-
Notifications
You must be signed in to change notification settings - Fork 2
/
main_adv.py
282 lines (225 loc) · 12.6 KB
/
main_adv.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
'''
Adversarial Training
'''
import os
import sys
from numpy.core.numeric import outer
import torch
import pickle
import argparse
import torch.optim
import torch.nn as nn
import torch.utils.data
import matplotlib.pyplot as plt
import torchvision.models as models
from utils import *
from sparselearning.core import Masking, CosineDecay
from sparselearning.pruning_utils import check_sparsity
parser = argparse.ArgumentParser(description='PyTorch Adversarial Sparse Training')
########################## data setting ##########################
parser.add_argument('--data', type=str, default='data/cifar10', help='location of the data corpus', required=True)
parser.add_argument('--dataset', type=str, default='cifar10', help='dataset [cifar10, cifar100, tinyimagenet]', required=True)
########################## model setting ##########################
parser.add_argument('--arch', type=str, default='resnet18', help='model architecture [resnet18, wideresnet, vgg16]', required=True)
parser.add_argument('--depth_factor', default=34, type=int, help='depth-factor of wideresnet')
parser.add_argument('--width_factor', default=10, type=int, help='width-factor of wideresnet')
########################## basic setting ##########################
parser.add_argument('--seed', default=1, type=int, help='random seed')
parser.add_argument('--gpu', type=int, default=0, help='gpu device id')
parser.add_argument('--resume', action="store_true", help="resume from checkpoint")
parser.add_argument('--resume_dir', help='The directory resume the trained models', default=None, type=str)
parser.add_argument('--pretrained', default=None, type=str, help='pretrained model')
parser.add_argument('--eval', action="store_true", help="evaluation pretrained model")
parser.add_argument('--print_freq', default=50, type=int, help='logging frequency during training')
parser.add_argument('--save_dir', help='The parent directory used to save the trained models', default=None, type=str)
########################## training setting ##########################
parser.add_argument('--batch_size', type=int, default=128, help='batch size')
parser.add_argument('--lr', default=0.1, type=float, help='initial learning rate')
parser.add_argument('--decreasing_lr', default='100,150', help='decreasing strategy')
parser.add_argument('--momentum', default=0.9, type=float, help='momentum')
parser.add_argument('--weight_decay', default=5e-4, type=float, help='weight decay')
parser.add_argument('--epochs', default=200, type=int, help='number of total epochs to run')
########################## attack setting ##########################
parser.add_argument('--norm', default='linf', type=str, help='linf or l2')
parser.add_argument('--train_eps', default=8, type=float, help='epsilon of attack during training')
parser.add_argument('--train_step', default=10, type=int, help='itertion number of attack during training')
parser.add_argument('--train_gamma', default=2, type=float, help='step size of attack during training')
parser.add_argument('--train_randinit', action='store_false', help='randinit usage flag (default: on)')
parser.add_argument('--test_eps', default=8, type=float, help='epsilon of attack during testing')
parser.add_argument('--test_step', default=20, type=int, help='itertion number of attack during testing')
parser.add_argument('--test_gamma', default=2, type=float, help='step size of attack during testing')
parser.add_argument('--test_randinit', action='store_false', help='randinit usage flag (default: on)')
########################## Flying Bird setting ##########################
parser.add_argument('--fb', action='store_true', help='Enable flying bird mode. Default: True.')
parser.add_argument('--fix', action='store_true', help='Fix topology during training. Default: True.')
parser.add_argument('--sparse_init', type=str, default='igq', help='sparse initialization')
parser.add_argument('--reset', action='store_true', help='Fix topology during training. Default: True.')
parser.add_argument('--growth', type=str, default='random', help='Growth mode. Choose from: momentum, random, and momentum_neuron.')
parser.add_argument('--death', type=str, default='magnitude', help='Death mode / pruning mode. Choose from: magnitude, SET, threshold, CS_death.')
