2 releases
0.1.1 | Mar 15, 2023 |
---|---|
0.1.0 | Mar 7, 2023 |
#716 in Machine learning
2MB
280 lines
rustygrad
Inspired by Andrej Karpathy's micrograd. This repo implements a tiny Autograd engine in Rust:
- with a friendly API
- an easy-to-understand implementation
- in a minimal amount of code
The engine and the neural net are implemented in about 150 and 100 lines of code respectively (vs Andrej's 100 and 50)! About twice as long, but twice as fast!
Example usage
Value API
Value
can be used to construct arbitrary DAG Neural Nets (directed acyclic graph). One way to think about it, is it can be used to model common mathematical expressions. For example,
$$ g = [(a + b) * (c + d)] ^ 2 $$
can be translated into a neural net as such:
use rustygrad::Value;
fn main() {
let a = Value::from(1.0);
let b = Value::from(2.0);
let c = Value::from(3.0);
let d = Value::from(4.0);
let g = ((&a + &b) * (&c + &d)).pow(2.0);
g.backwards(); // compute gradients
}
Below is more complicated example (from micrograd) designed to show most of the supported Value operations // and their Python micrograd version
fn main() {
// a = Value(-4.0)
// b = Value(2.0)
let a = Value::from(-4.0);
let b = Value::from(2.0);
// c = a + b
// d = a * b + b**3
let mut c = &a + &b;
let mut d = &a * &b + &b.pow(3.0);
// c += c + 1
// c += 1 + c + (-a)
// d += d * 2 + (b + a).relu()
// d += 3 * d + (b - a).relu()
c += &c + 1.0;
c += 1.0 + &c + (-&a);
d += &d * 2.0 + (&b + &a).relu();
d += 3.0 * &d + (&b - &a).relu();
// e = c - d
// f = e**2
// g = f / 2.0
// g += 10.0 / f
let e = &c - &d;
let f = e.pow(2.0);
let mut g = &f / 2.0;
g += 10.0 / &f;
// print(f'{g.data:.4f}') # prints 24.7041, the outcome of this forward pass
println!("{:.4}", g.borrow().data); // 24.7041
// g.backward()
// print(f'{a.grad:.4f}') # prints 138.8338, i.e. the numerical value of dg/da
// print(f'{b.grad:.4f}') # prints 645.5773, i.e. the numerical value of dg/db
g.backward();
println!("{:.4}", a.borrow().grad); // 138.8338
println!("{:.4}", b.borrow().grad); // 645.5773
}
cargo run --example engine
Neuron and MLP API
The library also exposes a Neuron
and Multilayer Perceptron MLP
use rustygrad::{Neuron, MLP};
fn main() {
// Create a Neuron
// With input size of 2
// With ReLu layer (true)
let neuron = Neuron::new(2, true);
// Output node
let g = &neuron.forward(&vec![Value::from(7.0)]);
// Create a 2x2x1 MLP net:
// Input layer of size 2
// Hidden layer of size 2
// Ouput layer of size 1
let model = MLP::new(2, vec![2, 1]);
// Some input vector of size 2
let x = vec![Value::from(7.0), Value::from(8.0)];
// Output Value node
let g = &model.forward(x)[0];
}
cargo run --example graphviz
Neuron:
MLP:
Training a neural net
The file mlp.rs
trains a MLP binary classifier (with 2 16-node hidden layers) on a toy make_moons.csv
dataset. Since plots in rust are hard, for now, here is an ascii representation of the learned solution space:
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . * * * * * . . . . . . . . . . . . . . . . *
. . . . . . . . . . . . . . . . . * * * * * * * . . . . . . . . . . . . . . * *
. . . . . . . . . . . . . . . . . * * * * * * * * . . . . . . . . . . . . * * *
. . . . . . . . . . . . . . . . . * * * * * * * * * . . . . . . . . . . * * * *
. . . . . . . . . . . . . . . . . * * * * * * * * * * . . . . . . . . . * * * *
. . . . . . . . . . . . . . . . * * * * * * * * * * * . . . . . . . . * * * * *
. . . . . . . . . . . . . . . . * * * * * * * * * * * * . . . . . . * * * * * *
. . . . . . . . . . . . . . . . * * * * * * * * * * * * * . . . . * * * * * * *
. . . . . . . . . . . . . . . * * * * * * * * * * * * * * * * * * * * * * * * *
. . . . . . . . . . . . . . . * * * * * * * * * * * * * * * * * * * * * * * * *
. . . . . . . . . . . . . . * * * * * * * * * * * * * * * * * * * * * * * * * *
. . . . . . . . . . . . . . * * * * * * * * * * * * * * * * * * * * * * * * * *
. . . . . . . . . . . . . . * * * * * * * * * * * * * * * * * * * * * * * * * *
. . . . . . . . . . . . . * * * * * * * * * * * * * * * * * * * * * * * * * * *
. . . . . . . . . . . . . * * * * * * * * * * * * * * * * * * * * * * * * * * *
. . . . . . . . . . . . * * * * * * * * * * * * * * * * * * * * * * * * * * * *
. . . . . . . . . . . . * * * * * * * * * * * * * * * * * * * * * * * * * * * *
. . . . . . . . . . . * * * * * * * * * * * * * * * * * * * * * * * * * * * * *
. . . . . . . . . . . * * * * * * * * * * * * * * * * * * * * * * * * * * * * *
. . . . . . . . . . . * * * * * * * * * * * * * * * * * * * * * * * * * * * * *
. . . . . . . . . . * * * * * * * * * * * * * * * * * * * * * * * * * * * * * *
. . . . . . . . . . * * * * * * * * * * * * * * * * * * * * * * * * * * * * * *
. . . . . . . . . * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * *
. . . . . . . . . * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * *
. . . . . . . . * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * *
. . . . . . . . * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * * *
cargo run --example mlp
Running tests
cargo test