-
Notifications
You must be signed in to change notification settings - Fork 4
/
main_raw.py
173 lines (139 loc) · 5.3 KB
/
main_raw.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
# -*- coding:utf-8 -*-
# Created Time: Thu 05 Jul 2018 10:00:41 PM CST
# Author: Taihong Xiao <[email protected]>
import torch
import torchvision
from torch.autograd import Variable
from torch.nn.parameter import Parameter
import torch.nn.functional as F
from torchvision import transforms
import numpy as np
import os
import argparse
train_dir = 'train_log'
cuda = True
batch_size = 30
w1, h1 = 224, 224
w2, h2 = 28, 28
lmd = 5e-5
lr = 0.005
decay = 0.99
max_epoch = 30
restore = False
if not os.path.exists(train_dir):
os.makedirs(train_dir)
def imagenet_label2_mnist_label(imagenet_label):
return imagenet_label[:,:10]
def tensor2var(tensor, requires_grad=False, cuda=False, volatile=False):
if cuda:
with torch.cuda.device(0):
tensor = tensor.cuda()
var = Variable(tensor, requires_grad=requires_grad, volatile=volatile)
return var
def generator(dataloader):
while True:
for data in dataloader:
yield data
transform = transforms.Compose([
# you can add other transformations in this list
transforms.ToTensor()
])
resnet50 = torchvision.models.resnet50(pretrained=False)
resnet50.load_state_dict(torch.load('./models/resnet50-19c8e357.pth'))
resnet50.eval()
kwargs = {'num_workers': 1, 'pin_memory': True, 'drop_last': True}
train_set = torchvision.datasets.MNIST('./datasets/mnist/', train=True, transform=transform, download=True)
test_set = torchvision.datasets.MNIST('./datasets/mnist/', train=False, transform=transform, download=True)
train_loader = torch.utils.data.DataLoader(train_set, batch_size=batch_size, shuffle=True, **kwargs)
test_loader = torch.utils.data.DataLoader(test_set, batch_size=batch_size, shuffle=True, **kwargs)
# train_generator = generator(train_loader)
# test_generator = generator(test_loader)
# test image read
# from PIL import Image
# import time
# t1 = time.time()
# img1 = np.asanyarray(Image.open('dog.jpg').resize((224,224)))
# img1 = np.transpose(img1, (2,0,1)) / 255.
# t2 = time.time()
# import cv2
# img = cv2.cvtColor(cv2.imread('cat1.jpg'), cv2.COLOR_BGR2RGB)
# img = cv2.resize(img, (224,224))
# img = np.transpose(img, (2,0,1)) / 255.
# mean = np.array([0.485, 0.456, 0.406])
# mean = mean[..., np.newaxis, np.newaxis]
# std = np.array([0.229, 0.224, 0.225])
# std = std[..., np.newaxis, np.newaxis]
# img = (img - mean) / std
# img = img[np.newaxis, ...].astype(np.float32)
# x = tensor2var(torch.from_numpy(img))
# out = resnet50(x)
# print(np.argmax(out.data.numpy()))
# from IPython import embed; embed();
# mean and std for input
mean = np.array([0.485, 0.456, 0.406], dtype=np.float32)
mean = mean[..., np.newaxis, np.newaxis]
std = np.array([0.229, 0.224, 0.225],dtype=np.float32)
std = std[..., np.newaxis, np.newaxis]
mean = tensor2var(torch.from_numpy(mean), cuda=cuda)
std = tensor2var(torch.from_numpy(std), cuda=cuda)
# create mask M
M = np.ones((3, h1, w1), dtype=np.float32)
c_w, c_h = int(np.ceil(w1/2.)), int(np.ceil(h1/2.))
M[:,c_h-h2//2:c_h+h2, c_w-w2//2:c_w+w2//2] = 0
M = tensor2var(torch.from_numpy(M), cuda=cuda)
# Learnable parameter W
if restore:
W = torch.load(os.path.join(train_dir, 'W_{:03d}.pt'.format(restore))).data
else:
W = torch.randn(M.shape)
W = tensor2var(W, requires_grad=True, cuda=cuda)
# optimizer
BCE = torch.nn.BCELoss()
optimizer = torch.optim.Adam([W], lr=lr, betas=(0.5, 0.999))
# optimizer = torch.optim.SGD([W], lr=lr)
lr_scheduler = torch.optim.lr_scheduler.StepLR(optimizer, step_size=100, gamma=decay)
if cuda:
with torch.cuda.device(0):
BCE.cuda()
resnet50.cuda()
# start training
for i in range(max_epoch):
for j, (image, label) in enumerate(train_loader):
lr_scheduler.step()
image = np.tile(image, (1,3,1,1))
label = torch.zeros(batch_size, 10).scatter_(1, label.view(-1,1), 1)
label = tensor2var(label, cuda=cuda)
X = np.zeros((batch_size, 3, h1, w1), dtype=np.float32)
X[:,:,(h1-h2)//2:(h1+h2)//2, (w1-w2)//2:(w1+w2)//2] = image
X = tensor2var(torch.from_numpy(X), cuda=cuda)
P = torch.sigmoid(W * M)
X_adv = X + P # range [0, 1]
X_adv = (X_adv - mean) / std
Y_adv = resnet50(X_adv)
Y_adv = F.softmax(Y_adv, 1)
out = imagenet_label2_mnist_label(Y_adv)
loss = BCE(out, label) + lmd * torch.norm(W) ** 2
optimizer.zero_grad()
loss.backward()
optimizer.step()
print('epoch %03d/%03d, batch %06d, loss %.6f' % (i + 1, max_epoch, j + 1, loss.data.cpu().numpy()))
# test
acc = 0.0
for j, (image, label) in enumerate(test_loader):
image = np.tile(image, (1,3,1,1))
# label = torch.zeros(batch_size, 10).scatter_(1, label.view(-1,1), 1)
# label = tensor2var(label)
X = torch.zeros(batch_size, 3, h1, w1)
X[:,:,(h1-h2)//2:(h1+h2)//2, (w1-w2)//2:(w1+w2)//2] = torch.from_numpy(image)
X = tensor2var(X)
if cuda:
with torch.cuda.device(0):
X = X.cuda()
P = torch.sigmoid(W * M)
X_adv = X + P # range [0, 1]
Y_adv = resnet50(X_adv)
Y_adv = F.softmax(Y_adv, 1)
out = imagenet_label2_mnist_label(Y_adv)
pred = out.data.cpu().numpy().argmax(1)
acc += sum(label.numpy() == pred) / float(len(label) * len(test_loader))
print('epoch %03d/%03d, test accuracy %.6f' % (i, max_epoch, acc))