Non-square kernels #109
Description
Hello! I have a specific use case that requires the use of non-square kernels, e.g. (2,1), not supported in ESCNN by default.
From Cohen & Welling 2016, it seems this should be no problem in principle, simply by masking an appropriate square kernel.
In the ESCNN approach though, since the weights serve as expansion coefficients, its not clear to me how I should do the masking in a principled way.
Any suggestions? Is looking at e.g. https://github.com/diningeachox/G-CNN instead my best option?