🧠 Neural Network from Scratch (Pure Python Implementation)

A fully functional neural network built from scratch in Python, without relying on deep learning frameworks like TensorFlow or PyTorch.
This project demonstrates the core mechanics behind neural networks — from forward propagation to backpropagation and gradient descent — with clean, modular, and well-documented code.

🚀 Features

✅ Custom Neural Network Class – Pure Python + NumPy implementation
✅ Multiple Activation Functions – ReLU, Sigmoid, and Tanh
✅ Configurable Architecture – Define any number of layers and neurons
✅ Supports Both Tasks – Classification & Regression
✅ Mini-Batch Gradient Descent – Efficient training
✅ Real-Time Training Visualization – Loss and accuracy plots
✅ Comprehensive Evaluation Metrics – Accuracy, MSE, and R² score

🧩 Project Structure

    neural_network_demo/
    ├── neural_network.py      # Core neural network implementation (classification)
    ├── main.py                # Demo for classification and regression
    ├── requirements.txt       # Dependencies
    ├── run.bat                # Run script for Windows
    ├── run.sh                 # Run script for Linux/Mac
    ├── README.md              # Project documentation
    └── .gitignore             # Ignore unnecessary files

⚙️ Installation

1️⃣ Clone the repository:

git clone https://github.com/sankhya007/demo_neural.git
cd demo_neural

2️⃣ Install dependencies:

pip install -r requirements.txt

▶️ Usage

Run the demo: bash python main.py

You’ll see: - Training progress with live updates every 100 epochs
- Real-time loss and accuracy visualization
- Final evaluation metrics for both classification and regression tasks

📊 Results

• Achieves ~78% accuracy on multi-class classification datasets
• Smooth convergence curves for both loss and accuracy
• Regression model yields strong R² performance and low MSE

🧠 Concepts Covered

• Forward Propagation
• Backpropagation
• Activation Functions (ReLU, Sigmoid, Tanh)
• Softmax Output for Multi-Class Problems
• Mini-Batch Gradient Descent
• Loss Functions (Cross-Entropy & MSE)
• Model Evaluation Metrics

📈 Example Visualizations

• Classification: Loss & Accuracy curves over epochs
• Regression: Predicted vs Actual scatter plot with R² score
• Training Diagnostics: Validation loss tracking

💡 Future Improvements

• Add dropout and batch normalization
• Implement momentum and Adam optimizers
• Save and load trained model weights
• Extend to support convolutional layers

🧑‍💻 Author

Sankhyapriyo Dey
📧 GitHub Profile

🪪 License

This project is released under the MIT License – free to use and modify.