Final project for MAIS 202: Accelerated Introduction to Machine Learning bootcamp, hosted by the McGill Artificial Intelligence Society.

Goal: Generate tonal piano music using a LSTM neural network with attention

Information about model architecture, challenges, data preprocessing and process can be found here: DevPost

Examples in can be found here or in the "examples" folder of the repository.

The start of each example is the starting sequence which was used to generated it (see devpost link above for details). A very high note is played to separate the starting sequence from the music generated by the model.

To get necessary packages: pip3 install -r "requirements.txt"

• Run generate_data.py with default parameters

• Run train.py

• Run generate_data.py with default parameters if not already done

• Run generate_music.py, changing weights_path, midi_output_dir and wav_output_dir as desired

• Change model architecture in model.py -> MusicGenerator -> create_model()

• Change loss functions or loss weights in model.py -> MusicGenerator -> train.py

• Change number of timesteps in params.py (requires rerunning generate_data.py)

Once changes are made, run train.py to train new model

• run generate_composer_data.py

• run app.py

Dataset can be found already separated in 'dataset/train-data' and 'dataset/test-data' folders

A subset of the dataset containing only pieces by Mozart and Haydn (useful for quickly testing models) can be found in 'dataset/smaller-dataset-test' and 'dataset/smaller-dataset-train'.

Data taken from here: http://www.piano-midi.de/

Train/test split can be done differently, but pretrained weights provided assume current train/test split

Deliverables for McGill Artificial Intelligence Society's Accelerated Introduction to Machine Learning Bootcamp found in 'Deliverables' folder

• Experiment with the use of transformers
• Augment dataset by slightly increasing/decreasing pitch/duration of each note
• Use the MAESTRO dataset and have velocity as a regression problem instead of a binary classification problem to emulate human performance