-
Notifications
You must be signed in to change notification settings - Fork 2
/
train.py
241 lines (194 loc) · 10.4 KB
/
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
import numpy as np
import os, sys
import pickle
import argparse
import time
import torch
import torch.nn as nn
import torch.backends.cudnn as cudnn
import torchvision.transforms as trn
import torchvision.datasets as dset
import torch.nn.functional as F
from wrn import WideResNet
from utils import get_measures, print_measures
parser = argparse.ArgumentParser(description='Watermarking for OOD Detection', formatter_class=argparse.ArgumentDefaultsHelpFormatter)
# Experimental Setup
parser.add_argument('dataset', type=str, choices=['cifar10', 'cifar100'])
parser.add_argument('--score', type=str, default='MSP', choices=['MSP', 'energy'])
# Hyerparameter
parser.add_argument('--epochs', type=int, default=50, help='epochs')
parser.add_argument('--batch_size', type=int, default=128, help='batch size.')
parser.add_argument('--test_bs', type=int, default=200, help='batch size.')
parser.add_argument('--alpha', type=float, default=0.0, help='learning rate.')
parser.add_argument('--sigma', type=float, default=0.0, help='covaraince.')
parser.add_argument('--rho', type=float, default=0.0, help='constraint.')
parser.add_argument('--beta', type=float, default=0.0, help='trade-off.')
parser.add_argument('--T_in', type=float, default=0.0, help='Scale for energy ID.')
parser.add_argument('--T_out', type=float, default=0.0, help='Scale for energy OOD.')
parser.add_argument('--T', default=1., type=float, help='temperature: energy')
# WRN Architecture
parser.add_argument('--layers', default=40, type=int)
parser.add_argument('--widen-factor', default=2, type=int)
parser.add_argument('--droprate', default=0.3, type=float)
# Others
parser.add_argument('--load', type=str, default='./ckpt')
parser.add_argument('--prefetch', type=int, default=4)
parser.add_argument('--out_as_pos', action = 'store_true')
parser.add_argument('--num_to_avg', type=int, default=1)
args = parser.parse_args()
# pre defined hyperparameter
if args.alpha == 0 and args.sigma == 0 and args.rho == 0 and args.beta == 0:
print('loading predefined hyper parameters')
args.alpha = 0.01
if args.dataset == 'cifar10' and args.score == 'MSP':
args.sigma = 0.4
args.rho = 1.0
args.beta = 3.5
if args.dataset == 'cifar100' and args.score == 'MSP':
args.sigma = 1.0
args.rho = 0.2
args.beta = 2.5
if args.dataset == 'cifar10' and args.score == 'energy':
args.sigma = 1.5
args.rho = .05
args.beta = 0.1
args.T_in = 0.2
args.T_out = 0.7
if args.dataset == 'cifar100' and args.score == 'energy':
args.sigma = 0.4
args.rho = 0.1
args.beta = 0.1
args.T_in = 0.9
args.T_out = 0.1
cudnn.benchmark = True
mean = [x / 255 for x in [125.3, 123.0, 113.9]]
std = [x / 255 for x in [63.0, 62.1, 66.7]]
train_transform = trn.Compose([trn.RandomHorizontalFlip(), trn.ToTensor(), trn.Normalize(mean, std)])
test_transform = trn.Compose([trn.ToTensor(), trn.Normalize(mean, std)])
if args.dataset == 'cifar10':
train_data = dset.CIFAR10('../data/cifarpy', train=True, transform=train_transform, download=True)
test_data = dset.CIFAR10('../data/cifarpy', train=False, transform=test_transform)
num_classes = 10
else:
