#automatic-differentiation #reverse #mode #function

macro reverse_differentiable

Automatic differentiation of functions

3 releases

0.1.2 Aug 2, 2021
0.1.1 Aug 2, 2021
0.1.0 Aug 2, 2021

#39 in #automatic-differentiation

MIT/Apache

7KB
81 lines

reverse

Crates.io Documentation License

Reverse mode automatic differentiation in Rust.

To use this in your crate, add the following to Cargo.toml:

[dependencies]
reverse = "0.1"

Examples

use reverse::*;

fn main() {
  let graph = Graph::new();
  let a = graph.add_var(2.5);
  let b = graph.add_var(14.);
  let c = (a.sin().powi(2) + b.ln() * 3.) - 5.;
  let gradients = c.grad();

  assert_eq!(gradients.wrt(&a), (2. * 2.5).sin());
  assert_eq!(gradients.wrt(&b), 3. / 14.);
}

Differentiable Functions

There is an optional diff feature that activates a macro to transform functions to the right type so that they are differentiable. That is, functions that act on f64s can be used on differentiable variables without change, and without needing to specify the (not simple) correct type.

To use this, add the following to Cargo.toml:

reverse = { version = "0.1", features = ["diff"] }

Functions must have the type Fn(&[f64], &[&[f64]]) -> f64, where the first argument contains the differentiable parameters and the second argument contains arbitrary arrays of data.

Example

Here is an example of what the feature allows you to do:

use reverse::*;

fn main() {
    let graph = Graph::new();
    let a = graph.add_var(5.);
    let b = graph.add_var(2.);

    // you can track gradients through the function as usual!
    let res = addmul(&[a, b], &[&[4.]]);
    let grad = res.grad();

    assert_eq!(grad.wrt(&a), 1.);
    assert_eq!(grad.wrt(&b), 4.);
}

// function must have these argument types but can be arbitrarily complex
#[differentiable]
fn addmul(params: &[f64], data: &[&[f64]]) -> f64 {
    params[0] + data[0][0] * params[1]
}

Dependencies

~1.5MB
~35K SLoC