Tensorflowtts: Problem when training Multiband MelGAN

[tienanh-1999]

Some early steps, it is okay

After that, I got a problem

I also fix that like https://github.com/TensorSpeech/TensorFlowTTS/issues/113, but it doesn't work

Thank you !

[dathudeptrai]

sub_sc_loss = tf.where(sub_sc_loss >= 10.0, 0.0, sub_sc_loss)
sub_mag_loss = tf.where(sub_mag_loss >= 10.0, 0.0, sub_mag_loss)
full_sc_loss = tf.where(full_sc_loss >= 10.0, 0.0, full_sc_loss)
full_mag_loss = tf.where(full_mag_loss >= 10.0, 0.0, full_mag_loss)
gen_loss = 0.5 * (sub_sc_loss + sub_mag_loss) + 0.5 * (
     full_sc_loss + full_mag_loss
)

try this. Let read the loss function here (https://github.com/TensorSpeech/TensorFlowTTS/blob/master/examples/multiband_melgan/train_multiband_melgan.py#L109). I think 10.0 is enough, in case the loss don't decrease, let try to set it by 20.0

[tienanh-1999]

thank you, I will try this !

[tienanh-1999]

I solved this by decreasing the learning rate (similar to the paper).
Thank you !

[dathudeptrai]

I solved this by decreasing the learning rate (similar to the paper).
Thank you !

okay, let share some results if you can :)).

[tienanh-1999]

I solved this by decreasing the learning rate (similar to the paper).
Thank you !

okay, let share some results if you can :)).

https://drive.google.com/drive/folders/1IST9UElMdxFj2SOJwQbUcNDp1PJ4Gdju?usp=sharing, here is the result after 400k steps :D

[dathudeptrai]

I solved this by decreasing the learning rate (similar to the paper).
Thank you !

okay, let share some results if you can :)).

https://drive.google.com/drive/folders/1IST9UElMdxFj2SOJwQbUcNDp1PJ4Gdju?usp=sharing, here is the result after 400k steps :D

it's good :)), let see the quality at 1M steps.