3 ms·
TLDR; Make 1x1 convolutions sparse, write fast Sparse Matrix Multiplication kernels, get a nearly 2x speedup with smaller models.
by ekelsen 7y ago
TLDR; Make 1x1 convolutions sparse, write fast Sparse Matrix Multiplication kernels, get a nearly 2x speedup with smaller models.