LDP: A Large Diffuse Filter Pruning to Slim the CNN

In recent years, filter pruning has become one of the most promising methods for CNN compression. However, most of the existing approaches need to prune filters iteratively, which increases the cost and complexity of the pruning process. In this paper, we propose a novel filter pruning strategy to conduct the removal operation only once, which avoids reconstructing new thinner models over and over. To accomplish this, we focus on the information amount of post-features, and replace the iterative removals with equally efficacious but simpler deactivation. The network repeatedly goes through filter deactivation and accuracy recovery, with more and more inefficient filters deactivated, just like a gas diffusing in a container. Therefore the proposed method is named as a Large Diffuse Pruning (LDP). Besides, we simplify the entropy weight calculation procedures by eliminating the repetitive Min-Max Normalization. Extensive experiments on VGGNet and ResNet show that our approach could achieve higher or comparable performance than some of the state-of-the-art methods.