MusicTransformer

A Julia-based implementation of Music Transformer: Generating Music with Long-Term Structure.

The Music Transformer is an attention-based neural network that can generate music with improved long-term coherence. Using the event-based Performance representation, the Music Transformer generates expressive performances directly without first generating a score. A Julia-based implementation of the Performance representation can be found at NoteSequences.jl

Installation

MusicTransformer.jl uses NoteSequences.jl internally. Please install both of them to use the package.

] add https://github.com/VasanthManiVasi/NoteSequences.jl
] add https://github.com/VasanthManiVasi/MusicTransformer.jl

Example

using FileIO
using MusicTransformer
ENV["DATADEPS_ALWAYS_ACCEPT"] = true

music_transformer = pretrained"unconditional_model_16"

midi = generate(music_transformer, numsteps=1500)

save("generated.mid", midi)

See examples/Melody Conditioning for an example of using the Melody-Conditioned Music Transformer to generate an accompaniment given a melody.

Training

To train a new music transformer or to fine-tune one of the available models on your own collection of midi files, the midi files must first be converted to one-hot indices, which can then be fed to the model as inputs. Please refer to examples/MAESTRO Language Modelling/maestro_datagen.jl for an example script on converting midifiles to model inputs. The example provided performs data generation for training a piano performance language model on the MAESTRO dataset. Finally, see examples/MAESTRO Language Modelling/maestro_train.jl for a training script to train a language model on the processed MAESTRO dataset we obtained from data generation. The examples can be modified for training on your own data.

References

1. Music Transformer: Generating Music with Long-Term Structure - arxiv.org
2. Music Transformer - Google
3. Tensor2Tensor - Google