PyTorch sparse convlution

Hi, did anyone worked with sparse convolutions in PyTorch? Should I expect a feed forward speed up increase when using a sparse cnn on gpu/cpu?


I don’t think such thing exists.
It would be tricky to implement efficiently and most likely will require a very very high level of sparsity to be faster for gpu.

Nvidia Ampere Architecture has supported the sparisity inference :smiley: