-
Notifications
You must be signed in to change notification settings - Fork 1
/
Copy pathsimple_adv_training.py
166 lines (142 loc) · 7.41 KB
/
simple_adv_training.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
import torch
import os
from depth_model import import_depth_model
from torchvision import transforms
from my_utils import device0, eval_depth_diff, get_mean_depth_diff, ori_H, ori_W, object_dataset_root, save_depth_model
from torchattacks import PGD_depth, Phy_obj_atk, Phy_obj_atk_l0
from torch.utils.data.dataloader import DataLoader
from dataLoader import KittiLoader
import torch.optim as optim
from tensorboardX import SummaryWriter
import datetime
import numpy as np
import random
from options import getCLIOptions
from image_preprocess import process_car_img
args = getCLIOptions()
print(args)
scene_size = (1024, 320)
original_size = (ori_W, ori_H)
trans = transforms.Resize([int(scene_size[1]), int(scene_size[0])])
# log_dir = os.path.join('/data/cheng443/model_harden', 'logs', datetime.datetime.now().strftime('%b%d_%H-%M-%S') + f"_{args['log_postfix']}")
# log_dir = os.path.join('/home/jcl3689/zhiyuan/tmp/model_harden_log', datetime.datetime.now().strftime('%b%d_%H-%M-%S') + f"_{args['log_postfix']}")
log_dir = os.path.join('/home/jcl3689/zhiyuan/tmp/model_harden_log', f"{args['log_postfix']}")
os.makedirs(log_dir)
logger = SummaryWriter(log_dir)
logger.add_text('Options', str(args), 0)
def setup_seed(seed):
torch.manual_seed(seed)
torch.cuda.manual_seed_all(seed)
np.random.seed(seed)
random.seed(seed)
torch.backends.cudnn.deterministic = True
def get_atk_model(model_rob):
if args['adv_type'] == 'image':
depth_atk = PGD_depth(model_rob, eps=args['epsilon'], alpha=args['alpha'],steps=args['step'])
depth_atk._targeted = True
else:
car_img_resize, car_mask_np, _ = process_car_img('BMW.png', '-2')
obj_tensor = transforms.ToTensor()(car_img_resize)[:3,:,:].unsqueeze(0).float().to(device0)
mask_tensor = torch.from_numpy(car_mask_np).unsqueeze(0).unsqueeze(0).float().to(device0)
if args['adv_type'] == 'object':
depth_atk = Phy_obj_atk(model_rob, obj_tensor, mask_tensor,\
eps=args['epsilon'], alpha=args['alpha'], steps=args['step'])
elif args['adv_type'] == 'object_l0':
depth_atk = Phy_obj_atk_l0(model_rob, obj_tensor, mask_tensor, \
adam_lr=args['adam_lr'], steps=args['step'], mask_wt=args['mask_wt'], l0_thresh=args['l0_thresh'])
else:
raise NotImplementedError('adv_type not implemented')
# logger.add_image('input/obj_img', obj_tensor[0], 0)
# logger.add_image('input/obj_mask', mask_tensor[0], 0)
return depth_atk
def eval_atk_perf(model_gt, model, data_loader, epoch):
depth_atk = get_atk_model(model)
model.eval()
model_gt.eval()
model_acc, atk_perf = 0, 0
eval_count = 50
start_idx = 42
for i, (scene_img, _) in enumerate(data_loader):
if i < start_idx:
continue
if i - start_idx >= eval_count:
break
scene_img = scene_img.to(device0)
# scene_img = trans(scene_img)
if args['adv_type'] == 'object' or args['adv_type'] == 'object_l0':
adv_images, ben_images, obj_masks_out, obj_img_adv = depth_atk(scene_img, args['batch_size'], eval=True)
elif args['adv_type'] == 'image':
adv_images, ben_images = depth_atk(scene_img)
obj_masks_out = None
with torch.no_grad():
disp_gt = model_gt(ben_images)
disp_pre = model(ben_images)
disp_atk = model(adv_images)
if i == start_idx:
result_img_model, _, _ = eval_depth_diff(ben_images[[0]], ben_images[[0]], model, disp1=disp_pre[[0]], disp2=disp_gt[[0]])
