Fast Sparse ConvNets (opens in new tab)

(arxiv.org)

4 pointsMarat_Dukhan6y ago1 comments

1 comments

1 comments · 1 top-level

TLDR; Make 1x1 convolutions sparse, write fast Sparse Matrix Multiplication kernels, get a nearly 2x speedup with smaller models.

j / k navigate · click thread line to collapse