architecture shows bad results

[e0xextazy]

Hi, i have completely repeated your steps for learning. During training, style speech loss fell down, but after learning began, meta style speech loss began to grow up. Can you help with training the model? I can describe my steps in more detail.

[keonlee9420]

Hi @e0xextazy, thanks for the report. I found that the current implementation updates both generator and discriminator when the discriminator is updated by ScheduledOptimDisc, but it should be the discriminator only. I fixed the code and pushed to main branch, and my training log shows the expected results as follows. (grey: previous version, blue: fixed version)

The pre-trained model is just updated as well.

[keonlee9420]

close due to inactivity