Exploring the Impact of Hyperparameters on the Generation Quality of CycleGAN

Abstract

Cycle-Consistent Generative Adversarial Networks (CycleGAN) has demonstrated remarkable proficiency in its ability to undertake image transference between differing domains without the need for paired examples, overcoming a major limitation of traditional image-to-image translation methods. These advantages make it a valuable addition to the toolbox of computer vision researchers and practitioners. However, there persists ample potential for further advancements. This article explores the effectiveness of refining the loss function and training parameters of CycleGAN, aiming to explore how these adjustments affect the outcomes. Specifically, this work plans to improve the loss weight allocation, balancing the translation from different domains. Additionally, the author also explores the number of training epochs and learning rate, halting the process at specific intervals to observe the impact of gradually decreasing the learning rate, as opposed to reducing it to zero. To assess the enhanced performance of CycleGAN, the author will employ both direct human visual perception and cycle consistency loss as evaluation metrics.