-
Notifications
You must be signed in to change notification settings - Fork 2
/
Copy pathdataloader_cifar.py
293 lines (264 loc) · 15.4 KB
/
dataloader_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
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
from torch.utils.data import Dataset, DataLoader, WeightedRandomSampler
import torchvision.transforms as transforms
import random
import numpy as np
from PIL import Image
import json
import os
from autoaugment import CIFAR10Policy
import torch
from collections import Counter
def unpickle(file):
import _pickle as cPickle
with open(file, 'rb') as fo:
dict = cPickle.load(fo, encoding='latin1')
return dict
class cifar_dataset(Dataset):
def __init__(self, dataset, r, noise_mode, root_dir, transform, mode, noise_file='', mask=[], conf=[], transform_strong=None):
self.r = r # noise ratio
self.transform = transform
self.transform_strong = transform_strong
self.mode = mode
self.transition = {0:0,2:0,4:7,7:7,1:1,9:1,3:5,5:3,6:6,8:8} # class transition for asymmetric noise
if self.mode=='test':
if dataset=='cifar10':
test_dic = unpickle('%s/test_batch'%root_dir)
self.test_data = test_dic['data']
self.test_data = self.test_data.reshape((10000, 3, 32, 32))
self.test_data = self.test_data.transpose((0, 2, 3, 1))
self.test_label = test_dic['labels']
elif dataset=='cifar100':
test_dic = unpickle('%s/test'%root_dir)
self.test_data = test_dic['data']
self.test_data = self.test_data.reshape((10000, 3, 32, 32))
self.test_data = self.test_data.transpose((0, 2, 3, 1))
self.test_label = test_dic['fine_labels']
else:
train_data=[]
train_label=[]
if dataset=='cifar10':
for n in range(1,6):
dpath = '%s/data_batch_%d'%(root_dir,n)
data_dic = unpickle(dpath)
train_data.append(data_dic['data'])
train_label = train_label+data_dic['labels']
train_data = np.concatenate(train_data)
elif dataset=='cifar100':
train_dic = unpickle('%s/train'%root_dir)
train_data = train_dic['data']
train_label = train_dic['fine_labels']
train_data = train_data.reshape((50000, 3, 32, 32))
train_data = train_data.transpose((0, 2, 3, 1))
gt_noise_file = noise_file + "_groundtruth"
if os.path.exists(noise_file) and os.path.exists(gt_noise_file):
noise_label = json.load(open(noise_file,"r"))
gt_noise_label = json.load(open(gt_noise_file,"r")) #train_label
else: #inject noise
noise_label = []
gt_noise_label = []
idx = list(range(50000))
random.shuffle(idx)
num_noise = int(self.r*50000)
noise_idx = idx[:num_noise]
for i in range(50000):
if i in noise_idx:
if noise_mode=='sym':
if dataset=='cifar10':
noiselabel = random.randint(0,9)
elif dataset=='cifar100':
noiselabel = random.randint(0,99)
noise_label.append(noiselabel)
elif noise_mode=='asym':
noiselabel = self.transition[train_label[i]]
noise_label.append(noiselabel)
else:
noise_label.append(train_label[i])
gt_noise_label.append(train_label[i])
json.dump(noise_label,open(noise_file,"w"))
json.dump(gt_noise_label,open(gt_noise_file,"w"))
if self.mode == 'all' or self.mode == 'all_ssl':
self.train_data = train_data
self.noise_label = noise_label
self.gt_noise_label = gt_noise_label
elif self.mode == 'all_correction':
pred_idx = mask.nonzero()[0]
self.conf = [conf[i] for i in pred_idx]
self.train_data = train_data[pred_idx]
self.noise_label = [noise_label[i] for i in pred_idx]
self.gt_noise_label = [gt_noise_label[i] for i in pred_idx]
elif self.mode == 'clean_eval_train':
self.train_data = train_data
self.noise_label = noise_label
self.gt_noise_label = gt_noise_label
self.conf = conf
else:
pred_idx = mask.nonzero()[0]
self.conf = [conf[i] for i in pred_idx]
self.train_data = train_data[pred_idx]
self.noise_label = [noise_label[i] for i in pred_idx]
self.gt_noise_label = [gt_noise_label[i] for i in pred_idx]
# self.train_data = train_data
# self.noise_label = noise_label
# self.gt_noise_label = gt_noise_label
def __getitem__(self, index):
if self.mode=='train':
img, target = self.train_data[index], self.noise_label[index]
gt_target = self.gt_noise_label[index]
img = Image.fromarray(img)
