recurrent-neural-notes

Exam project in the Deep Learning course at DTU.

A GRU-network model for next-step prediction of notes in the Nottingham folk-melody dataset.

The model is also expanded to feeding the previous prediction from the output-network as input to the GRU-input-network, so as to be able to generate new sequences of notes from an initial one.

The final documentation of the project and results can be found here.

Data pre-processing

From ABC-format to Music21 objects to zero-padded one-hot encoded vectors for each note (pitch and duration) in each melody (list of lists of vectors --> numpy array X=[M, N, F]=[Melodies, Notes, Features]). Here all notes in all melodies are represented by a duration tensor, X_d, with 14 features and a pitch tensor, X_d, with 35 features.

The next-step prediction GRU network models (with and without [orange] prediction input)

The learning curves for the 2 models with and without regularization.

Regularizing the GRU networks reduce overfitting, as seen by less span between training and validation accuracy curves.

Dropout: By leaving out notes along the melodies, a lossy noise and therefore a completion task is introduced to the models, so during training the next-step prediction will rely more on the previous GRU activations h^t-1 and the horizontal connections will be enhanced to make up for the missing input.

Prediction input: By feeding in the previous prediction, a stronger loss signal will traverse across the horizontal connections and vanishing gradients can be avoided.

Histogram over pitch (left) and duration (right) usage in all melodies.

Reconstructions of "The Fiddle Hill Jig" by Model 1 and 2.

GRU activations for two units in model 1.