-
Notifications
You must be signed in to change notification settings - Fork 8
/
Copy pathdatasets.py
180 lines (145 loc) · 8.61 KB
/
datasets.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
import numpy as np
from os.path import isfile
import torchvision, torch, copy
import torch
from torch.utils import data
from pathlib import Path
from utils import get_targeted_classes
def get_labels(dataset):
dataloader = torch.utils.data.DataLoader(dataset, batch_size=1, shuffle=False, num_workers=2)
labels = np.zeros(len(dataset), dtype=int)
num, max_val = 0, -100000
print('==> Getting label array..')
for images, targets in dataloader:
maxx_img_val = torch.max(images)
max_val = max(max_val, maxx_img_val)
labels[num] = targets.item()
num+=1
return labels, max_val
def load_dataset(dataset, root='../data/'):
# Step 1: Load Transformations and Normalizations
if dataset in ['CIFAR10','CIFAR100']:
train_augment = [torchvision.transforms.RandomCrop(32, padding=4), torchvision.transforms.RandomHorizontalFlip()]
test_augment = []
mean, std = (0.4914, 0.4822, 0.4465), (0.2470, 0.2435, 0.2616)
elif dataset in ['PCAM']:
train_augment = [torchvision.transforms.CenterCrop(32), torchvision.transforms.RandomCrop(32, padding=4), torchvision.transforms.RandomHorizontalFlip()]
test_augment = [torchvision.transforms.CenterCrop(32)]
mean, std = (0.5, 0.5, 0.5), (0.5, 0.5, 0.5)
elif dataset in ['LFWPeople', 'CelebA', 'DermNet', 'Pneumonia']:
train_augment = [torchvision.transforms.RandomResizedCrop(224), torchvision.transforms.RandomHorizontalFlip()]
test_augment = [torchvision.transforms.Resize(256), torchvision.transforms.CenterCrop(224)]
mean, std = (0.485, 0.456, 0.406), (0.229, 0.224, 0.225)
train_transforms = torchvision.transforms.Compose(train_augment + [torchvision.transforms.ToTensor(), torchvision.transforms.Normalize(mean, std)])
test_transforms = torchvision.transforms.Compose(test_augment + [torchvision.transforms.ToTensor(), torchvision.transforms.Normalize(mean, std)])
# Step 2: Load Train, Test and Evaluation Train Sets
if dataset in ['CIFAR10','CIFAR100']:
train_set = getattr(torchvision.datasets, dataset)(root=root, train=True, download=True, transform=train_transforms)
test_set = getattr(torchvision.datasets, dataset)(root=root, train=False, download=True, transform=test_transforms)
eval_train_set = getattr(torchvision.datasets, dataset)(root=root, train=True, download=True, transform=test_transforms)
elif dataset in ['PCAM', 'LFWPeople', 'CelebA']:
train_set = getattr(torchvision.datasets, dataset)(root=root, split='train', download=True, transform=train_transforms)
test_set = getattr(torchvision.datasets, dataset)(root=root, split='test', download=True, transform=test_transforms)
eval_train_set = getattr(torchvision.datasets, dataset)(root=root, split='train', download=True, transform=test_transforms)
elif dataset in ['DermNet','Pneumonia']:
train_set = torchvision.datasets.ImageFolder(root=root+'/'+dataset+'/train', transform=train_transforms)
test_set = torchvision.datasets.ImageFolder(root=root+'/'+dataset+'/test', transform=test_transforms)
eval_train_set = torchvision.datasets.ImageFolder(root=root+'/'+dataset+'/train', transform=test_transforms)
# If found, cache values of labels and max_val else compute them
if isfile(root+'/'+dataset+'_labels.npy'):
train_labels = np.load(root+'/'+dataset+'_labels.npy')
max_val = np.load(root+'/'+dataset+'_maxval.npy')
else:
train_labels, max_val = get_labels(eval_train_set)
np.save(root+'/'+dataset+'_labels.npy', train_labels)
np.save(root+'/'+dataset+'_maxval.npy', max_val)
return train_set, eval_train_set, test_set, train_labels, max_val
def manip_dataset(dataset, train_labels, method, manip_set_size, save_dir='../saved_models'):
assert(method in ['randomlabelswap', 'interclasslabelswap', 'poisoning'])
manip_idx_path = save_dir+'/'+dataset+'_'+method+'_'+str(manip_set_size)+'_manip.npy'
if method == 'randomlabelswap' or method == 'poisoning': # Shuffle labels of a selected subset of samples
