Closed orbisAI closed 5 years ago
Ok, so the problem was r. I set r to be gradually training as per config below: "gradual_training": [[0, 7, 32], [10000, 5, 32], [50000, 3, 32], [130000, 2, 16], [290000, 1, 8]],
I trained model to 200k, so r should have been 2.
But I was loading the dataset and the model with r = 7, causing weird spectrogram outputs. Now I get this after changing r to 2:
Which makes much more sense.
When I try to use the notebook to generate spectrogram for training a vocoder, I get the following results as spectrogram (plz note it's upside down):
What's causing the vertical gaps in between the spectrogram data? I do not see such gaps when checking spectrograms during training. When we do GL with this spectrogram, as expected the sound is super jittery and broken.
Here is the spectrogram I can see in tensorboard.
FYI the notebook has been modified slightly to be used with libritts:
This is a notebook to generate mel-spectrograms from a TTS model to be used for WaveRNN training.