This is the official repository for the paper, MidiBERT-Piano: Large-scale Pre-training for Symbolic Music Understanding.

With this repository, you can

• pre-train a MidiBERT-Piano with your customized pre-trained dataset
• fine-tune & evaluate on 4 downstream tasks
• extract melody (mid to mid) using pre-trained MidiBERT-Piano

All the datasets employed in this work are publicly available.

If you'd like to reproduce the results (MidiBERT) shown in the paper,

1. Please download the checkpoints, and rename files like the following

   (Note: we only provide checkpoints for models in CP representations)

result/
└── finetune/
	└── melody_default/
		└── model_best.ckpt
	└── velocity_default/
		└── model_best.ckpt
	└── composer_default/
		└── model_best.ckpt
	└── emotion_default/
		└── model_best.ckpt

2. Run ./scripts/eval.sh

   Or refer to Readme in MidiBERT folder for more details.

   No gpu is needed for evaluation

Edit scripts/melody_extraction.sh and modify song_path to your midi path. The midi file to predicted melody will be saved at the root folder.

./scripts/melody_extraction.sh

# modify this line (export PYTHONPATH='.') to the following
set PYTHONPATH='.'
# print the environment variable to make sure it's working
echo %PYTHONPATH%

I've experimented this on Adele hello (piano cover), and I think it's good.
But for non-pop music like Mozart sonata, I feel like the model is pretty confused. This is expected. As the training data is POP909 Dataset, the model knows very little about classical music.

Side note: I try to make it more friendly for non-programmers. Feel free to open an issue if there's any problem.

• Python3
• Install generally used packages for MidiBERT-Piano:

git clone https://github.com/wazenmai/MIDI-BERT.git
cd MIDI-BERT
pip install -r requirements.txt

Please see scripts folder, which includes bash file for

• prepare data
• pretrain
• finetune
• evaluation
• melody extraction

You may need to change the folder/file name or any config settings you prefer.

Data/
└── Dataset/       
    └── pop909/       
    └── .../
└── CP_data/
    └── pop909_train.npy
    └── *.npy

data_creation/
└── preprocess_pop909/
└── prepare_data/       # convert midi to CP_data 
    └── dict/           # CP dictionary 

melody_extraction/
└── skyline/
└── midibert/

MidiBERT/
└── *py

For more details on

• data preparation, please go to data_creation and follow Readme
• MidiBERT pretraining, finetuning, evaluation, please go to MidiBERT and follow Readme.
• skyline, please go to melody_extraction/skyline and follow Readme.
• pianoroll figure generation, please go to melody_extraction/pianoroll and follow Readme. We also provide clearer pianoroll pictures of the paper.
• listening to melody extraction results, please go to melody_extraction/audio and read Readme for more details.

Note that Baseline (LSTM) and code in remi versions are removed for cleaness. But you could find them in main branch.

If you find this useful, please cite our paper.

@article{midibertpiano,
  title={{MidiBERT-Piano}: Large-scale Pre-training for Symbolic Music Understanding},
  author={Yi-Hui Chou and I-Chun Chen and Chin-Jui Chang and Joann Ching, and Yi-Hsuan Yang},
  journal={arXiv preprint arXiv:2107.05223},
  year={2021}
}