Micrograd in Rust

Reimplementing Micrograd in Rust. Based on Andrej Karpathy's youtube series

This is an attempt to implement Micrograd in Rust while preserving a friendly looking API. This ships with WebAssemlby bindings and a web based playground allowing to step through forward and back prop passes.

Rust API preview

// define an expression: a * b + c * f
let mut expression = (value!(2.0, "a") * value!(-3.0, "b") + value!(10.0, "c")) * value!(-2.0, "f");
// backprop, compute gradients
expression.backward();

// define a basic 2-layer MLP
let mut net = MLP {
    layers: vec![Layer::new(1, 1), Layer::new(1, 1)],
    result: None,
};

// forward pass on the MLP
net.forward(vec![value!(2.0, "x1")]);

// compute loss based on the expected output value
let mut loss = net.loss(vec![0.5]);

// backprop
loss.backward();

Web preview

Check playground example here