Filter Pruning via Automatic Pruning Rate Search

Qiming Sun, Shan Cao, Zhixiang Chen; Proceedings of the Asian Conference on Computer Vision (ACCV), 2022, pp. 4293-4309

Model pruning is important for deploying models on devices with limited resources. However, the searching of optimal pruned model is still a significant challenge due to the large space to be exploited. In this paper, we propose an Automatic Pruning Rate Search(APRS) method to achieve automatic pruning. We reveal the connection between the model performance and Wasserstein distance to automatic searching optimal pruning rate. To reduce the search space, we quantify the sensitivity of each filter layer by layer and reveal the connection between model performance and Wasserstein distance. We introduce an end-to-end optimization method called Pareto plane to automatically search for the pruning rate to fit the overall size of the model. APRS can obtain more compact and efficient pruning models. To verify the effectiveness of our method, we conduct extensive experiments on ResNet, VGG and DenseNet, and the results show that our method outperforms the state-of-the-art methods under different parameter settings.