Multi-GANs with Shared Generator: An Approach for Handling Mode Collapse Issue

The Generative Adversarial Network (GAN) is a highly effective member of the generative models' category and is widely utilized for generating realistic samples across various domains. The fundamental concept behind GAN involves two networks - a generator and a discriminator, engaged in a competitive process. During the training process, issues arise that can potentially impact the quality and diversity of the generated samples produced by GAN. One significant issue is mode collapse, where the generator fails to create diverse samples. In order to tackle this problem, we propose an approach known as Multi-Generative Adversarial Networks with Single Generator (MultiGAN-SG), which involves pitting a single generator against multiple discriminators within different GAN models. In this approach, GAN models are categorized into two types: master GAN and subsidiary GANs. During training, the master GAN updates the weights of the subsidiary GANs, hence updating the generator's weights. The purpose of this approach is to prevent the generator from receiving fixed feedback from a discriminator. The proposed approach has been validated using four distinct benchmarks representing Mnist, Fashion-Mnist, Cifar-10, and CelebA. The outcomes of the experiments indicate that the proposed approach outperforms the state-of-the-art GAN models when assessed using the FID metric, which measures the diversity of the generated samples.