Issue in reproducing results from the paper

[tripathiarpan20]

Hi!
Thanks for the open-source work.

I tried training an LLVC with a custom voice using a pretrained model from weights.gg, however, there were several discrepancies that lead to failure in reproduction of the results:

• The number of epochs are hardcoded to 10000, as opposed to 53 as mentioned in paper.
• The training takes more than 1 hour on a H100 instance just for 250 global steps, and ~7-8 hours per 250 global steps on an A100 instance (with the default arguments in the experiments/llvc/config.json, with only the log_interval and checkpoint_interval altered). This is way off as compared to the information in the paper claiming a training time of 3 days on RTX 3090 (for 500000 steps), any suggestions to fix the same?:
  
• Is there a way to efficiently finetune a new voice model with the pretrained G_500000.pth checkpoint in a parameter efficient manner that can be open-sourced?

[ksadov]

The number of epochs is hardcoded to 10000, but we manually stopped training at 53. Which batch size are you using for training on the H100?

As for efficiently fine-tuning from G_500000.pt: the training procedure should be the same as starting from an untrained checkpoint, and from my informal experience should be slightly quicker to converge.