A Filter Rank Based Pruning Method for Convolutional Neural Networks

The use of convolutional neural networks which achieved success in many area nowadays is hardly applicable in embedded devices directly because of the high computation and storage cost. Many previous methods have been proposed to reduce the FLOPs and parameters of network. However, many methods aren't effective enough and suffer the loss of accuracy. In this paper, we propose a novel pruning method according to filter rank. Comparative experiments are carried out to validate the effectiveness. Results show that it can achieve 93.55% accuracy with VGG-16 on CIFAR-10 pruning 49.48% FLOPs and 81.17% parameters. Besides, we reduce 50.02% FLOPs and 42.35% parameters of ResNet-56 on CIFAR-10 achieving 92.56% accuracy.