-
Notifications
You must be signed in to change notification settings - Fork 299
/
datasets.py
555 lines (451 loc) · 21.2 KB
/
datasets.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
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved
import os
import torch
from PIL import Image, ImageFile
from torchvision import transforms
import torchvision.datasets.folder
from torch.utils.data import TensorDataset, Subset, ConcatDataset, Dataset
from torchvision.datasets import MNIST, ImageFolder
from torchvision.transforms.functional import rotate
from wilds.datasets.camelyon17_dataset import Camelyon17Dataset
from wilds.datasets.fmow_dataset import FMoWDataset
ImageFile.LOAD_TRUNCATED_IMAGES = True
DATASETS = [
# Debug
"Debug28",
"Debug224",
# Small images
"ColoredMNIST",
"RotatedMNIST",
# Big images
"VLCS",
"PACS",
"OfficeHome",
"TerraIncognita",
"DomainNet",
"SVIRO",
# WILDS datasets
"WILDSCamelyon",
"WILDSFMoW",
# Spawrious datasets
"SpawriousO2O_easy",
"SpawriousO2O_medium",
"SpawriousO2O_hard",
"SpawriousM2M_easy",
"SpawriousM2M_medium",
"SpawriousM2M_hard",
]
def get_dataset_class(dataset_name):
"""Return the dataset class with the given name."""
if dataset_name not in globals():
raise NotImplementedError("Dataset not found: {}".format(dataset_name))
return globals()[dataset_name]
def num_environments(dataset_name):
return len(get_dataset_class(dataset_name).ENVIRONMENTS)
class MultipleDomainDataset:
N_STEPS = 5001 # Default, subclasses may override
CHECKPOINT_FREQ = 100 # Default, subclasses may override
N_WORKERS = 8 # Default, subclasses may override
ENVIRONMENTS = None # Subclasses should override
INPUT_SHAPE = None # Subclasses should override
def __getitem__(self, index):
return self.datasets[index]
def __len__(self):
return len(self.datasets)
class Debug(MultipleDomainDataset):
def __init__(self, root, test_envs, hparams):
super().__init__()
self.input_shape = self.INPUT_SHAPE
self.num_classes = 2
self.datasets = []
for _ in [0, 1, 2]:
self.datasets.append(
TensorDataset(
torch.randn(16, *self.INPUT_SHAPE),
torch.randint(0, self.num_classes, (16,))
)
)
class Debug28(Debug):
INPUT_SHAPE = (3, 28, 28)
ENVIRONMENTS = ['0', '1', '2']
class Debug224(Debug):
INPUT_SHAPE = (3, 224, 224)
ENVIRONMENTS = ['0', '1', '2']
class MultipleEnvironmentMNIST(MultipleDomainDataset):
def __init__(self, root, environments, dataset_transform, input_shape,
num_classes):
super().__init__()
if root is None:
raise ValueError('Data directory not specified!')
