Creating a custom loss function that depends on previous embedding layer

Question

I am trying to implement a simple version of the YouTube recommendation model described here. However, instead of movie recommendations, I\'m trying to perform item recommen