This package contains a pipeline for analyzing/processing a set of midi files and building a dataset from the analysis.

It also provides a compact interface for accessing such a dataset.

Analysis pipeline

The analysis pipeline requires a source directory containing MIDI files as input and outputs a dataset directory containing analysis results.

Before running the pipeline, you need to:

1. Install this package (for local installation, pip install -e .)
2. Prepare a soundfont file on your disk
3. Prepare the source folder

Run the pipeline:

python run.py --path <path_to_dataset_dir> --src <path_to_src_dir> --verbose\
 -sape\
 --soundfont <path_to_soundfont>

The pipeline provides 4 actions:

1. --sync, -s: Sync the notes to the beat
   • Input: midi
   • Output: synced_midi
2. --align, -a: Align the MIDI files to C major/A minor
   • Input: midi
   • Output: synced_midi (same as -s. If -a is used together with -s, their effects will both be applied to synced_midi)
3. --to_pianoroll, -p: Convert the processed MIDI files to pianoroll format
   • Input: synced_midi
   • Output: pianoroll
4. --extract_features, -e: Extract features from the pianorolls
   • Input: pianoroll
   • Output: many

Combinations of actions:

• -sape full pipeline
• -spe full pipeline without alignment
• -sa outputs synced and aligned midi. No conversion to pianorolls

Dataset interface

In the dataset directory, each subdirectory contains one attribute of all pieces.

dataset_dir/
    midi/
        piece1.mid
        piece2.mid
    pianoroll/
        piece1.npz
        piece2.npz
    key/
        piece1.json
        piece2.json
    density/
        piece1.json
        piece2.json
    ...

The music_data_analysis.Dataset class provides an interface for accessing attributes of the pieces.

from music_data_analysis import Dataset

dataset = Dataset('path_to_dataset')
songs = dataset.songs() # list[music_data_analysis.Song]

# Get midi of the first song
midi = songs[0].read_midi("synced_midi") # miditoolkit.MidiFile

# Get pianoroll of the first song
pianoroll = songs[0].read_pianoroll("pianoroll") # music_data_analysis.Pianoroll

# Get key of the first song
key = songs[0].read_json("key") # dict

# Get note density of the first song
density = songs[0].read_json("density") # list[float], one value per bar

# Get song by name
song = dataset.get_song("piece1") # music_data_analysis.Song