Python code for reproducing music genre translation experiments presented in the paper Leveraging knowledge bases and parallel annotations for music genre translation at ISMIR 2019.
- The directory "scripts" contains standalone python scripts.
- The directory "tag_translation" contains the python module used to perform tag translation.
- The directory "docker" contains the Dockerfile that defines the image in which the experiment's code can be run.
Make sure you have Docker installed, and run
git clone [email protected]:deezer/MusicGenreTranslation.git
cd MusicGenreTranslation
./build_and_run.sh
This will:
- build a docker image called
tag_translation_researchin which the experiments will be run. - launch a container where all the scripts to produce the articles' figures will be run.
- download all the groundtruth files for train and validation data in the common recordings dataset, along with the taxonomies provided by acousticbrainz for the 3 sources lastfm, discogs and tagtraum. Additionally, a mapping between recodingmbids and artistmbids will also be downloaded. The artist information is need to perform the stratified sampling when generating folds.
- run all the necessary scripts to generate the final plots
After this, if everything went well, the figures will be in data/plots.
/!\ Because running all the results is quite long, the plots are generated only
for the lowest amounts of data, but you can change that setting in
tag_translation/conf.py.
To have better control of how the scripts are executed as well as the
environment you can refer to tag_translation/conf.py.