/piano-reduction-with-midibert

Primary LanguageJupyter Notebook

TOWARDS PRACTICAL AUTOMATIC PIANO REDUCTION USING BERT WITH SEMI-SUPERVISED LEARNING

Environment setup

Please first create an environment using conda or venv with python3.7. Then install the dependencies as in requirements.txt.

The baseline DBM method

To run the DBM method,

  1. Prepare midi pieces to be used for building the database in the data folder.
  2. Prepare also the piece to be reduced.
  3. go to DBM/build_up.ipynb and modify the directories, then run the code!

Pretraining

(Note: If you need to run the training code, a powerful GPU is recommended.)

  1. Prepare midi used for pretraining
  2. Follow MidiBERT/Pretraining.ipynb and modify relevant arguments

The MB-NR method

Please follow MidiBERT/MBNR.ipynb for more instructions regarding training and inferencing.

The MB-R2F method

  1. Prepare data following MidiBERT/skinlineTokenize.ipynb.
  2. Run python MidiBERT/CP/main.py --mode seq2seq and add other arguments as required.

Evaluation

All objective evaluation codes are included within the eval folder.

  1. Make any necessary directory changes in eval/eval.py
  2. Run the code. You will get a pickle file which contains a directionary of the