/
OAT.py
314 lines (280 loc) · 13.8 KB
/
OAT.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
306
307
308
309
310
311
312
313
314
import os, argparse, time
import numpy as np
import matplotlib.pyplot as plt
import torch
import torch.nn.functional as F
from torch.optim import SGD, Adam, lr_scheduler
from models.cifar10.resnet_OAT import ResNet34OAT
from models.svhn.wide_resnet_OAT import WRN16_8OAT
from models.stl10.wide_resnet_OAT import WRN40_2OAT
from dataloaders.cifar10 import cifar10_dataloaders
from dataloaders.svhn import svhn_dataloaders
from dataloaders.stl10 import stl10_dataloaders
from utils.utils import *
from utils.context import ctx_noparamgrad_and_eval
from utils.sample_lambda import element_wise_sample_lambda, batch_wise_sample_lambda
from attacks.pgd import PGD
parser = argparse.ArgumentParser(description='cifar10 Training')
parser.add_argument('--gpu', default='7')
parser.add_argument('--cpus', default=4)
# dataset:
parser.add_argument('--dataset', '--ds', default='cifar10', choices=['cifar10', 'svhn', 'stl10'], help='which dataset to use')
# optimization parameters:
parser.add_argument('--batch_size', '-b', default=128, type=int, help='mini-batch size')
parser.add_argument('--epochs', '-e', default=200, type=int, help='number of total epochs to run')
parser.add_argument('--decay_epochs', '--de', default=[50,150], nargs='+', type=int, help='milestones for multisteps lr decay')
parser.add_argument('--opt', default='sgd', choices=['sgd', 'adam'], help='which optimizer to use')
parser.add_argument('--decay', default='cos', choices=['cos', 'multisteps'], help='which lr decay method to use')
parser.add_argument('--lr', default=0.1, type=float, help='initial learning rate')
parser.add_argument('--momentum', default=0.9, type=float, help='momentum')
parser.add_argument('--wd', default=5e-4, type=float, help='weight decay')
# adv parameters:
parser.add_argument('--targeted', action='store_true', help='If true, targeted attack')
parser.add_argument('--eps', type=int, default=8)
parser.add_argument('--steps', type=int, default=7)
# OAT parameters:
parser.add_argument('--distribution', default='disc', choices=['disc'], help='Lambda distribution')
parser.add_argument('--lambda_choices', default=[0.0,0.1,0.2,0.3,0.4,1.0], nargs='*', type=float, help='possible lambda values to sample during training')
parser.add_argument('--probs', default=-1, type=float, help='the probs for sample 0, if not uniform sample')
parser.add_argument('--encoding', default='rand', choices=['none', 'onehot', 'dct', 'rand'], help='encoding scheme for Lambda')
parser.add_argument('--dim', default=128, type=int, help='encoding dimention for Lambda')
parser.add_argument('--use2BN', action='store_true', help='If true, use dual BN')
parser.add_argument('--sampling', default='ew', choices=['ew', 'bw'], help='sampling scheme for Lambda')
# others:
parser.add_argument('--resume', action='store_true', help='If true, resume from early stopped ckpt')
args = parser.parse_args()
args.efficient = True
print(args)
os.environ['CUDA_VISIBLE_DEVICES'] = args.gpu
torch.backends.cudnn.benchmark = True
