-
Notifications
You must be signed in to change notification settings - Fork 6
/
loader.py
245 lines (205 loc) · 9.67 KB
/
loader.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
from pathlib import Path
from random import randint
import PIL
from torch.utils.data import Dataset, DataLoader
from torchvision import transforms as T
from torchvision.datasets import ImageFolder, FakeData
from pytorch_lightning import LightningDataModule
import torch
from typing import Any, Tuple
import webdataset as wds
from PIL import Image
from io import BytesIO
#To prevent truncated error
from PIL import ImageFile
ImageFile.LOAD_TRUNCATED_IMAGES = True
def web_dataset_helper(path):
if Path(path).is_dir():
DATASET = [str(p) for p in Path(path).glob("**/*") if ".tar" in str(p).lower()] # .name
assert len(DATASET) > 0, 'The directory ({}) does not contain any WebDataset/.tar files.'.format(path)
print('Found {} WebDataset .tar(.gz) file(s) under given path {}!'.format(len(DATASET), path))
elif ('http://' in path.lower()) | ('https://' in path.lower()):
DATASET = f"pipe:curl -L -s {path} || true"
print('Found {} http(s) link under given path!'.format(len(DATASET), path))
elif 'gs://' in path.lower():
DATASET = f"pipe:gsutil cat {path} || true"
print('Found {} GCS link under given path!'.format(len(DATASET), path))
elif '.tar' in path:
DATASET = path
print('Found WebDataset .tar(.gz) file under given path {}!'.format(path))
else:
raise Exception('No folder, no .tar(.gz) and no url pointing to tar files provided under {}.'.format(path))
return DATASET
def identity(x):
return x
class Grayscale2RGB:
def __init__(self):
pass
def __call__(self, img):
if img.mode != 'RGB':
return img.convert('RGB')
else:
return img
def __repr__(self):
return self.__class__.__name__ + '()'
class ImageDataModule(LightningDataModule):
def __init__(self, train_dir, val_dir, batch_size, num_workers, img_size, resize_ratio=0.75,
fake_data=False, web_dataset=False, world_size = 1, dataset_size = [int(1e9)]):
super().__init__()
self.train_dir = train_dir
self.val_dir = val_dir
self.batch_size = batch_size
self.num_workers = num_workers
self.fake_data = fake_data
self.img_size = img_size
self.web_dataset = web_dataset
if len(dataset_size) == 1:
self.train_dataset_size = dataset_size[0]
self.val_dataset_size = dataset_size[0]
else:
self.train_dataset_size = dataset_size[0]
self.val_dataset_size = dataset_size[1]
self.world_size = world_size
self.transform_train = T.Compose([
Grayscale2RGB(),
T.RandomResizedCrop(img_size,
scale=(resize_ratio, 1.),ratio=(1., 1.)),
T.ToTensor(),
T.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5)),
])
self.transform_val = T.Compose([
Grayscale2RGB(),
T.Resize(img_size),
T.CenterCrop(img_size),
T.ToTensor(),
T.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5)),
])
def imagetransform(self, b):
return Image.open(BytesIO(b))
def dummy(self, s):
return torch.zeros(1)
def setup(self, stage=None):
if self.fake_data:
self.train_dataset = FakeData(12000000, (3, self.img_size, self.img_size), 1000, self.transform_train)
self.val_dataset = FakeData(50000, (3, self.img_size, self.img_size), 1000, self.transform_val)
self.transform_train = None
self.transform_val = None
else:
if self.web_dataset:
DATASET_TRAIN = web_dataset_helper(self.train_dir)
DATASET_VAL = web_dataset_helper(self.val_dir)
self.train_dataset = (
wds.WebDataset(DATASET_TRAIN, handler=wds.warn_and_continue)
.shuffle(1000, handler=wds.warn_and_continue)
.decode("pil", handler=wds.warn_and_continue)
.to_tuple("jpg;png;jpeg", handler=wds.warn_and_continue)
.map_tuple(self.transform_train, handler=wds.warn_and_continue)
.batched(self.batch_size, partial=False) # It is good to avoid partial batches when using Distributed training
)
self.val_dataset = (
wds.WebDataset(DATASET_VAL, handler=wds.warn_and_continue)
.decode("pil", handler=wds.warn_and_continue)
