dim=params.decoder_dim, + joiner_dim=params.joiner_dim, + vocab_size=params.vocab_size, + ) + return model + + +def load_checkpoint_if_available( + params: AttributeDict, + model: nn.Module, + model_avg: nn.Module = None, + optimizer: Optional[torch.optim.Optimizer] = None, + scheduler: Optional[LRSchedulerType] = None, +) -> Optional[Dict[str, Any]]: + """Load checkpoint from file. + + If params.start_batch is positive, it will load the checkpoint from + `params.exp_dir/checkpoint-{params.start_batch}.pt`. Otherwise, if + params.start_epoch is larger than 1, it will load the checkpoint from + `params.start_epoch - 1`. + + Apart from loading state dict for `model` and `optimizer` it also updates + `best_train_epoch`, `best_train_loss`, `best_valid_epoch`, + and `best_valid_loss` in `params`. + + Args: + params: + The return value of :func:`get_params`. + model: + The training model. + model_avg: + The stored model averaged from the start of training. + optimizer: + The optimizer that we are using. + scheduler: + The scheduler that we are using. + Returns: + Return a dict containing previously saved training info. + """ + if params.start_batch > 0: + filename = params.exp_dir / f"checkpoint-{params.start_batch}.pt" + elif params.start_epoch > 1: + filename = params.exp_dir / f"epoch-{params.start_epoch-1}.pt" + else: + return None + + assert filename.is_file(), f"{filename} does not exist!" + + saved_params = load_checkpoint( + filename, + model=model, + model_avg=model_avg, + optimizer=optimizer, + scheduler=scheduler, + ) + + keys = [ + "best_train_epoch", + "best_valid_epoch", + "batch_idx_train", + "best_train_loss", + "best_valid_loss", + ] + for k in keys: + params[k] = saved_params[k] + + if params.start_batch > 0: + if "cur_epoch" in saved_params: + params["start_epoch"] = saved_params["cur_epoch"] + + return saved_params + + +def save_checkpoint( + params: AttributeDict, + model: Union[nn.Module, DDP], + model_avg: Optional[nn.Module] = None, + optimizer: Optional[torch.optim.Optimizer] = None, + scheduler: Optional[LRSchedulerType] = None, + sampler: Optional[CutSampler] = None, + scaler: Optional[GradScaler] = None, + rank: int = 0, +) -> None: + """Save model, optimizer, scheduler and training stats to file. + + Args: + params: + It is returned by :func:`get_params`. + model: + The training model. + model_avg: + The stored model averaged from the start of training. + optimizer: + The optimizer used in the training. + sampler: + The sampler for the training dataset. + scaler: + The scaler used for mix precision training. + """ + if rank != 0: + return + filename = params.exp_dir / f"epoch-{params.cur_epoch}.pt" + save_checkpoint_impl( + filename=filename, + model=model, + model_avg=model_avg, + params=params, + optimizer=optimizer, + scheduler=scheduler, + sampler=sampler, + scaler=scaler, + rank=rank, + ) + + if params.best_train_epoch == params.cur_epoch: + best_train_filename = params.exp_dir / "best-train-loss.pt" + copyfile(src=filename, dst=best_train_filename) + + if params.best_valid_epoch == params.cur_epoch: + best_valid_filename = params.exp_dir / "best-valid-loss.pt" + copyfile(src=filename, dst=best_valid_filename) + + +def compute_heat_loss(x_masked, batch, num_channels=2) -> Tensor: + """ + Compute HEAT loss for separated sources using the output of mask encoder. + Args: + x_masked: + The output of mask encoder. It is a tensor of shape (B, T, C). + batch: + A batch of data. See `lhotse.dataset.K2SurtDatasetWithSources()` + for the content in it. + num_channels: + The number of output branches in the SURT model. + """ + B, T, D = x_masked[0].shape + device = x_masked[0].device + + # Create training targets for each channel. + targets = [] + for i in range(num_channels): + target = torch.ones_like(x_masked[i]) * LOG_EPSILON + targets.append(target) + + source_feats = batch["source_feats"] + source_boundaries = batch["source_boundaries"] + input_lens = batch["input_lens"].to(device) + # Assign sources to channels based on the HEAT criteria + for b in range(B): + cut_source_feats = source_feats[b] + cut_source_boundaries = source_boundaries[b] + last_seg_end = [0 for _ in range(num_channels)] + for source_feat, (start, end) in zip(cut_source_feats, cut_source_boundaries): + assigned = False + for i in range(num_channels): + if start >= last_seg_end[i]: + targets[i][b, start:end, :] += source_feat.to(device) + last_seg_end[i] = max(end, last_seg_end[i]) + assigned = True + break + if not assigned: + min_end_channel = last_seg_end.index(min(last_seg_end)) + targets[min_end_channel][b, start:end, :] += source_feat + last_seg_end[min_end_channel] = max(end, last_seg_end[min_end_channel]) + + # Get padding mask based on input lengths + pad_mask = torch.arange(T, device=device).expand(B, T) > input_lens.unsqueeze(1) + pad_mask = pad_mask.unsqueeze(-1) + + # Compute masked loss for each channel + losses = torch.zeros((num_channels, B, T, D), device=device) + for i in range(num_channels): + loss = nn.functional.mse_loss(x_masked[i], targets[i], reduction="none") + # Apply padding mask to loss + loss.masked_fill_(pad_mask, 0) + losses[i] = loss + + # loss: C x B x T x D. pad_mask: B x T x 1 + # We want to compute loss for each item in the batch. Each item has loss given + # by the sum over C, and average over T and D. For T, we need to use the padding. + loss = losses.sum(0).mean(-1).sum(-1) / batch["input_lens"].to(device) + return loss + + +def compute_loss( + params: AttributeDict, + model: Union[nn.Module, DDP], + sp: spm.SentencePieceProcessor, + batch: dict, + is_training: bool, +) -> Tuple[Tensor, MetricsTracker]: + """ + Compute RNN-T loss given the model