MNIST Handwritten Digit Classifier (99% Accuracy)

A custom CNN achieving over 99% test accuracy on MNIST — from data augmentation to deployment-ready model.

Features

• Custom CNN architecture (2 conv blocks + FC head)
• Data augmentation (rotation, perspective) for robustness
• Adam optimizer with dropout for 99%+ accuracy in 10 epochs
• M3 Air optimized (MPS backend) — trains in <5 minutes locally

Results

• Test Accuracy: 99.29%
• Training Loss Curve: results/loss-curve.png
• Access saved model path results/mnist_cnn_99.pth

Quick Start

# Install
pip install -r requirements.txt

# Run notebook
jupyter notebook code/mnist_cnn.ipynb