Music Composition with Linear Transformers

Install Requirements

pip install -r requirements.txt

1. Data Download

1.1 Download VGMIDI dataset

Clone the VGMIDI dataset repository, go to the unabelled pieces directory and extract the midi.zip file:

git clone
cd vgmidi/unlabelled
unzip midi.zip

1.2 Split the dataset

Go to the src directory of unabelled pieces and run the midi_split script:

cd unlabelled/src
python3 midi_split.py --csv ../../vgmidi_unlabelled.csv --midi ../midi

2. Data Pre-processing

2.1 Data Augmentation

Augment the VGMIDI dataset with tranposition and time strech as defined by Oore et al. 2017:

python3 augment.py --midi <path_to_midi_data>

2.2 Data Encoding

Encode the VGMIDI dataset with the encoding scheme as defined by Oore et al. 2017:

python3 encoder.py --midi <path_to_midi_data>

3. Train Model

Train a fast-tranformer on the VGMIDI dataset:

python3 train.py --train vgmidi_augmented/train/ --test vgmidi_augmented/test --seq_len 2048 --lr 1e-04 --epochs 50 --d_query 32 --n_heads 8 --n_layers 8 --batch_size 16 --save_to trained/model_{}.pth

4. Generate Music

This repository includes a pre-trained model (trained/model40.pth). To generate music with that model:

python3 generate.py --seq_len 2048 --n_layers 8 --model trained/model40.pth --vocab_size 390 --k 45 --p 0.9

To generate music with a model you trained yourself, simply change the --model argument to point to the model saved during training.

lucasnfe/fast-transformer-music