FMA: A Dataset For Music Analysis

Michaël Defferrard, Kirell Benzi, Pierre Vandergheynst, Xavier Bresson.
International Society for Music Information Retrieval Conference (ISMIR), 2017.

We introduce the Free Music Archive (FMA), an open and easily accessible dataset suitable for evaluating several tasks in MIR, a field concerned with browsing, searching, and organizing large music collections. The community's growing interest in feature and end-to-end learning is however restrained by the limited availability of large audio datasets. The FMA aims to overcome this hurdle by providing 917 GiB and 343 days of Creative Commons-licensed audio from 106,574 tracks from 16,341 artists and 14,854 albums, arranged in a hierarchical taxonomy of 161 genres. It provides full-length and high-quality audio, pre-computed features, together with track- and user-level metadata, tags, and free-form text such as biographies. We here describe the dataset and how it was created, propose a train/validation/test split and three subsets, discuss some suitable MIR tasks, and evaluate some baselines for genre recognition. Code, data, and usage examples are available at https://github.com/mdeff/fma.

• Paper: arXiv:1612.01840 (latex and reviews)
• Slides: doi:10.5281/zenodo.1066119
• Poster: doi:10.5281/zenodo.1035847

Data

All metadata and features for all tracks are distributed in fma_metadata.zip (342 MiB). The below tables can be used with pandas or any other data analysis tool. See the paper or the usage.ipynb notebook for a description.

• tracks.csv: per track metadata such as ID, title, artist, genres, tags and play counts, for all 106,574 tracks.
• genres.csv: all 163 genre IDs with their name and parent (used to infer the genre hierarchy and top-level genres).
• features.csv: common features extracted with librosa.
• echonest.csv: audio features provided by Echonest (now Spotify) for a subset of 13,129 tracks.

Then, you got various sizes of MP3-encoded audio data:

1. fma_small.zip: 8,000 tracks of 30s, 8 balanced genres (GTZAN-like) (7.2 GiB)
2. fma_medium.zip: 25,000 tracks of 30s, 16 unbalanced genres (22 GiB)
3. fma_large.zip: 106,574 tracks of 30s, 161 unbalanced genres (93 GiB)
4. fma_full.zip: 106,574 untrimmed tracks, 161 unbalanced genres (879 GiB)

Code

The following notebooks, scripts and modules have been developed for the dataset.

1. usage.ipynb: shows how to load the datasets and develop, train and test your own models with it.
2. analysis.ipynb: exploration of the metadata, data and features. Creates the figures used in the paper.
3. baselines.ipynb: baseline models for genre recognition, both from audio and features.
4. features.py: features extraction from the audio (used to create features.csv).
5. webapi.ipynb: query the web API of the FMA. Can be used to update the dataset.
6. creation.ipynb: creation of the dataset (used to create tracks.csv and genres.csv).
7. creation.py: creation of the dataset (long-running data collection and processing).
8. utils.py: helper functions and classes.

Usage

1. Download some data, verify its integrity, and uncompress the archives.

   curl -O https://os.unil.cloud.switch.ch/fma/fma_metadata.zip
   curl -O https://os.unil.cloud.switch.ch/fma/fma_small.zip
   curl -O https://os.unil.cloud.switch.ch/fma/fma_medium.zip
   curl -O https://os.unil.cloud.switch.ch/fma/fma_large.zip
   curl -O https://os.unil.cloud.switch.ch/fma/fma_full.zip
   
   echo "f0df49ffe5f2a6008d7dc83c6915b31835dfe733  fma_metadata.zip" | sha1sum -c -
   echo "ade154f733639d52e35e32f5593efe5be76c6d70  fma_small.zip"    | sha1sum -c -
   echo "c67b69ea232021025fca9231fc1c7c1a063ab50b  fma_medium.zip"   | sha1sum -c -
   echo "497109f4dd721066b5ce5e5f250ec604dc78939e  fma_large.zip"    | sha1sum -c -
   echo "0f0ace23fbe9ba30ecb7e95f763e435ea802b8ab  fma_full.zip"     | sha1sum -c -
   
   unzip fma_metadata.zip
   unzip fma_small.zip
   unzip fma_medium.zip
   unzip fma_large.zip
   unzip fma_full.zip

   If you get any error while decompressing the archives (especially with the Windows and macOS system unzippers), please try 7zip. That is probably an unsupported compression issue.