and its inputs. + + Args: + params: + Parameters for training. See :func:`get_params`. + model: + The model for training. It is an instance of Conformer in our case. + batch: + A batch of data. See `lhotse.dataset.K2SpeechRecognitionDataset()` + for the content in it. + is_training: + True for training. False for validation. When it is True, this + function enables autograd during computation; when it is False, it + disables autograd. + """ + device = model.device if isinstance(model, DDP) else next(model.parameters()).device + feature = batch["inputs"].to(device) + feature_lens = batch["input_lens"].to(device) + + # at entry, feature is (N, T, C) + assert feature.ndim == 3 + + # The dataloader returns text as a list of cuts, each of which is a list of channel + # text. We flatten this to a list where all channels are together, i.e., it looks like + # [utt1_ch1, utt2_ch1, ..., uttN_ch1, utt1_ch2, ...., uttN,ch2]. + text = [val for tup in zip(*batch["text"]) for val in tup] + assert len(text) == len(feature) * params.num_channels + + # Convert all channel texts to token IDs and create a ragged tensor. + y = sp.encode(text, out_type=int) + y = k2.RaggedTensor(y).to(device) + + batch_idx_train = params.batch_idx_train + warm_step = params.model_warm_step + + with torch.set_grad_enabled(is_training): + (simple_loss, pruned_loss, ctc_loss, x_masked) = model( + x=feature, + x_lens=feature_lens, + y=y, + prune_range=params.prune_range, + am_scale=params.am_scale, + lm_scale=params.lm_scale, + reduction="none", + subsampling_factor=params.subsampling_factor, + ) + simple_loss_is_finite = torch.isfinite(simple_loss) + pruned_loss_is_finite = torch.isfinite(pruned_loss) + ctc_loss_is_finite = torch.isfinite(ctc_loss) + + # Compute HEAT loss + if is_training and params.heat_loss_scale > 0.0: + heat_loss = compute_heat_loss( + x_masked, batch, num_channels=params.num_channels + ) + else: + heat_loss = torch.tensor(0.0, device=device) + + heat_loss_is_finite = torch.isfinite(heat_loss) + is_finite = ( + simple_loss_is_finite + & pruned_loss_is_finite + & ctc_loss_is_finite + & heat_loss_is_finite + ) + if not torch.all(is_finite): + # logging.info( + # "Not all losses are finite!\n" + # f"simple_losses: {simple_loss}\n" + # f"pruned_losses: {pruned_loss}\n" + # f"ctc_losses: {ctc_loss}\n" + # f"heat_losses: {heat_loss}\n" + # ) + # display_and_save_batch(batch, params=params, sp=sp) + simple_loss = simple_loss[simple_loss_is_finite] + pruned_loss = pruned_loss[pruned_loss_is_finite] + ctc_loss = ctc_loss[ctc_loss_is_finite] + heat_loss = heat_loss[heat_loss_is_finite] + + # If either all simple_loss or pruned_loss is inf or nan, + # we stop the training process by raising an exception + if ( + torch.all(~simple_loss_is_finite) + or torch.all(~pruned_loss_is_finite) + or torch.all(~ctc_loss_is_finite) + or torch.all(~heat_loss_is_finite) + ): + raise ValueError( + "There are too many utterances in this batch " + "leading to inf or nan losses." + ) + + simple_loss_sum = simple_loss.sum() + pruned_loss_sum = pruned_loss.sum() + ctc_loss_sum = ctc_loss.sum() + heat_loss_sum = heat_loss.sum() + + 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_sum + + pruned_loss_scale * pruned_loss_sum + + params.ctc_loss_scale * ctc_loss_sum + + params.heat_loss_scale * heat_loss_sum + ) + + assert loss.requires_grad == is_training + + info = MetricsTracker() + with warnings.catch_warnings(): + warnings.simplefilter("ignore") + # info["frames"] is an approximate number for two reasons: + # (1) The acutal subsampling factor is ((lens - 1) // 2 - 1) // 2 + # (2) If some utterances in the batch lead to inf/nan loss, they + # are filtered out. + info["frames"] = (feature_lens // params.subsampling_factor).sum().item() + + # `utt_duration` and `utt_pad_proportion` would be normalized by `utterances` # noqa + info["utterances"] = feature.size(0) + # averaged input duration in frames over utterances + info["utt_duration"] = feature_lens.sum().item() + # averaged padding proportion over utterances + info["utt_pad_proportion"] = ( + ((feature.size(1) - feature_lens) / feature.size(1)).sum().item() + ) + + # Note: We use reduction=sum while computing the loss. + info["loss"] = loss.detach().cpu().item() + info["simple_loss"] = simple_loss_sum.detach().cpu().item() + info["pruned_loss"] = pruned_loss_sum.detach().cpu().item() + if params.ctc_loss_scale > 0.0: + info["ctc_loss"] = ctc_loss_sum.detach().cpu().item() + if params.heat_loss_scale > 0.0: + info["heat_loss"] = heat_loss_sum.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. + """ + torch.cuda.empty_cache() + model.train() + + tot_loss = MetricsTracker() + + cur_batch_idx = params.get("cur_batch_idx", 0) + + for batch_idx, batch in enumerate(train_dl): + if batch_idx < cur_batch_idx: + continue + cur_batch_idx = batch_idx + + params.batch_idx_train += 1 + batch_size = batch["inputs"].shape[0] + + 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() + set_batch_count(model, params.batch_idx_train) + scheduler.step_batch(params.batch_idx_train) + + scaler.step(optimizer) + scaler.update() + optimizer.zero_grad() + except: # noqa + display_and_save_batch(batch, params=params, 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 < 1.0 or (cur_grad_scale < 8.0 and batch_idx % 400 == 0): + scaler.update(cur_grad_scale * 2.0) + if cur_grad_scale < 0.01: + logging.warning(f"Grad scale is small: {cur_grad_scale}") + if cur_grad_scale < 1.0e-05: + raise RuntimeError( + f"grad_scale is too small, exiting: {cur_grad_scale}" + ) + + if batch_idx % params.log_interval == 0: + cur_lr = scheduler.get_last_lr()[0] + cur_grad_scale = scaler._scale.item() if