Progressive Kernel Pruning Based on the Information Mapping Sparse Index for CNN Compression

Network pruning can effectively reduce a model's capacity and computational load, thereby making model deployment in mobile devices less difficult than that without pruning. To improve the rate of while maintaining kernel's feature extraction ability, this paper designs progressive kernel method for CNN compression based on proposed information mapping sparsity index. This first prunes kernels filter then convolution layer when reaches certain ratio. The whole process is called (PKP). For process, defines sparse index (IMSI), which used to measure ability related amount transferred by operation. When filters layers, according IMSI, with strongest abilities are retained transfer as much possible least number kernels. Progressive make use characteristic easy optimize filter, it avoids having easily fall into local optima directly. experimental results CIFAR-10/100 ImageNet datasets show compared existing methods, IMSI-based exhibits better performance processing tasks currently popular. In particular, VGG-16 CIFAR-10 our achieves ratio 80.8x an acceleration 14.8×, 5.8× 4.2× higher best at present, respectively, classification accuracy decreases only 0.59% relative baseline.