The idea being that pytorch can just be a high-level API executing lower-level tensorflow under the hood.