How to implement PyTorch's nn.MaxPool2d(2, 1, ceil_mode=True) in Burn?

[dfsfdfse]

How to implement PyTorch's nn.MaxPool2d(2, 1, ceil_mode=True) in Burn? Since PaddingConfig::Same does not support asymmetric padding, how to resolve this?

[laggui]

I'm afraid this is not currently supported, iirc ceil_mode might add padding to the right and down of the input (asymmetric, as you point out). So there are two parts to this: 1) asymmetric padding and 2) ceil_mode, both of which are not currently implemented. Asymmetric padding support has to come first (#2676), but feel free to open an issue for ceil mode!

In the meantime, you might be able to manually pad the input with tensor.pad((left, right, top, bottom), value) before applying the maxpool.

[dfsfdfse]

Thank you for taking the time to respond。

[laggui]

There's already an issue for asymmetric padding, but if you'd like to see ceil_mode added following that you can always make a feature request :)