Research On Fine-Grained Model Pruning Algorithm

Deep neural networks have made breakthroughs in the fields of computer vision,speech,natural language processing and etc.With the development of deep learning algorithms,the demand is dramatically increasing to deploy neural network applications on cloud,device and edge scenarios.Neural network model is both computation and memory intensive,and it poses challenge to the deployment on mobile devices where resources and power consumptions are strictly constrained.Fine-grained model pruning can simultaneously obtain a high compression rate and negligible loss of accuracy,which makes it easier to deploy neural networks.This paper conducts research on the existing pruning methods.To overcome the problems,such as how to properly set hyper parameters which are hard to be optimized and how to further improve compression rates,an automatic fine-grained pruning framework and a pruning algorithm based on L1/2 penalty are separately proposed.To overcome the first problem,the key steps in pruning algorithm are abstracted and the hyper parameters are automatically optimized based on reinforce learning.Based on the ideas mentioned above,an automatic fine-grained pruning framework is proposed.Compared with existing methods,the proposed method can further improve compression rates with an automatic procedure.By introducing the L1/2 penalty to improve the sparsity of the pretrained model,the negative effect caused by incorrect pruning is effectively reduced.Compared with existing methods,the proposed method further improves compression rates and achieves state-of-the-art compression rates on several models.A software/hardware co-design optimization for memory-fetching system is also proposed to utilize the sparsity of the pruned model.Experimental results verify the importance of fine-grained pruning algorithm to hardware acceleration.