Abstract
Generative Adversarial Networks (GANs) have become the most used network models towards solving the problem of image generation. In recent years, self-supervised GANs are proposed to aid stabilized GAN training without the catastrophic forgetting problem and to improve the image generation quality without the need for the class labels of the data. However, the generalizability of the self-supervision tasks on different GAN architectures is not studied before. To that end, we extensively analyze the contribution of the deshuffling task of DeshuffleGANs in the generalizability context. We assign the deshuffling task to two different GAN discriminators and study the effects of the deshuffling on both architectures. We also evaluate the performance of DeshuffleGANs on various datasets that are mostly used in GAN benchmarks: LSUN-Bedroom, LSUN-Church, and CelebA-HQ. We show that the DeshuffleGAN obtains the best FID results for LSUN datasets compared to the other self-supervised GANs. Furthermore, we compare the deshuffling with the rotation prediction that is firstly deployed to the GAN training and demonstrate that its contribution exceeds the rotation prediction. Lastly, we show the contribution of the self-supervision tasks to the GAN training on loss landscape and present that the effects of the self-supervision tasks may not be cooperative to the adversarial training in some settings. Our code can be found at this https URL.
Abstract (translated)
URL
https://arxiv.org/abs/2011.01730
