Music Composition with Bayesian Program Learning

We apply the Bayesian Program Learning framework to a music composition context.

For details, check out our writeup.

Music samples are available here.

Running System

The core pipeline is contained in main.py. The other scripts also require this script to have been run and properly cached the various objects. To execute:

python3 main.py

Or to choose between HMM and GRU based type generation run either of following:

python3 main.py gru
python3 main.py hmm

To run the GRU based generator, one must first run the above, and then run:

python3 gru_generator.py

Computing Metric

After having run main.py, one can now compare a generated score to the scores cached. One does so as follows:

python3 score.py path/to/generated/score.json
python3 score.py path/to/directory/of/scores/

Other utilities

python3 baseline.py uses the currently cached training scores to do an independent empirical drawing of words to produce a piece of music. This represents a system which has learned nothing about how a piece of music works together.