d-li14/octconv.pytorch

Computation time around 3 times longer with OctConv

niam06 opened this issue · 3 comments

niam06 commented

Thanks for your implementation of OctConv!

I see that the OctConv paper reports reduced GFLOPs usage as well as reduced computation time per image. However, running your implementation on ImageNet dataset, it seems the computation time per image actually increases by a multiple of 3! This happens both during training and during validation. I am comparing PyTorch's ResNet50 vs your OctResNet-50 with default alpha parameters.

d-li14 commented

duplicate of #4

  • 👍1
niam06 commented

Thanks!

niam06 commented

Continuing on my experiments on ImageNet, I tried testing OctConv Resnet50 on CPU, and it seems that it has runtime around 25% higher to PyTorch's vanilla ResNet50 (~0.79 seconds with ResNet50 vs. ~1.1 seconds with OctConvResNet50 for each 160 images). I am using default alpha_in and alpha_out at 0.5 each. Any thoughts on this? Thanks!