Scholarlink

<p>Training Generative Adversarial Networks (GANs) with an efficient optimizer can improve image generation quality in an unsupervised setting. Most existing networks employ first-order derivative optimizers due to their fast computation, but these optimizers rely on short-term optimization with a fixed learning rate. In this paper, a novel two-phase switching optimizer known as Stochastic Diagonal Approximate Greatest Descent (SDAGD) by incorporating a decaying factor is proposed to optimize XingGAN network for pose transfer applications. The optimizer adaptively controls the relative step length to reach the greatest descent at the boundary of each local spherical search region. When the relative step length approaches zero, it switches to Newton’s method to achieve the optimal solution at fast descending speed. The incorporation of decaying factor can ensure stability of the relative step length and guarantees convergence to the optimal solution. To evaluate performance, the proposed optimizer is compared with the benchmark Adam optimizer and the vanilla SDAGD in term of quantitative and qualitative results. A very small decaying factor shows negligible impact on step length decay, while a very large one triggers premature switching to Newton’s method before reaching the optimum. With a decaying factor of 0.3, the proposed optimizer outperforms Adam by 0.270%, 5.473%, 0.702%, and 8.335% in SSIM, IS, PCKh, and FID, respectively. It also surpasses the vanilla SDAGD by 0.405%, 7.979%, 2.135%, and 22.929% in the same metrics. In addition, the qualitative results demonstrate that SDAGD with a decaying factor of 0.3 produces better appearance and shape consistency in the pose transfer applications.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Decaying Factor in Two-Phase Switching Optimization Strategy for Pose Transfer Network

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2025年10月20日

查看全文
  • King Hann Lim,
  • Zong Qi Ooi

摘要

Training Generative Adversarial Networks (GANs) with an efficient optimizer can improve image generation quality in an unsupervised setting. Most existing networks employ first-order derivative optimizers due to their fast computation, but these optimizers rely on short-term optimization with a fixed learning rate. In this paper, a novel two-phase switching optimizer known as Stochastic Diagonal Approximate Greatest Descent (SDAGD) by incorporating a decaying factor is proposed to optimize XingGAN network for pose transfer applications. The optimizer adaptively controls the relative step length to reach the greatest descent at the boundary of each local spherical search region. When the relative step length approaches zero, it switches to Newton’s method to achieve the optimal solution at fast descending speed. The incorporation of decaying factor can ensure stability of the relative step length and guarantees convergence to the optimal solution. To evaluate performance, the proposed optimizer is compared with the benchmark Adam optimizer and the vanilla SDAGD in term of quantitative and qualitative results. A very small decaying factor shows negligible impact on step length decay, while a very large one triggers premature switching to Newton’s method before reaching the optimum. With a decaying factor of 0.3, the proposed optimizer outperforms Adam by 0.270%, 5.473%, 0.702%, and 8.335% in SSIM, IS, PCKh, and FID, respectively. It also surpasses the vanilla SDAGD by 0.405%, 7.979%, 2.135%, and 22.929% in the same metrics. In addition, the qualitative results demonstrate that SDAGD with a decaying factor of 0.3 produces better appearance and shape consistency in the pose transfer applications.