original_dataset_tr = MNIST(root, train=True, download=True)
original_dataset_te = MNIST(root, train=False, download=True)
original_images = torch.cat((original_dataset_tr.data,
original_dataset_te.data))
original_labels = torch.cat((original_dataset_tr.targets,
original_dataset_te.targets))
shuffle = torch.randperm(len(original_images))
original_images = original_images[shuffle]
original_labels = original_labels[shuffle]
self.datasets = []
for i in range(len(environments)):
images = original_images[i::len(environments)]
labels = original_labels[i::len(environments)]
self.datasets.append(dataset_transform(images, labels, environments[i]))
self.input_shape = input_shape
self.num_classes = num_classes
class ColoredMNIST(MultipleEnvironmentMNIST):
ENVIRONMENTS = ['+90%', '+80%', '-90%']
def __init__(self, root, test_envs, hparams):
super(ColoredMNIST, self).__init__(root, [0.1, 0.2, 0.9],
self.color_dataset, (2, 28, 28,), 2)
self.input_shape = (2, 28, 28,)
self.num_classes = 2
def color_dataset(self, images, labels, environment):
# # Subsample 2x for computational convenience
# images = images.reshape((-1, 28, 28))[:, ::2, ::2]
# Assign a binary label based on the digit
labels = (labels < 5).float()
# Flip label with probability 0.25
labels = self.torch_xor_(labels,
self.torch_bernoulli_(0.25, len(labels)))
# Assign a color based on the label; flip the color with probability e
colors = self.torch_xor_(labels,
self.torch_bernoulli_(environment,
len(labels)))
images = torch.stack([images, images], dim=1)
# Apply the color to the image by zeroing out the other color channel
images[torch.tensor(range(len(images))), (
1 - colors).long(), :, :] *= 0
x = images.float().div_(255.0)
y = labels.view(-1).long()
return TensorDataset(x, y)
def torch_bernoulli_(self, p, size):
return (torch.rand(size) < p).float()
def torch_xor_(self, a, b):
return (a - b).abs()
class RotatedMNIST(MultipleEnvironmentMNIST):
ENVIRONMENTS = ['0', '15', '30', '45', '60', '75']
def __init__(self, root, test_envs, hparams):
super(RotatedMNIST, self).__init__(root, [0, 15, 30, 45, 60, 75],
self.rotate_dataset, (1, 28, 28,), 10)
def rotate_dataset(self, images, labels, angle):
rotation = transforms.Compose([
transforms.ToPILImage(),
transforms.Lambda(lambda x: rotate(x, angle, fill=(0,),
interpolation=torchvision.transforms.InterpolationMode.BILINEAR)),
transforms.ToTensor()])
x = torch.zeros(len(images), 1, 28, 28)
for i in range(len(images)):
x[i] = rotation(images[i])
y = labels.view(-1)
return TensorDataset(x, y)
class MultipleEnvironmentImageFolder(MultipleDomainDataset):
def __init__(self, root, test_envs, augment, hparams):
super().__init__()
environments = [f.name for f in os.scandir(root) if f.is_dir()]
environments = sorted(environments)
transform = transforms.Compose([
transforms.Resize((224,224)),
transforms.ToTensor(),
transforms.Normalize(
mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])
])
augment_transform = transforms.Compose([
# transforms.Resize((224,224)),
transforms.RandomResizedCrop(224, scale=(0.7, 1.0)),
transforms.RandomHorizontalFlip(),
transforms.ColorJitter(0.3, 0.3, 0.3, 0.3),
transforms.RandomGrayscale(),
transforms.ToTensor(),
transforms.Normalize(
mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),
])
self.datasets = []
for i, environment in enumerate(environments):
if augment and (i not in test_envs):
env_transform = augment_transform
else:
env_transform = transform
path = os.path.join(root, environment)
env_dataset = ImageFolder(path,
transform=env_transform)
self.datasets.append(env_dataset)
self.input_shape = (3, 224, 224,)
self.num_classes = len(self.datasets[-1].classes)
class VLCS(MultipleEnvironmentImageFolder):
CHECKPOINT_FREQ = 300
ENVIRONMENTS = ["C", "L", "S", "V"]
def __init__(self, root, test_envs, hparams):
self.dir = os.path.join(root, "VLCS/")
