FGPGAN: a finer-grained CNN pruning via generative adversarial network

Model pruning has gained increasing attention in the field of network model compression. Although current methods combine GANs with model pruning, their pruning granularity is insufficient to achieve better compression ratios for models. Moreover, they cannot recognize similar filters, leading to redundancy and an inability to learn the optimal pruning structure. To address these issues, we propose a fine-grained generative adversarial network (GAN) for pruning models in this paper. Specifically, we utilize shape masks to learn the shape properties of the model to achieve larger compression ratios. Additionally, we employ mean-shift clustering to prune similar filters and tackle the redundancy issue they cause. Extensive experiments demonstrate the effectiveness of our method, which we validate on the CIFAR-10/100 and ILSVRC-2012 datasets. The results show that our approach outperforms state-of-the-art methods significantly.