No clamp for input after Gaussian data augumentation?

[xinghua-qu]

# augment inputs with noise
inputs = inputs + torch.randn_like(inputs, device='cuda') * noise_sd

Line 109 in train.py.

If there is no torch.clamp() after line 109, it is possible that the input image will exceed its allowed pixel value range (i.e., [0,1])

[vietvo89]

I have the same question. Why the implementation of this approach does not clip the image within the range [0,1]. Without clarification, I think it is a pitfall. When using torch.clamp(), the results of the paper are incorrect.