super().__init__(self.dir, test_envs, hparams['data_augmentation'], hparams)
class PACS(MultipleEnvironmentImageFolder):
CHECKPOINT_FREQ = 300
ENVIRONMENTS = ["A", "C", "P", "S"]
def __init__(self, root, test_envs, hparams):
self.dir = os.path.join(root, "PACS/")
super().__init__(self.dir, test_envs, hparams['data_augmentation'], hparams)
class DomainNet(MultipleEnvironmentImageFolder):
CHECKPOINT_FREQ = 1000
ENVIRONMENTS = ["clip", "info", "paint", "quick", "real", "sketch"]
def __init__(self, root, test_envs, hparams):
self.dir = os.path.join(root, "domain_net/")
super().__init__(self.dir, test_envs, hparams['data_augmentation'], hparams)
class OfficeHome(MultipleEnvironmentImageFolder):
CHECKPOINT_FREQ = 300
ENVIRONMENTS = ["A", "C", "P", "R"]
def __init__(self, root, test_envs, hparams):
self.dir = os.path.join(root, "office_home/")
super().__init__(self.dir, test_envs, hparams['data_augmentation'], hparams)
class TerraIncognita(MultipleEnvironmentImageFolder):
CHECKPOINT_FREQ = 300
ENVIRONMENTS = ["L100", "L38", "L43", "L46"]
def __init__(self, root, test_envs, hparams):
self.dir = os.path.join(root, "terra_incognita/")
super().__init__(self.dir, test_envs, hparams['data_augmentation'], hparams)
class SVIRO(MultipleEnvironmentImageFolder):
CHECKPOINT_FREQ = 300
ENVIRONMENTS = ["aclass", "escape", "hilux", "i3", "lexus", "tesla", "tiguan", "tucson", "x5", "zoe"]
def __init__(self, root, test_envs, hparams):
self.dir = os.path.join(root, "sviro/")
super().__init__(self.dir, test_envs, hparams['data_augmentation'], hparams)
class WILDSEnvironment:
def __init__(
self,
wilds_dataset,
metadata_name,
metadata_value,
transform=None):
self.name = metadata_name + "_" + str(metadata_value)
metadata_index = wilds_dataset.metadata_fields.index(metadata_name)
metadata_array = wilds_dataset.metadata_array
subset_indices = torch.where(
metadata_array[:, metadata_index] == metadata_value)[0]
self.dataset = wilds_dataset
self.indices = subset_indices
self.transform = transform
def __getitem__(self, i):
x = self.dataset.get_input(self.indices[i])
if type(x).__name__ != "Image":
x = Image.fromarray(x)
y = self.dataset.y_array[self.indices[i]]
if self.transform is not None:
x = self.transform(x)
return x, y
def __len__(self):
return len(self.indices)
class WILDSDataset(MultipleDomainDataset):
INPUT_SHAPE = (3, 224, 224)
def __init__(self, dataset, metadata_name, test_envs, augment, hparams):
super().__init__()
transform = transforms.Compose([
transforms.Resize((224, 224)),
transforms.ToTensor(),
transforms.Normalize(
mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])
])
augment_transform = transforms.Compose([
transforms.Resize((224, 224)),
transforms.RandomResizedCrop(224, scale=(0.7, 1.0)),
transforms.RandomHorizontalFlip(),
transforms.ColorJitter(0.3, 0.3, 0.3, 0.3),
transforms.RandomGrayscale(),
transforms.ToTensor(),
transforms.Normalize(
mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),
])
self.datasets = []
for i, metadata_value in enumerate(
self.metadata_values(dataset, metadata_name)):
if augment and (i not in test_envs):
env_transform = augment_transform
else:
env_transform = transform
env_dataset = WILDSEnvironment(
dataset, metadata_name, metadata_value, env_transform)
self.datasets.append(env_dataset)
self.input_shape = (3, 224, 224,)
self.num_classes = dataset.n_classes
def metadata_values(self, wilds_dataset, metadata_name):
metadata_index = wilds_dataset.metadata_fields.index(metadata_name)
metadata_vals = wilds_dataset.metadata_array[:, metadata_index]
return sorted(list(set(metadata_vals.view(-1).tolist())))
class WILDSCamelyon(WILDSDataset):
ENVIRONMENTS = [ "hospital_0", "hospital_1", "hospital_2", "hospital_3",
"hospital_4"]
def __init__(self, root, test_envs, hparams):
dataset = Camelyon17Dataset(root_dir=root)
super().__init__(
