-
Notifications
You must be signed in to change notification settings - Fork 103
/
imbalance_cifar.py
208 lines (180 loc) · 8.06 KB
/
imbalance_cifar.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
# To ensure fairness, we use the same code in LDAM (https://github.com/kaidic/LDAM-DRW) to produce long-tailed CIFAR datasets.
import torchvision
import torchvision.transforms as transforms
import numpy as np
from PIL import Image
import random
class IMBALANCECIFAR10(torchvision.datasets.CIFAR10):
cls_num = 10
def __init__(self, mode, cfg, root = './datasets/imbalance_cifar10', imb_type='exp',
transform=None, target_transform=None, download=True):
train = True if mode == "train" else False
super(IMBALANCECIFAR10, self).__init__(root, train, transform, target_transform, download)
self.cfg = cfg
self.train = train
self.dual_sample = True if cfg.TRAIN.SAMPLER.DUAL_SAMPLER.ENABLE and self.train else False
rand_number = cfg.DATASET.IMBALANCECIFAR.RANDOM_SEED
if self.train:
np.random.seed(rand_number)
random.seed(rand_number)
imb_factor = self.cfg.DATASET.IMBALANCECIFAR.RATIO
img_num_list = self.get_img_num_per_cls(self.cls_num, imb_type, imb_factor)
self.gen_imbalanced_data(img_num_list)
self.transform = transforms.Compose([
transforms.RandomCrop(32, padding=4),
transforms.RandomHorizontalFlip(),
transforms.ToTensor(),
transforms.Normalize((0.4914, 0.4822, 0.4465), (0.2023, 0.1994, 0.2010)),
])
else:
self.transform = transforms.Compose([
transforms.ToTensor(),
transforms.Normalize((0.4914, 0.4822, 0.4465), (0.2023, 0.1994, 0.2010)),
])
print("{} Mode: Contain {} images".format(mode, len(self.data)))
if self.dual_sample or (self.cfg.TRAIN.SAMPLER.TYPE == "weighted sampler" and self.train):
self.class_weight, self.sum_weight = self.get_weight(self.get_annotations(), self.cls_num)
self.class_dict = self._get_class_dict()
def get_img_num_per_cls(self, cls_num, imb_type, imb_factor):
img_max = len(self.data) / cls_num
img_num_per_cls = []
if imb_type == 'exp':
for cls_idx in range(cls_num):
num = img_max * (imb_factor**(cls_idx / (cls_num - 1.0)))
img_num_per_cls.append(int(num))
elif imb_type == 'step':
for cls_idx in range(cls_num // 2):
img_num_per_cls.append(int(img_max))
for cls_idx in range(cls_num // 2):
img_num_per_cls.append(int(img_max * imb_factor))
else:
img_num_per_cls.extend([int(img_max)] * cls_num)
return img_num_per_cls
def sample_class_index_by_weight(self):
rand_number, now_sum = random.random() * self.sum_weight, 0
for i in range(self.cls_num):
now_sum += self.class_weight[i]
if rand_number <= now_sum:
return i
def reset_epoch(self, cur_epoch):
self.epoch = cur_epoch
def _get_class_dict(self):
class_dict = dict()
for i, anno in enumerate(self.get_annotations()):
cat_id = anno["category_id"]
if not cat_id in class_dict:
class_dict[cat_id] = []
class_dict[cat_id].append(i)
return class_dict
def get_weight(self, annotations, num_classes):
num_list = [0] * num_classes
cat_list = []
for anno in annotations:
category_id = anno["category_id"]
num_list[category_id] += 1
cat_list.append(category_id)
max_num = max(num_list)
class_weight = [max_num / i for i in num_list]
sum_weight = sum(class_weight)
return class_weight, sum_weight
def __getitem__(self, index):
"""
Args:
index (int): Index
Returns:
tuple: (image, target) where target is index of the target class.
