def _create_master_and_worker(self,
service_endpoint=None,
embedding_dims={}):
model_inst = custom_model()
master = MasterServicer(
2,
2,
tf.optimizers.SGD(0.1),
None,
init_var=model_inst.trainable_variables,
embedding_service_endpoint=service_endpoint,
embedding_dims=embedding_dims,
checkpoint_filename_for_init=None,
checkpoint_service=None,
evaluation_service=None,
)
worker = Worker(
1,
JobType.TRAINING_ONLY,
2,
_model_zoo_path,
model_def="test_module.custom_model",
channel=None,
)
worker.set_model(model_inst)
worker._stub = InProcessMaster(master)
return master, worker
def _create_master_and_worker(self,
service_endpoint=None,
embedding_dims={}):
model_inst = custom_model()
master = MasterServicer(
2,
2,
tf.optimizers.SGD(0.1),
None,
init_var=model_inst.trainable_variables,
embedding_service_endpoint=service_endpoint,
embedding_dims=embedding_dims,
checkpoint_filename_for_init=None,
checkpoint_service=None,
evaluation_service=None,
)
arguments = [
"--worker_id",
1,
"--job_type",
JobType.TRAINING_ONLY,
"--minibatch_size",
2,
"--model_zoo",
_model_zoo_path,
"--model_def",
"test_module.custom_model",
]
args = parse_worker_args(arguments)
worker = Worker(args)
worker.set_model(model_inst)
worker._stub = InProcessMaster(master)
return master, worker
def distributed_train_and_evaluate(
feature_shape,
model_zoo_path,
model_def,
model_params="",
eval_metrics_fn="eval_metrics_fn",
loss="loss",
training=True,
dataset_name=DatasetName.IMAGE_DEFAULT,
callback_classes=[],
use_async=False,
get_model_steps=1,
ps_channels=None,
pservers=None,
distribution_strategy=DistributionStrategy.PARAMETER_SERVER,
):
"""Runs distributed training and evaluation with a local master. Grpc
calls are mocked by local master call.
Args:
feature_shape: The shape of model input.
model_zoo_path: The directory that contains user-defined model files
or a specific model file.
model_def: The import path to the model definition function/class in
the model zoo, e.g. "cifar10_subclass.CustomModel".
model_params: The dictionary of model parameters in a string that will
be used to instantiate the model, e.g. "param1=1,param2=2".
eval_metrics_fn: The name of the evaluation metrics function defined
in the model file.
loss: The name of the loss function defined in the model file.
training: True for job type `TRAIN_WITH_EVALUATION`, False for
job type `EVALUATION`.
dataset_name: A dataset name from `DatasetName`.
callback_classes: A List of callbacks that will be called at given
stages of the training procedure.
use_async: A bool. True if using asynchronous updates.
get_model_steps: Worker will perform `get_model` from the parameter
server every this many steps.
ps_channels: A channel list to all parameter server pods.
pservers: A list of parameter server pods.
distribution_strategy: The distribution startegy used by workers, e.g.
DistributionStrategy.PARAMETER_SERVER or
DistributionStrategy.AllreduceStrategy.
Returns:
An integer indicating the model version after the distributed training
and evaluation.
"""
job_type = (JobType.TRAINING_WITH_EVALUATION
if training else JobType.EVALUATION_ONLY)
evaluation_steps = 1 if job_type == JobType.TRAINING_WITH_EVALUATION else 0
batch_size = 8 if dataset_name == DatasetName.IMAGENET else 16
pservers = pservers or []
ps_channels = ps_channels or []
model_module = load_module(get_module_file_path(model_zoo_path,
model_def)).__dict__
for channel in ps_channels:
grpc.channel_ready_future(channel).result()
worker_arguments = [
"--worker_id",
"1",
"--job_type",
job_type,
"--minibatch_size",
batch_size,
"--model_zoo",
model_zoo_path,
"--model_def",
model_def,
"--model_params",
model_params,
"--loss",
loss,
"--get_model_steps",
get_model_steps,
"--distribution_strategy",
distribution_strategy,
]
args = parse_worker_args(worker_arguments)
worker = Worker(args, ps_channels=ps_channels)
if dataset_name in [DatasetName.IMAGENET, DatasetName.FRAPPE]:
record_num = batch_size
else:
record_num = 128
shards = {
create_recordio_file(record_num, dataset_name, feature_shape): (
0,
record_num,
)
}
if training:
training_shards = shards
evaluation_shards = shards
else:
training_shards = {}
evaluation_shards = shards
task_d = _TaskDispatcher(
training_shards,
evaluation_shards,
{},
records_per_task=64,
num_epochs=1,
)
if training:
evaluation_service = EvaluationService(
None,
task_d,
0,
0,
evaluation_steps,
False,
model_module[eval_metrics_fn],
)
else:
evaluation_service = EvaluationService(
None,
task_d,
0,
0,
evaluation_steps,
True,
model_module[eval_metrics_fn],
)
task_d.set_evaluation_service(evaluation_service)
master = MasterServicer(
batch_size,
task_d,
evaluation_service=evaluation_service,
)
callbacks = [
callback_class(master, worker) for callback_class in callback_classes
]
in_process_master = InProcessMaster(master, callbacks)
worker._stub = in_process_master
for pservicer in pservers:
pservicer._master_stub = in_process_master
worker.run()
req = elasticdl_pb2.GetTaskRequest()
req.worker_id = 1
task = master.get_task(req, None)
