Digit Identifier

A handwritten digit identifier built from scratch using Python and NumPy.

This project trains a simple neural network on MNIST-style digit data and predicts digits from a 28x28 image input. The model logic is implemented manually without using PyTorch or TensorFlow.

Features

• Neural network built from scratch using NumPy
• Manual forward propagation
• ReLU activation
• Softmax output layer
• One-hot encoding
• Backpropagation
• Gradient descent
• Model parameter saving using .npz
• Local image-based digit prediction
• Popup result display

Project Structure

DigitIdentifier/
├── data/
│   └── mnist.npz
│
├── mainNeuralNetwork/
│   ├── main.py
│   └── app.py
│
├── digit.png
└── README.md

How It Works

main.py trains the neural network and saves the learned parameters:

w1, b1, w2, b2

These parameters are stored in:

data/mnist.npz

app.py loads the saved model parameters, reads digit.png, converts it into a 28x28 normalized input, runs forward propagation, and displays the predicted digit in a popup.

Installation

Install the required packages:

pip install numpy pandas pillow matplotlib

Usage

Train the model:

python mainNeuralNetwork/main.py

Run digit prediction:

python mainNeuralNetwork/app.py

Result

The model reached around 95% training accuracy.

Notes

This project was built to understand the internal working of neural networks, including forward propagation, backpropagation, gradient descent, and parameter updates.