def start_tensorboard_service(self):
tb_client = TensorBoardClient(self._k8s_client)
tb_client.create_tensorboard_service()
logger.info("Waiting for the URL for TensorBoard service...")
tb_url = tb_client.get_tensorboard_url()
if tb_url:
logger.info("TensorBoard service is available at: %s" % tb_url)
else:
logger.warning("Unable to get the URL for TensorBoard service")
def test_create_tensorboard_service(self):
tb_client = TensorBoardClient(
image_name=None,
namespace="default",
job_name="test-job-%d-%d" %
(int(time.time()), random.randint(1, 101)),
event_callback=None,
)
tb_client._create_tensorboard_service(port=80,
service_type="LoadBalancer")
time.sleep(1)
service = tb_client._get_tensorboard_service()
self.assertTrue("load_balancer" in service["status"])
self.assertEqual(service["spec"]["ports"][0]["port"], 80)
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()
class Master(object):
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 _set_completed_steps_by_checkpoint(self, checkpoint_dir_for_init):
if not checkpoint_dir_for_init:
return
if not CheckpointSaver.check_checkpoint_valid(checkpoint_dir_for_init):
raise ValueError(
"Invalid checkpoint directory {}".format(
checkpoint_dir_for_init
)
)
model_verion = CheckpointSaver.get_version_from_checkpoint(
checkpoint_dir_for_init
)
for callback in self.callbacks_list.callbacks:
if isinstance(callback, MaxStepsStopping):
callback.set_completed_steps(model_verion)
def request_stop(self, err_msg=None):
"""Request master to quit"""
self._should_stop = True
if err_msg:
self.logger.error(err_msg)
# TODO (chengfu.wcy) create meaningful status codes
self._exit_code = -1
def prepare(self):
"""
Start the components one by one. Make sure that it is ready to run.
"""
# Start the evaluation service if requested
if self.evaluation_service:
self.logger.info("Starting evaluation service")
self.evaluation_service.start()
self.logger.info("Evaluation service started")
# Start the master GRPC server
self.logger.info("Starting master RPC server")
self.server.start()
self.logger.info("Master RPC server started")
# Start the worker manager if requested
if self.instance_manager:
self.instance_manager.update_status(InstanceManagerStatus.PENDING)
if self.distribution_strategy == DistributionStrategy.ALLREDUCE:
# Start rendezvous server for workers to initialize Horovod
self.rendezvous_server.start()
else:
self.instance_manager.start_parameter_servers()
self.instance_manager.start_workers()
self.instance_manager.update_status(InstanceManagerStatus.RUNNING)
# Start TensorBoard k8s Service if requested
if self.tb_service and self.tb_client:
self.logger.info("Starting tensorboard service")
self.tb_service.start()
self.tb_client.start_tensorboard_service()
self.logger.info("Tensorboard service started")
def run(self):
"""
The main loop of master.
Dispatch the tasks to the workers until all the tasks are completed.
"""
try:
while True:
if self.instance_manager.all_workers_failed:
raise Exception(
"All workers fail with unrecoverable errors"
)
break
if self.task_d.finished():
if self.instance_manager:
self.instance_manager.update_status(
InstanceManagerStatus.FINISHED
)
break
if self._should_stop:
break
time.sleep(30)
except KeyboardInterrupt:
self.logger.warning("Server stopping")
finally:
self._stop()
return self._exit_code
def _stop(self):
"""
Stop all the components.
Make sure that the created services and components are shut down.
"""
self.logger.info("Stopping master")
if self.evaluation_service:
self.logger.info("Stopping evaluation service")
self.evaluation_service.stop()
self.logger.info("Evaluation service stopped")
self.logger.info("Stopping RPC server")
self.server.stop(None) # grace = None
self.logger.info("RPC server stopped")
# Keep TensorBoard running when all the tasks are finished
if self.tb_service:
self.logger.info(
"All tasks finished. Keeping TensorBoard service running..."
)
while True:
if self.tb_service.is_active():
time.sleep(10)
else:
self.logger.warning(
"Unable to keep TensorBoard running. "
"It has already terminated"
)
break
self.logger.info("Master stopped")
@staticmethod
def _get_job_type(args):
if all(
(
args.training_data,
args.validation_data,
args.evaluation_throttle_secs or args.evaluation_steps,
)
):
job_type = JobType.TRAINING_WITH_EVALUATION
elif all(
(
args.validation_data,
not args.training_data,
not args.prediction_data,
)
):
job_type = JobType.EVALUATION_ONLY
elif all(
(
args.prediction_data,
not args.validation_data,
not args.training_data,
)
):
job_type = JobType.PREDICTION_ONLY
else:
job_type = JobType.TRAINING_ONLY
return job_type
def _create_tensorboard_service(self, tensorboard_log_dir, master_ip):
tb_service = None
if tensorboard_log_dir:
self.logger.info(
"Creating TensorBoard service with log directory %s",
tensorboard_log_dir,
)
# Start TensorBoard CLI
tb_service = TensorboardService(tensorboard_log_dir, master_ip)
return tb_service
def _create_evaluation_service(self, args):
evaluation_service = None
if (
self.job_type == JobType.TRAINING_WITH_EVALUATION
or self.job_type == JobType.EVALUATION_ONLY
):
self.logger.info(
"Creating evaluation service with throttle seconds %d "
" and evaluation steps %d",
args.evaluation_throttle_secs,
args.evaluation_steps,
)
evaluation_service = EvaluationService(
self.tb_service,
self.task_d,
args.evaluation_start_delay_secs,
args.evaluation_throttle_secs,
args.evaluation_steps,
self.job_type == JobType.EVALUATION_ONLY,
self.model_module[args.eval_metrics_fn],
)
self.task_d.set_evaluation_service(evaluation_service)
return evaluation_service
def _create_master_service(self, args):
self.logger.info("Creating 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.minibatch_size,
evaluation_service=self.evaluation_service,
master=self,
)
elasticdl_pb2_grpc.add_MasterServicer_to_server(
master_servicer, server
)
server.add_insecure_port("[::]:{}".format(args.port))
self.logger.info("The port of the master server is: %d", args.port)
return master_servicer, server
def _create_instance_manager(self, args):
instance_manager = None
container_command = ["/bin/bash"]
if args.num_workers:
assert args.worker_image, "Worker image cannot be empty"
worker_client_command = (
BashCommandTemplate.SET_PIPEFAIL
+ " python -m elasticdl.python.worker.main"
)
worker_args = [
"--master_addr",
self.master_addr,
"--job_type",
self.job_type,
]
worker_args.extend(
build_arguments_from_parsed_result(args, filter_args=["envs"])
)
worker_args = wrap_python_args_with_string(worker_args)
worker_args.insert(0, worker_client_command)
if args.use_go_ps:
opt_type, opt_args = get_optimizer_info(self.optimizer)
ps_command = "elasticdl_ps"
ps_command_args = [
"-job_name=" + args.job_name,
"-namespace=" + args.namespace,
"-master_addr=" + self.master_addr,
"-port=2222",
"-use_async=" + ("true" if args.use_async else "false"),
"-grads_to_wait=" + str(args.grads_to_wait),
"-lr_staleness_modulation="
+ ("true" if args.lr_staleness_modulation else "false"),
"-sync_version_tolerance="
+ str(args.sync_version_tolerance),
"-evaluation_steps=" + str(args.evaluation_steps),
"-num_ps_pods=" + str(args.num_ps_pods),
"-num_workers=" + str(args.num_workers),
"-checkpoint_dir=" + str(args.checkpoint_dir),
"-checkpoint_steps=" + str(args.checkpoint_steps),
"-keep_checkpoint_max=" + str(args.keep_checkpoint_max),
"-checkpoint_dir_for_init="
+ str(args.checkpoint_dir_for_init),
"-opt_type=" + opt_type,
"-opt_args=" + opt_args,
]
ps_command_args = wrap_go_args_with_string(ps_command_args)
# Execute source /root/.bashrc to add the file path
# of `elasticdl_ps` into the PATH environment variable.
ps_args = [
"source",
"/root/.bashrc_elasticdl",
"&&",
ps_command,
]
ps_args.extend(ps_command_args)
else:
ps_command = (
BashCommandTemplate.SET_PIPEFAIL
+ " python -m elasticdl.python.ps.main"
)
ps_command_args = [
"--grads_to_wait",
str(args.grads_to_wait),
"--lr_staleness_modulation",
str(args.lr_staleness_modulation),
"--sync_version_tolerance",
str(args.sync_version_tolerance),
"--use_async",
str(args.use_async),
"--model_zoo",
args.model_zoo,
"--model_def",
args.model_def,
"--job_name",
args.job_name,
"--port",
"2222",
"--master_addr",
self.master_addr,
"--namespace",
args.namespace,
"--evaluation_steps",
str(args.evaluation_steps),
"--checkpoint_dir",
str(args.checkpoint_dir),
"--checkpoint_steps",
str(args.checkpoint_steps),
"--keep_checkpoint_max",
str(args.keep_checkpoint_max),
"--num_ps_pods",
str(args.num_ps_pods),
"--checkpoint_dir_for_init",
str(args.checkpoint_dir_for_init),
"--num_workers",
str(args.num_workers),
"--log_level",
str(args.log_level),
"--minibatch_size",
str(args.minibatch_size),