result_img_atk, _, _ = eval_depth_diff(adv_images[[0]], ben_images[[0]], model_gt, disp1=disp_atk[[0]], disp2=disp_gt[[0]])
model_acc += get_mean_depth_diff(disp_pre, disp_gt, scene_car_mask=None, use_abs=True) # the lower, the better model performance
atk_perf += get_mean_depth_diff(disp_atk, disp_gt, scene_car_mask=obj_masks_out, use_abs=True) # the higher, the better attack, the lower robustness
# data_size = len(data_loader)
logger.add_image('eval/model_comp', transforms.ToTensor()(result_img_model), epoch)
logger.add_image('eval/atk_comp', transforms.ToTensor()(result_img_atk), epoch)
model_perf = model_acc / eval_count
atk_perf = atk_perf / eval_count
logger.add_scalar('eval/model_perf', model_perf, epoch)
logger.add_scalar('eval/atk_perf', atk_perf, epoch)
return model_perf, atk_perf
def do_adv_training(model_rob, model):
model_ori = model
model_ori.eval()
total_epoch = 20
# if args['adv_type'] == 'object':
# loader_batch_size = 1
# elif args['adv_type'] == 'image':
# loader_batch_size = args['batch_size']
loader_batch_size = args['batch_size']
kitti_loader_train = KittiLoader(mode='train', root_dir=object_dataset_root, train_list='trainval.txt', val_list='test.txt', size=original_size)
train_loader = DataLoader(kitti_loader_train, batch_size=loader_batch_size, shuffle=True, num_workers=3, pin_memory=True, drop_last=True)
kitti_loader_test = KittiLoader(mode='val', root_dir=object_dataset_root, train_list='trainval.txt', val_list='test.txt', size=original_size)
test_loader = DataLoader(kitti_loader_test, batch_size=loader_batch_size, shuffle=False, num_workers=3, pin_memory=True, drop_last=True)
print(f"Train set size: {len(kitti_loader_train)}, Test set size: {len(kitti_loader_test)}")
optimizer = optim.Adam(model_rob.parameters(), lr=0.0001)
depth_atk = get_atk_model(model_rob)
loss_creteria = torch.nn.MSELoss()
model_perf, atk_perf = eval_atk_perf(model_ori, model_rob, test_loader, 0)
print(f"Initial performance: model perf: {model_perf}, attack perf: {atk_perf}")
for epoch in range(total_epoch):
print('Current epoch: ', epoch)
# train
model_rob.train()
for i, (scene_img_ori, _) in enumerate(train_loader):
scene_img_ori = scene_img_ori.to(device0)
if args['adv_type'] == 'object' or args['adv_type'] == 'object_l0':
adv_images, ben_images, obj_masks_out, obj_img_adv = depth_atk(scene_img_ori, args['batch_size'])
elif args['adv_type'] == 'image':
adv_images, ben_images = depth_atk(scene_img_ori)
with torch.no_grad():
disp_gt = model_ori(ben_images)
pre_disp = model_rob(adv_images)
loss = loss_creteria(disp_gt, pre_disp)
optimizer.zero_grad()
loss.backward()
optimizer.step()
if i % 30 == 0:
print("Current step: ", i)
# evaluate
model_rob.eval()
model_perf, atk_perf = eval_atk_perf(model_ori, model_rob, test_loader, epoch+1)
print(f"Performance: model perf: {model_perf}, attack perf: {atk_perf}")
# save model
if epoch % 2 == 0:
save_depth_model(model_rob, log_dir, epoch+1)
# torch.save(model_rob.state_dict(), os.path.join(log_dir, f'model_rob_ep_{epoch+1}.pt'))
# end training
save_depth_model(model_rob, log_dir, 'final')
# torch.save(model_rob.state_dict(), os.path.join(log_dir, f'model_rob_ep_final.pt'))
if __name__ == "__main__":
setup_seed(args['random_seed'])
model = import_depth_model(scene_size).to(device0)
model_rob = import_depth_model(scene_size).to(device0)
# torch.save(model_rob.state_dict(), os.path.join(log_dir, f'model_rob_ep_test.pt'))
do_adv_training(model_rob, model)