irithyll 10.0.1

Streaming ML in Rust -- gradient boosted trees, neural architectures (TTT/KAN/MoE/Mamba/SNN), AutoML, kernel methods, and composable pipelines
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

//! Logistic loss for binary classification.
//!
//! L(y, f) = -y * ln(sigma(f)) - (1-y) * ln(1 - sigma(f))
//!
//! where sigma(f) = 1 / (1 + exp(-f)) is the sigmoid function,
//! y in {0, 1} is the binary target, and f is the raw model output (logit).
//!
//! Also known as binary cross-entropy or log loss.

pub use super::{Loss, LossType};
pub use irithyll_core::loss::logistic::*;

#[cfg(test)]
mod tests {
    use super::*;

    const EPS: f64 = 1e-10;

    #[test]
    fn test_n_outputs() {
        assert_eq!(LogisticLoss.n_outputs(), 1);
    }

    #[test]
    fn test_gradient_target_1_pred_0() {
        let loss = LogisticLoss;
        // sigmoid(0) = 0.5, so gradient = 0.5 - 1.0 = -0.5
        let g = loss.gradient(1.0, 0.0);
        assert!((g - (-0.5)).abs() < EPS);
    }

    #[test]
    fn test_gradient_target_0_pred_0() {
        let loss = LogisticLoss;
        // sigmoid(0) = 0.5, gradient = 0.5 - 0.0 = 0.5
        let g = loss.gradient(0.0, 0.0);
        assert!((g - 0.5).abs() < EPS);
    }

    #[test]
    fn test_gradient_perfect_prediction() {
        let loss = LogisticLoss;
        // target=1, prediction very large => sigmoid ~ 1.0, gradient ~ 0
        let g = loss.gradient(1.0, 20.0);
        assert!(g.abs() < 1e-6);
    }

    #[test]
    fn test_hessian_positive() {
        let loss = LogisticLoss;
        // Hessian should always be positive for logistic loss
        assert!(loss.hessian(0.0, 0.0) > 0.0);
        assert!(loss.hessian(1.0, 5.0) > 0.0);
        assert!(loss.hessian(0.0, -5.0) > 0.0);
        assert!(loss.hessian(1.0, 100.0) > 0.0); // extreme, but still > 0 (clamped)
    }

    #[test]
    fn test_hessian_max_at_pred_zero() {
        let loss = LogisticLoss;
        // p*(1-p) is maximized at p=0.5, i.e. prediction=0
        let h_zero = loss.hessian(0.0, 0.0);
        let h_five = loss.hessian(0.0, 5.0);
        assert!((h_zero - 0.25).abs() < EPS);
        assert!(h_five < h_zero);
    }

    #[test]
    fn test_loss_value() {
        let loss = LogisticLoss;
        // At prediction=0, sigmoid=0.5, loss = -ln(0.5) = ln(2) for both classes
        let l1 = loss.loss(1.0, 0.0);
        let l0 = loss.loss(0.0, 0.0);
        let ln2 = 2.0_f64.ln();
        assert!((l1 - ln2).abs() < 1e-8);
        assert!((l0 - ln2).abs() < 1e-8);
    }

    #[test]
    fn test_predict_transform_is_sigmoid() {
        let loss = LogisticLoss;
        assert!((loss.predict_transform(0.0) - 0.5).abs() < EPS);
        assert!(loss.predict_transform(10.0) > 0.99);
        assert!(loss.predict_transform(-10.0) < 0.01);
    }

    #[test]
    fn test_initial_prediction_balanced() {
        let loss = LogisticLoss;
        // Balanced dataset: mean=0.5, log-odds = ln(1) = 0
        let targets = [0.0, 1.0, 0.0, 1.0];
        assert!(loss.initial_prediction(&targets).abs() < EPS);
    }

    #[test]
    fn test_initial_prediction_skewed() {
        let loss = LogisticLoss;
        // 75% positive: mean=0.75, log-odds = ln(3) ~ 1.0986
        let targets = [1.0, 1.0, 1.0, 0.0];
        let init = loss.initial_prediction(&targets);
        let expected = (0.75_f64 / 0.25).ln();
        assert!((init - expected).abs() < 1e-8);
    }

    #[test]
    fn test_initial_prediction_empty() {
        let loss = LogisticLoss;
        assert!((loss.initial_prediction(&[])).abs() < EPS);
    }

    #[test]
    fn test_gradient_is_derivative_of_loss() {
        let loss = LogisticLoss;
        let target = 1.0;
        let pred = 1.5;
        let h = 1e-7;
        let numerical = (loss.loss(target, pred + h) - loss.loss(target, pred - h)) / (2.0 * h);
        let analytical = loss.gradient(target, pred);
        assert!(
            (numerical - analytical).abs() < 1e-5,
            "numerical={numerical}, analytical={analytical}"
        );
    }
}