doing a convolve2d in valid mode but return same size.

[nickylimjj]

I understand that there are 3 modes for convolving and correlating. For valid mode, it outputs a smaller image as it only convolves the valid elements. I believe it should be a core functionality to convolves in valid mode but output the same size image padded with fill_value

[rgommers]

I understand that there are 3 modes for convolving and correlating.

For signal.convolve/correlate, indeed. ndimage.convolve/correlate always returns the same size, which is usually what you want for image processing type applications. what is your use case? I can imagine why your proposal would be useful, but only fill_value=np.nan would be a sensible default.

[nickylimjj]

My use case is for image convolution. I am a student in a computer vision class and we are using the scipy.signal library for convolve2d and wanted to get the same output size for valid convolution and pad the rest with 0s. Our hacky solution is to return a smaller image and embed it in a new image of original size. Perhaps it is better for the ndimage library, but it seems like that doesn't have a simple function call for that behavior either.

[rgommers]

Perhaps it is better for the ndimage library, but it seems like that doesn't have a simple function call for that behavior either.

Indeed. No one has asked for it before. The reason I think is simply that filling with zeros is arbitrary, and should be strictly worse than using ndimage.convolve with mode='constant', cval=0. That way you get results that are between zero and the "right" answer.

[rfezzani]

What about returning the values of the input array unchanged where the convolution is not 'valid'?