-
Notifications
You must be signed in to change notification settings - Fork 1
/
train_model.py
138 lines (116 loc) · 5.16 KB
/
train_model.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
import torch
import torch.nn as nn
from torch.autograd import Variable
from tqdm import tqdm
from setup.trades import trades_loss
from setup.utils import loaddata, loadmodel, savefile
from setup.setup_pgd import to_var, pred_batch, adv_train, LinfPGDAttack, attack_over_test_data
def trainClassifier(model, train_loader, test_loader, args, use_cuda=True):
adversary = LinfPGDAttack(epsilon=args['epsilon'], k=args['num_k'], a=args['alpha'])
if use_cuda:
model = model.cuda()
optimizer = torch.optim.SGD(model.parameters(),lr=args['lr'],momentum=0.9, weight_decay=args['weight_decay'])
criterion = nn.CrossEntropyLoss()
for epoch in range(args['num_epoch']):
# trainning
ave_loss = 0
step = 0
for x, target in tqdm(train_loader):
x, target = to_var(x), to_var(target)
loss = criterion(model(x),target)
# Adversarial training or no defense training
if args['method'] == 'madry':
target_pred = pred_batch(x, model)
x_adv = adv_train(x, target_pred, model, criterion, adversary)
x_adv = to_var(x_adv)
loss = criterion(model(x_adv),target)
elif args['method'] == 'trades':
loss = trades_loss(model=model,
x_natural=x,
y=target,
optimizer=optimizer,
step_size=args['alpha'],
epsilon=args['epsilon'],
perturb_steps=args['num_k'],
beta=args['beta'],
distance='l_inf')
else:
loss = criterion(model(x),target)
ave_loss = ave_loss * 0.9 + loss.item() * 0.1
optimizer.zero_grad()
loss.backward()
optimizer.step()
step += 1
if (step + 1) % args['print_every'] == 0:
print("Epoch: [%d/%d], step: [%d/%d], Average Loss: %.4f" %
(epoch + 1, args['num_epoch'], step + 1, len(train_loader), ave_loss))
savefile(args['file_name'], model, args['dataset'])
return model
def testClassifier(test_loader, model, use_cuda=True, batch_size=100):
model.eval()
correct_cnt = 0
total_cnt = 0
for batch_idx, (x, target) in enumerate(test_loader):
if use_cuda:
x, target = x.cuda(), target.cuda()
x, target = Variable(x), Variable(target)
out = model(x)
_, pred_label = torch.max(out.data, 1)
total_cnt += x.data.size()[0]
correct_cnt += (pred_label == target.data).sum()
acc = float(correct_cnt.double()/total_cnt)
print("The prediction accuracy on testset is {}".format(acc))
return acc
def testattack(classifier, test_loader, epsilon, k, alpha, batch_size, use_cuda=True):
classifier.eval()
adversary = LinfPGDAttack(classifier, epsilon=epsilon, k=k, a=alpha)
param = {
'test_batch_size': batch_size,
'epsilon': epsilon,
}
acc = attack_over_test_data(classifier, adversary, param, test_loader, use_cuda=use_cuda)
return acc
def main(args):
use_cuda = torch.cuda.is_available()
print('==> Loading data..')
train_loader, test_loader = loaddata(args)
print('==> Loading model..')
model = loadmodel(args)
print('==> Training starts..')
model = trainClassifier(model, train_loader, test_loader, args, use_cuda=use_cuda)
testClassifier(test_loader,model,use_cuda=use_cuda,batch_size=args['batch_size'])
testattack(model, test_loader, epsilon=args['epsilon'], k=args['num_k'], alpha=args['alpha'],
batch_size=args['batch_size'], use_cuda=use_cuda)
if __name__ == "__main__":
import argparse
parser = argparse.ArgumentParser(description='Training defense models')
parser.add_argument("-d", '--dataset', choices=["mnist", "cifar10"], default="mnist")
parser.add_argument("-n", "--num_epoch", type=int, default=60)
parser.add_argument("-f", "--file_name", required=True)
parser.add_argument("-l", "--lr", type=float, default=1e-4)
parser.add_argument("--model", default="cnn", choices=["cnn", "noise"])
parser.add_argument("--method", default="no_defense", choices=["no_defense", "madry", "trades"])
parser.add_argument("--init", default=None)
parser.add_argument("--weight_decay", type=float, default=5e-4)
parser.add_argument("--root", required=True)
args = vars(parser.parse_args())
if args['dataset'] == 'mnist':
args['alpha'] = 0.02
args['num_k'] = 40
args['epsilon'] = 0.3
args['batch_size'] = 100
args['std0'] = 0.8
args['std'] = 0.4
args['print_every'] = 300
elif args['dataset'] == 'cifar10':
args['alpha'] = 0.01
args['num_k'] = 20
args['epsilon'] = 0.03
args['std0'] = 0.2
args['std'] = 0.1
args['batch_size'] = 100
args['print_every'] = 250
else:
print('invalid dataset')
print(args)
main(args)