# data loader:
if args.dataset == 'cifar10':
train_loader, val_loader, _ = cifar10_dataloaders(train_batch_size=args.batch_size, num_workers=args.cpus)
elif args.dataset == 'svhn':
train_loader, val_loader, _ = svhn_dataloaders(train_batch_size=args.batch_size, num_workers=args.cpus)
elif args.dataset == 'stl10':
train_loader, val_loader = stl10_dataloaders(train_batch_size=args.batch_size, num_workers=args.cpus)
# model:
if args.encoding in ['onehot', 'dct', 'rand']:
FiLM_in_channels = args.dim
else: # non encoding
FiLM_in_channels = 1
if args.dataset == 'cifar10':
model_fn = ResNet34OAT
elif args.dataset == 'svhn':
model_fn = WRN16_8OAT
elif args.dataset == 'stl10':
model_fn = WRN40_2OAT
model = model_fn(use2BN=args.use2BN, FiLM_in_channels=FiLM_in_channels).cuda()
model = torch.nn.DataParallel(model)
# for name, p in model.named_parameters():
# print(name, p.size())
# mkdirs:
model_str = os.path.join(model_fn.__name__)
if args.use2BN:
model_str += '-2BN'
if args.opt == 'sgd':
opt_str = 'e%d-b%d_sgd-lr%s-m%s-wd%s' % (args.epochs, args.batch_size, args.lr, args.momentum, args.wd)
elif args.opt == 'adam':
opt_str = 'e%d-b%d_adam-lr%s-wd%s' % (args.epochs, args.batch_size, args.lr, args.wd)
if args.decay == 'cos':
decay_str = 'cos'
elif args.decay == 'multisteps':
decay_str = 'multisteps-%s' % args.decay_epochs
attack_str = 'targeted' if args.targeted else 'untargeted' + '-pgd-%s-%d' % (args.eps, args.steps)
lambda_str = '%s-%s-%s' % (args.distribution, args.sampling, args.lambda_choices)
if args.probs > 0:
lambda_str += '-%s' % args.probs
if args.encoding in ['onehot', 'dct', 'rand']:
lambda_str += '-%s-d%s' % (args.encoding, args.dim)
save_folder = os.path.join('/hdd1/haotao/OAT_results', 'cifar10', model_str, '%s_%s_%s_%s' % (attack_str, opt_str, decay_str, lambda_str))
print(save_folder)
create_dir(save_folder)
# encoding matrix:
if args.encoding == 'onehot':
I_mat = np.eye(args.dim)
encoding_mat = I_mat
elif args.encoding == 'dct':
from scipy.fftpack import dct
dct_mat = dct(np.eye(args.dim), axis=0)
encoding_mat = dct_mat
elif args.encoding == 'rand':
rand_mat = np.random.randn(args.dim, args.dim)
np.save(os.path.join(save_folder, 'rand_mat.npy'), rand_mat)
rand_otho_mat, _ = np.linalg.qr(rand_mat)
np.save(os.path.join(save_folder, 'rand_otho_mat.npy'), rand_otho_mat)
encoding_mat = rand_otho_mat
elif args.encoding == 'none':
encoding_mat = None
# val_lambdas:
if args.distribution == 'disc':
val_lambdas = args.lambda_choices
else:
val_lambdas = [0,0.2,0.5,1]
# optimizer:
if args.opt == 'sgd':
optimizer = SGD(model.parameters(), lr=args.lr, momentum=args.momentum, weight_decay=args.wd)
elif args.opt == 'adam':
optimizer = Adam(model.parameters(), lr=args.lr, weight_decay=args.wd)
if args.decay == 'cos':
scheduler = lr_scheduler.CosineAnnealingLR(optimizer, args.epochs)
elif args.decay == 'multisteps':
scheduler = lr_scheduler.MultiStepLR(optimizer, args.decay_epochs, gamma=0.1)
# load ckpt:
if args.resume:
last_epoch, best_TA, best_ATA, training_loss, val_TA, val_ATA \
= load_ckpt(model, optimizer, scheduler, os.path.join(save_folder, 'latest.pth'))
start_epoch = last_epoch + 1
else:
