Minimal Autograd Engine in Rust.
Based on https://github.com/karpathy/micrograd
But in Rust.
Simple:
use rustygrad::v;
let x = v(0.0);
let mut y = (x.clone() * - 3.0 - 3.0).pow(2.0).relu();
y.backward();
// y = relu((-3 * x - 3)^2)
// y = relu(9x^2 + 18x +9)
// dy/dx = drelu(18x + 18)
// dy/dx(0) = 18
assert_eq!(x.grad(), 18.0);Train a classifier:
use rustygrad::{v, Mlp, Module};
fn main() {
let mut mlp = Mlp::new(2, &[8, 8, 1]);
println!("{} - {} parameters", mlp, mlp.parameters().len());
let xs = [[1.0, 1.0], [1.0, -1.0], [-1.0, 1.0], [-1.0, -1.0]];
let ys = [-1.0, 1.0, 1.0, -1.0];
let step_size = 0.1;
for step in 0..100 {
// Always zero_grad.
mlp.zero_grad();
// Reset loss to zero.
let mut loss = v(0.0);
// Run over all examples.
for (x, y) in xs.iter().zip(ys) {
let x = x.map(v);
let pred = mlp.call(&x).pop().unwrap();
loss = loss + (pred - y).pow(2.0);
}
// Compute mean loss.
loss = loss / xs.len() as f64;
// Compute gradients.
loss.backward();
// Apply gradient step.
for mut p in mlp.parameters() {
p.set_data(p.data() + p.grad() * -step_size);
}
// Log progress.
if step % 10 == 0 {
println!("step {} - loss = {:0.4} step_size = {}", step, loss.data(), step_size);
}
}
}