Wall-clock performance issues

[qyllan]

LambdaResNet50 in comparison to a normal ResNet50 from my testing is 7x slower. I am training on ImageNet, and I am comparing your LambdaResNet50 to a torch.hub ResNet50.

Is this to be expected?

[Myeongjun Kim]

Perhaps the torch.einsum computation time is too slow.
I will find another way to reduce the computation time.

Thanks

[qyllan]

yeah, that was the conclusion i came to as well. it's significantly faster than torch's einops still.