def main():
client = skein.ApplicationClient.from_current()
task = cluster.get_task()
task_type, task_id = cluster.get_task_description()
event.init_event(client, task, "127.0.0.1:0")
_task_commons._setup_container_logs(client)
if task_type == "evaluator":
evaluator_fn(client)
else:
logger.info(f"{task_type}:{task_id}: nothing to do")
event.stop_event(client, task, None)
def main():
client = skein.ApplicationClient.from_current()
task_type, task_id = cluster.get_task_description()
task = cluster.get_task()
event.init_event(client, task, f"127.0.0.1:0")
_task_commons._setup_container_logs(client)
if task_type in ['chief', 'worker']:
_worker_fn(task_type, task_id, client)
elif task_type == 'evaluator':
_evaluator_fn(client)
else:
logger.error(f'Unknown task type {task_type}')
event.stop_event(client, task, None)
def start_cluster(
host_port: typing.Tuple[str, int], client: skein.ApplicationClient,
all_tasks: typing.List[str]) -> typing.Dict[str, typing.List[str]]:
# There is a race condition between acquiring a TCP port for
# ``tf.train.Server``, and calling ``train_and_evaluate``.
# There is no TensorFlow API to get rid of the race condition
# completely, but the window of opportunity can be reduced by
# preempting the server.
# See https://github.com/tensorflow/tensorflow/issues/21492
cluster_spec: typing.Dict = dict()
host, port = host_port
event.init_event(client, get_task(),
f"{socket.gethostbyname(host)}:{port}")
cluster_spec = aggregate_spec(client, all_tasks)
return cluster_spec
def main():
client = skein.ApplicationClient.from_current()
task_type, task_id = get_task_description()
task = get_task()
event.init_event(client, task, "127.0.0.1:0")
_task_commons._setup_container_logs(client)
net_if = get_net_if()
if task_type == 'chief':
_driver_fn(client, net_if)
if task_type in ['worker', 'chief']:
_worker_fn(client, task, net_if)
elif task_type == 'evaluator':
evaluator_fn(client)
else:
logger.error(f'Unknown task type {task_type}')
event.stop_event(client, task, None)