PaddleScience/ppsci/data/__init__.py at develop · PaddlePaddle/PaddleScience · GitHub

History

207 lines (183 loc) · 8.27 KB

Raw

100

101

102

103

104

105

106

107

108

109

110

111

112

113

114

115

116

117

118

119

120

121

122

123

124

125

126

127

128

129

130

131

132

133

134

135

136

137

138

139

140

141

142

143

144

145

146

147

148

149

150

151

152

153

154

155

156

157

158

159

160

161

162

163

164

165

166

167

168

169

170

171

172

173

174

175

176

177

178

179

180

181

182

183

184

185

186

187

188

189

190

191

192

193

194

195

196

197

198

199

200

201

202

203

204

205

206

207

# 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 copy

import random

from functools import partial

from typing import Callable

from typing import Optional

import numpy as np

import paddle.distributed as dist

from paddle import device

from paddle import io

from ppsci.data import dataloader

from ppsci.data import dataset

from ppsci.data import process

from ppsci.data.dataset import register_to_dataset

from ppsci.data.process import batch_transform

from ppsci.data.process import transform

from ppsci.utils import logger

__all__ = [

"dataset",

"process",

"dataloader",

"build_dataloader",

"transform",

"batch_transform",

"register_to_dataset",

]

def worker_init_fn(worker_id: int, num_workers: int, rank: int, base_seed: int) -> None:

"""Callback function on each worker subprocess after seeding and before data loading.

Args:

worker_id (int): Worker id in [0, num_workers - 1].

num_workers (int): Number of subprocesses to use for data loading.

rank (int): Rank of process in distributed environment. If in non-distributed

environment, it is a constant number `0`.

base_seed (int): Base random seed.

"""

# The seed of each worker equals to: user_seed + num_worker * rank + worker_id

worker_seed = base_seed + num_workers * rank + worker_id

np.random.seed(worker_seed)

random.seed(worker_seed)

def build_dataloader(_dataset, cfg):

world_size = dist.get_world_size()

# just return IterableDataset as dataloader

if isinstance(_dataset, io.IterableDataset):

return _dataset

cfg = copy.deepcopy(cfg)

# build sampler

sampler_cfg = cfg.pop("sampler", None)

if sampler_cfg is not None:

batch_sampler_cls = sampler_cfg.pop("name")

if batch_sampler_cls == "BatchSampler":

if world_size > 1:

batch_sampler_cls = "DistributedBatchSampler"

logger.warning(

f"Automatically use 'DistributedBatchSampler' instead of "

f"'BatchSampler' when world_size({world_size}) > 1."

)

sampler_cfg["batch_size"] = cfg["batch_size"]

batch_sampler = getattr(io, batch_sampler_cls)(_dataset, **sampler_cfg)

else:

batch_sampler_cls = "BatchSampler"

if world_size > 1:

batch_sampler_cls = "DistributedBatchSampler"

logger.warning(

f"Automatically use 'DistributedBatchSampler' instead of "

f"'BatchSampler' when world_size({world_size}) > 1."

)

batch_sampler = getattr(io, batch_sampler_cls)(

_dataset,

batch_size=cfg["batch_size"],

shuffle=False,

drop_last=False,

)

logger.message(

"'shuffle' and 'drop_last' are both set to False in default as sampler config is not specified."

)

# build collate_fn if specified

batch_transforms_cfg = cfg.pop("batch_transforms", None)

collate_fn: Optional[Callable] = cfg.pop("collate_fn", None)

if isinstance(batch_transforms_cfg, (list, tuple)):

collate_fn = batch_transform.build_batch_transforms(

batch_transforms_cfg, collate_fn

)

# build init function

_DEFAULT_NUM_WORKERS = 1

_DEFAULT_SEED = 42

init_fn = partial(

worker_init_fn,

num_workers=cfg.get("num_workers", _DEFAULT_NUM_WORKERS),

rank=dist.get_rank(),

base_seed=cfg.get("seed", _DEFAULT_SEED),

)

# build dataloader

if getattr(_dataset, "use_pgl", False):

