-
Notifications
You must be signed in to change notification settings - Fork 9
/
sinkhorn.py
175 lines (130 loc) · 6.31 KB
/
sinkhorn.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
import torch
import torch.nn as nn
from data import str2dataset
from model import str2model
from utils import *
from wasserstein_attack import WassersteinAttack
from projected_sinkhorn import projected_sinkhorn
from projected_sinkhorn import wasserstein_cost
class Sinkhorn(WassersteinAttack):
def __init__(self,
predict, loss_fn,
eps, kernel_size,
lr, nb_iter, lam, sinkhorn_max_iter, stop_abs, stop_rel,
device="cuda",
postprocess=False,
verbose=True
):
super().__init__(predict=predict, loss_fn=loss_fn,
eps=eps, kernel_size=kernel_size,
device=device,
postprocess=postprocess,
verbose=verbose,
)
self.lr = lr
self.nb_iter = nb_iter
self.lam = lam
self.sinkhorn_max_iter = sinkhorn_max_iter
self.stop_abs = stop_abs
self.stop_rel = stop_rel
def perturb(self, X, y):
return self.perturb_fix(X, y)
def perturb_fix(self, X, y):
batch_size = X.size(0)
epsilon = X.new_ones(batch_size) * self.eps
C = wasserstein_cost(X, p=1, kernel_size=self.kernel_size)
normalization = X.sum(dim=(1, 2, 3), keepdim=True)
X_ = X.clone().detach().requires_grad_(True)
for t in range(self.nb_iter):
# scores = self.predict(X_)
adv_example = X_.clamp(min=self.clip_min, max=self.clip_max)
scores = self.predict(adv_example)
loss = self.loss_fn(scores, y)
loss.backward()
with torch.no_grad():
self.lst_loss.append(loss.item())
self.lst_acc.append((scores.max(dim=1)[1] == y).sum().item())
X_ += self.lr * torch.sign(X_.grad)
start = torch.cuda.Event(enable_timing=True)
end = torch.cuda.Event(enable_timing=True)
start.record()
X_, num_iter = projected_sinkhorn(X.clone().detach() / normalization,
X_.clone().detach() / normalization,
C,
epsilon,
self.lam,
verbose=False,
maxiters=self.sinkhorn_max_iter,
termination=(self.stop_abs, self.stop_rel),
return_iters=True)
if X_ is None and num_iter is None:
self.overflow = True
return X
if num_iter >= self.sinkhorn_max_iter:
self.converge = False
X_ *= normalization
end.record()
torch.cuda.synchronize()
self.run_time += start.elapsed_time(end)
self.num_iter += num_iter
self.func_calls += 1
if self.verbose and (t + 1) % 10 == 0:
print("num of iters : {:4d}, ".format(t + 1),
"loss : {:12.6f}, ".format(loss.item()),
"acc : {:5.2f}%, ".format((scores.max(dim=1)[1] == y).sum().item() / batch_size * 100),
"dual iter : {:3d}, ".format(num_iter),
"per iter time : {:7.3f}ms".format(start.elapsed_time(end) / num_iter))
X_ = X_.clone().detach().requires_grad_(True)
check_hypercube(X_, verbose=self.verbose)
return X_
if __name__ == "__main__":
import argparse
parser = argparse.ArgumentParser()
parser.add_argument('--dataset', type=str, default='MNIST')
parser.add_argument('--checkpoint', type=str, default='mnist_vanilla')
parser.add_argument('--batch_size', type=int, default=20)
parser.add_argument('--num_batch', type=int_or_none, default=5)
parser.add_argument('--eps', type=float, default=0.5, help='the perturbation size')
parser.add_argument('--kernel_size', type=int_or_none, default=5)
parser.add_argument('--lr', type=float, default=0.1, help='gradient step size')
parser.add_argument('--nb_iter', type=int, default=20)
parser.add_argument('--lam', type=float, default=1000, help='entropic regularization constanst')
parser.add_argument('--sinkhorn_max_iter', type=int, default=400)
parser.add_argument('--stop_abs', type=float, default=1e-4)
parser.add_argument('--stop_rel', type=float, default=1e-4)
parser.add_argument('--save_img_loc', type=str_or_none, default=None)
parser.add_argument('--save_info_loc', type=str_or_none, default=None)
parser.add_argument('--seed', type=int, default=0)
parser.add_argument('--postprocess', type=str2bool, default=False)
args = parser.parse_args()
print(args)
device = "cuda"
set_seed(args.seed)
testset, normalize, unnormalize = str2dataset(args.dataset)
testloader = torch.utils.data.DataLoader(testset, batch_size=args.batch_size, shuffle=False, num_workers=0)
net = str2model(args.checkpoint, dataset=args.dataset, pretrained=True).eval().to(device)
for param in net.parameters():
param.requires_grad = False
sinkhorn = Sinkhorn(predict=lambda x: net(normalize(x)),
loss_fn=nn.CrossEntropyLoss(reduction="sum"),
eps=args.eps,
kernel_size=args.kernel_size,
lr=args.lr,
nb_iter=args.nb_iter,
lam=args.lam,
sinkhorn_max_iter=args.sinkhorn_max_iter,
stop_abs=args.stop_abs,
stop_rel=args.stop_rel,
device=device,
postprocess=False,
verbose=True)
acc = test(lambda x: net(normalize(x)),
testloader,
device=device,
attacker=sinkhorn,
num_batch=args.num_batch,
save_img_loc=args.save_img_loc)
sinkhorn.print_info(acc)
# print(sinkhorn.lst_acc)
if args.save_info_loc is not None:
sinkhorn.save_info(acc, args.save_info_loc)