Padding in Tensorflow vs Pytorch

[zanussbaum]

Hey I had a quick question regarding padding. I don't have a ton of experience with Pytorch, but as far as I can tell from this, it seems that the layer adds the parameter padding of zeros to each axis of the input.

However, in your implementation, you use padding='same'. In the original implementation in pytorch, the author uses padding=1. I'm under the impression that to get the padding the same, we would need to specify the padding manually in tensorflow.

Am I missing something here where Tensorflow is taking care of something under the hood that I'm not understanding?

[asmith26]

Hi @zanussbaum,

Thanks for the info. It's been a while since I wrote this, and I'm not familiar with PyTorch syntax so I'm not sure if there is a discrepancy. I remember when I originally wrote this I followed the original Lua code, https://github.com/szagoruyko/wide-residual-networks/blob/master/models/wide-resnet.lua.

If you think there is a discrepancy and can improve on the model results currently obtained, I happily accept pull requests.

Thanks again! :)

[zanussbaum]

Thanks I'll try and see what's going on. If there's not a discrepancy, at least I'll hopefully understand everything a little more :)