Towards Compressing Efficient Generative Adversarial Networks for Image Translation via Pruning and Distilling

Luqi Gong, Chao Li, Hailong Hong, Hui Zhu, Tangwen Qian, Yongjun Xu

Abstract: Deploying GANs (Generative Adversarial Networks) for Image Translation tasks on edge devices is plagued with the constraints of storage and computation. Compared to some methods like neural architecture search (NAS), filter pruning is an effective DNN (Deep Neural Network) compressing method. It can compressing DNNs in a short time. The filter importance is measured by the filter norm, the filters with low norm are pruned. As for image classification, the filter with larger norm has larger influence on the final classification scores. However, as illustrated in Fig. 4, the filter with large norm don’t always have a big impact on the quality of generated images for GANs. Based on the observation that the filter close to the filters’ center in the same convolution layer can be represented by others in [8], we develop a distance-based pruning criterion. We prune the filters which are close to the filters’ center in a convolution layer. KD (Knowledge distillation) trains the compressed model and improves its performance. The most common KD method ignores the transformation information across the feature maps, which is important for GANs. We take them as additional knowledge and transfer it from the uncompressed GAN to the pruned GAN. Our experiments on CycleGan, Pix2pix, and GauGan achieved excellent performance. Without losing image quality, we obtain 51.68 $$\times $$ and 36.20 $$\times $$ compression on parameters and MACs (Multiply-Accumulate Operations) respectively on CycleGan. Our code (We will open source within one week after the paper being received) will be made available at github.