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

use crate::errors::ForustError;
use crate::utils::items_to_strings;
use rand::rngs::StdRng;
use rand::Rng;
use serde::{Deserialize, Serialize};
use std::str::FromStr;

#[derive(Serialize, Deserialize)]
pub enum SampleMethod {
    None,
    Random,
    Goss,
}

impl FromStr for SampleMethod {
    type Err = ForustError;

    fn from_str(s: &str) -> Result<Self, Self::Err> {
        match s {
            "random" => Ok(SampleMethod::Random),
            "goss" => Ok(SampleMethod::Goss),
            _ => Err(ForustError::ParseString(
                s.to_string(),
                "SampleMethod".to_string(),
                items_to_strings(vec!["random", "goss"]),
            )),
        }
    }
}

// A sampler can be used to subset the data prior to fitting a new tree.
pub trait Sampler {
    /// Sample the data, returning a tuple, where the first item is the samples
    /// chosen for training, and the second are the samples excluded.
    fn sample(
        &mut self,
        rng: &mut StdRng,
        index: &[usize],
        grad: &mut [f32],
        hess: &mut [f32],
    ) -> (Vec<usize>, Vec<usize>);
}

pub struct RandomSampler {
    subsample: f32,
}

impl RandomSampler {
    #[allow(dead_code)]
    pub fn new(subsample: f32) -> Self {
        RandomSampler { subsample }
    }
}

impl Sampler for RandomSampler {
    fn sample(
        &mut self,
        rng: &mut StdRng,
        index: &[usize],
        _grad: &mut [f32],
        _hess: &mut [f32],
    ) -> (Vec<usize>, Vec<usize>) {
        let subsample = self.subsample;
        let mut chosen = Vec::new();
        let mut excluded = Vec::new();
        for i in index {
            if rng.gen_range(0.0..1.0) < subsample {
                chosen.push(*i);
            } else {
                excluded.push(*i)
            }
        }
        (chosen, excluded)
    }
}

#[allow(dead_code)]
pub struct GossSampler {
    a: f64, // https://lightgbm.readthedocs.io/en/latest/Parameters.html#top_rate
    b: f64, // https://lightgbm.readthedocs.io/en/latest/Parameters.html#other_rate
}

impl Default for GossSampler {
    fn default() -> Self {
        GossSampler { a: 0.2, b: 0.1 }
    }
}

#[allow(dead_code)]
impl GossSampler {
    pub fn new(a: f64, b: f64) -> Self {
        GossSampler { a, b }
    }
}

impl Sampler for GossSampler {
    fn sample(
        &mut self,
        rng: &mut StdRng,
        index: &[usize],
        grad: &mut [f32],
        hess: &mut [f32],
    ) -> (Vec<usize>, Vec<usize>) {
        let fact = ((1. - self.a) / self.b) as f32;
        let top_n = (self.a * index.len() as f64) as usize;
        let rand_n = (self.b * index.len() as f64) as usize;

        // sort gradient by absolute value from highest to lowest
        let mut sorted = (0..index.len()).collect::<Vec<_>>();
        sorted.sort_unstable_by(|&a, &b| grad[b].abs().total_cmp(&grad[a].abs()));

        // select the topN largest gradients
        let mut used_set = sorted[0..top_n].to_vec();

        // sample the rest based on randN
        let subsample = rand_n as f64 / (index.len() as f64 - top_n as f64);

        // weight the sampled "small gradients" by fact and append indices to used_set
        for i in &sorted[top_n..sorted.len()] {
            if rng.gen_range(0.0..1.0) < subsample {
                grad[*i] *= fact;
                hess[*i] *= fact;
                used_set.push(*i);
            }
        }

        (used_set, Vec::new())
    }
}