Re-GAN: Data-Efficient GANs Training via Architectural Reconfiguration

Figure 1: Results of our proposed Re-GAN where dynamically GANs architecture is reconfigured to explore different GANs subnetwork structures during training time. (left) Image generation trained on multiple few-shot generation datasets, such as 100-shot Obama [1], Panda [1], and Animal Face-Cat (A-Cat) [2]; (right) FID scores vs. CIFAR-10 [3] training set size with training cost, FLOPs. Best viewed in color.