def test_multi_data_model_deploy_pretrained_models_local_mode(container_image, sagemaker_session):
timestamp = sagemaker_timestamp()
endpoint_name = "test-multimodel-endpoint-{}".format(timestamp)
model_name = "test-multimodel-{}".format(timestamp)
# Define pretrained model local path
pretrained_model_data_local_path = os.path.join(DATA_DIR, "sparkml_model", "mleap_model.tar.gz")
with timeout(minutes=30):
model_data_prefix = os.path.join(
"s3://", sagemaker_session.default_bucket(), "multimodel-{}/".format(timestamp)
)
multi_data_model = MultiDataModel(
name=model_name,
model_data_prefix=model_data_prefix,
image_uri=container_image,
role=ROLE,
sagemaker_session=sagemaker_session,
)
# Add model before deploy
multi_data_model.add_model(pretrained_model_data_local_path, PRETRAINED_MODEL_PATH_1)
# Deploy model to an endpoint
multi_data_model.deploy(1, "local", endpoint_name=endpoint_name)
# Add models after deploy
multi_data_model.add_model(pretrained_model_data_local_path, PRETRAINED_MODEL_PATH_2)
endpoint_models = []
for model_path in multi_data_model.list_models():
endpoint_models.append(model_path)
assert PRETRAINED_MODEL_PATH_1 in endpoint_models
assert PRETRAINED_MODEL_PATH_2 in endpoint_models
predictor = Predictor(
endpoint_name=endpoint_name,
sagemaker_session=multi_data_model.sagemaker_session,
serializer=NumpySerializer(),
deserializer=string_deserializer,
)
data = numpy.zeros(shape=(1, 1, 28, 28))
result = predictor.predict(data, target_model=PRETRAINED_MODEL_PATH_1)
assert result == "Invoked model: {}".format(PRETRAINED_MODEL_PATH_1)
result = predictor.predict(data, target_model=PRETRAINED_MODEL_PATH_2)
assert result == "Invoked model: {}".format(PRETRAINED_MODEL_PATH_2)
# Cleanup
multi_data_model.sagemaker_session.sagemaker_client.delete_endpoint_config(
EndpointConfigName=endpoint_name
)
multi_data_model.sagemaker_session.delete_endpoint(endpoint_name)
multi_data_model.delete_model()
with pytest.raises(Exception) as exception:
sagemaker_session.sagemaker_client.describe_model(ModelName=multi_data_model.name)
assert "Could not find model" in str(exception.value)
sagemaker_session.sagemaker_client.describe_endpoint_config(name=endpoint_name)
assert "Could not find endpoint" in str(exception.value)
def test_multi_data_model_deploy_pretrained_models_update_endpoint(
container_image, sagemaker_session, cpu_instance_type, alternative_cpu_instance_type
):
timestamp = sagemaker_timestamp()
endpoint_name = "test-multimodel-endpoint-{}".format(timestamp)
model_name = "test-multimodel-{}".format(timestamp)
# Define pretrained model local path
pretrained_model_data_local_path = os.path.join(DATA_DIR, "sparkml_model", "mleap_model.tar.gz")
with timeout_and_delete_endpoint_by_name(endpoint_name, sagemaker_session):
model_data_prefix = os.path.join(
"s3://", sagemaker_session.default_bucket(), "multimodel-{}/".format(timestamp)
)
multi_data_model = MultiDataModel(
name=model_name,
model_data_prefix=model_data_prefix,
image_uri=container_image,
role=ROLE,
sagemaker_session=sagemaker_session,
)
# Add model before deploy
multi_data_model.add_model(pretrained_model_data_local_path, PRETRAINED_MODEL_PATH_1)
# Deploy model to an endpoint
multi_data_model.deploy(1, cpu_instance_type, endpoint_name=endpoint_name)
# Add model after deploy
multi_data_model.add_model(pretrained_model_data_local_path, PRETRAINED_MODEL_PATH_2)
# List model assertions
endpoint_models = []
for model_path in multi_data_model.list_models():
endpoint_models.append(model_path)
assert PRETRAINED_MODEL_PATH_1 in endpoint_models
assert PRETRAINED_MODEL_PATH_2 in endpoint_models
predictor = Predictor(
endpoint_name=endpoint_name,
sagemaker_session=sagemaker_session,
serializer=NumpySerializer(),
deserializer=string_deserializer,
)
data = numpy.zeros(shape=(1, 1, 28, 28))
result = predictor.predict(data, target_model=PRETRAINED_MODEL_PATH_1)
assert result == "Invoked model: {}".format(PRETRAINED_MODEL_PATH_1)
result = predictor.predict(data, target_model=PRETRAINED_MODEL_PATH_2)
assert result == "Invoked model: {}".format(PRETRAINED_MODEL_PATH_2)
endpoint_desc = sagemaker_session.sagemaker_client.describe_endpoint(
EndpointName=endpoint_name
)
old_config_name = endpoint_desc["EndpointConfigName"]
# Update endpoint
predictor.update_endpoint(
initial_instance_count=1, instance_type=alternative_cpu_instance_type
)
endpoint_desc = sagemaker_session.sagemaker_client.describe_endpoint(
EndpointName=endpoint_name
)
new_config_name = endpoint_desc["EndpointConfigName"]
new_config = sagemaker_session.sagemaker_client.describe_endpoint_config(
EndpointConfigName=new_config_name
)
assert old_config_name != new_config_name
assert new_config["ProductionVariants"][0]["InstanceType"] == alternative_cpu_instance_type
assert new_config["ProductionVariants"][0]["InitialInstanceCount"] == 1
# Cleanup
sagemaker_session.sagemaker_client.delete_endpoint_config(
EndpointConfigName=old_config_name
)
sagemaker_session.sagemaker_client.delete_endpoint_config(
EndpointConfigName=new_config_name
)
multi_data_model.delete_model()
with pytest.raises(Exception) as exception:
sagemaker_session.sagemaker_client.describe_model(ModelName=model_name)
assert "Could not find model" in str(exception.value)
sagemaker_session.sagemaker_client.describe_endpoint_config(name=old_config_name)
assert "Could not find endpoint" in str(exception.value)
sagemaker_session.sagemaker_client.describe_endpoint_config(name=new_config_name)
assert "Could not find endpoint" in str(exception.value)
def test_multi_data_model_deploy_train_model_from_amazon_first_party_estimator(
container_image, sagemaker_session, cpu_instance_type
):
timestamp = sagemaker_timestamp()
endpoint_name = "test-multimodel-endpoint-{}".format(timestamp)
model_name = "test-multimodel-{}".format(timestamp)
with timeout_and_delete_endpoint_by_name(endpoint_name, sagemaker_session):
rcf_model_v1 = __rcf_training_job(
sagemaker_session, container_image, cpu_instance_type, 50, 20
)
model_data_prefix = os.path.join(
"s3://", sagemaker_session.default_bucket(), "multimodel-{}/".format(timestamp)
)
multi_data_model = MultiDataModel(
name=model_name,
model_data_prefix=model_data_prefix,
model=rcf_model_v1,
sagemaker_session=sagemaker_session,
)
# Add model before deploy
multi_data_model.add_model(rcf_model_v1.model_data, PRETRAINED_MODEL_PATH_1)
# Deploy model to an endpoint
multi_data_model.deploy(1, cpu_instance_type, endpoint_name=endpoint_name)
# Train another model
rcf_model_v2 = __rcf_training_job(
sagemaker_session, container_image, cpu_instance_type, 70, 20
)
# Deploy newly trained model
multi_data_model.add_model(rcf_model_v2.model_data, PRETRAINED_MODEL_PATH_2)
# List model assertions
endpoint_models = []
for model_path in multi_data_model.list_models():
endpoint_models.append(model_path)
