octconv-pytorch icon indicating copy to clipboard operation
octconv-pytorch copied to clipboard

Published 20 hours ago • verified publisher icon iacolippo

Strided convolutions not implemented

Open andravin opened this issue 6 years ago • 2 comments

It appears that the stride parameter does not have the desired effect in this implementation of OctConv. It is used to compute the padding, but otherwise has no effect on the convolution (ie does not cause downsampling of the feature map).

andravin avatar Apr 21 '19 04:04 andravin

For reference, see https://github.com/terrychenism/OctaveConv/issues/4

andravin avatar Apr 23 '19 18:04 andravin

Hi Andrew, thanks for opening the issue. You are correct.

I will have a look at the issue you linked and will try to fix this. Matching the upsampled and the downsampled maps when the stride is different from one and the filter size is even is quite a puzzle. I'm curious to see how it's handled when the authors will release their implementation.

iacolippo avatar Apr 23 '19 19:04 iacolippo