search

NVIDIA / nv-wavenet

Reference implementation of real-time autoregressive wavenet inference
BSD 3-Clause "New" or "Revised" License
735 stars 126 forks source link

Does nv-wavenet support higher mu_quantization value? #45

Open nanma opened 6 years ago

nanma commented 6 years ago

The default mu_quantization value is 256, and I think higher value of mu_quantization may produce better quality audio. I did these modifications to test mu_quantization value of 512:

  1. modify values of mu_quantization, n_in_channels, n_out_channels in config.json to 512;
  2. modify value of const int A in wavenet_infer.cu to 512, and re-build the project;
  3. modify the 256 value in inference.py to 512.

But after training the model and inferencing, the produced audio files contain no sound. Did I miss anything else to modify, or nv-wavenet doesn't support other mu_quantization value?

gsoul commented 6 years ago

@nanma did you resolve your issue?

nanma commented 6 years ago

@gsoul No, I didn't further study this issue.

adc360 commented 6 years ago

@nanma I found it can work sometimes but often just gives a constant non-zero value. Running with CUDA_LAUNCH_BLOCKING=1 seems to improve the odds of success.