iRealPro-Chart-Generator

Generating jazz music charts for the iRealPro mobile app.

Dependencies

1. numpy
2. pandas
3. keras
4. sklearn
5. pychord
6. pyRealParser
7. pyrealpro
8. progress

How to run

1. Open chart_generator.py
2. Modify props at the bottom of the file, specifying:
   1. key The key of the chart you want to generate
   2. starting_chords The first four starting chords of the chart
   3. style The Style of the Chart or None
   4. influenced_composer The Composer of the chart or None
   5. generated_chord_count The number of generated chords, on top of the provided 4
3. Run the file, and check the console for the iRealPro url. This can be opened in Safari to load the chart into your app.

Sample Output

irealbook://AI%20Generated%20Chart%20%231=M.C.%20ChartGeneratorAI=Latin=C=n=%5BT44C69%2C%20%2C%20%2C%20%7CDbdi-7%2C%20%2C%20%2C%20%7CD-7%2C%20%2C%20%2C%20%7CG13b9%2C%20%2C%20%2C%20%7CF6%2C%20%2C%20%2C%20%7CE-7%2C%20%2C%20%2C%20%7CA7b9%2C%20%2C%20%2C%20%7CD-7%2C%20%2C%20%2C%20%7CD-7%2C%20%2C%20%2C%20%7CF-7%2C%20%2C%20%2C%20%7CBb7%2C%20%2C%20%2C%20%7CEb7%2C%20%2C%20%2C%20%7CDb7%2311%2C%20%2C%20%2C%20%7CC7%2C%20%2C%20%2C%20%7CC7%2C%20%2C%20%2C%20%7CC7%2C%20%2C%20%2C%20Z

How it works

Model Input

The model works by trying to predict the fifth chord based on the previous 4 chords. This accumulates to 8 inputs in the Neural Network as each chords is identified by its root note, and its quality. Another 2 inputs are added, the style (i.e. Latin, Medium Swing) and the composer.

Model Output

The outputs works best by using a categorical approach. the total output layer is comprised of 114 neurons (12 notes 102 Qualities).

Dataset

This structure results in a training dataset size of 43575, and a test dataset size of 18675. The provided model is trained based on the jazz 1400 library from the iRealPro Forum, but this could be used for any style of music that iRealPro supports.

Improvements

1. When training, all of the pieces are transposed to a common key of C Major (or A minor), as it makes the model only cares about the chords relations, not the absolute chords. This means that when generating a chart, the provided chords are transposed, then output charts is inverse transposed back to your desired key.

2. I originaly encountered an issue where the generated chords will get stuck in a loop, and then the same chord will get generated over and over again. A solution to fix this is (yet to implement):

   1. Don't train the neural network on repeated chords
   2. Create Multiple models with different lags (default 4), then a lag 6, lag 8, lag 10 and so on. When the chart is being generated, use the appropriate model based on how many chords already exist in the chart.

Supported Styles

inspect dp.styles_encoder.classes_

Supported Composers

inspect dp.composers_encoder.classes_