BananaCheck - MLOps Learning Project

A modern ML project showcasing MLOps best practices with banana ripeness classification.

Live Demo: https://daystobananadeath.streamlit.app/

Project Goals

This project was built as a learning exercise to master key MLOps technologies:

• Configuration Management with Hydra
• Experiment Tracking & Model Versioning with Weights & Biases
• API Development with FastAPI
• Model Deployment with ONNX Runtime
• Web Interface with Streamlit

Features

• AI-powered banana ripeness estimation (days remaining before spoilage)
• REST API for programmatic access
• Web interface for easy image upload and prediction
• Model versioning and artifact management via W&B
• Production-ready deployment with Docker support

Tech Stack

Core ML/MLOps

• PyTorch Lightning - Training framework
• Hydra - Configuration management
• Weights & Biases - Experiment tracking, model versioning
• ONNX Runtime - Model inference optimization

API & Web

• FastAPI - REST API with automatic documentation
• Streamlit - Interactive web interface
• Uvicorn - ASGI server

Data & Vision

• Albumentations - Image augmentation pipeline
• PIL/OpenCV - Image processing

Model Performance

• Architecture: ResNet50 & Vision Transformer (ViT-B/16)
• Dataset: 20K+ banana images across 5 ripeness stages
• Classes: unripe, ripe, overripe, rotten, unknowns
• Deployment: ONNX optimized for CPU inference

Quick Start

API Usage

# Start the API
uvicorn src.api:app --host 0.0.0.0 --port 8000

# Test prediction
curl -X POST "http://localhost:8000/predict" \
     -H "Content-Type: multipart/form-data" \
     -F "file=@banana.jpg"

Web Interface

# Start Streamlit app
streamlit run streamlit_app.py

# Navigate to http://localhost:8501

Environment Setup

pip install -r requirements.txt

# For W&B model loading
export WANDB_API_KEY=your_api_key

Project Structure

├── src/
│   ├── api.py              # FastAPI REST endpoint
│   ├── model.py            # Model architectures
│   ├── train.py            # Training pipeline
│   └── inference.py        # Inference utilities
├── configs/                # Hydra configuration files
├── streamlit_app.py        # Web interface
├── requirements.txt        # Python dependencies
└── Procfile               # Deployment configuration

Key Learning Outcomes

Hydra Configuration Management

• Structured config files for different components (model, data, training)
• Easy experiment switching with config overrides
• Environment-specific configurations

W&B MLOps Pipeline

• Automated experiment tracking with metrics logging
• Model artifact versioning and storage
• Seamless model deployment from artifacts
• Training curve visualization and comparison

Production API Design

• RESTful API with proper error handling
• Automatic model loading from W&B artifacts
• Health checks and monitoring endpoints
• ONNX optimization for fast inference

Dataset Sources

Built using established vision datasets:

• BananaRipeness Dataset by Chuquimarca et al.
• Additional synthetic and augmented data

Deployment

The project includes production-ready deployment configurations:

• Heroku/Railway: via Procfile
• Docker: containerized deployment
• Environment variables: for secure API key management

---

Built to explore modern MLOps practices and deployment strategies