[FastPitch/Pytorch] Using WaveGlow model train from Tacotron2 repository

[keto33]

Related to FastPitch/Pytorch

I trained a WaveGlow by the main repository https://github.com/NVIDIA/waveglow, and was able to make the model readable by FastPitch using the provided convert_model.py script.

However, I was unable to use the WaveGlow model trained using this repository (DeepLearningExamples/PyTorch/SpeechSynthesis/Tacotron2).

Loaded WaveGlow _IncompatibleKeys(missing_keys=['WN.0.cond_layer.bias', 'WN.0.cond_layer.weight_g', ....

1. Why is the model architecture of this repository is different from that of main WaveGlow repo? Is there any benefit?

2. Shouldn't a WaveGlow model train by DeepLearningExamples/PyTorch/SpeechSynthesis/Tacotron2 usable in FastPitch as they are part of the same project?

I use the direct codes rather than docker, this is probably the reason for the module issue. Nonetheless, I believe the WaveGlow model should be directly usable in FastPitch.

[jmmarko]

I am having the same issue and I am using Docker. Any solution or workaround so far?