Why a copy is made if conv has padding='same' and even kernels

[grudloff]

Hello, I know this has more to do with pytorch than with pytorch-lightning, but I couldn't find any issues regarding this there and I don't think this is a bug so I don't think it deserves one.

I would like to know why a convolution with padding='same' and even kernels requires a copy of its input, as stated by this warning:

UserWarning: Using padding='same' with even kernel lengths and odd dilation may require a zero-padded copy of the input be created (Triggered internally at  ..\aten\src\ATen\native\Convolution.cpp:647.)

Thanks!