Cloud-Based Deep Learning for Handwritten Digit Recognition
About | Overview | Demo | Features | Tech Stack | Architecture | Installation | Evaluation | Future Directions
Authors: Ruida Zeng, Aadarsh Jha, Ashwin Kumar, Terry Luo
Professor & Faculty Mentor: Dr. Aniruddha Gokhale, Professor of Computer Science, Vanderbilt University
Vanderbilt University Fall 2021 Immersion Vanderbilt Project – Program Information
Stratus is a cloud-deployed, end-to-end deep learning solution designed to recognize handwritten digits in real time. This project integrates cloud computing and distributed deep learning principles to deliver a scalable, production-grade pipeline for digit recognition. It leverages an ensemble of cloud-based tools and modern full-stack technologies to streamline real-time predictions via a web interface.
Try the live demo at stratus-final.ml
Stratus is an end-to-end, full-stack application built for handwritten digit classification. Users draw a digit in the web interface, which is then sent to the cloud infrastructure where the deep learning model makes a prediction. The cloud-based backend processes and returns the result in real time.
The system combines:
- Deep Learning for accurate digit recognition using the MNIST dataset.
- Cloud Computing for distributed, scalable processing.
- Full-Stack Architecture with a user-friendly web interface.
- End-to-End Machine Learning Pipeline: From user input to model prediction, all components are integrated and productionized.
- Distributed Architecture: Scales to handle variable user traffic with load balancing and real-time data processing.
- Intuitive User Interface: A canvas for digit drawing, real-time feedback, and prediction probability graph.
- Front-end: HTML, JavaScript
- Back-end: Python Flask, Apache Kafka
- Data Processing: Apache Spark ML
- Cloud Infrastructure: Chameleon Cloud, Kubernetes, Docker
- Orchestration & Automation: Ansible, Vagrant
- Database: CouchDB
- Web Server: Nginx
The Stratus architecture comprises three main components:
- Cloud-Based Infrastructure: Managed through Docker and Kubernetes across multiple VMs, ensuring scalability and robustness.
- Full-Stack Web Application: A user interface to draw digits and view predictions, with Flask managing backend requests.
- Deep Learning Model: A Convolutional Neural Network (CNN) trained on the MNIST dataset, deployed via Spark for distributed model inference.
- Python 3.x
- Docker, Kubernetes, Ansible, Vagrant, Nginx, CouchDB
-
Clone the repository:
git clone https://github.com/ruidazeng/stratus.git cd stratus -
Set up Virtual Machines:
vagrant up
-
Deploy services using Ansible:
ansible-playbook playbook_demo_master.yml
-
Run the Flask server:
python3 fullstack-python/app.py
After deployment, navigate to the live link or access the application locally on localhost:5000. Draw a digit on the canvas and press "Predict" to see the model's prediction along with a probability graph.
The project has been rigorously tested for:
- Model Accuracy: Achieved 97.4% accuracy on the MNIST dataset.
- Load Balancing: Stress tests were conducted to determine the optimal user load for fast and reliable predictions.
- Response Time: System is optimized for minimal response time under various user loads.
For more details on testing and evaluation, see the technical report.
Potential improvements include:
- Support for multi-digit recognition.
- Enhanced model robustness and accuracy for varied input styles.
- Real-world deployment in a large-scale production environment.
This project is licensed under the MIT License. See the LICENSE file for details.