Extension request or Training improvement to automate Textual Inversion trainings

[hyppyhyppo]

Hello,

When doing TI, I'm now making several tries with the same dataset with several batch / Gradient accumulation steps then compare them.
Ex: with a dataset of 10 pics, I'll try several Batch/GAS combos like 10x1, 5x2, 2x5, 1x10. With such ratio it's always the same Learning rate progression and max steps. I then compare the generated TI (with same init text and vectors number), pick up the best and eventually merge them.
These combos batch / GAS really matters depending on a dataset, there is no ideal values and need to be benchmarked for good TI. I now do it manually but feel like a robot :)

I see (dream of) an user interface where I'd had to input an embedding prefix, init text, vector size, learning rate progression and max steps and several lines for the Batch / GAS values.
For 4 Batch / GAS combo it would create 4 TI with name embedding prefix_number from 1 to 4.

Does such extension exist already ? I guess no. Would it be something doable ? I guess yes. Would it be usefull ? Yes, for sure !!!
It even could be a great improvement on the actual training feature :)

What do you think ?
Thanks for reading anyway :)