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
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106

use super::*;
use ndarray::Array1;
use std::ops::MulAssign;

const EPS: f64 = 1e-7;

/// Adaptive moment estimation gradient descent.
///
/// https://arxiv.org/pdf/1412.6980.pdf
#[derive(Clone, Debug, PartialEq)]
#[cfg_attr(
    feature = "serde",
    derive(Serialize, Deserialize),
    serde(crate = "serde_crate")
)]
pub struct Adam {
    beta1: f64,
    beta1_prod: f64,

    beta2: f64,
    beta2_prod: f64,

    learning_rate: f64,

    exp_avg: Array1<f64>,
    exp_avg_sq: Array1<f64>,
}

impl Adam {
    pub fn new(n_params: usize, learning_rate: f64, beta1: f64, beta2: f64) -> Self {
        Adam {
            beta1,
            beta1_prod: beta1,

            beta2,
            beta2_prod: beta2,

            learning_rate,

            exp_avg: Array1::zeros(n_params),
            exp_avg_sq: Array1::zeros(n_params),
        }
    }
}

impl Optimiser<Features> for Adam {
    fn step_scaled(
        &mut self,
        weights: &mut ArrayViewMut1<f64>,
        features: &Features,
        scale_factor: f64,
    ) -> Result<()>
    {
        self.beta1_prod *= self.beta1;
        self.beta2_prod *= self.beta2;

        match features {
            Features::Dense(da) => {
                let m1 = self.exp_avg.as_slice_memory_order_mut().unwrap();
                let m2 = self.exp_avg_sq.as_slice_memory_order_mut().unwrap();

                for (i, a) in da.indexed_iter() {
                    let g = a * scale_factor;

                    let m1_new = self.beta1 * m1[i] + (1.0 - self.beta1) * g;
                    let m2_new = self.beta2 * m2[i] + (1.0 - self.beta2) * g * g;

                    let m1_unbiased = m1_new / (1.0 - self.beta1_prod);
                    let m2_unbiased = m2_new / (1.0 - self.beta2_prod);

                    m1[i] = m1_new;
                    m2[i] = m2_new;
                    weights[i] += self.learning_rate * m1_unbiased / (m2_unbiased.sqrt() + EPS);
                }
            },
            Features::Sparse(sa) => {
                self.exp_avg.mul_assign(self.beta1);
                self.exp_avg_sq.mul_assign(self.beta2);

                let m1 = self.exp_avg.as_slice_memory_order_mut().unwrap();
                let m2 = self.exp_avg_sq.as_slice_memory_order_mut().unwrap();

                for (&i, a) in sa.iter() {
                    let g = a * scale_factor;

                    let m1_new = m1[i] + (1.0 - self.beta1) * g;
                    let m2_new = m2[i] + (1.0 - self.beta2) * g * g;

                    let m1_unbiased = m1_new / (1.0 - self.beta1_prod);
                    let m2_unbiased = m2_new / (1.0 - self.beta2_prod);

                    m1[i] = m1_new;
                    m2[i] = m2_new;
                    weights[i] += self.learning_rate * m1_unbiased / (m2_unbiased.sqrt() + EPS);
                }
            },
        }

        Ok(())
    }

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