Loading
Cookies help us deliver our services. By using our services, you agree to our use of cookies. Learn more

crowdAI is shutting down - please read our blog post for more information

WWW 2018 Challenge: Learning to Recognize Musical Genre

Learning to Recognize Musical Genre from Audio on the Web


Completed
728
Submissions
416
Participants
23778
Views

Precomputed features

Posted by stereo over 2 years ago

Is there a precomputed features file? I was trying to run the SVM baseline and used the features.csv from fma_metadata but the indices don’t match and the assert fails. Computing the features from scratch would take more than 8 hours unfortunately..

Posted by Michaël_Defferrard  over 2 years ago |  Quote

Yeah you should not use the features.csv from fma_metadata.csv. Those features were computed on the full subset of FMA (on tracks longer than 30s) and obviously don’t contain the test set for this challenge. Features can be computed with the features.py script from the starter kit. It took me ~4h to extract them for the whole training and test tracks. If that’s too much I can share with you the features I generated with that script, but no guarantee that they’re the best features or anything, it’s just a baseline. ;)

Posted by Michaël_Defferrard  over 2 years ago |  Quote

By popular demand, I’ve put the output of the features.py script on the dataset page. Enjoy! Again, no guarantee on how good those features are. ;-)