Filter Pruning Based on Coordination of Layers for Deep Neural Networks

Convolutional neural networks have achieved remarkable success in the field of computer vision. However, deploying them to edge devices is challenging due to their numerous parameters and computational complexity. In this paper, we propose a new network pruning method called Filter Pruning based on Coordination of Layers (FPCL), which reduces the size of the network by pruning some relatively unimportant filters while ensuring low accuracy loss. In order to maintain the coordination of the feature extraction by filters in the adjacent network layers, we take the influence of the filter on the loss function during the images of dataset passing through the network as the attribute of the filter, cluster filters in the adjacent layers, and then prune the filters in the cluster. During pruning, we ensure that filters from different layers within the same cluster are properly preserved. To avoid the loss of model generalization after pruning, we fine-tune the pruned model and then re-evaluate all filters, including the pruned ones, to prune again. The process of pruning and fine-tuning is iterated multiple times, with an increasing number of filters pruned each time, until the desired number of pruned filters is reached. To validate the effectiveness of the FPCL method, we pruned various models on multiple datasets. Compared to other advanced methods, the pruned model by FPCL achieves higher testing accuracy with fewer parameters.