Another question #3824
IIEleven11
started this conversation in
General
Replies: 0 comments
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
-
I know you guys threw in the towel and this will most likely go unanswered. Which is fine, I'm more doing it to just create a log.
What happened between tortoise and xttsv2 where it lost the ability to understand the prompt engineering that tortoise was capable of? Things like [sad], [happy], or [angry], etc... I can see parts of it possibly having been in the code and some easter eggs. I've been working to restore it, if you could possibly shed some insight I would much appreciate it.
I totally understand why one would remove it to try and monetize. I'm going to take a guess here and say open source is fundamentally flawed and a terrible business model. It appears you were a victim of that, even after helping lead the AI voice model community for so long. You have my respect for what you accomplished. Thank you.
That being said there's some really questionable portions of your code specifically surrounding the xtts model that could use some insight otherwise it could appear intentional.
For example, the LR scheduler and batch size/gradient accum. These two combined result in an almost immediate overfitting of the xttsv2 model and therefore forced early stopping. Once I adjusted these to more sensible values the fine tuning continues as normal. I refuse to believe competent engineers were under the impression that 900k epochs was a solid first milestone. There's a chance you thought it was 900k steps maybe? In my attempts to debunk myself I figured maybe the model does best set this way but it doesn't appear to be the case.
On your way out the community is still looking to you for your expertise and guidance. The big tech companies will likely win, no doubt. But in my opinion the best part of all this lies within the chase to the top. If you could give us one last push with your wisdom it would be much appreciated.
Either way thank you and I wish you well in whatever you choose to chase next.
Beta Was this translation helpful? Give feedback.
All reactions