if isfile(manip_idx_path):
manip_idx = np.load(manip_idx_path)
else:
manip_idx = np.random.choice(len(train_labels), manip_set_size, replace=False)
p = Path(save_dir)
p.mkdir(exist_ok=True)
np.save(manip_idx_path, manip_idx)
idxes_in_manipidx = copy.deepcopy(manip_idx)
idxes_in_manipidx.sort()
manip_dict = {}
for i in range(len(idxes_in_manipidx)):
manip_dict[idxes_in_manipidx[i]] = train_labels[manip_idx[i]] if method == 'randomlabelswap' else 0
elif method == 'interclasslabelswap':
classes = get_targeted_classes(dataset)
if isfile(manip_idx_path):
manip_idx = np.load(manip_idx_path)
else:
assert(manip_set_size%2==0)
idx1 = np.asarray(train_labels==classes[0]).nonzero()[0][:manip_set_size//2]
idx2 = np.asarray(train_labels==classes[1]).nonzero()[0][:manip_set_size//2]
manip_idx = np.concatenate([idx1, idx2])
p = Path(save_dir)
p.mkdir(exist_ok=True)
np.save(manip_idx_path, manip_idx)
manip_dict = {}
for i in range(len(manip_idx)):
if i < manip_set_size//2:
manip_dict[manip_idx[i]] = classes[1]
else:
manip_dict[manip_idx[i]] = classes[0]
full_idx = np.arange(len(train_labels))
untouched_idx = np.setdiff1d(full_idx, manip_idx)
manip_idx, untouched_idx = torch.from_numpy(manip_idx), torch.from_numpy(untouched_idx)
return manip_dict, manip_idx, untouched_idx
def get_deletion_set(deletion_size, manip_dict, train_size, dataset, method, save_dir='../saved_models'):
delete_idx_path = save_dir+'/'+dataset+'_'+method+'_'+str(len(manip_dict))+'_'+str(deletion_size)+'_deletion.npy'
if isfile(delete_idx_path):
delete_idx = np.load(delete_idx_path)
else:
delete_idx = np.random.choice(np.array(list(manip_dict.keys())), deletion_size, replace=False)
p = Path(save_dir)
p.mkdir(exist_ok=True)
np.save(delete_idx_path, delete_idx)
full_idx = np.arange(train_size)
retain_idx = np.setdiff1d(full_idx, delete_idx)
delete_idx, retain_idx = torch.from_numpy(delete_idx), torch.from_numpy(retain_idx)
return delete_idx, retain_idx
class DatasetWrapper(data.Dataset):
def __init__(self, dataset, manip_dict, mode='pretrain', corrupt_val=None, corrupt_size=3, delete_idx=None):
self.dataset = dataset
self.manip_dict = manip_dict
self.mode = mode
if corrupt_val is not None: corrupt_val = torch.from_numpy(corrupt_val)
self.corrupt_val = corrupt_val
self.corrupt_size = corrupt_size
self.delete_idx = delete_idx
assert(mode in ['pretrain', 'unlearn', 'manip', 'test', 'test_adversarial'])
def __getitem__(self, index):
image, label = self.dataset.__getitem__(index)
if self.mode == 'pretrain':
if int(index) in self.manip_dict: # Do nasty things while selecting samples from the manip set
label = self.manip_dict[int(index)]
if self.corrupt_val is not None:
image[:,-self.corrupt_size:,-self.corrupt_size:] = self.corrupt_val # Have the bottom right corner of the image as the poison
if self.delete_idx is None:
self.delete_idx = torch.tensor(list(self.manip_dict.keys()))
indel = int(index in self.delete_idx)
if self.mode in ['test', 'test_adversarial']:
if self.mode == 'test_adversarial':
image[:,-self.corrupt_size:,-self.corrupt_size:] = self.corrupt_val
return image, label
else:
return image, label, indel
def __len__(self):
return len(self.dataset)
if __name__ == "__main__":
train_set, test_set, mean, std, train_labels = load_dataset(dataset='CIFAR10', root='../data/')
manip_dict = manip_dataset(dataset='CIFAR10', train_labels=train_labels, method='randomlabelswap', manip_set_size=10000, save_dir='../saved_models')
#print(manip_dict)
train_set = DatasetWrapper(train_set, manip_dict, mode='pretrain')
badt_set = DatasetWrapper(train_set, manip_dict, mode='badt')
train_set, test_set, mean, std, train_labels = load_dataset(dataset='CIFAR100', root='../data/')
manip_dict = manip_dataset(dataset='CIFAR100', train_labels=train_labels, method='randomlabelswap', manip_set_size=1000, save_dir='../saved_models')
#print(manip_dict)
train_set = DatasetWrapper(train_set, manip_dict, mode='pretrain')