start_epoch = 0
# training curve lists:
training_loss, val_TA, val_ATA, best_TA, best_ATA = [], {}, {}, {}, {}
for val_lambda in val_lambdas:
val_TA[val_lambda], val_ATA[val_lambda], best_TA[val_lambda], best_ATA[val_lambda] = [], [], 0, 0
# attacker:
attacker = PGD(eps=args.eps/255, steps=args.steps, use_FiLM=True)
## training:
for epoch in range(start_epoch, args.epochs):
train_fp = open(os.path.join(save_folder, 'train_log.txt'), 'a+')
val_fp = open(os.path.join(save_folder, 'val_log.txt'), 'a+')
start_time = time.time()
## training:
model.train()
requires_grad_(model, True)
accs, accs_adv, losses, lps = AverageMeter(), AverageMeter(), AverageMeter(), AverageMeter()
for i, (imgs, labels) in enumerate(train_loader):
imgs, labels = imgs.cuda(), labels.cuda()
# sample _lambda:
if args.sampling == 'ew':
_lambda_flat, _lambda, num_zeros = element_wise_sample_lambda(args.distribution, args.lambda_choices, encoding_mat,
batch_size=args.batch_size, probs=args.probs)
if args.use2BN:
idx2BN = num_zeros
else:
idx2BN = None
# logits for clean imgs:
logits = model(imgs, _lambda, idx2BN)
# clean loss:
lc = F.cross_entropy(logits, labels, reduction='none')
if args.efficient:
# generate adversarial images:
with ctx_noparamgrad_and_eval(model):
if args.use2BN:
imgs_adv = attacker.attack(model, imgs[num_zeros:], labels=labels[num_zeros:], _lambda=_lambda[num_zeros:], idx2BN=0)
else:
imgs_adv = attacker.attack(model, imgs[num_zeros:], labels=labels[num_zeros:], _lambda=_lambda[num_zeros:], idx2BN=None)
# logits for adv imgs:
logits_adv = model(imgs_adv.detach(), _lambda[num_zeros:], idx2BN=0)
# loss and update:
la = F.cross_entropy(logits_adv, labels[num_zeros:], reduction='none')
la = torch.cat([torch.zeros((num_zeros,)).cuda(), la], dim=0)
else:
# generate adversarial images:
with ctx_noparamgrad_and_eval(model):
imgs_adv = attacker.attack(model, imgs, labels=labels, _lambda=_lambda, idx2BN=idx2BN)
# logits for adv imgs:
logits_adv = model(imgs_adv.detach(), _lambda, idx2BN=idx2BN)
# loss and update:
la = F.cross_entropy(logits_adv, labels, reduction='none')
wc = (1-_lambda_flat)
wa = _lambda_flat
loss = torch.mean(wc * lc + wa * la)
optimizer.zero_grad()
loss.backward()
optimizer.step()
# get current lr:
current_lr = scheduler.get_lr()[0]
# metrics:
accs.append((logits.argmax(1) == labels).float().mean().item())
if args.efficient:
accs_adv.append((logits_adv.argmax(1) == labels[num_zeros:]).float().mean().item())
else:
accs_adv.append((logits_adv.argmax(1) == labels).float().mean().item())
losses.append(loss.item())
if i % 100 == 0:
train_str = 'Epoch %d-%d | Train | Loss: %.4f, TA: %.4f, ATA: %.4f' % (
epoch, i, losses.avg, accs.avg, accs_adv.avg)
print(train_str)
# print('idx2BN:', idx2BN)
train_fp.write(train_str + '\n')
# if i % 100 == 0:
# print('_lambda_flat:', _lambda_flat.size(), _lambda_flat[0:10].data.data.cpu().numpy().squeeze())
# print('_lambda:', _lambda.size(), _lambda[0:5,:].data.cpu().numpy().squeeze())
# lr schedualr update at the end of each epoch:
scheduler.step()
## validation:
model.eval()
requires_grad_(model, False)
print(model.training)
