Editorial for Pattern Recognition Letters Special Issue on Face-Based Emotion Understanding
Jingting Li,Moi Hoon Yap,Wen-Huang Cheng,John See,Hong Xiao,Xiaobai Li,Sujing Wang
DOI: https://doi.org/10.1016/j.patrec.2023.02.022
IF: 4.757
2023-01-01
Pattern Recognition Letters
Abstract:Network pruning extends a promising prospect to neural network compression. However, applications of existing methods to various vision tasks, e.g., human pose estimation, are limited by the parameter distribution difference, shortage of general pruning strategies and high pruning cost. As a complex and widely used backbone network, HRNet is an effective benchmark for studying these issues. We propose an adaptive Hard channel pruning method with an Adaptive Compression Rate to replace manual experience in pruned model selection, called HACR. It first designs a hard criterion that takes γ and β to measure channel importance with the insight of the consistency relationship between the distribution adjustment of the channel and the final output distribution. Then, a partial pruning strategy is advanced for the complex structural network by only pruning channels without cross-layer integration and keeping the network infrastructure. Next, we adaptively predict an optimal compression rate based on the validation metric of unfine-tuned models, providing a reference for compression parameter selection. The processes of HACR are simplified to adaptive pruned model selection and fine-tuning with lower pruning cost. Furthermore, the numerical distribution of γ and β is used to extend HACR in common pruning task benchmarks, based on the a principle that lower level channels should be preferentially retained. We conduct HRNet pruning experiments on COCO2017 and MPII, and promote extension study on CIFAR-10 to verify the generality of strategy. The results show that HACR is effective on multiple tasks, e.g. HRNet-W48 reduces 58.2% Params and 50.1% FLOPS, and VGG-16 reduces 91.5% Params and 67.5% FLOPS with only 0.3% precision loss.