pynnet 0.0.2

A simple neural network library built from scratch in Python ensuring easy application and understanding of artificial neural networks.

PyNNet: A Pythonic Neural Network Library

Welcome to PyNNet! 👋

PyNNet is a beginner-friendly neural network library that helps you learn and understand deep learning from the ground up. Built in pure Python and NumPy, it provides a clean, intuitive API similar to popular frameworks while being transparent about what's happening under the hood.

🎯 Perfect for:

• Learning how neural networks work
• Experimenting with deep learning concepts
• Educational projects and assignments
• Small to medium-sized machine learning tasks

🚀 Why Choose PyNNet?

• Simple, Keras-like API that's easy to learn
• Clear, documented implementation you can understand
• Minimal dependencies (just NumPy!)
• Great for learning deep learning fundamentals

📚 Getting Started

Installation

Install PyNNet using pip:

pip install pynnet

Quick Start Guide

Building a neural network with PyNNet is as simple as 1-2-3:

1. Create a model:

from pynnet.network import sequential
model = Sequential()

2. Add layers:

from pynnet.layers import dense
from pynnet.activation import relu, sigmoid

# Input layer with ReLU activation
model.add(dense(input_size=2, output_size=4, weight_init='he'))
model.add(relu)

# Output layer with Sigmoid activation
model.add(dense(input_size=4, output_size=1, weight_init='xavier'))
model.add(sigmoid)

3. Compile and train:

from pynnet.optimizer import Adam
from pynnet.loss import mse, mse_derivative

# Compile the model
model.compile(
    loss=mse,
    loss_derivative=mse_derivative,
    optimizer=Adam(learning_rate=0.01)
)

# Train the model
model.fit(x_train, y_train, epochs=1000, verbose=True)

🎓 Learning by Example

We provide several example implementations to help you get started:

1. XOR Gate (examples/01_xor_example.py)

Learn how to create your first neural network by implementing the XOR logic gate. This is a perfect starting point for beginners!

# XOR Truth Table:
# Input  Output
# 0 0 => 0
# 0 1 => 1
# 1 0 => 1
# 1 1 => 0

2. Binary Classification (examples/02_binary_classification.py)

Learn how to classify points into two categories using a simple dataset of concentric circles. Great for understanding:

• Binary classification problems
• Using multiple layers
• Working with 2D input data

3. Regression (examples/03_regression.py)

Learn how to predict continuous values by fitting a sine wave. Demonstrates:

• Regression problems
• Using different activation functions
• Handling continuous output
• Data visualization

🛠 Features

Network Types

• Sequential: Build networks by stacking layers one after another

Layers

• Dense: Fully connected layer with customizable features:
  • Weight Initialization:
    • 'he': Best for ReLU activation (default)
    • 'xavier': Best for tanh/sigmoid
    • 'lecun': For normalized inputs
    • 'identity': For deep networks
    • 'orthogonal': For better training
    • 'random': Simple random initialization
  • Bias Initialization:
    • 'zeros': All zeros (default)
    • 'ones': All ones
    • 'random': Random values
    • 'constant': Custom value

Activation Functions

• relu: Rectified Linear Unit
• sigmoid: Sigmoid function (0 to 1)
• tanh: Hyperbolic tangent (-1 to 1)
• linear: No transformation (for regression)

Optimizers

• SGD: Stochastic Gradient Descent
• Adam: Adaptive Moment Estimation

Loss Functions

• mse: Mean Squared Error (for regression)
• binary_cross_entropy: For binary classification

💡 Tips for Success

Choosing Layer Sizes

• Input Layer: Must match your data's feature count
• Hidden Layers: Generally start with powers of 2 (e.g., 32, 64, 128)
• Output Layer:
  • Binary classification: 1 unit with sigmoid
  • Regression: 1 unit with linear activation
  • Multi-class: One unit per class with softmax

Picking Initialization Methods

• With ReLU: Use 'he' initialization
• With Sigmoid/Tanh: Use 'xavier' initialization
• Deep Networks: Try 'orthogonal' initialization
• When in Doubt: Start with 'he' initialization

Training Tips

1. Start Small: Begin with a simple network and gradually add complexity
2. Monitor Loss: Use verbose=True to watch training progress
3. Learning Rate:
   • Start with 0.01
   • If loss is unstable: decrease it
   • If learning is slow: increase it
4. Save Your Models: Use model.save_weights() to save progress

Common Patterns

Binary Classification:

model = Sequential()
model.add(Dense(input_size=n_features, output_size=64, weight_init='he'))
model.add(relu)
model.add(Dense(input_size=64, output_size=1, weight_init='xavier'))
model.add(sigmoid)

Regression:

model = Sequential()
model.add(dense(input_size=n_features, output_size=64, weight_init='he'))
model.add(relu)
model.add(dense(input_size=64, output_size=1, weight_init='he'))
model.add(linear)

🚀 Advanced Features

Model Persistence

Save and load your trained models:

# Save model weights
model.save_weights('my_model.npz')

# Load model weights
model.load_weights('my_model.npz')

Custom Training Loop

Monitor and control the training process:

for epoch in range(n_epochs):
    loss = model.fit(X, y, epochs=1, verbose=False)
    if epoch % 100 == 0:
        predictions = model.predict(X_test)
        print(f"Epoch {epoch}: Loss = {loss:.4f}")

📂 Project Structure

pynnet/
├── layers/         # Neural network layer implementations
│   ├── base.py    # Base layer class
│   └── dense.py   # Dense layer implementation
├── activation.py   # Activation functions
├── loss.py        # Loss functions
├── network.py     # Core neural network implementation
├── optimizer.py   # Optimization algorithms
└── test.py        # Unit tests

🤝 Getting Help

• Check the examples in the examples/ directory
• Review the docstrings in the code
• Create an issue on GitHub
• Send an email to the author

Requirements

• Python 3.7 or higher
• NumPy >= 1.20.0

Contributing

Contributions are welcome! Please feel free to submit a Pull Request.

License

This project is licensed under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International License - see the LICENSE file for details.

Author

• Zain Qamar - GitHub
• Email: [e-mail]

Acknowledgments

• Thanks to all contributors who help improve this library
• Special thanks to the NumPy community for providing the foundation for numerical computations

Citation

If you use PyNNet in your research, please cite it as:

@software{pynnet2025,
  author = {Qamar, Zain},
  title = {PyNNet: A Pythonic Neural Network Library},
  year = {2025},
  publisher = {GitHub},
  url = {https://github.com/prime-programmer-ar/pynnet_project.git}
}