Performance issues in Octave Convolution ResNet

(Miguel Varela Ramos) #1

Hello pytorch community,

I’ve been working on an implementation of Octave Convolution in pytorch, and now I’m trying to benchmark it.

This new octave convolution layer is supposed to be a “drop-in” replacement for nn.Conv2d and according to the paper it should have better performance, both in accuracy and speed, than its vanilla counterpart.

In my implementation, I benchmarked the convolutions individually and the OctConv2d is indeed faster than nn.Conv2d. However, the same is not true for the ResNet implementation, which I modified from the original torchvision implementation.

The benchmark code can be found here

Any clue of why this may be?