parser.add_argument('--redistribution', type=str, default='none', help='Redistribution mode. Choose from: momentum, magnitude, nonzeros, or none.')
parser.add_argument('--death-rate', type=float, default=0.50, help='The pruning rate / death rate.')
parser.add_argument('--density', type=float, default=0.2, help='The density of the overall sparse network.')
parser.add_argument('--dynamic_epoch', default=100, type=int, help='The dynamic sparse start epoch')
parser.add_argument('--decay-schedule', type=str, default='cosine', help='The decay schedule for the pruning rate. Default: cosine. Choose from: cosine, linear.')
parser.add_argument('--update_frequency', type=int, default=900, metavar='N', help='how many iterations to train between mask update')
########################## Flying Bird+ sparse ##########################
parser.add_argument('--fbp', action='store_true', help='Enable fling bird+ mode. Default: True.')
parser.add_argument('--epoch_range', type=int, default=4, help='epoch range to decide sparse action')
parser.add_argument('--prune_ratio', type=float, default=0.4, help='The ratio of dynamic prune ')
parser.add_argument('--growth_ratio', type=float, default=0.05, help='The ratio of dynamic growth ')
parser.add_argument('--update_threshold', type=float, default=0.5, help='The update threshold of dynamic prune or growth')
########################## Dynamic frequency #############################
parser.add_argument('--dynamic_fre', action='store_true', help='Enable dynamic frequency mode. Default: True.')
parser.add_argument('--second_frequency', type=int, default=1200, metavar='N', help='how many iterations to train between mask update in second stage')
def main():
args = parser.parse_args()
args.train_eps = args.train_eps / 255
args.train_gamma = args.train_gamma / 255
args.test_eps = args.test_eps / 255
args.test_gamma = args.test_gamma / 255
save_path = get_save_path(args)
os.makedirs(save_path, exist_ok=True)
setup_logger(args)
print_args(args)
print_and_log(args)
torch.cuda.set_device(int(args.gpu))
if args.seed:
print('set random seed = ', args.seed)
setup_seed(args.seed)
train_loader, val_loader, test_loader, model = setup_dataset_models(args)
model.cuda()
########################## optimizer and scheduler ##########################
criterion = nn.CrossEntropyLoss()
optimizer = torch.optim.SGD(model.parameters(), args.lr,
momentum=args.momentum,
weight_decay=args.weight_decay)
scheduler = torch.optim.lr_scheduler.MultiStepLR(optimizer, milestones=[int(lr) for lr in args.decreasing_lr.split(',')], last_epoch=-1)
########################## sparse mask ####################################
mask = None
if args.fb or args.fbp:
decay = CosineDecay(args.death_rate, len(train_loader) * (args.epochs))
mask = Masking(optimizer, death_rate=args.death_rate, death_mode=args.death, death_rate_decay=decay, growth_mode=args.growth,
redistribution_mode=args.redistribution, args=args)
mask.add_module(model, sparse_init=args.sparse_init, density=args.density)
# mask.set_dst_start_epoch(0.5*args.epochs)
mask.set_dst_start_epoch(args.dynamic_epoch)
######################### only evaluation ###################################
if args.eval:
assert args.pretrained
pretrained_model = torch.load(args.pretrained, map_location = torch.device('cuda:'+str(args.gpu)))
print_and_log('loading from state_dict')
if 'state_dict' in pretrained_model.keys():
pretrained_model = pretrained_model['state_dict']
model.load_state_dict(pretrained_model)
test(test_loader, model, criterion, args)
test_adv(test_loader, model, criterion, args)
return
########################## resume ##########################