# No more task.
if task.shard_name:
raise RuntimeError(
"There are some tasks unfinished after worker exits.")
return master._version
def distributed_train_and_evaluate(
feature_shape,
model_zoo_path,
model_def,
model_params="",
eval_metrics_fn="eval_metrics_fn",
training=True,
dataset_name=DatasetName.IMAGE_DEFAULT,
callback_classes=[],
use_async=False,
get_model_steps=1,
):
"""Runs distributed training and evaluation with a local master. Grpc
calls are mocked by local master call.
Args:
feature_shape: The shape of model input.
model_zoo_path: The directory that contains user-defined model files
or a specific model file.
model_def: The import path to the model definition function/class in
the model zoo, e.g. "cifar10_subclass.CustomModel".
model_params: The dictionary of model parameters in a string that will
be used to instantiate the model, e.g. "param1=1,param2=2".
training: True for job type `TRAIN_WITH_EVALUATION`, False for
job type `EVALUATION`.
dataset_name: A dataset name from `DatasetName`.
callback_classes: A List of callbacks that will be called at given
stages of the training procedure.
use_async: A python bool. True if using asynchronous updates.
get_model_steps: Worker will perform `get_model` from the parameter
server every this many steps.
Returns:
An integer indicating the model version after the distributed training
and evaluation.
"""
job_type = (JobType.TRAINING_WITH_EVALUATION
if training else JobType.EVALUATION_ONLY)
batch_size = 8 if dataset_name == DatasetName.IMAGENET else 16
arguments = [
"--worker_id",
"1",
"--job_type",
job_type,
"--minibatch_size",
batch_size,
"--model_zoo",
model_zoo_path,
"--model_def",
model_def,
"--model_params",
model_params,
"--get_model_steps",
get_model_steps,
]
args = parse_worker_args(arguments)
worker = Worker(args)
if dataset_name in [DatasetName.IMAGENET, DatasetName.FRAPPE]:
record_num = batch_size
else:
record_num = 128
shards = {
create_recordio_file(record_num, dataset_name, feature_shape): (
0,
record_num,
)
}
if training:
training_shards = shards
evaluation_shards = shards
else:
training_shards = {}
evaluation_shards = shards
task_d = _TaskDispatcher(
training_shards,
evaluation_shards,
{},
records_per_task=64,
num_epochs=1,
)
model_module = load_module(get_module_file_path(model_zoo_path,
model_def)).__dict__
checkpoint_service = CheckpointService("", 0, 0, True)
if training:
evaluation_service = EvaluationService(
checkpoint_service,
None,
task_d,
0,
0,
1,
False,
model_module[eval_metrics_fn],
)
else:
evaluation_service = EvaluationService(
checkpoint_service,
None,
task_d,
0,
0,
0,
True,
model_module[eval_metrics_fn],
)
task_d.set_evaluation_service(evaluation_service)
grads_to_wait = 1 if use_async else 2
master = MasterServicer(
grads_to_wait,
batch_size,
worker._opt_fn(),
task_d,
init_var=[],
checkpoint_filename_for_init="",
checkpoint_service=checkpoint_service,
evaluation_service=evaluation_service,
use_async=use_async,
)
callbacks = [
callback_class(master, worker) for callback_class in callback_classes
]
worker._stub = InProcessMaster(master, callbacks)
for var in worker._model.trainable_variables:
master.set_model_var(var.name, var.numpy())
worker.run()
req = elasticdl_pb2.GetTaskRequest()
req.worker_id = 1
task = master.GetTask(req, None)
# No more task.
if task.shard_name:
raise RuntimeError(
"There are some tasks unfinished after worker exits.")
