gprx 0.2.0

Gaussian process regression in Rust: exact GPR, sparse GPR (VFE), and SVGP.
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
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
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
//! Optimizer-facing objective traits.
//!
//! Each model implements them on its own adapter (`GprObjective` in
//! `src/gpr/`, `SgprObjective` in `src/sgpr/`), which borrows the fitted
//! model during `fit` / `refit` and forwards concatenated kernel-then-
//! likelihood `θ` to it. Capability is split so a derivative-free solver
//! can require only [`Objective`], L-BFGS can require [`Differentiable`],
//! and a Hessian-based solver ([`crate::TrustRegion`]) can require [`TwiceDifferentiable`].

use crate::error::GprError;
use crate::param::Interval;

/// Calls the model adapters make, for the `compare/perf` optimizer tables.
///
/// Present only with the non-default feature `bench-internals`. One global
/// pair of counters: fits must not run concurrently while it is read.
#[cfg(feature = "bench-internals")]
pub(crate) mod call_counts {
    use std::sync::atomic::{AtomicU64, Ordering};

    static VALUE_CALLS: AtomicU64 = AtomicU64::new(0);
    static JOINT_CALLS: AtomicU64 = AtomicU64::new(0);

    pub(crate) fn count_value() {
        VALUE_CALLS.fetch_add(1, Ordering::Relaxed);
    }

    pub(crate) fn count_joint() {
        JOINT_CALLS.fetch_add(1, Ordering::Relaxed);
    }

    pub(crate) fn reset() {
        VALUE_CALLS.store(0, Ordering::Relaxed);
        JOINT_CALLS.store(0, Ordering::Relaxed);
    }

    pub(crate) fn read() -> (u64, u64) {
        (
            VALUE_CALLS.load(Ordering::Relaxed),
            JOINT_CALLS.load(Ordering::Relaxed),
        )
    }
}

#[cfg(feature = "bench-internals")]
macro_rules! count_value_call {
    () => {
        $crate::objective::call_counts::count_value()
    };
}
#[cfg(not(feature = "bench-internals"))]
macro_rules! count_value_call {
    () => {};
}
#[cfg(feature = "bench-internals")]
macro_rules! count_joint_call {
    () => {
        $crate::objective::call_counts::count_joint()
    };
}
#[cfg(not(feature = "bench-internals"))]
macro_rules! count_joint_call {
    () => {};
}
pub(crate) use {count_joint_call, count_value_call};

