Compressing neural networks by pruning weights with small magnitudes can significantly reduce the computation and storage cost. Although pruning makes the model smaller, it is difficult to get practical speedup in modern computing platforms such as CPU and GPU due to the irregularity. Structural pruning has attract a lot of research interest to make sparsity hardware-friendly. Increasing the sparsity granularity can lead to better hardware utilization, but it will compromise the sparsity for maintaining accuracy. In this work, we propose a novel method, TETRIS, to achieve both better hardware utilization and higher sparsity. Just like a tile-matching game, we cluster the irregularly distributed weights with small value into structured groups by reordering the input/output dimension and structurally prune them. Results show that it can achieve comparable sparsity with the irregular element-wise pruning and demonstrate negligible accuracy loss. The experiments also shows ideal speedup, which is proportional to the sparsity, on GPU platforms. Our proposed method provides a new solution toward algorithm and architecture co-optimization for accuracy-efficiency trade-off.

This paper deals with sparsifying the weights of a neural net to reduce memory requirements and speed up inference. Simply pruning small weights yields unstructured sparsity, which is hard to exploit with standard libraries and hardware. This paper imposes block sparsity, where each weight tensor is divided into fixed blocks (of size 32 x 32, for example) and non-zero weights are specified in only a fraction of the blocks. The paper's innovation is an iterative algorithm for reordering the rows and columns of a tensor to group together the large weights, reducing the accuracy loss from block pruning. Experiments on the VGG16 network for ImageNet show that this method achieves better speed-accuracy trade-offs than either unstructured weight pruning or block pruning without reordering. Update: I appreciate the authors' responses.

Compressing neural networks by pruning weights with small magnitudes can significantly reduce the computation and storage cost. Although pruning makes the model smaller, it is difficult to get practical speedup in modern computing platforms such as CPU and GPU due to the irregularity. Structural pruning has attract a lot of research interest to make sparsity hardware-friendly. Increasing the sparsity granularity can lead to better hardware utilization, but it will compromise the sparsity for maintaining accuracy. In this work, we propose a novel method, TETRIS, to achieve both better hardware utilization and higher sparsity.