"""
if self.cfg.TRAIN.SAMPLER.TYPE == "weighted sampler" and self.train:
assert self.cfg.TRAIN.SAMPLER.WEIGHTED_SAMPLER.TYPE in ["balance", "reverse"]
if self.cfg.TRAIN.SAMPLER.WEIGHTED_SAMPLER.TYPE == "balance":
sample_class = random.randint(0, self.cls_num - 1)
elif self.cfg.TRAIN.SAMPLER.WEIGHTED_SAMPLER.TYPE == "reverse":
sample_class = self.sample_class_index_by_weight()
sample_indexes = self.class_dict[sample_class]
index = random.choice(sample_indexes)
img, target = self.data[index], self.targets[index]
meta = dict()
# doing this so that it is consistent with all other datasets
# to return a PIL Image
img = Image.fromarray(img)
if self.dual_sample:
if self.cfg.TRAIN.SAMPLER.DUAL_SAMPLER.TYPE == "reverse":
sample_class = self.sample_class_index_by_weight()
sample_indexes = self.class_dict[sample_class]
sample_index = random.choice(sample_indexes)
elif self.cfg.TRAIN.SAMPLER.DUAL_SAMPLER.TYPE == "balance":
sample_class = random.randint(0, self.cls_num-1)
sample_indexes = self.class_dict[sample_class]
sample_index = random.choice(sample_indexes)
elif self.cfg.TRAIN.SAMPLER.DUAL_SAMPLER.TYPE == "uniform":
sample_index = random.randint(0, self.__len__() - 1)
sample_img, sample_label = self.data[sample_index], self.targets[sample_index]
sample_img = Image.fromarray(sample_img)
sample_img = self.transform(sample_img)
meta['sample_image'] = sample_img
meta['sample_label'] = sample_label
if self.transform is not None:
img = self.transform(img)
if self.target_transform is not None:
target = self.target_transform(target)
return img, target, meta
def get_num_classes(self):
return self.cls_num
def reset_epoch(self, epoch):
self.epoch = epoch
def get_annotations(self):
annos = []
for target in self.targets:
annos.append({'category_id': int(target)})
return annos
def gen_imbalanced_data(self, img_num_per_cls):
new_data = []
new_targets = []
targets_np = np.array(self.targets, dtype=np.int64)
classes = np.unique(targets_np)
# np.random.shuffle(classes)
self.num_per_cls_dict = dict()
for the_class, the_img_num in zip(classes, img_num_per_cls):
self.num_per_cls_dict[the_class] = the_img_num
idx = np.where(targets_np == the_class)[0]
np.random.shuffle(idx)
selec_idx = idx[:the_img_num]
new_data.append(self.data[selec_idx, ...])
new_targets.extend([the_class, ] * the_img_num)
new_data = np.vstack(new_data)
self.data = new_data
self.targets = new_targets
def get_cls_num_list(self):
cls_num_list = []
for i in range(self.cls_num):
cls_num_list.append(self.num_per_cls_dict[i])
return cls_num_list
class IMBALANCECIFAR100(IMBALANCECIFAR10):
"""`CIFAR100 <https://www.cs.toronto.edu/~kriz/cifar.html>`_ Dataset.
This is a subclass of the `CIFAR10` Dataset.
"""
base_folder = 'cifar-100-python'
url = "https://www.cs.toronto.edu/~kriz/cifar-100-python.tar.gz"
filename = "cifar-100-python.tar.gz"
tgz_md5 = 'eb9058c3a382ffc7106e4002c42a8d85'
train_list = [
['train', '16019d7e3df5f24257cddd939b257f8d'],
]
test_list = [
['test', 'f0ef6b0ae62326f3e7ffdfab6717acfc'],
]
meta = {
'filename': 'meta',
'key': 'fine_label_names',
'md5': '7973b15100ade9c7d40fb424638fde48',
}
cls_num = 100
if __name__ == '__main__':
transform = transforms.Compose(
[transforms.ToTensor(),
transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5))])
trainset = IMBALANCECIFAR100(root='./data', train=True,
download=True, transform=transform)
trainloader = iter(trainset)
data, label = next(trainloader)
import pdb; pdb.set_trace()