2. Optionally, use pyenv to install Python 3.6 and create a virtual environment.

   pyenv install 3.6.0
   pyenv virtualenv 3.6.0 fma
   pyenv activate fma

3. Clone the repository.

   git clone https://github.com/mdeff/fma.git
   cd fma

4. Checkout the revision matching the data you downloaded (e.g., beta, rc1, v1). See the history of the dataset.

   git checkout rc1

5. Install the Python dependencies from requirements.txt. Depending on your usage, you may need to install ffmpeg or graphviz. Install CUDA if you want to train neural networks on GPUs (see Tensorflow's instructions).

   make install

6. Fill in the configuration.

   cat .env
   AUDIO_DIR=/path/to/audio
   FMA_KEY=IFIUSETHEAPI

7. Open Jupyter or run a notebook.

   jupyter notebook
   make baselines.ipynb

Coverage and resources

• Using CNNs and RNNs for Music Genre Recognition, Towards Data Science, 2018-12-13.
• Over 1.5 TB’s of Labeled Audio Datasets, Towards Data Science, 2018-11-13.
• Genre recognition challenge at the web conference, Lyon, 2018-04.
• Discovering Descriptive Music Genres Using K-Means Clustering, Medium, 2018-04-09.
• 25 Open Datasets for Deep Learning Every Data Scientist Must Work With, Analytics Vidhya, 2018-03-29.
• music2vec: Generating Vector Embeddings for Genre-Classification Task, Medium, 2017-11-28.
• Slides presented at the Data Jam days, Lausanne, 2017-11-24.
• Poster presented at ISMIR 2017, China, 2017-10-24.
• Slides for the Open Science in Practice summer school at EPFL, 2017-09-29.
• A Music Information Retrieval Dataset, Made With FMA, freemusicarchive.org, 2017-05-22.
• Pre-publication release announced, twitter.com, 2017-05-09.
• FMA: A Dataset For Music Analysis, tensorflow.blog, 2017-03-14.
• Beta release discussed, twitter.com, 2017-02-08.
• FMA Data Set for Researchers Released, freemusicarchive.org, 2016-12-15.

Research papers (see also citations on Google Scholar):

• Zero-shot Learning for Audio-based Music Classification and Tagging
• One deep music representation to rule them all? A comparative analysis of different representation learning strategies
• Deep Learning for Audio-Based Music Classification and Tagging: Teaching Computers to Distinguish Rock from Bach
• Learning Discrete Structures for Graph Neural Networks
• A context encoder for audio inpainting
• OpenMIC-2018: An Open Data-set for Multiple Instrument Recognition
• Detecting Music Genre Using Extreme Gradient Boosting
• Transfer Learning of Artist Group Factors to Musical Genre Classification
• Learning to Recognize Musical Genre from Audio: Challenge Overview
• Representation Learning of Music Using Artist Labels

Dataset lists:

• https://github.com/caesar0301/awesome-public-datasets
• https://archive.ics.uci.edu/ml/datasets/FMA:+A+Dataset+For+Music+Analysis
• http://deeplearning.net/datasets
• http://www.audiocontentanalysis.org/data-sets
• https://github.com/ismir/mir-datasets
• https://teachingmir.wikispaces.com/Datasets
• https://en.wikipedia.org/wiki/List_of_datasets_for_machine_learning_research
• https://cloudlab.atlassian.net/wiki/display/datasets/FMA:+A+Dataset+For+Music+Analysis
• https://www.datasetlist.com

History

2017-05-09 pre-publication release

• paper: arXiv:1612.01840v2
• code: git tag rc1
• fma_metadata.zip sha1: f0df49ffe5f2a6008d7dc83c6915b31835dfe733
• fma_small.zip sha1: ade154f733639d52e35e32f5593efe5be76c6d70
• fma_medium.zip sha1: c67b69ea232021025fca9231fc1c7c1a063ab50b
• fma_large.zip sha1: 497109f4dd721066b5ce5e5f250ec604dc78939e
• fma_full.zip sha1: 0f0ace23fbe9ba30ecb7e95f763e435ea802b8ab

2016-12-06 beta release

• paper: arXiv:1612.01840v1
• code: git tag beta
• fma_small.zip sha1: e731a5d56a5625f7b7f770923ee32922374e2cbf
• fma_medium.zip sha1: fe23d6f2a400821ed1271ded6bcd530b7a8ea551

Contributing

Please open an issue or a pull request if you want to contribute. Let's try to keep this repository the central place around the dataset! Links to resources related to the dataset are welcome.

License & co

• Please cite our paper if you use our code or data.

  @inproceedings{fma_dataset,
    title = {{FMA}: A Dataset for Music Analysis},
    author = {Defferrard, Micha\"el and Benzi, Kirell and Vandergheynst, Pierre and Bresson, Xavier},
    booktitle = {18th International Society for Music Information Retrieval Conference (ISMIR)},
    year = {2017},
    archiveprefix = {arXiv},
    eprint = {1612.01840},
    url = {https://arxiv.org/abs/1612.01840},
  }
  

• The code in this repository is released under the terms of the MIT license.
• The metadata is released under the terms of the Creative Commons Attribution 4.0 International License (CC BY 4.0).
• We do not hold the copyright on the audio and distribute it under the terms of the license chosen by the artist.
• The dataset is meant for research purposes.
• We are grateful to the Swiss Data Science Center (EPFL and ETHZ) for hosting the dataset.