Neural Network from Scratch

A simple, educational implementation of a feed-forward neural network in pure Python.
This project demonstrates core concepts—forward propagation, backpropagation, activation functions.

Features

• Build multi-layer feed-forward networks (input, hidden, output layers)
• Support for common activation functions:
  • Sigmoid
  • ReLU
  • Tanh
  • Softmax
• Manual forward propagation and backpropagation
• Multi-threaded neuron computation for acceleration

Requirements

• Python 3.7+
• NumPy
• tqdm
• Numba (optional, for JIT compilation speedups)

Installation

git clone https://github.com/B-R-P/Neural-Network.git
cd Neural-Network
pip install numpy tqdm numba

Usage

Import the main module:

from NN import Network, save_model, regain_model, check_err

Creating a Network

# Create a network with:
# - 4 inputs
# - two hidden layers with 8 and 5 neurons respectively (ReLU)
# - 3 outputs (Softmax)
net = Network(
    input_size=4,
    hidden_layers=[8, 5],
    output_size=3,
    hidden_activation='reLU',
    output_activation='softmax'
)

Training

# Single-sample training
x = [0.5, 0.2, 0.1, 0.7]
y_true = [0, 0, 1]  # One-hot target for class 2
net.train(x, y_true, learning_rate=0.1)

# Batch training over multiple epochs and samples
data = [([0.1,0.2,0.3,0.4], [1,0,0]), ...]
net.traina(data, iterations=1000, initial_lr=0.1)

Prediction

x_new = [0.6, 0.1, 0.8, 0.3]
pred_class, pred_confidence = net.predict(x_new)
print(f"Predicted class: {pred_class}, Confidence: {pred_confidence:.4f}")

Saving & Loading Models

# Save the trained model to ZIP
save_model(net, 'mymodel.zip')

# Later, load it back
loaded_net = regain_model('mymodel.zip')

API Reference

Class: Network

• Network(input_size, hidden_layers, output_size, hidden_activation, output_activation)
  Construct a new network.

• train(x, y, learning_rate)
  Train on a single input–target pair.

• traina(data, iterations, initial_lr)
  Train over multiple samples and iterations. data is a list of (x, y).

• predict(x) → (class_index, confidence)
  Run inference and return the predicted class.

• printer() / wprinter()
  Print network outputs or weights/biases.

• check_err(data) → float
  Compute total error over a dataset.

Functions

• save_model(net, filename)
• regain_model(filename)

Activation functions

• sigmoid, reLU, tanh, softmax (found in actdic)

Contributing

Contributions, issues, and feature requests are welcome!
Please open a GitHub issue or submit a pull request.

License

This project is licensed under the MIT License.
See LICENSE for details.