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

Decrease num_buckets & remove shuffle_buffer_size (#1955)

Browse Source
This commit is contained in:
Wei Kang 2025-06-19 12:26:37 +08:00 committed by GitHub
parent 3587c4b3b7
commit f80a2ee110
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
7 changed files with 13 additions and 495 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

5
egs/aishell/ASR/tdnn_lstm_ctc/asr_datamodule.py
View File

@ -92,7 +92,7 @@ class AishellAsrDataModule:
group.add_argument(
"--num-buckets",
type=int,
default=30,
default=15,
help="The number of buckets for the DynamicBucketingSampler"
"(you might want to increase it for larger datasets).",
)
@ -275,8 +275,7 @@ class AishellAsrDataModule:
max_duration=self.args.max_duration,
shuffle=self.args.shuffle,
num_buckets=self.args.num_buckets,
buffer_size=self.args.num_buckets * 2000,
shuffle_buffer_size=self.args.num_buckets * 5000,
buffer_size=self.args.num_buckets * 5000,
drop_last=self.args.drop_last,
)
else:

5
egs/aishell2/ASR/pruned_transducer_stateless5/asr_datamodule.py
View File

@ -104,7 +104,7 @@ class AiShell2AsrDataModule:
group.add_argument(
"--num-buckets",
type=int,
default=30,
default=15,
help="The number of buckets for the DynamicBucketingSampler"
"(you might want to increase it for larger datasets).",
)
@ -296,8 +296,7 @@ class AiShell2AsrDataModule:
max_duration=self.args.max_duration,
shuffle=self.args.shuffle,
num_buckets=self.args.num_buckets,
buffer_size=self.args.num_buckets * 2000,
shuffle_buffer_size=self.args.num_buckets * 5000,
buffer_size=self.args.num_buckets * 5000,
drop_last=self.args.drop_last,
)
else:

5
egs/gigaspeech/ASR/pruned_transducer_stateless2/asr_datamodule.py
View File

@ -101,7 +101,7 @@ class GigaSpeechAsrDataModule:
group.add_argument(
"--num-buckets",
type=int,
default=30,
default=15,
help="The number of buckets for the DynamicBucketingSampler"
"(you might want to increase it for larger datasets).",
)
@ -294,8 +294,7 @@ class GigaSpeechAsrDataModule:
max_duration=self.args.max_duration,
shuffle=self.args.shuffle,
num_buckets=self.args.num_buckets,
buffer_size=self.args.num_buckets * 2000,
shuffle_buffer_size=self.args.num_buckets * 5000,
buffer_size=self.args.num_buckets * 5000,
drop_last=True,
)
else:

7
egs/gigaspeech/ASR/zipformer/asr_datamodule.py
View File

@ -105,7 +105,7 @@ class GigaSpeechAsrDataModule:
group.add_argument(
"--num-buckets",
type=int,
default=100,
default=15,
help="The number of buckets for the DynamicBucketingSampler"
"(you might want to increase it for larger datasets).",
)
@ -311,8 +311,7 @@ class GigaSpeechAsrDataModule:
max_duration=self.args.max_duration,
shuffle=self.args.shuffle,
num_buckets=self.args.num_buckets,
buffer_size=self.args.num_buckets * 2000,
shuffle_buffer_size=self.args.num_buckets * 5000,
buffer_size=self.args.num_buckets * 5000,
drop_last=self.args.drop_last,
)
else:
@ -369,7 +368,7 @@ class GigaSpeechAsrDataModule:
cuts_valid,
max_duration=self.args.max_duration,
num_buckets=self.args.num_buckets,
buffer_size=self.args.num_buckets * 2000,
buffer_size=self.args.num_buckets * 5000,
shuffle=False,
)
logging.info("About to create dev dataloader")

