Replies: 6 comments
-
>>> dkreutz |
Beta Was this translation helpful? Give feedback.
-
>>> rui |
Beta Was this translation helpful? Give feedback.
-
>>> nmstoker |
Beta Was this translation helpful? Give feedback.
-
>>> rui |
Beta Was this translation helpful? Give feedback.
-
>>> nmstoker |
Beta Was this translation helpful? Give feedback.
-
>>> nmstoker |
Beta Was this translation helpful? Give feedback.
-
>>> rui
[August 8, 2020, 8:44pm]
Hi,
I am testing the approach with Tacotron2 + Multiband Melgan in the dev
branch (commit 8c07ae7). slash
I was wondering if anyone has an example of what a healthy loss function
looks like when training the vocoder.
In my case, (with a smallish custom dataset) I observed some weird
behavior the moment we introduce the discriminator
(
steps_to_start_discriminator
). slashBefore this point the generator losses were trending down (as expected).
After this point they somewhat blew up.
Is this expected? Just a matter of training for longer? Maybe setting a
larger
steps_to_start_discriminator
?472
Thanks!
[This is an archived TTS discussion thread from discourse.mozilla.org/t/example-of-healthy-loss-functions-in-tacotron2-multiband-melgan]
Beta Was this translation helpful? Give feedback.
All reactions