micrograd

This is a small C implementation of Andrej Karpathy's micrograd, based on his spelled-out intro to neural networks and backpropagation video.

It is a learning project, not a production autodiff library.

Build

cmake --preset debug
cmake --build --preset debug
ctest --preset debug
./build/debug/micrograd_basic

Documentation

cmake --preset docs
cmake --build --preset docs
open build/docs/html/index.html

The documentation target requires Doxygen. HTML output uses the vendored Doxygen Awesome CSS theme in docs/doxygen-awesome.

Example

mg_graph* g = mg_graph_new();
 
mg_value* a = mg_scalar(g, 2.0f);
mg_value* b = mg_scalar(g, 3.0f);
mg_value* c = mg_mul(g, a, b);
 
mg_backward(g, c);
 
printf("%f %f\n", mg_data(c), mg_grad(a));
 
mg_graph_free(g);

See examples/basic.c for a small MLP training example.

Notes

mg_graph owns all mg_values. Neurons, layers, and MLPs own their arrays, but not scalar values; free them before freeing the graph.

mg_backward resets gradients before backprop. Hidden MLP layers use tanh; the output layer is linear. mg_pow supports exponent gradients for positive bases.