mirrors/icefall
1
0
Fork 0
mirror of https://github.com/k2-fsa/icefall.git synced 2025-08-08 09:32:20 +00:00
Code Issues Packages Projects Releases Wiki Activity
icefall/egs/libricss/SURT/dprnn_zipformer/asr_datamodule.py

373 lines
12 KiB
Python
Raw Blame History

# Copyright 2021 Piotr Żelasko
# Copyright 2022 Xiaomi Corporation (Author: Mingshuang Luo)
# Copyright 2023 Johns Hopkins Univrtsity (Author: 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.
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 LibriCssAsrDataModule:
"""
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,
- 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,
return_sources: bool = True,
strict: bool = True,
) -> 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=return_sources,
strict=strict,
)
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 SimpleCutSampler.")
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,
max_cuts=self.args.max_cuts,
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 = 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,
)
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 lsmix_cuts(
self,
rvb_affix: str = "clean",
type_affix: str = "full",
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}_{type_affix}{source_affix}.jsonl.gz"
)
cs = cs.filter(lambda c: c.duration >= 1.0 and c.duration <= 30.0)
return cs
@lru_cache()
def libricss_cuts(self, split="dev", type="sdm") -> CutSet:
logging.info(f"About to get LibriCSS {split} {type} cuts")
cs = load_manifest_lazy(
self.args.manifest_dir / f"cuts_{split}_libricss-{type}.jsonl.gz"
)
return cs
Reference in New Issue View Git Blame Copy Permalink