Supervisors: Dr. Zhiwu Huang, Prof. Luc Van Gool
The network design generally impacts the performance of Generative Adversarial Nets (GAN) for image generation. However, traditional GAN architecture designs are often merely hand-crafted which may lead to a locally optimized solution. To overcome this drawback, the semester project aims to introduce an efficient Differentiable Architecture Search (DARTS) algorithm to automatically search for GAN's generator and discriminator architectures. In particular, we exploit the methodology of DARTS to parameterize the generator and discriminator architecture operations in a differential formulation. Based on such formulation, we propose a four-level optimization to update the operation probabilities as well as their parameters. The effectiveness and efficiency of the proposed AutoGAN model are validated on the image generation task on CIFAR-10. The experiments show that our proposed model expresses a high architecture search efficiency (takes less than 20 GPU hours ) and achieves a highly competitive image generation result (FID 11.9 and IS 8.49), compared to the state-of-the-art hand-crafted GAN models and the existing auto-searched GAN.