nanogbm 0.2.0

A small, pure-Rust gradient boosting library (GBDT, binary classification, CPU only).
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

use nanogbm::metric::{BinaryLogloss, Metric};
use nanogbm::{Config, DatasetBuilder, GbdtTrainer, Model};
use rand::SeedableRng;
use rand::prelude::*;
use rand_chacha::ChaCha8Rng;

/// Generate a noisy linearly-separable binary classification problem. `train_n` rows
/// for training, `valid_n` for validation, with the same underlying weights/bias.
fn make_classification(
    train_n: usize,
    valid_n: usize,
    d: usize,
    seed: u64,
) -> (Vec<f64>, Vec<f32>, Vec<f64>, Vec<f32>) {
    let mut rng = ChaCha8Rng::seed_from_u64(seed);
    let weights: Vec<f64> = (0..d).map(|_| rng.gen_range(-1.0..1.0)).collect();
    let bias: f64 = rng.gen_range(-0.5..0.5);
    let mut sample = |n: usize| -> (Vec<f64>, Vec<f32>) {
        let mut features = vec![0.0; n * d];
        let mut labels = vec![0f32; n];
        for i in 0..n {
            let mut z = bias;
            for j in 0..d {
                let x: f64 = rng.gen_range(-3.0..3.0);
                features[i * d + j] = x;
                z += weights[j] * x;
            }
            z += rng.gen_range(-0.4..0.4);
            let p = 1.0 / (1.0 + (-z).exp());
            labels[i] = if rng.r#gen::<f64>() < p { 1.0 } else { 0.0 };
        }
        (features, labels)
    };
    let (tx, ty) = sample(train_n);
    let (vx, vy) = sample(valid_n);
    (tx, ty, vx, vy)
}

#[test]
fn trains_and_reduces_logloss_on_synthetic() {
    let n_train = 4000;
    let n_valid = 1000;
    let d = 8;
    let (train_x, train_y, valid_x, valid_y) = make_classification(n_train, n_valid, d, 42);

    let mut cfg = Config::default();
    cfg.num_iterations = 200;
    cfg.learning_rate = 0.1;
    cfg.num_leaves = 31;
    cfg.min_data_in_leaf = 20;
    cfg.max_bin = 64;
    cfg.feature_fraction = 1.0;
    cfg.bagging_fraction = 1.0;
    cfg.lambda_l2 = 1.0;
    cfg.early_stopping_round = 20;
    cfg.seed = 0;

    let train_ds = DatasetBuilder::from_rows(&train_x, n_train, d, &train_y, &cfg).unwrap();
    let valid_ds = DatasetBuilder::from_rows(&valid_x, n_valid, d, &valid_y, &cfg).unwrap();
    let model = GbdtTrainer::new(&cfg)
        .fit(&train_ds, Some(&valid_ds))
        .unwrap();

    assert!(model.n_trees() > 0);

    // Compare logloss of init prediction vs. trained model on validation set.
    let metric = BinaryLogloss;
    let init_only_scores = vec![model.init_score(); n_valid];
    let init_loss = metric.evaluate(&init_only_scores, &valid_y);

    let final_scores = model.predict_raw_scores(&valid_x, n_valid);
    let final_loss = metric.evaluate(&final_scores, &valid_y);

    println!("init_loss={init_loss:.5} final_loss={final_loss:.5}");
    assert!(
        final_loss < init_loss * 0.9,
        "model failed to improve appreciably: {init_loss} -> {final_loss}"
    );
    assert!(final_loss < 0.6, "final loss too high: {final_loss}");
}

