lfa 0.15.0

Native rust implementations of linear function approximators.
Documentation 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61

use super::*;
use ndarray::Array1;

const EPS: f64 = 1e-7;

#[derive(Clone, Debug, PartialEq)]
#[cfg_attr(
    feature = "serde",
    derive(Serialize, Deserialize),
    serde(crate = "serde_crate")
)]
pub struct Adagrad {
    learning_rate: f64,
    accumulator: Array1<f64>,
}

impl Adagrad {
    pub fn new(n_params: usize, learning_rate: f64) -> Self {
        Adagrad {
            learning_rate,
            accumulator: Array1::zeros(n_params),
        }
    }
}

impl Optimiser<Features> for Adagrad {
    fn step_scaled(
        &mut self,
        weights: &mut ArrayViewMut1<f64>,
        features: &Features,
        scale_factor: f64,
    ) -> Result<()>
    {
        let lr = self.learning_rate;

        match features {
            Features::Dense(da) => da
                .iter()
                .zip(weights.iter_mut())
                .zip(self.accumulator.iter_mut())
                .for_each(|((a, w), ss)| {
                    let g = a * scale_factor;

                    *ss = *ss + g.powi(2);
                    *w = *w + lr * g / (ss.sqrt() + EPS);
                }),
            Features::Sparse(sa) => {
                for (i, ss) in self.accumulator.indexed_iter_mut() {
                    let g = sa.activations.get(&i).cloned().unwrap_or(0.0) * scale_factor;

                    *ss = *ss + g.powi(2);
                    weights[i] += lr * g / (ss.sqrt() + EPS);
                }
            },
        }

        Ok(())
    }

    fn reset(&mut self) { self.accumulator.fill(0.0); }
}