-
Notifications
You must be signed in to change notification settings - Fork 560
/
dataloader.py
305 lines (245 loc) · 11.1 KB
/
dataloader.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
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
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
# Copyright 2019-2020 Stanislav Pidhorskyi
#
# 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 dareblopy as db
import random
import numpy as np
import torch
import torch.tensor
import torch.utils
import torch.utils.data
import time
import math
cpu = torch.device('cpu')
class TFRecordsDataset:
def __init__(self, cfg, logger, rank=0, world_size=1, buffer_size_mb=200, channels=3, seed=None, train=True, needs_labels=False):
self.cfg = cfg
self.logger = logger
self.rank = rank
self.last_data = ""
if train:
self.part_count = cfg.DATASET.PART_COUNT
self.part_size = cfg.DATASET.SIZE // self.part_count
else:
self.part_count = cfg.DATASET.PART_COUNT_TEST
self.part_size = cfg.DATASET.SIZE_TEST // self.part_count
self.workers = []
self.workers_active = 0
self.iterator = None
self.filenames = {}
self.batch_size = 512
self.features = {}
self.channels = channels
self.seed = seed
self.train = train
self.needs_labels = needs_labels
assert self.part_count % world_size == 0
self.part_count_local = self.part_count // world_size
if train:
path = cfg.DATASET.PATH
else:
path = cfg.DATASET.PATH_TEST
for r in range(2, cfg.DATASET.MAX_RESOLUTION_LEVEL + 1):
files = []
for i in range(self.part_count_local * rank, self.part_count_local * (rank + 1)):
file = path % (r, i)
files.append(file)
self.filenames[r] = files
self.buffer_size_b = 1024 ** 2 * buffer_size_mb
self.current_filenames = []
def reset(self, lod, batch_size):
assert lod in self.filenames.keys()
self.current_filenames = self.filenames[lod]
self.batch_size = batch_size
img_size = 2 ** lod
if self.needs_labels:
self.features = {
# 'shape': db.FixedLenFeature([3], db.int64),
'data': db.FixedLenFeature([self.channels, img_size, img_size], db.uint8),
'label': db.FixedLenFeature([], db.int64)
}
else:
self.features = {
# 'shape': db.FixedLenFeature([3], db.int64),
'data': db.FixedLenFeature([self.channels, img_size, img_size], db.uint8)
}
buffer_size = self.buffer_size_b // (self.channels * img_size * img_size)
if self.seed is None:
seed = np.uint64(time.time() * 1000)
else:
seed = self.seed
self.logger.info('!' * 80)
self.logger.info('! Seed is used for to shuffle data in TFRecordsDataset!')
self.logger.info('!' * 80)
self.iterator = db.ParsedTFRecordsDatasetIterator(self.current_filenames, self.features, self.batch_size, buffer_size, seed=seed)
def __iter__(self):
return self.iterator
def __len__(self):
return self.part_count_local * self.part_size
def make_dataloader(cfg, logger, dataset, GPU_batch_size, local_rank, numpy=False):
class BatchCollator(object):
def __init__(self, device=torch.device("cpu")):
self.device = device
self.flip = cfg.DATASET.FLIP_IMAGES
self.numpy = numpy
def __call__(self, batch):
with torch.no_grad():
x, = batch
if self.flip:
flips = [(slice(None, None, None), slice(None, None, None), slice(None, None, random.choice([-1, None]))) for _ in range(x.shape[0])]
x = np.array([img[flip] for img, flip in zip(x, flips)])
if self.numpy:
return x
x = torch.tensor(x, requires_grad=True, device=torch.device(self.device), dtype=torch.float32)
return x
batches = db.data_loader(iter(dataset), BatchCollator(local_rank), len(dataset) // GPU_batch_size)
return batches
def make_dataloader_y(cfg, logger, dataset, GPU_batch_size, local_rank):
class BatchCollator(object):
def __init__(self, device=torch.device("cpu")):
self.device = device
self.flip = cfg.DATASET.FLIP_IMAGES
def __call__(self, batch):
with torch.no_grad():
x, y = batch
if self.flip:
flips = [(slice(None, None, None), slice(None, None, None), slice(None, None, random.choice([-1, None]))) for _ in range(x.shape[0])]
x = np.array([img[flip] for img, flip in zip(x, flips)])
x = torch.tensor(x, requires_grad=True, device=torch.device(self.device), dtype=torch.float32)
return x, y
batches = db.data_loader(iter(dataset), BatchCollator(local_rank), len(dataset) // GPU_batch_size)
return batches
class TFRecordsDatasetImageNet:
def __init__(self, cfg, logger, rank=0, world_size=1, buffer_size_mb=200, channels=3, seed=None, train=True, needs_labels=False):
