/
main.py
558 lines (492 loc) · 19.9 KB
/
main.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
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
from torch import nn
import torch as T
from torch.nn import functional as F
from models.selector import *
from utils.util import *
from data_loader import get_train_loader, get_test_loader, set_trojai_model_id
from at import AT
import config
import sys
import copy
import json
from config import get_arguments
from hook import Hook, hookwisemse, hookwisemse_imp_sample, hookwisemse_at
from timeit import default_timer as timer
from prune import prune
from functools import partial
from tqdm import tqdm
def loss_fn_kd(outputs, labels, teacher_outputs, params):
"""
Compute the knowledge-distillation (KD) loss given outputs, labels.
"Hyperparameters": temperature and alpha
NOTE: the KL Divergence for PyTorch comparing the softmaxs of teacher
and student expects the input tensor to be log probabilities! See Issue #2
"""
alpha = 1.0
T = 2
KD_loss = nn.KLDivLoss()(F.log_softmax(outputs/T, dim=1),
F.softmax(teacher_outputs/T, dim=1)) * (alpha * T * T)
return KD_loss
def train_step(opt, train_loader, nets, optimizer, criterions, epoch):
at_losses = AverageMeter()
hk_losses = AverageMeter()
top1 = AverageMeter()
top5 = AverageMeter()
snet = nets['snet']
tnet = nets['tnet']
snet.train()
if opt.nobndiv:
for m in tnet.modules():
if isinstance(m, nn.BatchNorm2d):
m.train()
m.track_running_stats = False
#for m1, m2 in zip(snet.modules(), tnet.modules()):
# if isinstance(m1, nn.BatchNorm2d):
# m1.running_mean = T.clone(m2.running_mean) * rand (m2.running_mean,0.9, 1.1)
# m1.running_var = T.clone(m2.running_var) * rand (m2.running_var,0.9, 1.1)
#for m in snet.modules():
# if isinstance(m, nn.BatchNorm2d):
# m.eval()
# m.track_running_stats = False
criterionCls = criterions['criterionCls']
criterionAT = criterions['criterionAT']
for idx, (img, target) in tqdm(enumerate(train_loader, start=1)):
if opt.cuda:
img = img.cuda()
target = target.cuda()
img.requires_grad=True
if opt.hook:
with shook as h2:
output_s = snet(img)
with thook as h1:
output_t = tnet(img)
else:
output_s = snet(img)
output_t = tnet(img)
at_loss = T.tensor(0.).cuda()
if opt.lwf:
at_loss += loss_fn_kd(output_s, target, output_t, None) * opt.hookweight
cls_loss = criterionCls(output_s, target)
if opt.hook and epoch>=opt.clone_start and epoch<=opt.clone_end:
if opt.isample == "l2":
_cls_loss = criterionCls(output_t, target) + cls_loss
hk_loss = hookwisemse_imp_sample(h1, h2, _cls_loss)
elif opt.isample == "":
hk_loss = hookwisemse(h1, h2)
elif opt.isample == "at":
hk_loss = hookwisemse_at(h1, h2)
else:
assert False
at_loss += hk_loss * config.opt.hookweight
else:
hk_loss = T.zeros((1,))
# For comparing NAD
if float(opt.beta3)!= 0.:
at3_loss = criterionAT(snet.activation3, tnet.activation3) * opt.beta3
else:
at3_loss = 0.
if float(opt.beta2)!=0.:
at2_loss = criterionAT(snet.activation2, tnet.activation2) * opt.beta2
else:
at2_loss = 0.
if float(opt.beta1)!=0.:
at1_loss = criterionAT(snet.activation1, tnet.activation1) * opt.beta1
else:
at1_loss = 0.
