File size: 14,616 Bytes
998bb30 |
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 |
import torch
import torchvision
import pandas as pd
import torch.nn as nn
import sys
import os
import numpy as np
import torchvision.transforms as transforms
import torchvision.datasets as datasets
from attacks.CGNC.image_transformer import TwoCropTransform
import timm
from torch.utils import model_zoo
import json
def load_robust_model(model_name):
if model_name in ['res50_sin', 'res50_sin_in', 'res50_sin_fine_in']:
model_urls = {
'res50_sin': 'https://bitbucket.org/robert_geirhos/texture-vs-shape-pretrained-models/raw/6f41d2e86fc60566f78de64ecff35cc61eb6436f/resnet50_train_60_epochs-c8e5653e.pth.tar',
'res50_sin_in': 'https://bitbucket.org/robert_geirhos/texture-vs-shape-pretrained-models/raw/60b770e128fffcbd8562a3ab3546c1a735432d03/resnet50_train_45_epochs_combined_IN_SF-2a0d100e.pth.tar',
'res50_sin_fine_in': 'https://bitbucket.org/robert_geirhos/texture-vs-shape-pretrained-models/raw/60b770e128fffcbd8562a3ab3546c1a735432d03/resnet50_finetune_60_epochs_lr_decay_after_30_start_resnet50_train_45_epochs_combined_IN_SF-ca06340c.pth.tar',
}
model_t = torchvision.models.resnet50(pretrained=False)
model_t = torch.nn.DataParallel(model_t).cuda()
checkpoint = model_zoo.load_url(model_urls[model_name])
model_t.load_state_dict(checkpoint["state_dict"])
elif model_name == 'adv_incv3':
model_t = timm.create_model('adv_inception_v3', pretrained=True)
elif model_name == 'ens_inc_res_v2':
model_t = timm.create_model('ens_adv_inception_resnet_v2', pretrained=True)
return model_t
# Load ImageNet model to evaluate
def load_model(model_name):
# Load Targeted Model
if model_name == 'dense201':
model_t = torchvision.models.densenet201(pretrained=True)
elif model_name == 'vgg19':
model_t = torchvision.models.vgg19(pretrained=True)
elif model_name == 'vgg16':
model_t = torchvision.models.vgg16(pretrained=True)
elif model_name == 'googlenet':
model_t = torchvision.models.googlenet(pretrained=True)
elif model_name == 'incv3':
model_t = torchvision.models.inception_v3(pretrained=True)
elif model_name == 'res152':
model_t = torchvision.models.resnet152(pretrained=True)
elif model_name == 'dense121':
model_t = torchvision.models.densenet121(pretrained=True)
elif model_name == "incv4":
model_t = timm.create_model('inception_v4', pretrained=True)
elif model_name == "inc_res_v2":
model_t = timm.create_model('inception_resnet_v2', pretrained=True)
elif model_name in ['res50_sin', 'res50_sin_in', 'res50_sin_fine_in', 'adv_incv3', 'ens_inc_res_v2']:
model_t = load_robust_model(model_name)
else:
raise ValueError
return model_t
def fix_labels(args, test_set):
val_dict = {}
with open("val.txt") as file:
for line in file:
(key, val) = line.split(',')
val_dict[key.split('.')[0]] = int(val.strip())
new_data_samples = []
for i, j in enumerate(test_set.samples):
org_label = val_dict[test_set.samples[i][0].split('/')[-1].split('.')[0]]
new_data_samples.append((test_set.samples[i][0], org_label))