self.cfg = cfg
self.logger = logger
self.rank = rank
self.last_data = ""
self.part_count = cfg.DATASET.PART_COUNT
if train:
self.part_size = cfg.DATASET.SIZE // cfg.DATASET.PART_COUNT
else:
self.part_size = cfg.DATASET.SIZE_TEST // cfg.DATASET.PART_COUNT
self.workers = []
self.workers_active = 0
self.iterator = None
self.filenames = {}
self.batch_size = 512
self.features = {}
self.channels = channels
self.seed = seed
self.train = train
self.needs_labels = needs_labels
assert self.part_count % world_size == 0
self.part_count_local = cfg.DATASET.PART_COUNT // world_size
if train:
path = cfg.DATASET.PATH
else:
path = cfg.DATASET.PATH_TEST
for r in range(2, cfg.DATASET.MAX_RESOLUTION_LEVEL + 1):
files = []
for i in range(self.part_count_local * rank, self.part_count_local * (rank + 1)):
file = path % (r, i)
files.append(file)
self.filenames[r] = files
self.buffer_size_b = 1024 ** 2 * buffer_size_mb
self.current_filenames = []
def reset(self, lod, batch_size):
assert lod in self.filenames.keys()
self.current_filenames = self.filenames[lod]
self.batch_size = batch_size
if self.train:
img_size = 2 ** lod + 2 ** (lod - 3)
else:
img_size = 2 ** lod
if self.needs_labels:
self.features = {
'data': db.FixedLenFeature([3, img_size, img_size], db.uint8),
'label': db.FixedLenFeature([], db.int64)
}
else:
self.features = {
'data': db.FixedLenFeature([3, img_size, img_size], db.uint8)
}
buffer_size = self.buffer_size_b // (self.channels * img_size * img_size)
if self.seed is None:
seed = np.uint64(time.time() * 1000)
else:
seed = self.seed
self.logger.info('!' * 80)
self.logger.info('! Seed is used for to shuffle data in TFRecordsDataset!')
self.logger.info('!' * 80)
self.iterator = db.ParsedTFRecordsDatasetIterator(self.current_filenames, self.features, self.batch_size, buffer_size, seed=seed)
def __iter__(self):
return self.iterator
def __len__(self):
return self.part_count_local * self.part_size
def make_imagenet_dataloader(cfg, logger, dataset, GPU_batch_size, target_size, local_rank, do_random_crops=True):
class BatchCollator(object):
def __init__(self, device=torch.device("cpu")):
self.device = device
self.flip = cfg.DATASET.FLIP_IMAGES
self.size = target_size
p = math.log2(target_size)
self.source_size = 2 ** p + 2 ** (p - 3)
self.do_random_crops = do_random_crops
def __call__(self, batch):
with torch.no_grad():
x, = batch
if self.do_random_crops:
images = []
for im in x:
deltax = self.source_size - target_size
deltay = self.source_size - target_size
offx = np.random.randint(deltax + 1)
offy = np.random.randint(deltay + 1)
im = im[:, offy:offy + self.size, offx:offx + self.size]
images.append(im)
x = np.stack(images)
if self.flip:
flips = [(slice(None, None, None), slice(None, None, None), slice(None, None, random.choice([-1, None]))) for _ in range(x.shape[0])]
x = np.array([img[flip] for img, flip in zip(x, flips)])
x = torch.tensor(x, requires_grad=True, device=torch.device(self.device), dtype=torch.float32)
return x
batches = db.data_loader(iter(dataset), BatchCollator(local_rank), len(dataset) // GPU_batch_size)
return batches
def make_imagenet_dataloader_y(cfg, logger, dataset, GPU_batch_size, target_size, local_rank, do_random_crops=True):
class BatchCollator(object):
def __init__(self, device=torch.device("cpu")):
self.device = device
self.flip = cfg.DATASET.FLIP_IMAGES
self.size = target_size
p = math.log2(target_size)
self.source_size = 2 ** p + 2 ** (p - 3)
self.do_random_crops = do_random_crops
def __call__(self, batch):
with torch.no_grad():
x, y = batch
if self.do_random_crops:
images = []
for im in x:
deltax = self.source_size - target_size
deltay = self.source_size - target_size
offx = np.random.randint(deltax + 1)
offy = np.random.randint(deltay + 1)
im = im[:, offy:offy+self.size, offx:offx+self.size]
images.append(im)
x = np.stack(images)
if self.flip:
flips = [(slice(None, None, None), slice(None, None, None), slice(None, None, random.choice([-1, None]))) for _ in range(x.shape[0])]
x = np.array([img[flip] for img, flip in zip(x, flips)])
x = torch.tensor(x, requires_grad=True, device=torch.device(self.device), dtype=torch.float32)
return x, y
batches = db.data_loader(iter(dataset), BatchCollator(local_rank), len(dataset) // GPU_batch_size)
return batches