477
egs/gigaspeech/KWS/zipformer/asr_datamodule.py
View File

@ -1,477 +0,0 @@
# Copyright 2021 Piotr Żelasko
# Copyright 2024 Xiaomi Corporation (Author: Wei Kang)
#
# See ../../../../LICENSE for clarification regarding multiple authors
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import argparse
import glob
import inspect
import logging
import re
from functools import lru_cache
from pathlib import Path
from typing import Any, Dict, Optional
import lhotse
import torch
from lhotse import CutSet, Fbank, FbankConfig, load_manifest, load_manifest_lazy
from lhotse.dataset import (
CutConcatenate,
CutMix,
DynamicBucketingSampler,
K2SpeechRecognitionDataset,
PrecomputedFeatures,
SimpleCutSampler,
SpecAugment,
)
from lhotse.dataset.input_strategies import AudioSamples, OnTheFlyFeatures
from lhotse.utils import fix_random_seed
from torch.utils.data import DataLoader
from icefall.utils import str2bool
class _SeedWorkers:
def __init__(self, seed: int):
self.seed = seed
def __call__(self, worker_id: int):
fix_random_seed(self.seed + worker_id)
class 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(
"--drop-last",
type=str2bool,
default=True,
help="Whether to drop last batch. Used by sampler.",
)
group.add_argument(
"--return-cuts",
type=str2bool,
default=True,
help="When enabled, each batch will have the "
"field: batch['supervisions']['cut'] with the cuts that "
"were used to construct it.",
)
group.add_argument(
"--num-workers",
type=int,
default=2,
help="The number of training dataloader workers that "
"collect the batches.",
)
group.add_argument(
"--enable-spec-aug",
type=str2bool,
default=True,
help="When enabled, use SpecAugment for training dataset.",
)
group.add_argument(
"--spec-aug-time-warp-factor",
type=int,
default=80,
help="Used only when --enable-spec-aug is True. "
"It specifies the factor for time warping in SpecAugment. "
"Larger values mean more warping. "
"A value less than 1 means to disable time warp.",
)
group.add_argument(
"--enable-musan",
type=str2bool,
default=True,
help="When enabled, select noise from MUSAN and mix it"
"with training dataset. ",
)
group.add_argument(
"--input-strategy",
type=str,
default="PrecomputedFeatures",
help="AudioSamples or PrecomputedFeatures",
)
# 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,
sampler_state_dict: Optional[Dict[str, Any]] = None,
) -> DataLoader:
"""
Args:
cuts_train:
CutSet for training.
sampler_state_dict:
The state dict for the training sampler.
"""
transforms = []
if self.args.enable_musan:
logging.info("Enable MUSAN")
logging.info("About to get Musan cuts")
cuts_musan = load_manifest(self.args.manifest_dir / "musan_cuts.jsonl.gz")
transforms.append(
CutMix(cuts=cuts_musan, p=0.5, snr=(10, 20), preserve_id=True)
)
else:
logging.info("Disable MUSAN")
if self.args.concatenate_cuts:
logging.info(
f"Using cut concatenation with duration factor "
f"{self.args.duration_factor} and gap {self.args.gap}."
)
# Cut concatenation should be the first transform in the list,
# so that if we e.g. mix noise in, it will fill the gaps between
# different utterances.
transforms = [
CutConcatenate(
duration_factor=self.args.duration_factor, gap=self.args.gap
)
] + transforms
input_transforms = []
if self.args.enable_spec_aug:
logging.info("Enable SpecAugment")
logging.info(f"Time warp factor: {self.args.spec_aug_time_warp_factor}")
# Set the value of num_frame_masks according to Lhotse's version.
# In different Lhotse's versions, the default of num_frame_masks is
# different.
num_frame_masks = 10
num_frame_masks_parameter = inspect.signature(
SpecAugment.__init__
).parameters["num_frame_masks"]
if num_frame_masks_parameter.default == 1:
num_frame_masks = 2
logging.info(f"Num frame mask: {num_frame_masks}")
input_transforms.append(
SpecAugment(
time_warp_factor=self.args.spec_aug_time_warp_factor,
num_frame_masks=num_frame_masks,
features_mask_size=27,
num_feature_masks=2,
frames_mask_size=100,
)
)
else:
logging.info("Disable SpecAugment")
logging.info("About to create train dataset")
train = K2SpeechRecognitionDataset(
input_strategy=eval(self.args.input_strategy)(),
cut_transforms=transforms,
input_transforms=input_transforms,
return_cuts=self.args.return_cuts,
)
if self.args.on_the_fly_feats:
# NOTE: the PerturbSpeed transform should be added only if we
# remove it from data prep stage.
# Add on-the-fly speed perturbation; since originally it would
# have increased epoch size by 3, we will apply prob 2/3 and use
# 3x more epochs.
# Speed perturbation probably should come first before
# concatenation, but in principle the transforms order doesn't have
# to be strict (e.g. could be randomized)
# transforms = [PerturbSpeed(factors=[0.9, 1.1], p=2/3)] + transforms # noqa
# Drop feats to be on the safe side.
train = K2SpeechRecognitionDataset(
