Variational Automatic Channel Pruning Algorithm Based on Structure Optimization for Convolutional Neural Networks

Abstract

Channel pruning has achieved remarkable success in solving the significant computation and memory consumption in Convolutional Neural Networks (CNN). Most existing methods measure the importance of channels with manually designed algorithms and pruning unimportant channels rely on heuristics or expertise during the processing, which are labourious and subjective. In this paper, we proposed a Variational Automatic Channel Pruning Algorithm based on structure optimization (VA-CPSO) which can automatically optimize channel numbers via channel scales in end-to-end manner through variational inference. Firstly, a weights generator controlled by channel scales is built to produce weights for various pruned structure of CNN. And then, the channel scales with truncated factorized log-uniform prior and log-normal posterior are optimized by variational inference for optimal pruning structure. Meanwhile, parameters of the weights generator are optimized synchronously. Finally, the acquired optimal structure and corresponding generated weights are deployed in the pruned CNN for further training to achieve high-performance compression. The experimental results demonstrate that our proposed VA-CPSO acquires better compression performance compared to existing pruning algorithms. The VA-CPSO achieve a compression of 34.60×, 4.28×, 1.96× and a speedup of 28.20×, 2.03×, 2.02× on LeNet-5, VGGNet, and ResNet-110 with no more than 0.5% loss of accuracy.