Conv1D much slower than linear layers on CPU?

Hi everyone,

I had a simple autoencoder implemented with fully-connected linear layers and I tried switching them to Conv1d layers. The number of parameters went down by two orders of magnitude, but the compute time on the CPU went up by more than 10x. Anyone know why this would be the case? Why are conv layers so slow, at least on the CPU? Am I doing this correctly?

1 Like