autograd

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

Documentation: https://docs.rs/autograd/

Overview

• Automatic differentiation
• Pure Rust
• Neural net first APIs
• Dynamic/static graph construction with shared variables

Examples

Here we are computing partial derivatives of z = 2x^2 + 3y + 1.

extern crate ndarray;
extern crate autograd as ag;

let mut graph = ag::Graph::new();
let ref x = graph.placeholder();
let ref y = graph.variable(ndarray::arr1(&[0]));
let ref z = 2*x*x + 3*y + 1;

// dz/dy
let ref g1 = ag::gradients(&[z], &[y], None)[0];

// dz/dx
let ref g2 = ag::gradients(&[z], &[x], None)[0];

// ddz/dx (differentiates `z` again)
let ref gg = ag::gradients(&[g2], &[x], None)[0];

// evaluation of symbolic gradients
assert_eq!(3., graph.eval(&[g1])[0][0]);
assert_eq!(4., graph.eval(&[gg])[0][0]);

// dz/dx requires to fill the placeholder `x`
graph.feed(x, ndarray::arr1(&[2.]));
assert_eq!(8., graph.eval(g2)[0][0]);

Another example: multi layer perceptron for MNIST.

// -- graph def --
let mut graph = ag::Graph::new();

let ref x = graph.placeholder();
let ref y = graph.placeholder();
let ref w = graph.variable(ag::ndarray_ext::glorot_uniform(&[28 * 28, 10]));
let ref b = graph.variable(ag::ndarray_ext::zeros(&[1, 10]));

let ref z = ag::matmul(x, w) + b;
let ref loss = ag::sparse_softmax_cross_entropy(z, y);
let ref grads = ag::gradients(loss, &[w, b], None);
let ref predictions = ag::argmax(z, -1, true);
let ref accuracy = ag::reduce_mean(&ag::equals(predictions, y), 0, false);

// -- dataset --
let ((x_train, y_train), (x_test, y_test)) = dataset::load();

// -- training method --
let mut optimizer = ag::sgd::optimizers::Adam { ..Default::default() };

// -- training loop --
for epoch in 0..max_epoch {
    ...
}

Available operations in rust-autograd are listed here

For more, see examples or tests.

License

MIT