Mapping text to MIDI tokens

• Examples of our process for mapping tokenized text to corresponding MIDI tokens can be found in the notebook file tokenization_examples.ipynb

To fine-tune MidiBERT-Piano on GLUE tasks

• First clone the MidiBERT-Piano repository and install required packages:

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

• Then copy the midi_glue.py and finetune_trainer.py files from our repository into the directory
• Download the pre-trained model checkpoint from the MidiBERT-Piano repo
• Now you can fine-tune MidiBERT-Piano on GLUE tasks:

python3 midi_glue.py --task='mrpc' --epochs=3 --ckpt='pretrain_model.ckpt' --lr=2e-5

• The fine-tuning runs for our experiments are in run_finetune_on_midi.sh

To fine-tune the Dutch model on GLUE tasks

• We used the run_glue.py script provided by Hugging Face to fine-tune GroNLP/bert-base-dutch-cased on the GLUE tasks.
• The Wordpiece tokenizer that we used for the English vocabulary is saved as a serialized JSON file on our repository as tokenizer.json and the vocabulary file is vocab.txt
• The fine-tuning runs for our experiments are in run_finetune_on_bert-base-cased.sh, run_finetune_on_bert-base-cased_random_weights.sh, and run_finetune_on_bert-base-dutch-cased.sh