test_set.samples = new_data_samples
return test_set
# This will fix labels for NIPS ImageNet
def fix_labels_nips(args, test_set, pytorch=False, target_flag=False):
'''
:param pytorch: pytorch models have 1000 labels as compared to tensorflow models with 1001 labels
'''
filenames = [i.split('/')[-1] for i, j in test_set.samples]
# Load provided files and get image labels and names
image_classes = pd.read_csv(os.path.join(args.data_dir, "images.csv"))
image_metadata = pd.DataFrame({"ImageId": [f[:-4] for f in filenames]}).merge(image_classes, on="ImageId")
true_classes = image_metadata["TrueLabel"].tolist()
target_classes = image_metadata["TargetClass"].tolist()
val_dict = {}
for f, i in zip(filenames, range(len(filenames))):
val_dict[f] = [true_classes[i], target_classes[i]]
new_data_samples = []
for i, j in enumerate(test_set.samples):
if target_flag:
org_label = val_dict[test_set.samples[i][0].split('/')[-1]][1]
else:
org_label = val_dict[test_set.samples[i][0].split('/')[-1]][0]
if pytorch:
new_data_samples.append((test_set.samples[i][0], org_label-1))
else:
new_data_samples.append((test_set.samples[i][0], org_label))
test_set.samples = new_data_samples
return test_set
def get_classes(label_flag):
if label_flag == 'N8':
label_set = np.array([150, 426, 843, 715, 952, 507, 590, 62])
elif label_flag == 'CL':
label_set = np.array([68, 56, 78, 8, 23, 84, 90, 65, 74, 76])
elif label_flag == 'C20':
label_set = np.array([4, 65, 70, 160, 249, 285, 334, 366, 394, 396, 458, 580, 593, 681, 815, 822, 849,
875, 964, 986])
elif label_flag == 'C50':
label_set = np.array([9, 71, 74, 86, 102, 141, 150, 181, 188, 223, 245, 275, 308, 332, 343, 352, 386,
405, 426, 430, 432, 450, 476, 501, 510, 521, 529, 546, 554, 567, 588, 597, 640,
643, 688, 712, 715, 729, 817, 830, 853, 876, 878, 883, 894, 906, 917, 919, 940,
988])
elif label_flag == 'C100':
label_set = np.array([6, 8, 31, 41, 43, 47, 48, 50, 56, 57, 66, 89, 93, 107, 121, 124, 130, 156, 159,
168, 170, 172, 178, 180, 202, 206, 214, 219, 220, 230, 248, 252, 269, 304, 323,
325, 339, 351, 353, 356, 368, 374, 379, 387, 395, 401, 435, 449, 453, 464, 472,
496, 504, 505, 509, 512, 527, 530, 542, 575, 577, 604, 636, 638, 647, 682, 683,
687, 704, 711, 713, 730, 733, 739, 746, 747, 763, 766, 774, 778, 783, 799, 809,
832, 843, 845, 846, 891, 895, 907, 930, 937, 946, 950, 961, 963, 972, 977, 984,
998])
elif label_flag == 'C200':
label_set = np.array([7, 12, 13, 14, 16, 22, 25, 36, 49, 58, 75, 84, 88, 104, 105, 112, 113, 114, 115,
117, 120, 134, 140, 143, 144, 155, 158, 165, 173, 182, 183, 194, 196, 200, 204,
207, 212, 218, 225, 231, 242, 244, 250, 261, 262, 266, 270, 277, 282, 288, 292,
297, 301, 310, 316, 320, 321, 327, 330, 348, 357, 359, 361, 365, 371, 375, 381,
382, 389, 407, 409, 411, 412, 413, 414, 418, 422, 436, 437, 445, 446, 448, 456,
461, 468, 470, 471, 474, 475, 480, 484, 486, 489, 491, 495, 500, 502, 506, 511,
514, 515, 526, 531, 535, 544, 547, 549, 561, 562, 566, 582, 591, 598, 603, 605,
610, 611, 612, 613, 616, 618, 619, 621, 627, 635, 641, 648, 653, 654, 656, 657,
