Skip to content

Latest commit

 

History

History

simple_machine_learning_deployments

README.md

machine-learning-operations-test

Originally a test for mid level Machine learning Engineer for a company in Thailand.

Since no data and model were actually provided, I synthesized them and adapted the test into my portfolio.

.env file has been adjusted accordingly and for the purpose of this portfolio, will be included in git.

Instruction

A junior data scientist has prepared a prediction model ready to predict client's orders for a restaurant - given that a scientist has no prior knowledge on machine learning operations and bringing the models up for production, you, as a machine learning engineer, has to bring this model to live by yourself. Please demonstrate how would you do it.

Fork this repo and make this model ready to deploy on GCP on any suitable service of your choice

Important

Will require .env file with corresponding keys to run

Disclaimer

While there were attempts to ensure compatibilities across platforms, the main code was only tested on Windows.

Methods

  1. Streamlit: Simple, easy to use, web application tool suitable for small data application.
  2. Airflow Docker: While normally associated with data pipeline, Airflow can be adapted to schedule regular model predict/retraining service.
  3. Flask API: I once used the company's own application in place of Flask to enable API on cloud. I'm a newbie to Flask itself and may not be familiar with security protocol. Nevertheless I made a working Flask API application, and can be tested with your own API call program or my streamlit program.

How to run

  1. All methods: sh 00_prepare_demo.sh to prepare dummy models and synthetic data for demo runs on all methods.
  2. Streamlit: streamlit run simple_streamlit.py, then insert test data with categorical features already encoded (can be obtained from /data after preparing demo in step 0). The results will be displayed on the streamlit web app.
  3. Airflow Docker: (requires running docker engine) sh 02_prepare_airflow_demo.sh. Go to localhost:{AIRFLOW_PORT} on your browser (requires said port to be vacant), and login with user and password specified in .env. Click to enable the daily_prediction_run DAG and wait for scheduled run or activate more runs manually. The prediction results and generated test data will be in /results folder. When done run sh 10_clean_docker.sh.
  4. Flask API: (requires running docker engine) sh 03_prepare_flask_demo.sh. Then run streamlit run simple_streamlit.py, except this time choose 'Call Flask API' on the radio button, then uploaded test data. When done run sh 10_clean_docker.sh.

Mid-Development Updates Log

2023-11-17 21:53: Added data synthesizer to somewhat mimic the data in jupyter notebook.

2023-11-17 22:18: Added module to mimic the preprocessing in the jupyter notebook.

2023-11-17 22.40: Added 2nd part of preprocess for feature set 2.

2023-11-18 01:06: Prepared dummy model modules and model saving functionalities (Stacking doesn't work yet)

2023-11-18 13:35: Method 1 (Streamlit) up (still without Stacking model until further notice)

2023-11-18 16:20: Method 2 (Airflow Docker) up (with only prediction DAG).

2023-11-19 11:15: Restructured project in preparation for Method 3.

2023-11-19 16:10: Method 3 (Flask Web API) up and simplified running steps.