return master._version
def test_train_acceleration_with_embedding(self):
kv_store = MockKvStore()
model_inst = CustomModel()
master = MasterServicer(
2,
2,
tf.optimizers.SGD(0.1),
None,
init_var=model_inst.trainable_variables,
checkpoint_filename_for_init=None,
checkpoint_service=None,
evaluation_service=None,
)
arguments = [
"--worker_id",
1,
"--job_type",
JobType.TRAINING_ONLY,
"--minibatch_size",
32,
"--model_zoo",
_model_zoo_path,
"--model_def",
"embedding_test_module.EdlEmbeddingModel",
]
args = parse_worker_args(arguments)
worker = Worker(args)
worker._stub = InProcessMaster(master)
inputs_list = [
{
"f1": tf.constant([[0], [1], [2]], tf.int64),
"f2": tf.constant([[2], [1], [0]], tf.int64),
},
{
"f1": tf.constant([[3], [4], [3]], tf.int64),
"f2": tf.constant([[2], [1], [0]], tf.int64),
},
]
labels_list = [[0, 1, 0], [1, 1, 0]]
input_dim = 5
embedding_dim = 16
worker.set_model(model_inst)
# initialize kv store
for layer in model_inst.layers:
if isinstance(layer, Embedding):
name = layer.name
keys = [Embedding.get_key([name, i]) for i in range(input_dim)]
values = [
np.random.rand(embedding_dim).astype(np.float32)
for i in range(input_dim)
]
kv_store.update(keys, values)
with mock.patch.object(
EmbeddingService, "lookup_embedding", kv_store.lookup
), mock.patch.object(
EmbeddingService, "update_embedding", kv_store.update
):
worker._init_embedding_layer()
worker._run_model_call_before_training(inputs_list[0])
# run training process without tf.function
correct_grads = []
correct_ids_list = []
for features, labels in zip(inputs_list, labels_list):
loss, grads = worker.training_process_eagerly(features, labels)
correct_grads.append(grads)
ids = {}
for layer in worker._embedding_layers:
ids[layer.name] = layer.embedding_and_ids[0].batch_ids
correct_ids_list.append(ids)
worker._reset_embedding()
# run training process with tf.function
test_grads = []
test_ids_list = []
for features, labels in zip(inputs_list, labels_list):
self.assertFalse(worker._train_eagerly)
loss, grads = worker.training_process(features, labels)
test_grads.append(grads)
ids = {}
for layer in worker._embedding_layers:
ids[layer.name] = copy.deepcopy(
layer.embedding_and_ids[0].batch_ids
)
test_ids_list.append(ids)
worker._reset_embedding()
# compare the gradients
for test_g, correct_g in zip(test_grads, correct_grads):
for g1, g2 in zip(test_g, correct_g):
if isinstance(g1, tf.IndexedSlices):
self.assertTrue(np.isclose(g1.values, g2.values).all())
self.assertTrue(np.isclose(g1.indices, g2.indices).all())
else:
self.assertTrue(np.isclose(g1, g2).all())
for test_ids, correct_ids in zip(correct_ids_list, test_ids_list):
for layer_name in correct_ids.keys():
self.assertTrue(
tf.equal(test_ids[layer_name], correct_ids[layer_name])
.numpy()
.all()
)
def distributed_train_and_evaluate(
self,
feature_shape,
model_def,
model_params="",
training=True,
dataset="",
):
"""
Run distributed training and evaluation with a local master.
grpc calls are mocked by local master call.
"""
job_type = (JobType.TRAINING_ONLY
if training else JobType.EVALUATION_ONLY)
batch_size = 16
worker = Worker(
1,
job_type,
batch_size,
_model_zoo_path,
model_def=model_def,
model_params=model_params,
channel=None,
)
if dataset == "imagenet":
batch_size = 8
shards = {create_imagenet_recordio_file(8, feature_shape): (0, 8)}
elif dataset == "frappe":
shards = {
create_frappe_recordio_file(16, feature_shape, 5383): (0, 16)
}
else:
shards = {create_recordio_file(128, feature_shape): (0, 128)}
if training:
training_shards = shards
evaluation_shards = shards
else:
training_shards = {}
evaluation_shards = shards
task_d = _TaskDispatcher(
training_shards,
evaluation_shards,
{},
records_per_task=64,
num_epochs=1,
)
# Initialize checkpoint service
checkpoint_service = CheckpointService("", 0, 0, True)
if training:
evaluation_service = EvaluationService(checkpoint_service, None,
task_d, 0, 0, 1, False)
else:
evaluation_service = EvaluationService(checkpoint_service, None,
task_d, 0, 0, 0, True)
task_d.set_evaluation_service(evaluation_service)
# The master service
master = MasterServicer(
2,
batch_size,
worker._opt_fn(),
task_d,
init_var=[],
checkpoint_filename_for_init="",
checkpoint_service=checkpoint_service,
evaluation_service=evaluation_service,
)
worker._stub = InProcessMaster(master)
for var in worker._model.trainable_variables:
master.set_model_var(var.name, var.numpy())
worker.run()
req = elasticdl_pb2.GetTaskRequest()
req.worker_id = 1
task = master.GetTask(req, None)
