Filter Pruning via Attention Consistency on Feature Maps

Abstract

Due to the effective guidance of prior information, feature map-based pruning methods have emerged as promising techniques for model compression. In the previous works, the undifferentiated treatment of all information on feature maps amplifies the negative impact of noise and background information. To address this issue, a novel filter pruning strategy called Filter Pruning via Attention Consistency (FPAC) is proposed, and a simple and effective implementation method of FPAC is presented. FPAC is inspired by the notion that the attention of feature maps on one layer is in high consistency of spatial dimension. The experiments also show that feature maps with lower consistency are less important. Hence, FPAC measures the importance of filters by evaluating the attention consistency on the feature maps and then prunes the filters corresponding to feature maps with lower consistency. The present experiments on various datasets further confirm the effectiveness of FPAC. For instance, applying VGG-16 on CIFAR-10, the classification accuracy even increases from 93.96% to 94.03% with 58.1% FLOPs reductions. Furthermore, applying ResNet-50 on ImageNet achieves 45% FLOPs reductions with only 0.53% accuracy loss.