Skip That Beat: Augmenting Meter Tracking Models for Underrepresented Time Signatures

Paper | Demo | Code

Datasets

Supported datasets:

• Beatles
• BRID
• GTZAN
• GTZAN Rhythm
• RWC Classical
• RWC Jazz

Dataset parsing

We use a custom Dataset class based on mirdata's, so we are able to load original and augmented data. This class requires the data to be structured in audio and annotation folders.

We provide a bash script that parse all the datasets (parse_datasets.sh) to the correct structure and infer the meter from the beat annotation data. You can change the input values for the needed variables in the script and then run it with ./parse_datasets.sh. This will make the data be in the correct location.

Augmentation

Once the datasets are in the correct format, we can also augment them. To do so, run the augment_dataset.py script.

To augment all supported datasets to all meters:

python augment_dataset.py \
	--data_home /path/to/datasets

To augment specific datasets to specific meters:

python augment_dataset.py \
    --data_home /path/to/datasets \
    --datasets gtzan beatles \
    --target_aug 24

IMPORTANT your dataset name should match the folder name you chose when parsing the datasets, otherwise the script will break. We recommend keeping the folder names as "gtzan" (GTZAN), "beatles" (Beatles), "rwcc" (RWC Classical) and "rwcj" (RWC Jazz).

Create splits

To create the splits shown in the figure above, just run

python create_splits.py \
	--data_home /path/to/datasets \
	--splits_home /path/to/output/splits

This will save your splits in the splits_home folder.

Run BayesBeat experiments

1. Make sure you have MATLAB installed
2. Clone the bayesbeat GitHub repo
3. Inside bayesbeat_experiments/bayes_beat_training.m file, replace the base_path variable (L46) to the location where you saved the repo in step 2
4. Run it!

The BayesBeat inference is the same process, but you change the bayesbeat path on L41 of the inference.m file and run it.

Run TCN experiments

To train the TCN, run the tcn.py script providing the path to your dataset folder and the experiment you wish to reproduce (baseline, augmented_sampled or augmented_full)

python run_tcn_experiment.py \
	--data_home /path/to/datasets \
	--splits_home /path/to/splits \
	--experiment exp

This will train and test the TCN in both BRID and the regular test set.