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
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
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::{Bin, BinData, BinMapper, BinWidth, Dataset};
use crate::error::{Error, Result};
use crate::objective::binary::sigmoid;
use crate::tree::Tree;
/// A trained GBDT model: ensemble of trees plus boosting metadata.
///
/// Fields are crate-private. Inspect a model through the accessor methods or
/// the `predict_*` methods.
#[derive(Debug, Clone, Serialize, Deserialize)]
pub struct Model {
pub(crate) init_score: f64,
pub(crate) learning_rate: f64,
pub(crate) n_features: usize,
/// Per-feature [`BinMapper`] from the training [`Dataset`]. Required for
/// the binned inference path ([`Model::predict_proba_binned`]) — the
/// per-node `threshold_bin` values stored in trees are calibrated to
/// these specific mappers, so re-fitting them at predict time 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
}
/// 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
}
/// Predict raw additive scores (pre-sigmoid logits) for a row-major feature
/// matrix of shape `n_rows × self.n_features()`.
///
/// # Panics
/// Panics if `features.len() != n_rows * self.n_features()`.
pub fn predict_raw_scores(&self, features: &[f64], n_rows: usize) -> Vec<f64> {
let n_features = self.n_features;
assert_eq!(
features.len(),
n_rows * n_features,
"features.len() {} != n_rows {} * n_features {}",
features.len(),
n_rows,
n_features
);
let init = self.init_score;
(0..n_rows)
.map(|row| {
let r = &features[row * n_features..(row + 1) * n_features];
let mut s = init;
for tree in &self.trees {
s += self.learning_rate * tree.predict_raw(r);
}
s
})
.collect()
}
/// Predict probabilities (sigmoid of raw scores) for a row-major feature
/// matrix of shape `n_rows × self.n_features()`.
pub fn predict_proba(&self, features: &[f64], n_rows: usize) -> Vec<f64> {
let raw = self.predict_raw_scores(features, n_rows);
raw.into_iter().map(sigmoid).collect()
}
/// Predict raw additive scores against an already-binned dataset. Use this
/// for fast inference paths where you can afford to bin once and predict
/// many times.
///
/// Dispatches once on the dataset's bin width and pre-collects column
/// slices, then walks each tree on each row using a type-stable inner
/// loop ([`crate::tree::Tree::predict_on_columns`]). Avoids the per-node
/// `BinData::U8/U16` match that [`crate::tree::Tree::predict_on_dataset`]
/// otherwise incurs.
pub fn predict_raw_scores_on_dataset(&self, dataset: &Dataset) -> Vec<f64> {
let n = dataset.n_rows();
let mut scores = vec![self.init_score; n];
match dataset.bin_width() {
crate::dataset::BinWidth::U8 => {
let cols: Vec<&[u8]> = (0..dataset.n_features())
.map(|f| dataset.feature_column_u8(f))
.collect();
self.predict_into_with_columns(&cols, n, &mut scores);
}
crate::dataset::BinWidth::U16 => {
let cols: Vec<&[u16]> = (0..dataset.n_features())
.map(|f| dataset.feature_column_u16(f))
.collect();
self.predict_into_with_columns(&cols, n, &mut scores);
}
}
scores
}
/// Tree-outer / row-inner accumulation onto `scores`. The tree-outer
/// order keeps the current tree's nodes hot in L1 across the full
/// row sweep.
fn predict_into_with_columns<B: Bin>(
&self,
columns: &[&[B]],
n_rows: usize,
scores: &mut [f64],
) {
for tree in &self.trees {
for row in 0..n_rows {
scores[row] += self.learning_rate * tree.predict_on_columns(columns, row);
}
}
}
/// Predict probabilities against an already-binned dataset.
pub fn predict_proba_on_dataset(&self, dataset: &Dataset) -> Vec<f64> {
let raw = self.predict_raw_scores_on_dataset(dataset);
raw.into_iter().map(sigmoid).collect()
}
/// Predict raw scores by binning the input features once (using the
/// [`BinMapper`]s stored at train time), then walking trees on bin
/// codes — significantly faster than [`Model::predict_raw_scores`] for
/// batch inference (>~10K rows). Predictions are equivalent: the trees
/// carry both raw and binned thresholds, so the two paths produce the
/// same leaves.
///
/// Why it's faster:
/// - u8 / u16 comparison vs f64 comparison at every node visit.
/// - 47-byte rows fit in one cache line vs 376-byte f64 rows.
/// - No `is_finite` NaN check — missing maps to bin 0 once at binning time.
///
/// # Panics
/// Panics if `features.len() != n_rows * self.n_features()`.
pub fn predict_raw_scores_binned(&self, features: &[f64], n_rows: usize) -> Vec<f64> {
let dataset = self.bin_for_predict(features, n_rows);
self.predict_raw_scores_on_dataset(&dataset)
}
/// Like [`Model::predict_proba`] but uses the binned inference path. See
/// [`Model::predict_raw_scores_binned`] for the rationale and tradeoffs.
pub fn predict_proba_binned(&self, features: &[f64], n_rows: usize) -> Vec<f64> {
let raw = self.predict_raw_scores_binned(features, n_rows);
raw.into_iter().map(sigmoid).collect()
}
/// Bin `features` using `self.bin_mappers` and pack into a [`Dataset`]
/// suitable for the `*_on_dataset` predict paths. Width (u8 / u16) is
/// chosen by the max `num_bins` across mappers — same rule as
/// [`crate::dataset::DatasetBuilder`].
fn bin_for_predict(&self, features: &[f64], n_rows: usize) -> Dataset {
let n_features = self.n_features;
assert_eq!(
features.len(),
n_rows * n_features,
"features.len() {} != n_rows {} * n_features {}",
features.len(),
n_rows,
n_features
);
assert_eq!(
self.bin_mappers.len(),
n_features,
"model.bin_mappers.len() {} != n_features {}",
self.bin_mappers.len(),
n_features
);
let max_num_bins = self
.bin_mappers
.iter()
.map(|m| m.num_bins())
.max()
.unwrap_or(2);
let width = if max_num_bins <= 256 {
BinWidth::U8
} else {
BinWidth::U16
};
// Bin column-by-column from row-major raw features. Two passes per
// column: one to read the f64 column into a scratch, one to write
// bin codes. The scratch avoids re-striding the row-major buffer
// inside the inner loop.
let bin_data = match width {
BinWidth::U8 => BinData::U8(self.bin_columns::<u8>(features, n_rows, n_features)),
BinWidth::U16 => BinData::U16(self.bin_columns::<u16>(features, n_rows, n_features)),
};
Dataset {
n_rows,
n_features,
bin_data,
// The on_dataset predict path doesn't read bin_mappers — but the
// Dataset struct requires the field. Avoid the clone by handing
// out a reference-counted empty Vec? No — Dataset owns its
// mappers. Just clone; this is one-shot per predict batch.
bin_mappers: self.bin_mappers.clone(),
// Labels aren't read by predict; allocate an empty placeholder.
labels: Vec::new(),
}
}
fn bin_columns<B: Bin>(
&self,
features: &[f64],
n_rows: usize,
n_features: usize,
) -> Vec<Vec<B>> {
(0..n_features)
.map(|feat| {
let bm = &self.bin_mappers[feat];
let mut col: Vec<B> = Vec::with_capacity(n_rows);
for row in 0..n_rows {
let v = features[row * n_features + feat];
col.push(B::from_u16(bm.value_to_bin(v)));
}
col
})
.collect()
}
}