-
Notifications
You must be signed in to change notification settings - Fork 1
/
cl_scl_ST.py
executable file
·313 lines (231 loc) · 11.8 KB
/
cl_scl_ST.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
293
294
295
296
297
298
299
300
301
302
303
304
305
import torch
import torchvision
import torch.nn as nn
import torch.optim as optim
import torchvision.transforms as transforms
import torchvision.models as models
import torch.nn as nn
import torch.nn.functional as F
import numpy as np
import os
from networks import Our_ResNet
from loss import SupConLoss
from attacks import pgd_linf, pgd_linf_end2end
import argparse
def parse_option():
parser = argparse.ArgumentParser('argument for training and test')
parser.add_argument('--method', type=str, default='SimCLR',
choices=['SimCLR', 'SupCon'], help='contrastive learning methods')
parser.add_argument('--reload_encoder', type=bool, default= False, help='reloading the trained base encoder')
parser.add_argument('--reload_classifier', type=bool, default= False, help='reloading the trained linear classifier')
parser.add_argument('--batch_size', type=int, default=256,
help='batch_size')
parser.add_argument('--numEpochs', type=int, default=200,
help='number of training epochs')
parser.add_argument('--num_workers', type=int, default=4,
help='num of workers to use')
parser.add_argument('--projectionDim', type=int, default=100,help='projection dimension')
parser.add_argument('--temp', type=float, default=0.07,
help='temperature for loss function')
parser.add_argument('--learningRate', type=float, default=3e-4)
parser.add_argument('--featuresDim', type=int, default=2048)
parser.add_argument('--trial', type=int, default=0,help='id for recording runs')
parser.add_argument('--model', type=str, default='resnet50')
parser.add_argument('--dataset', type=str, default='cifar10',
choices=['cifar10', 'cifar100'], help='dataset')
parser.add_argument('--eps_t1', type=float, default=4/255, help='eps for adversarial:threat model I')
parser.add_argument('--iter_t1', type=int, default=40, help='number of iterations for generating adversarial:threat model I')
parser.add_argument('--eps_t2', type=float, default=4/255, help='eps for adversarial:threat model II')
parser.add_argument('--iter_t2', type=int, default=40, help='numer of iterations for generating adversarial:threat model II')
parser.add_argument('--alpha', type=float, default=1e-2, help= 'movement multiplier per iteration in adversarial examples')
opt = parser.parse_args()
# set the path according to the environment
opt.save_path = './save/ST/{}_models'.format(opt.dataset)
opt.model_name = '{}_{}_{}_bsz_{}_epoch_{}_trial_{}'.\
format(opt.method, opt.dataset, opt.model, opt.batch_size, opt.numEpochs, opt.trial)
if not os.path.isdir(opt.save_path):
os.makedirs(opt.save_path)
if opt.dataset == 'cifar10':
opt.n_classes = 10
elif opt.dataset == 'cifar100':
opt.n_classes = 100
else:
raise ValueError('dataset not supported: {}'.format(opt.dataset))
return opt
class TwoAugmentedTransform:
def __init__(self, transform):
self.transform = transform
def __call__(self, x):
return [self.transform(x), self.transform(x)]
def set_loader(opt):
trainCLTransform = torchvision.transforms.Compose(
[
torchvision.transforms.RandomResizedCrop(size=32),
torchvision.transforms.RandomHorizontalFlip(), # with 0.5 probability
torchvision.transforms.RandomApply([transforms.ColorJitter(0.4, 0.4, 0.4, 0.1)], p=0.8),
