Wrong Implementation of Discriminator

[jaywalnut310]

• 2nd AvgPooling: 4 strides(x) -> 2 strides(o)
• Downsampling rate is not an integer(2): Input data should be divisible by 4 x 256 (in training)

Although it seems to work with the above errors, I'll push the fixed version (and with more experimental options) when I get reasonable results.

[jaywalnut310]

While I fix those two problems in a new branch, I found synthesis quality's worse than before.

There are three things that I changed: 2 problems as above, and melspectrogram power from 2 to 1.
Does anyone have any idea why it gets worse?
Here is the sample audio G_172_new.zip.

I'm trying to do ablation tests to find the cause: keeping melspectrogram power to 2, keeping 2nd AvgPooling strides to 4.