mirror of
https://github.com/k2-fsa/icefall.git
synced 2025-08-08 09:32:20 +00:00
5fe58de43c
* initial commit * support download, data prep, and fbank * on-the-fly feature extraction by default * support BPE based lang * support HLG for BPE * small fix * small fix * chunked feature extraction by default * Compute features for GigaSpeech by splitting the manifest. * Fixes after review. * Split manifests into 2000 pieces. * set audio duration mismatch tolerance to 0.01 * small fix * add conformer training recipe * Add conformer.py without pre-commit checking * lazy loading and use SingleCutSampler * DynamicBucketingSampler * use KaldifeatFbank to compute fbank for musan * use pretrained language model and lexicon * use 3gram to decode, 4gram to rescore * Add decode.py * Update .flake8 * Delete compute_fbank_gigaspeech.py * Use BucketingSampler for valid and test dataloader * Update params in train.py * Use bpe_500 * update params in decode.py * Decrease num_paths while CUDA OOM * Added README * Update RESULTS * black * Decrease num_paths while CUDA OOM * Decode with post-processing * Update results * Remove lazy_load option * Use default `storage_type` * Keep the original tolerance * Use split-lazy * black * Update pretrained model Co-authored-by: Fangjun Kuang <csukuangfj@gmail.com>
374 lines
13 KiB
Python
374 lines
13 KiB
Python
# Copyright 2021 Piotr Żelasko
|
|
#
|
|
# See ../../../../LICENSE for clarification regarding multiple authors
|
|
#
|
|
# Licensed under the Apache License, Version 2.0 (the "License");
|
|
# you may not use this file except in compliance with the License.
|
|
# You may obtain a copy of the License at
|
|
#
|
|
# http://www.apache.org/licenses/LICENSE-2.0
|
|
#
|
|
# Unless required by applicable law or agreed to in writing, software
|
|
# distributed under the License is distributed on an "AS IS" BASIS,
|
|
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
|
# See the License for the specific language governing permissions and
|
|
# limitations under the License.
|
|
|
|
|
|
import argparse
|
|
import logging
|
|
from functools import lru_cache
|
|
from pathlib import Path
|
|
|
|
from lhotse import CutSet, Fbank, FbankConfig, load_manifest
|
|
from lhotse.dataset import (
|
|
BucketingSampler,
|
|
CutConcatenate,
|
|
CutMix,
|
|
DynamicBucketingSampler,
|
|
K2SpeechRecognitionDataset,
|
|
PrecomputedFeatures,
|
|
SingleCutSampler,
|
|
SpecAugment,
|
|
)
|
|
from lhotse.dataset.input_strategies import OnTheFlyFeatures
|
|
from torch.utils.data import DataLoader
|
|
|
|
from icefall.utils import str2bool
|
|
|
|
|
|
class GigaSpeechAsrDataModule:
|
|
"""
|
|
DataModule for k2 ASR experiments.
|
|
It assumes there is always one train and valid dataloader,
|
|
but there can be multiple test dataloaders (e.g. LibriSpeech test-clean
|
|
and test-other).
|
|
|
|
It contains all the common data pipeline modules used in ASR
|
|
experiments, e.g.:
|
|
- dynamic batch size,
|
|
- bucketing samplers,
|
|
- cut concatenation,
|
|
- augmentation,
|
|
- on-the-fly feature extraction
|
|
|
|
This class should be derived for specific corpora used in ASR tasks.
|
|
"""
|
|
|
|
def __init__(self, args: argparse.Namespace):
|
|
self.args = args
|
|
|
|
@classmethod
|
|
def add_arguments(cls, parser: argparse.ArgumentParser):
|
|
group = parser.add_argument_group(
|
|
title="ASR data related options",
|
|
description="These options are used for the preparation of "
|
|
"PyTorch DataLoaders from Lhotse CutSet's -- they control the "
|
|
"effective batch sizes, sampling strategies, applied data "
|
|
"augmentations, etc.",
|
|
)
|
|
group.add_argument(
|
|
"--manifest-dir",
|
|
type=Path,
|
|
default=Path("data/fbank"),
|
|
help="Path to directory with train/valid/test cuts.",
|
|
)
|
|
group.add_argument(
|
|
"--max-duration",
|
|
type=int,
|
|
default=200.0,
|
|
help="Maximum pooled recordings duration (seconds) in a "
|
|
"single batch. You can reduce it if it causes CUDA OOM.",
|
|
)
|
|
group.add_argument(
|
|
"--bucketing-sampler",
|
|
type=str2bool,
|
|
default=True,
|
|
help="When enabled, the batches will come from buckets of "
|
|
"similar duration (saves padding frames).",
|
|
)
|
|
group.add_argument(
|
|
"--num-buckets",
|
|
type=int,
|
|
default=30,
|
|
help="The number of buckets for the DynamicBucketingSampler"
|
|
"(you might want to increase it for larger datasets).",
|
|
)
|
|
group.add_argument(
|
|
"--concatenate-cuts",
|
|
type=str2bool,
|
|
default=False,
|
|
help="When enabled, utterances (cuts) will be concatenated "
|
|
"to minimize the amount of padding.",
|
|
)
|
|
group.add_argument(
|
|
"--duration-factor",
|
|
type=float,
|
|
default=1.0,
|
|
help="Determines the maximum duration of a concatenated cut "
|
|
"relative to the duration of the longest cut in a batch.",
|
|
)
|
|
group.add_argument(
|
|
"--gap",
|
|
type=float,
|
|
default=1.0,
|
|
help="The amount of padding (in seconds) inserted between "
|
|
"concatenated cuts. This padding is filled with noise when "
|
|
"noise augmentation is used.",
|
|
)
|
|
group.add_argument(
|
|
"--on-the-fly-feats",
|
|
type=str2bool,
|
|
default=False,
|
|
help="When enabled, use on-the-fly cut mixing and feature "
|
|
"extraction. Will drop existing precomputed feature manifests "
|
|
"if available.",
|
|
)
|
|
group.add_argument(
|
|
"--shuffle",
|
|
type=str2bool,
|
|
default=True,
|
|
help="When enabled (=default), the examples will be "
|
|
"shuffled for each epoch.",
|
|
)
|
|
group.add_argument(
|
|
"--return-cuts",
|
|
type=str2bool,
|
|
default=True,
|
|
help="When enabled, each batch will have the "
|
|
"field: batch['supervisions']['cut'] with the cuts that "
|
|
"were used to construct it.",
|
|
)
|
|
|
|
group.add_argument(
|
|
"--num-workers",
|
|
type=int,
|
|
default=2,
|
|
help="The number of training dataloader workers that "
|
|
"collect the batches.",
|
|
)
|
|
|
|
group.add_argument(
|
|
"--enable-spec-aug",
|
|
type=str2bool,
|
|
default=True,
|
|
help="When enabled, use SpecAugment for training dataset.",
|
|
)
|
|
|
|
group.add_argument(
|
|
"--spec-aug-time-warp-factor",
|
|
type=int,
|
|
default=80,
|
|
help="Used only when --enable-spec-aug is True. "
|
|
"It specifies the factor for time warping in SpecAugment. "
|
|
"Larger values mean more warping. "
|
|
"A value less than 1 means to disable time warp.",
|
|
)
|
|
|
|
group.add_argument(
|
|
"--enable-musan",
|
|
type=str2bool,
|
|
default=True,
|
|
help="When enabled, select noise from MUSAN and mix it "
|
|
"with training dataset. ",
|
|
)
|
|
|
|
# GigaSpeech specific arguments
|
|
group.add_argument(
|
|
"--subset",
|
|
type=str,
|
|
default="XL",
|
|
help="Select the GigaSpeech subset (XS|S|M|L|XL)",
|
|
)
|
|
group.add_argument(
|
|
"--small-dev",
|
|
type=str2bool,
|
|
default=False,
|
|
help="Should we use only 1000 utterances for dev "
|
|
"(speeds up training)",
|
|
)
|
|
|
|
def train_dataloaders(self, cuts_train: CutSet) -> DataLoader:
|
|
logging.info("About to get Musan cuts")
|
|
cuts_musan = load_manifest(
|
|
self.args.manifest_dir / "cuts_musan.json.gz"
|
|
)
|
|
|
|
transforms = []
|
|
if self.args.enable_musan:
|
|
logging.info("Enable MUSAN")
|
|
transforms.append(
|
|
CutMix(
|
|
cuts=cuts_musan, prob=0.5, snr=(10, 20), preserve_id=True
|
|
)
|
|
)
|
|
else:
|
|
logging.info("Disable MUSAN")
|
|
|
|
if self.args.concatenate_cuts:
|
|
logging.info(
|
|
f"Using cut concatenation with duration factor "
|
|
f"{self.args.duration_factor} and gap {self.args.gap}."
|
|
)
|
|
# Cut concatenation should be the first transform in the list,
|
|
# so that if we e.g. mix noise in, it will fill the gaps between
|
|
# different utterances.
|
|
transforms = [
|
|
CutConcatenate(
|
|
duration_factor=self.args.duration_factor, gap=self.args.gap
|
|
)
|
|
] + transforms
|
|
|
|
input_transforms = []
|
|
if self.args.enable_spec_aug:
|
|
logging.info("Enable SpecAugment")
|
|
logging.info(
|
|
f"Time warp factor: {self.args.spec_aug_time_warp_factor}"
|
|
)
|
|
input_transforms.append(
|
|
SpecAugment(
|
|
time_warp_factor=self.args.spec_aug_time_warp_factor,
|
|
num_frame_masks=2,
|
|
features_mask_size=27,
|
|
num_feature_masks=2,
|
|
frames_mask_size=100,
|
|
)
|
|
)
|
|
else:
|
|
logging.info("Disable SpecAugment")
|
|
|
|
logging.info("About to create train dataset")
|
|
train = K2SpeechRecognitionDataset(
|
|
cut_transforms=transforms,
|
|
input_transforms=input_transforms,
|
|
return_cuts=self.args.return_cuts,
|
|
)
|
|
|
|
if self.args.on_the_fly_feats:
|
|
# NOTE: the PerturbSpeed transform should be added only if we
|
|
# remove it from data prep stage.
|
|
# Add on-the-fly speed perturbation; since originally it would
|
|
# have increased epoch size by 3, we will apply prob 2/3 and use
|
|
# 3x more epochs.
|
|
# Speed perturbation probably should come first before
|
|
# concatenation, but in principle the transforms order doesn't have
|
|
# to be strict (e.g. could be randomized)
|
|
# transforms = [PerturbSpeed(factors=[0.9, 1.1], p=2/3)] + transforms # noqa
|
|
# Drop feats to be on the safe side.
|
|
train = K2SpeechRecognitionDataset(
|
|
cut_transforms=transforms,
|
|
input_strategy=OnTheFlyFeatures(
|
|
Fbank(FbankConfig(num_mel_bins=80))
|
|
),
|
|
input_transforms=input_transforms,
|
|
return_cuts=self.args.return_cuts,
|
|
)
|
|
|
|
if self.args.bucketing_sampler:
|
|
logging.info("Using DynamicBucketingSampler.")
|
|
train_sampler = DynamicBucketingSampler(
|
|
cuts_train,
|
|
max_duration=self.args.max_duration,
|
|
shuffle=self.args.shuffle,
|
|
num_buckets=self.args.num_buckets,
|
|
drop_last=True,
|
|
)
|
|
else:
|
|
logging.info("Using SingleCutSampler.")
|
|
train_sampler = SingleCutSampler(
|
|
cuts_train,
|
|
max_duration=self.args.max_duration,
|
|
shuffle=self.args.shuffle,
|
|
)
|
|
logging.info("About to create train dataloader")
|
|
|
|
train_dl = DataLoader(
|
|
train,
|
|
sampler=train_sampler,
|
|
batch_size=None,
|
|
num_workers=self.args.num_workers,
|
|
persistent_workers=False,
|
|
)
|
|
|
|
return train_dl
|
|
|
|
def valid_dataloaders(self, cuts_valid: CutSet) -> DataLoader:
|
|
transforms = []
|
|
if self.args.concatenate_cuts:
|
|
transforms = [
|
|
CutConcatenate(
|
|
duration_factor=self.args.duration_factor, gap=self.args.gap
|
|
)
|
|
] + transforms
|
|
|
|
logging.info("About to create dev dataset")
|
|
if self.args.on_the_fly_feats:
|
|
validate = K2SpeechRecognitionDataset(
|
|
cut_transforms=transforms,
|
|
input_strategy=OnTheFlyFeatures(
|
|
Fbank(FbankConfig(num_mel_bins=80))
|
|
),
|
|
return_cuts=self.args.return_cuts,
|
|
)
|
|
else:
|
|
validate = K2SpeechRecognitionDataset(
|
|
cut_transforms=transforms,
|
|
return_cuts=self.args.return_cuts,
|
|
)
|
|
valid_sampler = BucketingSampler(
|
|
cuts_valid,
|
|
max_duration=self.args.max_duration,
|
|
shuffle=False,
|
|
)
|
|
logging.info("About to create dev dataloader")
|
|
valid_dl = DataLoader(
|
|
validate,
|
|
sampler=valid_sampler,
|
|
batch_size=None,
|
|
num_workers=2,
|
|
persistent_workers=False,
|
|
)
|
|
|
|
return valid_dl
|
|
|
|
def test_dataloaders(self, cuts: CutSet) -> DataLoader:
|
|
logging.debug("About to create test dataset")
|
|
test = K2SpeechRecognitionDataset(
|
|
input_strategy=OnTheFlyFeatures(Fbank(FbankConfig(num_mel_bins=80)))
|
|
if self.args.on_the_fly_feats
|
|
else PrecomputedFeatures(),
|
|
return_cuts=self.args.return_cuts,
|
|
)
|
|
sampler = BucketingSampler(
|
|
cuts, max_duration=self.args.max_duration, shuffle=False
|
|
)
|
|
logging.debug("About to create test dataloader")
|
|
test_dl = DataLoader(
|
|
test,
|
|
batch_size=None,
|
|
sampler=sampler,
|
|
num_workers=self.args.num_workers,
|
|
)
|
|
return test_dl
|
|
|
|
@lru_cache()
|
|
def train_cuts(self) -> CutSet:
|
|
logging.info(f"About to get train_{self.args.subset} cuts")
|
|
path = self.args.manifest_dir / f"cuts_{self.args.subset}.jsonl.gz"
|
|
cuts_train = CutSet.from_jsonl_lazy(path)
|
|
return cuts_train
|
|
|
|
@lru_cache()
|
|
def dev_cuts(self) -> CutSet:
|
|
logging.info("About to get dev cuts")
|
|
cuts_valid = load_manifest(self.args.manifest_dir / "cuts_DEV.jsonl.gz")
|
|
if self.args.small_dev:
|
|
return cuts_valid.subset(first=1000)
|
|
else:
|
|
return cuts_valid
|
|
|
|
@lru_cache()
|
|
def test_cuts(self) -> CutSet:
|
|
logging.info("About to get test cuts")
|
|
return load_manifest(self.args.manifest_dir / "cuts_TEST.jsonl.gz")
|