XTTS v1.1 GPT Trainer

[Edresson]

XTTS GPT Trainer for XTTS v1.1.

ToDos:

• Update Docs
• Update recipe with the released checkpoints
• Rebase the PR to update the code with the changes done in XTTS v1.1 release.

[CLAassistant]

All committers have signed the CLA.

[erogol]

Looks good to me. Waiting for @WeberJulian

[WeberJulian]

I'm not 100% sure that we should use those values BATCH_SIZE 3 and GRAD_ACUMM_STEPS 84 for fine-tuning. In some extreme cases (small datasets) a single step will be more than an epoch. Why not use the values we use internally 8 and 1 ?

[WeberJulian]

Otherwise, it looks good to me too. Great work. In future PRs, we can focus on reducing the VRAM footprint of fine-tuning to something more reasonable that you can do on collab with 16Gb of VRAM.

[markrmiller]

That's interesting that you use 8/1 internally. The comments say you need an effective batch size of like 230 or something for good quality. Regardless, it would seem difficult to use a small dataset, in my limited experience, the model overfits pretty quickly, even with a much reduced learning rate, even with a decent sized data set. I'm still trying to get anywhere near a good voice match before it overfits.