Python package to tokenize MIDI music files, presented at the ISMIR 2021 LBD.
Using Deep Learning with symbolic music ? MidiTok can take care of converting (tokenizing) your MIDI files into tokens, ready to be fed to models such as Transformer, for any generation, transcription or MIR task. MidiTok features most known MIDI tokenizations (e.g. REMI, Compound Word...), and is built around the idea that they all share common parameters and methods. It supports Byte Pair Encoding (BPE) and data augmentation.
Documentation: miditok.readthedocs.com
pip install miditokMidiTok uses MIDIToolkit, which itself uses Mido to read and write MIDI files, and BPE is backed by Hugging Face 🤗tokenizers for super-fast encoding.
The most basic and useful methods are summarized here. And here is a simple notebook example showing how to use Hugging Face models to generate music, with MidiTok taking care of tokenizing MIDIs.
from miditok import REMI, TokenizerConfig
from miditoolkit import MidiFile
from pathlib import Path
# Creating a multitrack tokenizer configuration, read the doc to explore other parameters
config = TokenizerConfig(nb_velocities=16, use_chords=True, use_programs=True)
tokenizer = REMI(config)
# Loads a midi, converts to tokens, and back to a MIDI
midi = MidiFile('path/to/your_midi.mid')
tokens = tokenizer(midi) # calling the tokenizer will automatically detect MIDIs, paths and tokens
converted_back_midi = tokenizer(tokens) # PyTorch / Tensorflow / Numpy tensors supported
# Tokenize a whole dataset and save it at Json files
midi_paths = list(Path("path", "to", "dataset").glob("**/*.mid"))
data_augmentation_offsets = [2, 1, 1] # data augmentation on 2 pitch octaves, 1 velocity and 1 duration values
tokenizer.tokenize_midi_dataset(midi_paths, Path("path", "to", "tokens_noBPE"),
data_augment_offsets=data_augmentation_offsets)
# Constructs the vocabulary with BPE, from the token files
tokenizer.learn_bpe(
vocab_size=10000,
tokens_paths=list(Path("path", "to", "tokens_noBPE").glob("**/*.json")),
start_from_empty_voc=False,
)
# Saving our tokenizer, to retrieve it back later with the load_params method
tokenizer.save_params(Path("path", "to", "save", "tokenizer.json"))
# And pushing it to the Hugging Face hub (you can download it back with .from_pretrained)
tokenizer.push_to_hub("username/model-name", private=True, token="your_hugging_face_token")
# Applies BPE to the previous tokens
tokenizer.apply_bpe_to_dataset(Path('path', 'to', 'tokens_noBPE'), Path('path', 'to', 'tokens_BPE'))MidiTok implements the tokenizations: (links to original papers)
You can find short presentations in the documentation.
Contributions are gratefully welcomed, feel free to open an issue or send a PR if you want to add a tokenization or speed up the code. You can read the contribution guide for details.
- Extend unimplemented additional tokens to all compatible tokenizations;
- Control Change messages;
- Option to represent pitch values as pitch intervals, as it seems to improve performances;
- Speeding up MIDI read / load (using a Rust / C++ io library + Python binding ?);
- Data augmentation on duration values at the MIDI level.
If you use MidiTok for your research, a citation in your manuscript would be gladly appreciated. ❤️
[MidiTok paper] [MidiTok original ISMIR publication]
@inproceedings{miditok2021,
title={{MidiTok}: A Python package for {MIDI} file tokenization},
author={Fradet, Nathan and Briot, Jean-Pierre and Chhel, Fabien and El Fallah Seghrouchni, Amal and Gutowski, Nicolas},
booktitle={Extended Abstracts for the Late-Breaking Demo Session of the 22nd International Society for Music Information Retrieval Conference},
year={2021},
url={https://archives.ismir.net/ismir2021/latebreaking/000005.pdf},
}The BibTeX citations of all tokenizations can be found in the documentation
Special thanks to all the contributors. We acknowledge Aubay, the LIP6, LERIA and ESEO for the initial financing and support.