Catalog Models with Model Registry

ML 模型本质上就是一个容器. 所以逻辑上你可以用管理容器的方式来管理 ML. Amazon Sagemaker Model Registry 是 Amazon Sagemaker 的子服务, 本质上就是让数据科学家在不懂 Docker 的知识的情况下管理 ML 模型的版本.

Model Registry 有两个概念 Model Group 和 Model Package:

  • Model Group: 就是一个模型的抽象, 你可以理解为 Docker 里的 Repository, 也可以理解为面向对象中的 类.

  • Model Package: 则是一个模型的具体版本, 你可以理解为 Docker 里的 Tag, 也可以理解为面向对象中的 实例. 其中 Model Package 都有对应的版本号 Version, 版本号是从 1 开始的整数. 这一点跟 Lambda Function Version 类似.

Register Version

Reference:

# Specify the model source
model_url = "s3://your-bucket-name/model.tar.gz"

modelpackage_inference_specification =  {
    "InferenceSpecification": {
      "Containers": [
         {
            "Image": '257758044811.dkr.ecr.us-east-2.amazonaws.com/sagemaker-xgboost:1.2-1',
        "ModelDataUrl": model_url
         }
      ],
      "SupportedContentTypes": [ "text/csv" ],
      "SupportedResponseMIMETypes": [ "text/csv" ],
   }
 }

# Alternatively, you can specify the model source like this:
# modelpackage_inference_specification["InferenceSpecification"]["Containers"][0]["ModelDataUrl"]=model_url

create_model_package_input_dict = {
    "ModelPackageGroupName" : model_package_group_name,
    "ModelPackageDescription" : "Model to detect 3 different types of irises (Setosa, Versicolour, and Virginica)",
    "ModelApprovalStatus" : "PendingManualApproval"
}
create_model_package_input_dict.update(modelpackage_inference_specification)

create_model_package_response = sm_client.create_model_package(**create_model_package_input_dict)
model_package_arn = create_model_package_response["ModelPackageArn"]
print('ModelPackage Version ARN : {}'.format(model_package_arn))

Deploy Model

从一个已经注册号的 Model Package 部署一个模型有两种方式:

# 用 sagemaker SDK (高级封装 API)
from sagemaker import ModelPackage
from time import gmtime, strftime

model_package_arn = 'arn:aws:sagemaker:us-east-1:12345678901:model-package/modeltest/1'
model = ModelPackage(
    role=role,
    model_package_arn=model_package_arn,
    sagemaker_session=sagemaker_session,
)
model.deploy(initial_instance_count=1, instance_type='ml.m5.xlarge')

# 用 boto3 (底层 API)
# 1. Create a model object from the model version
model_name = (
    'DEMO-modelregistry-model-'
    + strftime("%Y-%m-%d-%H-%M-%S", gmtime()
)
print("Model name : {}".format(model_name))
container_list = [{'ModelPackageName': model_version_arn}]

create_model_response = sm_client.create_model(
    ModelName = model_name,
    ExecutionRoleArn = role,
    Containers = container_list
)
print("Model arn : {}".format(create_model_response["ModelArn"]))

# 2. Create an endpoint configuration
endpoint_config_name = (
    'DEMO-modelregistry-EndpointConfig-'
    + strftime("%Y-%m-%d-%H-%M-%S", gmtime())
)
print(endpoint_config_name)
create_endpoint_config_response = sm_client.create_endpoint_config(
    EndpointConfigName = endpoint_config_name,
    ProductionVariants=[
        {
            'InstanceType':'ml.m4.xlarge',
            'InitialVariantWeight':1,
            'InitialInstanceCount':1,
            'ModelName':model_name,
            'VariantName':'AllTraffic'
        }
    ]
)

# 3. Create the endpoint
endpoint_name = (
    'DEMO-modelregistry-endpoint-'
    + strftime("%Y-%m-%d-%H-%M-%S", gmtime()
    )
print("EndpointName={}".format(endpoint_name))

create_endpoint_response = sm_client.create_endpoint(
    EndpointName=endpoint_name,
    EndpointConfigName=endpoint_config_name,
)
print(create_endpoint_response['EndpointArn'])