params.use_fp16 else 1.0 + + logging.info( + f"Epoch {params.cur_epoch}, " + f"batch {batch_idx}, loss[{loss_info}], " + f"tot_loss[{tot_loss}], batch size: {batch_size}, " + f"lr: {cur_lr:.2e}, " + + (f"grad_scale: {scaler._scale.item()}" if params.use_fp16 else "") + ) + + if tb_writer is not None: + tb_writer.add_scalar( + "train/learning_rate", cur_lr, params.batch_idx_train + ) + + loss_info.write_summary( + tb_writer, "train/current_", params.batch_idx_train + ) + tot_loss.write_summary(tb_writer, "train/tot_", params.batch_idx_train) + if params.use_fp16: + tb_writer.add_scalar( + "train/grad_scale", cur_grad_scale, params.batch_idx_train + ) + + if batch_idx % params.valid_interval == 0 and not params.print_diagnostics: + logging.info("Computing validation loss") + valid_info = compute_validation_loss( + params=params, + model=model, + 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_surt_model(params) + + num_param = sum([p.numel() for p in model.parameters()]) + logging.info(f"Number of model parameters: {num_param}") + + assert params.save_every_n >= params.average_period + model_avg: Optional[nn.Module] = None + if rank == 0: + # model_avg is only used with rank 0 + model_avg = copy.deepcopy(model) + + assert params.start_epoch > 0, params.start_epoch + checkpoints = load_checkpoint_if_available( + params=params, model=model, model_avg=model_avg + ) + + model.to(device) + + if checkpoints is None and params.model_init_ckpt is not None: + logging.info( + f"Initializing model with checkpoint from {params.model_init_ckpt}" + ) + init_ckpt = torch.load(params.model_init_ckpt, map_location=device) + model.load_state_dict(init_ckpt["model"], strict=False) + + if world_size > 1: + logging.info("Using DDP") + model = DDP(model, device_ids=[rank], find_unused_parameters=True) + + parameters_names = [] + parameters_names.append( + [name_param_pair[0] for name_param_pair in model.named_parameters()] + ) + optimizer = ScaledAdam( + model.parameters(), + lr=params.base_lr, + clipping_scale=2.0, + parameters_names=parameters_names, + ) + + scheduler = Eden(optimizer, params.lr_batches, params.lr_epochs) + + if checkpoints and "optimizer" in checkpoints: + logging.info("Loading optimizer state dict") + optimizer.load_state_dict(checkpoints["optimizer"]) + + if ( + checkpoints + and "scheduler" in checkpoints + and checkpoints["scheduler"] is not None + ): + logging.info("Loading scheduler state dict") + scheduler.load_state_dict(checkpoints["scheduler"]) + + if params.print_diagnostics: + diagnostic = diagnostics.attach_diagnostics(model) + + ami = AmiAsrDataModule(args) + + train_cuts = ami.train_cuts() + train_cuts = train_cuts.filter(lambda c: 0.5 <= c.duration <= 35.0) + dev_cuts = ami.ami_cuts(split="dev", type="ihm-mix") + dev_cuts = dev_cuts.trim_to_supervision_groups(max_pause=0.0).filter( + lambda c: 0.2 <= c.duration <= 60.0 + ) + + if params.start_batch > 0 and checkpoints and "sampler" in checkpoints: + # We only load the sampler's state dict when it loads a checkpoint + # saved in the middle of an epoch + sampler_state_dict = checkpoints["sampler"] + else: + sampler_state_dict = None + + train_dl = ami.train_dataloaders( + train_cuts, + sampler_state_dict=sampler_state_dict, + ) + valid_dl = ami.valid_dataloaders(dev_cuts) + + 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) + + features = batch["inputs"] + + logging.info(f"features shape: {features.shape}") + + y = [sp.encode(text_ch) for text_ch in batch["text"]] + num_tokens = [sum(len(yi) for yi in y_ch) for y_ch in y] + logging.info(f"num tokens: {num_tokens}") + + +def main(): + parser = get_parser() + AmiAsrDataModule.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) +torch.multiprocessing.set_sharing_strategy("file_system") + +if __name__ == "__main__": + main() diff --git a/egs/ami/SURT/dprnn_zipformer/zipformer.py b/egs/ami/SURT/dprnn_zipformer/zipformer.py new file mode 120000 index 000000000..59b772024 --- /dev/null +++ b/egs/ami/SURT/dprnn_zipformer/zipformer.py @@ -0,0 +1 @@ +../../../libricss/SURT/dprnn_zipformer/zipformer.py \ No newline at end of file diff --git a/egs/ami/SURT/local/add_source_feats.py b/egs/ami/SURT/local/add_source_feats.py new file mode 100755 index 000000000..0917b88a6 --- /dev/null +++ b/egs/ami/SURT/local/add_source_feats.py @@ -0,0 +1,78 @@ +#!/usr/bin/env python3 +# Copyright 2022 Johns Hopkins University (authors: Desh Raj) +# +# See ../../../../LICENSE for clarification regarding multiple authors +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + + +""" +This file adds source features as temporal arrays to the mixture manifests. +It looks for manifests in the directory data/manifests. +""" +import logging +from pathlib import Path + +import numpy as np +from lhotse import CutSet, LilcomChunkyWriter, load_manifest, load_manifest_lazy +from tqdm import tqdm + + +def add_source_feats(): + src_dir = Path("data/manifests") + output_dir = Path("data/fbank") + + logging.info("Reading mixed cuts") + mixed_cuts_clean = load_manifest_lazy(src_dir / "cuts_train_clean.jsonl.gz") + mixed_cuts_reverb = load_manifest_lazy(src_dir / "cuts_train_reverb.jsonl.gz") + + logging.info("Reading source cuts") + source_cuts = load_manifest(src_dir / "ihm_cuts_train_trimmed.jsonl.gz") + + logging.info("Adding source features to the mixed cuts") + pbar = tqdm(total=len(mixed_cuts_clean), desc="Adding source features") + with CutSet.open_writer( + src_dir / "cuts_train_clean_sources.jsonl.gz" + ) as cut_writer_clean, CutSet.open_writer( + src_dir / "cuts_train_reverb_sources.jsonl.gz" + ) as cut_writer_reverb, LilcomChunkyWriter( + output_dir / "feats_train_clean_sources" + ) as