Neural Network Compression and Acceleration by Federated Pruning

In recent years, channel pruning is one of the important methods for deep model compression. But the resulting model still has tremendous redundant feature maps. In this paper, we propose a novel method, namely federated pruning algorithm, to achieve narrower model with negligible performance degradation. Different from many existing approaches, the federated pruning algorithm removes all filters in the pre-trained model together with their connecting feature map by combining the weights with the importance of the channels, rather than pruning the network in terms of a single criterion. Finally, we fine-tune the resulting model to restore network performance. Extensive experiments demonstrate the effectiveness of federated pruning algorithm. VGG-19 network pruned by federated pruning algorithm on CIFAR-10 achieves 92.5% reduction in total parameters and 13.58x compression ratio with only 0.23% decrease in accuracy. Meanwhile, tested on SVHN, VGG-19 achieves 94.5% reduction in total parameters and 18.01x compression ratio with only 0.43% decrease in accuracy.