train_data = dset.CIFAR100('../data/cifarpy', train=True, transform=train_transform, download=True)
test_data = dset.CIFAR100('../data/cifarpy', train=False, transform=test_transform)
num_classes = 100
train_loader = torch.utils.data.DataLoader(train_data, batch_size=args.batch_size, shuffle=True, num_workers=args.prefetch, pin_memory=False)
test_loader = torch.utils.data.DataLoader(test_data, batch_size=args.batch_size, shuffle=False, num_workers=args.prefetch, pin_memory=False)
texture_data = dset.ImageFolder(root="../watermarking_final/data/dtd/images", transform=trn.Compose([trn.Resize(32), trn.CenterCrop(32), trn.ToTensor(), trn.Normalize(mean, std)]))
svhn_data = dset.SVHN(root='../data/svhn/', split="test", transform=trn.Compose( [trn.ToTensor(), trn.Normalize(mean, std)]), download = True)
places365_data = dset.ImageFolder(root="../data/places365_standard/", transform=trn.Compose([trn.Resize(32), trn.CenterCrop(32), trn.ToTensor(), trn.Normalize(mean, std)]))
# lsunc_data = dset.ImageFolder(root="../data/LSUN", transform=trn.Compose([trn.Resize(32), trn.ToTensor(), trn.Normalize(mean, std)]))
lsunr_data = dset.ImageFolder(root="../data/LSUN_resize", transform=trn.Compose([trn.Resize(32), trn.ToTensor(), trn.Normalize(mean, std)]))
isun_data = dset.ImageFolder(root="../data/iSUN", transform=trn.Compose([trn.ToTensor(), trn.Normalize(mean, std)]))
texture_loader = torch.utils.data.DataLoader(texture_data, batch_size=args.test_bs, shuffle=True, num_workers=1, pin_memory=False)
svhn_loader = torch.utils.data.DataLoader(svhn_data, batch_size=args.test_bs, shuffle=True, num_workers=1, pin_memory=False)
places365_loader = torch.utils.data.DataLoader(places365_data, batch_size=args.test_bs, shuffle=True, num_workers=1, pin_memory=False)
# lsunc_loader = torch.utils.data.DataLoader(lsunc_data, batch_size=args.test_bs, shuffle=True, num_workers=1, pin_memory=False)
lsunr_loader = torch.utils.data.DataLoader(lsunr_data, batch_size=args.test_bs, shuffle=True, num_workers=1, pin_memory=False)
isun_loader = torch.utils.data.DataLoader(isun_data, batch_size=args.test_bs, shuffle=True, num_workers=1, pin_memory=False)
ood_num_examples = len(test_data) // 5
concat = lambda x: np.concatenate(x, axis=0)
to_np = lambda x: x.data.cpu().numpy()
def get_ood_scores(loader, watermark = None, in_dist=False):
_score = []
with torch.no_grad():
for batch_idx, (data, target) in enumerate(loader):
if batch_idx >= ood_num_examples // args.test_bs and in_dist is False:
break
data, target = data.cuda(), target.cuda()
if watermark is not None:
data = data + watermark.cuda()
output = net(data)
smax = to_np(F.softmax(output, dim=1))
if args.score == 'energy':
_score.append(-to_np((args.T*torch.logsumexp(output / args.T, dim=1))))
else:
_score.append(-np.max(smax, axis=1))
if in_dist:
return concat(_score).copy()
else:
return concat(_score)[:ood_num_examples].copy()
def get_and_print_results(ood_loader, in_score, watermark = None, num_to_avg=args.num_to_avg):
aurocs, auprs, fprs = [], [], []
for _ in range(num_to_avg):
out_score = get_ood_scores(ood_loader, watermark)
if args.out_as_pos:
measures = get_measures(out_score, in_score)
else:
measures = get_measures(-in_score, -out_score)
aurocs.append(measures[0]); auprs.append(measures[1]); fprs.append(measures[2])