source_feat_writer: + for cut_clean, cut_reverb in zip(mixed_cuts_clean, mixed_cuts_reverb): + assert cut_reverb.id == cut_clean.id + "_rvb" + source_feats = [] + source_feat_offsets = [] + cur_offset = 0 + for sup in sorted( + cut_clean.supervisions, key=lambda s: (s.start, s.speaker) + ): + source_cut = source_cuts[sup.id] + source_feats.append(source_cut.load_features()) + source_feat_offsets.append(cur_offset) + cur_offset += source_cut.num_frames + cut_clean.source_feats = source_feat_writer.store_array( + cut_clean.id, np.concatenate(source_feats, axis=0) + ) + cut_clean.source_feat_offsets = source_feat_offsets + cut_writer_clean.write(cut_clean) + # Also write the reverb cut + cut_reverb.source_feats = cut_clean.source_feats + cut_reverb.source_feat_offsets = cut_clean.source_feat_offsets + cut_writer_reverb.write(cut_reverb) + pbar.update(1) + + +if __name__ == "__main__": + formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s" + logging.basicConfig(format=formatter, level=logging.INFO) + add_source_feats() diff --git a/egs/ami/SURT/local/compute_fbank_aimix.py b/egs/ami/SURT/local/compute_fbank_aimix.py new file mode 100755 index 000000000..91b3a060b --- /dev/null +++ b/egs/ami/SURT/local/compute_fbank_aimix.py @@ -0,0 +1,185 @@ +#!/usr/bin/env python3 +# Copyright 2022 Johns Hopkins University (authors: Desh Raj) +# +# See ../../../../LICENSE for clarification regarding multiple authors +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + + +""" +This file computes fbank features of the synthetically mixed AMI and ICSI +train set. +It looks for manifests in the directory data/manifests. + +The generated fbank features are saved in data/fbank. +""" +import logging +import random +import warnings +from pathlib import Path + +import torch +import torch.multiprocessing +import torchaudio +from lhotse import ( + AudioSource, + LilcomChunkyWriter, + Recording, + load_manifest, + load_manifest_lazy, +) +from lhotse.audio import set_ffmpeg_torchaudio_info_enabled +from lhotse.cut import MixedCut, MixTrack, MultiCut +from lhotse.features.kaldifeat import ( + KaldifeatFbank, + KaldifeatFbankConfig, + KaldifeatFrameOptions, + KaldifeatMelOptions, +) +from lhotse.utils import fix_random_seed, uuid4 +from tqdm import tqdm + +# Torch's multithreaded behavior needs to be disabled or +# it wastes a lot of CPU and slow things down. +# Do this outside of main() in case it needs to take effect +# even when we are not invoking the main (e.g. when spawning subprocesses). +torch.set_num_threads(1) +torch.set_num_interop_threads(1) +torch.multiprocessing.set_sharing_strategy("file_system") +torchaudio.set_audio_backend("soundfile") +set_ffmpeg_torchaudio_info_enabled(False) + + +def compute_fbank_aimix(): + src_dir = Path("data/manifests") + output_dir = Path("data/fbank") + + sampling_rate = 16000 + num_mel_bins = 80 + + extractor = KaldifeatFbank( + KaldifeatFbankConfig( + frame_opts=KaldifeatFrameOptions(sampling_rate=sampling_rate), + mel_opts=KaldifeatMelOptions(num_bins=num_mel_bins), + device="cuda", + ) + ) + + logging.info("Reading manifests") + train_cuts = load_manifest_lazy(src_dir / "ai-mix_cuts_clean_full.jsonl.gz") + + # only uses RIRs and noises from REVERB challenge + real_rirs = load_manifest(src_dir / "real-rir_recordings_all.jsonl.gz").filter( + lambda r: "RVB2014" in r.id + ) + noises = load_manifest(src_dir / "iso-noise_recordings_all.jsonl.gz").filter( + lambda r: "RVB2014" in r.id + ) + + # Apply perturbation to the training cuts + logging.info("Applying perturbation to the training cuts") + train_cuts_rvb = train_cuts.map( + lambda c: augment( + c, perturb_snr=True, rirs=real_rirs, noises=noises, perturb_loudness=True + ) + ) + + logging.info("Extracting fbank features for training cuts") + _ = train_cuts.compute_and_store_features_batch( + extractor=extractor, + storage_path=output_dir / "ai-mix_feats_clean", + manifest_path=src_dir / "cuts_train_clean.jsonl.gz", + batch_duration=5000, + num_workers=4, + storage_type=LilcomChunkyWriter, + overwrite=True, + ) + + with warnings.catch_warnings(): + warnings.simplefilter("ignore") + _ = train_cuts_rvb.compute_and_store_features_batch( + extractor=extractor, + storage_path=output_dir / "ai-mix_feats_reverb", + manifest_path=src_dir / "cuts_train_reverb.jsonl.gz", + batch_duration=5000, + num_workers=4, + storage_type=LilcomChunkyWriter, + overwrite=True, + ) + + +def augment(cut, perturb_snr=False, rirs=None, noises=None, perturb_loudness=False): + """ + Given a mixed cut, this function optionally applies the following augmentations: + - Perturbing the SNRs of the tracks (in range [-5, 5] dB) + - Reverberation using a randomly selected RIR + - Adding noise + - Perturbing the loudness (in range [-20, -25] dB) + """ + out_cut = cut.drop_features() + + # Perturb the SNRs (optional) + if perturb_snr: + snrs = [random.uniform(-5, 5) for _ in range(len(cut.tracks))] + for i, (track, snr) in enumerate(zip(out_cut.tracks, snrs)): + if i == 0: + # Skip the first track since it is the reference + continue + track.snr = snr + + # Reverberate the cut (optional) + if rirs is not None: + # Select an RIR at random + rir = random.choice(rirs) + # Select a channel at random + rir_channel = random.choice(list(range(rir.num_channels))) + # Reverberate the cut + out_cut = out_cut.reverb_rir(rir_recording=rir, rir_channels=[rir_channel]) + + # Add noise (optional) + if noises is not None: + # Select a noise recording at random + noise = random.choice(noises).to_cut() + if isinstance(noise, MultiCut): + noise = noise.to_mono()[0] + # Select an SNR at random + snr = random.uniform(10, 30) + # Repeat the noise to match the duration of the