This is a generative model from the paper "Byte Pair Encoding for Symbolic Music" (EMNLP 2023). The model has been trained with Byte Pair Encoding (BPE) on the Maestro dataset to generate classical piano music with the REMI tokenizer.
It has a vocabulary of 20k tokens learned with Byte Pair Encoding (BPE) using MidiTok.
- Developed and shared by: Nathan Fradet
- Affiliations: Sorbonne University (LIP6 lab) and Aubay
- Model type: causal autoregressive Transformer
- Backbone model: GPT2
- Music genres: Classical piano 🎹
- License: Apache 2.0
- Repository: https://github.com/Natooz/BPE-Symbolic-Music
- Paper: https://arxiv.org/abs/2301.11975
The model is designed for autoregressive music generation. It generates the continuation of a music prompt.
Use the code below to get started with the model.
You will need the miditok
, transformers
and torch
packages to make it run, that can be installed with pip.
You will also need to manually download the tokenizer.conf
file from the repo files.
import torch
from transformers import AutoModelForCausalLM
from miditok import REMI
from miditoolkit import MidiFile
torch.set_default_device("cuda")
model = AutoModelForCausalLM.from_pretrained("Natooz/Maestro-REMI-bpe20k", trust_remote_code=True, torch_dtype="auto")
tokenizer = REMI(params="tokenizer.conf")
input_midi = MidiFile("path/to/file.mid")
input_tokens = tokenizer(input_midi)
generated_token_ids = model.generate(input_tokens.ids, max_length=200)
generated_midi = tokenizer(generated_token_ids)
generated_midi.dump("path/to/continued.mid")
The model has been trained on the Maestro dataset. The dataset contains about 200 hours of classical piano music. The tokenizer is trained with Byte Pair Encoding (BPE) to build a vocabulary of 20k tokens.
- Training regime: fp16 mixed precision on V100 PCIE 32GB GPUs
- Compute Region: France
The following hyperparameters were used during training:
- learning_rate: 0.0001
- train_batch_size: 64
- eval_batch_size: 96
- seed: 444
- optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
- lr_scheduler_type: cosine_with_restarts
- lr_scheduler_warmup_ratio: 0.3
- training_steps: 100000
We cannot estimate reliably the amount of CO2eq emitted, as we lack data on the exact power source used during training. However, we can highlight that the cluster used is mostly powered by nuclear energy, which is a low carbon energy source ensuring a reduced direct environmental impact.
BibTeX:
@inproceedings{bpe-symbolic-music,
title = "Byte Pair Encoding for Symbolic Music",
author = "Fradet, Nathan and
Gutowski, Nicolas and
Chhel, Fabien and
Briot, Jean-Pierre",
booktitle = "Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing",
month = dec,
year = "2023",
address = "Singapore",
publisher = "Association for Computational Linguistics",
url = "https://arxiv.org/abs/2301.11975",
}