Page 1 of 1

If you want to try multiple trainers, do you use a different model dir or the same one? Do they help each other?

Posted: Mon Jan 09, 2023 11:40 pm
by unkempt

And if you do, do you choose which to use for conversion or does it figure that out on its own?


Re: If you want to try multiple trainers, do you use a different model dir or the same one? Do they help each other?

Posted: Mon Jan 09, 2023 11:43 pm
by unkempt

So I answered part of the question. Adding a different trainer just created new files in the same folder with the name of the trainer. It seems to be starting over so we'll see what happens. I'll try converting with one and then the other (if it gives me the choice). Otherwise maybe conversion blends them? I don't know yet... we'll see.


Re: If you want to try multiple trainers, do you use a different model dir or the same one? Do they help each other?

Posted: Tue Jan 10, 2023 11:47 am
by torzdf

Models are not compatible with each other. When you create a model you create a "graph" which is fundamentally fixed. Changing the model or settings to a model won't change the existing model.

You should keep different models in different folders to avoid complication further down the line.


Re: If you want to try multiple trainers, do you use a different model dir or the same one? Do they help each other?

Posted: Thu Jan 12, 2023 12:58 am
by unkempt

Fair enough (and matches my experimentation thus far). I'm going to try training on most/each of the varients on the same input to see which does better for the scenario I'm trying to do. Wish me luck (so far, H128 seems best).