muzic/musicbert
曾明亮 de07092801
Update README.md
2022-07-20 22:16:38 +08:00
..
musicbert first commit 2021-09-22 23:54:56 +08:00
README.md Update README.md 2022-07-20 22:16:38 +08:00
binarize_genre.sh Fix typos discovered by codespell (#6) 2021-09-28 09:20:34 +08:00
binarize_nsp.sh first commit 2021-09-22 23:54:56 +08:00
binarize_pretrain.sh first commit 2021-09-22 23:54:56 +08:00
chord_mapping.pkl first commit 2021-09-22 23:54:56 +08:00
eval_genre.py first commit 2021-09-22 23:54:56 +08:00
eval_nsp.py first commit 2021-09-22 23:54:56 +08:00
gen_genre.py first commit 2021-09-22 23:54:56 +08:00
gen_nsp.py first commit 2021-09-22 23:54:56 +08:00
preprocess.py fix duration issue in encoding to MIDI 2021-11-28 22:58:12 +08:00
train_genre.sh first commit 2021-09-22 23:54:56 +08:00
train_mask.sh first commit 2021-09-22 23:54:56 +08:00
train_nsp.sh first commit 2021-09-22 23:54:56 +08:00

README.md

MusicBERT

MusicBERT: Symbolic Music Understanding with Large-Scale Pre-Training, by Mingliang Zeng, Xu Tan, Rui Wang, Zeqian Ju, Tao Qin, Tie-Yan Liu, ACL 2021, is a large-scale pre-trained model for symbolic music understanding. It has several mechanisms including OctupleMIDI encoding and bar-level masking strategy that are specifically designed for symbolic music data, and achieves state-of-the-art accuracy on several music understanding tasks, including melody completion, accompaniment suggestion, genre classification, and style classification.

Projects using MusicBERT:


Model structure of MusicBERT


OctupleMIDI encoding

1. Preparing datasets

1.1 Pre-training datasets

  • Prepare The Lakh MIDI Dataset (LMD-full) in zip format for pre-training. (say lmd_full.zip)

    wget http://hog.ee.columbia.edu/craffel/lmd/lmd_full.tar.gz
    tar -xzvf lmd_full.tar.gz
    zip -r lmd_full.zip lmd_full
    
  • Run the dataset processing script. (preprocess.py)

    python -u preprocess.py
    
  • The script should prompt you to input the path of the midi zip and the path for OctupleMIDI output.

    Dataset zip path: /xxx/xxx/MusicBERT/lmd_full.zip
    OctupleMIDI output path: lmd_full_data_raw
    SUCCESS: lmd_full/a/0000.mid
    ......
    
  • Binarize the raw text format dataset. (this script will read lmd_full_data_raw folder and output lmd_full_data_bin)

    bash binarize_pretrain.sh lmd_full
    

1.2 Melody completion and accompaniment suggestion datasets

  • Follow "PiRhDy: Learning Pitch-, Rhythm-, and Dynamics-aware Embeddings for Symbolic Music" (https://github.com/mengshor/PiRhDy) to generate datasets for melody completion task and accompaniment suggestion task, or download generated datasets directly.

    PiRhDy/dataset/context_next/train
    PiRhDy/dataset/context_next/test
    PiRhDy/dataset/context_acc/train
    PiRhDy/dataset/context_acc/test
    
  • Convert these two datasets to OctupleMIDI format with gen_nsp.py.

    python -u gen_nsp.py
    
  • The script should prompt you to input which downstream task to process. (next for melody task and acc for accompaniment task)

    task = next
    
  • Binarize the raw text format dataset. (this script will read next_data_raw folder and output next_data_bin)

    bash binarize_nsp.sh next
    

1.3 Genre and style classification datasets

  • Prepare The Lakh MIDI Dataset (LMD-full) in zip format. (say lmd_full.zip, ignore this step if you have lmd_full.zip)

    wget http://hog.ee.columbia.edu/craffel/lmd/lmd_full.tar.gz
    tar -xzvf lmd_full.tar.gz
    zip -r lmd_full.zip lmd_full
    
  • Get TOPMAGD and MASD midi to genre mapping midi_genre_map from "On large-scale genre classification in symbolically encoded music by automatic identification of repeating patterns".(DLfM 2018) (https://github.com/andrebola/patterns-genres)

    wget https://raw.githubusercontent.com/andrebola/patterns-genres/master/data/midi_genre_map.json
    
  • Generate these two datasets in OctupleMIDI format using the midi to genre mapping file with gen_genre.py.

    python -u gen_genre.py
    
  • The script should prompt you to input which downstream task to process. (topmagd for genre task and masd for style task)

    subset: topmagd
    LMD dataset zip path: lmd_full.zip
    sequence length: 1000
    
  • Binarize the raw text format dataset. (this script will read topmagd_data_raw folder and output topmagd_data_bin)

    bash binarize_genre.sh topmagd
    

2. Training

2.1 Pre-training

bash train_mask.sh lmd_full small
  • Download our pre-trained checkpoints here: small and base, and save in the checkpoints folder. (a newer version of fairseq is needed for using provided checkpoints: see issue-37 or issue-45)

2.2 Fine-tuning on melody completion task and accompaniment suggestion task

bash train_nsp.sh next checkpoints/checkpoint_last_musicbert_base.pt
bash train_nsp.sh acc checkpoints/checkpoint_last_musicbert_small.pt

2.3 Fine-tuning on genre and style classification task

bash train_genre.sh topmagd 13 0 checkpoints/checkpoint_last_musicbert_base.pt
bash train_genre.sh masd 25 4 checkpoints/checkpoint_last_musicbert_small.pt

3. Evaluation

3.1 Melody completion task and accompaniment suggestion task

python -u eval_nsp.py checkpoints/checkpoint_last_nsp_next_checkpoint_last_musicbert_base.pt next_data_bin

3.2 Genre and style classification task

python -u eval_genre.py checkpoints/checkpoint_last_genre_topmagd_x_checkpoint_last_musicbert_small.pt topmagd_data_bin/x