SE and SNL Diagrams: Flexible data structures for music-based data streams

A softward for generating Start-End and Start-Normalized Length Diagrams from Aligned Hierarchies for Music-Based Data.

Authors

Melissa R. McGuirl, Katherine M. Kinnaird, Claire Savard, Erin Bugbee.

Contact

For questions/comments please contact Melissa R. McGuirl at melissa_mcguirl@brown.edu.

About

This software addresses the cover-song task for music-based data.

Inputs

This software takes as input a collection of .mat files containing Aligned Hierarchies data stored in variables called 'full_key' and 'full_matrix_no,' where 'full_matrix_no' is a matrix indicating where the repeated structures start and 'full_key' is a vector containing the length of the repeated structures in each row.

Outputs

Precision and Recall values for mutual k-nearest neighbor matching on SE/SNL diagrams for the cover song task.

User options

1. Wasserstein Inner Norm: Value in [1.0, Infinity]
2. Wasserstein Outer Norm: Value in [1.0, Infinity] (note, when the outer norm =infinit you are using the Bottleneck metric. In this case the inner norm is automatically set to l-infinity)
3. Normalization (for run_SNLD.py): 'none' for no mormalization, 'std' for ordinary normalization, or 'cheb' for chebyshev normalization.
4. Alpha: Scaling parameter alpha, alpha = 1 indicates no scaling.

Usage

python run_SE.py [-h HELP_MENU] -I INPUT_DIRECTORY -w INNER_NORM -W OUTER_NORM python run_SNLD.py [-h HELP_MENU] -I INPUT_DIRECTORY -N NORMALIZATION -A ALPHA -w INNER_NORM -W OUTER_NORM

Dependencies

Hera (available here: https://bitbucket.org/grey_narn/hera.git), Python 2.7, numpy, scipy, sklearn

IMPORTANT: If you are using hera to compute distances between SL or SnL diagrams, you must first modify the hera source code to move the 'diagonal' to the y-axis. This should be done in basic_defs.cpp for both the Wasserstein and Bottleneck codes.

Examples

1. python run_E.py -I ../data/Thresh01_ShingleNumber6/ -w 2 -W 2 (SE diagrams with 2-2 Wasserstein metric)
2. python run_SNLD.py -I ../data/Thresh01_ShingleNumber6/ -N none -A 0 -w 2 -W 2 (SL diagrams with 2-2 Wasserstein metric)
3. python run_SNLD.py -I ../data/Thresh01_ShingleNumber6/ -N std -A 1 -w 2 -W 2 (SNL diagrams with standard normalization (M = maximum start time), alpha=1, and the 2-2 Wasserstein metric)
4. python run_SNLD.py -I ../data/Thresh01_ShingleNumber6/ -N cheb -A 10 -w inf -W inf (SNL diagrams with chebyshev normalization, alpha=10, and the infinity-Bottleneck metric)

Pipeline

1. Extract start-end diagrams or start-normalized length diagrams from Aligned Hierarchies
2. Normalize start times using chebyshev/standard norm (if specified)
3. Compute Wasserstein or Bottleneck distances between all diagrams in the dataset
4. Use mutual k-nearest neighbor to pair songs
5. Generate truth (exanded version of each song pairs with non-expanded version)
6. Compare to true pairing with SE/SNL mutual-KNN pairing
7. Compute precision and recall

Help

0. python TREE.py -h

Updates for future releases

1. Make sofware flexible for cases where truth is not known
2. Save mutual knn results to text file

References and citation for this software

This accompanying paper for this sofware is:

• http://ismir2018.ircam.fr/doc/pdfs/105_Paper.pdf

For more information on Aligned Hierarchies see:

• K. M. Kinnaird. Aligned Hierarchies: A Multi-Scale Structure-based Representation For Music-Based Data Streams. Proceedings of the 17th ISMIR Conference, New York City, 2016.