An Empirical Comparison of Generative Adversarial Network (GAN) Measures

Abstract

Generative adversarial network (GAN) is the most potent unsupervised learning generative model in deep learning. Though numerous impressive results have been published in computer vision tasks, it is still complex to compare and validate the performance of GAN algorithms. In this work, the review of quantitative evaluation measures of GAN is done with the performance of Frechet inception distance (FID) and the inception score (IS). Evaluations of several recently proposed GAN approaches are based on these two metrics. These evaluations demonstrate an evident variation in their performance based on key factors like training model and hyperparameters such as dimensions of the latent space, learning rate, and gradient penalty. This work discovers the proper dimension of latent space and compares FID and IS that are implemented for evaluation of generated data distribution. FID and IS are the best metrics for evaluating generated data distribution. This work gives an emphasis on appropriate dimension of the latent space and compares these two metrics concerning the improved GAN models. The experimental analysis shows that FID gives better performance compared to IS. NS-GAN and LS-GAN perform more precise. The generator generates better results for 10-dimensional latent spaces, which are not really distinct from the consequence of the normal 100-dimensions. It is recommended to use LS-GAN foe better performance and understanding of algorithms.