AttributeModelling

About

Code for the paper: Latent Space Regularization for Explicit Control of Musical Attributes. Implements the models and regularization technique to encode selected musical attributes along specific dimensions of the latent space of a VAE (Variational Auto-Encoder)-based model trained to reconstruct individual measures of music.

Installation and Use

1. Requires python3.x and pytorch1.0.0
2. Download or clone this repository. Navigate to the root folder and run python setup.py install.
3. Download the folder linked here. Unzip it and place the datasets and folk_raw_data folders in the AttributeModelling/data folder.
4. Run the train_measure_vae.py script with appropriate arguments to train or test the models.

Additonal Information

Training Parameters

All the model variants were trained using the same parameters to ensure consistency.

• Optimizer: Adam (b1=0.9, b2=0.999, e=1e-8)
• Learning Rate: 1e-4
• Batch-Size: 256
• Number of Epochs: 30
• Beta (for VAE training): 1e-3

Interpretability Metric

The scores were computed using the method proposed by Adel et al.. The computation steps are:

1. All data-points in the held-out test set are passed through the encoder of the trained model to obtain the corresponding latent vectors.
2. For each attribute a, the latent space dimension r which has the maximum mutual information with a is computed.
3. A simple linear regression model is then fit to predict a given z_r, i.e. the value of the latent code for dimension r.
4. The interpretability metric is finally the regression score (coefficienct of determination R2) for this regression model

The regression scores (higher is better) are shown in the table below:

Model Type	Rhythmic Complexity	Pitch Range	Average
RHY	0.8364	1.1E-06	0.42
PR	0.014	0.9625	0.49
RHY-PR	0.8339	0.9681	0.90
Base	4.2E-07	1.5E-05	7.9E-06

In our experiments, we found that r corresponded to the regularized dimension for attribute a for all the regularized models.

Attribution

This research work is published as as conference paper at the ICML ML4MD Workshop, 2019.

Ashis Pati, Alexander Lerch. "Latent Space Regularization for Explicit Control of Musical Attributes", ICML Workshop on Machine Learning for Music Discovery Workshop (ML4MD), Extended Abstract, Long Beach, CA, USA, 2019.

@inproceedings{pati19latent-reg,
  author = {Pati, Ashis and Lerch, Alexander},
  booktitle = {ICML Workshop on Machine Learning for Music Discovery Workshop (ML4MD), Extended Abstract},
  location = {Long Beach, CA, USA},
  title = {Latent Space Regularization for Explicit Control of Musical Attributes},
  year = {2019}
}

Please cite the above publication if you are using the code/data in this repository in any manner.

License

This work is licensed under a Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International License.