This repository contains python scripts which can be used to convert MIDI files to color-coded images (and vice-versa) for the purpose of training an image-generation model to write music.

• Python 3
• Python packages
  • Pillow
  • os
  • Music21
  • midiutil

To create images from a directory containing MIDI files, run Midi_to_Images.py, changing the filepaths to correspond to your input and ouput directories. Each track in the MIDI file will be color coded red, green, or blue.

Tiler.py takes these images and crops them into seperate smaller images of a specified length to be used as a dataset for training a generative image model (e.g. PixelCNN++). By default, Tiler.py will crop any notes with pitches outside of the range of the 88 keys on a piano to create smaller images which will lead to faster training times. You may need to change this if you are using MIDI files with other instruments that have a greater range in pitch.

Image_to_Midi.py converts a single image of height 128 to a MIDI file by horizontal lines of pixels as notes with pitch corresponding to the height in the image.

Below are some of the results I was able to achieve after training a PixelCNN model based on work by Tim Salimans, Andrej Karpathy, Xi Chen, and Diederik P. Kingma for 50 epochs on a dataset of over 2,000 training images created from Bach's Goldberg Variations, WTC books 1 and 2, The Art of Fugue, 400+ Chorales, two-part inventions, and other miscellaneous keyboard works.

This approach has also performed reasonably well with other piano music, the the two examples below were generated after training PixelCNN on a larger dataset of over 7,000 images created from Beethoven, Mozart, and Schubert's Sonatas, and other Romantic-era piano pieces.