eval_this_epoch = (epoch % 10 == 0) or (epoch>=int(0.75*args.epochs)) # boolean
if eval_this_epoch:
val_accs, val_accs_adv = {}, {}
for val_lambda in val_lambdas:
val_accs[val_lambda], val_accs_adv[val_lambda] = AverageMeter(), AverageMeter()
for i, (imgs, labels) in enumerate(val_loader):
imgs, labels = imgs.cuda(), labels.cuda()
for j, val_lambda in enumerate(val_lambdas):
# sample _lambda:
if args.distribution == 'disc' and encoding_mat is not None:
_lambda = np.expand_dims( np.repeat(j, labels.size()[0]), axis=1 ).astype(np.uint8)
_lambda = encoding_mat[_lambda,:]
else:
_lambda = np.expand_dims( np.repeat(val_lambda, labels.size()[0]), axis=1 )
_lambda = torch.from_numpy(_lambda).float().cuda()
if args.use2BN:
idx2BN = int(labels.size()[0]) if val_lambda==0 else 0
else:
idx2BN = None
# TA:
logits = model(imgs, _lambda, idx2BN)
val_accs[val_lambda].append((logits.argmax(1) == labels).float().mean().item())
# ATA:
# generate adversarial images:
with ctx_noparamgrad_and_eval(model):
imgs_adv = attacker.attack(model, imgs, labels=labels, _lambda=_lambda, idx2BN=idx2BN)
linf_norms = (imgs_adv - imgs).view(imgs.size()[0], -1).norm(p=np.Inf, dim=1)
logits_adv = model(imgs_adv.detach(), _lambda, idx2BN)
val_accs_adv[val_lambda].append((logits_adv.argmax(1) == labels).float().mean().item())
val_str = 'Epoch %d | Validation | Time: %.4f | lr: %s' % (epoch, (time.time()-start_time), current_lr)
if eval_this_epoch:
val_str += ' | linf: %.4f - %.4f\n' % (torch.min(linf_norms).data, torch.max(linf_norms).data)
for val_lambda in val_lambdas:
val_str += 'val_lambda%s: TA: %.4f, ATA: %.4f\n' % (val_lambda, val_accs[val_lambda].avg, val_accs_adv[val_lambda].avg)
print(val_str)
val_fp.write(val_str + '\n')
val_fp.close() # close file pointer
# save loss curve:
training_loss.append(losses.avg)
plt.plot(training_loss)
plt.grid(True)
plt.savefig(os.path.join(save_folder, 'training_loss.png'))
plt.close()
if eval_this_epoch:
for val_lambda in val_lambdas:
val_TA[val_lambda].append(val_accs[val_lambda].avg)
plt.plot(val_TA[val_lambda], 'r')
val_ATA[val_lambda].append(val_accs_adv[val_lambda].avg)
plt.plot(val_ATA[val_lambda], 'g')
plt.grid(True)
plt.savefig(os.path.join(save_folder, 'val_acc%s.png' % val_lambda))
plt.close()
else:
for val_lambda in val_lambdas:
val_TA[val_lambda].append(val_TA[val_lambda][-1])
plt.plot(val_TA[val_lambda], 'r')
val_ATA[val_lambda].append(val_ATA[val_lambda][-1])
plt.plot(val_ATA[val_lambda], 'g')
plt.grid(True)
plt.savefig(os.path.join(save_folder, 'val_acc%s.png' % val_lambda))
plt.close()
# save pth:
if eval_this_epoch:
for val_lambda in val_lambdas:
if val_accs[val_lambda].avg >= best_TA[val_lambda]:
best_TA[val_lambda] = val_accs[val_lambda].avg # update best TA
torch.save(model.state_dict(), os.path.join(save_folder, 'best_TA%s.pth' % val_lambda))
if val_accs_adv[val_lambda].avg >= best_ATA[val_lambda]:
best_ATA[val_lambda] = val_accs_adv[val_lambda].avg # update best ATA
torch.save(model.state_dict(), os.path.join(save_folder, 'best_ATA%s.pth' % val_lambda))
save_ckpt(epoch, model, optimizer, scheduler, best_TA, best_ATA, training_loss, val_TA, val_ATA,
os.path.join(save_folder, 'latest.pth'))