-
Notifications
You must be signed in to change notification settings - Fork 3
/
demo_mtlrobust.py
361 lines (279 loc) · 12.4 KB
/
demo_mtlrobust.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
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
#!/usr/bin/env python
# -*- coding: utf-8 -*-
# This code is the demo for the robustness under multitask attack on segmentation model.
import argparse
import json
import logging
import math
import os
from os.path import exists, join, split
import threading
import time
import numpy as np
import shutil
import sys
from PIL import Image
import torch
from torch import nn
import torch.backends.cudnn as cudnn
import torch.optim as optim
from torchvision import datasets, transforms
from torch.autograd import Variable
import models.drn as drn
from models.DRNSeg import DRNSeg
from models.FCN32s import FCN32s
import data_transforms as transforms
try:
from modules import batchnormsync
except ImportError:
pass
from learning.train_loop import *
from learning.validate import *
from learning.test import *
FORMAT = "[%(asctime)-15s %(filename)s:%(lineno)d %(funcName)s] %(message)s"
logging.basicConfig(format=FORMAT)
logger = logging.getLogger(__name__)
logger.setLevel(logging.DEBUG)
def get_torch_std(info):
std_array = np.asarray(info["std"])
tensor_std = torch.from_numpy(std_array)
tensor_std = tensor_std.unsqueeze(0)
tensor_std = tensor_std.unsqueeze(2)
tensor_std = tensor_std.unsqueeze(2).float()
return tensor_std
def PGD_drnseg_masked_attack_city(x, y, attack_mask, net, criterion, epsilon, steps, step_size, info):
tensor_std = get_torch_std(info)
GPU_flag = False
if torch.cuda.is_available():
GPU_flag = True
x_adv = x.clone()
epsilon = epsilon / 255.
step_size = step_size / 255.
ones_like_x_adv = torch.ones_like(x_adv)
if GPU_flag:
ones_like_x_adv = ones_like_x_adv.cuda()
tensor_std = tensor_std.cuda()
pert_epsilon = ones_like_x_adv * epsilon / tensor_std
pert_upper = x_adv + pert_epsilon
pert_lower = x_adv - pert_epsilon
upper_bound = torch.ones_like(x_adv)
lower_bound = torch.zeros_like(x_adv)
upper_bound = forward_transform(upper_bound, info)
lower_bound = forward_transform(lower_bound, info)
upper_bound = torch.min(upper_bound, pert_upper)
lower_bound = torch.max(lower_bound, pert_lower)
# TODO: print and check the bound
ones_x = torch.ones_like(x).float()
if GPU_flag:
# Loss = 0
x_adv = x_adv.cuda()
upper_bound = upper_bound.cuda()
lower_bound = lower_bound.cuda()
tensor_std = tensor_std.cuda()
ones_x = ones_x.cuda()
y = y.cuda()
step_size_tensor = ones_x * step_size / tensor_std
x_adv = Variable(x_adv, requires_grad=True)
for i in range(steps):
h_adv = net(x_adv) # dict{rep:float32,segmentasemantic:float32, depth_zbuffer:float32, reconstruct:float32}
ignore_value = 255
attack_mask = attack_mask.long()
y = y * attack_mask + ignore_value * (1 - attack_mask) # y is {auto:float,segsem:int64,deoth:float}
grad_total_loss = criterion(h_adv[0], y)
net.zero_grad()
if x_adv.grad is not None:
x_adv.grad.data.fill_(0)
grad_total_loss.backward()
x_adv.grad.sign_()
x_adv = x_adv + step_size_tensor * x_adv.grad
x_adv = clamp_tensor(x_adv, upper_bound, lower_bound)
x_adv = Variable(x_adv.data, requires_grad=True) # TODO: optimize, remove this variable init each
return x_adv
def test_drnseg_masked_attack(eval_data_loader, model):
model.eval()
batch_time = AverageMeter()
data_time = AverageMeter()
end = time.time()
# hist = np.zeros((num_classes, num_classes))
# exit(0)
if torch.cuda.is_available():
GPU_flag = True
else:
GPU_flag = False
# Number of points to be selected for masking - analogous to number of output dimensions. Only these many pixels will be considered to calculate the loss.
