1 unstable release
0.2.0 | Jun 25, 2022 |
---|
#647 in Machine learning
59KB
1K
SLoC
ES-Optimizer [esopt]
General Evolution-Strategy-Optimizer implementation according to https://arxiv.org/abs/1703.03864 in Rust.
Usage
There are examples in the examples folder and below is the simplest one:
Simple example (see examples/simple.rs):
extern crate esopt;
use esopt::*;
fn main()
{
//create required evaluator
let eval = ExampleEval { target: 25.0 };
//create Evolution-Strategy-Optimizer
let mut es = ES::new_with_sgd(eval, 0.75, 0.0, 0.0); //using evaluator, lr, beta(momentum), lambda(weight decay)
es.set_params(vec![0.0]) //initial parameters (important to specify the problem dimension, default is vec![0.0])
.set_std(50.0) //parameter noise standard deviation to approximate the gradient
.set_samples(50); //number of mirrored samples to use to approximate the gradient
//track the optimizer's results
for _ in 0..5
{
let res = es.optimize(2); //optimize for n steps
println!("(Score, Gradnorm): {:?}", res);
println!("Params: {:?}", es.get_params());
}
}
//carrier object for evaluator, which can include training data or target information
#[derive(Clone)]
struct ExampleEval
{
target:Float,
}
//implement Evaluator trait to allow usage in the optimizer
impl Evaluator for ExampleEval
{
//compute the negative absolute error (maximize to get close to target)
fn eval_test(&self, params:&[Float]) -> Float
{
let error = self.target - params[0];
-error.abs()
}
fn eval_train(&self, params:&[Float], _:usize) -> Float
{
self.eval_test(params)
}
}
Dependencies
~2.8–4MB
~80K SLoC