MLOps Demo

This repo shows how to use Azure Machine Learning (ML) Services and Azure DevOps to manage machine learning using DevOps principles and practices.

Model

A model is trained on Pima Indian Diabetes data to predict the likelihood of diabetes from health markers. The training is done using Azure Machine learning workspaces and every step of the process is in code. The trained model is deployed to Azure Container Instances (ACI) for consuming. Once approved, the model is deployed to AKS to show production-grade deployment.

AutoML

While a data scientist could run the training pipeline numerous times with a different c value for the LogisticRegression model, a better way to hypertune the model is to use AutoML. Using AutoML, Azure ML iterates the training/scoring automatically for scores of runs, selecting the best model (and parameters). This model is deployed in the same manner as the model that is "manually" trained.

Model API URLs

1. Scoring URL: Get this from the ML Portal under Endpoints.
2. For the Swagger, change /score to /swagger.json in the URL to get the OpenAPI definition.

Website

A simple website is deployed that consumes a deployed model. See this page for a demo.

Pipeline Status

Model Pipeline

Stage	Status
Provision Workspace
Train Model
Deploy to ACI
Deploy to AKS

Site Pipeline

Stage	Status
Build Site
Provision Infra and Deploy to DEV
Provision Infra and Deploy to PROD

AutoML Pipeline

Stage	Status
Provision Workspace
Train Model
Deploy Model to ACI