dxpr

Differentiable expression templates in Rust.

Examples

At runtime:

use dxpr::{Eval, var};
let x = 4;
let a = var(&x);
let expression = -a;
let value = expression.eval();
assert_eq!(-4, value);

At compile time:

#![feature(const_trait_impl)]
use dxpr::{ops, Eval, Expr, var};
const X: i32 = 4;
const A: Expr<&i32> = var(&X);
const EXPRESSION: Expr<ops::Neg<Expr<&i32>>> = -A;
const VALUE: i32 = EXPRESSION.eval();
assert_eq!(-4, VALUE);

Automatic differentiation:

use dxpr::{var, Eval, Grad};
let a = 4;
let b = 4;
let ga = (var(&a) - var(&b)).grad(&a);
let gb = (var(&a) - var(&b)).grad(&b);
assert_eq!(1, ga.eval());
assert_eq!(-1, gb.eval());

We can reuse an expression without copying (e.g. for machine learning) by calling eval on a reference:

use dxpr::{ops, Eval, EvalRef, Expr, var};
let x = 4;
let expression = -var(&x);
assert_eq!(-4, (&expression).eval());
assert_eq!(-4, (&expression).eval());
assert_eq!(-4, (&expression).eval());
// still movable the last time:
assert_eq!(-4, expression.eval());
// can't use it again:    ------ `expression` moved due to this method call
// /* bad! */ assert_eq!(-4, expression.eval());
//                           ^^^^^^^^^^ value used here after move

To do:

• Differentiation on arbitrary expressions
• Index and RangeBound operators (i.e. a[b] and a..b)