[Feature request] Support SAME padding for convolutional layers

[zuoxingdong]

It would be very convenient to have SAME padding supported for convolutional layers

[zou3519]

What is "same" padding? Could you give an example?

[zuoxingdong]

@zou3519 For example, it could be easier to have the output size identical to input size if the stride is 1. More generally,

out_H = torch.ceil(in_H/stride)
out_W = torch.ceil(in_W/stride)

[crcrpar]

"SAME" padding option is used in TensorFlow. It's explained in below document.
https://www.tensorflow.org/api_guides/python/nn#Convolution

[ssnl]

I'm personally not a fan of "same" as a padding option. It hides some degrees of freedom from user, e.g. when kernel size is even.

[fmassa]

Duplicate of #3867
I personally feel that we should not support this except if the behavior exactly matches tensorflow.
But for that, we would need to have asymmetric padding for our operations, which would involve some non-trivial changes. I've discussed those points in more detail in #3867

Closing as a duplicate.