Filter pruning via annealing decaying for deep convolutional neural networks acceleration

Filter pruning has been used on a large scale to compress and accelerate convolutional neural networks. The goal of filter pruning is to find the optimal network substructure from the unpruned network. Most previous pruning methods directly remove the redundant filters from the network or set them to zero. It is not reasonable because these redundant filters still contain information. If these filters are removed directly, the performance of the model may be drastically reduced. To solve this problem, this paper proposes a new filter pruning method, namely Filter Pruning via Annealing Decaying (FPAD), for a fast and efficient search of the optimal substructure. Our proposed FPAD effectively preserves the pre-trained information of filters in the pruning process. In addition, FPAD can improve the convergence speed of the model to achieve the compression target faster. We introduce an annealing function to control the amount of filter decay during the pruning process. To demonstrate the validity of our method, we apply FPAD to three image classification benchmarks. The results show that FPAD outperforms the state-of-the-art pruning methods. In particular, on ILSVRC-2012, our FPAD reduces 42.2% FLOPs in ResNet-50 with 0.17% loss of top-1 accuracy and with only 0.03% loss of top-5 accuracy.