Generated samples are non-negative about tensorflow-wavenet

[ibab]

Do you mean the numbers that are printed while it's generating?
Those are the outputs of the neural network, which we transform back into a waveform before saving the audio.

[hoonyoung]

no, I used the following command.. and 'generated.wav' file looks like that.

$ python generate.py --wav_out_path=generated.wav --save_every 2000 --samples 16000 model.ckpt-250

Maybe it's somehow related to 'librosa'. The package was not nicely installed in my case, so I had to try a few stuff to make it work. I'll check it again.

[ibab]

Okay, this used to be a bug and I thought I had fixed it.
Is it possible that you are running with a version that doesn't include 9f85246?
I'll look into reproducing this.

[hoonyoung]

You're correct. I was using the older version of 'wavenet.py'.
I'm now retraining the network and will check if it is resolved. Thank you.. :-)

[hoonyoung]

closing this because current version doesn't include this issue anymore.