This project uses a CNN autoencoder model to generate trill suggestions from a midi input primer. You can upload and convert your own MIDI data in main.py and model() will generate a prediction of where a trill could go in the midi data.

The goal of this project is to be able to provide ornamentation suggestions for composers, when they are composing music on MIDI supported score notation software. This is currently a proof of concept unfinished project - at the moment, trill suggestions are given by training a model using the velocities given in MIDI data as a positional marker - which is an arduous and painstaking process as you have to create your own training dataset marking the position of trills in a scorewriter, e.g. Musescore, Sibelius, or a DAW, e.g. Logic, Ableton.

As it stands, the output suggests trills, but also completely changes your MIDI input, SO TAKE NOTE.

• Download the repository 'CC/' using

• cd to 'CC/' and activate virtual environment

• Install required libraries/packages using pip install -r requirements.txt in shell

• Check to make sure the following package is installed: https://github.com/craffel/pretty-midi

For generation of trills from a primer MIDI input, you will need to provide a MIDI data file, that can be inputted in main.py. A collection of random MIDI files can be found in the CC/midi_primer directory, which can fufill this purpose.

If you would like to train the model using your own dataset of midi files, these files need to be pre-pre-processed, as in, all the velocities in the MIDI data need to be set to 80 (comes up as 0 in Musescore) for the X input, and 80 for 'non-trills' and 127 for trills in the Y label input.

These MIDI files should then be uploaded to 'data/' and running preprocess_data.py will create a pickle file
containing all the training data to upload into the model.

If you are using the pre-trained model to generate trill suggestions, upload a MIDI file to main.py, and you should be able to generate a MIDI output containing the trill suggestions marked by velcity values of 127 (as opposed to 'non-trills' - which have a velocity value of 80) which can be shown when uploaded to a scorewriter like Musescore.

The model is best trained using GPUs with CUDA.

To train the model on your own dataset, please pre-process your data in the preprocess_data.py script. Training takes place in model.py and various hyperparameters can be adjusted.

Output of the model can be obtained by running the main.py script. This is only possible if there is a trained model available called my_trained_model.pt in the CC directory.