dataset, "hospital", test_envs, hparams['data_augmentation'], hparams)
class WILDSFMoW(WILDSDataset):
ENVIRONMENTS = [ "region_0", "region_1", "region_2", "region_3",
"region_4", "region_5"]
def __init__(self, root, test_envs, hparams):
dataset = FMoWDataset(root_dir=root)
super().__init__(
dataset, "region", test_envs, hparams['data_augmentation'], hparams)
## Spawrious base classes
class CustomImageFolder(Dataset):
"""
A class that takes one folder at a time and loads a set number of images in a folder and assigns them a specific class
"""
def __init__(self, folder_path, class_index, limit=None, transform=None):
self.folder_path = folder_path
self.class_index = class_index
self.image_paths = [os.path.join(folder_path, img) for img in os.listdir(folder_path) if img.endswith(('.png', '.jpg', '.jpeg'))]
if limit:
self.image_paths = self.image_paths[:limit]
self.transform = transform
def __len__(self):
return len(self.image_paths)
def __getitem__(self, index):
img_path = self.image_paths[index]
img = Image.open(img_path).convert('RGB')
if self.transform:
img = self.transform(img)
label = torch.tensor(self.class_index, dtype=torch.long)
return img, label
class SpawriousBenchmark(MultipleDomainDataset):
ENVIRONMENTS = ["Test", "SC_group_1", "SC_group_2"]
input_shape = (3, 224, 224)
num_classes = 4
class_list = ["bulldog", "corgi", "dachshund", "labrador"]
def __init__(self, train_combinations, test_combinations, root_dir, augment=True, type1=False):
self.type1 = type1
train_datasets, test_datasets = self._prepare_data_lists(train_combinations, test_combinations, root_dir, augment)
self.datasets = [ConcatDataset(test_datasets)] + train_datasets
# Prepares the train and test data lists by applying the necessary transformations.
def _prepare_data_lists(self, train_combinations, test_combinations, root_dir, augment):
test_transforms = transforms.Compose([
transforms.Resize((self.input_shape[1], self.input_shape[2])),
transforms.transforms.ToTensor(),
transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),
])
if augment:
train_transforms = transforms.Compose([
transforms.Resize((self.input_shape[1], self.input_shape[2])),
transforms.RandomHorizontalFlip(),
transforms.ColorJitter(0.3, 0.3, 0.3, 0.3),
transforms.RandomGrayscale(),
transforms.ToTensor(),
transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),
])
else:
train_transforms = test_transforms
train_data_list = self._create_data_list(train_combinations, root_dir, train_transforms)
test_data_list = self._create_data_list(test_combinations, root_dir, test_transforms)
return train_data_list, test_data_list
# Creates a list of datasets based on the given combinations and transformations.
def _create_data_list(self, combinations, root_dir, transforms):
data_list = []
if isinstance(combinations, dict):
# Build class groups for a given set of combinations, root directory, and transformations.
for_each_class_group = []
cg_index = 0
for classes, comb_list in combinations.items():
for_each_class_group.append([])
for ind, location_limit in enumerate(comb_list):
if isinstance(location_limit, tuple):
location, limit = location_limit
else:
location, limit = location_limit, None
cg_data_list = []
for cls in classes:
path = os.path.join(root_dir, f"{0 if not self.type1 else ind}/{location}/{cls}")
data = CustomImageFolder(folder_path=path, class_index=self.class_list.index(cls), limit=limit, transform=transforms)
cg_data_list.append(data)
for_each_class_group[cg_index].append(ConcatDataset(cg_data_list))
cg_index += 1
for group in range(len(for_each_class_group[0])):
data_list.append(
ConcatDataset(
[for_each_class_group[k][group] for k in range(len(for_each_class_group))]
)
)
else:
for location in combinations:
path = os.path.join(root_dir, f"{0}/{location}/")
data = ImageFolder(root=path, transform=transforms)
data_list.append(data)
return data_list