658, 661, 662, 672, 673, 680, 686, 689, 691, 693, 697, 700, 705, 706, 707, 716,
725, 735, 743, 750, 752, 760, 768, 772, 776, 781, 790, 791, 796, 798, 800, 802,
811, 819, 823, 824, 828, 833, 834, 836, 848, 855, 874, 890, 893, 898, 903, 922,
923, 928, 931, 935, 936, 939, 943, 944, 945, 948, 955, 960, 967, 969, 970, 971,
980, 983, 990, 992, 999])
else:
raise ValueError
return label_set
def get_classes_cifar100(label_flag):
if label_flag == 'N8':
label_set = np.array([15, 42, 84, 71, 95, 50, 59, 6])
elif label_flag == 'C20':
label_set = np.array([4, 6, 7, 16, 24, 28, 33, 36, 39, 39, 45, 58, 59, 68, 81, 82, 84,
87, 96, 98])
elif label_flag == 'C50':
label_set = np.array([9, 71, 74, 86, 102, 141, 150, 181, 188, 223, 245, 275, 308, 332, 343, 352, 386,
405, 426, 430, 432, 450, 476, 501, 510, 521, 529, 546, 554, 567, 588, 597, 640,
643, 688, 712, 715, 729, 817, 830, 853, 876, 878, 883, 894, 906, 917, 919, 940,
988])
elif label_flag == 'C100':
label_set = np.array([6, 8, 31, 41, 43, 47, 48, 50, 56, 57, 66, 89, 93, 107, 121, 124, 130, 156, 159,
168, 170, 172, 178, 180, 202, 206, 214, 219, 220, 230, 248, 252, 269, 304, 323,
325, 339, 351, 353, 356, 368, 374, 379, 387, 395, 401, 435, 449, 453, 464, 472,
496, 504, 505, 509, 512, 527, 530, 542, 575, 577, 604, 636, 638, 647, 682, 683,
687, 704, 711, 713, 730, 733, 739, 746, 747, 763, 766, 774, 778, 783, 799, 809,
832, 843, 845, 846, 891, 895, 907, 930, 937, 946, 950, 961, 963, 972, 977, 984,
998])
elif label_flag == 'C200':
label_set = np.array([7, 12, 13, 14, 16, 22, 25, 36, 49, 58, 75, 84, 88, 104, 105, 112, 113, 114, 115,
117, 120, 134, 140, 143, 144, 155, 158, 165, 173, 182, 183, 194, 196, 200, 204,
207, 212, 218, 225, 231, 242, 244, 250, 261, 262, 266, 270, 277, 282, 288, 292,
297, 301, 310, 316, 320, 321, 327, 330, 348, 357, 359, 361, 365, 371, 375, 381,
382, 389, 407, 409, 411, 412, 413, 414, 418, 422, 436, 437, 445, 446, 448, 456,
461, 468, 470, 471, 474, 475, 480, 484, 486, 489, 491, 495, 500, 502, 506, 511,
514, 515, 526, 531, 535, 544, 547, 549, 561, 562, 566, 582, 591, 598, 603, 605,
610, 611, 612, 613, 616, 618, 619, 621, 627, 635, 641, 648, 653, 654, 656, 657,
658, 661, 662, 672, 673, 680, 686, 689, 691, 693, 697, 700, 705, 706, 707, 716,
725, 735, 743, 750, 752, 760, 768, 772, 776, 781, 790, 791, 796, 798, 800, 802,
811, 819, 823, 824, 828, 833, 834, 836, 848, 855, 874, 890, 893, 898, 903, 922,
923, 928, 931, 935, 936, 939, 943, 944, 945, 948, 955, 960, 967, 969, 970, 971,
980, 983, 990, 992, 999])
else:
raise ValueError
return label_set
def normalize(t):
mean = [0.485, 0.456, 0.406]
std = [0.229, 0.224, 0.225]
t[:, 0, :, :] = (t[:, 0, :, :] - mean[0])/std[0]
t[:, 1, :, :] = (t[:, 1, :, :] - mean[1])/std[1]
t[:, 2, :, :] = (t[:, 2, :, :] - mean[2])/std[2]
return t
def get_data(train_dir, scale_size, img_size):
data_transform = transforms.Compose([
transforms.Resize(scale_size),
transforms.CenterCrop(img_size),
transforms.ToTensor(),
])
