weixsong / waveglow Goto Github PK

I tried your code, hparams not changed, the wavs generated of 44k steps is understandable, but the line noise appear as many people reported in the nvidia code branch.
But your samples sounds great and suffer from this problem little, did you add post-process to the wavs? You have any tips ? You change the value of n_group or the sigma?

By the way, I tried the b04901014's branch, using lr=0.001, after about 30k, the model converge, and the total loss is about -5.8

hello thank you for great repository
after running training.. in logdir/waveglow/

by changing write_meta_graph = true, metafile would be recorded but I want to restore latest checkpoint default setting with repo.

But I wonder there is a way how to load and infer without meta
The structure of logdir is this.