drift

A rudimentary implementation of automatic differentiation written in Rust. Currently supports forward mode automatic differentiation using Operator Overloading. Currently supports functions of the form R^n -> R^m. Can compute individual partial derivatives via differential, gradient via gradient. derivative can be used when dealing with R -> R functions.

The idea is to use a custom float type, F64 which implements the num::Float trait and the usual binary operations (+,-,*,/). Then derivatives can be computed next to function outputs.

Example

    fn simple_arr(a: ndarray::Array1<F64>) -> F64 {
        // x^3 * y + 5x.
        // dx is 3x^2y + 5; dy is x^3. - at (5,2) dx is 155, 125
        a[0].powi(3) * a[1] + F64::constant(5.0) * a[0]
    }
    assert_eq!(
        differential(simple_arr, &ndarray::array![5.0, 2.0], 0).dx,
        155.0
    )

Todos

1. Reverse mode support. This is useful because I'm currently using drift in mmasque/translator where I'm implementing some basic neural networks. Backpropagation there ends up being a special case of reverse mode autodiff so it will likely simplify the code.

References

The autodiff crate was used to cross-check definitions and API setup.