-
|
Hey! I'm currently in the process of training my model and I'm using tensorboard to view my scalars. In this context, I see scalars for grad_norm_d, grad_norm_g, and loss scalars like d/total, g/fm, g/kl, g/lf0, g/mel, g/total and a lr scalar? What do those mean exactly and what is being measured? I know that loss is very important to look at to indicate over-fitting, but I don't know how to measure over-fitting with these scalars mentioned above. Can anyone kindly give an explanation? Thank you for taking your time to read this. :) |
Beta Was this translation helpful? Give feedback.
Replies: 1 comment
-
|
Hello. The loss was discussed extensively in the VITS repository, so you can take a look at it. Probably the same parameters are used. I don't have an exact understanding of loss, but I don't think overfitting generally happens that easily (like within 100k steps) (if using default params). I think it is problematic that the loss is not currently calculated in the validation steps. |
Beta Was this translation helpful? Give feedback.
Hello. The loss was discussed extensively in the VITS repository, so you can take a look at it. Probably the same parameters are used. I don't have an exact understanding of loss, but I don't think overfitting generally happens that easily (like within 100k steps) (if using default params).
I think it is problematic that the loss is not currently calculated in the validation steps.