autograd

A library to run the computation graphs, whose backend is rust-ndarray.

Overview

• (Higher order) automatic differentiation
• Neural net first APIs
• Pure Rust
• Dynamic/static graph construction with shared variables

Example

extern crate autograd as ag;

let ref x = ag::placeholder(&[1]);
let ref y = ag::variable(ag::init::zeros(&[1]));
// `z` is a target of partial differentiation.
let ref z = 2*x*x + 3*y + 1;
// dz/dy
let ref g1 = ag::gradients(z, &[y], None)[0];
// dz/dx (necessary to fill the placeholder `x`)
let ref g2 = ag::gradients(z, &[x], None)[0];
// ddz/dx (second order derivative)
let ref gg = ag::gradients(g2, &[x], None)[0];

// evaluation of symbolic gradients
assert_eq!(3., g1.eval()[0]);
let feed_dict = ag::Input::new().add(x, ag::init::from_scalar(2.));
assert_eq!(8., g2.eval_with_input(feed_dict)[0]);
assert_eq!(4., gg.eval()[0]);

For more, see examples or tests.

Documentation

WIP

License

MIT