We have investigated the classification of different textural elements in orchestral symbolic music data. A simple convolutional neural network (CNN) is utilized to perform the classification task in a track-wise and bar-wise manner. Preliminary results are reported, and different training parameters, including the use of contextual data and the combination of tracks, are also discussed.
- Piano roll dataset can be found at
/dataset - Modify different model settings in
train.py, then train and validate the model
| File / folder | Description |
|---|---|
train.py |
Trains and evaluates the model input and training settings can be modified in this file |
make_piano_rolls.py |
Parse .musicxml files and convert to piano rolls as numpy arrays |
organize_annotations.py |
Parse .orch files and store the labels as numpy arrays |
/data_processing/pianoRoll.py |
Object defined to handle, access and manipulate piano rolls |
/data_processing/PianoRollsDataset.py |
Object extend to Pytorch Dataset object to manage the selection of training examples by bar |
/src/ |
Contains code for parsing .orch files |
/model/ |
Stores trained models |
/dataset/ |
Including annotations, .musicxml, MIDI, and converted pianorolls (.npy files) |
/result/ |
Performance metrics saved here |
- We use the dataset provided by Le, Dinh-Viet-Toan, et al. "A Corpus Describing Orchestral Texture in First Movements of Classical and Early-Romantic Symphonies." Proceedings of the 9th International Conference on Digital Libraries for Musicology. 2022. [Paper] [Gitlab Repo]
- We parse .musicxml files and convert into pianorolls
| Folder | Content |
|---|---|
/annotaions/ |
Labels of each pieces (numpy arrays) are stored in .npy files, the original .orch files are in /{nameOfAuthor} folders |
/score_xml/ |
.musicxml files : original digital scores |
/scores_midi/ |
MIDI files converted from .musicxml files are sourced |
/score_pianoroll/ |
Piano roll (numpy array) are stored as .npy files |