A Model You Can Hear:

Audio Identification with Playable Prototypes

Description

Official PyTorch implementation of the paper "A Model You Can Hear: Audio Identification with Playable Prototypes".

Please visit our webpage for more details.

Installation 👷

1. Clone the repository in recursive mode

git clone git@github.com:romainloiseau/a-model-you-can-hear.git --recursive

2. Clone, create and activate conda environment

This implementation uses Pytorch.

Optional: some monitoring routines are implemented with tensorboard.

Note: this implementation uses pytorch_lightning for all training routines and hydra to manage configuration files and command line arguments.

How to use 🚀

Training the model

To train our best model, launch :

python main.py \
  +experiment={$dataset}_ours_{$supervision}

with dataset in {libri, sol} and supervision in {sup, unsup}

Testing the model

To test the model, launch :

python test.py \
  +experiment={$dataset}_ours_{$supervision} \
  model.load_weights="/path/to/trained/weights.ckpt"

Note: pretrained weights to come in pretrained_models/

Citation

@article{loiseau22online,
  title={A Model You Can Hear: Audio Identification with Playable Prototypes.},
  author={Romain Loiseau and Baptiste Bouvier and Yan Teytaut and Elliot Vincent and Mathieu Aubry and Loic Landrieu},
  journal={ISMIR},
  year={2022}
}