# Buils combination dictionary for o2o datasets
def build_type1_combination(self,group,test,filler):
total = 3168
counts = [int(0.97*total),int(0.87*total)]
combinations = {}
combinations['train_combinations'] = {
## correlated class
("bulldog",):[(group[0],counts[0]),(group[0],counts[1])],
("dachshund",):[(group[1],counts[0]),(group[1],counts[1])],
("labrador",):[(group[2],counts[0]),(group[2],counts[1])],
("corgi",):[(group[3],counts[0]),(group[3],counts[1])],
## filler
("bulldog","dachshund","labrador","corgi"):[(filler,total-counts[0]),(filler,total-counts[1])],
}
## TEST
combinations['test_combinations'] = {
("bulldog",):[test[0], test[0]],
("dachshund",):[test[1], test[1]],
("labrador",):[test[2], test[2]],
("corgi",):[test[3], test[3]],
}
return combinations
# Buils combination dictionary for m2m datasets
def build_type2_combination(self,group,test):
total = 3168
counts = [total,total]
combinations = {}
combinations['train_combinations'] = {
## correlated class
("bulldog",):[(group[0],counts[0]),(group[1],counts[1])],
("dachshund",):[(group[1],counts[0]),(group[0],counts[1])],
("labrador",):[(group[2],counts[0]),(group[3],counts[1])],
("corgi",):[(group[3],counts[0]),(group[2],counts[1])],
}
combinations['test_combinations'] = {
("bulldog",):[test[0], test[1]],
("dachshund",):[test[1], test[0]],
("labrador",):[test[2], test[3]],
("corgi",):[test[3], test[2]],
}
return combinations
## Spawrious classes for each Spawrious dataset
class SpawriousO2O_easy(SpawriousBenchmark):
def __init__(self, root_dir, test_envs, hparams):
group = ["desert","jungle","dirt","snow"]
test = ["dirt","snow","desert","jungle"]
filler = "beach"
combinations = self.build_type1_combination(group,test,filler)
super().__init__(combinations['train_combinations'], combinations['test_combinations'], root_dir, hparams['data_augmentation'], type1=True)
class SpawriousO2O_medium(SpawriousBenchmark):
def __init__(self, root_dir, test_envs, hparams):
group = ['mountain', 'beach', 'dirt', 'jungle']
test = ['jungle', 'dirt', 'beach', 'snow']
filler = "desert"
combinations = self.build_type1_combination(group,test,filler)
super().__init__(combinations['train_combinations'], combinations['test_combinations'], root_dir, hparams['data_augmentation'], type1=True)
class SpawriousO2O_hard(SpawriousBenchmark):
def __init__(self, root_dir, test_envs, hparams):
group = ['jungle', 'mountain', 'snow', 'desert']
test = ['mountain', 'snow', 'desert', 'jungle']
filler = "beach"
combinations = self.build_type1_combination(group,test,filler)
super().__init__(combinations['train_combinations'], combinations['test_combinations'], root_dir, hparams['data_augmentation'], type1=True)
class SpawriousM2M_easy(SpawriousBenchmark):
def __init__(self, root_dir, test_envs, hparams):
group = ['desert', 'mountain', 'dirt', 'jungle']
test = ['dirt', 'jungle', 'mountain', 'desert']
combinations = self.build_type2_combination(group,test)
super().__init__(combinations['train_combinations'], combinations['test_combinations'], root_dir, hparams['data_augmentation'])
class SpawriousM2M_medium(SpawriousBenchmark):
def __init__(self, root_dir, test_envs, hparams):
group = ['beach', 'snow', 'mountain', 'desert']
test = ['desert', 'mountain', 'beach', 'snow']
combinations = self.build_type2_combination(group,test)
super().__init__(combinations['train_combinations'], combinations['test_combinations'], root_dir, hparams['data_augmentation'])
class SpawriousM2M_hard(SpawriousBenchmark):
ENVIRONMENTS = ["Test","SC_group_1","SC_group_2"]
def __init__(self, root_dir, test_envs, hparams):
group = ["dirt","jungle","snow","beach"]
test = ["snow","beach","dirt","jungle"]
combinations = self.build_type2_combination(group,test)
super().__init__(combinations['train_combinations'], combinations['test_combinations'], root_dir, hparams['data_augmentation'])