Problem with overlapping patches when utilizing the "combine_tensor_patches" function

[kiratben]

I encounter a problem with overlapping patches when utilizing the "combine_tensor_patches" function. I'm encountering a NotImplementedError: Only stride == window_size is supported, but I have (1, 1) stride and (3, 3) window size. If you have any suggestions for resolving this issue, please let me know.

[edgarriba]

can you how do you use the function ? the error you report seems to be raising this one
https://github.com/kornia/kornia/blob/master/kornia/contrib/extract_patches.py#L246

[kiratben]

patches = kornia.contrib.extract_tensor_patches(x, window_size=(3, 3), stride=(1, 1))
It seems that the issue arises when I use “combine_tensor_patches.”
output = kornia.contrib.combine_tensor_patches(patches, original_size=x.shape[2], window_size=(3,3), stride=(1,1))

[edgarriba]

what's your input x shape ?

[kiratben]

My input data is organized in a tensor with dimensions B (batch size), C (number of channels), H (height), and W (width), specifically in the shape of 32, 32, 64, 64.

[edgarriba]

not sure if the input shape should be divisible by window_size and stride? @shijianjian @ashnair1

[ashnair1]

Currently CombineTensorPatches only works if stride == window_size. In your case, your stride and window_size is (1,1) and (3,3) respectively. This is a known weakness of the implementation. #2650 aims to lift this restriction.