forust-ml 0.5.0

A lightweight gradient boosting implementation in Rust.
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
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194

use crate::{data::FloatData, metric::Metric};
use serde::{Deserialize, Serialize};

type ObjFn = fn(&[f64], &[f64], &[f64]) -> (Vec<f32>, Vec<f32>);

#[derive(Debug, Deserialize, Serialize)]
pub enum ObjectiveType {
    LogLoss,
    SquaredLoss,
}

pub fn gradient_hessian_callables(objective_type: &ObjectiveType) -> ObjFn {
    match objective_type {
        ObjectiveType::LogLoss => LogLoss::calc_grad_hess,
        ObjectiveType::SquaredLoss => SquaredLoss::calc_grad_hess,
    }
}

pub fn calc_init_callables(objective_type: &ObjectiveType) -> fn(&[f64], &[f64]) -> f64 {
    match objective_type {
        ObjectiveType::LogLoss => LogLoss::calc_init,
        ObjectiveType::SquaredLoss => SquaredLoss::calc_init,
    }
}

pub trait ObjectiveFunction {
    fn calc_loss(y: &[f64], yhat: &[f64], sample_weight: &[f64]) -> Vec<f32>;
    fn calc_grad_hess(y: &[f64], yhat: &[f64], sample_weight: &[f64]) -> (Vec<f32>, Vec<f32>);
    fn calc_init(y: &[f64], sample_weight: &[f64]) -> f64;
    fn default_metric() -> Metric;
}

#[derive(Default)]
pub struct LogLoss {}

impl ObjectiveFunction for LogLoss {
    #[inline]
    fn calc_loss(y: &[f64], yhat: &[f64], sample_weight: &[f64]) -> Vec<f32> {
        y.iter()
            .zip(yhat)
            .zip(sample_weight)
            .map(|((y_, yhat_), w_)| {
                let yhat_ = f64::ONE / (f64::ONE + (-*yhat_).exp());
                (-(*y_ * yhat_.ln() + (f64::ONE - *y_) * ((f64::ONE - yhat_).ln())) * *w_) as f32
            })
            .collect()
    }

    fn calc_init(y: &[f64], sample_weight: &[f64]) -> f64 {
        let mut ytot: f64 = 0.;
        let mut ntot: f64 = 0.;
        for i in 0..y.len() {
            ytot += sample_weight[i] * y[i];
            ntot += sample_weight[i];
        }
        f64::ln(ytot / (ntot - ytot))
    }

    #[inline]
    fn calc_grad_hess(y: &[f64], yhat: &[f64], sample_weight: &[f64]) -> (Vec<f32>, Vec<f32>) {
        y.iter()
            .zip(yhat)
            .zip(sample_weight)
            .map(|((y_, yhat_), w_)| {
                let yhat_ = f64::ONE / (f64::ONE + (-*yhat_).exp());
                (
                    ((yhat_ - *y_) * *w_) as f32,
                    (yhat_ * (f64::ONE - yhat_) * *w_) as f32,
                )
            })
            .unzip()
    }

    fn default_metric() -> Metric {
        Metric::LogLoss
    }
}

#[derive(Default)]
pub struct SquaredLoss {}

impl ObjectiveFunction for SquaredLoss {
    #[inline]
    fn calc_loss(y: &[f64], yhat: &[f64], sample_weight: &[f64]) -> Vec<f32> {
        y.iter()
            .zip(yhat)
            .zip(sample_weight)
            .map(|((y_, yhat_), w_)| {
                let s = *y_ - *yhat_;
                (s * s * *w_) as f32
            })
            .collect()
    }

    fn calc_init(y: &[f64], sample_weight: &[f64]) -> f64 {
        let mut ytot: f64 = 0.;
        let mut ntot: f64 = 0.;
        for i in 0..y.len() {
            ytot += sample_weight[i] * y[i];
            ntot += sample_weight[i];
        }

        ytot / ntot
    }

    #[inline]
    fn calc_grad_hess(y: &[f64], yhat: &[f64], sample_weight: &[f64]) -> (Vec<f32>, Vec<f32>) {
        y.iter()
            .zip(yhat)
            .zip(sample_weight)
            .map(|((y_, yhat_), w_)| (((yhat_ - *y_) * *w_) as f32, *w_ as f32))
            .unzip()
    }
    fn default_metric() -> Metric {
        Metric::RootMeanSquaredLogError
    }
}

#[cfg(test)]
mod tests {
    use super::*;
    #[test]
    fn test_logloss_loss() {
        let y = vec![0.0, 0.0, 0.0, 1.0, 1.0, 1.0];
        let yhat1 = vec![-1.0, -1.0, -1.0, 1.0, 1.0, 1.0];
        let w = vec![1.; y.len()];
        let l1 = LogLoss::calc_loss(&y, &yhat1, &w);
        let yhat2 = vec![0.0, 0.0, -1.0, 1.0, 0.0, 1.0];
        let l2 = LogLoss::calc_loss(&y, &yhat2, &w);
        assert!(l1.iter().sum::<f32>() < l2.iter().sum::<f32>());
    }

    #[test]
    fn test_logloss_grad() {
        let y = vec![0.0, 0.0, 0.0, 1.0, 1.0, 1.0];
        let yhat1 = vec![-1.0, -1.0, -1.0, 1.0, 1.0, 1.0];
        let w = vec![1.; y.len()];
        let (g1, _) = LogLoss::calc_grad_hess(&y, &yhat1, &w);
        let yhat2 = vec![0.0, 0.0, -1.0, 1.0, 0.0, 1.0];
        let (g2, _) = LogLoss::calc_grad_hess(&y, &yhat2, &w);
        assert!(g1.iter().sum::<f32>() < g2.iter().sum::<f32>());
    }

    #[test]
    fn test_logloss_hess() {
        let y = vec![0.0, 0.0, 0.0, 1.0, 1.0, 1.0];
        let yhat1 = vec![-1.0, -1.0, -1.0, 1.0, 1.0, 1.0];
        let w = vec![1.; y.len()];
        let (_, h1) = LogLoss::calc_grad_hess(&y, &yhat1, &w);
        let yhat2 = vec![0.0, 0.0, -1.0, 1.0, 0.0, 1.0];
        let (_, h2) = LogLoss::calc_grad_hess(&y, &yhat2, &w);
        assert!(h1.iter().sum::<f32>() < h2.iter().sum::<f32>());
    }

    #[test]
    fn test_logloss_init() {
        let y = vec![0.0, 0.0, 0.0, 1.0, 1.0, 1.0];
        let w = vec![1.; y.len()];
        let l1 = LogLoss::calc_init(&y, &w);
        assert!(l1 == 0.);

        let y = vec![1.0; 6];
        let l2 = LogLoss::calc_init(&y, &w);
        assert!(l2 == f64::INFINITY);

        let y = vec![0.0; 6];
        let l3 = LogLoss::calc_init(&y, &w);
        assert!(l3 == f64::NEG_INFINITY);

        let y = vec![0., 0., 0., 0., 1., 1.];
        let l4 = LogLoss::calc_init(&y, &w);
        assert!(l4 == f64::ln(2. / 4.));
    }

    #[test]
    fn test_mse_init() {
        let y = vec![0.0, 0.0, 0.0, 1.0, 1.0, 1.0];
        let w = vec![1.; y.len()];
        let l1 = SquaredLoss::calc_init(&y, &w);
        assert!(l1 == 0.5);

        let y = vec![1.0, 1.0, 1.0, 1.0, 1.0, 1.0];
        let l2 = SquaredLoss::calc_init(&y, &w);
        assert!(l2 == 1.);

        let y = vec![-1.0, -1.0, -1.0, -1.0, -1.0, -1.0];
        let l3 = SquaredLoss::calc_init(&y, &w);
        assert!(l3 == -1.);

        let y = vec![-1.0, -1.0, -1.0, 1., 1., 1.];
        let l4 = SquaredLoss::calc_init(&y, &w);
        assert!(l4 == 0.);
    }
}