README.md

metadata

title: MNIST Digit Classifier
emoji: 🔢
colorFrom: blue
colorTo: purple
sdk: gradio
sdk_version: 3.50.0
app_file: app.py
pinned: false

🔢 MNIST Handwritten Digit Classifier

A Convolutional Neural Network (CNN) trained to recognize handwritten digits (0-9) with 99.17% accuracy on the MNIST test set.

🎯 Model Details

• Architecture: 2-layer CNN with batch normalization and dropout
• Parameters: 421,066
• Training Data: 60,000 MNIST handwritten digits
• Validation Data: 9,000 images (stratified split)
• Test Accuracy: 99.17%
• Framework: PyTorch 2.0.1

Architecture Breakdown

Input (28x28x1)
→ Conv2D(32, 3x3) + BatchNorm + ReLU
→ Conv2D(64, 3x3) + BatchNorm + ReLU + MaxPool(2x2)
→ Dropout(0.25)
→ Fully Connected(128) + ReLU + Dropout(0.5)
→ Output(10) + Softmax

🚀 Usage

Draw a digit in the canvas or upload an image. The model will predict the digit and show confidence scores for all 10 classes.

Tips for Best Results

• Draw the digit large and centered
• Use clear, bold strokes
• Keep digits similar to MNIST style (white on black background)
• Numbers should be recognizable and not ambiguous

📊 Performance Metrics

Metric	Value
Test Accuracy	99.17%
Training Time	20 epochs (~10 mins on GPU)
Inference Time	<50ms per image

Per-Class Performance

All digits achieve >98% accuracy, with digit '1' being the easiest (99.7%) and digit '5' being the most challenging (98.3%).

🛠️ Technical Stack

• Deep Learning: PyTorch 2.0.1
• Interface: Gradio 3.50.0
• Experiment Tracking: MLflow
• Deployment: Docker + Hugging Face Spaces

📁 Project Structure

├── app.py                 # Gradio interface
├── models/
│   └── best_model.pt      # Trained CNN weights
├── scripts/
│   ├── models.py          # CNN architecture
│   ├── inference.py       # Inference wrapper
│   └── data_loader.py     # MNIST data utilities
├── examples/              # Sample digits (0-9)
└── requirements.txt       # Python dependencies

🔬 Training Details

• Optimizer: Adam (lr=0.001)
• Loss Function: Cross-Entropy
• Data Augmentation: Rotation (±10°), scaling (0.9-1.1x)
• Regularization: Dropout (0.25, 0.5) + Batch Normalization
• Early Stopping: Patience=5 epochs on validation loss

📈 Training Results

• Converged in 20 epochs
• No overfitting observed (train/val gap <0.5%)
• Stable learning curve with LR scheduling

🎓 About MNIST

The MNIST dataset is a classic benchmark in computer vision, containing 70,000 grayscale images of handwritten digits (28×28 pixels). It's widely used for training and testing image classification models.

📝 License

MIT License - This project is for educational purposes.

🙏 Acknowledgments

• MNIST dataset by Yann LeCun et al.
• Built as part of an AI Engineering course focusing on MLOps best practices