7 releases

0.313.4 Dec 10, 2022
0.313.3 Dec 10, 2022
0.313.0 Nov 25, 2022
0.3.13 Nov 25, 2022
0.0.1 Oct 21, 2018

#328 in Machine learning

24 downloads per month

AGPL-3.0-or-later

17KB
321 lines

epsilon - Fast autograd using dual numbers

Dual numbers are a straightforward way of doing forward gradient propagation, i.e. keep track of derivatives for all expressions, to automatically differentiate a function without storing a computation graph.

Using dual numbers, one can augment their numbers with a "dual" part, representing the derivative of the term with respect to some input variable. The input variable has a unit dual part of 1, and each resulting expression has a dual part with the derivaite up to that point.

This can be trivially extended to multiple variables by storing one dual part per input variable.

One can find a more in-depth at Wikipedia


This crate statically generates code for dual numbers using macros, meaning one can provide names for each dependent variable, and corresponding methods will be generated with names reflecting the name of the variable.

The interface exposed by the types generated by this crate is very similar to that of the standard numerical rust types, meaning most code using f64 should be very straightforward to convert to using dual numbers.

Example usage:

use epsilon::make_dual;
// We want to compute dz/dx and dz/dy for z = x^2+y*sin(y) at x=5, y=7
make_dual! { MyDual, x, y } // Create a dual number with terms `x` and `y`

let (x, y) = (MyDual::x(5.), MyDual::y(7.)); // Perform the calculations, and compute the derivative at x=5, y=7

let z = x.pow(2.) + y * y.sin();

let dzdx = z.d_dx();
let dzdy = z.d_dy();

assert_eq!(dzdx, 10.); // 2 * x
assert_eq!(dzdy, 5.934302379121921); // y*cos(y) + sin(y)

Dependencies