def main():
image_name = "sagemaker-sklearn-rf-regressor-local"
# Prepare data for model inference - we use the Boston housing dataset
print('Preparing data for model inference')
data = fetch_california_housing()
X_train, X_test, y_train, y_test = train_test_split(data.data,
data.target,
test_size=0.25,
random_state=42)
# we don't train a model, so we will need only the testing data
testX = pd.DataFrame(X_test, columns=data.feature_names)
# Download a pre-trained model file
print('Downloading a pre-trained model file')
s3.download_file(
'aws-ml-blog',
'artifacts/scikit_learn_bring_your_own_model/model.joblib',
'model.joblib')
# Creating a model.tar.gz file
tar = tarfile.open('model.tar.gz', 'w:gz')
tar.add('model.joblib')
tar.close()
model = Model(image_uri=image_name,
role=DUMMY_IAM_ROLE,
model_data='file://./model.tar.gz')
print('Deploying endpoint in local mode')
endpoint = model.deploy(initial_instance_count=1,
instance_type='local',
endpoint_name="my-local-endpoint")
predictor = Predictor(endpoint_name="my-local-endpoint",
sagemaker_session=sagemaker_session,
serializer=CSVSerializer(),
deserializer=CSVDeserializer())
predictions = predictor.predict(testX[data.feature_names].head(5).to_csv(
header=False, index=False))
print(f"Predictions: {predictions}")
print('About to delete the endpoint to stop paying (if in cloud mode).')
predictor.delete_endpoint(predictor.endpoint_name)
def _deploy_inference_pipeline(
self,
inference_containers,
initial_instance_count,
instance_type,
name=None,
sagemaker_session=None,
endpoint_name=None,
tags=None,
wait=True,
update_endpoint=False,
vpc_config=None,
enable_network_isolation=False,
model_kms_key=None,
):
"""Deploy a SageMaker Inference Pipeline.
Args:
inference_containers (list): a list of inference container definitions
initial_instance_count (int): The initial number of instances to run
in the ``Endpoint`` created from this ``Model``.
instance_type (str): The EC2 instance type to deploy this Model to.
For example, 'ml.p2.xlarge'.
name (str): The pipeline model name. If None, a default model name will
be selected on each ``deploy``.
sagemaker_session (sagemaker.session.Session): A SageMaker Session
object, used for SageMaker interactions (default: None). If not
specified, one is created using the default AWS configuration
chain.
endpoint_name (str): The name of the endpoint to create (default:
None). If not specified, a unique endpoint name will be created.
tags (List[dict[str, str]]): The list of tags to attach to this
specific endpoint.
wait (bool): Whether the call should wait until the deployment of
model completes (default: True).
update_endpoint (bool): Flag to update the model in an existing
Amazon SageMaker endpoint. If True, this will deploy a new
EndpointConfig to an already existing endpoint and delete
resources corresponding to the previous EndpointConfig. If
False, a new endpoint will be created. Default: False
vpc_config (dict): information about vpc configuration, optionally
contains "SecurityGroupIds", "Subnets"
model_kms_key (str): KMS key ARN used to encrypt the repacked
model archive file if the model is repacked
"""
# construct Model objects
models = []
for container in inference_containers:
image = container["Image"]
model_data = container["ModelDataUrl"]
env = container["Environment"]
model = Model(
image=image,
model_data=model_data,
role=self.role,
env=env,
vpc_config=vpc_config,
sagemaker_session=sagemaker_session or self.sagemaker_session,
enable_network_isolation=enable_network_isolation,
model_kms_key=model_kms_key,
)
models.append(model)
pipeline = PipelineModel(
models=models,
role=self.role,
name=name,
vpc_config=vpc_config,
sagemaker_session=sagemaker_session or self.sagemaker_session,
)
return pipeline.deploy(
initial_instance_count=initial_instance_count,
instance_type=instance_type,
endpoint_name=endpoint_name,
tags=tags,
wait=wait,
update_endpoint=update_endpoint,
)
def create_model(
self,
name,
sagemaker_session=None,
candidate=None,
vpc_config=None,
enable_network_isolation=False,
model_kms_key=None,
predictor_cls=None,
inference_response_keys=None,
):
"""Creates a model from a given candidate or the best candidate
from the automl job
Args:
name (str): The pipeline model name.
sagemaker_session (sagemaker.session.Session): A SageMaker Session
object, used for SageMaker interactions (default: None). If not
specified, the one originally associated with the ``AutoML`` instance is used.:
candidate (CandidateEstimator or dict): a CandidateEstimator used for deploying
to a SageMaker Inference Pipeline. If None, the best candidate will
be used. If the candidate input is a dict, a CandidateEstimator will be
created from it.
