Forest Fires Prediction - MLOps Project

Project Overview

This project demonstrates an end-to-end MLOps pipeline for wildfire prediction, with a primary focus on implementing and showcasing MLOps best practices and tools. While it includes a wildfire prediction model, the main objective is to illustrate a robust MLOps workflow rather than to create a production-grade prediction system.

The implementation covers the full machine learning lifecycle, including data processing, model training, experiment tracking, model deployment, monitoring, and serving predictions through a web API. This serves as a comprehensive example of how to operationalize machine learning models in a production environment.

Dataset

• Source: UCI Machine Learning Repository
• Name: Algerian Forest Fires Dataset
• Link: Dataset Information

Tech Stack

• MLflow: Experiment tracking and model registry
• Optuna: Hyperparameter optimization
• Prefect: Workflow orchestration
• Evidently: Model monitoring
• Docker: Containerization
• PostgreSQL: Database management
• FastAPI: Web API development

Project Structure

.
├── database/           # Database configuration
├── deploy-batch/       # Batch deployment
├── experiment-tracking/ # MLflow experiment tracking
├── monitoring/         # Model monitoring
├── orchestration/      # Workflow orchestration
├── train-deploy/       # Model training and deployment
├── tests/              # Test files
└── web-api/           # Web API service

Getting Started

Prerequisites

• Docker
• Docker Compose

Installation

1. Clone the repository
2. Build and start the services:

   docker compose up -d --build

Usage

Model Training

1. Access the orchestration container:

   docker compose exec -it orchestration /bin/bash
   python flow.py

2. Monitor training progress at http://localhost:5000
3. View experiment runs in MLflow UI at http://localhost:4200

Batch Processing

1. Run the batch service:

   docker compose --profile batch-service up --build batch-service

2. Monitor batch processing in the Prefect UI at http://localhost:4200
3. Predictions are saved in the batch-data volume at /batch-data/output

Generate Additional Data for Monitoring

To create sufficient data for meaningful monitoring, repeat the model training and batch processing steps:

1. Run the orchestration flow again:

   docker compose exec -it orchestration /bin/bash
   python flow.py

2. Run the batch service again:

   docker compose --profile batch-service up --build batch-service

This will generate additional data points that will be used for model monitoring and drift detection.

API Testing

Test the prediction API with sample requests. The test results will be displayed directly in the terminal:

docker compose up -d
docker compose --profile test up test-api

The test script will output the results of:

• API health check
• High-risk fire scenario prediction
• Low-risk fire scenario prediction

Model Monitoring

1. Start the monitoring service:

   docker compose --profile monitoring up monitoring -d

2. View monitoring reports in the monitoring/reports directory

3. Reports include:

   • CSV files with detailed metrics
   • PNG visualizations of model metrics

   Note about the visualization: The PNG graph might show what appears to be a single line because the batch processing was run back-to-back with minimal time difference between runs. This causes the data points to overlap in the visualization. In a production environment with more time between runs, you would see distinct data points showing the model's performance over time.

Monitoring

Model performance and data drift are monitored using Evidently. Reports are generated after each batch prediction run and can be found in the monitoring/reports directory.