Scholarlink

Generative adversarial networks, popularly known as GANs, have been finding profound use in different application areas. GAN models’ performance has always been of interest to researchers and programmers alike in a bid to devise more efficient and better trained models. The role of different hyperparameters in the network is a crucial study domain as its optimization often yields better performing model. This paper analyzes the influence of different parameters like the choice of batch size, requirement of normalization, choice of activation function, learning rate, and network optimizer. The paper trains different GAN variations on two publicly available datasets: the digit and fashion MNIST dataset. It can be observed from the experimental simulation that large batch sizes perform slightly better in terms of training time efficiency but the same is not true with regard to accuracy. It was also observed that other than smoothening the training process, batch normalization has no significant impact on the output of the model. The images generated through different GAN models have been compared based on two metrics, the Structural Similarity Index Metric (SSIM) and the Peak Signal-to-Noise Ratio (PSNR) and were found to be distinguishable.

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

An Empirical Analysis on the Effect of Different Parameters in Training a Generative Adversarial Network to Produce Realistic Images

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Shamama Anwar,
  • Nandan Tiwari

摘要

Generative adversarial networks, popularly known as GANs, have been finding profound use in different application areas. GAN models’ performance has always been of interest to researchers and programmers alike in a bid to devise more efficient and better trained models. The role of different hyperparameters in the network is a crucial study domain as its optimization often yields better performing model. This paper analyzes the influence of different parameters like the choice of batch size, requirement of normalization, choice of activation function, learning rate, and network optimizer. The paper trains different GAN variations on two publicly available datasets: the digit and fashion MNIST dataset. It can be observed from the experimental simulation that large batch sizes perform slightly better in terms of training time efficiency but the same is not true with regard to accuracy. It was also observed that other than smoothening the training process, batch normalization has no significant impact on the output of the model. The images generated through different GAN models have been compared based on two metrics, the Structural Similarity Index Metric (SSIM) and the Peak Signal-to-Noise Ratio (PSNR) and were found to be distinguishable.