What is the fastest way to do:
(x - m) / sqrt(v + eps)
where x, m and v are tensors of same shape, and eps is a scalar ?
Can I borrow the functional of some normalization layer? It seems not possible because if I give running_mean=m and running_var=v those functions complain about the dimension.
Using torch.rsqrt for 1/sqrt() does not make it much faster.