wekws/kws/dataset/dataset.py
2021-11-10 18:48:57 +08:00

163 lines
5.0 KiB
Python

# Copyright (c) 2021 Binbin Zhang
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import random
import torch
import torch.distributed as dist
from torch.utils.data import IterableDataset
import kws.dataset.processor as processor
from kws.utils.file_utils import read_lists
class Processor(IterableDataset):
def __init__(self, source, f, *args, **kw):
assert callable(f)
self.source = source
self.f = f
self.args = args
self.kw = kw
def set_epoch(self, epoch):
self.source.set_epoch(epoch)
def __iter__(self):
""" Return an iterator over the source dataset processed by the
given processor.
"""
assert self.source is not None
assert callable(self.f)
return self.f(iter(self.source), *self.args, **self.kw)
def apply(self, f):
assert callable(f)
return Processor(self, f, *self.args, **self.kw)
class DistributedSampler:
def __init__(self, shuffle=True, partition=True):
self.epoch = -1
self.update()
self.shuffle = shuffle
self.partition = partition
def update(self):
assert dist.is_available()
if dist.is_initialized():
self.rank = dist.get_rank()
self.world_size = dist.get_world_size()
else:
self.rank = 0
self.world_size = 1
worker_info = torch.utils.data.get_worker_info()
if worker_info is None:
self.worker_id = 0
self.num_workers = 1
else:
self.worker_id = worker_info.id
self.num_workers = worker_info.num_workers
return dict(rank=self.rank,
world_size=self.world_size,
worker_id=self.worker_id,
num_workers=self.num_workers)
def set_epoch(self, epoch):
self.epoch = epoch
def sample(self, data):
""" Sample data according to rank/world_size/num_workers
Args:
data(List): input data list
Returns:
List: data list after sample
"""
data = data.copy()
if self.partition:
if self.shuffle:
random.Random(self.epoch).shuffle(data)
data = data[self.rank::self.world_size]
data = data[self.worker_id::self.num_workers]
return data
class DataList(IterableDataset):
def __init__(self, lists, shuffle=True, partition=True):
self.lists = lists
self.sampler = DistributedSampler(shuffle, partition)
def set_epoch(self, epoch):
self.sampler.set_epoch(epoch)
def __iter__(self):
sampler_info = self.sampler.update()
lists = self.sampler.sample(self.lists)
for src in lists:
# yield dict(src=src)
data = dict(src=src)
data.update(sampler_info)
yield data
def Dataset(data_list_file, conf, partition=True):
""" Construct dataset from arguments
We have two shuffle stage in the Dataset. The first is global
shuffle at shards tar/raw file level. The second is global shuffle
at training samples level.
Args:
data_type(str): raw/shard
partition(bool): whether to do data partition in terms of rank
"""
lists = read_lists(data_list_file)
shuffle = conf.get('shuffle', True)
dataset = DataList(lists, shuffle=shuffle, partition=partition)
dataset = Processor(dataset, processor.parse_raw)
filter_conf = conf.get('filter_conf', {})
dataset = Processor(dataset, processor.filter, **filter_conf)
resample_conf = conf.get('resample_conf', {})
dataset = Processor(dataset, processor.resample, **resample_conf)
speed_perturb = conf.get('speed_perturb', False)
if speed_perturb:
dataset = Processor(dataset, processor.speed_perturb)
fbank_conf = conf.get('fbank_conf', {})
dataset = Processor(dataset, processor.compute_fbank, **fbank_conf)
spec_aug = conf.get('spec_aug', True)
if spec_aug:
spec_aug_conf = conf.get('spec_aug_conf', {})
dataset = Processor(dataset, processor.spec_aug, **spec_aug_conf)
if shuffle:
shuffle_conf = conf.get('shuffle_conf', {})
dataset = Processor(dataset, processor.shuffle, **shuffle_conf)
batch_conf = conf.get('batch_conf', {})
dataset = Processor(dataset, processor.batch, **batch_conf)
dataset = Processor(dataset, processor.padding)
return dataset
if __name__ == '__main__':
import sys
dataset = Dataset(sys.argv[1], {})
for data in dataset:
print(data)