Album Art Classifier

A pre-trained tensorflow model to predict music genre
Created by Amanda Bowers

About the Model/ Training Set

This model was trained on Inception V3 via transfer learning:

iTunes genres and corresponding subgenres were grouped together to create 36 image classes:

Explore some of the artwork below:



Blues - B. B. King: Paying the Cost to be the Boss (2000)
Classical - Brahms: Violin Sonata in G Major, Op. 78 (1878)
Country - Kenny Chesney: Somewhere With You (2010)
Dance - La Roux: Bulletproof (2015)
Disney - Ariel: Happy Birthday, Princess (2005)
Electronic - Kuba: Hole In My Sitar (2012)
Jazz - Gregory Porter: Consequence of Love (2017)
Pop - Adele: Hello (2015)
R&B/Soul - Destiny's Child: Say My Name (1999)
Hip-Hop - Drake: Controlla (2015)
Reggae - Bob Marley: One Love/People Get Ready
Rock - Elvis Presley: Jailhouse Rock (1958)

More Artwork

Model Performance

Model Performance was evaluated for the training, validation and test sets


Training Set: 36 classes each with an average of 750 images (unbalanced classes)

To see how the model performed on each image displayed above, click here
Training accuracy: ~45% (varies iteratively)

Validation Set: 36 classes each with an average of 94 images

Validation accuracy: 28% (varies iteratively)

Test Set: 36 classes each with an average of 94 images

Test accuracy: 17.1% (final hold-out set)

What if your favorite album cover was in the test set?


Upload your image below to seee how the model would perform





Contact

Feel free get in touch, ask questions, or even just say hello!

bowersa@hawaii.edu