HScore: A Target Attention-based Filter Pruning with Visualizing the Area of Feature Map

Abstract

Neural network pruning plays an important role in the deployment on resource-constrained devices by reducing the scale of the network and the computational complexity. Different from existing pruning methods that only consider the amount of information filters contain, we think the distribution of information is more important to the model accuracy. Then we use visualization of feature map to consider the impact of the attention area on the network decision. In order to directly relate the attention area with confidence score, we propose HScore to use interpolation and normalization to visualize the attention area of feature map and calculate the confidence score by putting only the attention area into the model. The principle behind is that if the feature map focuses on an area which can achieve a higher classification confidence, then the corresponding channel is important to keep the accuracy of model. The confidence score of attention area represents the importance of the channel to the model and filters with smaller confidences would be pruned. The experiment results prove that HScore can prune more redundant filters while maintaining better accuracy. Notably, our method can reduce 68.6% parameters and 73.0% FLOPs of ResNet-110 even with 0.14% top-1 accuracy increase on CIFAR-10. With ResNet-50, we achieve a 55% FLOPs reduction with only a loss of 1.03% in the top-1 accuracy on LSVRC-2012, which has advanced the state-of-the-art.