vpc_config (dict): Specifies a VPC that your training jobs and hosted models have
access to. Contents include "SecurityGroupIds" and "Subnets".
enable_network_isolation (bool): Isolates the training container. No inbound or
outbound network calls can be made, except for calls between peers within a
training cluster for distributed training. Default: False
model_kms_key (str): KMS key ARN used to encrypt the repacked
model archive file if the model is repacked
predictor_cls (callable[string, sagemaker.session.Session]): A
function to call to create a predictor (default: None). If
specified, ``deploy()`` returns the result of invoking this
function on the created endpoint name.
inference_response_keys (list): List of keys for response content. The order of the
keys will dictate the content order in the response.
Returns:
PipelineModel object
"""
sagemaker_session = sagemaker_session or self.sagemaker_session
if candidate is None:
candidate_dict = self.best_candidate()
candidate = CandidateEstimator(candidate_dict, sagemaker_session=sagemaker_session)
elif isinstance(candidate, dict):
candidate = CandidateEstimator(candidate, sagemaker_session=sagemaker_session)
inference_containers = candidate.containers
self.validate_and_update_inference_response(inference_containers, inference_response_keys)
# construct Model objects
models = []
for container in inference_containers:
image_uri = container["Image"]
model_data = container["ModelDataUrl"]
env = container["Environment"]
model = Model(
image_uri=image_uri,
model_data=model_data,
role=self.role,
env=env,
vpc_config=vpc_config,
sagemaker_session=sagemaker_session or self.sagemaker_session,
enable_network_isolation=enable_network_isolation,
model_kms_key=model_kms_key,
)
models.append(model)
pipeline = PipelineModel(
models=models,
role=self.role,
predictor_cls=predictor_cls,
name=name,
vpc_config=vpc_config,
sagemaker_session=sagemaker_session or self.sagemaker_session,
)
return pipeline
def test_conditional_pytorch_training_model_registration(
sagemaker_session,
role,
cpu_instance_type,
pipeline_name,
region_name,
):
base_dir = os.path.join(DATA_DIR, "pytorch_mnist")
entry_point = os.path.join(base_dir, "mnist.py")
input_path = sagemaker_session.upload_data(
path=os.path.join(base_dir, "training"),
key_prefix="integ-test-data/pytorch_mnist/training",
)
inputs = TrainingInput(s3_data=input_path)
instance_count = ParameterInteger(name="InstanceCount", default_value=1)
instance_type = ParameterString(name="InstanceType",
default_value="ml.m5.xlarge")
good_enough_input = ParameterInteger(name="GoodEnoughInput",
default_value=1)
in_condition_input = ParameterString(name="Foo", default_value="Foo")
pytorch_estimator = PyTorch(
entry_point=entry_point,
role=role,
framework_version="1.5.0",
py_version="py3",
instance_count=instance_count,
instance_type=instance_type,
sagemaker_session=sagemaker_session,
)
step_train = TrainingStep(
name="pytorch-train",
estimator=pytorch_estimator,
inputs=inputs,
)
step_register = RegisterModel(
name="pytorch-register-model",
estimator=pytorch_estimator,
model_data=step_train.properties.ModelArtifacts.S3ModelArtifacts,
content_types=["*"],
response_types=["*"],
inference_instances=["*"],
transform_instances=["*"],
description="test-description",
)
model = Model(
image_uri=pytorch_estimator.training_image_uri(),
model_data=step_train.properties.ModelArtifacts.S3ModelArtifacts,
sagemaker_session=sagemaker_session,
role=role,
)
model_inputs = CreateModelInput(
instance_type="ml.m5.large",
accelerator_type="ml.eia1.medium",
)
step_model = CreateModelStep(
name="pytorch-model",
model=model,
inputs=model_inputs,
)
step_cond = ConditionStep(
name="cond-good-enough",
conditions=[
ConditionGreaterThanOrEqualTo(left=good_enough_input, right=1),
ConditionIn(value=in_condition_input, in_values=["foo", "bar"]),
],
if_steps=[step_train, step_register],
else_steps=[step_model],
)
