This repository contains unconditinal WaveNet structure from:
-WaveNet: A Generative Model for Raw Audio[https://arxiv.org/abs/1609.03499]
The dataset is from:
-The MAESTRO Dataset V1.0.0[https://magenta.tensorflow.org/datasets/maestro#dataset] which stands for MIDI and Audio Edited for Synchronous TRacks and Organization.
Section | Description |
---|---|
Theory | Basic Theory |
Requirements | How to install the required package |
Usage | Quickstart examples |
GPU | GPU requirement and memory |
The dilated convolution is seen as below:
The network architecture is seen below:
In order to decrease the complexity of computation, we change the settings as follows:
Total layers: 8 residual channels: 32 skip channels: 128 max dilation: 128
This repo was tested on Python 3.7.3 with PyTorch 1.1 and Scipy 1.3.0
PyTorch can be installed by conda as follows:
conda install pytorch torchvision cudatoolkit=9.0 -c pytorch
Scipy can be installed by conda as follows:
conda install -c anaconda scipy
If you want to reproduce the results music reconstruction, you can run the command:
python train.py
If you want to train on different dataset, you should change the config.json file and train_files.txt
If you want to reproduce our results with the defult settings, you need a GPU with more than 10GB memory. Otherwise you need to decrease the number of layers.