-
Notifications
You must be signed in to change notification settings - Fork 2
/
Copy pathmain_attack.py
336 lines (291 loc) · 17.2 KB
/
main_attack.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
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
# -*- coding: utf-8 -*-
"""
Created on Wed Mar 13 10:50:25 2019
@author: Keshik
"""
import argparse
import torch
import numpy as np
from torchvision import transforms
import torchvision.models as models
from torch.utils.data import DataLoader
from dataset import PascalVOC_Dataset, CocoDetection, CutoutPIL
from randaugment import RandAugment
import torch.optim as optim
from train import train_model, test
from attack import tkmlap
from baseline_attacks import baselineap
from utils import encode_labels, plot_history
import os
import torch.utils.model_zoo as model_zoo
import utils
from models.inception import Inception3
os.environ['TORCH_HOME'] = '.'
# os.environ["CUDA_VISIBLE_DEVICES"] = '1'
def main(args):
"""
Main function
Args:
data_dir: directory to download Pascal VOC data
model_name: resnet18, resnet34 or resnet50
num: model_num for file management purposes (can be any postive integer. Your results stored will have this number as suffix)
lr: initial learning rate list [lr for resnet_backbone, lr for resnet_fc]
epochs: number of training epochs
batch_size: batch size. Default=16
download_data: Boolean. If true will download the entire 2012 pascal VOC data as tar to the specified data_dir.
Set this to True only the first time you run it, and then set to False. Default False
save_results: Store results (boolean). Default False
Returns:
test-time loss and average precision
Example way of running this function:
if __name__ == '__main__':
main('../data/', "resnet34", num=1, lr = [1.5e-4, 5e-2], epochs = 15, batch_size=16, download_data=False, save_results=True)
"""
data_dir = args.data
model_name = args.arch
num = args.num
lr = args.lr
epochs = args.epochs
batch_size = args.batch_size
download_data = args.download_data
save_results = args.save_results
model_dir = os.path.join(args.results, args.arch)
if not os.path.isdir(model_dir):
os.makedirs(model_dir)
model_urls = {
'resnet18': 'https://download.pytorch.org/models/resnet18-5c106cde.pth',
'resnet34': 'https://download.pytorch.org/models/resnet34-333f7ec4.pth',
'resnet50': 'https://download.pytorch.org/models/resnet50-19c8e357.pth',
'inception_v3': 'https://download.pytorch.org/models/inception_v3_google-1a9a5a14.pth'
}
model_collections_dict = {
"resnet18": models.resnet18(),
"resnet34": models.resnet34(),
"resnet50": models.resnet50(),
"inception_v3": models.inception_v3()
}
# Initialize cuda parameters
use_cuda = torch.cuda.is_available()
np.random.seed(2019)
torch.manual_seed(2019)
device = torch.device("cuda" if use_cuda else "cpu")
# device = torch.device("cpu")
print("Available device = ", device)
if model_name in ['resnet18', 'resnet34', 'resnet50', 'inception_v3']:
model = model_collections_dict[model_name]
model.avgpool = torch.nn.AdaptiveAvgPool2d(1)
model.load_state_dict(model_zoo.load_url(model_urls[model_name]))
num_ftrs = model.fc.in_features
model.fc = torch.nn.Linear(num_ftrs, args.num_classes)
else:
model = Inception3(num_classes=args.num_classes)
model.to(device)
if args.opt == 'SGD':
optimizer = optim.SGD([
{'params': list(model.parameters())[:-1], 'lr': lr[0], 'momentum': 0.9},
{'params': list(model.parameters())[-1], 'lr': lr[1], 'momentum': 0.9}
])
elif args.opt == 'Adam':
optimizer = optim.Adam([
{'params': list(model.parameters())[:-1], 'lr': lr[0], 'momentum': 0.9},
{'params': list(model.parameters())[-1], 'lr': lr[1], 'momentum': 0.9}
])
scheduler = optim.lr_scheduler.CosineAnnealingLR(optimizer, 12, eta_min=0, last_epoch=-1)
# Imagnet values
# mean=[0.457342265910642, 0.4387686270106377, 0.4073427106250871]
# std=[0.26753769276329037, 0.2638145880487105, 0.2776826934044154]
if args.normalize == 'mean_std':
mean = [0.457342265910642, 0.4387686270106377, 0.4073427106250871]
std = [0.26753769276329037, 0.2638145880487105, 0.2776826934044154]
elif args.normalize == 'boxmaxmin':
if args.boxmax == 1 and args.boxmin == 0:
mean = [0, 0, 0]
std = [1.0, 1.0, 1.0]
elif args.boxmax == -(args.boxmin):
mean = [0.5, 0.5, 0.5]
std = [0.5 / args.boxmax, 0.5 / args.boxmax, 0.5 / args.boxmax]
else:
return
if args.dataset =='COCO':
# COCO DataLoader
instances_path_val = os.path.join(args.data, 'annotations/instances_val2014.json')
instances_path_train = os.path.join(args.data, 'annotations/instances_train2014.json')
# data_path_val = args.data
# data_path_train = args.data
data_path_val = f'{args.data}/val2014' # args.data
data_path_train = f'{args.data}/train2014' # args.data
COCO_val_dataset = CocoDetection(data_path_val,
instances_path_val,
transforms.Compose([
transforms.Resize((args.image_size, args.image_size)),
transforms.ToTensor(),
transforms.Normalize(mean=mean, std=std),
# normalize, # no need, toTensor does normalization
]))
COCO_train_dataset = CocoDetection(data_path_train,
instances_path_train,
transforms.Compose([
transforms.Resize((args.image_size, args.image_size)),
CutoutPIL(cutout_factor=0.5),
RandAugment(),
transforms.ToTensor(),
transforms.Normalize(mean=mean, std=std),
# normalize,
]))
print('Using COCO dataset')
print("COCO len(val_dataset)): ", len(COCO_val_dataset))
print("COCO len(train_dataset)): ", len(COCO_train_dataset))
train_loader = torch.utils.data.DataLoader(
COCO_train_dataset, batch_size=args.batch_size, shuffle=True,
num_workers=args.workers, pin_memory=True)
# valid_loader = torch.utils.data.DataLoader(
# COCO_val_dataset, batch_size=args.batch_size, shuffle=False,
# num_workers=args.workers, pin_memory=False)
valid_loader = torch.utils.data.DataLoader(
COCO_val_dataset, batch_size=args.batch_size,
num_workers=args.workers)
elif args.dataset == 'VOC':
# Create VOC train dataloader
transformations = transforms.Compose([transforms.Resize((300, 300)),
transforms.RandomChoice([
transforms.ColorJitter(brightness=(0.80, 1.20)),
transforms.RandomGrayscale(p = 0.25)
]),
transforms.RandomHorizontalFlip(p = 0.25),
transforms.RandomRotation(25),
transforms.ToTensor(),
transforms.Normalize(mean = mean, std = std),
])
VOC_dataset_train = PascalVOC_Dataset(data_dir,
year='2012',
image_set='train',
download=download_data,
transform=transformations,
target_transform=encode_labels)
# VOC validation dataloader
transformations_valid = transforms.Compose([transforms.Resize(330),
transforms.CenterCrop(300),
transforms.ToTensor(),
transforms.Normalize(mean = mean, std = std),
])
VOC_dataset_valid = PascalVOC_Dataset(data_dir,
year='2012',
image_set='val',
download=download_data,
transform=transformations_valid,
target_transform=encode_labels)
train_loader = DataLoader(VOC_dataset_train, batch_size=batch_size, num_workers=4, shuffle=True)
valid_loader = DataLoader(VOC_dataset_valid, batch_size=batch_size, num_workers=4)
# VOC testing loader
transformations_test = transforms.Compose([transforms.Resize(330),
transforms.FiveCrop(300),
transforms.Lambda(lambda crops: torch.stack([transforms.ToTensor()(crop) for crop in crops])),
transforms.Lambda(lambda crops: torch.stack([transforms.Normalize(mean = mean, std = std)(crop) for crop in crops])),
])
dataset_test = PascalVOC_Dataset(data_dir,
year='2012',
image_set='val',
download=download_data,
transform=transformations_test,
target_transform=encode_labels)
#---------------Test your model here---------------------------------------
# Load the best weights before testing
if args.app == 'train':
log_file = open(os.path.join(model_dir, "log-{}.txt".format(num)), "w+")
log_file.write("----------Experiment {} - {}-----------\n".format(num, model_name))
# log_file.write("transformations == {}\n".format(transformations.__str__()))
trn_hist, val_hist = train_model(model, device, optimizer, scheduler, train_loader, valid_loader, model_dir, num, epochs, log_file)
torch.cuda.empty_cache()
plot_history(trn_hist[0], val_hist[0], "Loss", os.path.join(model_dir, "loss-{}".format(num)))
plot_history(trn_hist[1], val_hist[1], "Accuracy", os.path.join(model_dir, "accuracy-{}".format(num)))
log_file.close()
elif args.app == 'test':
weights_file_path = os.path.join(model_dir, "model-{}.pth".format(num))
if os.path.isfile(weights_file_path):
print("Loading best weights")
model.load_state_dict(torch.load(weights_file_path))
### we use val dataset without 5 crops for testing
# test_loader = DataLoader(valid_loader, batch_size=int(batch_size), num_workers=0, shuffle=False)
if args.save_results:
loss, ap, scores, gt = test(model, device, valid_loader, returnAllScores=True, num_classes = args.num_classes)
gt_path, scores_path, scores_with_gt_path = os.path.join(model_dir, "gt-{}.csv".format(num)), os.path.join(model_dir, "scores-{}.csv".format(num)), os.path.join(model_dir, "scores_wth_gt-{}.csv".format(num))
utils.save_results(valid_loader.dataset.images, gt, utils.object_categories, gt_path)
utils.save_results(valid_loader.dataset.images, scores, utils.object_categories, scores_path)
utils.append_gt(gt_path, scores_path, scores_with_gt_path)
utils.get_classification_accuracy(gt_path, scores_path, os.path.join(model_dir, "clf_vs_threshold-{}.png".format(num)))
ap_1_list = []
for i in range(len(gt)):
score = utils.average_precision_score(gt[i], scores[i])
if score ==1:
ap_1_list.append(i)
np.save('ap_{}_list'.format(args.dataset), ap_1_list)
print('Testing AP: {}'.format(scores/len(gt)))
elif 'attack' in args.app:
weights_file_path = os.path.join(model_dir, "model-{}.pth".format(num))
if os.path.isfile(weights_file_path):
print("Loading best weights")
model.load_state_dict(torch.load(weights_file_path))
tkmlap(args, model, device, valid_loader)
elif 'baseline' in args.app:
weights_file_path = os.path.join(model_dir, "model-{}.pth".format(num))
if os.path.isfile(weights_file_path):
print("Loading best weights")
model.load_state_dict(torch.load(weights_file_path))
baselineap(args, model, device, valid_loader)
