Fashion MNIST Classifier 👗

A production-ready image classification system for the Fashion MNIST dataset, featuring a CNN model with ~92% accuracy, REST API, and interactive web interface.

🔴 Live Demo

You can deploy this application directly to Streamlit Cloud to see it in action:

Note: After deploying, update the link above to your specific app URL.

🎯 Quick Start

Option 1: Docker (Recommended)

# Clone repository
git clone https://github.com/millenniumsingha/StyleNet.git
cd StyleNet

# Train model and start services
docker-compose --profile training up train
docker-compose up -d api streamlit

# Access:
# - API Docs: http://localhost:8000/docs
# - Web App:  http://localhost:8501

Option 2: Local Installation

# Install dependencies
pip install -r requirements.txt

# Train model
python -m src.train --model cnn --epochs 15

# Start API
uvicorn api.main:app --reload

# Start Web App (new terminal)
streamlit run app/streamlit_app.py

📊 Model Performance

Model	Test Accuracy	Parameters
Simple (Original)	~88%	~101K
CNN (Current)	~92%	~400K

🏗️ Architecture

Input (28x28x1)
    │
    ▼
┌─────────────────┐
│ Conv Block 1    │  32 filters, BatchNorm, MaxPool, Dropout
└────────┬────────┘
         │
┌────────▼────────┐
│ Conv Block 2    │  64 filters, BatchNorm, MaxPool, Dropout
└────────┬────────┘
         │
┌────────▼────────┐
│ Conv Block 3    │  128 filters, BatchNorm, MaxPool, Dropout
└────────┬────────┘
         │
┌────────▼────────┐
│ Dense (256)     │  BatchNorm, Dropout
└────────┬────────┘
         │
┌────────▼────────┐
│ Output (10)     │  Softmax
└─────────────────┘

🔌 API Reference

Endpoints

Method	Endpoint	Description
GET	/	API information
GET	/health	Health check
GET	/classes	List class names
POST	/predict	Classify single image
POST	/predict/batch	Classify multiple images

Example Request

curl -X POST "http://localhost:8000/predict" \
  -H "Content-Type: multipart/form-data" \
  -F "file=@your_image.png"

Example Response

{
  "success": true,
  "predicted_class": "Ankle boot",
  "predicted_index": 9,
  "confidence": 0.97,
  "top_predictions": [
    {"class_name": "Ankle boot", "class_index": 9, "confidence": 0.97},
    {"class_name": "Sneaker", "class_index": 7, "confidence": 0.02},
    {"class_name": "Sandal", "class_index": 5, "confidence": 0.01}
  ]
}

📁 Project Structure

├── src/                 # Core ML code
│   ├── model.py         # CNN architecture
│   ├── train.py         # Training script
│   └── predict.py       # Inference utilities
├── api/                 # FastAPI backend
│   └── main.py          # REST API endpoints
├── app/                 # Streamlit frontend
│   └── streamlit_app.py # Web interface
├── tests/               # Unit tests
├── models/              # Saved models
├── notebooks/           # Jupyter notebooks
└── legacy/              # Original project files

🧪 Testing

# Run all tests
pytest tests/ -v

# With coverage
pytest tests/ --cov=src --cov=api

📈 Training Your Own Model

# CNN model (recommended)
python -m src.train --model cnn --epochs 15

# Simple model (for comparison)
python -m src.train --model simple --epochs 10

🎨 Supported Classes

Index	Class
0	T-shirt/top
1	Trouser
2	Pullover
3	Dress
4	Coat
5	Sandal
6	Shirt
7	Sneaker
8	Bag
9	Ankle boot

🚀 Future Improvements

• Model versioning with MLflow
• CI/CD pipeline with GitHub Actions
• Kubernetes deployment configs
• Model monitoring and drift detection
• A/B testing framework

🔄 CI/CD Pipeline

The project uses GitHub Actions for Continuous Integration and Deployment:

• Build & Test: runs on every push to verify code quality.
• Docker Build: pushes new images to GitHub Container Registry (GHCR) on release.

📦 Model Versioning (MLflow)

Experiments are tracked using MLflow.

# Start MLflow server
docker-compose up -d mlflow

# View dashboard at http://localhost:5000

📊 Monitoring (Evidently AI)

Data drift is monitored comparing production traffic against training data.

• Drift Logic: Implemented in src/monitoring.py.
• Logs: Predictions are logged to monitoring/current_data.csv.

☸️ Kubernetes Deployment

Deploy to a cluster using the manifests in k8s/:

kubectl apply -f k8s/

Includes:

• API: 2 Replicas, Health Checks.
• Streamlit: LoadBalancer Service.

🔀 A/B Testing

Traffic is routed between stable and canary models based on environment variables:

• Set AB_TEST_RATIO=0.2 to send 20% of traffic to the canary model.

🙏 Acknowledgments

• Fashion MNIST Dataset by Zalando Research
• TensorFlow and Keras teams
• FastAPI and Streamlit communities

---

Originally created as a learning project, upgraded to production-ready status.