This project implements a Convolutional Neural Network (CNN) for MNIST digit classification with a complete CI/CD pipeline using GitHub Actions. The model architecture uses batch normalization, dropout, and global average pooling to achieve >95% accuracy on the test set.

• Input Layer: Accepts 28x28 grayscale images
• 5 Convolutional layers with batch normalization and dropout
• Global Average Pooling
• Final Dense layer with 10 outputs (one for each digit)
• Total parameters: = 19,050

• Automated training and testing pipeline
• Model validation checks:
  • Parameter count verification (< 20,000 parameters)
  • Input shape validation (28x28)
  • Output shape validation (10 classes)
  • Accuracy threshold (> 99.3% on test set)
• Automatic model versioning with timestamp and accuracy
• CPU-only training support for both local and CI environments

• Python 3.8 or higher
• PyTorch (CPU version)
• Other dependencies listed in requirements.txt

1. Clone the repository:

git clone https://github.com/walnashgit/NNAndBasicCICD.git
cd NNAndBasicCICD

2. Create and activate a virtual environment:

python -m venv venv
source venv/bin/activate 
On Windows: 
venv\Scripts\activate

3. Install dependencies:

pip install -r requirements.txt

4. Train the model:

python train.py

5. Run tests:

python -m pytest test_model.py

The project includes a GitHub Actions workflow that automatically:

1. Sets up a Python environment
2. Installs dependencies (CPU-only versions)
3. Trains the model
4. Runs validation tests
5. Archives the trained model as an artifact

The pipeline is triggered on every push to the repository.

• Dataset: MNIST (60,000 training images, 10,000 test images)
• Batch Size: 128
• Optimizer: SGD (lr=0.01)
• Loss Function: Negative Log Likelihood
• Training: Single epoch with progress bar showing loss and accuracy
• Model Saving: Automatic with timestamp and accuracy in filename

• Training Accuracy: 99.59%

• Test Accuracy: 99.37%%

The test suite (test_model.py) verifies:

• Model architecture compatibility with 28x28 input images
• Total parameter count (< 20,000)
• Model accuracy on test set (> 99.3%)
• Output shape (10 classes)

Trained models are saved in the models/ directory with the name mnist_model_YYYYMMDD_HHMMSS_accXX.X.pth where:

• YYYYMMDD: Date
• HHMMSS: Time
• XX.X: Achieved accuracy