select_num_list = [1, 2, 5, 10, 50, 100, 200] + [i * 500 for i in range(1, 15)]
result_list = []
for select_num in select_num_list:
print("********")
print("selecting {} of output".format(select_num))
import random
acc_sample_avg_sum = 0
if select_num < 400:
MCtimes = 3
else:
MCtimes = 1
MCtimes = 1
# Monte Carlo Sampling - MCTimes is the number of times that we sample
for inner_i in range(MCtimes):
acc_sum = 0
cnt = 0
print("MC time {}".format(inner_i))
for iter, (image, label, name) in enumerate(eval_data_loader):
# break if 50 images (batches) done to speed up the demo
if cnt > 50:
break
data_time.update(time.time() - end)
if torch.cuda.is_available():
image_var = Variable(image.cuda(), requires_grad=True)
else:
image_var = Variable(image, requires_grad=True)
# Generate mask for attack
# For this image, sample select_num number of pixels
temp = [i for i in range(image_var.size(2) * image_var.size(3))]
selected = random.sample(temp, select_num)
attack_mask = np.zeros((image_var.size(2) * image_var.size(3)), dtype=np.uint8)
for iii in range(select_num):
attack_mask[selected[iii]] = 1
attack_mask = attack_mask.reshape(1, 1, image_var.size(2), image_var.size(3))
attack_mask = torch.from_numpy(attack_mask)
if GPU_flag:
attack_mask = attack_mask.cuda()
label = label.cuda()
attack_mask = Variable(attack_mask)
# TODO - get the things reqd for its arguments such as criteria and tasks
info = {"mean": [0.29010095242892997,0.32808144844279574,0.28696394422942517],
"std": [0.1829540508368939, 0.18656561047509476, 0.18447508988480435]}
criteria = cross_entropy2d
adv_image_var = PGD_drnseg_masked_attack_city(image_var, label, attack_mask, model, criteria,
epsilon=8, steps=5,
step_size=2, info=info)
final = model(adv_image_var)[0]
_, pred = torch.max(final, 1)
def accuracy_masked_attack(preds, label, mask):
valid_label = (label >= 0) * (label <= 18)
valid = valid_label * mask.bool()
acc_sum = (valid * (preds == label)).sum()
valid_sum = valid.sum()
acc = float(acc_sum) / (valid_sum + 1e-10)
return acc
acc = accuracy_masked_attack(pred, label, attack_mask)
acc_sum += acc
cnt += 1
batch_time.update(time.time() - end)
end = time.time()
acc_batch_avg = acc_sum / cnt # Represents the gradient average for batch. cnt is the number of samples in a batch.
acc_sample_avg_sum += acc_batch_avg # For each sampling this is the sum of avg gradients in that sample.
acc_sample_avg_sum /= MCtimes
result_list.append(acc_sample_avg_sum)
print(select_num, 'middle result', result_list)
print('Final', result_list)
def test_grad_diffoutdim(eval_data_loader, model):
"""
Evaluates the effect of increasing output dimension on the norm of the gradient.
Monte Carlo sampling will be used and the result would be averaged.
First choose the number of pixels to calculate the loss for (output dimension) --> select_num.
For each select_num, we do the following MC_times(as Monte Carlo sampling):
Calculate the loss for select_num pixels chosen, backpropagate and get the input gradient.
Average all these.
