Group Fisher Pruning for Practical Network Compression

Network compression has been widely studied since it is able to reduce the memory and computation cost during inference. However, previous methods seldom deal with complicated structures like residual connections, group/depthwise convolution and feature pyramid network, where channels of multiple layers are coupled and need to be pruned simultaneously. In this paper, we present a general channel pruning approach that can be applied to various complicated structures. Particularly, we propose a layer grouping algorithm to find coupled channels automatically. Then we derive a unified metric based on Fisher information to evaluate the importance of a single channel and coupled channels. Moreover, we find that inference speedup on GPUs is more correlated with the reduction of memory6 rather than FLOPs, and thus we employ the memory reduction of each channel to normalize the importance. Our method can be used to prune any structures including those with coupled channels. We conduct extensive experiments on various backbones, including the classic ResNet and ResNeXt, mobilefriendly MobileNetV2, and the NAS-based RegNet, both on image classification and object detection which is under-explored. Experimental results validate that our method can effectively prune sophisticated networks, boosting inference speed without sacrificing accuracy. Equal contribution Shenzhen International Graduate School/Department of Electronic Engineering, Tsinghua University, Beijing, China Shanghai AI Laboratory, Shanghai, China SenseTime Research, Hong Kong, China Department of Statistical Science, University College London, London, United Kingdom Qing Yuan Research Institute, Shanghai, China. Correspondence to: Wenming Yang <yang.wenming@sz.tsinghua.edu.cn>. Proceedings of the 38 th International Conference on Machine Learning, PMLR 139, 2021. Copyright 2021 by the author(s). We use ”memory” to denote the number of elements in the output feature maps of all layers. The code will be available at https://github.com/jshilong/FisherPruning. 0 1 2 3 4 FLOPs (G) 66 68 70 72 74 76 78 To p1 A cc (% ) ThiNet SSS IE Meta GBN Ours