FB-SKP: Front-Back Structured Kernel Pruning via Group Convolution without Increasing Memory

Recently, deep convolutional neural network (DCNN) models have achieved remarkable results in the field of computer vision; however, as the size of the network has gradually increased, the application of compression techniques has become essential for practical use. Among compression methods, pruning is a widely used technique, and weight pruning and filter pruning are mainly being studied. Weight pruning is a representative unstructured pruning technique with high compression rate; however, it has a limitation in that it is difficult to lead to real network acceleration. Filter pruning is a structured pruning technique with high acceleration performance because it directly removes parameters; however, it has the limitation of having a relatively low compression rate. Kernel pruning is an intermediate technique and has the advantages of both methods but has the problem of being difficult to implement as structured pruning. To solve this problem, this paper proposed Front-Back Structured Kernel Pruning (FB-SKP). FB-SKP can implement structured kernel pruning without increasing network memory through two convolution groups. Additionally, to avoid increasing the complexity of pruning implementation, the filter was divided into two parts, front and back, and one of the two parts was removed. As a result, 45.2% and 47.4% of parameters were removed from ResNet56 and ResNet110, respectively, without performance degradation on CIFAR-100.