2 unstable releases

0.2.0 Dec 29, 2022
0.1.0 Sep 19, 2022

#37 in #automatic-differentiation

31 downloads per month
Used in aegir

Custom license

11KB
124 lines

aegir

Crates.io Build Status Coverage Status

Overview

Strongly-typed, compile-time autodifferentiation in Rust.

aegir is an experimental autodifferentiation framework designed to leverage the powerful type-system in Rust and avoid runtime as much as humanly possible. The approach taken resembles that of expression templates, as commonly used in linear-algebra libraries written in C++.

Key Features

  • Built-in arithmetic, linear-algebraic, trigonometric and special operators.
  • Infinitely differentiable: Jacobian, Hessian, etc...
  • Custom DSL for operator expansion.
  • Decoupled/generic tensor type.

Installation

[dependencies]
aegir = "1.0"

Example

#[macro_use]
extern crate aegir;
extern crate rand;

use aegir::{Differentiable, Function, Identifier, Node, ids::{X, Y, W}};

db!(Database { x: X, y: Y, w: W });

fn main() {
    let mut weights = [0.0, 0.0];

    let x = X.into_var();
    let y = Y.into_var();
    let w = W.into_var();

    let model = x.dot(w);

    // Using standard method calls...
    let sse = model.sub(y).squared();
    let adj = sse.adjoint(W);

    // ...or using aegir! macro
    let sse = aegir!((model - y) ^ 2);
    let adj = sse.adjoint(W);

    for _ in 0..100000 {
        let [x1, x2] = [rand::random::<f64>(), rand::random::<f64>()];

        let g = adj.evaluate(Database {
            // Independent variables:
            x: [x1, x2],

            // Dependent variable:
            y: x1 * 2.0 - x2 * 4.0,

            // Model weights:
            w: &weights,
        }).unwrap();

        weights[0] -= 0.01 * g[0][0];
        weights[1] -= 0.01 * g[0][1];
    }

    println!("{:?}", weights.to_vec());
}

Dependencies

~1.5MB
~33K SLoC