MNIST Classification with GitHub Actions

This repository contains a lightweight CNN model for MNIST digit classification that achieves >95% accuracy in a single epoch while maintaining fewer than 25,000 parameters.

Model Architecture

The model uses a carefully designed CNN architecture with the following key features:

• Multiple convolutional layers with batch normalization
• Max pooling to reduce spatial dimensions
• Efficient channel progression (8 → 16 → 32 → 10)
• Global average pooling instead of dense layers
• Total parameters: < 25,000

Requirements

• Python 3.8+
• PyTorch 1.7.0+
• torchvision 0.8.1+
• pytest 6.0.0+
• tqdm 4.50.0+

Install dependencies:

pip install -r requirements.txt

Training

The model includes:

• Image augmentation (rotation and translation)
• Batch normalization for faster convergence
• SGD optimizer with momentum
• One epoch training to achieve >95% accuracy

To train the model:

python train.py

Tests

The repository includes comprehensive tests that verify:

1. Model parameter count (< 25,000)
2. Output shape correctness
3. Forward pass stability
4. Training accuracy (>95% in one epoch)
5. Image augmentation effectiveness

Run tests:

pytest test_model.py -v

GitHub Actions

The repository uses GitHub Actions to automatically:

• Run all tests on every push and pull request
• Verify model meets parameter constraints
• Confirm accuracy requirements
• Validate image augmentation
• Display build status badge

Image Augmentation

The training pipeline includes robust image augmentation:

• Random rotation (-7° to 7°)
• Random translation (±10% in both dimensions)
• Verified on 100+ images during testing

License

MIT