This seems too good to be true, setting min_snr_gamma = 5 has been giving really good sample images at the first epoch, and they only get better. The loss rate goes very low even with large diverse image sets that had higher loss rates when I've trained with them before.
Should I lower the num_epochs or dataset_repeats when using this feature?
This seems too good to be true, setting
min_snr_gamma = 5
has been giving really good sample images at the first epoch, and they only get better. The loss rate goes very low even with large diverse image sets that had higher loss rates when I've trained with them before.Should I lower the
num_epochs
ordataset_repeats
when using this feature?