-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathimagenet.py
201 lines (138 loc) · 7.78 KB
/
imagenet.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
import torchvision
import numpy as np
import os
from copy import deepcopy
from data.data_utils import subsample_instances
from config import imagenet_root
class ImageNetBase(torchvision.datasets.ImageFolder):
def __init__(self, root, transform):
super(ImageNetBase, self).__init__(root, transform)
self.uq_idxs = np.array(range(len(self)))
def __getitem__(self, item):
img, label = super().__getitem__(item)
uq_idx = self.uq_idxs[item]
return img, label, uq_idx
def subsample_dataset(dataset, idxs):
imgs_ = []
for i in idxs:
imgs_.append(dataset.imgs[i])
dataset.imgs = imgs_
samples_ = []
for i in idxs:
samples_.append(dataset.samples[i])
dataset.samples = samples_
# dataset.imgs = [x for i, x in enumerate(dataset.imgs) if i in idxs]
# dataset.samples = [x for i, x in enumerate(dataset.samples) if i in idxs]
dataset.targets = np.array(dataset.targets)[idxs].tolist()
dataset.uq_idxs = dataset.uq_idxs[idxs]
return dataset
def subsample_classes(dataset, include_classes=list(range(1000))):
cls_idxs = [x for x, t in enumerate(dataset.targets) if t in include_classes]
target_xform_dict = {}
for i, k in enumerate(include_classes):
target_xform_dict[k] = i
dataset = subsample_dataset(dataset, cls_idxs)
dataset.target_transform = lambda x: target_xform_dict[x]
return dataset
def get_train_val_indices(train_dataset, val_split=0.2):
train_classes = list(set(train_dataset.targets))
# Get train/test indices
train_idxs = []
val_idxs = []
for cls in train_classes:
cls_idxs = np.where(np.array(train_dataset.targets) == cls)[0]
v_ = np.random.choice(cls_idxs, replace=False, size=((int(val_split * len(cls_idxs))),))
t_ = [x for x in cls_idxs if x not in v_]
train_idxs.extend(t_)
val_idxs.extend(v_)
return train_idxs, val_idxs
def get_imagenet_100_datasets(train_transform, test_transform, train_classes=range(80),
prop_train_labels=0.8, split_train_val=False, seed=0):
np.random.seed(seed)
# Subsample imagenet dataset initially to include 100 classes
subsampled_100_classes = np.random.choice(range(1000), size=(100,), replace=False)
subsampled_100_classes = np.sort(subsampled_100_classes)
print(f'Constructing ImageNet-100 dataset from the following classes: {subsampled_100_classes.tolist()}')
cls_map = {i: j for i, j in zip(subsampled_100_classes, range(100))}
# Init entire training set
imagenet_training_set = ImageNetBase(root=os.path.join(imagenet_root, 'train'), transform=train_transform)
whole_training_set = subsample_classes(imagenet_training_set, include_classes=subsampled_100_classes)
# Reset dataset
whole_training_set.samples = [(s[0], cls_map[s[1]]) for s in whole_training_set.samples]
whole_training_set.targets = [s[1] for s in whole_training_set.samples]
whole_training_set.uq_idxs = np.array(range(len(whole_training_set)))
whole_training_set.target_transform = None
# Get labelled training set which has subsampled classes, then subsample some indices from that
train_dataset_labelled = subsample_classes(deepcopy(whole_training_set), include_classes=train_classes)
subsample_indices = subsample_instances(train_dataset_labelled, prop_indices_to_subsample=prop_train_labels)
train_dataset_labelled = subsample_dataset(train_dataset_labelled, subsample_indices)
# Split into training and validation sets
train_idxs, val_idxs = get_train_val_indices(train_dataset_labelled)
train_dataset_labelled_split = subsample_dataset(deepcopy(train_dataset_labelled), train_idxs)
val_dataset_labelled_split = subsample_dataset(deepcopy(train_dataset_labelled), val_idxs)
val_dataset_labelled_split.transform = test_transform
