-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathdataset.py
164 lines (130 loc) · 6.33 KB
/
dataset.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
import numpy as np
import torch
from torch.utils.data import Dataset
from torch.utils.data import DataLoader
import logging
import os
import copy
import os.path as osp
import pickle
from collections import Counter
def _load_datafile(filename):
with open(filename, 'rb') as fo:
data_dict = pickle.load(fo, encoding='bytes')
#print(data_dict.keys())
assert data_dict[b'data'].dtype == np.uint8
image_data = data_dict[b'data']
image_data = image_data.reshape((image_data.shape[0], 3, 32, 32)).transpose(0, 2, 3, 1)
return image_data, np.array(data_dict[b'labels'])
class CIFAR_Dataset(object):
def __init__(self, root, n_labeled=1000, n_unlabeled=50000,
transform=None, seed=None, pos_labels=[0,1,8,9], neg_labels=[2,3,4,5,6,7]):
self.root = root
self.pos_labels = pos_labels
self.neg_labels = neg_labels
self.n_labeled = n_labeled
self.n_unlabeled = n_unlabeled
self.transform = transform
self.X_tr, self.Y_tr, self.X_te, self.Y_te = self.get_cifar(self.root)
def binarize(self, labels):
Y_ = - np.ones_like(labels)
for label in self.pos_labels:
Y_[labels==label] = 1
return Y_
def get_dataset(self):
X_l, y_l, X_u, y_u = self.process_pu(self.X_tr, self.Y_tr)
return SimpleDataSet(X_l, y_l, self.transform['train'], mode='L'), SimpleDataSet(X_u, y_u, self.transform['train'], mode='U'), SimpleDataSet(self.X_te, self.Y_te, self.transform['val'], mode='T'), SimpleDataSet(self.X_tr, self.Y_tr, self.transform['train'], mode='T')
def get_cifar(self, root):
train_filenames = ['data_batch_{}'.format(ii + 1) for ii in range(5)]
eval_filename = 'test_batch'
x_tr = np.zeros((50000, 32, 32, 3), dtype='uint8')
y_tr = np.zeros(50000, dtype='int32')
for ii, fname in enumerate(train_filenames):
cur_images, cur_labels = _load_datafile(osp.join(root, fname))
x_tr[ii * 10000 : (ii+1) * 10000, ...] = cur_images
y_tr[ii * 10000 : (ii+1) * 10000, ...] = cur_labels
x_te, y_te = _load_datafile(osp.join(root, eval_filename))
y_tr = self.binarize(y_tr)
y_te = self.binarize(y_te)
return x_tr, y_tr, x_te, y_te
def process_pu(self, X, Y):
rand_idx = np.arange(X.shape[0])
np.random.shuffle(rand_idx)
X_, Y_ = X[rand_idx], Y[rand_idx]
pos_idx = np.where(Y_==1)[0]
label_idx = pos_idx[:self.n_labeled]
unlabel_idx = np.concatenate([pos_idx[self.n_labeled:], np.where(Y_==-1)[0]], 0)
unlabel_idx = unlabel_idx[:self.n_unlabeled]
X_l, Y_l = X_[label_idx], Y_[label_idx]
X_u, Y_u = X_[unlabel_idx], Y_[unlabel_idx]
return X_l, Y_l, X_u, Y_u
class SimpleDataSet(Dataset):
def __init__(self, data, label, transform, mode='train'):
self.data = data
self.label = label
self.transform = transform
self.mode = mode
def __len__(self):
return self.data.shape[0]
def __getitem__(self, index):
data = self.data[index]
label = self.label[index]
if self.mode == 'L':
return self.transform(data), 1
elif self.mode == 'U':
return self.transform(data), -1
else:
return self.transform(data), label
class DebugDataSet(Dataset):
def __init__(self, ori_loader, teacher, oracle, num_clean_delete=1000, num_noisy_delete=1000):
clean_data, clean_label, noisy_data, noisy_label = self.filter_data(ori_loader, teacher, oracle)
print('In total {} clean samples, {} noisy samples'.format(clean_data.shape[0], noisy_data.shape[0]))
#assert num_delete < noisy_data.shape[0]
clean_idx = torch.randperm(clean_data.shape[0])
clean_data, clean_label = clean_data[clean_idx][:-num_clean_delete], clean_label[clean_idx][:-num_clean_delete]
noisy_idx = torch.randperm(noisy_data.shape[0])
noisy_data, noisy_label = noisy_data[noisy_idx][:-num_noisy_delete], noisy_label[noisy_idx][:-num_noisy_delete]
self.data = torch.cat([clean_data, noisy_data], 0)
self.label = torch.cat([clean_label, noisy_label], 0)
print('After deletion, {} samples in total, {} clean samples, {} noisy samples'.format(self.data.shape[0], clean_data.shape[0], noisy_data.shape[0]))
def filter_data(self, ori_loader, teacher, oracle):
clean_data, clean_label, noisy_data, noisy_label = [], [], [], []
for i, (data, label) in enumerate(ori_loader):
data, label = data.cuda(), label.cuda()
with torch.no_grad():
teacher_logits = teacher(data)
oracle_logits = oracle(data)
teacher_pred = (teacher_logits>0).float().squeeze()
oracle_pred = (oracle_logits>0).float().squeeze()
clean_index = torch.nonzero(teacher_pred==oracle_pred)[:, 0]
noisy_index = torch.nonzero(teacher_pred!=oracle_pred)[:, 0]
clean_data.append(data[clean_index].cpu())
clean_label.append(teacher_pred[clean_index].detach().cpu())
noisy_data.append(data[noisy_index].cpu())
noisy_label.append(teacher_pred[noisy_index].detach().cpu())
clean_data = torch.cat(clean_data, 0)
clean_label = torch.cat(clean_label, 0)
noisy_data = torch.cat(noisy_data, 0)
noisy_label = torch.cat(noisy_label, 0)
return clean_data, clean_label, noisy_data, noisy_label
def __len__(self):
return self.data.shape[0]
def __getitem__(self, idx):
return self.data[idx], self.label[idx]
class IndexDataset(Dataset):
def __init__(self, dataset):
self.data = dataset.data
self.label = dataset.label
self.transform = dataset.transform
self.mode = dataset.mode
def __len__(self):
return self.data.shape[0]
def __getitem__(self, index):
data = self.data[index]
label = self.label[index]
if self.mode == 'L':
return self.transform(data), 1, index
elif self.mode == 'U':
return self.transform(data), -1, index
else:
return self.transform(data), label, index