/// Represents the optimizer-facing scalar objective (`value` only).
///
/// Default [`Differentiable::value_and_gradient_into`] is not on this trait.
/// A solver that needs derivatives takes [`Differentiable`], not a runtime
/// flag.
///
/// The built-in optimizers read each parameter on a positive interval as
/// `log` of a user-unit value inside that [`Interval`] (log-`θ` for the model
/// fits), and any other parameter as the user-unit value itself.
///
/// # Examples
///
/// The smallest objective:
///
/// ```rust
/// use gprx::Objective;
///
/// struct Zero;
///
/// impl Objective for Zero {
///     fn num_params(&self) -> usize {
///         0
///     }
///
///     fn value(&mut self, _params: &[f64]) -> Result<f64, gprx::GprError> {
///         Ok(0.0)
///     }
/// }
///
/// # fn main() -> Result<(), gprx::GprError> {
/// let mut objective = Zero;
/// assert_eq!(objective.value(&[])?, 0.0);
/// # Ok(())
/// # }
/// ```
///
/// A bowl `f(θ) = Σ (θ_i − c_i)²` with every capability, minimized by each
/// built-in optimizer:
///
/// ```rust
/// use gprx::{
///     Differentiable, FastSimulatedAnnealing, GprError, IncrementalObjective, Interval, Lbfgs,
///     NelderMead, Objective, Optimizer, TrustRegion, TwiceDifferentiable,
/// };
///
/// struct Bowl {
///     center: Vec<f64>,
/// }
///
/// impl Bowl {
///     fn check(&self, params: &[f64]) -> Result<(), GprError> {
///         if params.len() == self.center.len() {
///             Ok(())
///         } else {
///             Err(GprError::LengthMismatch {
///                 reason: format!("expected {} parameters", self.center.len()),
///             })
///         }
///     }
/// }
///
/// impl Objective for Bowl {
///     fn num_params(&self) -> usize {
///         self.center.len()
///     }
///
///     fn value(&mut self, params: &[f64]) -> Result<f64, GprError> {
///         self.check(params)?;
///         Ok(params.iter().zip(&self.center).map(|(p, c)| (p - c).powi(2)).sum())
///     }
///
///     // The default `fill_intervals` writes `Interval::DEFAULT_POSITIVE`, and
///     // the default `value_at_changes` calls `value`.
/// }
///
/// impl Differentiable for Bowl {
///     fn gradient_into(&mut self, params: &[f64], out: &mut [f64]) -> Result<(), GprError> {
///         self.check(params)?;
///         for ((g, p), c) in out.iter_mut().zip(params).zip(&self.center) {
///             *g = 2.0 * (p - c);
///         }
///         Ok(())
///     }
/// }
///
/// impl TwiceDifferentiable for Bowl {
///     fn hessian_into(&mut self, params: &[f64], out: &mut [f64]) -> Result<(), GprError> {
///         self.check(params)?;
///         let n = params.len();
///         for (k, h) in out.iter_mut().enumerate() {
///             *h = if k / n == k % n { 2.0 } else { 0.0 };
///         }
///         Ok(())
///     }
/// }
///
/// impl IncrementalObjective for Bowl {
///     // Every term depends on one coordinate, so nothing is cached to skip.
///     fn value_with_changes(&mut self, params: &[f64], _indices: &[usize]) -> Result<f64, GprError> {
///         self.value(params)
///     }
/// }
///
/// # fn main() -> Result<(), GprError> {
/// // The minimum is at user values 2 and 0.5, so at their logs.
/// let center = vec![2.0_f64.ln(), 0.5_f64.ln()];
/// let mut bowl = Bowl { center: center.clone() };
/// let init = [0.0, 0.0];
///
/// let mut intervals = [Interval::DEFAULT_POSITIVE; 2];
/// bowl.fill_intervals(&mut intervals)?;
/// assert_eq!(intervals, [Interval::DEFAULT_POSITIVE; 2]);
/// let start = bowl.value(&init)?;
/// assert!((bowl.value_at_changes(&[center[0], 0.0], &[0])? - center[1].powi(2)).abs() < 1e-12);
/// assert!((bowl.value_with_changes(&init, &[0])? - start).abs() < 1e-12);
///
/// let close = |params: &[f64]| params.iter().zip(&center).all(|(p, c)| (p - c).abs() < 1e-3);
/// assert!(close(&NelderMead::new().minimize(&mut bowl, &init)?.params));
/// assert!(close(&Lbfgs::new().minimize(&mut bowl, &init)?.params));
/// assert!(close(&TrustRegion::new().minimize(&mut bowl, &init)?.params));
/// // Annealing is a random search: it keeps the best point it visited.
/// let annealed = FastSimulatedAnnealing::new().minimize(&mut bowl, &init)?;
/// assert!(annealed.value <= start);
/// # Ok(())
/// # }
/// ```
pub trait Objective {
    /// Returns the concatenated parameter count.
    ///
    /// See the example on [`Objective`].
    fn num_params(&self) -> usize;

    /// Returns the objective at `params`.
    ///
    /// # Errors
    ///
    /// An implementation chooses its errors. The exact GPR fit objective
    /// returns [`GprError::LengthMismatch`] when `params` is not
    /// [`Self::num_params`] long, [`GprError::InvalidHyperparameter`] or
    /// [`GprError::InvalidNoiseVariance`] when a coordinate is not finite or
    /// its `exp` overflows or underflows, [`GprError::InvalidInterval`] when a
    /// value leaves its [`Interval`], and [`GprError::CholeskyFailed`] when
    /// `A = K + σn² I` does not factor at `params`.
    ///
    /// See the example on [`Objective`].
    fn value(&mut self, params: &[f64]) -> Result<f64, GprError>;

    /// Returns the objective after the coordinates in `indices` changed.
    ///
    /// `indices` lists **every** coordinate of `params` that differs from
    /// the `params` of the previous evaluation on this objective, not from
    /// the optimizer's accepted point. After a rejected proposal, the next
    /// step lists the coordinate it reverted as well.
    ///
    /// The default rebuilds everything through [`Self::value`]. The GPR fit
    /// objective rebuilds only the touched kernel leaves when the optimizer
    /// sets [`crate::Optimizer::USES_CHANGE_INDICES`] and the fit keeps a
    /// dedicated `W` ([`crate::CholeskyBuffer::Retain`]).
    ///
    /// # Errors
    ///
    /// Same as [`Self::value`]. The exact GPR fit objective, when it
    /// rebuilds only the touched leaves, also returns
    /// [`GprError::IndexOutOfRange`] for an empty, duplicate, or out-of-range
    /// `indices`, and for a step that changed a coordinate `indices` does not
    /// list.
    ///
    /// See the example on [`Objective`].
    fn value_at_changes(&mut self, params: &[f64], indices: &[usize]) -> Result<f64, GprError> {
        let _ = indices;
        self.value(params)
    }