:param eval_data_loader:
:param model:
:param num_classes:
:param output_dir:
:param has_gt:
:param save_vis:
:param downsize_scale:
:param args:
:return:
"""
model.eval()
if torch.cuda.is_available():
GPU_flag = True
else:
GPU_flag = False
# Number of points to be selected for masking - analogous to number of output dimensions. Only these many pixels will be considered to calculate the loss.
select_num_list = [1] + [i * 4 for i in range(1, 100)] + [400 + i*200 for i in range(100)]
result_list = []
for select_num in select_num_list:
print("********")
print("selecting {} of output".format(select_num))
import random
grad_sample_avg_sum = 0
MCtimes = 20
MCtimes = 1 # for demo speed up
# Monte Carlo Sampling - MCTimes is the number of times that we sample
for inner_i in range(MCtimes):
grad_sum = 0
cnt = 0
print("MC time {}".format(inner_i))
for iter, (image, label, name) in enumerate(eval_data_loader):
# break if 50 images (batches) done
if cnt > 200:
break
if torch.cuda.is_available():
image_var = Variable(image.cuda(), requires_grad=True)
else:
image_var = Variable(image, requires_grad=True)
final = model(image_var)[0]
_, pred = torch.max(final, 1)
# for this image, sample select_num number of pixels
temp = [i for i in range(image_var.size(2) * image_var.size(3))]
selected = random.sample(temp, select_num)
# Build mask for image -
mask = np.zeros((image_var.size(2) * image_var.size(3)), dtype=np.uint8)
for iii in range(select_num):
mask[selected[iii]] = 1
mask = mask.reshape(1, 1, image_var.size(2), image_var.size(3))
mask = torch.from_numpy(mask)
mask = mask.float()
mask_target = mask.long()
label = label.long()
if GPU_flag:
mask = mask.cuda()
mask_target = mask_target.cuda()
label = label.cuda()
target, mask = Variable(label), Variable(mask)
loss = cross_entropy2d(final * mask, target * mask_target, size_average=False)
loss.backward()
data_grad = image_var.grad
np_data_grad = data_grad.cpu().numpy()
L2_grad_norm = np.linalg.norm(np_data_grad) / select_num # the 1/M \sum_M \partial{Loss_i}/\partial{input}
grad_sum += L2_grad_norm
# increment the batch # counter
cnt += 1
grad_avg = grad_sum / cnt # Represents the gradient average for batch. cnt is the number of samples in a batch.
grad_sample_avg_sum += grad_avg # For each sampling this is the sum of avg gradients in that sample.
grad_sample_avg_sum /= MCtimes
result_list.append(grad_sample_avg_sum)
print(select_num, 'graident L2 Norm List:', result_list)
np.save('grad_mtl.npy', result_list)
print('Final', result_list)
np.save('grad_mtl.npy', result_list)
def test_seg(data_dir, modeldir, test_acc_output_dim, get_grad):
test_batch_size =2
single_model = DRNSeg("drn_d_22", 19, pretrained_model=None,
pretrained=False)
single_model.load_state_dict(torch.load(modeldir))
model = torch.nn.DataParallel(single_model)
if torch.cuda.is_available():
model.cuda()
normalize = transforms.Normalize(mean=[0.29010095242892997,0.32808144844279574,0.28696394422942517],
std=[0.1829540508368939, 0.18656561047509476, 0.18447508988480435])
test_loader = torch.utils.data.DataLoader(
SegList(data_dir, 'val', transforms.Compose([
transforms.ToTensor(),
normalize,
]), list_dir=None, out_name=True),
batch_size=test_batch_size, shuffle=False, num_workers=8,
pin_memory=True, drop_last=True
)
if test_acc_output_dim:
test_drnseg_masked_attack(test_loader, model)
elif get_grad:
_ = test_grad_diffoutdim(test_loader, model)
from learning.adv_train_loop import train_seg_adv
def main():
datadir = '/local/rcs/ECCV/Cityscape/cityscape_dataset'
modeldir = '/local/rcs/ECCV/Cityscape/model_zoo/DRN/drn_d_22_cityscapes.pth'
test_seg(datadir, modeldir, test_acc_output_dim=False, get_grad=True)
if __name__ == '__main__':
main()