def __init__(self, args):
self.logger = get_logger("master", level=args.log_level.upper())
self.num_ps_pods = args.num_ps_pods
self.checkpoint_output_path = args.checkpoint_dir
self.distribution_strategy = args.distribution_strategy
# Master addr
master_ip = os.getenv("MY_POD_IP", "localhost")
self.master_addr = "%s:%d" % (master_ip, args.port)
self.job_type = Master._get_job_type(args)
self.rendezvous_server = None
if self.distribution_strategy == DistributionStrategy.ALLREDUCE:
self.rendezvous_server = HorovodRendezvousServer(master_ip)
# Initialize TensorBoard service if requested
self.tb_service = self._create_tensorboard_service(
args.tensorboard_log_dir, master_ip
)
if self.tb_service:
self.tb_client = TensorBoardClient(
job_name=args.job_name,
image_name=args.worker_image,
namespace=args.namespace,
)
# Initialize the components from the model definition
self.model_module = load_module(
get_module_file_path(args.model_zoo, args.model_def)
).__dict__
self.model_inst = load_model_from_module(
args.model_def, self.model_module, args.model_params
)
self.optimizer = self.model_module[args.optimizer]()
self._create_data_reader_fn = create_data_reader
if args.custom_data_reader in self.model_module:
self._create_data_reader_fn = self.model_module[
args.custom_data_reader
]
# Initialize the callbacks
self.callbacks_list = load_callbacks_from_module(
args.callbacks, self.model_module
)
self.callbacks_list.set_model(self.model_inst)
set_callback_parameters(
self.callbacks_list,
batch_size=args.minibatch_size,
saved_model_path=args.output,
checkpoint_path=args.checkpoint_dir,
)
self._set_completed_steps_by_checkpoint(args.checkpoint_dir_for_init)
# Start task queue
records_per_task = args.minibatch_size * args.num_minibatches_per_task
self.task_d = _make_task_dispatcher(
args.training_data,
args.validation_data,
args.prediction_data,
records_per_task,
args.num_epochs,
args.data_reader_params,
self._create_data_reader_fn,
self.callbacks_list,
)
self.task_d.add_deferred_callback_create_train_end_task()
self.evaluation_service = self._create_evaluation_service(args)
# Initialize instance manager
self.instance_manager = self._create_instance_manager(args)
# Initialize master service
self.master_servicer, self.server = self._create_master_service(args)
self._should_stop = False
self._exit_code = 0
threading.Thread(
target=self._check_timeout_tasks,
name="check_timeout_tasks",
daemon=True,
).start()
def __init__(self, args):
self.logger = get_logger("master", level=args.log_level.upper())
self.num_ps_pods = args.num_ps_pods
self.checkpoint_output_path = args.checkpoint_dir
# Master addr
master_ip = os.getenv("MY_POD_IP", "localhost")
self.master_addr = "%s:%d" % (master_ip, args.port)
self.job_type = Master._get_job_type(args)
# Initialize TensorBoard service if requested
self.tb_service = self._create_tensorboard_service(
args.tensorboard_log_dir, master_ip)
if self.tb_service:
self.tb_client = TensorBoardClient(
job_name=args.job_name,
image_name=args.worker_image,
namespace=args.namespace,
)
# Initialize the components from the model definition
self.model_module = load_module(
get_module_file_path(args.model_zoo, args.model_def)).__dict__
self.model_inst = load_model_from_module(args.model_def,
self.model_module,
args.model_params)
model_handler = ModelHandler.get_model_handler(
args.distribution_strategy, checkpoint_dir=args.checkpoint_dir)
self.model_inst = model_handler.get_model_to_train(self.model_inst)
self.optimizer = self.model_module[args.optimizer]()
self._create_data_reader_fn = create_data_reader
if args.custom_data_reader in self.model_module:
self._create_data_reader_fn = self.model_module[
args.custom_data_reader]
# Start task queue
records_per_task = args.minibatch_size * args.num_minibatches_per_task
self.task_d = _make_task_dispatcher(
args.training_data,
args.validation_data,
args.prediction_data,
records_per_task,