.to_tuple("jpg;png;jpeg", handler=wds.warn_and_continue)
.map_tuple(self.transform_val, handler=wds.warn_and_continue)
.batched(self.batch_size, partial=False) # It is good to avoid partial batches when using Distributed training
)
else:
self.train_dataset = ImageDataset(self.train_dir, self.transform_train)
self.val_dataset = ImageDataset(self.val_dir, self.transform_val)
def train_dataloader(self):
if self.web_dataset:
dl = wds.WebLoader(self.train_dataset, batch_size=None, num_workers=self.num_workers)
number_of_batches = self.train_dataset_size // (self.batch_size * self.world_size)
dl = dl.repeat(9999999999).slice(number_of_batches)
dl.length = number_of_batches
return dl
else:
return DataLoader(self.train_dataset, batch_size=self.batch_size, num_workers=self.num_workers, shuffle=True)
def val_dataloader(self):
if self.web_dataset:
dl = wds.WebLoader(self.val_dataset, batch_size=None, num_workers=self.num_workers)
number_of_batches = self.val_dataset_size // (self.batch_size * self.world_size)
dl = dl.repeat(9999999999).slice(number_of_batches)
dl.length = number_of_batches
return dl
else:
return DataLoader(self.val_dataset, batch_size=self.batch_size, num_workers=self.num_workers)
def test_dataloader(self):
#simply reuse val_dataloader for test.
if self.web_dataset:
dl = wds.WebLoader(self.val_dataset, batch_size=None, num_workers=self.num_workers)
number_of_batches = self.val_dataset_size // (self.batch_size * self.world_size)
dl = dl.repeat(9999999999).slice(number_of_batches)
dl.length = number_of_batches
return dl
else:
return DataLoader(self.val_dataset, batch_size=self.batch_size, num_workers=self.num_workers)
IMG_EXTENSIONS = ('.jpg', '.jpeg', '.png', '.ppm', '.bmp', '.pgm', '.tif', '.tiff', '.webp')
class ImageDataset(ImageFolder):
def random_sample(self):
return self.__getitem__(randint(0, self.__len__() - 1))
def sequential_sample(self, ind):
if ind >= self.__len__() - 1:
return self.__getitem__(0)
return self.__getitem__(ind + 1)
def skip_sample(self, ind):
return self.random_sample()
def __getitem__(self, index: int) -> Tuple[Any, Any]:
"""
Args:
index (int): Index
Returns:
tuple: (sample, target) where target is class_index of the target class.
"""
try:
path, target = self.samples[index]
sample = self.loader(path)
except (PIL.UnidentifiedImageError, OSError) as corrupt_image_exceptions:
print(corrupt_image_exceptions)
print(f"An exception occurred trying to load file {path}.")
print(f"Skipping index {index}")
return self.skip_sample(index)
if self.transform is not None:
sample = self.transform(sample)
if self.target_transform is not None:
target = self.target_transform(target)
return sample, target
class ImageDataset2(Dataset):
def __init__(self,
folder,
transform=None,
shuffle=False,
):
"""
@param folder: Folder containing images and text files matched by their paths' respective "stem"
@param truncate_captions: Rather than throw an exception, captions which are too long will be truncated.
"""
super().__init__()
self.shuffle = shuffle
path = Path(folder)
self.image_files = [
*path.glob('**/*.png'), *path.glob('**/*.jpg'),
*path.glob('**/*.jpeg'), *path.glob('**/*.bmp')
]
self.transform = transform
def __len__(self):
return len(self.image_files)
def random_sample(self):
return self.__getitem__(randint(0, self.__len__() - 1))
def sequential_sample(self, ind):
if ind >= self.__len__() - 1:
return self.__getitem__(0)
return self.__getitem__(ind + 1)
def skip_sample(self, ind):
if self.shuffle:
return self.random_sample()
return self.sequential_sample(ind=ind)
def __getitem__(self, ind):
try:
image_tensor = self.transform(PIL.Image.open(self.image_files[ind]))
except (PIL.UnidentifiedImageError, OSError) as corrupt_image_exceptions:
print(corrupt_image_exceptions)
print(f"An exception occurred trying to load file {self.image_files[ind]}.")
print(f"Skipping index {ind}")
return self.skip_sample(ind)
return image_tensor