Why is original resnet50 faster than octave-resnet50?

[BlueAnthony]

Why is original resnet50 faster than octave-resnet50?
In inference on gtx-1080, original resnet50 is about 70 fps and octave-resnet50 is around 42 fps.

[miguelvr]

I noticed the same thing during training.

It actually takes less GPU memory, but training is slower.

Not sure if this is an issue from pytorch (static graph vs dynamic graphs) or it is related to the fact that the authors use TVM to compile the model.

[zimenglan-sysu-512]

i train a model of mask-rcnn using octave-resnet-50, and find that it's slower than resnet-50 and the performance is 3 points lower as well.

[peri044]

@BlueAnthony I've observed the same thing from my side here facebookresearch/OctConv#4

My guess is Oct Conv graph is not well suited on GPU (Could be due to two different branch structure, upsampling layers, more number of convolutions, more kernel launches and more mem copies). On CPU, Octave conv performance is better as reported in the paper.

[BlueAnthony]

@peri044 Yeah, I also try Octave Conv on CPU, it really has some improvement on CPU.
I agree with you. Maybe Oct Conv graph is not well suited on GPU.
And, thank you for ur recommendation! I will try AutoTVM.