Style Transfer + Model Distillation (TensorFlow/Keras)

This project implements a Neural Style Transfer pipeline that uses a high-quality VGG19-based teacher model and a lightweight, real-time student model trained via knowledge distillation. Designed for deployment, it features a Flask web interface and is containerized with Docker for cloud readiness.

Deployed on: https://styliqueai.azurewebsites.net/

---

Features

• VGG19 teacher for detailed artistic stylization
• Lightweight student model (U-Net/CNN) for fast inference
• Custom training pipeline using teacher-generated data
• Flask web app for interactive upload & preview
• Modular codebase with separate components for cleaning, preprocessing, training, and inference
• Dockerized for easy deployment (Azure-ready)

---

Project Structure

├── app.py                     # Main Flask app (frontend + inference logic)
├── DataCleaner.py             # Image filtering & dataset cleaning utilities
├── DataProcessor.py           # Data loader and augmentation logic
├── Dockerfile                 # Docker setup for full app deployment
├── ModelRunner.py             # Abstraction to run teacher/student models
├── ModelTrainer.py            # Training script for student model
├── style_transfer_logic.py    # Inference logic (used by Flask)
├── utils.py                   # Helper functions (e.g., image loading)
├── style_transfer_model.png   # Architecture diagram or sample image
├── requirements.txt           # Python dependency list
├── README.md
│
├── models/                    # Saved models (e.g., .keras files)
├── static/                    # Static web assets (uploads, styles, results)
└── templates/                 # HTML templates (Jinja2: index.html, result.html)

---

Requirements

Install dependencies with:

pip install -r requirements.txt

Key Libraries:

• TensorFlow 2.x
• Flask
• Pillow
• NumPy
• gunicorn
• Werkzeug

---

Running the App

1. Local (Flask Dev Server)

python app.py

Then go to http://localhost in your browser.

---

2. Docker Build & Run

docker build -t style-transfer-app .
docker run -p 80:80 style-transfer-app

---

3. Deploying on Azure App Services

• Push the Docker image to Azure Container Registry
• Deploy via Azure App Services with container option
• Use environment variable: FLASK_SECRET_KEY (optional but recommended)

---

Training Overview

The teacher model (VGG19) produces stylized outputs from content/style pairs.
The student model is trained on this output using perceptual loss + image reconstruction, allowing:

• Faster inference (real-time)
• Smaller model size (deployable anywhere)
• Maintain near-teacher visual quality

Use ModelTrainer.py for training and ModelRunner.py for running both models in isolation.

---

Web Interface

The app provides:

• Drag-and-drop or file upload for content and style images
• Option to select from predefined styles
• Dual output view (teacher and student results)
• Configurable iteration counts for both models
• Downloadable results with performance stats

---

Preview

---