img11 = self.transform(img)
img12 = self.transform(img)
img2 = self.transform_strong(img)
conf = self.conf[index]
return img11, img12, img2, target, gt_target, conf, index
elif self.mode=='all_ssl':
img, target = self.train_data[index], self.noise_label[index]
gt_target = self.gt_noise_label[index]
img = Image.fromarray(img)
img11 = self.transform(img)
img12 = self.transform(img)
img2 = self.transform_strong(img)
return img11, img12, img2, target, gt_target, index
elif self.mode=='all':
img, target = self.train_data[index], self.noise_label[index]
gt_target = self.gt_noise_label[index]
img = Image.fromarray(img)
img = self.transform(img)
return img, target, gt_target, index
elif self.mode=='all_correction':
img, target = self.train_data[index], self.noise_label[index]
gt_target = self.gt_noise_label[index]
img = Image.fromarray(img)
img = self.transform(img)
conf = self.conf[index]
return img, target, gt_target, conf, index
elif self.mode=='test':
img, target = self.test_data[index], self.test_label[index]
img = Image.fromarray(img)
img = self.transform(img)
return img, target, index
elif self.mode == 'clean_eval_train':
img, target = self.train_data[index], self.noise_label[index]
gt_target = self.gt_noise_label[index]
img = Image.fromarray(img)
img = self.transform(img)
conf = self.conf[index]
return img, target, gt_target, conf, index
def __len__(self):
if self.mode!='test':
return len(self.train_data)
else:
return len(self.test_data)
class cifar_dataloader():
def __init__(self, dataset, r, noise_mode, batch_size, num_workers, root_dir, noise_file=''):
self.dataset = dataset
self.r = r
self.noise_mode = noise_mode
self.batch_size = batch_size
self.num_workers = num_workers
self.root_dir = root_dir
self.noise_file = noise_file
if self.dataset=='cifar10':
self.transform_train = 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)),
])
self.transform_test = transforms.Compose([
transforms.ToTensor(),
transforms.Normalize((0.4914, 0.4822, 0.4465),(0.2023, 0.1994, 0.2010)),
])
self.transform_train_strong = transforms.Compose(
[
transforms.RandomCrop(32, padding=4),
transforms.RandomHorizontalFlip(),
CIFAR10Policy(),
transforms.ToTensor(),
transforms.Normalize((0.4914, 0.4822, 0.4465), (0.2023, 0.1994, 0.2010)),
]
)
elif self.dataset=='cifar100':
self.transform_train = transforms.Compose([
transforms.RandomCrop(32, padding=4),
transforms.RandomHorizontalFlip(),
transforms.ToTensor(),
transforms.Normalize((0.507, 0.487, 0.441), (0.267, 0.256, 0.276)),
])
self.transform_test = transforms.Compose([
transforms.ToTensor(),
transforms.Normalize((0.507, 0.487, 0.441), (0.267, 0.256, 0.276)),
])
self.transform_train_strong = transforms.Compose(
[
transforms.RandomCrop(32, padding=4),
transforms.RandomHorizontalFlip(),
CIFAR10Policy(),
transforms.ToTensor(),
transforms.Normalize((0.507, 0.487, 0.441), (0.267, 0.256, 0.276)),
]
)
def run(self, mode, batch_size, conf=[], conf_mask=[], lowconf_mask=[]):
if mode=='warmup':
all_dataset = cifar_dataset(dataset=self.dataset, noise_mode=self.noise_mode, r=self.r, root_dir=self.root_dir, transform=self.transform_train, mode="all", noise_file=self.noise_file)
trainloader = DataLoader(dataset=all_dataset, batch_size=batch_size, shuffle=True, num_workers=self.num_workers)
return trainloader
elif mode=='train':
if np.sum(conf_mask) > 0:
labeled_dataset = cifar_dataset(dataset=self.dataset, noise_mode=self.noise_mode, r=self.r, root_dir=self.root_dir, transform=self.transform_train, transform_strong=self.transform_train_strong, mode="train", noise_file=self.noise_file, mask=conf_mask, conf=conf)
labeled_trainloader = DataLoader(dataset=labeled_dataset, batch_size=batch_size, shuffle=True, num_workers=self.num_workers, drop_last=True)
else:
labeled_trainloader = None
if np.sum(lowconf_mask) > 0:
unlabeled_dataset = cifar_dataset(dataset=self.dataset, noise_mode=self.noise_mode, r=self.r, root_dir=self.root_dir, transform=self.transform_train, transform_strong=self.transform_train_strong, mode="train", noise_file=self.noise_file, mask=lowconf_mask, conf=conf)
