Performance issues in Octave Convolution ResNet

vision
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?

Thanks!
Miguel

2

Hi

Were you able to resolve this?

I think with CUDA, the timings being measured by your code might not be accurate:
(https://github.com/braincreators/octconv/blob/oct-resnet152/benchmarks/benchmark.py#L49)

You would either need cuda synchronize events or the autograd profiler (How to measure time in PyTorch)

3

Hii
Can u please mention the challenges phased by octave convolution

4

Hii
can u please mention what is the methodology used for implementing octave convolution

5

You can check the implementation in the repository linked in the first post

6

Can u share any proposed diagram for octave convolution

7

it is also in the repository, but you should read the paper for that.

8

u didn’t mentioned any proposed diagram in the link

9

Can u please send the paper regarding octaveconvolution