Convolutional Neural Network Channel Pruning Based on Regularized Sparse

Abstract

Aiming at the problem that the parameters of the deep learning model are too large and the application is limited at the embedded end, a channel pruning model compression method based on channel sparsity is proposed. This method defines the sparsity function of each channel in the convolutional neural network. By calculating the sparsity index of each channel, the channel sparsity characteristics of the convolutional neural network are obtained, and the channel sparsity is used to perform channel pruning. Furthermore, the L2 regularization is added to the sparsity degree, and the channel function is constructed by combining with the sparsity function. The channel sparsity index is obtained through the channel function operation. This pruning method is applied to the three classic convolutional neural networks of VGGNet, GoogleNet and ResNet on the CIFAR-10 and CIFAR-100 data sets. While maintaining the accuracy of the model, the model sizes are compressed to 2.1MB, 1. 7MB and 0. 64MB respectively.