Questions about using PSNR , SSIM as metric in vanilla-GAN in training and test time

Soheil T 0
Viewed 15

Suppose that we have a vanilla-GAN training on MNIST dataset :

  1. How could we use SSIM or PSNR as evaluation metrics(not loss function) in training time? should we average on PSNR(fake,real) during loop?(I mean that fake image compares with only single paired-related real image)

  2. How could we use PSNR or SSIM in test time in a vanilla-GAN network? because GAN acts unsupervised and in test time we don't have related real image and seems that it is not good idea to compare each "fake image (generated in test stage)" with all dataset

Thank you for helping me,

generative-adversarial-network deep-learning machine-learning
0 Answers
Related