Difference of computational efficiency between Conv1D and Conv2D

The original network is constructed by Con1D, while Conv1D can’t be supported in edge devices. So I use Conv2D to implement Conv1D and change some operators’ parameters. Conv2D took much more time to train for a epoch (9mins vs 2hs).

I don’t know which device and backend you are using, but e.g cuDNN uses the same approach of unsqueezing a spatial dimension as seen here while e.g. MKL seems to natively support 1D convs.