Rule of thumb for kimg?

[lebeli]

I have a dataset with ~4000 images (CAD images without any noise). Is there a rule of thumb for choosing the kimg hyperparameter? Also, the kimg hyperparameter is basically how one can controll the number of epochs, right?

[thinkercache]

General understanding is : 1kimg =1000inmg. This means 1000 images are shown to the network during the training. According to my experience, I would suggest 4000kimg (--kimg=4000) in the training configuration, as a good starting point to observe how the G and D behaves. After that, you may go for lesser kimg or higher, depends on your dataset. To satisfy the built-in plugins in PyTorch, kimg is used.

[lebeli]

Thank you, for the explanation. What's a good metric to observe both the generator and discriminator? FID for the generator and logits for the discriminator? Or simply logits for both?

[thinkercache]

@lebeli KID is good for small datasets as the original KID paper (Demystifying MMD GANs Mikołaj Bińkowski, Danica J. Sutherland, Michael Arbel, Arthur Gretton https://arxiv.org/abs/1801.01401) suggests. FID is widely used and good fit for large datasets such as over 10k images, according to my understanding at the moment. One metric for both.