Feature Variance Ratio-Guided Channel Pruning for Deep Convolutional Network Acceleration

Junjie He, Bohua Chen, Yinzhang Ding, Dongxiao Li; Proceedings of the Asian Conference on Computer Vision (ACCV), 2020

Most existing channel pruning approaches utilize the magnitude of network parameters to guide the pruning process. However, these methods suffer from some limitations in modern networks, where the magnitude of parameters can vary independently of the importance of corresponding channels. To recognize redundancies more accurately and therefore, accelerate networks better, we propose a novel channel pruning criterion based on the Pearson correlation coefficient. The criterion preserves the features that are essentially informative to the given task and avoids the influence of useless parameter scales. Based on this criterion, we further establish our channel pruning framework named Feature Variance Ratio-guided Channel Pruning (FVRCP). FVRCP prunes channels globally with little human intervention. Moreover, it can automatically find important layers in the network. Extensive numerical experiments on CIFAR-10 and ImageNet with widely varying architectures present state-of-the-art performance of our method.