MLflow Example Project + Notebook

This repository provides an example of dataset preprocessing, model training and evaluation, model tuning and finally model serving (REST API) in a containerized environment using MLflow tracking, projects and models modules.

This project contains an MLflow project that trains a GBRT (Gradient Boosted Regression Tree) model on UC Irvine's Bike Sharing Dataset Data Set and uses a Docker image to capture the dependencies needed to run training and inference code.

Talk: @PyDataRiyadh:

• https://twitter.com/PyDataRiyadh/status/1291043529146466304
• https://twitter.com/PyDataRiyadh/status/1314841078999154689

Talk slides: MLflow-presentation.pdf

Notebook: MLflow-example-notebook.ipynb

Structure of this project

• MLproject specifies the Docker container environment to run the project and defines command and parameters in entry_points
• Dockerfile used to build the image referenced by the MLproject
• requirements.txt: defined python dependencies needed to build training and inference docker image
• mlflow_project_driver.py: creates an MLflow experiment for model training and tuning and launches MLflow runs in parallel in docker containers.
• mlflow_model_driver.py: finds best training run and starts a REST API model server based on MLflow Models in docker containers.
• train.py : contains a file that trains a scikit-learn model and uses MLflow Tracking APIs to log the model and its metadata (e.g., hyperparameters and metrics)
• data/hour.csv: Bike Sharing Dataset Data Set

Getting started

Prerequisites:

1. Python 3
2. Install Docker per instructions at https://docs.docker.com/install/overview/
3. Install mlflow pip install mlflow

Dockerized model training and training with MLflow:

1. clone this repo: git clone https://github.com/alfozan/mlflow-example
2. build the image for the project's Docker container environment: docker build -t mlflow_example -f Dockerfile .
3. Start training and tracking: python3 mlflow_project_driver.py

Run MLflow tracking UI:

In the same repo directory, run mlflow ui --host 0.0.0.0 --port 5000 UI is accessible at http://localhost:5000/

Dockerized MLflow model serving (REST API)

In the same repo directory, run python3 mlflow_model_driver.py

Inference request:

curl --silent --show-error 'http://localhost:5001/invocations' -H 'Content-Type: application/json' -d '{
    "columns": ["season", "year", "month", "hour_of_day", "is_holiday", "weekday", "is_workingday", "weather_situation", "temperature", "feels_like_temperature", "humidity", "windspeed"],
    "data": [[1, 0, 1, 0, 0, 6, 0, 1, 0.24, 0.2879, 0.81, 0.0000]]
}'