gal1eo / dt2119

This is a Music Generation Implementation using WaveNet

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.

Content

Section	Description
Theory	Basic Theory
Requirements	How to install the required package
Usage	Quickstart examples
GPU	GPU requirement and memory

Theory

The Dilated Convolution

The dilated convolution is seen as below:

The Music WaveNet

The network architecture is seen below:

Reduced Version

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

Requirements

This repo was tested on Python 3.7.3 with PyTorch 1.1 and Scipy 1.3.0

Installation

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

Usage

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

GPU

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.