cut + noise = repeat_cut(noise, out_cut.duration) + out_cut = MixedCut( + id=out_cut.id, + tracks=[ + MixTrack(cut=out_cut, type="MixedCut"), + MixTrack(cut=noise, type="DataCut", snr=snr), + ], + ) + + # Perturb the loudness (optional) + if perturb_loudness: + target_loudness = random.uniform(-20, -25) + out_cut = out_cut.normalize_loudness(target_loudness, mix_first=True) + return out_cut + + +def repeat_cut(cut, duration): + while cut.duration < duration: + cut = cut.mix(cut, offset_other_by=cut.duration) + return cut.truncate(duration=duration) + + +if __name__ == "__main__": + formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s" + logging.basicConfig(format=formatter, level=logging.INFO) + + fix_random_seed(42) + compute_fbank_aimix() diff --git a/egs/ami/SURT/local/compute_fbank_ami.py b/egs/ami/SURT/local/compute_fbank_ami.py new file mode 100755 index 000000000..351b41765 --- /dev/null +++ b/egs/ami/SURT/local/compute_fbank_ami.py @@ -0,0 +1,94 @@ +#!/usr/bin/env python3 +# Copyright 2022 Johns Hopkins University (authors: Desh Raj) +# +# See ../../../../LICENSE for clarification regarding multiple authors +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + + +""" +This file computes fbank features of the AMI dataset. +We compute features for full recordings (i.e., without trimming to supervisions). +This way we can create arbitrary segmentations later. + +The generated fbank features are saved in data/fbank. +""" +import logging +import math +from pathlib import Path + +import torch +import torch.multiprocessing +from lhotse import CutSet, LilcomChunkyWriter +from lhotse.features.kaldifeat import ( + KaldifeatFbank, + KaldifeatFbankConfig, + KaldifeatFrameOptions, + KaldifeatMelOptions, +) +from lhotse.recipes.utils import read_manifests_if_cached + +# Torch's multithreaded behavior needs to be disabled or +# it wastes a lot of CPU and slow things down. +# Do this outside of main() in case it needs to take effect +# even when we are not invoking the main (e.g. when spawning subprocesses). +torch.set_num_threads(1) +torch.set_num_interop_threads(1) +torch.multiprocessing.set_sharing_strategy("file_system") + + +def compute_fbank_ami(): + src_dir = Path("data/manifests") + output_dir = Path("data/fbank") + + sampling_rate = 16000 + num_mel_bins = 80 + + extractor = KaldifeatFbank( + KaldifeatFbankConfig( + frame_opts=KaldifeatFrameOptions(sampling_rate=sampling_rate), + mel_opts=KaldifeatMelOptions(num_bins=num_mel_bins), + device="cuda", + ) + ) + + logging.info("Reading manifests") + manifests = {} + for part in ["ihm-mix", "sdm", "mdm8-bf"]: + manifests[part] = read_manifests_if_cached( + dataset_parts=["train", "dev", "test"], + output_dir=src_dir, + prefix=f"ami-{part}", + suffix="jsonl.gz", + ) + + for part in ["ihm-mix", "sdm", "mdm8-bf"]: + for split in ["train", "dev", "test"]: + logging.info(f"Processing {part} {split}") + cuts = CutSet.from_manifests( + **manifests[part][split] + ).compute_and_store_features_batch( + extractor=extractor, + storage_path=output_dir / f"ami-{part}_{split}_feats", + manifest_path=src_dir / f"cuts_ami-{part}_{split}.jsonl.gz", + batch_duration=5000, + num_workers=4, + storage_type=LilcomChunkyWriter, + ) + + +if __name__ == "__main__": + formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s" + logging.basicConfig(format=formatter, level=logging.INFO) + + compute_fbank_ami() diff --git a/egs/ami/SURT/local/compute_fbank_icsi.py b/egs/ami/SURT/local/compute_fbank_icsi.py new file mode 100755 index 000000000..4e2ff3f3b --- /dev/null +++ b/egs/ami/SURT/local/compute_fbank_icsi.py @@ -0,0 +1,95 @@ +#!/usr/bin/env python3 +# Copyright 2022 Johns Hopkins University (authors: Desh Raj) +# +# See ../../../../LICENSE for clarification regarding multiple authors +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + + +""" +This file computes fbank features of the ICSI dataset. +We compute features for full recordings (i.e., without trimming to supervisions). +This way we can create arbitrary segmentations later. + +The generated fbank features are saved in data/fbank. +""" +import logging +import math +from pathlib import Path + +import torch +import torch.multiprocessing +from lhotse import CutSet, LilcomChunkyWriter +from lhotse.features.kaldifeat import ( + KaldifeatFbank, + KaldifeatFbankConfig, + KaldifeatFrameOptions, + KaldifeatMelOptions, +) +from lhotse.recipes.utils import read_manifests_if_cached + +# Torch's multithreaded behavior needs to be disabled or +# it wastes a lot of CPU and slow things down. +# Do this outside of main() in case it needs to take effect +# even when we are not invoking the main (e.g. when spawning subprocesses). +torch.set_num_threads(1) +torch.set_num_interop_threads(1) +torch.multiprocessing.set_sharing_strategy("file_system") + + +def compute_fbank_icsi(): + src_dir = Path("data/manifests") + output_dir = Path("data/fbank") + + sampling_rate = 16000 + num_mel_bins = 80 + + extractor = KaldifeatFbank( + KaldifeatFbankConfig( + frame_opts=KaldifeatFrameOptions(sampling_rate=sampling_rate), + mel_opts=KaldifeatMelOptions(num_bins=num_mel_bins), + device="cuda", + ) + ) + + logging.info("Reading manifests") + manifests = {} + for part in ["ihm-mix", "sdm"]: + manifests[part] = read_manifests_if_cached( + dataset_parts=["train"], + output_dir=src_dir, + prefix=f"icsi-{part}", + suffix="jsonl.gz", + ) + + for part in ["ihm-mix", "sdm"]: + for split in ["train"]: + logging.info(f"Processing {part} {split}") + cuts = CutSet.from_manifests( + **manifests[part][split] + ).compute_and_store_features_batch( + extractor=extractor, + storage_path=output_dir / f"icsi-{part}_{split}_feats", + manifest_path=src_dir / f"cuts_icsi-{part}_{split}.jsonl.gz", + batch_duration=5000, + num_workers=4, + storage_type=LilcomChunkyWriter, + overwrite=True, + ) + + +if __name__ == "__main__": + formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s" + logging.basicConfig(format=formatter, level=logging.INFO) + + compute_fbank_icsi() diff --git a/egs/ami/SURT/local/compute_fbank_ihm.py b/egs/ami/SURT/local/compute_fbank_ihm.py new file mode 100755 index 000000000..56f54aa21 --- /dev/null +++ b/egs/ami/SURT/local/compute_fbank_ihm.py @@ -0,0 +1,101 @@ +#!