-
Notifications
You must be signed in to change notification settings - Fork 0
/
DHBE_train.py
executable file
·270 lines (191 loc) · 9.85 KB
/
DHBE_train.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
import os
import sys
import random
import argparse
import pickle as pkl
import numpy as np
import cv2
import torch
import torch.nn.functional as F
import torch.optim as optim
import gan
import models
from dataloader import get_dataset, get_norm_trans, get_norm_trans_inv
import my_utils as utils
from backdoor.base import SimpleSubset, TriggerPastedTestDataset
import vis
def train(args, teacher, student, generator, pert_generator,
norm_trans, norm_trans_inv,
optimizer, epoch, plotter=None,
):
teacher.eval()
student.train()
generator.train()
pert_generator.train()
optimizer_S, optimizer_G, optimizer_Gp = optimizer
for i in range( args.epoch_itrs ):
for k in range(5):
z = torch.randn((args.batch_size, args.nz)).cuda()
z2 = torch.randn((args.batch_size, args.nz2)).cuda()
optimizer_S.zero_grad()
fake_data = generator(z).detach()
pert = pert_generator(z2)
pert = pert_generator.random_pad(pert).detach()
fake_img = norm_trans_inv(fake_data)
patched_img = fake_img + pert
patched_data = norm_trans(patched_img)
t_logit = teacher(fake_data)
s_logit = student(fake_data)
s_logit_pret = student(patched_data)
loss_S1 = F.l1_loss(s_logit, t_logit.detach())
loss_S2 = F.l1_loss(s_logit_pret, s_logit.detach())
loss_S = loss_S1 + loss_S2 * args.loss_weight_d1
loss_S.backward()
optimizer_S.step()
z = torch.randn((args.batch_size, args.nz)).cuda()
optimizer_G.zero_grad()
fake_data = generator(z)
t_logit = teacher(fake_data)
s_logit = student(fake_data)
loss_G1 = - F.l1_loss( s_logit, t_logit )
loss_G2 = utils.get_image_prior_losses_l1(fake_data)
loss_G3 = utils.get_image_prior_losses_l2(fake_data)
loss_G = loss_G1 + args.loss_weight_tvl1 * loss_G2 + args.loss_weight_tvl2 * loss_G3
loss_G.backward()
optimizer_G.step()
z = torch.randn((args.batch_size, args.nz)).cuda()
z2 = torch.randn((args.batch_size, args.nz2)).cuda()
optimizer_Gp.zero_grad()
fake_data = generator(z).detach()
pert = pert_generator(z2)
pert = pert_generator.random_pad(pert)
fake_img = norm_trans_inv(fake_data)
patched_img = fake_img + pert
patched_data = norm_trans(patched_img)
s_logit = student(fake_data)
s_logit_pret = student(patched_data)
loss_Gp = - F.l1_loss(s_logit, s_logit_pret)
loss_Gp.backward()
optimizer_Gp.step()
if i % args.log_interval == 0:
print('Train Epoch: {} [{}/{} ({:.0f}%)]\tG_Loss: {:.6f} S_loss: {:.6f}'.format(
epoch, i, args.epoch_itrs, 100*float(i)/float(args.epoch_itrs), loss_G.item(), loss_S.item()))
if plotter is not None:
plotter.scalar('Loss_S', (epoch-1)*args.epoch_itrs+i, loss_S.item())
plotter.scalar('Loss_S1', (epoch-1)*args.epoch_itrs+i, loss_S1.item())
plotter.scalar('Loss_S2', (epoch-1)*args.epoch_itrs+i, loss_S2.item())
plotter.scalar('Loss_G', (epoch-1)*args.epoch_itrs+i, loss_G.item())
plotter.scalar('Loss_G_l1', (epoch-1)*args.epoch_itrs+i, loss_G1.item())
plotter.scalar('Loss_G_tvl1', (epoch-1)*args.epoch_itrs+i, loss_G2.item())
plotter.scalar('Loss_G_tvl2', (epoch-1)*args.epoch_itrs+i, loss_G3.item())
plotter.scalar('Loss_Gp', (epoch-1)*args.epoch_itrs+i, loss_Gp.item())
def main():
# Training settings
parser = argparse.ArgumentParser(description='DHBE CIFAR')
