search

Kyubyong / tacotron

A TensorFlow Implementation of Tacotron: A Fully End-to-End Text-To-Speech Synthesis Model
Apache License 2.0
1.83k stars 436 forks source link

No DynamicAttentionWrapper in Tensorflow 1.2 #46

Open markisus opened 7 years ago

markisus commented 7 years ago

The readme says the code requires Tensorflow >= 1.1, but it crashes on TF 1.2 because of missing DynamicAttentionWrapper. https://github.com/Kyubyong/tacotron/blob/master/modules.py#L247

I substituted for AttentionWrapper https://www.tensorflow.org/api_docs/python/tf/contrib/seq2seq/AttentionWrapper and the code runs. However I'm not getting good results.

YisenWang commented 7 years ago

You can downgrade to TF 1.1.0

jaron commented 7 years ago

Looking at the commit history , it should just be a rename.

Can you explain what you mean by not getting good results?

markisus commented 7 years ago

screenshot of my tensorboard sample from eval.py The network does not seem to be making progress and the sample files outputted by eval.py sound like repeating noise. Could it be a matter of just not training enough?

jaron commented 7 years ago

Pull and try with the latest commits. I did get noisy results initially too, but it should get better and produce recognisable sounds if the model can learn. You might want to try training with the new sanity_check flag.