I have a tensor which has empty values. I want to apply a 2d blurring convolution kernel just to the non-empty values. For simplicity let’s assume kernel weights are all equal.
Maybe something like dynamic kernel weights, so kernel weights = 1/n, where n is the number of valid pixels in the region.
Is there currently pytorch functionality which supports something like this?
I think you can just apply a convolution using a mask. In the end a convolution is product + sum between the elements the kernel and the tensor. If you just multiply your tensor by a binary mask (ones for the non-empty values and zeros for empty values) the result should be that.
Besides you have torch.fold and unfold which you can use to emulate an sliding kernel implementing ur own operation.
https://pytorch.org/docs/stable/generated/torch.nn.Fold.html
But basically masking either the input or the output of the convolution should work.
Ooor maybe you can use sparse tensors (Not really aware how deep they have been implemented as i don’t use them).