# Use special dataloader from "Paddle Graph Learning" toolkit.

try:

from pgl.utils import data as pgl_data

except ModuleNotFoundError as e:

logger.error("Please install pgl with `pip install pgl`.")

raise ModuleNotFoundError(str(e))

if collate_fn is None:

collate_fn = batch_transform.default_collate_fn

dataloader_ = pgl_data.Dataloader(

dataset=_dataset,

batch_size=cfg["batch_size"],

drop_last=sampler_cfg.get("drop_last", False),

shuffle=sampler_cfg.get("shuffle", False),

num_workers=cfg.get("num_workers", _DEFAULT_NUM_WORKERS),

collate_fn=collate_fn,

)

elif getattr(_dataset, "use_graph_grid_mesh", False):

# Use special dataloader `GridMeshAtmosphericDataset`.

if collate_fn is None:

collate_fn = batch_transform.default_collate_fn

dataloader_ = io.DataLoader(

dataset=_dataset,

places=device.get_device(),

batch_sampler=batch_sampler,

collate_fn=collate_fn,

num_workers=cfg.get("num_workers", _DEFAULT_NUM_WORKERS),

use_shared_memory=cfg.get("use_shared_memory", False),

worker_init_fn=init_fn,

)

else:

if (

cfg.get("auto_collation", not getattr(_dataset, "batch_index", False))

is False

if "transforms" in cfg["dataset"] and "auto_collation" not in cfg:

logger.warning(

"'transforms' and batch indexing(auto_collation=False) are both "

"enabled. If you do want to apply transforms to the batch samples, "

"please explicitly set 'auto_collation' to False in dataloader_cfg;"

" otherwise, the 'transforms' will be retained, but batch indexing "

"will be disabled."

)

else:

# 1. wrap batch_sampler again into BatchSampler for disabling auto collation,

# which can speed up the process of batch samples indexing from dataset. See

# details at: https://discuss.pytorch.org/t/efficiency-of-dataloader-and-collate-for-large-array-like-datasets/59569/8

batch_sampler = io.BatchSampler(sampler=batch_sampler, batch_size=1)

if collate_fn is not None:

raise NotImplementedError(

"Detected collate_fn is not None for 'batch_transforms' might "

"be specified in 'dataloader_cfg', which is not supported yet "

"with 'auto_collation' is False at the same time"

)

# 2. disable auto collation by given identity collate_fn which return the first

# (also the only) batch data in batch list, or there will be a redundant

# axis at the first dimension returned by dataloader. This step is necessary

# because paddle do not support 'sampler' as instantiation argument of 'io.DataLoader'

collate_fn = lambda batch: batch[0] # noqa: E731

_DEFAULT_NUM_WORKERS = 0

logger.info(

"Auto collation is disabled and set num_workers to "

f"{_DEFAULT_NUM_WORKERS} to speed up batch sampling."

)

dataloader_ = io.DataLoader(

dataset=_dataset,

places=device.get_device(),

batch_sampler=batch_sampler,

collate_fn=collate_fn,

num_workers=cfg.get("num_workers", _DEFAULT_NUM_WORKERS),

use_shared_memory=cfg.get("use_shared_memory", False),

worker_init_fn=init_fn,

# TODO: Do not enable 'persistent_workers' below for

# 'IndexError: pop from empty list ...' will be raised in certain cases

# persistent_workers=cfg.get("num_workers", _DEFAULT_NUM_WORKERS) > 0,

)

if len(dataloader_) == 0:

raise ValueError(

f"batch_size({sampler_cfg['batch_size']}) should not bigger than number of "

f"samples({len(_dataset)}) when drop_last is {sampler_cfg.get('drop_last', False)}."

)

return dataloader_

PaddleScience/ppsci/data/init.py at develop · PaddlePaddle/PaddleScience · GitHub

Uh oh!

PaddleScience/ppsci/data/__init__.py at develop · PaddlePaddle/PaddleScience · GitHub

Uh oh!

Expand file tree

Breadcrumbs

__init__.py

Latest commit

History

Breadcrumbs

__init__.py

File metadata and controls

PaddleScience/ppsci/data/init.py at develop · PaddlePaddle/PaddleScience · GitHub

init.py

init.py