args.num_epochs,
args.data_reader_params,
self._create_data_reader_fn,
)
saved_model_path = args.output
if saved_model_path is not None and self.job_type in [
JobType.TRAINING_ONLY,
JobType.TRAINING_WITH_EVALUATION,
]:
self.task_d.add_deferred_callback_create_save_model_task(
saved_model_path)
self.evaluation_service = self._create_evaluation_service(args)
# Initialize master service
self.master_servicer, self.server = self._create_master_service(args)
# Initialize instance manager
self.instance_manager = self._create_instance_manager(args)
self._should_stop = False
self._exit_code = 0
def main():
args = parse_master_args()
logger = get_logger("master", level=args.log_level.upper())
# Master addr
master_ip = os.getenv("MY_POD_IP", "localhost")
master_addr = "%s:%d" % (master_ip, args.port)
# Start TensorBoard service if requested
if args.tensorboard_log_dir:
logger.info(
"Starting TensorBoard service with log directory %s",
args.tensorboard_log_dir,
)
# Start TensorBoard CLI
tb_service = TensorboardService(args.tensorboard_log_dir, master_ip)
tb_service.start()
else:
tb_service = None
# Start task queue
logger.debug(
"Starting task queue with training data directory %s, "
"evaluation data directory %s, "
"and prediction data directory %s",
args.training_data_dir,
args.evaluation_data_dir,
args.prediction_data_dir,
)
records_per_task = args.minibatch_size * args.num_minibatches_per_task
task_d = _make_task_dispatcher(
args.training_data_dir,
args.evaluation_data_dir,
args.prediction_data_dir,
records_per_task,
args.num_epochs,
)
model_module = load_module(
get_module_file_path(args.model_zoo, args.model_def)
).__dict__
model_inst = load_model_from_module(
args.model_def, model_module, args.model_params
)
optimizer = model_module[args.optimizer]()
if all(
(
args.training_data_dir,
args.evaluation_data_dir,
args.evaluation_throttle_secs or args.evaluation_steps,
)
):
job_type = JobType.TRAINING_WITH_EVALUATION
elif all(
(
args.evaluation_data_dir,
not args.training_data_dir,
not args.prediction_data_dir,
)
):
job_type = JobType.EVALUATION_ONLY
elif all(
(
args.prediction_data_dir,
not args.evaluation_data_dir,
not args.training_data_dir,
)
):
job_type = JobType.PREDICTION_ONLY
else:
job_type = JobType.TRAINING_ONLY
# Initialize checkpoint service
if args.checkpoint_steps or job_type == JobType.TRAINING_WITH_EVALUATION:
logger.info("Starting checkpoint service")
checkpoint_service = CheckpointService(
args.checkpoint_dir,
args.checkpoint_steps,
args.keep_checkpoint_max,
job_type == JobType.TRAINING_WITH_EVALUATION,
)
else:
checkpoint_service = None
# Initialize evaluation service
evaluation_service = None
if (
job_type == JobType.TRAINING_WITH_EVALUATION
or job_type == JobType.EVALUATION_ONLY
):
logger.info(
"Starting evaluation service with throttle seconds %d "
" and evaluation steps %d",
args.evaluation_throttle_secs,
args.evaluation_steps,
)
evaluation_service = EvaluationService(
checkpoint_service,
tb_service,
task_d,
args.evaluation_start_delay_secs,
args.evaluation_throttle_secs,
args.evaluation_steps,
job_type == JobType.EVALUATION_ONLY,
)
evaluation_service.start()
task_d.set_evaluation_service(evaluation_service)
embedding_service_endpoint = None
embedding_dims = {}
# Search for embedding layers in the model,
# if found, initialize embedding service
layers = find_layer(model_inst, Embedding)
if layers:
embedding_service = EmbeddingService()
embedding_service_endpoint = embedding_service.start_embedding_service(
job_name=args.job_name,
image_name=args.worker_image,
namespace=args.namespace,
resource_request=args.master_resource_request,
resource_limit=args.master_resource_limit,
pod_priority=args.worker_pod_priority,
volume=args.volume,
image_pull_policy=args.image_pull_policy,
restart_policy=args.restart_policy,
cluster_spec=args.cluster_spec,
)
logger.info(
"Embedding service start succeeded. The endpoint is %s."