cut_transforms=transforms,
input_strategy=OnTheFlyFeatures(Fbank(FbankConfig(num_mel_bins=80))),
input_transforms=input_transforms,
return_cuts=self.args.return_cuts,
)
if self.args.bucketing_sampler:
logging.info("Using DynamicBucketingSampler.")
train_sampler = DynamicBucketingSampler(
cuts_train,
max_duration=self.args.max_duration,
shuffle=self.args.shuffle,
num_buckets=self.args.num_buckets,
drop_last=self.args.drop_last,
buffer_size=self.args.num_buckets * 2000,
shuffle_buffer_size=self.args.num_buckets * 5000,
)
else:
logging.info("Using SimpleCutSampler.")
train_sampler = SimpleCutSampler(
cuts_train,
max_duration=self.args.max_duration,
shuffle=self.args.shuffle,
)
logging.info("About to create train dataloader")
if sampler_state_dict is not None:
logging.info("Loading sampler state dict")
train_sampler.load_state_dict(sampler_state_dict)
# 'seed' is derived from the current random state, which will have
# previously been set in the main process.
seed = torch.randint(0, 100000, ()).item()
worker_init_fn = _SeedWorkers(seed)
train_dl = DataLoader(
train,
sampler=train_sampler,
batch_size=None,
num_workers=self.args.num_workers,
persistent_workers=False,
worker_init_fn=worker_init_fn,
)
return train_dl
def valid_dataloaders(self, cuts_valid: CutSet) -> DataLoader:
transforms = []
if self.args.concatenate_cuts:
transforms = [
CutConcatenate(
duration_factor=self.args.duration_factor, gap=self.args.gap
)
] + transforms
logging.info("About to create dev dataset")
if self.args.on_the_fly_feats:
validate = K2SpeechRecognitionDataset(
cut_transforms=transforms,
input_strategy=OnTheFlyFeatures(Fbank(FbankConfig(num_mel_bins=80))),
return_cuts=self.args.return_cuts,
)
else:
validate = K2SpeechRecognitionDataset(
cut_transforms=transforms,
return_cuts=self.args.return_cuts,
)
valid_sampler = DynamicBucketingSampler(
cuts_valid,
max_duration=self.args.max_duration,
num_buckets=self.args.num_buckets,
buffer_size=self.args.num_buckets * 2000,
shuffle=False,
)
logging.info("About to create dev dataloader")
valid_dl = DataLoader(
validate,
sampler=valid_sampler,
batch_size=None,
num_workers=2,
persistent_workers=False,
)
return valid_dl
def test_dataloaders(self, cuts: CutSet) -> DataLoader:
logging.debug("About to create test dataset")
test = K2SpeechRecognitionDataset(
input_strategy=OnTheFlyFeatures(Fbank(FbankConfig(num_mel_bins=80)))
if self.args.on_the_fly_feats
else eval(self.args.input_strategy)(),
return_cuts=self.args.return_cuts,
)
sampler = DynamicBucketingSampler(
cuts,
max_duration=self.args.max_duration,
shuffle=False,
)
logging.debug("About to create test dataloader")
test_dl = DataLoader(
test,
batch_size=None,
sampler=sampler,
num_workers=self.args.num_workers,
)
return test_dl
@lru_cache()
def train_cuts(self) -> CutSet:
logging.info(f"About to get train {self.args.subset} cuts")
if self.args.subset == "XL":
filenames = glob.glob(
f"{self.args.manifest_dir}/XL_split/gigaspeech_cuts_XL.*.jsonl.gz"
)
pattern = re.compile(r"gigaspeech_cuts_XL.([0-9]+).jsonl.gz")
idx_filenames = ((int(pattern.search(f).group(1)), f) for f in filenames)
idx_filenames = sorted(idx_filenames, key=lambda x: x[0])
sorted_filenames = [f[1] for f in idx_filenames]
logging.info(
f"Loading GigaSpeech {len(sorted_filenames)} splits in lazy mode"
)
cuts_train = lhotse.combine(
lhotse.load_manifest_lazy(p) for p in sorted_filenames
)
else:
path = (
self.args.manifest_dir / f"gigaspeech_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_lazy(
self.args.manifest_dir / "gigaspeech_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_lazy(
self.args.manifest_dir / "gigaspeech_cuts_TEST.jsonl.gz"
)
@lru_cache()
def fsc_train_cuts(self) -> CutSet:
logging.info("About to get fluent speech commands train cuts")
return load_manifest_lazy(
self.args.manifest_dir / "fluent_speech_commands_cuts_train.jsonl.gz"
)
@lru_cache()
def fsc_valid_cuts(self) -> CutSet:
logging.info("About to get fluent speech commands valid cuts")
return load_manifest_lazy(
self.args.manifest_dir / "fluent_speech_commands_cuts_valid.jsonl.gz"
)
@lru_cache()
def fsc_test_small_cuts(self) -> CutSet:
logging.info("About to get fluent speech commands small test cuts")
return load_manifest_lazy(
self.args.manifest_dir / "fluent_speech_commands_cuts_small.jsonl.gz"
)
@lru_cache()
def fsc_test_large_cuts(self) -> CutSet:
logging.info("About to get fluent speech commands large test cuts")
return load_manifest_lazy(
self.args.manifest_dir / "fluent_speech_commands_cuts_large.jsonl.gz"
)