at_loss += at1_loss + at2_loss + at3_loss + cls_loss
prec1, prec5 = accuracy(output_s, target, topk=(1, 5))
at_losses.update(at_loss.item(), img.size(0))
top1.update(prec1.item(), img.size(0))
top5.update(prec5.item(), img.size(0))
if opt.hook:
hk_losses.update(hk_loss.item(), img.size(0))
optimizer.zero_grad()
at_loss.backward()
if opt.rw:
for param in snet.parameters():
if param.requires_grad:
rv = T.normal(T.zeros_like(param), T.ones_like(param))
nm = T.sqrt(T.sum(T.square(rv)))
nm_p = T.sqrt(T.sum(T.square(param.grad)))
lr = optimizer.param_groups[0]["lr"]
rv = rv / nm *0.1
parrallel = T.dot(T.flatten(rv), T.flatten(param.grad)) *param.grad
rv = rv * nm_p
indp = rv - parrallel
param.grad.data = indp.detach()
optimizer.step()
if idx % opt.print_freq == 0:
print('Epoch[{0}]:[{1:03}/{2:03}] '
'AT_loss:{losses.val:.4f}({losses.avg:.4f}) '
'HK_loss:{hk_losses.val:.4f}({hk_losses.avg:.4f}) '
'prec@1:{top1.val:.2f}({top1.avg:.2f}) '
'prec@5:{top5.val:.2f}({top5.avg:.2f})'.format(epoch, idx, len(train_loader),hk_losses=hk_losses, losses=at_losses, top1=top1, top5=top5))
def test(opt, test_clean_loader, test_bad_loader, nets, criterions, epoch):
test_process = []
top1 = AverageMeter()
top5 = AverageMeter()
snet = nets['snet']
#tnet = nets['tnet']
criterionCls = criterions['criterionCls']
snet.eval()
for idx, (img, target) in enumerate(test_clean_loader, start=1):
img = img.cuda()
target = target.cuda()
with torch.no_grad():
output_s = snet(img)
prec1, prec5 = accuracy(output_s, target, topk=(1, 5))
top1.update(prec1.item(), img.size(0))
top5.update(prec5.item(), img.size(0))
acc_clean = [top1.avg, top5.avg]
cls_losses = AverageMeter()
at_losses = AverageMeter()
top1 = AverageMeter()
top5 = AverageMeter()
for idx, (img, target) in enumerate(test_bad_loader, start=1):
img = img.cuda()
img.requires_grad=True
target = target.cuda()
at_loss = T.tensor(0.).cuda()
with torch.no_grad():
output_s = snet(img)
cls_loss = criterionCls(output_s, target)
at_loss = 0.
prec1, prec5 = accuracy(output_s, target, topk=(1, 5))
cls_losses.update(cls_loss.item(), img.size(0))
at_losses.update(at_loss, img.size(0)) #.item()
top1.update(prec1.item(), img.size(0))
top5.update(prec5.item(), img.size(0))
acc_bd = [top1.avg, top5.avg, cls_losses.avg, at_losses.avg]
print('[clean]Prec@1: {:.2f}'.format(acc_clean[0]))
print('[bad]Prec@1: {:.2f}'.format(acc_bd[0]))
# save training progress
log_root = opt.log_root + '/results.csv'
test_process.append(
(epoch, acc_clean[0], acc_bd[0], acc_bd[2], acc_bd[3]))
df = pd.DataFrame(test_process, columns=(
"epoch", "test_clean_acc", "test_bad_acc", "test_bad_cls_loss", "test_bad_at_loss"))
df.to_csv(log_root, mode='a', index=False, encoding='utf-8')
return acc_clean, acc_bd
def adjust_learning_rate_medic(optimizer, epoch, lr):
# Cloning requires a bit larger learning rate than normal finetuneing
if epoch < 40:
lr = 0.01
elif epoch < 70:
lr = 0.001
elif epoch < 100:
lr = 0.0001
else:
lr = 0.0001
print('epoch: {} lr: {:.4f}'.format(epoch, lr))
for param_group in optimizer.param_groups:
param_group['lr'] = lr
def train(opt, teacher = None, student = None):
# Load models
ret = {}
print('----------- Network Initialization --------------')
if teacher is None:
teacher = select_model(dataset=opt.data_name,
model_name=opt.t_name,
pretrained=True,
pretrained_models_path=opt.t_model,
n_classes=opt.num_class).to(opt.device)
print('finished teacher model init...')
if opt.pretrain:
pretrained = True
else:
if opt.lwf or opt.hook or opt.scratch:
pretrained = False
else:
pretrained = True
if student is None:
student = select_model(dataset=opt.data_name,
model_name=opt.s_name,
pretrained=pretrained,
pretrained_models_path=opt.s_model,
n_classes=opt.num_class).to(opt.device)
print('finished student model init...')
teacher.eval()
nets = {'snet': student, 'tnet': teacher}
for param in teacher.parameters():
param.requires_grad = False
# initialize optimizer
if opt.hook:
optimizer = torch.optim.Adam(student.parameters(),
lr=opt.lr,
weight_decay=opt.weight_decay,
)
else:
optimizer = torch.optim.SGD(student.parameters(),
lr=opt.lr,
momentum=opt.momentum,
weight_decay=opt.weight_decay,
nesterov=True)
if opt.converge:
schedule=T.optim.lr_scheduler.CosineAnnealingLR(optimizer, opt.epochs)
# define loss functions
if opt.cuda:
criterionCls = nn.CrossEntropyLoss().cuda()
criterionAT = AT(opt.p)
else:
criterionCls = nn.CrossEntropyLoss()
criterionAT = AT(opt.p)
print('----------- DATA Initialization --------------')
train_loader = get_train_loader(opt)
if opt.fineprune:
prune(partial(student,prune=True),train_loader,student.setprunemask,leastk=opt.prune_num)
test_clean_loader, test_bad_loader = get_test_loader(opt)
if opt.hook and opt.keepstat:
thook.train()
for idx, (img, target) in tqdm(enumerate(train_loader, start=1)):
if opt.cuda:
img = img.cuda()
target = target.cuda()
img.requires_grad=True
if opt.hook:
with thook as h1:
output_t = teacher(img)
if idx>100:
break
thook.eval()
print('----------- Train Initialization --------------')
for epoch in range(0, opt.epochs):
if not opt.converge:
if opt.lwf:
adjust_learning_rate_medic(optimizer, epoch, opt.lr)
elif opt.hook:
adjust_learning_rate_medic(optimizer, epoch, opt.lr)
else:
adjust_learning_rate(optimizer, epoch, opt.lr)
# train every epoch
criterions = {'criterionCls': criterionCls, 'criterionAT': criterionAT}
if epoch == 0:
# before training test firstly
test(opt, test_clean_loader, test_bad_loader, nets,
criterions, epoch)
train_step(opt, train_loader, nets, optimizer, criterions, epoch+1)
# evaluate on testing set
print('testing the models......')
acc_clean, acc_bad = test(opt, test_clean_loader, test_bad_loader, nets, criterions, epoch+1)
if opt.converge:
schedule.step()
# remember best precision and save checkpoint
# save_root = opt.checkpoint_root + '/' + opt.s_name
if opt.save:
os.makedirs(opt.checkpoint_root,exist_ok=True)
is_best = acc_clean[0] > opt.threshold_clean
print("saving", is_best)
opt.threshold_clean = min(acc_bad[0], opt.threshold_clean)
best_clean_acc = acc_clean[0]
best_bad_acc = acc_bad[0]
ret["best_clean_acc"] = best_clean_acc
ret["best_bad_acc"] = best_bad_acc
save_checkpoint({
'epoch': epoch,
'state_dict': student.state_dict(),
'best_clean_acc': best_clean_acc,
'best_bad_acc': best_bad_acc,
'optimizer': optimizer.state_dict(),
}, is_best, opt.checkpoint_root, opt.s_name)
return student, ret
def finetune(opt):
ret = {}
# Load models
print('----------- Network Initialization --------------')
teacher = select_model(dataset=opt.data_name,
model_name=opt.t_name,
pretrained=True,
pretrained_models_path=opt.t_model,
n_classes=opt.num_class).to(opt.device)
print('finished teacher model init...')
student = select_model(dataset=opt.data_name,
model_name=opt.s_name,
pretrained=True,
pretrained_models_path=opt.s_model,
n_classes=opt.num_class).to(opt.device)
print('finished student model init...')
teacher.eval()
nets = {'snet': student, 'tnet': teacher}
for param in teacher.parameters():
param.requires_grad = False
# initialize optimizer
optimizer = torch.optim.SGD(student.parameters(),
lr=opt.lr,
momentum=opt.momentum,
weight_decay=opt.weight_decay,
nesterov=True)
# define loss functions
if opt.cuda:
criterionCls = nn.CrossEntropyLoss().cuda()
criterionAT = AT(opt.p)
else:
criterionCls = nn.CrossEntropyLoss()
criterionAT = AT(opt.p)
print('----------- DATA Initialization --------------')
train_loader = get_train_loader(opt)
if opt.fineprune:
prune(partial(student,prune=True),train_loader,student.setprunemask)
test_clean_loader, test_bad_loader = get_test_loader(opt)
print('----------- Train Initialization --------------')
for epoch in range(0, opt.epochs):
adjust_learning_rate(optimizer, epoch, opt.lr)
# train every epoch
criterions = {'criterionCls': criterionCls, 'criterionAT': criterionAT}
if epoch == 0:
# before training test firstly
test(opt, test_clean_loader, test_bad_loader, nets,
criterions, epoch)
train_step(opt, train_loader, nets, optimizer, criterions, epoch+1)
# evaluate on testing set
print('testing the models......')
acc_clean, acc_bad = test(opt, test_clean_loader, test_bad_loader, nets, criterions, epoch+1)
# remember best precision and save checkpoint
# save_root = opt.checkpoint_root + '/' + opt.s_name
if opt.save:
os.makedirs(opt.checkpoint_root,exist_ok=True)
is_best = acc_clean[0] > opt.threshold_clean
opt.threshold_clean = min(acc_bad[0], opt.threshold_clean)
best_clean_acc = acc_clean[0]
best_bad_acc = acc_bad[0]
ret["best_clean_acc"] = best_clean_acc
ret["best_bad_acc"] = best_bad_acc
save_checkpoint({
'epoch': epoch,
'state_dict': student.state_dict(),
'best_clean_acc': best_clean_acc,
'best_bad_acc': best_bad_acc,
'optimizer': optimizer.state_dict(),
}, is_best, opt.checkpoint_root, opt.s_name)
return student, ret
def test_only(opt, model, epoch):
test_clean_loader, test_bad_loader = get_test_loader(opt)
criterionCls = nn.CrossEntropyLoss().cuda()
criterions = {'criterionCls': criterionCls}
nets = {'snet':model}
acc_clean, acc_bad = test(opt, test_clean_loader, test_bad_loader, nets, criterions, epoch)
ret = {}
ret["best_clean_acc"] = acc_clean
ret["best_bad_acc"] = acc_bad
return ret
def main():
# Prepare arguments
opt = get_arguments().parse_args()
config.opt = opt
start = timer()
global shook, thook
shook = Hook(train=True,keepstat=opt.keepstat, nmode=opt.sbn)
thook = Hook(keepstat=opt.keepstat, nmode=opt.tbn)
shook.train()
thook.eval()
if opt.NAD:
opt1 = copy.copy(opt)
opt1.beta1 = 0
opt1.beta2 = 0
opt1.beta3 = 0
opt1.epochs = 10
opt1.save = 0
opt1.checkpoint_root += "_finetune"
config.opt = opt1
_name = config.name(opt)
student, _ = finetune(opt1)
opt1 = copy.copy(opt)
opt1.save = 1
