normalize problem

[PaParaZz1]

I found in your implementation you normalize the total kernel with softmax. But actually, in origin paper it said:
each k up × k up reassembly kernel is normalized with a softmax function spatially
so you only need to normalize kernel in the channel dimension.
And I reimplement a new version of CARAFE. We can discuss how to speed up this basic implementation.
https://github.com/PaParaZz1/CARAFE_pytorch

[sephirothhua]

You reminded me that i normal the wrong dimension. I'll update later, Thanks for your reply.