micrograd++

A lightweight C++ neural network library inspired by Andrej Karpathy's micrograd. This implementation provides a minimalistic yet powerful neural network framework with automatic differentiation capabilities.

Features

• Automatic differentiation engine
• Neural network primitives:
  • Neurons with configurable inputs
  • Layers with multiple neurons
  • Multi-layer perceptron (MLP) architecture
• Modern C++ implementation (C++17)
• Header-only library
• Efficient memory management using smart pointers

Getting Started

Prerequisites

• C++17 compatible compiler
• CMake (for building)

Building the Project

mkdir build
cd build
cmake ..
make

Usage Example

#include "core/Neuron.h"

// Create a simple neural network
size_t inputs = 3;
std::vector<size_t> layer_sizes = {4, 2, 1};  // 4 neurons in first layer, 2 in second, 1 in output
MultiLayerPerceptron mlp(inputs, layer_sizes);

// Create input values
std::vector<ValuePtr> input_values = {
    create_value(1.0),
    create_value(0.5),
    create_value(-1.0)
};

// Forward pass
auto output = mlp(input_values);
// Backpropagate
output[0]->backpropagate();
// Visualize 
output[0]->visualize("neural_network");

Architecture

Core Components

1. Value Class

   • Represents a value in the computation graph
   • Handles automatic differentiation

2. Neuron Class

   • Basic computational unit
   • Contains weights and bias
   • Applies ReLU activation function

3. Layer Class

   • Collection of neurons
   • Handles forward propagation through neurons

4. MultiLayerPerceptron Class

   • Complete neural network
   • Manages multiple layers
   • Provides forward propagation through the entire network

Contributing

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

License

This project is licensed under the MIT License - see the LICENSE file for details.

Acknowledgments

• Inspired by Andrej Karpathy's micrograd
• Built with modern C++ features for efficiency and ease of use

TODO

• Add backward propagation examples
• Implement additional activation functions
• Add serialization support
• Include more comprehensive testing
• Add benchmarking suite
• Improve documentation with more examples
• Optimize + Add more features to the NN