torchvision.transforms.RandomGrayscale(p=0.2),
torchvision.transforms.ToTensor()])
trainEvalTransform = transforms.Compose([
transforms.ToTensor()])
testTransform = transforms.Compose([
transforms.ToTensor()])
if opt.dataset == 'cifar10':
opt.n_cls = 10
trainCLDataset = torchvision.datasets.CIFAR10(root='./data/' ,train=True, transform=TwoAugmentedTransform(trainCLTransform), download=True)
trainEvalDataset = torchvision.datasets.CIFAR10(root='./data/' ,train=True, transform=trainEvalTransform, download=True)
testDataset = torchvision.datasets.CIFAR10(root='./data/' ,train=False, transform=testTransform)
elif opt.dataset == 'cifar100':
opt.n_cls = 100
trainCLDataset = torchvision.datasets.CIFAR100(root='./data/' ,train=True, transform=TwoAugmentedTransform(trainCLTransform), download=True)
trainEvalDataset = torchvision.datasets.CIFAR100(root='./data/' ,train=True, transform=trainEvalTransform, download=True)
testDataset = torchvision.datasets.CIFAR100(root='./data/' ,train=False, transform=testTransform)
else:
raise ValueError('dataset not supported: {}'.format(opt.dataset))
trainCLLoader = torch.utils.data.DataLoader(dataset=trainCLDataset, batch_size = opt.batch_size, num_workers= opt.num_workers, pin_memory=True, shuffle=True , drop_last=True)
trainEvalLoader = torch.utils.data.DataLoader(dataset=trainEvalDataset, batch_size = opt.batch_size, num_workers= opt.num_workers, pin_memory=True, shuffle=True , drop_last=True)
testLoader = torch.utils.data.DataLoader(dataset=testDataset, batch_size = opt.batch_size, num_workers= opt.num_workers, pin_memory=True, shuffle=False, drop_last=True )
return trainCLLoader,trainEvalLoader,testLoader
def set_models(opt,device):
ResNet = Our_ResNet()
Encoder = ResNet.to(device)
MLP = nn.Sequential( nn.Linear(opt.featuresDim, opt.featuresDim ),
nn.ReLU(inplace=True),
nn.Linear(opt.featuresDim, opt.projectionDim ) )
MLP = MLP.to(device)
Linear = nn.Linear(opt.featuresDim,opt.n_classes)
Linear = Linear.to(device)
class EncoderWithHead(nn.Module):
def __init__(self, encoder, head):
super(EncoderWithHead, self).__init__()
self.encoder = encoder
self.head = head
def forward(self, x):
out = F.normalize(self.head(self.encoder(x)),dim=1)
return out
CLNet = EncoderWithHead(Encoder, MLP)
EvalNet = EncoderWithHead(Encoder, Linear)
return CLNet,EvalNet
def trainCLNet(opt,trainCLLoader,CLNet,criterion,optimizer,device):
totalStep = len(trainCLLoader)
CLNet.encoder.train()
CLNet.head.train()
for epoch in range(opt.numEpochs):
for i, (X, labels) in enumerate(trainCLLoader):
x1 = X[0].to(device)
x2 = X[1].to(device)
labels = labels.to(device)
# Forward pass
z1 = CLNet(x1)
z2 = CLNet(x2)
features = torch.cat([z1.unsqueeze(1), z2.unsqueeze(1)], dim=1)
if opt.method == 'SupCon':
loss = criterion(features, labels).to(device)
elif opt.method == 'SimCLR':
loss = criterion(features).to(device)
# Backward and optimize
optimizer.zero_grad()
loss.backward()
optimizer.step()
if (i+1) % 1 == 0:
test_Accuracy = 0 #testAccuracy()
print ("Epoch [{}/{}], Step [{}/{}] Loss: {:.4f}".format(epoch+1, opt.numEpochs, i+1, totalStep, loss.item()),flush=True)
PATH = opt.save_path+'/CLNet_'+opt.model_name+'.pt'
torch.save(CLNet.state_dict(), PATH)
def trainEvalNet(opt,trainEvalLoader,EvalNet,criterion,optimizer,device):
totalStep = len(trainEvalLoader)
EvalNet.encoder.eval()
EvalNet.head.train()
for epoch in range(opt.numEpochs):
