NanoNet – Neural Network From Scratch (NumPy Only)

This project implements a fully functional multi-layer neural network from scratch using only NumPy.
No TensorFlow, no PyTorch — just pure mathematics and linear algebra.

The network is trained on the MNIST handwritten digits dataset (data/train.csv format) and supports:

• He initialization
• ReLU activation
• Softmax output
• Cross-entropy loss
• Full forward + backward propagation
• Gradient descent optimization
• Accuracy + loss tracking
• A clean NeuralNet class with .train(), .predict(), .evaluate()

---

🚀 Features

• Modular architecture – define arbitrary layers like [784, 128, 64, 10]
• Pure NumPy implementation – complete understanding of the internals
• Training loop with loss + accuracy logging
• Easy prediction + evaluation
• Object-oriented design (NeuralNet.py)

---

📁 Project Structure

/
├── NeuralNet.py        # Neural network implementation (class)
├── train.py            # Script to load MNIST, train, and evaluate
└── data/
    └── train.csv       # MNIST dataset (Kaggle digit recognizer format)

---

🧠 How It Works

1. Weights initialized using He initialization
2. Forward pass
   • Hidden layers → ReLU
   • Output layer → Softmax
3. Loss computed using cross-entropy
4. Backpropagation
   • Softmax gradient
   • ReLU derivative
   • Layer-by-layer weight updates
5. Gradient descent training

This mimics the internal logic of modern deep learning libraries — but written manually.

---

🏁 Usage

Train the model

Run:

python3 train.py

Your training output will log:

Iteration 0   Accuracy: 0.07   Loss: 2.38
Iteration 10  Accuracy: 0.45   Loss: 1.95
...

---

🧪 Evaluate

Inside train.py:

model.evaluate(X_test, Y_test)

Example output:

Test Accuracy: 0.92

---

🔍 Sample Prediction

model.test_predictions(10, X_test, Y_test)

Output:

Model Prediction: 5
True Label: 5

---

📚 Requirements

• Python 3.9+
• NumPy
• Pandas
• Matplotlib (optional, for visualization)

Install:

python3 -m venv venv
source venv/bin/activate
pip install numpy pandas

---

🌟 Future Improvements

• Add mini-batch gradient descent
• Add Adam optimizer
• Add regularization (L2 / dropout)
• Add plotting utilities for loss + accuracy
• Save/load model weights
• Implement momentum-based training

---

📜 License

MIT License – free to use, modify, and learn from.