lcn

lcn copied to clipboard

https://github.com/anthonymlortiz/lcn/blob/master/local_context_norm.py#L40 if squares < sums*sums/n var becomes negative so when you take square root later in padded_vars it leads to nan results

would adding var = torch.abs(var) make sense

The variance should never be negative. I had the same problem, in my case it was caused by overflow because I was using float16.