auroc = np.mean(aurocs); aupr = np.mean(auprs); fpr = np.mean(fprs)
print_measures(auroc, aupr, fpr, '')
return auroc, aupr, fpr
def cal_loss(x, in_set):
loss = F.cross_entropy(x[:len(in_set[0])], in_set[1])
if args.score == 'energy':
T_out, T_in = args.T_out, args.T_in
Ec_out = (x[len(in_set[0]):] * T_out).exp().mean(1)
Ec_in = (-x[:len(in_set[0])] * T_in).exp().mean(1)
loss += args.beta * (Ec_out.mean() + Ec_in.mean())
else:
loss_oe = -(x[len(in_set[0]):].mean(1) - torch.logsumexp(x[len(in_set[0]):], dim=1))
loss += args.beta * loss_oe.mean()
return loss
def train(epoch, watermark, alpha):
loss_avg = 0.0
for batch_idx, in_set in enumerate(train_loader):
in_set[0], in_set[1] = in_set[0].cuda(), in_set[1].cuda()
out_set = torch.ones_like(in_set[0]) * 0 + torch.randn_like(in_set[0]) * args.sigma
data = torch.cat((in_set[0], out_set), 0)
# [STAGE 1] for perturbation
watermark.requires_grad_()
optim.zero_grad()
logits = net(data + watermark)
loss = cal_loss(logits, in_set)
grad = torch.autograd.grad(loss, [watermark])[0].detach()
perturb = grad.sign() * grad.abs() / (grad ** 2).sum().sqrt()
# [STAGE 2] for updating
watermark.requires_grad_()
optim.zero_grad()
logits = net(data + watermark + perturb * args.rho)
loss = cal_loss(logits, in_set)
grad = torch.autograd.grad(loss, [watermark])[0].detach()
watermark = watermark.detach() - alpha * torch.sign(grad)
# LOGS
loss_avg = loss_avg * 0.8 + float(loss) * 0.2
sys.stdout.write('\r E%d [%d / %d] loss %.2f ' %(epoch, batch_idx + 1, len(train_loader), loss_avg))
return watermark
net = WideResNet(args.layers, num_classes, args.widen_factor, dropRate=args.droprate).cuda()
optim = torch.optim.SGD(net.parameters(), lr = 1)
watermark = torch.zeros(1,3,32,32).cuda()
# Restore model
if args.dataset == 'cifar10':
model_path = args.load + '/cifar10_wrn_pretrained_epoch_99.pt'
else:
model_path = args.load + '/cifar100_wrn_pretrained_epoch_99.pt'
net.load_state_dict(torch.load(model_path))
net.eval()
print(' Score: ' + args.score)
print(' FPR95 AUROC AUPR')
in_score = get_ood_scores(test_loader, in_dist=True)
metric_ll = []
metric_ll.append(get_and_print_results(isun_loader, in_score))
metric_ll.append(get_and_print_results(places365_loader, in_score))
metric_ll.append(get_and_print_results(texture_loader, in_score))
metric_ll.append(get_and_print_results(svhn_loader, in_score))
metric_ll.append(get_and_print_results(lsunr_loader, in_score))
avg_r = torch.Tensor(metric_ll).mean(0)
print_measures(avg_r[0], avg_r[1], avg_r[2], '')
alpha = args.alpha
for epoch in range(0, args.epochs + 1):
if epoch == 25: alpha /= 10
watermark = train(epoch, watermark, alpha)
if epoch % 5 == 0:
print('\n FPR95 AUROC AUPR')
in_score = get_ood_scores(test_loader, watermark, in_dist=True)
metric_ll = []
metric_ll.append(get_and_print_results(isun_loader, in_score, watermark))
metric_ll.append(get_and_print_results(places365_loader, in_score, watermark))
metric_ll.append(get_and_print_results(texture_loader, in_score, watermark))
metric_ll.append(get_and_print_results(svhn_loader, in_score, watermark))
metric_ll.append(get_and_print_results(lsunr_loader, in_score, watermark))
avg_r = torch.Tensor(metric_ll).mean(0)
print_measures(avg_r[0], avg_r[1], avg_r[2], '')