Use torch.*_like()

[adsharma]

b = torch.zeros_like(a)

should be preferred over

b = torch.zeros(...).to(a)

to avoid copies.