Ideally: we do this by parsing native boilerplate python implementation of the model, rather than writing it with tensorflow's clunky api