Training parameters and code

[980202006]

This project is very inspiring. Could you provide the training parameters and code? For example, the weights of KL loss and SISNR loss.
Thank you!

[fakerybakery]

+1

[bigpon]

We directly took the training code from DAC. The modifications are adding the KL loss and removing RVQ-related losses. (That is, we didn't use SISNR loss for training but for evaluations only)

We also found that the model is not sensitive to the KL loss weights (We have tried a large range of the KL weight, and the results are almost the same), so it is OK to use any common value used in a standard VAE training.