# Execute main function here.
if __name__ == '__main__':
parser = argparse.ArgumentParser(description='TKMLAP')
parser.add_argument('--data', default='./data', help='path to dataset') ## ./data or ./data/COCO_2014
parser.add_argument('--dataset', default='VOC', type=str, choices={'VOC', 'COCO'}, help='path to dataset')
parser.add_argument('--results', default='woSigmoid-BCE-Adam-bs64-box_-1_1', help='path to dataset') ##woSigmoid-BCE-Adam-bs64-box_-1_1, woSigmoid-BCE-Adam-bs128-box_-1_1_COCO_fixMem
parser.add_argument('--num_classes', default=20, type=int, help='number of classes') ##20 or 80
parser.add_argument('--arch', default='inception_v3',
help='model architecture: ' +
' (default: inception_v3, resnet50)')
parser.add_argument('-j', '--workers', default=1, type=int, metavar='N',
help='number of data loading workers (default: 4)')
parser.add_argument('--epochs', default=1, type=int, metavar='N',
help='number of total epochs to run')
parser.add_argument('--start-epoch', default=0, type=int, metavar='N',
help='manual epoch number (useful on restarts)')
parser.add_argument('-b', '--batch-size', default=1, type=int,
metavar='N',
help='mini-batch size (default: 256), this is the total '
'batch size of all GPUs on the current node when '
'using Data Parallel or Distributed Data Parallel')
parser.add_argument('--image_size', default=300, type=int,
metavar='N', help='input image size (default: 300, 224)')
parser.add_argument('--lr', '--learning-rate', default=[1.5e-4, 5e-2], type=float,
metavar='LR', help='initial learning rate', dest='lr')
parser.add_argument('--momentum', default=0.9, type=float, metavar='M',
help='momentum')
parser.add_argument('--gpu', default=0, type=int,
help='GPU id to use.')
parser.add_argument('--normalize', default='boxmaxmin', type=str, choices={'mean_std', 'boxmaxmin'},
help='optimizer for training')
parser.add_argument('--opt', default='Adam', type=str, choices={'Adam', 'SGD'}, help='optimizer for training')
parser.add_argument('--num', default=1, type=int, help='num to resume')
parser.add_argument('--download_data', default=False, type=bool, help='download data')
parser.add_argument('--save_results', default=True, type=bool, help='save results')
parser.add_argument('--k_value', default=3, type=int, help='k-value')
parser.add_argument('--eps', default=10, type=int, help='eps')
parser.add_argument('--maxiter', default=1000, type=int, help='max iteration to attack')
parser.add_argument('--remove_tier_para', default=0, type=float, help='remove_tier_para')
parser.add_argument('--boxmax', default=1, type=float, help='max value of input')
parser.add_argument('--boxmin', default=-1, type=float, help='min value of input')
parser.add_argument('--lr_attack', default=1e-2, type=float, help='learning rate of attacks')
parser.add_argument('--ufr_lower_bound', default=0.8, type=float, help='ufr lower bound')
parser.add_argument('--max_iter_uni', default=np.inf, type=int, help='max iter in universal')
parser.add_argument('--uap_train_index_end', default=3000, type=int, help='tain index in universal')
parser.add_argument('--uap_test_index_start', default=3000, type=int, help='test index start in universal')
parser.add_argument('--uap_test_index_end', default=4000, type=int, help='test index end in universal')
parser.add_argument('--uap_norm', default=2, help='2 or np.inf')
parser.add_argument('--uap_eps', default=100, type=int, help='eps for uap. 2000 for l_2 norm, 10 for l_infty norm.')
parser.add_argument('--label_difficult', default='best', type=str, choices={'best', 'random', 'worst'}, help='difficult types')
parser.add_argument('--app', default='none_target_attack', type=str, \
choices={'target_attack', 'none_target_attack', 'UAP_attack', 'baseline_rank', 'baseline_kfool', 'baseline_kUAP', 'test', 'train'}, \
help='attack types')
args = parser.parse_args()
os.environ["CUDA_VISIBLE_DEVICES"] = '{}'.format(args.gpu)
# main('./data/', "inception_v3", num=1, lr = [1.5e-4, 5e-2], epochs = 100, batch_size=1, download_data=False, save_results=True)
main(args)