% str(embedding_service_endpoint)
)
embedding_dims = dict(
[(layer.name, layer.output_dim) for layer in layers]
)
# The master service
logger.info("Starting master service")
server = grpc.server(
futures.ThreadPoolExecutor(max_workers=64),
options=[
("grpc.max_send_message_length", GRPC.MAX_SEND_MESSAGE_LENGTH),
(
"grpc.max_receive_message_length",
GRPC.MAX_RECEIVE_MESSAGE_LENGTH,
),
],
)
master_servicer = MasterServicer(
args.grads_to_wait,
args.minibatch_size,
optimizer,
task_d,
init_var=model_inst.trainable_variables if model_inst.built else [],
embedding_dims=embedding_dims,
checkpoint_filename_for_init=args.checkpoint_filename_for_init,
checkpoint_service=checkpoint_service,
evaluation_service=evaluation_service,
embedding_service_endpoint=embedding_service_endpoint,
lr_staleness_modulation=args.lr_staleness_modulation,
use_async=args.use_async,
)
elasticdl_pb2_grpc.add_MasterServicer_to_server(master_servicer, server)
server.add_insecure_port("[::]:{}".format(args.port))
server.start()
logger.info("Server started at port: %d", args.port)
worker_manager = None
if args.num_workers:
assert args.worker_image, "Worker image cannot be empty"
worker_command = ["python"]
worker_args = [
"-m",
"elasticdl.python.worker.main",
"--master_addr",
master_addr,
"--job_type",
job_type,
"--embedding_service_endpoint",
str(embedding_service_endpoint),
]
worker_args.extend(build_arguments_from_parsed_result(args))
env_dict = parse_envs(args.envs)
env = []
for key in env_dict:
env.append(V1EnvVar(name=key, value=env_dict[key]))
worker_manager = WorkerManager(
task_d,
job_name=args.job_name,
image_name=args.worker_image,
command=worker_command,
args=worker_args,
namespace=args.namespace,
num_workers=args.num_workers,
worker_resource_request=args.worker_resource_request,
worker_resource_limit=args.worker_resource_limit,
pod_priority=args.worker_pod_priority,
volume=args.volume,
image_pull_policy=args.image_pull_policy,
restart_policy=args.restart_policy,
cluster_spec=args.cluster_spec,
envs=env,
)
worker_manager.update_status(WorkerManagerStatus.PENDING)
logger.info("Launching %d workers", args.num_workers)
worker_manager.start_workers()
worker_manager.update_status(WorkerManagerStatus.RUNNING)
# Start TensorBoard k8s Service if requested
if tb_service:
TensorBoardClient(
job_name=args.job_name,
image_name=args.worker_image,
namespace=args.namespace,
).start_tensorboard_service()
try:
while True:
if task_d.finished():
if worker_manager:
worker_manager.update_status(WorkerManagerStatus.FINISHED)
if args.output:
master_servicer.save_latest_checkpoint(args.output)
break
time.sleep(30)
except KeyboardInterrupt:
logger.warning("Server stopping")
if evaluation_service:
logger.info("Stopping evaluation service")
evaluation_service.stop()
logger.info("Stopping RPC server")
server.stop(0)
# Keep TensorBoard running when all the tasks are finished
if tb_service:
logger.info(
"All tasks finished. Keeping TensorBoard service running..."
)
while True:
if tb_service.is_active():
time.sleep(10)
else:
logger.warning(
"Unable to keep TensorBoard running. "
"It has already terminated"
)
break
logger.info("Master stopped")
def _create_model_instance(model_def):
module_file = get_module_file_path(_get_model_zoo_path(), model_def)
model_module = load_module(module_file).__dict__
return load_model_from_module(model_def, model_module, None)