/usr/bin/env python3 +# Copyright 2022 Johns Hopkins University (authors: Desh Raj) +# +# See ../../../../LICENSE for clarification regarding multiple authors +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + + +""" +This file computes fbank features of the trimmed sub-segments which will be +used for simulating the training mixtures. + +The generated fbank features are saved in data/fbank. +""" +import logging +import math +from pathlib import Path + +import torch +import torch.multiprocessing +import torchaudio +from lhotse import CutSet, LilcomChunkyWriter, load_manifest +from lhotse.audio import set_ffmpeg_torchaudio_info_enabled +from lhotse.features.kaldifeat import ( + KaldifeatFbank, + KaldifeatFbankConfig, + KaldifeatFrameOptions, + KaldifeatMelOptions, +) +from lhotse.recipes.utils import read_manifests_if_cached +from tqdm import tqdm + +# Torch's multithreaded behavior needs to be disabled or +# it wastes a lot of CPU and slow things down. +# Do this outside of main() in case it needs to take effect +# even when we are not invoking the main (e.g. when spawning subprocesses). +torch.set_num_threads(1) +torch.set_num_interop_threads(1) +torch.multiprocessing.set_sharing_strategy("file_system") +torchaudio.set_audio_backend("soundfile") +set_ffmpeg_torchaudio_info_enabled(False) + + +def compute_fbank_ihm(): + src_dir = Path("data/manifests") + output_dir = Path("data/fbank") + + sampling_rate = 16000 + num_mel_bins = 80 + + extractor = KaldifeatFbank( + KaldifeatFbankConfig( + frame_opts=KaldifeatFrameOptions(sampling_rate=sampling_rate), + mel_opts=KaldifeatMelOptions(num_bins=num_mel_bins), + device="cuda", + ) + ) + + logging.info("Reading manifests") + manifests = {} + for data in ["ami", "icsi"]: + manifests[data] = read_manifests_if_cached( + dataset_parts=["train"], + output_dir=src_dir, + types=["recordings", "supervisions"], + prefix=f"{data}-ihm", + suffix="jsonl.gz", + ) + + logging.info("Computing features") + for data in ["ami", "icsi"]: + cs = CutSet.from_manifests(**manifests[data]["train"]) + cs = cs.trim_to_supervisions(keep_overlapping=False) + cs = cs.normalize_loudness(target=-23.0, affix_id=False) + cs = cs + cs.perturb_speed(0.9) + cs.perturb_speed(1.1) + _ = cs.compute_and_store_features_batch( + extractor=extractor, + storage_path=output_dir / f"{data}-ihm_train_feats", + manifest_path=src_dir / f"{data}-ihm_cuts_train.jsonl.gz", + batch_duration=5000, + num_workers=4, + storage_type=LilcomChunkyWriter, + overwrite=True, + ) + + +if __name__ == "__main__": + formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s" + logging.basicConfig(format=formatter, level=logging.INFO) + + compute_fbank_ihm() diff --git a/egs/ami/SURT/local/prepare_ami_train_cuts.py b/egs/ami/SURT/local/prepare_ami_train_cuts.py new file mode 100755 index 000000000..72fced70d --- /dev/null +++ b/egs/ami/SURT/local/prepare_ami_train_cuts.py @@ -0,0 +1,146 @@ +#!/usr/bin/env python3 +# Copyright 2022 Johns Hopkins University (authors: Desh Raj) +# +# See ../../../../LICENSE for clarification regarding multiple authors +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + + +""" +This file creates AMI train segments. +""" +import logging +import math +from pathlib import Path + +import torch +import torch.multiprocessing +from lhotse import LilcomChunkyWriter, load_manifest_lazy +from lhotse.cut import Cut, CutSet +from lhotse.utils import EPSILON, add_durations +from tqdm import tqdm + + +def cut_into_windows(cuts: CutSet, duration: float): + """ + This function takes a CutSet and cuts each cut into windows of roughly + `duration` seconds. By roughly, we mean that we try to adjust for the last supervision + that exceeds the duration, or is shorter than the duration. + """ + res = [] + with tqdm() as pbar: + for cut in cuts: + pbar.update(1) + sups = cut.index_supervisions()[cut.id] + sr = cut.sampling_rate + start = 0.0 + end = duration + num_tries = 0 + while start < cut.duration and num_tries < 2: + # Find the supervision that are cut by the window endpoint + hitlist = [iv for iv in sups.at(end) if iv.begin < end] + # If there are no supervisions, we are done + if not hitlist: + res.append( + cut.truncate( + offset=start, + duration=add_durations(end, -start, sampling_rate=sr), + keep_excessive_supervisions=False, + ) + ) + # Update the start and end for the next window + start = end + end = add_durations(end, duration, sampling_rate=sr) + else: + # find ratio of durations cut by the window endpoint + ratios = [ + add_durations(end, -iv.end, sampling_rate=sr) / iv.length() + for iv in hitlist + ] + # we retain the supervisions that have >50% of their duration + # in the window, and discard the others + retained = [] + discarded = [] + for iv, ratio in zip(hitlist, ratios): + if ratio > 0.5: + retained.append(iv) + else: + discarded.append(iv) + cur_end = max(iv.end for iv in retained) if retained else end + res.append( + cut.truncate( + offset=start, + duration=add_durations(cur_end, -start, sampling_rate=sr), + keep_excessive_supervisions=False, + ) + ) + # For the next window, we start at the earliest discarded