MLOPs for Azure Machine Learning Vision Models with Automated ML

Introduction

This guide provides step-by-step instructions for setting up and configuring your Azure Machine Learning environment using both PowerShell and Bash. Please replace placeholder values (e.g., <subscription-id>, <workspace>, <resource-group>, <location>) with your actual Azure subscription and resource details.

Local Development Environment Setup

• Install AZ CLI
• Install AML Python SDK V2
• Install az ml CLI v2
  • Run the following command to see the az extension

    az extension list

  • If you see the azure-cli-ml extension, remove it by running the following command:

    az extension remove -n azure-cli-ml

  • If you see ml extension, remove it by running the following command:

    az extension remove -n ml

  • Install the latest az ml CLI v2 extension by running the following command:

    az extension add -n ml -y 

• git clone repo git clone https://github.com/anildwarepo/amlvision-mlops

Azure Login

PowerShell:

az login
az account set -s <subscription-id>
az configure --defaults workspace=<workspace> group=<resource-group> location=<location>
az ml workspace show -n <workspace> -g <resource-group>

Bash:

az login
az account set -s <subscription-id>
az configure --defaults workspace=<workspace> group=<resource-group> location=<location>
az ml workspace show -n <workspace> -g <resource-group>

Set Environment Variables

PowerShell:

$env:aml_workspace="<workspace>"
$env:aml_resource_group="<resource-group>"

Bash:

export aml_workspace=<workspace>
export aml_resource_group=<resource-group>

Github Secret Setup

This task requires the following:

• Creating a Personal Access Token (PAT) in Github
• Adding a Service Principal (SP) to your forked repository in Github

To create PAT (Personal Access Token), go to Settings of your account, NOT repo setting.

• Go to Settings > Developer Settings > Personal Access Tokens
• Click on Generate new token
• Give it a name and select the following scopes:

Create a Service Principal (SP) in Azure and assign it a role. Replace placeholders with your specific values. This requires contributor permission to the subscription.

az ad sp create-for-rbac --name {REPLACE_SPNAME} --role contributor --scopes /subscriptions/{REPLACE_SUBSCRIPTIONID}/resourceGroups/{REPLACE_RESOURCEGROUPNAME} --sdk-auth

Create self hosted runners in Github (optional)

Follow the steps in the link below to create self hosted runners in Github. This is required for the Github Actions to run.

Self-hosted runners

Create AML Environment

PowerShell:

cd C:\source\repos\amlvision-mlops

Bash:

cd /source/repos/amlvision-mlops

az ml environment create --file ./src/environment/create_aml_env.yml -w $aml_workspace -g $aml_resource_group

Upload Data from Local Path to Azure ML Datastore

az ml data create -f ./src/upload_data/upload.yml -w $aml_workspace -g $aml_resource_group

Create Compute Cluster

az ml compute create -f .\src\compute\create-cpu-cluster.yml -w $aml_workspace -g $aml_resource_group

Create GPU Compute Cluster

Update the parameters in the create-gpu-cluster.yml file as per your requirement and training data volume.

az ml compute create -f .\src\compute\create-gpu-cluster.yml -w $aml_workspace -g $aml_resource_group

Data Prep Job (Optional)

az ml job create -f .\src\dataprep\dataprep.yml -w $aml_workspace -g $aml_resource_group

Create MLTable for AutoML Training

Update the following parameters in the create_mltable.yml file:

In this step, we convert the coco format dataset to MLTable format. This is required for training using AutoML. This will also create the train and validation MLTable datasets.

Update the following parameters in the create_mltable.yml file.

• training_data_folder_path e.g. azureml://datastores/workspaceblobstore/paths/LocalUpload/8c920d5d5994f81cc82b286c429a1667/training_data/training-mltable-folder/

• azure_storage_url - Azure Storage url to the images folder. e.g "azureml://subscriptions/840b5c5c-3f4a-459a-94fc-6bad2a969f9d/resourcegroups/ml/workspaces/ws02ent/datastores/workspaceblobstore/paths/LocalUpload/8c920d5d5994f81cc82b286c429a1667/training_data/augmented-samples/images/"

• output_folder same as training_data_folder_path

• environment - This is the environment name created in the create environment step. e.g. azureml:mltable_env:1

• compute - This is the compute name created in the create compute step. e.g. azureml:cpu-cluster

az ml job create -f .\src\create_mltable\create_mltable.yml  -w $aml_workspace -g $aml_resource_group

Create Train and Validation MLTable Datasets

Update path to the MLTable folder in the create_mltable_train_dataset.yml and create_mltable_validation_dataset.yml files.

az ml data create -f .\src\create_mltable\create_mltable_train_dataset.yml -w $aml_workspace -g $aml_resource_group
az ml data create -f .\src\create_mltable\create_mltable_validation_dataset.yml -w $aml_workspace -g $aml_resource_group

Create AutoML Training Job

Create pipeline job. You can run multiple experiments by changing tile_grid_size to either "5x5" or "9x9" to see if we are able to get better results.

az ml job create -f .\src\training\automl_training_pipeline.yml -w $aml_workspace -g $aml_resource_group

Deploy model

az ml online-endpoint create -f .\src\deployment\endpoint.yml -w $aml_workspace -g $aml_resource_group

az ml online-deployment create -f .\src\deployment\deployment.yml -w $aml_workspace -g $aml_resource_group

Inference

Update api-key in the score.py file and run the following command to test the endpoint.

./src/inference/score.py

Result: