Group Pruning with Group Sparse Regularization for Deep Neural Network Compression

Network pruning is important for the deployment of deep neural network on hardware platforms. However, most pruning methods focus on coarse-grained pruning with high precision loss. In addition, many fine-grained pruning methods focus the pruning on the fully connected layer rather than the convolution layer. The group pruning technique is proposed, focusing only on CONV-layer that enables to keep the weight reduction rate consistent within each weight group. This helps solve some inefficiency problems including internal buffer misalignment and load imbalance after fine-grained pruning. In the pre-training, the strategy of group sparse regularization (GSR) and standardizing the weight distribution is added to alleviate the loss of precision under high sparsity. Finally, we used the MNIST and CIFAR-10 data sets to test the LeNet and VGG-16 for experiments. The reduction rates of convolution layer are 87.5% and 62.5% respectively in the range of 0.14% accuracy loss.