🧠 MNIST Digit Recognizer — From Scratch in NumPy

This project is a handwritten digit recognizer built entirely from scratch using NumPy — no external machine learning libraries like TensorFlow, PyTorch, or Keras were used. It replicates a basic neural network with forward/backward propagation, softmax, ReLU, and cross-entropy loss.

---

✨ Features

• Fully custom neural network with:
  • Arbitrary number of hidden layers
  • Xavier/He initialization
  • ReLU + Softmax activations
  • Cross-entropy loss + Gradient descent
• Dropout regularization support
• Custom data loading (from folders of .png)
• Training on custom handwritten digits
• Test accuracy tracking
• Exportable and loadable models (.npz)
• CLI prediction tool that works on any image

---

🗂️ Project Structure

MNIST_FROM_SCRATCH/
├── code/ # All Python source code
│ ├── network.py # Core neural network logic
│ ├── predict.py # CLI tool for predictions
│ ├── data_loader.py # Loads MNIST/custom data
│ ├── nn_math.py # Activation functions & loss
│ ├── main.py # Training script
│ ├── preprocessing.py # Preprocesses custom data to match MNIST preprocessing
├── custom_data/ # Your own 28×28 black-bg white-digit PNGs
├── data/ # Raw MNIST files (idx format)
├── models/ # Trained .npz model files
├── test_data/ # Digits to test the model
└── README.md

---

Predicting images

To classify images:

python predict.py <model_path> <image(s)_or_folder>

I suggest using the augmented_more_epochs.npz model from the models/ folder as it is the best one.