and keyword arguments when calling the layer.
*   Functional models now get constructed if *any* tensor in a layer call's
    arguments/keyword arguments comes from a keras input. Previously the
    functional api would only work if all of the elements in the first argument
    to the layer came from a keras input.