parser.add_argument('--batch_size', type=int, default=256, metavar='N', help='input batch size for training (default: 256)')
parser.add_argument('--test_batch_size', type=int, default=128, metavar='N', help='input batch size for testing (default: 128)')
parser.add_argument('--epochs', type=int, default=300, metavar='N', help='number of epochs to train (default: 500)')
parser.add_argument('--epoch_itrs', type=int, default=50)
parser.add_argument('--lr_S', type=float, default=0.1, metavar='LR', help='learning rate (default: 0.1)')
parser.add_argument('--lr_G', type=float, default=1e-3, help='learning rate (default: 0.1)')
parser.add_argument('--lr_decay', type=float, default=0.1)
parser.add_argument('--weight_decay', type=float, default=5e-4)
parser.add_argument('--seed', type=int, default=1, metavar='S', help='random seed (default: 1)')
parser.add_argument('--input_dir', type=str, default="train_teacher_badnets_cifar10_resnet18_e_200_tri1_3x3_t9_0_0_n300_results")
parser.add_argument('--dataset', type=str, default='cifar10', choices=['mnist', 'svhn', 'cifar10', 'cifar100', 'vggface2_subset', 'mini-imagenet'], help='dataset name (default: mnist)')
parser.add_argument('--log-interval', type=int, default=10, metavar='N', help='how many batches to wait before logging training status')
parser.add_argument('--nz', type=int, default=256)
parser.add_argument('--loss_weight_tvl1', type=float, default=0.0)
parser.add_argument('--loss_weight_tvl2', type=float, default=0.0001)
parser.add_argument('--lr_Gp', type=float, default=1e-3, help='learning rate (default: 0.1)')
parser.add_argument('--loss_weight_d1', type=float, default=0.1)
parser.add_argument('--patch_size', type=int, default=5)
parser.add_argument('--nz2', type=int, default=256)
parser.add_argument('--vis_generator', action='store_true', default=False)
args = parser.parse_args()
args.num_classes = {"cifar10":10, "cifar100":100, "mnist":10, "vggface2_subset":100, "svhn":10, "mini-imagenet":100}.get(args.dataset, 10)
args.img_size = {"cifar10":32, "cifar100":32, "mnist":28, "vggface2_subset":64, "svhn":32, "mini-imagenet":64}.get(args.dataset, 32)
args.img_channels = {"cifar10":3, "cifar100":3, "mnist":1, "vggface2_subset":3, "svhn":3, "mini-imagenet":3}.get(args.dataset, 3)
torch.manual_seed(args.seed)
torch.cuda.manual_seed(args.seed)
np.random.seed(args.seed)
random.seed(args.seed)
torch.backends.cudnn.deterministic = True
torch.backends.cudnn.benchmark = False
print(args)
param_string = "e_{}_ps_{}_wd1_{}_lwtvl1_{}_lwtvl2_{}_lrs_{}_lrg_{}_lrgp_{}_nz_{}_nz2_{}"
param_string = param_string.format(args.epochs, args.patch_size, args.loss_weight_d1, args.loss_weight_tvl1, args.loss_weight_tvl2, args.lr_S, args.lr_G, args.lr_Gp, args.nz, args.nz2)
output_dir = os.path.join(args.input_dir, __file__.split('.')[0] + "_{}_results".format(param_string))
if not os.path.exists(output_dir):
os.mkdir(output_dir)
os.makedirs(os.path.join(output_dir, "student"), exist_ok=True)
os.makedirs(os.path.join(output_dir, "generator"), exist_ok=True)
if args.vis_generator:
os.makedirs(os.path.join(output_dir, "images"), exist_ok=True)
if args.vis_sensitivity:
os.makedirs(os.path.join(output_dir, "sensitivity"), exist_ok=True)
if args.save_checkpoint:
os.makedirs(os.path.join(output_dir, "train_stats"), exist_ok=True)
