A hard fork of the G2P training code from OpenUTAU’s repository, extended with modifications to enable more flexible experimentation with training parameters.
- Integrated
pytorch-optimizer: This adds access to a wider variety of optimizers and learning rate schedulers, making it easier to experiment with training stability, performance improvements, or avoiding NaN issues.
-
Edit your
.yamlconfiguration file (or the default settings inside the input dictionary in the Colab notebook) to define parameters for optimizers and learning rate schedulers. -
For optimizers, you can use either:
pytorch_optimizer.<OptimizerName>torch.optim.<OptimizerName>
-
For learning rate schedulers, you can also use either:
torch.optim.lr_scheduler.<SchedulerName>pytorch_optimizer.<SchedulerName>
-
For more information, check the original repo.
Both optimizers and schedulers support full parameter customization through the .yaml file (or inside the Colab cell).
Please refer to the official docs of both PyTorch and pytorch-optimizer
Credits: Based on the original OpenUTAU repository. Social Preview made with Socialify, Tennoji Rina´s signature used there came from here.