unlabeled_trainloader = DataLoader(dataset=unlabeled_dataset, batch_size=batch_size, shuffle=True, num_workers=self.num_workers, drop_last=True)
else:
unlabeled_trainloader = None
return labeled_trainloader, unlabeled_trainloader
elif mode=='test':
test_dataset = cifar_dataset(dataset=self.dataset, noise_mode=self.noise_mode, r=self.r, root_dir=self.root_dir, transform=self.transform_test, mode='test')
test_loader = DataLoader(dataset=test_dataset, batch_size=batch_size, shuffle=False, num_workers=self.num_workers)
return test_loader
elif mode=='eval_train':
eval_dataset = cifar_dataset(dataset=self.dataset, noise_mode=self.noise_mode, r=self.r, root_dir=self.root_dir, transform=self.transform_test, mode='all', noise_file=self.noise_file)
eval_loader = DataLoader(dataset=eval_dataset, batch_size=batch_size, shuffle=True, num_workers=self.num_workers)
return eval_loader
elif mode=='warmup_ssl':
all_dataset = cifar_dataset(dataset=self.dataset, noise_mode=self.noise_mode, r=self.r, root_dir=self.root_dir, transform=self.transform_train, transform_strong=self.transform_train_strong, mode="all_ssl", noise_file=self.noise_file)
trainloader = DataLoader(dataset=all_dataset, batch_size=self.batch_size*2, shuffle=True, num_workers=self.num_workers)
return trainloader
elif mode=='train_correction':
corrected_dataset = cifar_dataset(dataset=self.dataset, noise_mode=self.noise_mode, r=self.r, root_dir=self.root_dir, transform=self.transform_train, transform_strong=self.transform_train_strong, mode="train", noise_file=self.noise_file, mask=conf_mask, conf=conf)
corrected_trainloader = DataLoader(dataset=corrected_dataset, batch_size=batch_size, shuffle=True, num_workers=self.num_workers, drop_last=True)
return corrected_trainloader
elif mode=='dirty_correction':
dirty_dataset = cifar_dataset(dataset=self.dataset, noise_mode=self.noise_mode, r=self.r, root_dir=self.root_dir, transform=self.transform_test, mode="all_correction", noise_file=self.noise_file, mask=conf_mask, conf=conf)
dirty_loader = DataLoader(dataset=dirty_dataset, batch_size=batch_size, shuffle=True, num_workers=self.num_workers)
return dirty_loader
elif mode=='clean_auxiliary':
clean_dataset = cifar_dataset(dataset=self.dataset, noise_mode=self.noise_mode, r=self.r, root_dir=self.root_dir, transform=self.transform_test, mode="all_correction", noise_file=self.noise_file, mask=conf_mask, conf=conf)
# freq = Counter(clean_dataset.noise_label)
# class_weight = {x : 1.0 / freq[x] for x in freq}
# sample_weights = [class_weight[x] for x in clean_dataset.noise_label]
# # sample_weights = [class_weight[x]*conf[x] for x in clean_dataset.noise_label]
# sampler = WeightedRandomSampler(sample_weights, len(clean_dataset.noise_label))
if self.dataset == 'cifar10':
class_num = 10
elif self.dataset == 'cifar100':
class_num = 100
selected_num = 1024//class_num
selected_mask = np.zeros((len(clean_dataset.noise_label),), dtype=bool)
for i in range(class_num):
idx = np.where(np.array(clean_dataset.noise_label) == i)[0]
p = np.array(conf[idx])
p = p/np.sum(p)
sampled_items = np.random.choice(idx, size=selected_num, replace=False, p=p)
selected_mask[sampled_items] = True
# print(np.unique(np.array(clean_dataset.noise_label)[selected_mask], return_counts=True))
clean_dataset = cifar_dataset(dataset=self.dataset, noise_mode=self.noise_mode, r=self.r, root_dir=self.root_dir, transform=self.transform_test, mode="all_correction", noise_file=self.noise_file, mask=selected_mask, conf=conf)
clean_loader = DataLoader(dataset=clean_dataset, batch_size=batch_size,
num_workers=self.num_workers)
return clean_loader
elif mode=='clean_eval_train':
eval_dataset = cifar_dataset(dataset=self.dataset, noise_mode=self.noise_mode, r=self.r, root_dir=self.root_dir, transform=self.transform_test, mode='clean_eval_train', noise_file=self.noise_file, conf=conf)
eval_loader = DataLoader(dataset=eval_dataset, batch_size=batch_size, shuffle=True, num_workers=self.num_workers)
return eval_loader