Operationalizing Machine Learning

Overview

I created and deployed a Machine Learning model in this project, first using Azure ML Studio and then using Azure Python SDK. I used Swagger UI to interact with the REST API resources of the deployed model.

The dataset used for training the model contained data about the marketing campaigns of a banking institution. My goal was to predict whether the client will subscribe to the bank or not.

Architectural Diagram

In this project, we followed the following steps:

1. Automated ML experiment
2. Deploy the best Model
3. Enable logging
4. Swagger Documentation
5. Consume Model endpoints
6. Create and publish a pipeline
7. Documentation

Key Steps and Screenshots

1. Automated ML Experiment: This step involves creating an experiment using Automated ML, configuring a compute cluster, and using that cluster to run the experiment.

   

   Fig. 1: Registered dataset

   ---

   

   Fig 2: A complete run from the AutoML experiment

   ---

   

   Fig 3: Top 5 AutoML models

   ---

   

   Fig 4: Best AutoML Model Metrics

   ---

2. Deploy the Best Model: Based on the accuracy and AUC values, I decided the best model (it is at the top of the Models tab in Azure). I deployed the best model using Azure Container Instance (ACI).

3. Logging: In this step, we enabled Application Insights and retrieved logs.

   

   FIg 5: Image showing that Applications Insights is enabled

   ---

   

   Fig 6: Output of logs.py

   ---

4. Swagger Documentation - "Swagger" is an open-source tool that helps design, build, document, and consume REST APIs. In this project, we have used a tool called Swagger UI for generating interactive API documentation. Using Swagger UI, a user can try out API calls directly in the browser.

   Azure creates a swagger.json file for all the deployed models. I used this JSON file to consume the deployed model using Swagger. To that end, I executed the shell script swagger.sh, which starts the swagger docker container. Next, I created a Python server on Port 8000 using serve.py.

   

   Fig 7: Running Python server on Port 8000

   ---

   Finally, I opened the Swagger-UI tool and clicked on Live Demo to interact with the deployed model's API. Below are some screenshots taken from the Swagger-UI tool.

   

   Fig 8: Swagger UI - Live Explore

   ---

   

   Fig 9: Swagger UI - GET Method

   ---

   

   Fig 10: Swagger UI - POST Method

5. Consume Model Endpoints: After deploying the best model, I consumed its REST API endpoint using endpoint.py after modifying the scoring_uri and key values to match what appears in the Consume tab of the endpoint.

   

   Fig 11: Output of endpoint.py

6. Create, Publish and Consume a Pipeline: In this step, I used the aml-pipelines-with-automated-machine-learning-step.ipynb notebook to create, publish and cosume a pipeline using the Azure Python SDK.

   

   Fig 12: Created a pipeline

   ---

   

   Fig 13: Pipeline Run Complete

   ---

   

   Fig 14: Created pipeline endpoint

   ---

   

   Fig 15: Image showing the Active status of Pipeline Endpoint and the created swagger.json file

   ---

   

   Fig 16: Output of the RunDetails Widget in the Jupyter Notebook

Areas of Improvement

• Tackle class imbalance by oversampling of minority class
• Try other custom algorithms
• Run AutoML for a longer time to get better results.

Screen Recording

https://youtu.be/srqP9J4dv0k