train_set = datasets.ImageFolder(train_dir, TwoCropTransform(data_transform, img_size))
train_size = len(train_set)
print('Training data size:', train_size)
return train_set
def getImageNetClassIndex():
file = open('imagenet_class_index.json', 'r')
load_dic = json.load(file)
class_list = []
for item in load_dic:
cls = []
cls.append(load_dic[item][0])
cls.append(load_dic[item][1])
class_list.append(cls)
return class_list
def getCIFAR100ClassIndex():
class_list = [
('apple', 0), ('aquarium_fish', 1), ('orange', 2), ('peacock', 3), ('pear', 4),
('pickup_truck', 5), ('pine_tree', 6), ('plain', 7), ('plate', 8), ('pomegranate', 9),
('possum', 10), ('rabbit', 11), ('raccoon', 12), ('ray', 13), ('refrigerator', 14),
('rocket', 15), ('rose', 16), ('sea_horse', 17), ('sea_shell', 18), ('seal', 19),
('skeleton', 20), ('skyscraper', 21), ('snake', 22), ('spider', 23), ('squirrel', 24),
('streetcar', 25), ('sunflower', 26), ('sweet_pepper', 27), ('table', 28), ('tank', 29),
('telephone', 30), ('television', 31), ('tiger', 32), ('tractor', 33), ('train', 34),
('trout', 35), ('tulip', 36), ('umbrella', 37), ('watch', 38), ('water_lilly', 39),
('whale', 40), ('wheelchair', 41), ('wolf', 42), ('woman', 43), ('worm', 44),
('yellow_ladybug', 45), ('zebra', 46), ('bottle', 47), ('breakfast_cereal', 48), ('breezeblock', 49),
('brick', 50), ('bridge', 51), ('broom', 52), ('bucket', 53), ('bulldozer', 54),
('bus', 55), ('cabinet', 56), ('camera', 57), ('can', 58), ('cardigan', 59),
('carrot', 60), ('caterpillar', 61), ('cattle', 62), ('cello', 63), ('chandelier', 64),
('chicken', 65), ('clock', 66), ('cloud', 67), ('cockroach', 68), ('couch', 69),
('crab', 70), ('crane', 71), ('crocodile', 72), ('cup', 73), ('diamond', 74),
('dining_table', 75), ('dolphin', 76), ('donkey', 77), ('dragonfly', 78), ('electric_guitar', 79),
('elephant', 80), ('emu', 81), ('elevator', 82), ('envelope', 83), ('fire_engine', 84),
('flamingo', 85), ('flashlight', 86), ('floor_lamp', 87), ('flute', 88), ('forest', 89),
('frog', 90), ('furniture', 91), ('garbage_truck', 92), ('guitar', 93), ('hamburger', 94),
('harp', 95), ('harmonica', 96), ('helicopter', 97), ('horn', 98), ('hotel', 99)
]
return class_list
def get_mask(batch_perturb, mask_ratio, device, patch_size=32):
N, C, H, W = batch_perturb.shape
assert patch_size <= H and patch_size <= W
num_patch_h = H // patch_size
num_path_w = W // patch_size
mask = torch.zeros(patch_size, patch_size).unsqueeze(0).repeat(C, 1, 1).to(device) # 3 channel noise
mask_patch_num = int(num_patch_h * num_path_w * mask_ratio)
if mask_patch_num <= 0:
return batch_perturb
noise = torch.rand(N, num_patch_h * num_path_w).to(device)
mask_path = torch.argsort(noise, dim=1)[:, :mask_patch_num] # mask the first several patches in ascending order
for i in range(len(batch_perturb)):
for patch_idx in mask_path[i]:
row = patch_idx // num_path_w
col = patch_idx - row * num_path_w
batch_perturb[i, :, row * patch_size: (row + 1) * patch_size,
col * patch_size: (col + 1) * patch_size] *= mask
return batch_perturb
|