Realistic labeled data generation

[ringw]

We have a primitive random measure generator using VexFlow. This should be improved to generate realistic music according to some prior distribution. We also need realistic noise; as a starting point, we should replicate the noise in the sequence-to-sequence paper.

That paper uses MuseScore, but VexFlow will be preferable for ease of adding hooks for writing the coordinates of each symbol. This will enable more types of models than the fully end-to-end model described in the paper.

The node.js script should also be set up in Bazel, and we need an open source Apache Beam pipeline binary for running it.