Tacotron2: HParams for training the model from readme

[gsoul]

Hello

If that's possible, could you please kindly share the parameters that were used to train the model that produced the sample from the readme (ideally both Tacotron2 and WaveNet).

@rafaelvalle, you indicated here that this might be a possibility: https://github.com/NVIDIA/tacotron2/issues/31#issuecomment-400345054

Thank you in advance.

[rafaelvalle]

Yes, we will share the params and possibly a model checkpoint soon after we're done refactoring the model.

[gsoul]

If that's possible, could you please give a hint on when that might happen?

[rafaelvalle]

Probably this week.

[rafaelvalle]

We're evaluating models that starts learning attention as fast as 1500 iterations on a single GPU as one can see in the picture below. We're in the process of comparing these models to others.

[DLZML001]

Hi, did you share the latest hparams for tacotron2 training?

[gsoul]

Hey, @rafaelvalle, if that's a possibility, could you please share the hparams that you used for training demo samples now, and then post the new code when it's ready. It would be extremely beneficial to the ones who try to reproduce your work with smaller hardware setups, as it takes ~2 weeks to train WaveNet to ~1mln steps...

[rafaelvalle]

We'll be releasing the new model and hprams in a branch soon.

[alexpeattie]

Any update on this @rafaelvalle ☺️ ?

[mikhail-barg]

Would be really nice to see actual hparams for the published sample.

We've tried running ttraining with default hparams on LJ dataset as explained in readme, and got incomparably worse results even after whole week of learning. Alignments seem to be OK, but the voice itself it very shaky, which doe not seem to be an effect of GL.

So we are not sure, is this because we did something wrong, or because we need to guess proper hparams, or something else

[rafaelvalle]

We've updated the repo with pre-trained tacotron 2 model weights.
Please check out the README.