SDWD: STYLE DIVERSITY WEIGHTED DISTANCE EVALUATES THE INTRA-CLASS DATA DIVERSITY OF DISTRIBUTED GANS

Due to the distributed storage of massive data, efficient deployment of Generative Adversarial Networks (GANs) in distributed scenarios has become a hot topic. This paper analyzes the characteristics of existing distributed GANs. Moreover, it discusses the evaluation methods of these models. Through experiments, we found that Frechet Inception Distance (FID) cannot sensitively reflect the diversity of intra-class data, and the evaluation effect of different distributedGANs is greatly affected by different datasets, so we cannot accurately rank the models. To more accurately evaluate the generation capability of distributed GANs, we propose a novel evaluation model, Style Diversity Weighted Distance (SDWD), which introduces a Siamese network to measure data similarity. We conduct experiments on two datasets, and the results prove that SDWD can sensitively reflect the intra-class diversity of data and can reliably evaluate different models and frameworks.