Guidance Needed for Selecting Best Epoch/Weights in Pix2PixHD Training

[yagobilel]

Hello,

I've been training a Pix2PixHD model for an image-to-image translation task and have reached a point where I need some advice on selecting the best epoch/weights based on the training metrics. My main goal is to achieve high-quality image generation that closely matches the target domain while maintaining a good balance in the adversarial training process.

Here are the types of metrics I'm observing at various epochs (example from epoch 1):

Epoch: 1
Iters: 3200
Time: 0.021
G_GAN: 1.442
G_GAN_Feat: 9.491
G_VGG: 7.601
D_real: 1.121
D_fake: 1.027
I understand that lower G_GAN, G_GAN_Feat, and G_VGG values are desirable as they indicate that the generator is effectively fooling the discriminator and generating images that are close in feature and perceptual quality to the target images. However, I'm unsure how to balance these metrics optimally and what to consider an acceptable range for D_real and D_fake to ensure the discriminator is functioning effectively without overpowering the generator.

Questions:
1-How should I interpret the balance between G_GAN, G_GAN_Feat, and G_VGG losses in selecting the best epoch for my model?
2-What are the ideal ranges or balance for D_real and D_fake losses to ensure effective learning without causing the discriminator to overpower the generator or vice versa?
3-Beyond the numeric metrics, what visual quality aspects should I prioritize in the generated images when manually reviewing them for selecting the best model weights?
4-Are there any community recommendations or best practices for monitoring and deciding on the best epoch/weights in Pix2PixHD or similar GAN-based models?
Any insights, recommendations, or references to relevant resources would be greatly appreciated. Thank you for your time and assistance.

Best regards,