search

lyakaap / VAT-pytorch

Virtual Adversarial Training (VAT) implementation for PyTorch
296 stars 43 forks source link

random unit tensor strictly positive #2

Closed Anierin closed 6 years ago

Anierin commented 6 years ago

Hello,

When you "prepare random unit tensor" in the VAT code, you use this :

d = torch.rand(x.shape).to( torch.device('cuda' if torch.cuda.is_available() else 'cpu')) d = _l2_normalize(d)

It mean that d is strictly positive. It's intended? Sometimes we need to have a partially negative perturbation.

Why you didn't use something like the following code? d = torch.rand(x.shape) - 0.5 d = d.to('cuda' if torch.cuda.is_available() else 'cpu') d = _l2_normalize(d)

lyakaap commented 6 years ago

Thanks for pointing! Changed the code.