micrograd-rust 0.1.0

A Rust-backed scalar autograd engine. Drop-in replacement for Karpathy's micrograd, but 13-75x faster.

micrograd-rust

A Rust-backed scalar autograd engine and neural network library for Python. Drop-in replacement for Karpathy's micrograd, but 13-75x faster.

Install

pip install micrograd-rust

Usage

from micrograd_rust import Value, MLP

# scalar autograd
a = Value(2.0)
b = Value(3.0)
c = (a * b + a).tanh()
c.backward()
print(a.grad)  # dc/da
print(b.grad)  # dc/db

# neural network
mlp = MLP(3, [4, 4, 1])
x = [Value(1.0), Value(2.0), Value(3.0)]
out = mlp(x)
out[0].backward()

API

Value(data, label="")

• .data — the scalar value
• .grad — the gradient (populated after .backward())
• Supports +, *, **, -, /, .tanh(), .exp(), .backward()

MLP(nin, nouts)

• mlp(x) / mlp.forward(x) — forward pass, returns list of Value
• mlp.parameters() — list of all Value parameters
• mlp.zero_grad() — reset all gradients to 0

How it works

Uses PyO3 to call into a Rust autograd engine under the hood. All computation happens in Rust. See the GitHub repo for benchmarks and source.