[Feat]: Disable random flip by default

[dxqb]

Describe your use-case.

The default of using "Random flip" augmentation can be dangerous, because it hurts several use cases. Those who need it can enable it

What would you like to see as a solution?

Disable default

Have you considered alternatives? List them here.

No response

[hyppyhyppo]

That would be a quick win:

• Low effort I guess.
• This setting can be really dangerous and many users falls into that trap.

[ciscokid009]

maybe it's an idea to pop up warning/hints after the start training push, to avoid traps. OT has many options which can have dependecies one might not be clear when setting. why not catch up some obvious concerns before training starts, with options to correct/change, or point/link to wiki?

for example, with a small dataset image/text-augmentation can make good sense, and OT knows about the amount of samples, number of epochs/batchsize it will process.

or, using prodigy with a LR <> 1 the training will fail.

with more optimizer-/model-types to expect this would be a conceptual aproach and can be expected to have a degree of effort in rmplementing at first, but it could be done stepwise when based on a simple-kept global structure, and in consequence easy to maintain when newly given defaults and the expected results are already known by OT/Devs when implementing new tech.

[O-J1]

Cisco whilst the intent is sound, thats off topic to this feature request and if it were easy to implement we would have. For any hyperparamaters off defaults and not occuring because of misclick we expect users know what they are doing, there is no abstracting away this is a technical field.

For moderate and lower UI issues see #220, gathered them there, as an aside you will notice OT now has basic UI data type validation now, try entering random text into LR 🙂

[ciscokid009]

Yep, off topic, understood. As said, just an idea, OT is complex, technical, but offers the magic to simplify with an intuitive UI, and beside, good to see LR dependency thingies are aware and adressed, coz I trapped in LR<>1 and I swiped away one of the best optimizers available I've seen so far, sadly for a long time . Wiki somehow confuses me sometimes, regarding defaults, I tend to miss important settings..

To get on topic again, my use case is graphjical art. I'm drawing graphic novels and get inspired by the generalization base models can achieve when trained on my personal style. Therefore I could say "default on" is better for me.

But my topic is kind of special interest, I'm using older sdxl-tech, augmentation helps a lot here, so, likewise as you said, I should know what I'm doing, and I'm not really representative here, I guess. Anyway, maybe this helps somehow.