The purpose of this project is to create a model which will predict the genre of a track 'Hip-Hop':0, 'Pop':1, 'Country':2, 'Rock':3, 'R&B':4, 'Dance/Electronic':5, 'Indie':6, 'Sleep':7, 'Jazz':8, 'Soul':9, 'Metal':10 based on the 13 audio features provided from the Spotify API

Music Genre Classification Link

Music classification has a wide range of uses from tagging new songs that were uploaded and profiling them automatically at scale to figuring out which genre a user listens to more and possibly having it as an input for a recommendation system

The dataset is created using the Spotify API. It downloads all the tracks from a playlist as Tracks object and for each track gets the features.

Each playlist is tagged a genre based on the name of the playlist

There are 13 audio features for each track, including confidence measures like acousticness, liveness, speechiness and instrumentalness, perceptual measures like energy, loudness, danceability and valence (positiveness), and descriptors like duration, tempo, key, time signature and mode.

The data is fetched from the analytics zone of big query dataset. Currently the model was built on 37015 tracks

• The Target variable was created based on the genre
• Data was cleaned to make sure there were not any NAs
• Duplicate tracks were removed (since more than one playlist can have the same track)
• Remove columns that were not required for modelling such as type,id,uri,track_href,analysis_url etc.

An Xgboost algorithm was used to predict the target variable

Grid search methodology was used for hyperparameter tuning. The optimal parameters came out to be:

{'max_depth': 100, 'max_features': 2, 'min_samples_leaf': 3, 'min_samples_split': 8, 'n_estimators': 100}

The model can be deployed locally using Docker and the following commands

docker build -t musicpred .
docker run -it --rm -p 9698:9698 musicpred

1. Create a project and enable cloud build and cloud run api
2. Add below code to docker file at the end

CMD exec gunicorn --bind :$PORT --workers 1 --threads 8 --timeout 0 predict:app

3. Change directory to where the Docker file resides i.e. 03_model_deployment/ and run the below commands

gcloud builds submit --tag gcr.io/{project_id}/predict
gcloud run deploy --image gcr.io/{project_id}/predict --platform managed

4. Change the url in the predict_test.py file to {url from google cloud run}/predict and run the script to get the prediction