norm_trans = get_norm_trans(args)
norm_trans_inv = get_norm_trans_inv(args)
train_ds, test_ds = get_dataset(args)
trigger, target_class = utils.infer_trigger_from_path(args.input_dir, train_ds, args.img_size)
args.model = utils.infer_model_from_path(args.input_dir)
# For testing ASR and ACC
backdoored_test_ds = TriggerPastedTestDataset(test_ds, trigger, target_class=target_class)
ckpt_path = os.path.join(args.input_dir, "teacher", "{}-{}.pt".format(args.dataset, args.model))
teacher = models.get_model(args)
student = models.get_model(args)
generator = gan.GeneratorB(nz=args.nz, nc=args.img_channels, img_size=args.img_size)
pert_generator = gan.PatchGeneratorPreBN(nz=args.nz2, nc=args.img_channels, patch_size=args.patch_size, out_size=args.img_size)
teacher.load_state_dict(torch.load(ckpt_path))
print("Teacher restored from %s"%(ckpt_path))
student.load_state_dict(torch.load(ckpt_path))
print("Student restored from %s"%(ckpt_path))
teacher = teacher.cuda()
student = student.cuda()
generator = generator.cuda()
pert_generator = pert_generator.cuda()
teacher.eval()
optimizer_S = optim.SGD( student.parameters(), lr=args.lr_S, weight_decay=args.weight_decay, momentum=0.9 )
optimizer_G = optim.Adam( generator.parameters(), lr=args.lr_G )
optimizer_Gp = optim.Adam( pert_generator.parameters(), lr=args.lr_Gp )
lr_decay_steps = [0.6, 0.8]
lr_decay_steps = [int(e * args.epochs) for e in lr_decay_steps]
scheduler_S = optim.lr_scheduler.MultiStepLR(optimizer_S, lr_decay_steps, args.lr_decay)
scheduler_G = optim.lr_scheduler.MultiStepLR(optimizer_G, lr_decay_steps, args.lr_decay)
scheduler_Gp = optim.lr_scheduler.MultiStepLR(optimizer_Gp, lr_decay_steps, args.lr_decay)
plotter = vis.Plotter()
for epoch in range(1, args.epochs + 1):
# Train
if epoch == 1:
acc, asr = utils.test_model_acc_and_asr(args, student, backdoored_test_ds)
plotter.scalar("test_acc", 0, acc)
plotter.scalar("test_asr", 0, asr)
train(args, teacher=teacher, student=student, generator=generator,
pert_generator=pert_generator,
norm_trans=norm_trans, norm_trans_inv=norm_trans_inv,
optimizer=[optimizer_S, optimizer_G,
optimizer_Gp
], epoch=epoch, plotter=plotter,
# trigger=trigger
)
scheduler_S.step()
scheduler_G.step()
scheduler_Gp.step()
# Test
if args.vis_generator:
utils.test_generators(args, {'img':generator}, args.nz, epoch, output_dir, plotter, norm_trans_inv=norm_trans_inv)
utils.test_generators(args, {'pert':pert_generator}, args.nz2, epoch, output_dir, plotter, norm_trans_inv=lambda x:(x+1.0)/2.0)
acc, asr = utils.test_model_acc_and_asr(args, student, backdoored_test_ds)
print("-"*30+"\n"+"Epoch {}: acc : {:.3f}, asr : {:.3f}".format(epoch, acc, asr))
plotter.scalar("test_acc", epoch, acc)
plotter.scalar("test_asr", epoch, asr)
if epoch % 10 == 0 or epoch == args.epochs:
torch.save(student.state_dict(), os.path.join(output_dir, "student/%s-%s.pt"%(args.dataset, args.model)))
torch.save(generator.state_dict(), os.path.join(output_dir, "generator/%s-%s-generator.pt"%(args.dataset, args.model)))
torch.save(pert_generator.state_dict(), os.path.join(output_dir, "generator/%s-%s-pert_generator.pt"%(args.dataset, args.model)))
plotter.to_csv(output_dir)
plotter.to_html_report(os.path.join(output_dir, "index.html"))
if __name__ == '__main__':
main()