Fine Tuning Loss Validation

[madrooky]

Is there a way of validation in OneTrainer as explained in this article?

https://medium.com/@damian0815/fine-tuning-stable-diffusionwith-validation-3fe1395ab8c3

Can't find anything. And without such a validation it seems to me the Tensorboard is rather useless and the right amount of training a very vague undertaking.

[mx]

No, OneTrainer does not have built-in support for this sort of validation. It's a pretty good idea though, so a pull request to add it would be quite welcome.

[IndigoDosSantos]

I've proposed a way to implement validation loss (using FID scores) in PR #326 .

[Seedmanc]

I've seen a discussion on GH claiming that the loss reported for Loras is quite a useless param and it does not reflect the quality of learning. Instead they recommended to turn on Scale Weight Norms and watch the reported amount of keys scaled. When it goes above 0, it's a sign of overfitting. I can't vouch for or against this approach but it sounded reasonable.