Filter Pruning via Structural Similarity Index for Deep Convolutional Neural Networks Acceleration

Abstract

Pruning can reduce the size of the model without reducing its performance. After pruning, the model can run in a small terminal flexibly. This paper proposes a new filter pruning method that uses soft filter pruning via a structural similarity index(FPSSI) to compress and prune the network. FPSSI uses the structural similarity index to measuring the difference between different filters, the filters with similar structures are pruned to achieve the purpose of compressing the Deep Convolutional Neural Networks(DNN) model. Compared to the norm-based approach to remove "relatively low" importance filters, the proposed method takes into account the structure between the filters. When applied to the different classification benchmarks, our method validates its usefulness and advantages. In CIFAR10, the ResNet network uses the SFP-SSIM method to reduce 52% of FLOPs and has better accuracy.