第二种策略，代码会报错
RuntimeError Traceback (most recent call last)
<ipython-input-4-df7add50a6f4> in <module>()
15 'task': {'type': 'worker', 'index': 0}
16 })
---> 17 strategy = tf.distribute.experimental.MultiWorkerMirroredStrategy()
18 batch_size = batch_size_per_replica * num_workers
19
6 frames
/usr/local/lib/python3.6/dist-packages/tensorflow/python/eager/context.py in configure_collective_ops(self, collective_leader, scoped_allocator_enabled_ops, use_nccl_communication, device_filters)
727
728 if self._context_handle is not None:
--> 729 raise RuntimeError("Collective ops must be configured at program startup")
730
731 self._collective_leader = collective_leader
RuntimeError: Collective ops must be configured at program startup