Best way for fixing overfit model

[dgrcode]

If I used too many steps with dreambooth (overfit model). Can I remove steps? or do I have to start from scratch?

[TheLastBen]

I will add a solution to that soon, but for now, you can do it manually by replacing the text_encoder "bin" file with the original one. then resume training for a small amount of steps (500) with the text_enc set to 100%

[dgrcode]

Thanks for the super fast response! 🙌

I'll try that tomorrow, since Google told me I'm out of resources for today haha

So that method would be equivalent to 500 steps? Or removing 500 steps? I'm a bit lost with the internals. Even if you can point me to documentation about steps/text_encoding/etc would be super helpful.

Thanks!

[TheLastBen]

500 steps would be the average steps that the tex_enc should be finetuned for, because if you replace the existing text_enc with the non-trained one, you will need to train it, but this time at lower steps to avoid overfitting again.