MIDI2vec: Learning Embeddings for MIDI Vector Space Representations

Compute graph embeddings from MIDI.

This library is described in detail in the following paper (bib)

Lisena P., Meroño-Peñuela A. & Troncy R. MIDI2vec: Learning MIDI Embeddings for Reliable Prediction of Symbolic Music Metadata. In Semantic Web Journal, Special issue on Deep Learning for Knowledge Graphs, vol.13, no.3, pp. 357-377, IOSPress, 6 April 2022 http://doi.org/10.3233/SW-210446

The experiments described in the paper are available here.

Pre-computed MIDI embeddings used in the paper are available in Zenodo.

Usage

The library is composed by 2 parts.

midi2edgelist

Convert a MIDI file into a graph, in edgelist format.

Requires Node.js:

cd midi2edgelist

npm install

node index.js -i <midi_folder>

Optional arguments:

• -i, --input. Input directory containing MIDI files. REQUIRED
• -o, --output Output directory for the edgelists. Default: ./edgelist
• -n, --note-groups Number of groups of simultaneous notes to be taken in account for each MIDI. For example, -n 300 uses the first 300 groups. Default: all.

The output is formed by 4 edgelist (notes, program, tempo and time signature) and a csv containing the mapping between the file names and the given identifiers.

edgelist2vec

It uses node2vec for computing the embeddings from the egelists.

pip install -r edgelist2vec/requirements.txt

python edgelist2vec/embed.py

Optional arguments:

• -i, --input Input graph (edgelists) path. Default: .\edgelist;
• -o, --output Output file name. Default: embeddings.bin;
• --walk_length Length of walk per source. Default: 10;
• --num_walks Number of walks per source. Default: 40;
• -p Return hyper-parameter (as in node2vec). Default: 1;
• -q Inout hyper-parameter (as in node2vec). Default: 1;
• --dimensions Number of dimensions. Default: 100;
• --window-size Context size for optimization. Default: 5;
• --iter Number of epochs in word2vec. Default: 5;
• --workers Number of parallel workers. Default: 0 (full use);
• --exclude Edgelists to be excluded from the computation.