Combinatorial Music generation model with song structure graph analysis

Paper Preprint: https://arxiv.org/pdf/2312.15400v1.pdf

For training from scratch...

1. download lpd_17 dataset in https://salu133445.github.io/lakh-pianoroll-dataset/dataset, make root as song_structure_graph/lpd_17

2. download lmd_full dataset in https://colinraffel.com/projects/lmd/, make root as song_structure_graph/dataset/lmd_full

3. run following commands

    python main.py dataset_conformity
    python main.py process_CBIR
    python main.py train_CBIR
    python main.py process_pattern
    python main.py train_AE
    python main.py preprocess_npz
    python main.py train_graph2vec
    python main.py preprocess_unet
    python main.py train_unet
    python main.py generate_music

simple explanation for each part is in main.py. note that you should change checkpoint loading in some codes..! (especially, preprocess_npz, preprocess_unet, generate_music...)

Note that folder "datasets" from various site, to get genre label for musics.

For generate from checkpoints only...

1. download lpd_17 dataset in https://salu133445.github.io/lakh-pianoroll-dataset/dataset, make root as pattern_representation/lpd_17

2. download lmd_full dataset in https://colinraffel.com/projects/lmd/, make root as pattern_representation/dataset/lmd_full

3. Download Checkpoints from

https://zenodo.org/record/8333802 and unzip it, make root as song_structure_graph/checkpoints

    python main.py dataset_conformity
    python main.py process_CBIR
    python main.py process_pattern
    python main.py preprocess_npz
    python main.py generate_music

Some progress bar updates slowly(dataset_conformity, process_CBIR, process_pattern) because of parallel processing.

This project, which has been shared as an arXiv preprint and on GitHub, unfortunately, did not yield the expected results. Consequently, there are no plans for further submissions to conferences or journals. So, feel free to use any of the ideas or content in this project as you wish!