Feature mapping loss increasing during training

[Ziyi6]

Hello, I run the training script with config v1 and LJ-Speech dataset and I plotted all the loss separately during training. I found out that the feature mapping loss of mpd and msd kept increasing during training. Is it normal? How can one understand such a phenomenon? Should the loss decreasing during training?

[TheDevloper2023]

It should lower after a few thousand steps