Music Generation Project

This project aims to generate music using a deep learning model trained on MIDI files. It involves processing MIDI files, training a recurrent neural network (RNN) model, and using the trained model to generate new music sequences, which are then converted back into MIDI format.

Overview

The project is structured into several key scripts:

• midi_to_preprocess.py: Converts raw MIDI files into a processed format suitable for training, including encoding of MIDI events and creating training sequences.
• augmentation.py: Augments data to desired augmentation or dimunition to artificially increase training set
• features.py: Extracts features and appends them to .npz files. These features include harmony and intensity profile. More may be added.
• validations/midi_to_human_readable.py: Converts MIDI files into a human-readable format for debugging and understanding the MIDI data structure.
• validations/check_npz.py: Converts chosen .npz file into text files for debugging.
• validations/intensity_profile.py: Examines intensity of piece(s) as a function of velocity, tempo, tension, and pitch aggregation. For research purposes.
• train_model.py: Trains a multi-output RNN model on the processed MIDI data, capable of predicting pitch, velocity, and time delta for the next event in a sequence.
• rnn_model.py: Defines the RNN model architecture used for training in train_model.py.
• generate.py: Uses the trained model(s) to generate new music sequences based on a random seed or a given input sequence.
• rules/...: These are "dissolvable" rules which aid music generation. They are dissolvable because they will eventually not be invoked as training improves.
• seeds/...: These are different starting sequences that can be invoked.
• midi_conversion.py: Converts the generated music sequences back into MIDI format, ready for playback.

Prerequisites

• Python 3.6 or newer
• TensorFlow 2.x
• NumPy
• mido
• pretty_midi (for midi_to_human_readable.py only)

Installation

1. Clone this repository to your local machine.

2. Install the required Python packages by running:

   pip install -r requirements.txt
   

Usage

Preprocessing MIDI Files

1. Place raw MIDI files in the ./raw directory.

2. Run midi_to_preprocess.py to convert raw MIDI files into processed sequences:

   python midi_to_preprocess.py
   

3. Run additional preprocessing such as augmentation.py and features.py:

   python augmentation.py
   python features.py
   

Training the Model

1. After preprocessing, use train_model.py to train the RNN model:

   python train_model.py
   

2. Trained models are saved in the ./outputs/models directory.

Generating Music

1. Use generate.py to generate new music sequences with the trained model:

   python generate.py
   

2. Generated MIDI files are saved in the ./outputs/generated_music directory.

Converting MIDI to Human-Readable Format

1. Run midi_to_human_readable.py to convert MIDI files into a human-readable text format for debugging:

   python midi_to_human_readable.py
   

Customization

• You can adjust model parameters, sequence length, and other configurations by modifying the respective scripts.
• The generation script (generate.py) allows for customization of the seed sequence, number of generated events, and ensemble model usage.

License

MIT License