/
trainc_awp.py
495 lines (415 loc) · 19.3 KB
/
trainc_awp.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
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
import time, datetime, shutil, os
import numpy as np
import mlconfig
import models
import argparse
import utils
from dataset import *
from thop import profile
from trades import TradesLoss
from pgdat import PGDATLoss
from trades_awp import TradesAWP
import knowledge
import torch
import cv2
import torch.nn as nn
from torch.utils import data
import torch.optim as optim
import torch.nn.functional as F
from collections import defaultdict
from torch.cuda.amp import GradScaler, autocast
import torchattacks
from ema import EMA
mlconfig.register(VOCDataset)
mlconfig.register(CocoDataset)
mlconfig.register(FaceVOCDataset)
mlconfig.register(TradesLoss)
mlconfig.register(PGDATLoss)
def get_arguments():
"""Parse all the arguments provided from the CLI.
Returns:
A list of parsed arguments.
"""
parser = argparse.ArgumentParser(description="Training a part segmentation network.")
parser.add_argument('--config_path', type=str, default='configs')
parser.add_argument('--exp_name', type=str, default="/home/lixiao/data/partdefense_nips")
parser.add_argument('--version', type=str, default="newtrial")
parser.add_argument('--seed', type=int, default=2333)
parser.add_argument('--load_model', action='store_true', default=False)
parser.add_argument('--load_best_model', action='store_true', default=False)
parser.add_argument('--data_parallel', action='store_true', default=False)
parser.add_argument('--train', action='store_true', default=False)
parser.add_argument('--epsilon', default=1.0, type=float, help='perturbation')
parser.add_argument('--attack_choice', default='BA', choices=['BA', 'AA', 'None', "Query"])
parser.add_argument('--attack_type', default='PGD', choices=["PGD", "FGSM", "MIM", "PGD400", "SPSA", "NES"])
parser.add_argument('--attack_steps', type=int, default=40)
parser.add_argument('--dump_path', default='None', type = str)
parser.add_argument('--other_data', default=None)
return parser.parse_args()
args = get_arguments()
# Set up
if args.exp_name == '':
args.exp_name = 'exp_' + datetime.datetime.now()
exp_path = os.path.join(args.exp_name, args.version)
log_file_path = os.path.join(exp_path, args.version)
checkpoint_path = os.path.join(exp_path, 'checkpoints')
search_results_checkpoint_file_name = None
checkpoint_path_file = os.path.join(checkpoint_path, args.version)
utils.build_dirs(exp_path)
utils.build_dirs(checkpoint_path)
logger = utils.setup_logger(name=args.version, log_file=log_file_path + ".log")
torch.manual_seed(args.seed)
np.random.seed(args.seed)
if torch.cuda.is_available():
torch.backends.cudnn.enabled = True
torch.backends.cudnn.benchmark = True
device = torch.device('cuda')
device_list = [torch.cuda.get_device_name(i) for i in range(0, torch.cuda.device_count())]
logger.info("GPU List: %s" % (device_list))
else:
device = torch.device('cpu')
config_file = os.path.join(args.config_path, args.version)+'.yaml'
config = mlconfig.load(config_file)
shutil.copyfile(config_file, os.path.join(exp_path, args.version+'.yaml'))
scaler = GradScaler()
def adjust_learning_rate(optimizer, epoch):
if epoch >= 0.75 * config.epochs:
for param_group in optimizer.param_groups:
param_group['lr'] = config.learning_rate * 0.1
elif epoch >= 0.85 * config.epochs:
for param_group in optimizer.param_groups:
param_group['lr'] = config.learning_rate * 0.01
return
def train(epoch, model, optimizer, trainloader, criterion, ENV, awp_adversary, ema):
logger.info("="*20 + "Training Epoch %d" % (epoch) + "="*20)
model.train()
# interp = nn.Upsample(size=(224, 224), mode='bilinear', align_corners=True)
log_frequency = config.log_frequency if config.log_frequency is not None else 100
category_loss= 0
category_correct = 0
category_total = 0
for batch_idx, batch in enumerate(trainloader):
start = time.time()
imgs, labels, cids, names, has_part = batch
num_has_part = has_part.sum().item()
imgs, cids = imgs.to(device, non_blocking=True), cids.to(device)
optimizer.zero_grad()
if "amp" in config and config.amp:
with autocast():
if "adv_train_type" in config and epoch >= config.warm_epochs:
x_adv = criterion(model, imgs, labels, cids)
model.train()
if epoch >= config.awp_warmup:
awp = awp_adversary.calc_awp(inputs_adv=x_adv,
inputs_clean=imgs,
targets=cids,
beta=config.beta)
awp_adversary.perturb(awp)
optimizer.zero_grad()
x_natural = imgs
logits = model(x_natural)
pred = logits
adv_logits = model(x_adv)
loss_natural = criterion.base_criterion(logits, cids)
loss_robust = (1.0 / (logits.shape[0])) * criterion.criterion_kl(F.log_softmax(adv_logits, dim=1),
F.softmax(logits, dim=1))
optimizer.zero_grad()
loss = loss_natural + config.beta * loss_robust
else:
print("error")
exit(-1)
scaler.scale(loss).backward()
scaler.unscale_(optimizer)
if "grad_clip" in config:
grad_norm = torch.nn.utils.clip_grad_norm_(model.parameters(), config.grad_clip)
scaler.step(optimizer)
scaler.update()
else:
print("error")
exit(-1)
if epoch >= config.awp_warmup:
awp_adversary.restore(awp)
if "ema" in config and config.ema == True:
ema.update()
category_loss += loss.item()
_, predicted = pred.max(1)
category_total += cids.size(0)
category_correct += predicted.eq(cids).sum().item()
end = time.time()
time_used = end - start
if ENV["global_step"] % log_frequency == 0:
log_payload = {"loss": category_loss / category_total, "acc": 100.*(category_correct/category_total)}
display = utils.log_display(epoch=epoch,
global_step=ENV["global_step"],
time_elapse=time_used,
**log_payload)
logger.info(display)
ENV["global_step"] += 1
return category_correct/category_total, 100.*(category_loss/category_total)
def test(epoch, model, testloader, criterion, attacker, ENV, ema):
logger.info("="*20 + "Test Epoch %d" % (epoch) + "="*20)
if "ema" in config and config.ema == True:
ema.apply_shadow()
model.eval()
category_loss= 0
category_correct = 0
category_correct_adv = 0
category_total = 0
log_frequency = config.log_frequency if config.log_frequency is not None else 100
for batch_idx, batch in enumerate(testloader):
start = time.time()
imgs, labels, cids, names, has_part = batch
imgs, cids = imgs.to(device, non_blocking=True), cids.to(device)
with torch.no_grad():
pred = model(imgs)
loss = criterion(pred, cids)
_, predicted = pred.max(1)
category_total += cids.size(0)
category_loss += loss.item()
category_correct += predicted.eq(cids).sum().item()
if epoch >= config.adv_eval_epoch * config.epochs:
adv_imgs = attacker(imgs, cids)
with torch.no_grad():
pred = model(adv_imgs)
_, predicted = pred.max(1)
category_correct_adv += predicted.eq(cids).sum().item()
end = time.time()
time_used = end - start
if (batch_idx+1) % log_frequency == 0:
log_payload = {"category acc": 100.* (category_correct/category_total),
"adv category acc": 100.* (category_correct_adv/category_total)}
display = utils.log_display(epoch=epoch,
global_step=ENV["global_step"],
time_elapse=time_used,
**log_payload)
logger.info(display)
if "ema" in config and config.ema == True:
ema.restore()
return 100.* (category_correct/category_total), 100.* (category_correct_adv/category_total), category_loss / category_total
def query_single(model, x, ids, sigma = 0.001, q = 128, dist = "gauss"):
model.eval()
g = torch.zeros(x.shape).to(x.device)
with torch.no_grad():
for t in range(q):
if dist == "gauss":
u = torch.randn(x.shape).to(x.device)
elif dist == "Rademacher":
u = torch.empty(x.shape).uniform_(0, 1).to(x.device)
u = torch.bernoulli(u)
u = (u - 0.5) * 2
z1 = model(x + sigma * u)
z2 = model(x - sigma * u)
Jlist = []
for i in range(z1.shape[0]):
z1y = z1[i][ids[i]].clone()
z1[i][ids[i]] = -10000
J1 = torch.max(z1[i]) - z1y
z2y = z2[i][ids[i]].clone()
z2[i][ids[i]] = -10000
J2 = torch.max(z2[i]) - z2y
Jlist.append(J1 - J2)