"--num_minibatches_per_task",
str(args.num_minibatches_per_task),
]
ps_args = wrap_python_args_with_string(ps_command_args)
ps_args.insert(0, ps_command)
worker_args = ["-c", " ".join(worker_args)]
ps_args = ["-c", " ".join(ps_args)]
env_dict = parse_envs(args.envs)
env = []
for key in env_dict:
env.append(V1EnvVar(name=key, value=env_dict[key]))
kwargs = get_dict_from_params_str(args.aux_params)
disable_relaunch = kwargs.get("disable_relaunch", False)
cluster_spec = self._get_image_cluster_spec(args.cluster_spec)
instance_manager = InstanceManager(
self.task_d,
rendezvous_server=self.rendezvous_server,
job_name=args.job_name,
image_name=args.worker_image,
worker_command=container_command,
worker_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,
worker_pod_priority=args.worker_pod_priority,
num_ps=args.num_ps_pods,
ps_command=container_command,
ps_args=ps_args,
ps_resource_request=args.ps_resource_request,
ps_resource_limit=args.ps_resource_limit,
ps_pod_priority=args.ps_pod_priority,
volume=args.volume,
image_pull_policy=args.image_pull_policy,
restart_policy=args.restart_policy,
cluster_spec=cluster_spec,
envs=env,
disable_relaunch=disable_relaunch,
log_file_path=args.log_file_path,
)
return instance_manager
def _get_image_cluster_spec(self, cluster_spec):
if cluster_spec:
filename = os.path.basename(cluster_spec)
image_cluster_spec = os.path.join(
ClusterSpecConfig.CLUSTER_SPEC_DIR, filename
)
return image_cluster_spec
return cluster_spec
def _check_timeout_tasks(self):
while True:
doing_tasks = self.task_d._doing.copy()
cur_time = time.time()
avg_time = self.master_servicer.get_average_task_complete_time()
for task_id, (worker_id, task, start_time) in doing_tasks.items():
if task.type == elasticdl_pb2.TRAINING:
start_time = self.master_servicer.get_worker_liveness_time(
worker_id
)
if task.type in [
elasticdl_pb2.TRAINING,
elasticdl_pb2.EVALUATION,
]:
if (cur_time - start_time) > 3 * avg_time[task.type]:
self.logger.info(
"worker %d timeout, relaunch it" % worker_id
)
self.task_d.recover_tasks(worker_id)
# TODO: save worker logs before remove it
self.instance_manager._remove_worker(worker_id)
break
time.sleep(30)
def main():
args = parse_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,
)
task_d = _make_task_dispatcher(
args.training_data_dir,
args.evaluation_data_dir,
args.prediction_data_dir,
args.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",
"--model_zoo",
args.model_zoo,
"--master_addr",
master_addr,
"--log_level",
args.log_level,
"--dataset_fn",
args.dataset_fn,
"--loss",
args.loss,
"--optimizer",
args.optimizer,
"--eval_metrics_fn",
args.eval_metrics_fn,
"--model_def",
args.model_def,
"--job_type",
job_type,
"--minibatch_size",
str(args.minibatch_size),
"--embedding_service_endpoint",
str(embedding_service_endpoint),
"--get_model_steps",
str(args.get_model_steps),
]
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 __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
class Master(object):
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 request_stop(self, err_msg=None):
"""Request master to quit"""
self._should_stop = True
if err_msg:
self.logger.error(err_msg)
# TODO (chengfu.wcy) create meaningful status codes
self._exit_code = -1
def prepare(self):
"""
Start the components one by one. Make sure that it is ready to run.
"""
# Start the evaluation service if requested
if self.evaluation_service:
self.logger.info("Starting evaluation service")
self.evaluation_service.start()
self.logger.info("Evaluation service started")
# Start the master GRPC server
self.logger.info("Starting master RPC server")
self.server.start()
self.logger.info("Master RPC server started")
# Start the worker manager if requested
if self.instance_manager:
self.instance_manager.update_status(InstanceManagerStatus.PENDING)
self.instance_manager.start_parameter_servers()
self.instance_manager.start_workers()
self.instance_manager.update_status(InstanceManagerStatus.RUNNING)
# Start TensorBoard k8s Service if requested
if self.tb_service and self.tb_client:
self.logger.info("Starting tensorboard service")
self.tb_service.start()
self.tb_client.start_tensorboard_service()
self.logger.info("Tensorboard service started")
def run(self):
"""
The main loop of master.
Dispatch the tasks to the workers until all the tasks are completed.
"""
try:
while True:
if self.task_d.finished():
if self.instance_manager:
self.instance_manager.update_status(
InstanceManagerStatus.FINISHED)
break
if self._should_stop:
break
time.sleep(30)
except KeyboardInterrupt:
self.logger.warning("Server stopping")
finally:
self._stop()
return self._exit_code
def _stop(self):
"""
Stop all the components.
Make sure that the created services and components are shut down.
