Models trained in this project predict the genre of a track from its spectrogram using the FMA dataset and a Transfer Learning framework. This project replicates some of the findings of Kim et al (https://arxiv.org/pdf/1805.02043.pdf), but on a 25k-sample subset of the FMA dataset. We've borrowed initial feature extraction and some additional scripts from a similar project that attempted to replicate another portion of this work (https://github.com/falkaer/artist-group-factors). Our contribution is to experiment with transfer from different layers of the DCNNs to an MLP, to see if the more general features extracted at lower layers have the potential for better performance.
Links to Project Video and Project Paper