Template code for the genre classification project of ENSEA course "Electronique et Signal Musical".

Setup

The easiest way to run the code in this repository is to build the provided Dockerfile using Docker. The provided image contains all First, you need to install Docker. Then clone this project:

git clone https://github.com/romi1502/genre_classification_ensea_project.git

Then change to the directory of the repo and build the image :

cd genre_classification_ensea_project
docker build -t genre_classification_image docker

It may take a while since it will download the small version of the FMA dataset (about 8GB).

Once the image is built you can launch a container in interactive mode using:

docker run -ti --rm --name genre_classification_container genre_classification_image

Once in the container, you should be able to run the basic provided training script train_model.py like this:

python train_model.py

All used data and dependencies are self contained in the docker image.

Code details

The code is based on Tensorflow which is a very versatile and widely used machine learning framework. It especially makes use of the Tensorflow dataset API for data preprocessing and of Keras for Neural Network creation and training.

data_pipeline.py contains all data preprocessing: audio loading, features computation from audio, label formatting... For now it actually provides as features only a piece of the waveform, but you'll have to enhance it.
keras_model.pycontains the description of the neural network model that will be trained. For now it is a very basic fully connected one-layer network, but you'll have to enhance it.
train_model.py is the script that run model training. Once again it is very basic, and you'll have to enhance it.
utils.pycontains some utility functions that are used by the data_pipeline.py module
fma_small.csv contains information about the dataset: filename and genre for each music excerpt.

Docker tips

If you want to edit code in your favorite code editor and make it accessible within the docker container, you can mount a path of your computer filesystem within the docker filesystem when launching your container using the -voption:

docker run -ti --rm --name genre_classification_container -v /my/local/path/:/path/in/docker/container genre_classification_image

Then you should be able to access files located on your computer at /my/local/path/ within the container at path /path/in/docker/container.

You can let run a script within your docker container and check it afterwards:

you can detach from your docker container using the keyboard combination: 'Ctrl'+'P', 'Ctrl'+'Q'
you can attach back to your running docker container with the command:

docker attach genre_classification_container