Subclasses should construct a model/graph to be executed by Run. Specific to TPU execution, this may involve a @tpu_function.on_device_training_loop etc. Args: strategy: Intended for use by host-driven TPUStrategy programs, otherwise None.
(self, strategy=None)