Micro-Grad Engine

An implementation of Andrej Karpathy's micrograd from scratch. This is a minimal neural network and autodiff engine built entirely in Python.

Project Structure

.
├── engine.py # Core autodiff engine (Value class)
├── model.py # Example usage of training loop
├── neural_network.py # Neuron, Layer, and MLP classes
├── README.md
└── requirements.txt # Python dependencies

Features

• Autograd Engine
  • Supports addition, multiplication, power, tanh, ReLu, and more
  • Reverse mode automatic differentiation (backpropagation)
• Neural Network Module
  • Fully connected layers (Layer)
  • Multi-layer perceptrons (MLP)
• Training Loop
  • Vanilla stochastic gradient descent
  • Uses mean squared error loss

Installation

1. Clone the repository:

git clone https://github.com/eko-071/micrograd-engine
cd micrograd-engine

2. Install dependencies:

pip install -r requirements.txt

Requirements

• numpy

This is a pure Python project; no heavy ML libraries are required.