# No more task.
self.assertTrue(not task.shard_name)
def testMaxCheckpointVersions(self):
with tempfile.TemporaryDirectory() as tempdir:
chkp_dir = os.path.join(tempdir, "testMaxCheckpointVersions")
os.makedirs(chkp_dir)
# Save checkpoints every 2 steps, and keep 5 checkpoints at most
checkpointer = CheckpointService(chkp_dir, 2, 5, False)
self.assertTrue(checkpointer.is_enabled())
batch_size = 2
# Launch the training
arguments = [
"--worker_id",
1,
"--job_type",
JobType.TRAINING_ONLY,
"--minibatch_size",
batch_size,
"--model_zoo",
_model_zoo_path,
"--model_def",
"test_module.custom_model",
]
args = parse_worker_args(arguments)
worker = Worker(args)
filename = create_recordio_file(128, DatasetName.TEST_MODULE, 1)
task_d = _TaskDispatcher({filename: (0, 128)}, {}, {},
records_per_task=64,
num_epochs=1)
master = MasterServicer(
2,
batch_size,
worker._opt_fn(),
task_d,
init_var=worker._model.trainable_variables,
checkpoint_filename_for_init="",
checkpoint_service=checkpointer,
evaluation_service=None,
)
worker._stub = InProcessMaster(master)
worker.run()
# We should have 5 checkpoints when the training finishes
checkpoint_files = sorted(os.listdir(checkpointer._directory))
self.assertEqual(
checkpoint_files,
[
"model_v24.chkpt",
"model_v26.chkpt",
"model_v28.chkpt",
"model_v30.chkpt",
"model_v32.chkpt",
],
)
# Latest version should be 32
self.assertEqual(32, checkpointer.get_latest_checkpoint_version())
# Check all checkpoints
for version in [24, 26, 28, 30, 32]:
model = checkpointer.get_checkpoint_model(version)
self.assertEqual(version, model.version)
# Checkpoint not found
self.assertRaisesRegex(
RuntimeError,
"Failed to read model checkpoint from file",
checkpointer.get_checkpoint_model,
100,
)
def distributed_train_and_evaluate(
self,
training=True,
callback_classes=[],
use_async=False,
grads_to_wait=2,
get_model_steps=1,
):
"""
Run distributed training and evaluation with a local master.
grpc calls are mocked by local master call.
"""
if use_async and grads_to_wait > 1:
raise ValueError(
"grads_to_wait should be 1 when using asynchronous SGD."
)
job_type = (
JobType.TRAINING_ONLY if training else JobType.EVALUATION_ONLY
)
batch_size = 16
worker = Worker(
1,
job_type,
batch_size,
_model_zoo_path,
model_def="test_module.custom_model",
channel=None,
get_model_steps=get_model_steps,
)
shards = {create_recordio_file(128): (0, 128)}
if training:
training_shards = shards
evaluation_shards = {}
else:
training_shards = {}
evaluation_shards = shards
task_d = _TaskDispatcher(
training_shards,
evaluation_shards,
{},
records_per_task=64,
num_epochs=1,
)
if not training:
evaluation_service = EvaluationService(
None, None, task_d, 0, 0, 0, True
)
task_d.set_evaluation_service(evaluation_service)
else:
evaluation_service = None
master = MasterServicer(
grads_to_wait,
batch_size,
worker._opt_fn(),
task_d,
init_var=[],
checkpoint_filename_for_init="",
checkpoint_service=None,
evaluation_service=evaluation_service,
use_async=use_async,
)
callbacks = [
callback_class(master, worker, self)
for callback_class in callback_classes
]
worker._stub = InProcessMaster(master, callbacks)
for var in worker._model.trainable_variables:
master.set_model_var(var.name, var.numpy())
worker.run()
req = elasticdl_pb2.GetTaskRequest()
req.worker_id = 1
task = master.GetTask(req, None)
# No more task.
self.assertTrue(not task.shard_name)