#[test]
fn saved_model_round_trip_matches_predictions() {
    let n = 500;
    let d = 4;
    let (x, y, _, _) = make_classification(n, 1, d, 7);
    let mut cfg = Config::default();
    cfg.num_iterations = 10;
    cfg.num_leaves = 15;
    cfg.min_data_in_leaf = 10;
    cfg.max_bin = 32;
    let ds = DatasetBuilder::from_rows(&x, n, d, &y, &cfg).unwrap();
    let model = GbdtTrainer::new(&cfg).fit(&ds, None).unwrap();

    let tmp = std::env::temp_dir().join("nanogbm_model.bin");
    model.save(&tmp).unwrap();
    let loaded = Model::load(&tmp).unwrap();

    let p1 = model.predict_proba(&x, n);
    let p2 = loaded.predict_proba(&x, n);
    for (a, b) in p1.iter().zip(p2.iter()) {
        assert!((a - b).abs() < 1e-12);
    }
}

#[test]
fn predict_bin_and_raw_paths_agree() {
    let n = 400;
    let d = 3;
    let (x, y, _, _) = make_classification(n, 1, d, 9);

    let mut cfg = Config::default();
    cfg.num_iterations = 8;
    cfg.num_leaves = 11;
    cfg.min_data_in_leaf = 5;
    cfg.max_bin = 32;
    let ds = DatasetBuilder::from_rows(&x, n, d, &y, &cfg).unwrap();
    let model = GbdtTrainer::new(&cfg).fit(&ds, None).unwrap();

    let raw = model.predict_raw_scores(&x, n);
    let bin_raw = model.predict_raw_scores_on_dataset(&ds);

    let max_abs = raw
        .iter()
        .zip(bin_raw.iter())
        .map(|(a, b)| (a - b).abs())
        .fold(0.0f64, f64::max);
    assert!(
        max_abs < 1e-9,
        "raw vs bin predictions diverged by {max_abs}"
    );
}

#[test]
fn predict_proba_binned_matches_raw_proba() {
    // Binned predict path must produce the same predictions as the raw f64
    // predict path: the trees carry both `threshold` and `threshold_bin`, and
    // the model now stores `bin_mappers` so the binned input gets the same
    // bin codes the trees were trained against.
    let n_train = 800;
    let n_eval = 1500;
    let d = 6;
    let (tx, ty, ex, _ey) = make_classification(n_train, n_eval, d, 17);

    let mut cfg = Config::default();
    cfg.num_iterations = 15;
    cfg.num_leaves = 21;
    cfg.min_data_in_leaf = 5;
    cfg.max_bin = 64;
    cfg.lambda_l2 = 0.5;

    let train_ds = DatasetBuilder::from_rows(&tx, n_train, d, &ty, &cfg).unwrap();
    let model = GbdtTrainer::new(&cfg).fit(&train_ds, None).unwrap();

    let raw_proba = model.predict_proba(&ex, n_eval);
    let binned_proba = model.predict_proba_binned(&ex, n_eval);

    let max_abs = raw_proba
        .iter()
        .zip(binned_proba.iter())
        .map(|(a, b)| (a - b).abs())
        .fold(0.0f64, f64::max);
    assert!(
        max_abs < 1e-9,
        "raw vs binned proba diverged by {max_abs}"
    );
}

#[test]
fn binned_predict_survives_bincode_round_trip() {
    // Saved/restored Model must still produce identical binned predictions:
    // `bin_mappers` need to round-trip through bincode.
    let n = 600;
    let d = 4;
    let (x, y, ex, _) = make_classification(n, n / 2, d, 31);
    let mut cfg = Config::default();
    cfg.num_iterations = 12;
    cfg.num_leaves = 15;
    cfg.min_data_in_leaf = 10;
    cfg.max_bin = 48;
    let ds = DatasetBuilder::from_rows(&x, n, d, &y, &cfg).unwrap();
    let model = GbdtTrainer::new(&cfg).fit(&ds, None).unwrap();

    let tmp = std::env::temp_dir().join("nanogbm_model_binned.bin");
    model.save(&tmp).unwrap();
    let loaded = Model::load(&tmp).unwrap();

    let before = model.predict_proba_binned(&ex, n / 2);
    let after = loaded.predict_proba_binned(&ex, n / 2);
    for (a, b) in before.iter().zip(after.iter()) {
        assert!((a - b).abs() < 1e-12);
    }
}