def generate_un_ad_sample(img_path, img_label):
"""
Args Type: str, int
Result path: '../generated/adv_image_name'
"""
pretrained_model = models.resnet18(pretrained=True)
FGS_untargeted = FastGradientSignUntargeted(pretrained_model, 0.01)
(original_image, prep_img, target_class) = get_params(img_path, img_label)
img_name = img_path.split('/')[-1]
gampicpath = FGS_untargeted.generate(original_image, target_class, img_name)
return gampicpath
def generate_tar_ad_sample(img_path, img_label):
"""
Args Type: str, int
Result path: '../generated/adv_image_name'
"""
#print(1)
pretrained_model = models.resnet18(pretrained=True)
FGS_untargeted = FastGradientSignTargeted(pretrained_model, 0.01)
original_image, prep_img, org_class = get_params(img_path, img_label)
#print(type(original_image), type(prep_img), type(org_class))
target_class = org_class + 100
if (target_class > 999):
target_class = target_class % 1000
img_name = img_path.split('/')[-1]
#print(img_name)
gampic = FGS_untargeted.generate(original_image, org_class, target_class,
img_name)
#print(gampic)
return gampic
model_output = self.model(self.input_image)
# Zero grads
self.model.zero_grad()
# Target for backprop
one_hot_output = torch.FloatTensor(1, model_output.size()[-1]).zero_()
one_hot_output[0][self.target_class] = 1
# Backward pass
model_output.backward(gradient=one_hot_output)
# Convert Pytorch variable to numpy array
# [0] to get rid of the first channel (1,3,224,224)
gradients_as_arr = self.gradients.data.numpy()[0]
return gradients_as_arr
if __name__ == '__main__':
# Get params
target_example = 0 # Snake
(original_image, prep_img, target_class, file_name_to_export, pretrained_model) =\
get_params(target_example)
# Vanilla backprop
VBP = VanillaBackprop(pretrained_model, prep_img, target_class)
# Generate gradients
vanilla_grads = VBP.generate_gradients()
# Save colored gradients
save_gradient_images(vanilla_grads, file_name_to_export + '_Vanilla_BP_color')
# Convert to grayscale
grayscale_vanilla_grads = convert_to_grayscale(vanilla_grads)
# Save grayscale gradients
save_gradient_images(grayscale_vanilla_grads, file_name_to_export + '_Vanilla_BP_gray')
print('Vanilla backprop completed')
tv_reg_beta)
# Sum all to optimize
loss = euc_loss + reg_alpha + reg_total_variation
# Step
loss.backward()
optimizer.step()
# Generate image every 5 iterations
if i % 5 == 0:
print('Iteration:', str(i), 'Loss:', loss.data.numpy()[0])
x = recreate_image(opt_img)
cv2.imwrite('../generated/Inv_Image_Layer_' + str(target_layer) +
'_Iteration_' + str(i) + '.jpg', x)
# Reduce learning rate every 40 iterations
if i % 40 == 0:
for param_group in optimizer.param_groups:
param_group['lr'] *= 1/10
if __name__ == '__main__':
# Get params
target_example = 0 # Snake
(original_image, prep_img, target_class, file_name_to_export, pretrained_model) =\
get_params(target_example)
inverted_representation = InvertedRepresentation(pretrained_model)
image_size = 224 # width & height
target_layer = 12
inverted_representation.generate_inverted_image_specific_layer(prep_img,
image_size,
target_layer)
default="../dataset/msra/valid/label/im100.png")
parser.add_argument('--model_path', type=str, default="./runs/debug/")
parser.add_argument('--fold', type=str, default="unet16")
parser.add_argument('--model_type', type=str, default="UNet16")
args = parser.parse_args()
image_path = args.image_path
mask_path = args.mask_path
model_path = str(
Path(args.model_path).joinpath(
'model_{fold}.pt'.format(fold=args.fold)))
model_type = args.model_type
(original_image, prep_img, prep_mask, file_name_to_export,
pretrained_model, img_width,
img_height) = get_params(image_path, mask_path, model_path, model_type)
# Vanilla backprop
VBP = VanillaBackprop(pretrained_model)
# Generate gradients
vanilla_grads = VBP.generate_gradients(prep_img, prep_mask)
print(vanilla_grads.shape)
# Save colored gradients
# save_gradient_images(vanilla_grads, file_name_to_export + '_Vanilla_BP_color', img_width, img_height)
# Convert to grayscale
grayscale_vanilla_grads = convert_to_grayscale(vanilla_grads)
# Save grayscale gradients
'../generated/targeted/adv_img_from_' + image_path + '_' +
str(org_class) + '_to_' + str(confirmation_prediction) +
'.jpg', recreate_img)
#FGSM 생성한 이미지의 Xception 모델 결과값(Fake/Real).
#deepfake_output = self.model(torch.Tensor(recreate_img))
deepfake_output = self.model(
processed_image
) #torch.from_numpy(np.flip(recreate_img, axis=0).copy())
_, FGSM_result = deepfake_output.data.max(1)
print("FGSM result : ", FGSM_result.numpy()[0])
break
return 1
#targeted 말구 untargeted 로 하면 안되려나..?!
#내가 생각한 생성방법이랑 너무 다른듯... 모든 이미지에 해당하는 AE 가 생성되어야 하는거 아닌가?
if __name__ == '__main__':
print("------------ Targeted FGSM 실행 ----------")
folder_pth = 'D:/full_data/test/real'
target_class = 1 #fake 라 1임.
for i in range(len(os.listdir(folder_pth))):
#전체 파일 리스트가 for문 동안 계속 왔다갔다하면 비효율적인데...
(original_image, prep_img, org_class, _, pretrained_model,
img_path) = get_params(i)
FGS_untargeted = FastGradientSignTargeted(pretrained_model, 0.02)
FGS_untargeted.generate(original_image, org_class, target_class,
img_path)
#FGSM -> model 에 넣어서 결과 출력
