ACP: Automatic Channel Pruning Method by Introducing Additional Loss for Deep Neural Networks

Channel pruning is one of the main methods of model compression for the deep neural network. Some of the existing pruning methods manually set parameters based on experience, which is very time-consuming, and pruning channels by greedy algorithms or heuristic algorithms will bring local optimal solutions. Some methods prune channels by automated methods, but the lack of theoretical guidance on network characteristics makes the pruning efficiency extremely low. In this article, we propose a new automated channel pruning method: by introducing additional losses into the network, to obtain the channel's ability to discriminate information and adopt an automated pruning method based on reinforcement learning to iteratively perform channel selection and parameter optimization and prune the filtered channels. A large number of experiments have proved the effectiveness of our method. For example, for ResNet-110, on CIFAR-10, our method reduces FLOPs by 63.1%, the parameters are reduced by 62.7%, and there is almost no loss in accuracy, only 0.01% lower than the baseline model. On ImageNet, FLOPs are reduced by 50.3%, The parameters are reduced by 55.2%, and the accuracy is only lost by 0.22%.