for i, (X, labels) in enumerate(trainEvalLoader):
X = X.to(device)
labels = labels.to(device)
# Forward pass
with torch.no_grad():
h = EvalNet.encoder(X)
Z = EvalNet.head(h)
loss = criterion(Z, labels)
# Backward and optimize
optimizer.zero_grad()
loss.backward()
optimizer.step()
if (i+1) % 1 == 0:
test_Accuracy = 0 #testAccuracy()
print ("Epoch [{}/{}], Step [{}/{}] Loss: {:.4f}".format(epoch+1, opt.numEpochs, i+1, totalStep, loss.item()),flush=True)
PATH = opt.save_path+'/EvalNet_'+opt.model_name+'.pt'
torch.save(EvalNet.state_dict(), PATH)
def testEvalNet(opt,testLoader,EvalNet,device):
totalStep = len(testLoader)
EvalNet.encoder.eval()
EvalNet.head.eval()
total_acc_test = 0
for i, (X, labels) in enumerate(testLoader):
X = X.to(device)
labels = labels.to(device)
h = EvalNet.encoder(X)
# Forward pass
Z = EvalNet(X)
total_acc_test += (Z.max(dim=1)[1] == labels).sum().item()
print('Acc_Test =', total_acc_test / len(testLoader.dataset),sep="\t")
return total_acc_test / len(testLoader.dataset)
def testEvalNet_adv(opt,testLoader,CLNet,EvalNet,criterion_adv,device):
totalStep = len(testLoader)
EvalNet.encoder.eval()
CLNet.encoder.eval()
EvalNet.head.eval()
CLNet.head.eval()
total_acc_test = 0
for i, (X, labels) in enumerate(testLoader):
X = X.to(device)
delta = pgd_linf(CLNet, X, opt.eps_t1, opt.alpha, opt.iter_t1, criterion_adv, labels, opt.method, device)
X_adv = (X + delta)
labels = labels.to(device)
# Forward pass
Z2 = EvalNet(X_adv)
total_acc_test += (Z2.max(dim=1)[1] == labels).sum().item()
print('Acc_Test_under_Threat Model I =', total_acc_test / len(testLoader.dataset),sep="\t")
return total_acc_test/len(testLoader.dataset)
def testEvalNet_adv_end2end(opt,testLoader,EvalNet,device):
totalStep = len(testLoader)
EvalNet.encoder.eval()
EvalNet.head.eval()
total_acc_test = 0
for i, (X, labels) in enumerate(testLoader):
X = X.to(device)
labels = labels.to(device)
delta = pgd_linf_end2end(EvalNet, X,labels, opt.eps_t2, opt.alpha, opt.iter_t2)
X_adv = (X + delta)
# Forward pass
Z2 = EvalNet(X_adv)
total_acc_test += (Z2.max(dim=1)[1] == labels).sum().item()
print('Acc_Test_under_Threat Model II =', total_acc_test / len(testLoader.dataset),sep="\t")
return total_acc_test/len(testLoader.dataset)
def main():
opt = parse_option()
trainCLLoader,trainEvalLoader,testLoader = set_loader(opt)
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
CLNet,EvalNet =set_models(opt,device)
# Representation Learning Phase
if opt.reload_encoder == True:
PATH = opt.save_path+'/CLNet_'+opt.model_name+'.pt'
CLNet.load_state_dict(torch.load(PATH))
else:
criterion = SupConLoss(temperature=opt.temp)
optimizer = torch.optim.Adam(CLNet.parameters(), lr=opt.learningRate)
trainCLNet(opt,trainCLLoader,CLNet,criterion,optimizer,device)
# Linear Classification Phase
if opt.reload_classifier == True:
PATH = opt.save_path+'/EvalNet_'+opt.model_name+'.pt'
EvalNet.load_state_dict(torch.load(PATH))
else:
criterion = nn.CrossEntropyLoss()
optimizer = torch.optim.Adam(EvalNet.head.parameters(), lr=opt.learningRate)
trainEvalNet(opt,trainEvalLoader,EvalNet,criterion,optimizer,device)
# Test on Clean data
testEvalNet(opt,testLoader,EvalNet,device)
# Test under Threat Model I
criterion_adv = SupConLoss(opt.temp)
testEvalNet_adv(opt,testLoader,CLNet,EvalNet,criterion_adv,device)
# Test under Threat Model II
testEvalNet_adv_end2end(opt,testLoader,EvalNet,device)
if __name__ == '__main__':
main()