Workflow Overview

In main.py, there are many parameters to customize regarding the analytical styles, types of machine learning models, model architectures, and hyper-parameters. I will provide a chart introducing all the available combinations of these parameters in the section of “Parameter Adjustment”.

For now, the project can accept Bach chorales, and the corresponding annotations from here based on my co-authored paper with Nathaniel Condit-Schultz called “A Flexible Approach to Automated Harmonic Analysis: Multiple Annotations of Chorales by Bach and Praetorius”, where the music and annotations are encoded in .krn files. Afterward, a series of functions are applied to pre-process the data to feed the machine learning models as inputs and outputs for training, and then the model will predict non-chord tones and the corresponding chord labels on the test set, presented as musicXML files for users to see the end results. Specifically, in the script of main.py:

• extract_chord_labels function is used to extract the chord labels from the .krn files and save them as .txt files. For example, for the annotations of the maximally harmonic style our ISMIR paper used, they can be found in the repository as .krn files located under ./genos-corpus/answer-sheets/bach-chorales/New_annotation/ISMIR2019/. They are in the format of Chorales_Bach_XXX.krn, and the function extract_chord_labels generates .txt files (for chord labels) in the same directory with the same name.
• annotation_translation function is used to translate the chord syntax of the annotations into the ones recognized by music21, a package which this project will rely on to do various kinds of music processing. The resuting files are named as translated_Chorales_Bach_[CHORALE_ID]ISMIR2019.txt.
• provide_path_12keys will transpose the chord annotations into 12 possible keys, so we can use either the annotations in the key of C major or A minor (without data augmentation), or use the annotations in all keys (with data augmentation) to train the model. Correspondingly, transpose_polyphony is used to translate the music into 12 keys. These files will all be saved in the directory of ./genos-corpus/answer-sheets/bach-chorales/New_annotation/ISMIR2019/. The resulting files are named as transposed_KB[KEY]KEtranslated_Chorales_Bach_[CHORALE_ID]ISMIR2019.txt.
• Once we have the music and the chord labels ready, generate_data_windowing_non_chord_tone_new_annotation_12keys will translate all of them into an encoding (one-hot) format which can be directly used by the machine learning models. These files will be saved in the directory of ./data_for_ML.
• Last, train_and_predict_non_chord_tone will train the machine learning model, save all the models, output all the performance information into the text file (they are saved in the directory of ./ML_result/), and visualize all the predicted non-chord-tones as well as the chord labels through the generated musicXML files (saved in the directory of ./predicted_result).