REMI + Ph&BC

Installation

Clone the original REMI code (Pop Music Transformer):

git clone https://github.com/YatingMusic/remi.git
cd remi
git checkout 6d40725

Clone the diff file in this repository:

git clone https://github.com/mil-tokyo/phrase-length-designated-music-generation.git -b master --single-branch

Apply the diff file to the REMI code:

patch -p1 < phrase-length-designated-music-generation/diff.patch

Clone the POP909 dataset and its annotations:

git clone https://github.com/music-x-lab/POP909-Dataset.git
git clone https://github.com/Dsqvival/hierarchical-structure-analysis.git

Please refer to the original REMI repository for instructions on how to install dependencies.

Training

NOTE: This is training from scratch (sorry for confusing you).

python finetune.py

After training, please choose which checkpoint you want to try and change the checkpoint names you choose into "model".

• "model-xxx-y.zzz.data-00000-of-00001" → "model.data-00000-of-00001"
• "model-xxx-y.zzz.index" → "model.index"
• "model-xxx-y.zzz.meta" → "model.meta"

Inference

After changing the phrase configuration in the main.py as necessary, run it.

python main.py

"from_scratch.midi" will be generated in the "result" folder.

Reference

• REMI code (Pop Music Transformer)
  https://github.com/YatingMusic/remi
• POP909 dataset
  https://github.com/music-x-lab/POP909-Dataset
• POP909 annotation (Automatic Analysis and Influence of Hierarchical Structure on Melody, Rhythm and Harmony in Popular Music)
  https://github.com/Dsqvival/hierarchical-structure-analysis