supervision + next_start = min(iv.begin for iv in discarded) if discarded else end + next_end = add_durations(next_start, duration, sampling_rate=sr) + # It may happen that next_start is the same as start, in which case + # we will advance the window anyway + if next_start == start: + logging.warning( + f"Next start is the same as start: {next_start} == {start} for cut {cut.id}" + ) + start = end + EPSILON + end = add_durations(start, duration, sampling_rate=sr) + num_tries += 1 + else: + start = next_start + end = next_end + return CutSet.from_cuts(res) + + +def prepare_train_cuts(): + src_dir = Path("data/manifests") + + logging.info("Loading the manifests") + train_cuts_ihm = load_manifest_lazy( + src_dir / "cuts_ami-ihm-mix_train.jsonl.gz" + ).map(lambda c: c.with_id(f"{c.id}_ihm-mix")) + train_cuts_sdm = load_manifest_lazy(src_dir / "cuts_ami-sdm_train.jsonl.gz").map( + lambda c: c.with_id(f"{c.id}_sdm") + ) + train_cuts_mdm = load_manifest_lazy( + src_dir / "cuts_ami-mdm8-bf_train.jsonl.gz" + ).map(lambda c: c.with_id(f"{c.id}_mdm8-bf")) + + # Combine all cuts into one CutSet + train_cuts = train_cuts_ihm + train_cuts_sdm + train_cuts_mdm + + train_cuts_1 = train_cuts.trim_to_supervision_groups(max_pause=0.5) + train_cuts_2 = train_cuts.trim_to_supervision_groups(max_pause=0.0) + + # Combine the two segmentations + train_all = train_cuts_1 + train_cuts_2 + + # At this point, some of the cuts may be very long. We will cut them into windows of + # roughly 30 seconds. + logging.info("Cutting the segments into windows of 30 seconds") + train_all_30 = cut_into_windows(train_all, duration=30.0) + logging.info(f"Number of cuts after cutting into windows: {len(train_all_30)}") + + # Show statistics + train_all.describe(full=True) + + # Save the cuts + logging.info("Saving the cuts") + train_all.to_file(src_dir / "cuts_train_ami.jsonl.gz") + + +if __name__ == "__main__": + formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s" + logging.basicConfig(format=formatter, level=logging.INFO) + + prepare_train_cuts() diff --git a/egs/ami/SURT/local/prepare_icsi_train_cuts.py b/egs/ami/SURT/local/prepare_icsi_train_cuts.py new file mode 100755 index 000000000..818e26bfb --- /dev/null +++ b/egs/ami/SURT/local/prepare_icsi_train_cuts.py @@ -0,0 +1,67 @@ +#!/usr/bin/env python3 +# Copyright 2022 Johns Hopkins University (authors: Desh Raj) +# +# See ../../../../LICENSE for clarification regarding multiple authors +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + + +""" +This file creates ICSI train segments. +""" +import logging +from pathlib import Path + +from lhotse import load_manifest_lazy +from prepare_ami_train_cuts import cut_into_windows + + +def prepare_train_cuts(): + src_dir = Path("data/manifests") + + logging.info("Loading the manifests") + train_cuts_ihm = load_manifest_lazy( + src_dir / "cuts_icsi-ihm-mix_train.jsonl.gz" + ).map(lambda c: c.with_id(f"{c.id}_ihm-mix")) + train_cuts_sdm = load_manifest_lazy(src_dir / "cuts_icsi-sdm_train.jsonl.gz").map( + lambda c: c.with_id(f"{c.id}_sdm") + ) + + # Combine all cuts into one CutSet + train_cuts = train_cuts_ihm + train_cuts_sdm + + train_cuts_1 = train_cuts.trim_to_supervision_groups(max_pause=0.5) + train_cuts_2 = train_cuts.trim_to_supervision_groups(max_pause=0.0) + + # Combine the two segmentations + train_all = train_cuts_1 + train_cuts_2 + + # At this point, some of the cuts may be very long. We will cut them into windows of + # roughly 30 seconds. + logging.info("Cutting the segments into windows of 30 seconds") + train_all_30 = cut_into_windows(train_all, duration=30.0) + logging.info(f"Number of cuts after cutting into windows: {len(train_all_30)}") + + # Show statistics + train_all.describe(full=True) + + # Save the cuts + logging.info("Saving the cuts") + train_all.to_file(src_dir / "cuts_train_icsi.jsonl.gz") + + +if __name__ == "__main__": + formatter = "%(asctime)s %(levelname)s [%(filename)s:%(lineno)d] %(message)s" + logging.basicConfig(format=formatter, level=logging.INFO) + + prepare_train_cuts() diff --git a/egs/ami/SURT/local/prepare_lang_bpe.py b/egs/ami/SURT/local/prepare_lang_bpe.py new file mode 120000 index 000000000..36b40e7fc --- /dev/null +++ b/egs/ami/SURT/local/prepare_lang_bpe.py @@ -0,0 +1 @@ +../../../librispeech/ASR/local/prepare_lang_bpe.py \ No newline at end of file diff --git a/egs/ami/SURT/local/train_bpe_model.py b/egs/ami/SURT/local/train_bpe_model.py new file mode 120000 index 000000000..6fad36421 --- /dev/null +++ b/egs/ami/SURT/local/train_bpe_model.py @@ -0,0 +1 @@ +../../../librispeech/ASR/local/train_bpe_model.py \ No newline at end of file diff --git a/egs/ami/SURT/prepare.sh b/egs/ami/SURT/prepare.sh new file mode 100755 index 000000000..ea4e5baf2 --- /dev/null +++ b/egs/ami/SURT/prepare.sh @@ -0,0 +1,195 @@ +#!/usr/bin/env bash + +set -eou pipefail + +stage=-1 +stop_stage=100 + +# We assume dl_dir (download dir) contains the following +# directories and files. If not, they will be downloaded +# by this script automatically. +# +# - $dl_dir/ami +# You can find audio and transcripts for AMI in this path. +# +# - $dl_dir/icsi +# You can find audio and transcripts for ICSI in this path. +# +# - $dl_dir/rirs_noises +# This directory contains the RIRS_NOISES corpus downloaded from https://openslr.org/28/. +# +dl_dir=$PWD/download + +. shared/parse_options.sh || exit 1 + +# All files generated by this script are saved in "data". +# You can safely remove "data" and rerun this script to regenerate it. +mkdir -p data +vocab_size=500 + +log() { + # This function is from espnet + local fname=${BASH_SOURCE[1]##*/} + echo -e "$(date '+%Y-%m-%d %H:%M:%S') (${fname}:${BASH_LINENO[0]}:${FUNCNAME[1]}) $*" +} + +log "dl_dir: $dl_dir" + +if [ $stage -le 0 ] && [ $stop_stage -ge 0 ]; then + log "Stage 0: Download data" + + # If you have pre-downloaded it to /path/to/amicorpus, + # you can create a symlink + # + # ln -sfv /path/to/amicorpus $dl_dir/amicorpus + # + if [ ! -d $dl_dir/amicorpus ]; then + for mic in ihm ihm-mix sdm mdm8-bf; do + lhotse download ami --mic $mic $dl_dir/amicorpus + done + fi + + # If you have pre-downloaded it to /path/to/icsi, + # you can create a symlink + # + # ln -sfv /path/to/icsi $dl_dir/icsi + # + if [ ! -d $dl_dir/icsi ]; then + lhotse download icsi $dl_dir/icsi + fi + + # If you have pre-downloaded it to /path/to/rirs_noises, + # you can create a symlink + # + # ln -sfv /path/to/rirs_noises $dl_dir/ + # + if [ ! -d $dl_dir/rirs_noises ]; then + lhotse download rirs_noises $dl_dir + fi +fi + +if [ $stage -le 1 ] && [ $stop_stage -ge 1 ]; then + log "Stage 1: Prepare AMI manifests" + # We assume that you have downloaded the AMI corpus + # to $dl_dir/amicorpus. We perform text normalization for the transcripts. + mkdir -p data/manifests + for mic in ihm ihm-mix sdm mdm8-bf; do + log "Preparing AMI manifest for $mic" + lhotse prepare ami --mic $mic --max-words-per-segment 30 --merge-consecutive $dl_dir/amicorpus data/manifests/ + done +fi + +if [ $stage -le 2 ] && [ $stop_stage -ge 2 ]; then + log "Stage 2: Prepare ICSI manifests" + # We assume that you have downloaded the ICSI corpus + # to $dl_dir/icsi. We perform text normalization for the transcripts. + mkdir -p data/manifests + log "Preparing ICSI manifest" + for mic in ihm ihm-mix sdm; do + lhotse prepare icsi --mic $mic $dl_dir/icsi data/manifests/ + done +fi + +if [ $stage -le 3 ] && [ $stop_stage -ge 3 ]; then + log "Stage 3: Prepare RIRs" + # We assume that you have downloaded the RIRS_NOISES corpus + # to $dl_dir/rirs_noises + lhotse prepare rir-noise -p real_rir -p iso_noise $dl_dir/rirs_noises data/manifests +fi + +if [ $stage -le 4 ] && [ $stop_stage -ge 4 ]; then + log "Stage 3: Extract features for AMI and ICSI recordings" + python local/compute_fbank_ami.py + python local/compute_fbank_icsi.py +fi + +if [ $stage -le 5 ] && [ $stop_stage -ge 5 ]; then + log "Stage 5: Create sources for simulating mixtures" + # In the following script, we speed-perturb the IHM recordings and extract features. + python local/compute_fbank_ihm.py + lhotse combine data/manifests/ami-ihm_cuts_train.jsonl.gz \ + data/manifests/icsi-ihm_cuts_train.jsonl.gz - |\ + lhotse cut trim-to-alignments --type word --max-pause 0.5 - - |\ + lhotse filter 'duration<=12.0' - - |\ + shuf | gzip -c > data/manifests/ihm_cuts_train_trimmed.jsonl.gz +fi + +if [ $stage -le 6 ] && [ $stop_stage -ge 6 ]; then + log "Stage 6: Create training mixtures" + lhotse workflows simulate-meetings \ + --method conversational \ + --same-spk-pause 0.5 \ + --diff-spk-pause 0.5 \ + --diff-spk-overlap 1.0 \ + --prob-diff-spk-overlap 0.8 \ + --num-meetings 200000 \ + --num-speakers-per-meeting 2,3 \ + --max-duration-per-speaker 15.0 \ + --max-utterances-per-speaker 3 \ + --seed 1234 \ + --num-jobs 2 \ + data/manifests/ihm_cuts_train_trimmed.jsonl.gz \ + data/manifests/ai-mix_cuts_clean.jsonl.gz + + python local/compute_fbank_aimix.py + + # Add source features to the manifest (will be used for masking loss) + # This may take ~2 hours. + python local/add_source_feats.py + + # Combine clean and reverb + cat <(gunzip -c data/manifests/cuts_train_clean_sources.jsonl.gz) \ + <(gunzip -c data/manifests/cuts_train_reverb_sources.jsonl.gz) |\ + shuf | gzip -c > data/manifests/cuts_train_comb_sources.jsonl.gz +fi + +if [ $stage -le 7 ] && [ $stop_stage -ge 7 ]; then + log "Stage 7: Create training mixtures from real sessions" + python local/prepare_ami_train_cuts.py + python local/prepare_icsi_train_cuts.py + + # Combine AMI and ICSI + cat <(gunzip -c data/manifests/cuts_train_ami.jsonl.gz) \ + <(gunzip -c data/manifests/cuts_train_icsi.jsonl.gz) |\ + shuf | gzip -c > data/manifests/cuts_train_ami_icsi.jsonl.gz +fi + +if [ $stage -le 8 ] && [ $stop_stage -ge 8 ]; then + log "Stage 8: Dump transcripts for BPE model training (using AMI and ICSI)." + mkdir -p data/lm + cat <(gunzip -c data/manifests/ami-sdm_supervisions_train.jsonl.gz | jq '.text' | sed 's:"::g') \ + <(gunzip -c data/manifests/icsi-sdm_supervisions_train.jsonl.gz | jq '.text' | sed 's:"::g') \ + > data/lm/transcript_words.txt +fi + +if [ $stage -le 9 ] && [ $stop_stage -ge 9 ]; then + log "Stage 9: Prepare BPE based lang (combining AMI and ICSI)" + + lang_dir=data/lang_bpe_${vocab_size} + mkdir -p $lang_dir + + # Add special words to words.txt + echo " 0" > $lang_dir/words.txt + echo "!SIL 1" >> $lang_dir/words.txt + echo " 2" >> $lang_dir/words.txt + + # Add regular words to words.txt + cat data/lm/transcript_words.txt | grep -o -E '\w+' | sort -u | awk '{print $0,NR+2}' >> $lang_dir/words.txt + + # Add remaining special word symbols expected by LM scripts. + num_words=$(cat $lang_dir/words.txt | wc -l) + echo " ${num_words}" >> $lang_dir/words.txt + num_words=$(cat $lang_dir/words.txt | wc -l) + echo " ${num_words}" >> $lang_dir/words.txt + num_words=$(cat $lang_dir/words.txt | wc -l) + echo "#0 ${num_words}" >> $lang_dir/words.txt + + ./local/train_bpe_model.py \ + --lang-dir $lang_dir \ + --vocab-size $vocab_size \ + --transcript data/lm/transcript_words.txt + + if [ ! -f $lang_dir/L_disambig.pt ]; then + ./local/prepare_lang_bpe.py --lang-dir $lang_dir + fi +fi diff --git a/egs/ami/SURT/shared b/egs/ami/SURT/shared new file mode 120000 index 000000000..4cbd91a7e --- /dev/null +++ b/egs/ami/SURT/shared @@ -0,0 +1 @@ +../../../icefall/shared \ No newline at end of file