MNIST Digit Recognition with PyTorch

This project implements a Convolutional Neural Network (CNN) using PyTorch to classify handwritten digits from the MNIST dataset. It includes model training, evaluation, and inference on custom images.

Project Structure

ML_/
├── data/                # (Optional) Data directory
├── MNIST/               # MNIST data (raw)
├── src/
│   ├── model.py         # Model definition and training script
│   ├── test.py          # Script to test the model on a custom image
│   ├── mnist_model.pth  # Trained model weights
│   └── image.png        # Example image for inference
├── README.md            # Project documentation
└── ...

Requirements

• Python 3.8+
• PyTorch
• torchvision
• Pillow

Install dependencies:

pip install torch torchvision pillow

Training the Model

The model is defined and trained in src/model.py. To train the model on the MNIST dataset and save the weights:

cd src
python model.py

This will download the MNIST dataset (if not present), train the CNN, and save the model as mnist_model.pth.

Testing/Inference

To predict the digit in a custom image (image.png):

1. Place a 28x28 grayscale image of a digit as src/image.png (or modify the path in test.py).
2. Run:

cd src
python test.py

The script will output the predicted digit.

Model Architecture

• 1 convolutional layer (32 filters, 3x3 kernel)
• ReLU activations
• Max pooling
• Fully connected layer (output: 10 classes)

Notes

• The MNIST dataset is automatically downloaded to the MNIST/raw/ directory.
• The model expects grayscale images of size 28x28 for inference.

License

MIT License