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
use std::fs::File;
use std::io::{BufReader, BufWriter, Read, Write};
use std::path::Path;
use bincode::config::standard;
use serde::{Deserialize, Serialize};
use crate::dataset::BinMapper;
use crate::error::{Error, Result};
use crate::feature::FeatureBuilder;
use crate::loss::sigmoid;
use crate::tree::Tree;
/// A trained GBDT model: ensemble of trees plus boosting metadata.
#[derive(Debug, Clone, Serialize, Deserialize)]
pub struct Model {
pub(crate) init_score: f64,
pub(crate) learning_rate: f64,
pub(crate) n_features: usize,
/// Feature names in index order, copied from the training
/// [`FeatureBuilder`]. Predict asserts the supplied builder matches these.
pub(crate) feature_names: Vec<String>,
/// Required by the binned inference path: per-node `threshold_bin` values
/// in trees are calibrated to these specific mappers, so re-fitting would
/// produce wrong predictions.
pub(crate) bin_mappers: Vec<BinMapper>,
pub(crate) trees: Vec<Tree>,
}
impl Model {
/// Constant the boosting loop started from (the prior log-odds of the labels).
pub fn init_score(&self) -> f64 {
self.init_score
}
/// Shrinkage applied to every tree's leaf value at inference.
pub fn learning_rate(&self) -> f64 {
self.learning_rate
}
/// Number of input features this model expects per row.
pub fn n_features(&self) -> usize {
self.n_features
}
/// Feature names in index order, as declared by the training builder.
pub fn feature_names(&self) -> &[String] {
&self.feature_names
}
/// Number of trees in the ensemble. After early stopping, this equals
/// `best_iter + 1`, not the configured `num_iterations`.
pub fn n_trees(&self) -> usize {
self.trees.len()
}
/// The trees themselves, in fit order.
pub fn trees(&self) -> &[Tree] {
&self.trees
}
/// Bincode-serialize this model to `path`.
pub fn save<P: AsRef<Path>>(&self, path: P) -> Result<()> {
let f = File::create(path)?;
let mut w = BufWriter::new(f);
let bytes = bincode::serde::encode_to_vec(self, standard())
.map_err(|e| Error::Serde(e.to_string()))?;
w.write_all(&bytes)?;
w.flush()?;
Ok(())
}
/// Deserialize a model previously written by [`Model::save`].
pub fn load<P: AsRef<Path>>(path: P) -> Result<Self> {
let f = File::open(path)?;
let mut r = BufReader::new(f);
let mut buf = Vec::new();
r.read_to_end(&mut buf)?;
let (model, _) = bincode::serde::decode_from_slice::<Model, _>(&buf, standard())
.map_err(|e| Error::Serde(e.to_string()))?;
Ok(model)
}
/// Number of times each feature was used as a split, indexed by feature.
pub fn feature_importance_split(&self) -> Vec<u32> {
let mut counts = vec![0u32; self.n_features];
for tree in &self.trees {
for node in &tree.nodes {
counts[node.feature as usize] += 1;
}
}
counts
}
/// Total split gain attributed to each feature, indexed by feature.
pub fn feature_importance_gain(&self) -> Vec<f64> {
let mut gains = vec![0.0f64; self.n_features];
for tree in &self.trees {
for (node, gain) in tree.nodes.iter().zip(tree.node_gains.iter()) {
gains[node.feature as usize] += gain;
}
}
gains
}
/// Assert the supplied builder declares exactly the features this model was
/// trained on (same names, same order). Guards against predicting with a
/// mismatched extractor set.
fn check_features<T>(&self, features: &FeatureBuilder<T>) {
assert_eq!(
features.len(),
self.n_features,
"feature count mismatch: builder has {}, model expects {}",
features.len(),
self.n_features
);
for (i, (got, want)) in features.names().zip(self.feature_names.iter()).enumerate() {
assert_eq!(
got, want,
"feature {i} name mismatch: builder has {got:?}, model expects {want:?}"
);
}
}
/// Predict probabilities (sigmoid of the raw additive scores) for `rows`,
/// extracting features through `features` (the same builder used to train).
///
/// Bins each row through the training-time [`BinMapper`]s into a row-major
/// scratch buffer, then walks the trees on bin codes (`u16` comparisons, no
/// per-node NaN check). Walks tree-outer/row-inner so the current tree's
/// nodes stay hot in L1 across the row sweep. This is the only inference
/// path: it's the fastest at both per-request and bulk batch sizes.
///
/// # Panics
/// Panics if `features` doesn't match the model's declared features.
pub fn predict_proba<T>(&self, features: &FeatureBuilder<T>, rows: &[T]) -> Vec<f64> {
self.predict_raw_scores(features, rows).into_iter().map(sigmoid).collect()
}
/// Raw additive scores (pre-sigmoid logits). Private on purpose: the public
/// serving surface is [`Model::predict_proba`]; raw logits aren't exposed.
fn predict_raw_scores<T>(&self, features: &FeatureBuilder<T>, rows: &[T]) -> Vec<f64> {
self.check_features(features);
let nf = self.n_features.max(1);
let mut bins: Vec<u16> = Vec::new();
features.extract_bins_row_major(rows, &self.bin_mappers, &mut bins);
let lr = self.learning_rate;
let mut scores = vec![self.init_score; rows.len()];
for tree in &self.trees {
for (r, s) in scores.iter_mut().enumerate() {
*s += lr * tree.predict_on_row_bins(&bins[r * nf..(r + 1) * nf]);
}
}
scores
}
}