SSFP: A Structured Stripe-Filter Pruning Method for Deep Neural Networks

Network pruning has gained popularity for reducing storage and computational requirements in deep neural network models. However, common pruning techniques focus on the single granularity without sufficiently exploring collaborative research on different granularity levels. In order to explore the combination of different pruning granularities while effectively compressing the network in a structured manner, this paper proposes a structured stripe-filter pruning method (SSFP) that combines filter and stripe pruning. The fine-grained is embedded within the architecture of coarse-grained pruning, employing a mixed compression strategy with two pruning modules to accomplish pruning tasks. Experiments demonstrate better model compression results than the comparison techniques in the VGG-16 and RESNET-56 convolutional neural networks. When applying SSFP to VGG-16, FLOPs is reduced by 70.1%, the total number of parameters is reduced by 93.5%, and the accuracy of CIFAR-10 is improved from 93.40% to 93.63%.