1
egs/gigaspeech/KWS/zipformer/asr_datamodule.py Symbolic link
View File

@ -0,0 +1 @@
../../ASR/zipformer/asr_datamodule.py

3
egs/librispeech/ASR/tdnn_lstm_ctc/asr_datamodule.py
View File

@ -311,8 +311,7 @@ class LibriSpeechAsrDataModule:
max_duration=self.args.max_duration,
shuffle=self.args.shuffle,
num_buckets=self.args.num_buckets,
buffer_size=self.args.num_buckets * 2000,
shuffle_buffer_size=self.args.num_buckets * 5000,
buffer_size=self.args.num_buckets * 5000,
drop_last=self.args.drop_last,
)
else:

5
egs/wenetspeech/ASR/pruned_transducer_stateless2/asr_datamodule.py
View File

@ -106,7 +106,7 @@ class WenetSpeechAsrDataModule:
group.add_argument(
"--num-buckets",
type=int,
default=30,
default=15,
help="The number of buckets for the DynamicBucketingSampler"
"(you might want to increase it for larger datasets).",
)
@ -292,8 +292,7 @@ class WenetSpeechAsrDataModule:
max_duration=self.args.max_duration,
shuffle=self.args.shuffle,
num_buckets=self.args.num_buckets,
buffer_size=self.args.num_buckets * 2000,
shuffle_buffer_size=self.args.num_buckets * 5000,
buffer_size=self.args.num_buckets * 5000,
drop_last=True,
)
else: