def __getitem__(self, idx):
"""
Resizes and normalizes tensors for Inception_v3
:param idx: ``int``, index
:return: Tensor([C, H, W])
"""
img = self.data[idx]
img = resize(img, size=(299, 299))
img = normalize(img,
mean=torch.tensor([0.485, 0.456, 0.406]),
std=torch.tensor([0.229, 0.224, 0.225]))
return img.squeeze(0)
def preprocess(self, images):
"""
Preprocess images
Args:
images: (N, 3, H, W), Input images
Return
x: (N, 3, H, W), Preprocessed images
"""
x = images
if self.pretrained:
# Create a mask for padded pixels
mask = (x == 0)
# Match ResNet pretrained preprocessing
x = normalize(x, mean=self.norm_mean, std=self.norm_std)
# Make padded pixels = 0
x[mask] = 0
return x
def fgsm_attack(images, epsilon, gradients, means, stds):
# We should apply adversarial attack in unnormalized domain
images = denormalize(images, means, stds)
images = torch.clamp(images, 0, 1)
# Collect the element-wise sign of the data gradient
gradient_signs = gradients.sign().float()
# Create the perturbed image by adjusting each pixel of the input image
perturbed_images = images + epsilon * gradient_signs
# Adding clipping to maintain [0,1] range
perturbed_images = torch.clamp(perturbed_images, 0, 1)
# Normalize the images back
perturbed_images = normalize(perturbed_images, means, stds)
# Return the perturbed image
return perturbed_images