This uses the GrandPiano dataset to train a Transformer network to translate sheet music aka images, into a simplified *kern representation. The simplification of *kern consists mostly in droping any features not related playback per se, e.g. dropping ties, beams, etc. In addition, all voices, or chords are merged into a single "spine".
The FILES files has infos about the role & purpose of every file in the package.
This is not meant for public consumption yet.
Datasets:
- GrandStaff https://grfia.dlsi.ua.es/musicdocs/grandstaff.tgz
- Kern repository http://kern.ccarh.org/
- asap-dataset: https://github.com/fosfrancesco/asap-dataset
KernSheet dataset
The KernSheet dataset will eventually map bars within a score (aka pdf) into the corresponding bars within the kern file. Here is how to cook it:
# ./kernsheet.py make-kern-sheet TARGETTARGET is the (existing) directory in which you want to create the dataset. This command will fetch a bunch of .zip files from KernScore and copy them locally.
# ./kernsheet.py merge-asap ASAP TARGETASAP is a direxctory in which you've cloned the asap-dataset (see above) and TARGET is your target KernSheet dataset directory. This command will convert all .musicxml files into .krn file and move them to their appropriate spot in the final dataset directory.
From there on, kernsheet.sh has a bunch of command to help you edit the dataset, this includes:
- finding one or more pdf score for a given kern file,
- reviewing and validating the page layout of each pdf,
- ensuring the mapping to the kern file is correct.