Dynamic Block Sparse Reparameterization of Convolutional Neural Networks

Sparse neural networks are efficient in both memory and compute when compared to dense neural networks. But on parallel hardware such as GPU, sparse neural networks result in small or no runtime performance gains. On the other hand, structured sparsity patterns like filter, channel and block sparsity result in large performance gains due to regularity induced by structure. Among structured sparsities, block sparsity is a generic structured sparsity pattern with filter and channel sparsity being sub cases of block sparsity. In this work, we focus on block sparsity and generate efficient block sparse convolutional neural networks using our approach DBSR (Dynamic block sparse reparameterization). Our DBSR approach, when applied on image classification task over Imagenet dataset, decreases parameters and FLOPS of ResneXt50 by a factor of 2x with only increase of 0.48 in Top-1 error. And when extended to the task of semantic segmentation, our approach reduces parameters and FLOPS by 30% and 20% respectively with only 1% decrease in mIoU for ERFNet over Cityscapes dataset. To ease developments in this line of work, we open sourced our code on github (https://github.com/idharmateja/bsnn).