start_epoch = 0
if args.resume:
print_and_log('resume from checkpoint.pth.tar')
checkpoint = torch.load(os.path.join(args.resume_dir, 'checkpoint.pth.tar'), map_location = torch.device('cuda:'+str(args.gpu)))
best_sa = checkpoint['best_sa']
best_ra = checkpoint['best_ra']
best_test_sa = checkpoint['best_test_sa']
best_test_ra = checkpoint['best_test_ra']
start_epoch = checkpoint['epoch']
model.load_state_dict(checkpoint['state_dict'])
optimizer.load_state_dict(checkpoint['optimizer'])
scheduler.load_state_dict(checkpoint['scheduler'])
all_result = checkpoint['result']
if mask:
mask.load_info_resume(checkpoint['mask_info'])
set_ite_step(checkpoint['iter_step'])
else:
all_result = {}
all_result['train_acc'] = []
all_result['val_sa'] = []
all_result['val_ra'] = []
all_result['test_sa'] = []
all_result['test_ra'] = []
all_result['g_norm'] = []
all_result['rc_ratio'] = []
all_result['grad_cs'] = []
all_result['sparsity'] = 1.0
all_result['total_fired_weights'] = 0
all_result['best_ra_epoch'] = 0
best_sa = 0
best_ra = 0
best_test_sa = 0
best_test_ra = 0
is_sa_best = False
is_ra_best = False
is_test_sa_best = False
is_test_ra_best = False
########################## training process ##########################
for epoch in range(start_epoch, args.epochs):
current_sparsity = check_sparsity(model)
print_and_log(optimizer.state_dict()['param_groups'][0]['lr'])
if mask:
mask.set_dst_current_epoch(epoch)
print_and_log('baseline adversarial training')
train_acc = train_epoch_adv(train_loader, model, criterion, optimizer, epoch, args, mask)
all_result['train_acc'].append(train_acc)
scheduler.step()
all_result['sparsity'] = current_sparsity
###validation###
val_sa = test(val_loader, model, criterion, args)
val_ra, val_loss = test_adv(val_loader, model, criterion, args)
test_sa = test(test_loader, model, criterion, args)
test_ra, _= test_adv(test_loader, model, criterion, args)
if args.fbp:
mask.update_loss_info(val_loss)
mask.update_train_val_diff(train_acc, val_ra)
all_result['val_sa'].append(val_sa)
all_result['val_ra'].append(val_ra)
all_result['test_sa'].append(test_sa)
all_result['test_ra'].append(test_ra)
is_sa_best = val_sa > best_sa
best_sa = max(val_sa, best_sa)
is_ra_best = val_ra > best_ra
best_ra = max(val_ra, best_ra)
if is_ra_best:
all_result['best_ra_epoch'] = epoch
checkpoint_state = {
'best_sa': best_sa,
'best_ra': best_ra,
'epoch': epoch+1,
'state_dict': model.state_dict(),
'optimizer': optimizer.state_dict(),
'scheduler': scheduler.state_dict(),
'iter_step': get_ite_step(),
'result': all_result
}
if mask:
checkpoint_state.update({
'mask_info': mask.get_info_resume()
})
if args.fbp:
prune_epochs, growth_epochs = mask.get_dst_epochs()
all_result['prune_epochs'] = prune_epochs
all_result['growth_epochs'] = growth_epochs
if mask:
all_result['total_fired_weights'] = mask.total_fired_weights
checkpoint_state.update({
'result': all_result
})
save_checkpoint(checkpoint_state, is_sa_best, is_ra_best, save_path)
if epoch == args.epochs - 1:
getinfo(checkpoint_state)
get_generalization_gap(model, criterion, args)
plt.plot(all_result['train_acc'], label='train_acc')
plt.plot(all_result['test_sa'], label='SA')
plt.plot(all_result['test_ra'], label='RA')
if args.fbp :
prune_epochs, growth_epochs = mask.get_dst_epochs()
delta = 0.4
for epoch in prune_epochs:
plt.axvline(epoch - delta, linewidth = 0.8, color = 'black', linestyle='--')
for epoch in growth_epochs:
plt.axvline(epoch + delta, linewidth = 0.8, color = 'red', linestyle='--')
plt.legend()
plt.savefig(os.path.join(save_path, 'net_train.png'))
plt.close()
if __name__ == '__main__':
main()