https://github.com/shuokay/mxnet-wavenet

MXNet implementation of WaveNet
https://github.com/shuokay/mxnet-wavenet

Last synced: 2 months ago
JSON representation

MXNet implementation of WaveNet

• Host: GitHub
• URL: https://github.com/shuokay/mxnet-wavenet
• Owner: shuokay
• Created: 2016-10-18T07:32:09.000Z (about 8 years ago)
• Default Branch: master
• Last Pushed: 2016-10-20T02:35:57.000Z (about 8 years ago)
• Last Synced: 2024-08-01T22:41:51.930Z (5 months ago)
• Language: Python
• Homepage:
• Size: 4.88 KB
• Stars: 20
• Watchers: 6
• Forks: 13
• Open Issues: 1
• Metadata Files:
  • Readme: README.md

Awesome Lists containing this project

• Awesome-MXNet - wavenet - gluon]](https://github.com/seujung/WaveNet-gluon) (<a name="Speech"></a>4. Speech / 2.14 Misc)

README

        
# MXNet Implementation of WaveNet #

I am trying to reproduce the WaveNet result using MXNet. Here is the training code of generating without any condition, ~~but the training process can't convergence since the *mae* is always around 126. Hoping someone can raise your advices.~~

The model convergence in fact. The reason of "mae is always around 126" is the mxnet office mae evaluation metric does not match this net config. I have defined a new `EvalMetric`  

[Training log](https://gist.github.com/shuokay/28de2c02c7857ab6ab7be2cd26b76915)

## How to Run ##

1. Install mxnet and fix the dilate bug according to https://github.com/dmlc/mxnet/issues/3479

2. Download VCTK cprpus and extract to the root folder

3. Start training by `python train.py`

## Implementation Note ##

* Padding zero on right side by *Concat* operator since the convolution op of mxnet can't pad only on one side.