#opt1.epochs = 30
config.opt = opt1
_name = config.name(opt1)
logdir = os.path.join("logs",opt.log_name, _name)
os.makedirs(logdir,exist_ok=True)
opt1.log_root = logdir
opt1.checkpoint_root = os.path.join("weight",opt.log_name, _name)
_, ret = train(opt1, teacher=student)
elif opt.Finetune:
config.opt = opt
opt.beta1 = 0
opt.beta2 = 0
opt.beta3 = 0
#opt.epochs = 30
_name = config.name(opt)
logdir = os.path.join("logs",opt.log_name, _name)
os.makedirs(logdir,exist_ok=True)
opt.log_root = logdir
opt.checkpoint_root = os.path.join("weight",opt.log_name, _name)
student, ret = finetune(opt)
elif opt.hook:
_name = config.name(opt)
logdir = os.path.join("logs",opt.log_name, _name)
os.makedirs(logdir,exist_ok=True)
opt.log_root = logdir
opt.checkpoint_root = os.path.join("weight",opt.log_name, _name)
_, ret = train(opt)
elif opt.scratch or opt.lwf:
_name = config.name(opt)
logdir = os.path.join("logs", opt.log_name, _name)
os.makedirs(logdir, exist_ok=True)
opt.log_root = logdir
opt.checkpoint_root = os.path.join("weight", opt.log_name, _name)
_, ret = train(opt)
elif opt.fineprune:
config.opt = opt
opt.beta1 = 0
opt.beta2 = 0
opt.beta3 = 0
#opt.epochs = 30
_name = config.name(opt)
logdir = os.path.join("logs",opt.log_name, _name)
os.makedirs(logdir,exist_ok=True)
opt.log_root = logdir
opt.checkpoint_root = os.path.join("weight",opt.log_name, _name)
student, ret = finetune(opt)
elif opt.MCR:
config.opt = opt
opt.beta1 = 0
opt.beta2 = 0
opt.beta3 = 0
model_tot = select_MCR_model(model_name=opt.t_name,n_classes=opt.num_class)
model = model_tot.net
#opt.epochs = 30
_name = config.name(opt)
logdir = os.path.join("logs",opt.log_name, _name)
os.makedirs(logdir,exist_ok=True)
opt.log_root = logdir
opt.checkpoint_root = os.path.join("weight",opt.log_name, _name)
student, ret = finetune(opt)
teacher = select_model(dataset=opt.data_name,
model_name=opt.t_name,
pretrained=True,
pretrained_models_path=opt.t_model,
n_classes=opt.num_class).to(opt.device)
model.load_state_dict(student.state_dict(),point=0)
model.load_state_dict(teacher.state_dict(),point=2)
model_tot.init_linear()
model_tot.cuda()
train_loader = get_train_loader(opt)
from MCR.train import MCR_train
from MCR.utils import update_bn
model_tot.fix_t(None)
model_tot = MCR_train(model_tot,train_loader)
model_tot.fix_t(0.1)
model_tot.train()
update_bn(train_loader, model_tot)
ret = test_only(opt, model_tot,100)
elif opt.ANP:
from ANP.interface import prune
config.opt = opt
opt.beta1 = 0
opt.beta2 = 0
opt.beta3 = 0
teacher = select_model(dataset=opt.data_name,
model_name=opt.t_name,
pretrained=True,
pretrained_models_path=opt.t_model,
n_classes=opt.num_class).to(opt.device)
train_loader = get_train_loader(opt)
prune(teacher,train_loader)
ret = test_only(opt, teacher,100)
else:
assert False
end = timer()
js = {"arg": vars(opt), "ret":ret, "time":end-start}
with open(os.path.join("logs",opt.log_name,"Exp_log.txt"),"a") as fo:
fo.write(json.dumps(js) +"\r\n")
if (__name__ == '__main__'):
main()