Fixing seed?

[rjgoif]

When I was doing a lot of experimentation with Kohya Lora training, I learned that a lot of what I thought were important settings changes were really just random variation in training results based on the seed. I fixed my seeds and it resulted in more controlled scientific experiments on training parameters.

Is there a way to do this in Onetrainer? I don't see line for "seed" in the training config files. If there isn't a current option, is this something that can be implemented in the future?

Love the program, great work!

[SirTrippsalot]

OneTrainer has a fixed seed.

[rjgoif]

Good to know, but conversely, how can we test seed effect to determine consistency? Also, what is the fixed seed?

Is there any way to override the seed setting to vary it, either randomly or to another fixed value?

I feel like this is a valid parameter for iterative testing.

[SirTrippsalot]

Best to join the discord for questions.

I believe 42
No, but we've been discussing it