pipeline = Pipeline(
name=pipeline_name,
parameters=[
in_condition_input,
good_enough_input,
instance_count,
instance_type,
],
steps=[step_cond],
sagemaker_session=sagemaker_session,
)
try:
response = pipeline.create(role)
create_arn = response["PipelineArn"]
assert re.match(
rf"arn:aws:sagemaker:{region_name}:\d{{12}}:pipeline/{pipeline_name}",
create_arn,
)
execution = pipeline.start(parameters={})
assert re.match(
rf"arn:aws:sagemaker:{region_name}:\d{{12}}:pipeline/{pipeline_name}/execution/",
execution.arn,
)
execution = pipeline.start(parameters={"GoodEnoughInput": 0})
assert re.match(
rf"arn:aws:sagemaker:{region_name}:\d{{12}}:pipeline/{pipeline_name}/execution/",
execution.arn,
)
finally:
try:
pipeline.delete()
except Exception:
pass
def test_tuning_single_algo(
sagemaker_session,
role,
cpu_instance_type,
pipeline_name,
region_name,
):
base_dir = os.path.join(DATA_DIR, "pytorch_mnist")
entry_point = os.path.join(base_dir, "mnist.py")
input_path = sagemaker_session.upload_data(
path=os.path.join(base_dir, "training"),
key_prefix="integ-test-data/pytorch_mnist/training",
)
inputs = TrainingInput(s3_data=input_path)
instance_count = ParameterInteger(name="InstanceCount", default_value=1)
instance_type = ParameterString(name="InstanceType",
default_value="ml.m5.xlarge")
pytorch_estimator = PyTorch(
entry_point=entry_point,
role=role,
framework_version="1.5.0",
py_version="py3",
instance_count=instance_count,
instance_type=instance_type,
sagemaker_session=sagemaker_session,
enable_sagemaker_metrics=True,
max_retry_attempts=3,
)
min_batch_size = ParameterInteger(name="MinBatchSize", default_value=64)
max_batch_size = ParameterInteger(name="MaxBatchSize", default_value=128)
hyperparameter_ranges = {
"batch-size": IntegerParameter(min_batch_size, max_batch_size),
}
tuner = HyperparameterTuner(
estimator=pytorch_estimator,
objective_metric_name="test:acc",
objective_type="Maximize",
hyperparameter_ranges=hyperparameter_ranges,
metric_definitions=[{
"Name": "test:acc",
"Regex": "Overall test accuracy: (.*?);"
}],
max_jobs=2,
max_parallel_jobs=2,
)
step_tune = TuningStep(
name="my-tuning-step",
tuner=tuner,
inputs=inputs,
)
best_model = Model(
image_uri=pytorch_estimator.training_image_uri(),
model_data=step_tune.get_top_model_s3_uri(
top_k=0,
s3_bucket=sagemaker_session.default_bucket(),
),
sagemaker_session=sagemaker_session,
role=role,
)
model_inputs = CreateModelInput(
instance_type="ml.m5.large",
accelerator_type="ml.eia1.medium",
)
step_best_model = CreateModelStep(
name="1st-model",
model=best_model,
inputs=model_inputs,
)
second_best_model = Model(
image_uri=pytorch_estimator.training_image_uri(),
model_data=step_tune.get_top_model_s3_uri(
top_k=1,
s3_bucket=sagemaker_session.default_bucket(),
),
sagemaker_session=sagemaker_session,
role=role,
entry_point=entry_point,
source_dir=base_dir,
)
step_second_best_model = CreateModelStep(
name="2nd-best-model",
model=second_best_model,
inputs=model_inputs,
)
pipeline = Pipeline(
name=pipeline_name,
parameters=[
instance_count, instance_type, min_batch_size, max_batch_size
],
steps=[step_tune, step_best_model, step_second_best_model],
sagemaker_session=sagemaker_session,
)
try:
response = pipeline.create(role)
create_arn = response["PipelineArn"]
assert re.match(
rf"arn:aws:sagemaker:{region_name}:\d{{12}}:pipeline/{pipeline_name}",
create_arn,
)
for _ in retries(
max_retry_count=5,
exception_message_prefix=
"Waiting for a successful execution of pipeline",
seconds_to_sleep=10,
):
execution = pipeline.start(parameters={})
assert re.match(
rf"arn:aws:sagemaker:{region_name}:\d{{12}}:pipeline/{pipeline_name}/execution/",
execution.arn,
)
try:
execution.wait(delay=30, max_attempts=60)
except WaiterError:
pass
execution_steps = execution.list_steps()
assert len(execution_steps) == 3
for step in execution_steps:
assert step["StepStatus"] == "Succeeded"
break
finally:
try:
pipeline.delete()
except Exception:
pass