pipelines/components/contrib/azure/azureml/aml-deploy-model

Azure ML Deploy Model

This component deploys trained model with Azure Machine Learning Service. In Azure Machine Learning Service, trained machine learning models could be deployed as web services in the cloud or locally. You can use your model in web services for inferencing with improved quality. See more on MLOps: Model management, deployment, and monitoring with Azure Machine Learning

This Azure ML Deploy Model component, is a self-contained set of code to perform model deployment operation using Azure Machine Learning service.

It performs the following workflow:

• Prepare an inference configuration.
• Prepare an entry script.
• Choose a compute target.
• Deploy the model to the compute target.

Inputs

Name	Type	Required	Description
deployment_name	string	Y	Name for the deployed model endpoint
model_name	string	Y	Name for registered model and version
inference_config	string	Y	File Path for inferenceconfig.json
deployment_config	string	Y	File Path for deploymentconfig.json
tenant_id	string	Y	Azure AD tenant Id
service_principal_id	string	Y	Azure service principal (client) Id
service_principal_password	string	Y	Azure service principal password
subscription_id	string	Y	Azure subscription Id
resource_group	string	Y	Azure resource group
workspace	string	Y	Azure ML workspace name

Outputs

Output output_config holds description of the deployed web service for the machine learning model. Output score_uri holds endpoint for the deployed model.

Prerequisites

Ensure Kubeflow has been installed on AKS following instructions for deploying Kubeflow on Azure.

Azure Machine Learning service will be used to deploy the model, store the metadata and manage the active deployment. To create a Machine Learning Service workspace: 1. Go to the Azure portal and click on your resource group. 2. Select the add a new resource option. 3. Search for Machine Learning Studio Workspace and use the default options, taking note of the name you decide for it.

You need to create a container registry to store those images in the cloud so that Kubeflow can pull the images as they are needed. See here for step-by-step guidance to create an Azure Container Registry. Please ensure enable admin user, and change the SKU option to Premium.

In order for the AKS cluster to have access to pulling images created for execution of the pipeline, you will need to update your cluster so that it is able to pull the images from the container registry we just created by running:

az aks update -n {myAKSCluster} -g {MyResourceGroup} --attach-acr {REGISTRY_NAME}

A service principal is used to allow your pipeline to securely interface with your Azure services without having to directly login in the pipeline and use admin privileges. See here for a detailed guide to create a service principal. Also please ensure to add a Contributor role to the service principal to be able access Azure resources.

Usage

In ./src folder, adjust pre-configured inferenceconfig.json, deploymentconfig.json, and environment.yml to specify the resources for deployment like the compute type, number of cores etc. Additionaly, deploymentconfig_aks.json could be specified from parameters to use Azure Kubernetes Service to deploy the web service. score.py is the entry script for inference, where init() and run(data) are the functions to be edited to use the model. See here for a detailed instruction on constructing the entry script. See Deploy your existing model with Azure for an example of score.py. If additional pip packages are used for score.py, please ensure them being added to environment.yml. In the end, build and push the image to the Container Registry, the pipeline operation will directly pull the image from ACR.

A sample script to build and push the image:

docker build . -t {your_ACR_name}.azurecr.io/deploy/{your_image_name}:latest
docker push {your_ACR_name}.azurecr.io/deploy/{your_image_name}:latest

An example usage for the component is provided in the compile_pipeline.py file. It is a single-operation pipeline to use this component and use_azure_secret function from kfp.azure. You can also provide service principal parameters as pipeline parameters.

@dsl.pipeline(
    name='AML Component Sample',
    description='Deploy Model using Azure Machine learning'
)
def model_deploy(
    resource_group,
    workspace
):

    operation = deploy_operation(deployment_name='deploymentname',
                            model_name='model_name:1',
                            tenant_id='$(AZ_TENANT_ID)',
                            service_principal_id='$(AZ_CLIENT_ID)',
                            service_principal_password='$(AZ_CLIENT_SECRET)',
                            subscription_id='$(AZ_SUBSCRIPTION_ID)',
                            resource_group=resource_group,
                            workspace=workspace,
                            inference_config='scripts/inferenceconfig.json',
                            deployment_config='scripts/deploymentconfig.json'). \
                            apply(use_azure_secret()). \
                            apply(use_image(deploy_image_name))

if __name__ == '__main__':
    compiler.Compiler().compile(model_deploy,  __file__ + '.tar.gz')

Note: The deployment name requires alphanumeric characters and the model_name shall be a combination of "model name" and "version number".

Once the deployment has finished, navigate to your Azure Machine Learning Workspace and select the Endpoints tab to find your Real-Time endpoints. You can access your model as a web-service by the REST endpoint. It should look like "http://{your_rest_endpoint}.{region}.azurecontainer.io/score?{your_data}". Alternatively, the output score_uri should also contain the information for the model endpoint. The output_config contains detailed information for the deployed web service.