assert PRETRAINED_MODEL_PATH_1 in endpoint_models
assert PRETRAINED_MODEL_PATH_2 in endpoint_models
# Define a predictor to set `serializer` parameter with `NumpySerializer`
# instead of `JSONSerializer` in the default predictor returned by `MXNetPredictor`
# Since we are using a placeholder container image the prediction results are not accurate.
predictor = Predictor(
endpoint_name=endpoint_name,
sagemaker_session=sagemaker_session,
serializer=NumpySerializer(),
deserializer=string_deserializer,
)
data = numpy.random.rand(1, 14)
# Prediction result for the first model
result = predictor.predict(data, target_model=PRETRAINED_MODEL_PATH_1)
assert result == "Invoked model: {}".format(PRETRAINED_MODEL_PATH_1)
# Prediction result for the second model
result = predictor.predict(data, target_model=PRETRAINED_MODEL_PATH_2)
assert result == "Invoked model: {}".format(PRETRAINED_MODEL_PATH_2)
# Cleanup
sagemaker_session.sagemaker_client.delete_endpoint_config(EndpointConfigName=endpoint_name)
multi_data_model.delete_model()
with pytest.raises(Exception) as exception:
sagemaker_session.sagemaker_client.describe_model(ModelName=model_name)
assert "Could not find model" in str(exception.value)
sagemaker_session.sagemaker_client.describe_endpoint_config(name=endpoint_name)
assert "Could not find endpoint" in str(exception.value)
def test_multi_data_model_deploy_trained_model_from_framework_estimator(
container_image, sagemaker_session, cpu_instance_type):
timestamp = sagemaker_timestamp()
endpoint_name = "test-multimodel-endpoint-{}".format(timestamp)
model_name = "test-multimodel-{}".format(timestamp)
mxnet_version = "1.4.1"
with timeout_and_delete_endpoint_by_name(endpoint_name, sagemaker_session):
mxnet_model_1 = __mxnet_training_job(sagemaker_session,
container_image, mxnet_version,
cpu_instance_type, 0.1)
model_data_prefix = os.path.join("s3://",
sagemaker_session.default_bucket(),
"multimodel-{}/".format(timestamp))
multi_data_model = MultiDataModel(
name=model_name,
model_data_prefix=model_data_prefix,
model=mxnet_model_1,
sagemaker_session=sagemaker_session,
)
# Add model before deploy
multi_data_model.add_model(mxnet_model_1.model_data,
PRETRAINED_MODEL_PATH_1)
# Deploy model to an endpoint
multi_data_model.deploy(1,
cpu_instance_type,
endpoint_name=endpoint_name)
# Train another model
mxnet_model_2 = __mxnet_training_job(sagemaker_session,
container_image, mxnet_version,
cpu_instance_type, 0.01)
# Deploy newly trained model
multi_data_model.add_model(mxnet_model_2.model_data,
PRETRAINED_MODEL_PATH_2)
endpoint_models = []
for model_path in multi_data_model.list_models():
endpoint_models.append(model_path)
assert PRETRAINED_MODEL_PATH_1 in endpoint_models
assert PRETRAINED_MODEL_PATH_2 in endpoint_models
# Define a predictor to set `serializer` parameter with npy_serializer
# instead of `json_serializer` in the default predictor returned by `MXNetPredictor`
# Since we are using a placeholder container image the prediction results are not accurate.
predictor = RealTimePredictor(
endpoint=endpoint_name,
sagemaker_session=sagemaker_session,
serializer=npy_serializer,
deserializer=string_deserializer,
)
data = numpy.zeros(shape=(1, 1, 28, 28))
# Prediction result for the first model
result = predictor.predict(data, target_model=PRETRAINED_MODEL_PATH_1)
assert result == "Invoked model: {}".format(PRETRAINED_MODEL_PATH_1)
# Prediction result for the second model
result = predictor.predict(data, target_model=PRETRAINED_MODEL_PATH_2)
assert result == "Invoked model: {}".format(PRETRAINED_MODEL_PATH_2)
# Cleanup
sagemaker_session.sagemaker_client.delete_endpoint_config(
EndpointConfigName=endpoint_name)
multi_data_model.delete_model()
with pytest.raises(Exception) as exception:
sagemaker_session.sagemaker_client.describe_model(ModelName=model_name)
assert "Could not find model" in str(exception.value)
sagemaker_session.sagemaker_client.describe_endpoint_config(
name=endpoint_name)
assert "Could not find endpoint" in str(exception.value)