# Get unlabelled data
unlabelled_indices = set(whole_training_set.uq_idxs) - set(train_dataset_labelled.uq_idxs)
train_dataset_unlabelled = subsample_dataset(deepcopy(whole_training_set), np.array(list(unlabelled_indices)))
# Get test set for all classes
test_dataset = ImageNetBase(root=os.path.join(imagenet_root, 'val'), transform=test_transform)
test_dataset = subsample_classes(test_dataset, include_classes=subsampled_100_classes)
# Reset test set
test_dataset.samples = [(s[0], cls_map[s[1]]) for s in test_dataset.samples]
test_dataset.targets = [s[1] for s in test_dataset.samples]
test_dataset.uq_idxs = np.array(range(len(test_dataset)))
test_dataset.target_transform = None
# Either split train into train and val or use test set as val
train_dataset_labelled = train_dataset_labelled_split if split_train_val else train_dataset_labelled
val_dataset_labelled = val_dataset_labelled_split if split_train_val else None
all_datasets = {
'train_labelled': train_dataset_labelled,
'train_unlabelled': train_dataset_unlabelled,
'val': val_dataset_labelled,
'test': test_dataset,
}
return all_datasets
def get_imagenet_1k_datasets(train_transform, test_transform, train_classes=range(500),
prop_train_labels=0.5, split_train_val=False, seed=0):
np.random.seed(seed)
# Init entire training set
whole_training_set = ImageNetBase(root=os.path.join(imagenet_root, 'train'), transform=train_transform)
# Get labelled training set which has subsampled classes, then subsample some indices from that
train_dataset_labelled = subsample_classes(deepcopy(whole_training_set), include_classes=train_classes)
subsample_indices = subsample_instances(train_dataset_labelled, prop_indices_to_subsample=prop_train_labels)
train_dataset_labelled = subsample_dataset(train_dataset_labelled, subsample_indices)
# Split into training and validation sets
train_idxs, val_idxs = get_train_val_indices(train_dataset_labelled)
train_dataset_labelled_split = subsample_dataset(deepcopy(train_dataset_labelled), train_idxs)
val_dataset_labelled_split = subsample_dataset(deepcopy(train_dataset_labelled), val_idxs)
val_dataset_labelled_split.transform = test_transform
# Get unlabelled data
unlabelled_indices = set(whole_training_set.uq_idxs) - set(train_dataset_labelled.uq_idxs)
train_dataset_unlabelled = subsample_dataset(deepcopy(whole_training_set), np.array(list(unlabelled_indices)))
# Get test set for all classes
test_dataset = ImageNetBase(root=os.path.join(imagenet_root, 'val'), transform=test_transform)
# Either split train into train and val or use test set as val
train_dataset_labelled = train_dataset_labelled_split if split_train_val else train_dataset_labelled
val_dataset_labelled = val_dataset_labelled_split if split_train_val else None
all_datasets = {
'train_labelled': train_dataset_labelled,
'train_unlabelled': train_dataset_unlabelled,
'val': val_dataset_labelled,
'test': test_dataset,
}
return all_datasets
if __name__ == '__main__':
x = get_imagenet_100_datasets(None, None, split_train_val=False,
train_classes=range(50), prop_train_labels=0.5)
print('Printing lens...')
for k, v in x.items():
if v is not None:
print(f'{k}: {len(v)}')
print('Printing labelled and unlabelled overlap...')
print(set.intersection(set(x['train_labelled'].uq_idxs), set(x['train_unlabelled'].uq_idxs)))
print('Printing total instances in train...')
print(len(set(x['train_labelled'].uq_idxs)) + len(set(x['train_unlabelled'].uq_idxs)))
print(f'Num Labelled Classes: {len(set(x["train_labelled"].targets))}')
print(f'Num Unabelled Classes: {len(set(x["train_unlabelled"].targets))}')
print(f'Len labelled set: {len(x["train_labelled"])}')
print(f'Len unlabelled set: {len(x["train_unlabelled"])}')