FPC: Filter pruning via the contribution of output feature map for deep convolutional neural networks acceleration

Pruning is a very effective solution to alleviate the difficulty of deploying neural networks on resource-constrained devices. However, most of the existing methods focus on the inherent parameters of the network itself, but rarely consider the contribution of the output feature map. In this paper, we propose the FPC, a novel filter pruning method based on the contribution of the output feature map, which considers the diverse information carried by different output feature maps. According to the above characteristic, FPC can evaluate the contribution of output feature maps and then effectively delete low contribution part without reducing the performance of the model. In this paper, we firstly use Singular Value Decomposition (SVD) to decompose the output feature map. Then we analyze the contribution of the output feature map to the model performance. Finally, we delete the filters with lower contribution output feature maps. Extensive experimental results show that our proposed FPC can produce excellent compression results. For example, with VGG-16, we can reduce the FLOPs by 65.62% and increase the accuracy by 0.25% on CIFAR-10. With ResNet-110, we can reduce FLOPs by 50.66% and increase the accuracy by 0.09% on CIFAR-100.