-
Notifications
You must be signed in to change notification settings - Fork 3
/
unseenresnet152_ilsvrc.py
208 lines (142 loc) · 7.15 KB
/
unseenresnet152_ilsvrc.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
import argparse
import numpy as np
from os import listdir
import os
os.environ['GLOG_minloglevel'] = '2'
from PIL import Image
import h5py
import caffe
from matplotlib import pyplot as plt
from random import randint, uniform
caffe.set_device(0)
caffe.set_mode_gpu()
img_crop = 224
np.random.seed(1000)
base_acc = {'CaffeNet': 0.564,
'VGG_F': 0.584,
'GoogLeNet': 0.686,
'VGG16': 0.684,
'ResNet152': 0.79}
def compute_test_accuracy(y_pred, img_labl):
top1_err = 0
top5_err = 0
for val_id in range(y_pred.shape[0]):
assert(np.isnan(np.sum(y_pred[val_id,:]))==0),"\n Nan value found in prediction labels"
# print '\n class label '+str(np.argmax(y_pred[val_id,:]))+'\t GT label '+str(img_labl[val_id,0])
if (np.argmax(y_pred[val_id, :]) != img_labl[val_id]):
# if y_pred_ori[val_id]==1:
top1_err += 1
# y_pred_ori[val_id] = 0
for val_id in range(y_pred.shape[0]):
y_sort = np.sort(y_pred[val_id, :])[::-1]
assert(np.isnan(np.sum(y_sort))==0),"\n Nan value found in sorted prediction labels"
err_flag = 0
for s_id in range(5):
if (np.nonzero(y_pred[val_id, :] == y_sort[s_id])[0][0]== img_labl[val_id]):
err_flag = 1
break
if err_flag == 0:
# if y_pred_ori[val_id]==1:
top5_err += 1
top1_acc = 1.0 -(top1_err)/float(y_pred.shape[0])
top5_acc = 1.0 -(top5_err)/float(y_pred.shape[0])
print '\n top 1 acc : '+str(top1_acc)
print '\n top 5 acc : '+str(top5_acc)
return top1_acc, top5_acc
#
def get_test_acc(net, num_test=10, use_noise=False):
# path to Imagenet validation data
val_folder_loc = args.input + '/Class_'
batch_size = 200
mini_b = 20
print '\n -------------------Testing on IMAGENET validation set ---------------------\n'
ref_test_acc = np.empty((2,2), np.float32)
y_pred = np.empty((1000*num_test, 1000), np.float32)
y_label = np.empty((1000*num_test), np.float32)
indx = 0
print '\n------------------Original--------------------------\n'
for batch_id in range(1000 / batch_size):
print '\n processing batch ' + str(batch_id)
img_data = np.empty((num_test * batch_size, 3, img_crop, img_crop), np.float32)
img_label = np.empty((num_test * batch_size), np.float32)
for class_id in range(batch_id * batch_size, (batch_id + 1) * batch_size):
filelist = listdir(str(val_folder_loc + str(class_id) + '/'))
# file_id = np.random.permutation(num_test)
for img_id in range(num_test):
# im = imread(val_folder_loc + str(class_id) + '/' + filelist[img_id])
# im = cv2.resize(im,(256,256))
im = Image.open(val_folder_loc + str(class_id) + '/' + filelist[img_id]).convert(
"RGB")
w, h = im.size
if w >= h:
im = im.resize((256 * w // h, 256))
else:
im = im.resize((256, 256 * h // w))
# im = im.resize((256,256))
img_temp = np.transpose(np.asarray(im), (2, 0, 1))
img_temp = img_temp[[2, 1, 0], :, :]
img_temp = img_temp.astype(np.float32)
img_temp = img_temp[:, (img_temp.shape[1] - img_crop) // 2:(img_temp.shape[1] + img_crop) // 2,
(img_temp.shape[2] - img_crop) // 2:(img_temp.shape[2] + img_crop) // 2]
if use_noise==True :
if args.attack == 'GAP':
adv_img = h5py.File('Attack_samples/GAP/res152_guided_GAP_epoch_10.h5', 'r')
adv_img = adv_img['v']
adv_img = 16 * adv_img[0, :]
adv_img = adv_img[[2, 1, 0], :, :]
adv_img = adv_img[:, 30:254, 30:254]
img_temp += adv_img
elif args.attack == 'GUAP':
adv_img = np.load('Attack_samples/GUAP/resnet152_with_data.npy')
# print adv_img.shape
adv_img = adv_img[0, :]
adv_img = np.transpose(adv_img, (2, 0, 1))
adv_img = 1.5 * adv_img[[2, 1, 0], :, :]
img_temp += adv_img
elif args.attack == 'sPGD':
adv_img = h5py.File('Attack_samples/sPGD/spgd_resnet152.h5', 'r')
img_temp += adv_img['v'][:]
# adv_img = h5py.File('Attack_samples/UAP/CaffeNet/Linf_'+str(xi)+'/caffenet_adv_iter_L'+str(xi)+'_'+str(np.random.randint(1,5))+'.h5','r')
# # adv_img = h5py.File('/home/tejas/universal/matlab/caffenet_adv_iter' + str(np.random.randint(1, 11)) + '.h5', 'r')
# adv_img = np.transpose(adv_img['v'][:], (0,2,1))
# img_temp += adv_img
img_data[num_test * (class_id - batch_size * batch_id) + img_id, :, :, :] = img_temp.copy()
img_label[num_test * (class_id - batch_size * batch_id) + img_id] = class_id
indx+=1
img_data[:, 0, :, :] -= 102.98
img_data[:, 1, :, :] -= 115.947
img_data[:, 2, :, :] -= 122.772
for bsz in range(batch_size * num_test / mini_b):
batch_in = img_data[bsz * mini_b:(bsz + 1) * mini_b, :]
net.blobs['data'].reshape(*batch_in.shape)
net.blobs['data'].data[...] = batch_in
net.forward()
y_pred[
batch_id * num_test * batch_size + bsz * mini_b:batch_id * num_test * batch_size + (bsz + 1) * mini_b, :] = \
net.blobs['prob'].data
y_label[batch_size * num_test * batch_id:(batch_id + 1) * num_test * batch_size] = img_label
del img_data, img_label
ref_test_acc[0, 0], ref_test_acc[0, 1] = compute_test_accuracy(y_pred, y_label)
# out_file.create_dataset('dc_vgg', data=y_pred)
del y_pred, y_label
return ref_test_acc
parser = argparse.ArgumentParser()
parser.add_argument('--input',help='path to the ILSVRC2012 validation set root folder')
parser.add_argument('--load', help='path to trained caffemodel and weights')
parser.add_argument('--attack', help='Select unseen attack', default='GAP',choices=['GAP','GUAP','sPGD'])
parser.add_argument('--defense', default='True', choices=['True', 'False'], help='Switch between baseline (no defense) and our proposed defense (FRU)')
args = parser.parse_args()
if not args.load:
print('Error: Model weights not provided')
else:
if args.defense=='True':
model_proto = 'Prototxt/ResNet152/deploy_resnet152_FRU.prototxt'
else:
model_proto = 'Prototxt/ResNet152/deploy_resnet152.prototxt'
net = caffe.Classifier(model_proto, args.load, caffe.TEST)
print('\n Accuracy on original images')
clean_acc = get_test_acc(net,num_test=50,use_noise=False)
print('\n Accuracy on adversarial images')
adv_acc = get_test_acc(net, num_test=50,use_noise=True)
print('\n Restoration rate: '+str((float(clean_acc[0,0])+float(adv_acc[0,0]))
/(2*float(base_acc['ResNet152']))))