ashispati / dmelodies_dataset

Dataset to study disentanglement in the context of symbolic music. Published as an ISMIR'20 paper titled: "dMelodies: A Music Dataset for Disentanglement Learning"

dMelodies: A Music Dataset for Disentanglement Learning

This repository contains the dMelodies dataset, meant to explore disentanglement in the context of symbolic music. Please cite as follows if you are using the code/data in this repository in any manner.

Ashis Pati, Siddharth Gururani, Alexander Lerch. "dMelodies: A Music Dataset for Disentanglement Learning", 21st International Society for Music Information Retrieval Conference (ISMIR), Montréal, Canada, 2020.

@inproceedings{pati2020dmelodies,
  title={dMelodies: A Music Dataset for Disentanglement Learning},
  author={Pati, Ashis and Gururani, Siddharth and Lerch, Alexander},
  booktitle={21st International Society for Music Information Retrieval Conference (ISMIR)},
  year={2020},
  address={Montréal, Canada}
}

Motivation

Over the last few years, there has been significant research attention on representation learning focused on disentangling the underlying factors of variation in given data. However, most of the current/previous studies rely on datasets from the image/computer vision domain (such as the dSprites dataset). The purpose of this work is to be able to create a standardized dataset for conducting disentanglement studies on symbolic music data. The key motivation is that such a dataset would help researchers working on disentanglement problems demonstrate their algorithm on diverse domains.

Description

dMelodies is dataset of simple 2-bar melodies generated using 9 independent latent factors of variation where each data point represents a unique melody based on the following constraints:

• Each melody will correspond to a unique scale (major, minor, blues, etc.).
• Each melody plays the arpeggios using the standard I-IV-V-I cadence chord pattern.
• Bar 1 plays the first 2 chords (6 notes), Bar 2 plays the second 2 chords (6 notes).
• Each played note is an 8th note.

A typical example is shown below.

Factors of Variation

The following factors of variation are considered:

1. Tonic (Root): 12 options from C to B
2. Octave: 3 options from C4 through C6
3. Mode/Scale: 3 options (Major, Minor, Blues)
4. Rhythm Bar 1: 28 options based on where the 6 note onsets are located in the first bar.
5. Rhythm Bar 2: 28 options based on where the 6 note onsets are located in the second bar.
6. Arpeggiation Direction Chord 1: 2 options (up/down) based on how the arpreggio is played
7. Arpeggiation Direction Chord 2: 2 options (up/down)
8. Arpeggiation Direction Chord 3: 2 options (up/down)
9. Arpeggiation Direction Chord 4: 2 options (up/down)

Consequently, the total number of data-points are 1,354,752.

Provided Data

The data is provided as a numpy .npz archive with six fields:

1. score_array: (1354752 x 16, int32) tokenized score representation
2. latent_array: (1354752 x 9, int32) integer index of the latent factor values
3. note2index_dict: dictionary mapping the musical note names/symbols to token indices
4. index2note_dict: dictionary mapping the token indices to musical note names/symbols
5. latent_dicts: dictionary mapping the different latent factor values to the corresponding integer indices
6. metadata: additional information (title, authors, date of creation etc.)

Usage

Install anaconda or miniconda by following the instruction here.

Create a new conda environment using the enviroment.yml file located in the root folder of this repository. The instructions for the same can be found here.

Activate the dmelodies environment using the following command:

conda activate dmelodies

The DMelodiesDataset class (see dmelodies_dataset.py) is a wrapper around the dataset and provides methods to read and load the dataset. If you are using PyTorch, you can use the DMelodiesTorchDataset class (see dmelodies_torch_dataloader.py) which implements a torch DataLoader. Examples for using both can be found in the dmelodies_loading.ipynb file.

Explore the benchmarking experiments on dMelodies using different unsupervised learning methods.

Dataset Creation

In case you want to create your own version of the dataset (as a .npz file), delete the contents of data folder and then run script_create_dataset.py from the root folder of this repository. Additional arguments --save-midi and --save-xml can be used to save the individual melodies as .mid or .musicxml files. The files will be saved in a raw_data folder. Note: Saving individual melodies will require approx. 16.5GB of space (5.5 GB for .mid format and 11GB for .musicxml format)