Downsample / concatenate?

[biggzlar]

Any explanation as to why this function is so involved? It's simply copying the input tensor and reshaping it. Why do we need four steps to do this?

attention-cnn/models/transformer.py

Line 145 in 21483bb

def downsample_concatenate(X, kernel):

[jbcdnr]

I think the reshaping is not just one line because vertical neighboring pixels are not next to each other in memory. So for example for a 2x2 concatenation, we first concatenate the channels of the horizontal neighbors. Then permute the dimensions and concatenate the vertical neighbors that are contiguous in memory. I was a bit crazy with the .contiguous(), some are maybe not necessary but they would lead to a no op so should not be a problem ;)

I am curious if you have a better/simpler solution that passes a test to match the output of my function.

[jbcdnr]

It could be made simpler using unfold probably.