Skip to content

Latest commit

 

History

History
199 lines (133 loc) · 7.26 KB

README.md

File metadata and controls

199 lines (133 loc) · 7.26 KB
Raw

Large Scale Cover Song Identification

Source code for the article "Data Driven and Discriminative Projections for Large-Scale Cover Song Identification" by Eric J. Humphrey, Oriol Nieto, and Juan P. Bello. ISMIR, Curitiba, Brazil, 2013.

This project includes three main tasks:

  • Binary Task: Analyze 500 tracks as described in (Thierry et al. 2012).
  • Cover Song ID in Second Hand Song Training Set (~12,000 songs on themselves).
  • Cover Song ID in Second Hand Song Test Set (~5,000 songs on 1,000,000).

Running the Tasks

Binary Task

This task compares one given track against two others and decides which one of the two is actually a cover of the given track. The list of tracks for this task is found in the file SHS/list_500queries.txt.

To run the task without any dimensionality reduction (i.e. each track is represented by the full 12x75 vector), type:

./binary_task.py path_to_MSD

The result, as reported in (Thierry et al. 2012), should be 82.0%.

This package already provides a previously learned Principal Component Analysis transformation (models/pca_250Kexamples_900dim_nocovers.pkl). To run the task with a PCA of 50 components:

./binary_task.py path_to_MSD -pca models/pca_250Kexamples_900dim_nocovers.pkl 50

This should result in 82.2%, as reported in (Thierry et al. 2012).

To run the test using a dictionary to project the data into a (hopefully) more separable space, we include a dictionary (models/BasisProjection2_kE2045_actEdot_shkE0x200_anormETrue.pk) that can be used like this:

./binary_task.py path_to_MSD -dictfile models/BasisProjection2_kE2045_actEdot_shkE0x200_anormETrue.pk

The result should be 81.2%.

To add Linear Discriminative Analysis to the new projected space:

./binary_task.py path_to_MSD -dictfile models/BasisProjection2_kE2045_actEdot_shkE0x200_anormETrue.pk -lda models/lda-kE2045-shkE0x200.pk n

where n is the index number of the model. In the given model (models/lda-kE2045-shkE0x200.pk), n = 0 represents 50 dimensions, n = 1 is 100 dimensions, and n = 2 is 200 dimensions. As an example, with 200 dimensions (n = 2) the result should be 94.4%.

Cover Song ID in Training

This task computes the Mean Average Precision (MAP) and the Average Rank (AR) of 12,960 tracks from the Second Hand Song dataset. These tracks are the ones selected for training, and are listed in SHS/shs_dataset_train.txt.

To run the task without any dimensionality reduction (i.e. each track is represented by the full 12x75 vector), type:

./cover_id_train.py path_to_MSD

The result should be exactly the same as the one reported by (Thierry et al. 2012): AR = 3096.7, MAP = 8.91%.

To apply PCA of 200 on these full features:

./cover_id_train.py path_to_MSD -pca models/pca_250Kexamples_900dim_nocovers.pkl 200

which should result in AR = 3005.1, MAP = 9.475%, as reported in (Thierry et al. 2012).

To apply a dictionary of k = 2045:

./cover_id_train.py path_to_MSD -dictfile models/BasisProjection2_kE2045_actEdot_shkE0x200_anormETrue.pk

This should result in AR = 3026, MAP = 5.51%, as reported in (Humphrey et al. 2013).

To apply LDA of 200 dimensions:

./cover_id_train.py path_to_MSD -dictfile models/BasisProjection2_kE2045_actEdot_shkE0x200_anormETrue.pk -lda 
models/lda-kE2045-shkE0x200.pk 2

Which should result in a similar result than AR = 1880, MAP = 28.33%, as reported in (Humphrey et al. 2013).

Cover Song ID in Test

This task compares the 5236 tracks from the test SHS dataset against all the 1,000,000 tracks from the MSD. Since computing the features for the entire MSD takes a long time, the code is written to be run in N different threads, so that the work can be split in N different processors.

To compute all the 1,000,000 codes with 10 processors, type:

./cover_id_test.py path_to_MSD -dictfile models/BasisProjection2_kE2045_actEdot_shkE0x200_anormETrue.pk -lda models/lda-kE2045-shkE0x200.pk -outdir msd_feats -N 10

Once all the features are computed inside a dir (e.g. msd_feats), we can compute the score by:

./cover_id_test.py path_to_MSD -codes msd_feats/ n

where n is the index number of the LDA model. In the given model (models/lda-kE2045-shkE0x200.pk), n = 0 represents 50 dimensions, n = 1 is 100 dimensions, and n = 2 is 200 dimensions. For this script to optimally run, we will need as much RAM memory as possible. The script works fine with 20GB of RAM. We also tried to run it with 6GB, but it took around 2 hours to load all the features, instead of a couple of minutes.

Note that in this task we can't compute the entire full features code (12x75, or, if projected to a higher dimensional space, 2045), since it would take too long to run.

Requirements

Million Song Dataset: http://labrosa.ee.columbia.edu/millionsong/

Numpy: http://www.numpy.org/

Scipy: http://www.scipy.org/

PyLab: http://wiki.scipy.org/PyLab

Scikit-Learn: http://scikit-learn.org/stable/

Multi-core processor. At least 10 units are recommended.

And as much RAM as possible. 16GB is recommended.

References

Humphrey, E. J., Nieto, O., & Bello, J. P. (2013). Data Driven and Discriminative Projections for Large-Scale Cover Song Identification. In Proc. of the 14th International Society for Music Information Retrieval Conference. Curitiba, Brazil.

Bertin-Mahieux, T., & Ellis, D. P. W. (2012). Large-Scale Cover Song Recognition Using The 2D Fourier Transform Magnitude. In Proc. of the 13th International Society for Music Information Retrieval Conference (pp. 241-246). Porto, Portugal.

Acknowledgments

Thanks to Thierry Bertin-Mahieux ([email protected]) for sharing his code and letting us publish it under the LGPL.

License

This code is distributed under the GNU LESSER PUBLIC LICENSE (LGPL, see www.gnu.org).

Copyright (c) 2012-2013 MARL@NYU. All rights reserved.

Redistribution and use in source and binary forms, with or without modification, are permitted provided that the following conditions are met:

a. Redistributions of source code must retain the above copyright notice, this list of conditions and the following disclaimer. b. Redistributions in binary form must reproduce the above copyright notice, this list of conditions and the following disclaimer in the documentation and/or other materials provided with the distribution. c. Neither the name of MARL, NYU nor the names of its contributors may be used to endorse or promote products derived from this software without specific prior written permission.

THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.