Finding sources for input data is something I struggle with when building deep learning models. Out of curiosity, how did you go about programmatically accessing the music files for all 120M+ songs, in order to create your embedding vector? I can't imagine iTunes has an API which would let a person do that.