Midi2Vec

Midi2Vec is a framework for converting MIDI files to sequences of embeddings and using machine learning models to learn and predict from these sequences. It is composed of the following principle submodules:

• data_loading: module for loading collections of MIDI files and framing the musical sequences they contain as time-series prediction problems
• encoding: module for training vector representations on a collection of MIDI files and for converting between vector and textual representations of musical notes
• evaluation: module for evaluating sequence learning models
• optimization: module for optimizing over a range of model hyperparameter combinations
• sequence_learning: module for sequence learning and prediction of vector-encoded MIDI note sequences
• pipeline: high-level API that allows all other modules to be pieced together in an intuitive way

Examples

The following example shows how to configure and run a GenerativePipeline, which creates a vector model from a corpus of MIDI files, encodes a selection of MIDI files into a sequence of vectors using this model, learns the patterns in this second set of MIDI files and evaluates its accuracy:

training_docs = "../resources/encoder_training_docs/full_1_measure_20k.txt"

pipeline_params = {

    # Encoder (Doc2vec) settings:
    'doc2vec_docs': encoder_training_docs,
    'doc2vec_dm': 1,
    'doc2vec_dm_mean': 1,
    'doc2vec_epochs': 1,
    'doc2vec_hs': 0,
    'doc2vec_learning_rate_start': 0.025,
    'doc2vec_learning_rate_end': 0.2,
    'doc2vec_min_count': 5,
    'doc2vec_negative': 0,
    'doc2vec_vector_size': 8,
    'doc2vec_window': 5,

    # Sequence learning (Keras LSTM) settings:
    'nn_features': ['bpm', 'measure', 'beat'],
    'nn_batch_size': 16,
    'nn_dense_activation_function': "linear",
    'nn_dropout': 0,
    'nn_epochs': 1,
    'nn_hidden_neurons': 8,
    'nn_layers': 4,
    'nn_lstm_activation_function': "selu",
    'nn_lstm_n_prev': 16,
    'nn_loss': 'mean_absolute_error',
    'nn_optimizer': 'rmsprop'
}

# Define note mapper for MIDI file loading
note_mapping_config_path = "../settings/map-to-group.json"
note_mapper = NoteMapper(note_mapping_config_path)

# Data loader used to encode MIDI-format training files
data_loader = MidiDataLoader(note_mapper, params=pipeline_params)

# Define training documents for sequence learning
training_docs = ["../resources/midi/bach_chorales/01-AchGottundHerr.mid",
                 "../resources/midi/bach_chorales/02-AchLiebenChristen.mid",
                 "../resources/midi/bach_chorales/03-ChristederdubistTagundLicht.mid",
                 "../resources/midi/bach_chorales/04-ChristeDuBeistand.mid",
                 "../resources/midi/bach_chorales/05-DieNacht.mid",
                 "../resources/midi/bach_chorales/06-DieSonne.mid",
                 "../resources/midi/bach_chorales/07-HerrGott.mid",
                 "../resources/midi/bach_chorales/08-FuerDeinenThron.mid",
                 "../resources/midi/bach_chorales/09-Jesus.mid",
                 "../resources/midi/bach_chorales/10-NunBitten.mid"]

pipeline = GenerativePipeline(params=pipeline_params)
pipeline.set_data_loader(data_loader)
pipeline.set_training_docs(training_docs)
pipeline.set_k_fold_cross_eval(k=3)
pipeline.set_evaluator(F1Evaluator())

result_df = pipeline.run()

Several Jupyter notebooks demonstrating the use of Midi2Vec are available:

• Parameter sweep and evaluation of a GenerativePipeline: https://github.com/TaylorPeer/Midi2Vec/blob/master/notebooks/Hyperparameter%20Optimization%20and%20Evaluation.ipynb
• Encoder inspection: https://github.com/TaylorPeer/Midi2Vec/blob/master/notebooks/Midi2Vec%20Model%20Inspection.ipynb
• Sequence Generation: https://github.com/TaylorPeer/Midi2Vec/blob/master/notebooks/Sequence%20Generation.ipynb