Finding Tori: Self-supervised Learning for Analyzing Korean Folk Song (ISMIR 2023)

This is the official code repository for our ISMIR 2023 paper.

Anthology of Korean Traditional Folksongs

The metadata we collected from the website is available in metadata.csv, including the url for each song. You can download the audio files from the website.

Contour Dataset

We used CREPE to extract the F0 contours from the audio dataset. You can download pre-processed CSV files from the google drive link

gdown gdown 18G6QaIqruyQBdJeMX9bjBasi5PFDuxh-
unzip finding_tori_contour_csv.zip

This will make contour_csv/ directory that includes csv files.

Requirements

Pipenv Environment

We used pipenv to manage the python environment. You can install the environment by running the following command.

pip install pipenv
pipenv install
pipenv shell

(Python Packages)

If you don't want to use pipenv, you can install the packages that are listed in requirements.txt. You can install them by running the following command.

pip install -r requirements.txt

Reproduce the Result

You can get the result of table 1 in our paper by running python3 get_eval_result.py. For CNN models, pre-trained weights are available in pretrained_weights/ directory.

The script gives nDCG and RF classifier accuracy in two results. Including others class or not.
To get result of Pitch Histogram with 25 bins, run python3 get_eval_result.py --use_histogram --resolution=1
To get result of Pitch Histogram with 124 bins, run python3 get_eval_result.py --use_histogram --resolution=0.2
To get result of Region-supervised CNN model, run python3 get_eval_result.py --model=region-trained
To get result of Self-supervised CNN model, run python3 get_eval_result.py --model=self-supervised
If you want to use your own trained model, you can specify the path of the model by --ckpt_path=path_to_model_state.pt option.

Train the model

You can also train the CNN model again with CSV datasets.

python3 train.py                          # train the self-supervised model
python3 train.py exp="terrain_classifier" # train the region-supervised model

The train.py uses hydra package to handle configuration. The checkpoint will be saved in ./experiments_checkpoints/

After the training, you can get the evaluation result by running code below

python3 get_eval_result.py --ckpt_path=experiments_checkpoints/{date_time_model_code}/model_state.pt

danbinaerinHan/finding-tori