MusicAutoEncoder

author : Samuel Berrien

Requirements

Installed and exported in PATH :

ffmpeg

Python 3.6 pip packages :

torch
numpy
scipy
tqdm

For training only : a CUDA capable GPU with at least 2GB

Installation

First install all the dependencies listed before.

Then clone the repo :

$ git clone https://github.com/Ipsedo/MusicAutoEncoder.git

Usage

The argparse module gives help information with :

$ python script.py -h

Train

Create your set of wav files for training, in this example mp3 files are assumed to be in /path/to/mp3 and we will convert 100 files in wav to /path/to/wav out directory :

$ python read_audio.py process --mp3-root /path/to/mp3 --out-dir /path/to/wav -l 100

Now you are able to generate the torch.Tensor file :

$ python read_audio.py save --wav-root /path/to/wav --nb-wav 30 --out-tensor out_tensor.pt --nfft 49 --sample-rate 44100 --seconds 1

It creates a tensor in the file out_tensor.pt from 30 wav files at 44100Hz with 49 FFT values contained in /path/to/wav directory.

Finally start the training :

$ python train.py -h # TODO

Inference

TODO

Generation

TODO