deepbach-pytorch

pip install deepbach-pytorch==1.5.0

For detailed usage instructions for the package api, please refer to test.py

What's new in 1.5.0

No public API was removed or renamed. harmonize() gained two keyword arguments, both with defaults, so existing calls behave as before.

Input robustness. harmonize() now accepts a melody-only file. It picks the part that actually carries notes rather than trusting the part count, so an exported lead sheet that ships one real staff plus three empty ones no longer fixes silence and regenerates the wrong voice. A single-part input used to produce a (1, length) tensor and crash the sampler; the melody is now encoded with the trained range and vocabulary of the slot it will occupy and the other voices are filled with rests. Metadata is padded or truncated to the melody length, which fixes the assertion failure on imported scores with pickups or trailing measures, and empty or zero-length input now raises a named error instead of failing deep inside the sampler.

Fixing an inner voice. melody_voice was documented as 0/1/2/3 but the generation span was built with [min(others), max(others) + 1], which for melody_voice=1 or 2 expanded to all four voices and silently regenerated the melody it was supposed to hold fixed. The voices to regenerate are now enumerated explicitly through a new voice_indices argument on DeepBach.generation(), so any voice can be held fixed.

Cadential cue. An imported melody normally carries no fermata, which left the fermata metadata channel — the model's only explicit phrase-ending cue — at zero for the whole piece. derive_fermatas=True (default) fills it with the convention the model was trained on: the last beat of every two bars plus the final note. Worth roughly +3.3 points of note-exact reconstruction accuracy on Bach chorales (70.2% -> 67.9% of the gap to a true channel; leaving it zeroed is 64.6%). Pass False for the previous behaviour. The channel is a conditioning input, not part of the generated music, so fermata_marks controls how it is written out: "none" (default) writes the XML with no fermata marks, "input" keeps only the input file's own, "channel" draws the channel as-is. Purely notational — a music21 Fermata carries no duration, so removing one changes no note's offset, length or pitch.

Range diagnostic. DeepBach encodes any note outside the trained range of the voice it occupies as a single out-of-range token, losing the pitch. A melody well below C4 therefore reaches the model as a nearly constant line. harmonize() now reports how many notes fall outside the fixed voice's range and suggests the voice that fits better, instead of failing silently. The melody is never transposed for you.

Two metadata fixes. FermataMetadata.generate() tested len(note.expressions) == 1, which reports any single expression as a fermata and misses a real fermata sharing a note with another expression; it now tests for music21.expressions.Fermata directly. It also indexed list_notes[0] on a part with no notes and raised IndexError. The MuseScore server no longer returns the conditioning channel painted onto the soprano of its output.

DeepBach

This repository contains implementations of the DeepBach model described in

DeepBach: a Steerable Model for Bach chorales generation
Gaëtan Hadjeres, François Pachet, Frank Nielsen
ICML 2017 arXiv:1612.01010

The code uses python 3.9 together with PyTorch v2.0 and music21 libraries.

For the original Keras version, please checkout the original_keras branch.

Examples of music generated by DeepBach are available on this website

Models, Dataset caches and Deployment script are available on Google Drive

Installation

You can clone this repository, install dependencies using Anaconda and download a pretrained model together with a dataset
with the following commands:

git clone git@github.com:Ghadjeres/DeepBach.git
cd DeepBach
conda env create --name deepbach_pytorch -f environment.yml
bash dl_dataset_and_models.sh

This will create a conda env named deepbach_pytorch.

music21 editor

You might need to Open a four-part chorale. Press enter on the server address, a list of computed models should appear. Select and (re)load a model. Configure properly the music editor called by music21. On Ubuntu you can eg. use MuseScore:

sudo apt install musescore
python -c 'import music21; music21.environment.set("musicxmlPath", "/usr/bin/musescore")'

For usage on a headless server (no X server), just set it to a dummy command:

python -c 'import music21; music21.environment.set("musicxmlPath", "/bin/true")'

Usage

Usage: deepBach.py [OPTIONS]

Options:
  --note_embedding_dim INTEGER    size of the note embeddings
  --meta_embedding_dim INTEGER    size of the metadata embeddings
  --num_layers INTEGER            number of layers of the LSTMs
  --lstm_hidden_size INTEGER      hidden size of the LSTMs
  --dropout_lstm FLOAT            amount of dropout between LSTM layers
  --linear_hidden_size INTEGER    hidden size of the Linear layers
  --batch_size INTEGER            training batch size
  --num_epochs INTEGER            number of training epochs
  --train                         train or retrain the specified model
  --num_iterations INTEGER        number of parallel pseudo-Gibbs sampling
                                  iterations
  --sequence_length_ticks INTEGER
                                  length of the generated chorale (in ticks)
  --help                          Show this message and exit.

Examples

You can generate a four-bar chorale with the pretrained model and display it in MuseScore by simply running

python deepBach.py

You can train a new model from scratch by adding the --train flag.

Usage with NONOTO

The command

python flask_server.py

starts a Flask server listening on port 5000. You can then use NONOTO to compose with DeepBach in an interactive way.

This server can also been started using Docker with:

docker run -p 5000:5000 -it --rm ghadjeres/deepbach

(CPU version), with or

docker run --runtime=nvidia -p 5000:5000 -it --rm ghadjeres/deepbach

(GPU version, requires nvidia-docker.

Usage within MuseScore

Deprecated

Put deepBachMuseScore.qml file in your MuseScore plugins directory, and run

python musescore_flask_server.py

Open MuseScore and activate deepBachMuseScore plugin using the Plugin manager. You can then click on the Compose button without any selection to create a new chorale from scratch. You can then select a region in the chorale score and click on the Compose button to regenerated this region using DeepBach.

Issues

Music21 editor not set

music21.converter.subConverters.SubConverterException: Cannot find a valid application path for format musicxml. Specify this in your Environment by calling environment.set(None, '/path/to/application')

Either set it to MuseScore or similar (on a machine with GUI) to to a dummy command (on a server). See the installation section.

Citing

Please consider citing this work or emailing me if you use DeepBach in musical projects.

@InProceedings{pmlr-v70-hadjeres17a,
  title = 	 {{D}eep{B}ach: a Steerable Model for {B}ach Chorales Generation},
  author = 	 {Ga{\"e}tan Hadjeres and Fran{\c{c}}ois Pachet and Frank Nielsen},
  booktitle = 	 {Proceedings of the 34th International Conference on Machine Learning},
  pages = 	 {1362--1371},
  year = 	 {2017},
  editor = 	 {Doina Precup and Yee Whye Teh},
  volume = 	 {70},
  series = 	 {Proceedings of Machine Learning Research},
  address = 	 {International Convention Centre, Sydney, Australia},
  month = 	 {06--11 Aug},
  publisher = 	 {PMLR},
  pdf = 	 {http://proceedings.mlr.press/v70/hadjeres17a/hadjeres17a.pdf},
  url = 	 {http://proceedings.mlr.press/v70/hadjeres17a.html},
}