Bootleg Score Composer Recognition Challenge 🏆
To regenerate the datasets, navigate to dataset_creation.ipynb and change the data_path variable to wherever you would like to save the dataset files. Run the notebook all the way through to get your own 9_way_dataset.pkland 100_way_dataset.pkl. Note that regenerating the datasets requires you to have the predictions from the filler classifier, like ensemble_imslp.tsv found in the filler folder.
The 9-way dataset has 9 hand-selected composers for being generally well known, whereas the 100-way composers has the top 100 composers with the most registered bootleg score events. You can see the lists for both sets of composers in config.
When unpickled, these files store a tuple in the format of (x_train, y_train, meta_train, x_valid, y_valid, meta_valid, x_test, y_test, meta_test).
"x" represents the input features formatted as a list of numpy arrays.
"y" represents the labels formated as a list of integers. The integer values is assigned by the index of the composer after being sorting all of them alphabetically.
"meta" represents the metadata for the fragment, formatted as a tuple of (ID, start_offset). The ID represents the unique ID assigned to the PDF the fragment was grabbed from on the IMSLP archive. The start offset is the starting index of the fragment in the overall bootleg score of the PDF.