Two-phase Switching Optimization for Generative Adversarial Network in Pose Transfer

Optimizer is crucial for training Generative Adversarial Networks (GAN) to generate images with enhanced image quality and achieve better evaluation metrics. The most widely used optimizers in GAN are the first-order derivative optimization methods due to their simplicity and efficient computation. However, these methods contain flaws such as fixed learning rate, slow convergence speed, and sensitivity to hyperparameter selection. In this paper, a two-phase switching optimization strategy known as Stochastic Diagonal Approximate Greatest Descent (SDAGD) is proposed in optimizing a pose transfer network named as XingGAN. To evaluate the optimization performance of SDAGD, the quantitative and qualitative results are compared with Adam. In terms of quantitative results, the XingGAN with SDAGD applied to the appearance-guided discriminator achieved 4.583% and 3.733% better PCKh and FID than Adam for 20% of the training dataset while 18.039% and 18.111% better PCKh and FID than Adam for 10% of the training dataset. Additionally, the qualitative results for XingGAN with SDAGD applied to the appearance-guided discriminator also achieved the best appearance and shape consistency.