Channel pruning guided by global channel relation

Channel pruning approaches have achieved some success in compressing deep convolutional neural networks (CNNs). However, existing methods ignore the interdependence among channels within the same layer. In this paper, we propose a novel channel pruning approach called Channel Pruning guided by Global Channel Relation (CPGCR), which additionally takes the global channel relation into account in the channel pruning process. Considering that Squeeze-and-Excitation (SE) blocks have the ability to encode the channel relation, our method is mainly used to compress the CNNs with SE blocks. We also observe that SE blocks will enforce channel-level sparsity in the network, which is useful for the implementation of channel pruning algorithms. Extensive experiments with a variety of neural networks on five datasets clearly demonstrate the effectiveness of our proposed CPGCR method. The results show that on ImageNet, our method gives a 56% reductions in float-point-operation (FLOPs) for ResNet-50. On CIFAR-10, the CNNs compressed by CPGCR achieve comparable accuracy to that of the original models, but with significant reductions in FLOPs (61.7% for ResNet-56, 78.0% for VGG-16).