Question about Maximum/Minimum value of the pixels. about pytorch-adversarial-training

[ylsung]

You can replace clamp_ by torch.maximum and torch.minimum, which can do the element-wise comparison between tensors. The sample code might be like

transform_train = tv.transforms.Compose([ 
         tv.transforms.RandomCrop(32, padding=4, fill=0, padding_mode='constant'), 
         tv.transforms.RandomHorizontalFlip(), 
         tv.transforms.ToTensor(), 
         tv.transforms.Normalize(), # normalize by mean and std
     ]) 

min_tensor = torch.full((1, 3, 32, 32), 0)
max_tensor = torch.full((1, 3, 32, 32), 255)

min_tensor = transform_train(min_tensor)
max_tensor = transform_train(max_tensor)

def clamp(inputs, max, min):
    max = max.expand(inputs.shape[0], -1, -1, -1) # match the shape of max to be the same as which of inputs
    min = min.expand(inputs.shape[0], -1, -1, -1)
    
    return torch.maximum(torch.minimum(inputs, max), min)

# in fast_gradient_sign_untargeted.py
x = clamp(x, max_tensor, min_tensor)

The code snippet should fit in your needs.

[chenmc1996]

Thanks for your reply. It solves my question.