J = torch.Tensor(Jlist).to(x.device)
J = J.unsqueeze(1).unsqueeze(2).unsqueeze(3)
g += J * u
g = g / (q * 2 * sigma)
return g
def query(model, imgs, ids, query_type):
model.eval()
adv_imgs = imgs.clone()
for i in range(40):
print("step: ", i)
if query_type == "NES":
est_grad = query_single(model, adv_imgs, ids, dist = "gauss")
elif query_type == "SPSA":
est_grad = query_single(model, adv_imgs, ids, dist = "Rademacher")
adv_imgs = adv_imgs.detach() + (args.epsilon / (8*255)) * est_grad.sign()
delta = torch.clamp(adv_imgs - imgs, min=-args.epsilon / 255, max=args.epsilon / 255)
adv_imgs = torch.clamp(imgs + delta, min=0, max=1).detach()
return adv_imgs
def evaluate_attack(model, testloader, attacker, ENV):
logger.info("="*20 + "Evaluate %s" % (args.attack_type) + "="*20)
model.eval()
for param in model.parameters():
param.requires_grad = False
category_total = 0
category_correct = 0
category_correct_adv = 0
log_frequency = config.log_frequency if config.log_frequency is not None else 100
for batch_idx, batch in enumerate(testloader):
start = time.time()
imgs, labels, cids, names, has_part = batch
imgs, cids = imgs.to(device), cids.to(device)
with torch.no_grad():
pred = model(imgs)
_, predicted = pred.max(1)
category_total += cids.size(0)
category_correct += predicted.eq(cids).sum().item()
if args.attack_choice == "BA" or args.attack_choice == "Query":
if args.attack_type == "NES":
adv_imgs = query(model, imgs, cids, args.attack_type)
elif args.attack_type == "SPSA":
adv_imgs = query(model, imgs, cids, args.attack_type)
else:
adv_imgs = attacker(imgs, cids)
if args.dump_path != "None":
if not os.path.exists(args.dump_path):
os.mkdir(args.dump_path)
if args.dump_path != "None":
s_adv_imgs = adv_imgs * 255
for idx in range(len(names)):
name = names[idx].split("/")[-1]
name = name.split(".")[0]
name = os.path.join(args.dump_path,name) + ".jpg"
cv2.imwrite(name, s_adv_imgs[idx,:,:,:].cpu().detach().numpy().transpose((1,2,0)))
with torch.no_grad():
pred = model(adv_imgs)
_, predicted = pred.max(1)
category_correct_adv += predicted.eq(cids).sum().item()
end = time.time()
time_used = end - start
if (batch_idx) % log_frequency == 0:
log_payload = {"category acc": 100.* (category_correct/category_total),
"adv category acc": 100.* (category_correct_adv/category_total)}
display = utils.log_display(epoch=0,
global_step=ENV["global_step"],
time_elapse=time_used,
**log_payload)
logger.info(display)
return 100.* (category_correct/category_total), 100.* (category_correct_adv/category_total)
def main():
model = config.model().to(device)
proxy = config.model().to(device)
if "transformer" in config:
optimizer = optim.AdamW(model.parameters(), lr=config.learning_rate, weight_decay=0.03)
else:
optimizer = optim.SGD(model.parameters(), lr=config.learning_rate, momentum=0.9,weight_decay=5e-4)
criterion = nn.CrossEntropyLoss()
# define judgement
names = ["stat_train.txt", "partlabel.txt", "categorylabel.txt"]
contents = []
for n in names:
f = open(os.path.join(config.root, config.subroot, n), "r")
s = eval(f.read())
f.close()
contents.append(s)
stat, plabels_, clabels_ = contents[0], list(contents[1]), list(contents[2])
categ2pids = defaultdict(list)
pid2name = dict()
for pid, plabel in enumerate(plabels_):
categ = plabel.split('_')[0]
categ_id = clabels_.index(categ)
categ2pids[categ_id].append(pid+1)
pid2name[pid+1] = plabel
starting_epoch = 0
config.set_immutable()
for key in config:
logger.info("%s: %s" % (key, config[key]))
logger.info("PyTorch Version: %s" % (torch.__version__))
if torch.cuda.is_available():
device_list = [torch.cuda.get_device_name(i) for i in range(0, torch.cuda.device_count())]
logger.info("GPU List: %s" % (device_list))
ENV = { 'global_step': 0,
'best_acc': 0.0,
'curren_acc': 0.0,
'best_pgd_acc': 0.0,
'train_history': [],
'eval_history': [],
'pgd_eval_history': []}
if args.load_model or args.load_best_model:
filename = checkpoint_path_file + '_best.pth' if args.load_best_model else checkpoint_path_file + '.pth'
checkpoint = utils.load_model(filename=filename,
model=model,
optimizer=optimizer,
alpha_optimizer=None,
scheduler=None)
starting_epoch = checkpoint['epoch'] + 1
ENV = checkpoint['ENV']
logger.info("File %s loaded!" % (filename))
if args.data_parallel:
print('data_parallel')
model = torch.nn.DataParallel(model).to(device)
proxy = torch.nn.DataParallel(proxy).to(device)
proxy_optim = optim.SGD(proxy.parameters(), lr=config.learning_rate)
awp_adversary = TradesAWP(model=model, proxy=proxy, proxy_optim=proxy_optim,
gamma=config.awp_gamma, base_criterion = criterion, ctype = "cls")
if "ema" in config and config.ema == True:
ema = EMA(model, 0.995)
ema.register()
else:
ema = None
trainloader = data.DataLoader(config.traindataset(root = config.root),
batch_size= config.batch_per_gpu * len(device_list), shuffle=True,
num_workers=4, pin_memory=True)
testloader = data.DataLoader(config.testdataset(root = config.root, other_data = args.other_data),
batch_size= config.batch_per_gpu * len(device_list) // 2, shuffle=True,
num_workers=4, pin_memory=True)
logger.info("Starting Epoch: %d" % (starting_epoch))
if args.train:
# define attacker
eps, alpha = config.epsilon / 255, config.epsilon / (255*8)
if "single_step" in config:
alpha = config.single_step / 255
model.eval()
attacker = torchattacks.PGD(model, eps = eps, alpha = alpha, steps=40)
for epoch in range(starting_epoch, config.epochs):
adjust_learning_rate(optimizer, epoch)
if "adv_train_type" in config and epoch >= config.warm_epochs:
train_criterion = config.adv_train_type(base_criterion = criterion)
else:
train_criterion = criterion
tc_loss, tc_acc = train(epoch, model, optimizer, trainloader, train_criterion, ENV, awp_adversary, ema)
vc_acc, vc_acc_adv, vc_loss = test(epoch, model, testloader, criterion, attacker, ENV, ema)
if "ema" in config and config.ema == True:
ema.apply_shadow()
is_best = True if vc_acc > ENV['best_acc'] else False
ENV['train_history'].append(tc_acc)
ENV['eval_history'].append(vc_acc)
ENV['best_acc'] = max(ENV['best_acc'], vc_acc)
ENV['curren_acc'] = vc_acc
if epoch >= config.adv_eval_epoch * config.epochs:
ENV['best_pgd_acc'] = max(ENV['best_pgd_acc'], vc_acc_adv)
ENV['pgd_eval_history'].append(vc_acc_adv)
logger.info('Current loss: %.2f' % (vc_loss))
logger.info('Current accuracy: %.2f' % (vc_acc))
logger.info('Current PGD classification accuracy: %.2f' % (vc_acc_adv))
target_model = model.module if args.data_parallel else model
filename = checkpoint_path_file + '.pth'
utils.save_model(ENV=ENV,
epoch=epoch,
model=target_model,
optimizer=optimizer,
alpha_optimizer=None,
scheduler=None,
genotype=None,
save_best=is_best,
filename=filename)
logger.info('Model Saved at %s\n', filename)
if "ema" in config and config.ema == True:
ema.restore()
elif args.attack_choice == 'BA' or args.attack_choice == 'None' or args.attack_choice == "Query":
# define attacker
eps, alpha = args.epsilon / 255, args.epsilon / (255*8)
model.eval()
if args.attack_type == "PGD":
attacker = torchattacks.PGD(model, eps = eps, alpha = alpha, steps=args.attack_steps)
elif args.attack_type == "PGD400":
attacker = torchattacks.PGD(model, eps = eps, alpha = alpha, steps=400)
elif args.attack_type == "MIM":
attacker = torchattacks.MIFGSM(model, eps = eps, steps = args.attack_steps, decay = 1)
elif args.attack_type == "FGSM":
attacker = torchattacks.FGSM(model, eps = eps)
else:
attacker = None
vc_acc, vc_acc_adv = evaluate_attack(model, testloader, attacker, ENV)
logger.info('Current classification accuracy: %.2f' % (vc_acc))
print(vc_acc)
logger.info('Current PGD classification accuracy: %.2f' % (vc_acc_adv))
if __name__ == "__main__":
for arg in vars(args):
logger.info("%s: %s" % (arg, getattr(args, arg)))
start = time.time()
main()
end = time.time()
cost = (end - start) / 86400
payload = "Running Cost %.2f Days" % cost
logger.info(payload)