-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathdataset.py
200 lines (169 loc) · 7.06 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
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
import copy
import numpy as np
import torch
from PIL import Image
from torch.utils.data.dataset import Dataset
class PoisonLabelDataset(Dataset):
"""Poison-Label dataset wrapper.
Args:
dataset (Dataset): The dataset to be wrapped.
transform (callable): The backdoor transformations.
poison_idx (np.array): An 0/1 (clean/poisoned) array with
shape `(len(dataset), )`.
target_label (int): The target label.
"""
def __init__(self, dataset, transform, poison_idx, target_label):
super(PoisonLabelDataset, self).__init__()
self.dataset = copy.deepcopy(dataset)
self.train = self.dataset.train
if self.train:
self.data = self.dataset.data
self.targets = self.dataset.targets
self.poison_idx = poison_idx
else:
# Only fetch poison data when testing.
self.data = self.dataset.data[np.nonzero(poison_idx)[0]]
self.targets = self.dataset.targets[np.nonzero(poison_idx)[0]]
self.poison_idx = poison_idx[poison_idx == 1]
self.pre_transform = self.dataset.pre_transform
self.primary_transform = self.dataset.primary_transform
self.remaining_transform = self.dataset.remaining_transform
self.prefetch = self.dataset.prefetch
if self.prefetch:
self.mean, self.std = self.dataset.mean, self.dataset.std
self.bd_transform = transform
self.target_label = target_label
def __getitem__(self, index):
if isinstance(self.data[index], str):
with open(self.data[index], "rb") as f:
img = np.array(Image.open(f).convert("RGB"))
else:
img = self.data[index]
target = self.targets[index]
poison = 0
origin = target # original target
if self.poison_idx[index] == 1:
img = self.bd_first_augment(img, bd_transform=self.bd_transform)
target = self.target_label
poison = 1
else:
img = self.bd_first_augment(img, bd_transform=None)
item = {"img": img, "target": target, "poison": poison, "origin": origin}
return item
def __len__(self):
return len(self.data)
def bd_first_augment(self, img, bd_transform=None):
# Pre-processing transformation (HWC ndarray->HWC ndarray).
img = Image.fromarray(img)
img = self.pre_transform(img)
img = np.array(img)
# Backdoor transformation (HWC ndarray->HWC ndarray).
if bd_transform is not None:
img = bd_transform(img)
# Primary and the remaining transformations (HWC ndarray->CHW tensor).
img = Image.fromarray(img)
img = self.primary_transform(img)
img = self.remaining_transform(img)
if self.prefetch:
# HWC ndarray->CHW tensor with C=3.
img = np.rollaxis(np.array(img, dtype=np.uint8), 2)
img = torch.from_numpy(img)
return img
class MixMatchDataset(Dataset):
"""Semi-supervised MixMatch dataset.
Args:
dataset (Dataset): The dataset to be wrapped.
semi_idx (np.array): An 0/1 (labeled/unlabeled) array with shape ``(len(dataset), )``.
labeled (bool): If True, creates dataset from labeled set, otherwise creates from unlabeled
set (default: True).
"""
def __init__(self, dataset, semi_idx, labeled=True):
super(MixMatchDataset, self).__init__()
self.dataset = copy.deepcopy(dataset)
if labeled:
self.semi_indice = np.nonzero(semi_idx == 1)[0]
else:
self.semi_indice = np.nonzero(semi_idx == 0)[0]
self.labeled = labeled
self.prefetch = self.dataset.prefetch
self.mean, self.std = self.dataset.mean, self.dataset.std
def __getitem__(self, index):
if self.labeled:
item = self.dataset[self.semi_indice[index]]
item["labeled"] = True
else:
item1 = self.dataset[self.semi_indice[index]]
item2 = self.dataset[self.semi_indice[index]]
img1, img2 = item1.pop("img"), item2.pop("img")
item1.update({"img1": img1, "img2": img2})
item = item1
item["labeled"] = False
return item
def __len__(self):
return len(self.semi_indice)
class SelfPoisonDataset(Dataset):
"""Self-supervised poison-label contrastive dataset.
Args:
dataset (PoisonLabelDataset): The poison-label dataset to be wrapped.
transform (dict): Augmented transformation dict has three keys `pre`, `primary`
and `remaining` which corresponds to pre-processing, primary and the
remaining transformations.
"""
def __init__(self, dataset, transform):
super(SelfPoisonDataset, self).__init__()
self.dataset = copy.deepcopy(dataset)
self.data = self.dataset.data
self.targets = self.dataset.targets
self.poison_idx = self.dataset.poison_idx
self.bd_transform = self.dataset.bd_transform
self.target_label = self.dataset.target_label
self.pre_transform = transform["pre"]
self.primary_transform = transform["primary"]
self.remaining_transform = self.dataset.remaining_transform
self.prefetch = self.dataset.prefetch
if self.prefetch:
self.mean, self.std = self.dataset.mean, self.dataset.std
def __getitem__(self, index):
if isinstance(self.data[index], str):
with open(self.data[index], "rb") as f:
img = np.array(Image.open(f).convert("RGB"))
else:
img = self.data[index]
target = self.targets[index]
poison = 0
origin = target # original target
if self.poison_idx[index] == 1:
img1 = self.bd_first_augment(img, bd_transform=self.bd_transform)
img2 = self.bd_first_augment(img, bd_transform=self.bd_transform)
target = self.target_label
poison = 1
else:
img1 = self.bd_first_augment(img, bd_transform=None)
img2 = self.bd_first_augment(img, bd_transform=None)
item = {
"img1": img1,
"img2": img2,
"target": target,
"poison": poison,
"origin": origin,
}
return item
def __len__(self):
return len(self.data)
def bd_first_augment(self, img, bd_transform=None):
# Pre-processing transformations (HWC ndarray->HWC ndarray).
img = Image.fromarray(img)
img = self.pre_transform(img)
img = np.array(img)
# Backdoor transformationss (HWC ndarray->HWC ndarray).
if bd_transform is not None:
img = bd_transform(img)
# Primary and the remaining transformations (HWC ndarray->CHW tensor).
img = Image.fromarray(img)
img = self.primary_transform(img)
img = self.remaining_transform(img)
if self.prefetch:
# HWC ndarray->CHW tensor with C=3.
img = np.rollaxis(np.array(img, dtype=np.uint8), 2)
img = torch.from_numpy(img)
return img