4 ms·Fast Sparse ConvNets4 points by Marat_Dukhan 7y agoekelsen 7y agoTLDR; Make 1x1 convolutions sparse, write fast Sparse Matrix Multiplication kernels, get a nearly 2x speedup with smaller models.