"""
self.logger.info("Stopping master")
if self.evaluation_service:
self.logger.info("Stopping evaluation service")
self.evaluation_service.stop()
self.logger.info("Evaluation service stopped")
self.logger.info("Stopping RPC server")
self.server.stop(None) # grace = None
self.logger.info("RPC server stopped")
# Keep TensorBoard running when all the tasks are finished
if self.tb_service:
self.logger.info(
"All tasks finished. Keeping TensorBoard service running...")
while True:
if self.tb_service.is_active():
time.sleep(10)
else:
self.logger.warning("Unable to keep TensorBoard running. "
"It has already terminated")
break
self.logger.info("Master stopped")
@staticmethod
def _get_job_type(args):
if all((
args.training_data,
args.validation_data,
args.evaluation_throttle_secs or args.evaluation_steps,
)):
job_type = JobType.TRAINING_WITH_EVALUATION
elif all((
args.validation_data,
not args.training_data,
not args.prediction_data,
)):
job_type = JobType.EVALUATION_ONLY
elif all((
args.prediction_data,
not args.validation_data,
not args.training_data,
)):
job_type = JobType.PREDICTION_ONLY
else:
job_type = JobType.TRAINING_ONLY
return job_type
def _create_tensorboard_service(self, tensorboard_log_dir, master_ip):
tb_service = None
if tensorboard_log_dir:
self.logger.info(
"Creating TensorBoard service with log directory %s",
tensorboard_log_dir,
)
# Start TensorBoard CLI
tb_service = TensorboardService(tensorboard_log_dir, master_ip)
return tb_service
def _create_evaluation_service(self, args):
evaluation_service = None
if (self.job_type == JobType.TRAINING_WITH_EVALUATION
or self.job_type == JobType.EVALUATION_ONLY):
self.logger.info(
"Creating evaluation service with throttle seconds %d "
" and evaluation steps %d",
args.evaluation_throttle_secs,
args.evaluation_steps,
)
evaluation_service = EvaluationService(
self.tb_service,
self.task_d,
args.evaluation_start_delay_secs,
args.evaluation_throttle_secs,
args.evaluation_steps,
self.job_type == JobType.EVALUATION_ONLY,
self.model_module[args.eval_metrics_fn],
)
self.task_d.set_evaluation_service(evaluation_service)
return evaluation_service
def _create_master_service(self, args):
self.logger.info("Creating 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.minibatch_size,
self.task_d,
evaluation_service=self.evaluation_service,
)
elasticdl_pb2_grpc.add_MasterServicer_to_server(
master_servicer, server)
server.add_insecure_port("[::]:{}".format(args.port))
self.logger.info("The port of the master server is: %d", args.port)
return master_servicer, server
def _create_instance_manager(self, args):
instance_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",
self.master_addr,
"--job_type",
self.job_type,
]
worker_args.extend(build_arguments_from_parsed_result(args))
ps_command = ["python"]
ps_args = [
"-m",
"elasticdl.python.ps.main",
"--grads_to_wait",
str(args.grads_to_wait),
"--lr_staleness_modulation",
str(args.lr_staleness_modulation),
"--use_async",
str(args.use_async),
"--minibatch_size",
str(args.minibatch_size),
"--model_zoo",
args.model_zoo,
"--model_def",
args.model_def,
"--job_name",
args.job_name,
"--num_minibatches_per_task",
str(args.num_minibatches_per_task),
"--port",
"2222",
"--master_addr",
self.master_addr,
"--namespace",
args.namespace,
"--evaluation_steps",
str(args.evaluation_steps),
"--checkpoint_dir",
str(args.checkpoint_dir),
"--checkpoint_steps",
str(args.checkpoint_steps),
"--keep_checkpoint_max",
str(args.keep_checkpoint_max),
"--num_ps_pods",
str(args.num_ps_pods),
"--checkpoint_dir_for_init",
str(args.checkpoint_dir_for_init),
"--num_workers",
str(args.num_workers),
"--log_level",
str(args.log_level),
]
env_dict = parse_envs(args.envs)
env = []
for key in env_dict:
env.append(V1EnvVar(name=key, value=env_dict[key]))
instance_manager = InstanceManager(
self.task_d,
job_name=args.job_name,
image_name=args.worker_image,
worker_command=worker_command,
worker_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,
worker_pod_priority=args.worker_pod_priority,
num_ps=args.num_ps_pods,
ps_command=ps_command,
ps_args=ps_args,
ps_resource_request=args.ps_resource_request,
ps_resource_limit=args.ps_resource_limit,
ps_pod_priority=args.ps_pod_priority,
volume=args.volume,
image_pull_policy=args.image_pull_policy,
restart_policy=args.restart_policy,
cluster_spec=args.cluster_spec,
envs=env,
)
return instance_manager