    /// Writes the [`Interval`] of each parameter, in user units, into `out` (length [`Self::num_params`]).
    ///
    /// The built-in optimizers search inside these intervals, so a user
    /// optimizer can read them too and call a built-in one on the same
    /// objective. The GPR fit objectives write the intervals the kernel and
    /// likelihood declare. The default writes [`Interval::DEFAULT_POSITIVE`]
    /// for every parameter.
    ///
    /// # Errors
    ///
    /// Returns [`GprError::LengthMismatch`] when `out.len()` is not
    /// [`Self::num_params`].
    ///
    /// See the example on [`Objective`].
    fn fill_intervals(&self, out: &mut [Interval]) -> Result<(), GprError> {
        crate::data::require_count(out.len(), self.num_params(), "parameters")?;
        out.fill(Interval::DEFAULT_POSITIVE);
        Ok(())
    }
}

/// Represents the first-order objective.
///
/// [`Objective`] is its supertrait: an implementation also implements
/// [`Objective`].
///
/// The GPR fit objective overrides [`Self::value_and_gradient_into`] so one
/// Cholesky produces `L`, `α`, and `W`.
///
/// # Examples
///
/// ```rust
/// use gprx::{Differentiable, Objective};
///
/// struct Zero;
///
/// impl Objective for Zero {
///     fn num_params(&self) -> usize {
///         1
///     }
///
///     fn value(&mut self, _params: &[f64]) -> Result<f64, gprx::GprError> {
///         Ok(0.0)
///     }
/// }
///
/// impl Differentiable for Zero {
///     fn gradient_into(
///         &mut self,
///         _params: &[f64],
///         out: &mut [f64],
///     ) -> Result<(), gprx::GprError> {
///         out.fill(0.0);
///         Ok(())
///     }
/// }
///
/// # fn main() -> Result<(), gprx::GprError> {
/// let mut objective = Zero;
/// let mut grad = [1.0];
/// objective.gradient_into(&[0.0], &mut grad)?;
/// assert_eq!(grad, [0.0]);
/// // The default joint call runs `value`, then `gradient_into`.
/// let value = objective.value_and_gradient_into(&[0.0], &mut grad)?;
/// assert_eq!((value, grad), (0.0, [0.0]));
/// # Ok(())
/// # }
/// ```
pub trait Differentiable: Objective {
    /// Writes `∂L/∂θ` into `out`.
    ///
    /// # Errors
    ///
    /// Same as [`Objective::value`]. The exact GPR fit objective also
    /// returns [`GprError::LengthMismatch`] when `out` is not
    /// [`Objective::num_params`] long.
    ///
    /// See the example on [`Differentiable`].
    fn gradient_into(&mut self, params: &[f64], out: &mut [f64]) -> Result<(), GprError>;

    /// Returns the value and writes the gradient in one evaluation.
    ///
    /// The default calls [`Objective::value`] then [`Self::gradient_into`].
    ///
    /// # Errors
    ///
    /// Same as [`Self::gradient_into`].
    ///
    /// See the example on [`Differentiable`].
    fn value_and_gradient_into(
        &mut self,
        params: &[f64],
        out: &mut [f64],
    ) -> Result<f64, GprError> {
        let value = self.value(params)?;
        self.gradient_into(params, out)?;
        Ok(value)
    }
}

/// Represents the second-order objective.
///
/// [`Differentiable`] is its supertrait: an implementation also implements
/// [`Differentiable`] and [`Objective`].
///
/// The exact GPR fit objective implements it with the analytic Hessian that
/// [`crate::FittedGpr::hessian_into`] also returns, so [`crate::TrustRegion`]
/// takes it at compile time; there is no runtime "not implemented" error.
///
/// The second example on [`Objective`] minimizes an implementation with
/// [`crate::TrustRegion`].
///
/// # Examples
///
/// ```rust
/// use gprx::{Differentiable, Objective, TwiceDifferentiable};
///
/// struct Zero;
///
/// impl Objective for Zero {
///     fn num_params(&self) -> usize {
///         1
///     }
///
///     fn value(&mut self, _params: &[f64]) -> Result<f64, gprx::GprError> {
///         Ok(0.0)
///     }
/// }
///
/// impl Differentiable for Zero {
///     fn gradient_into(
///         &mut self,
///         _params: &[f64],
///         out: &mut [f64],
///     ) -> Result<(), gprx::GprError> {
///         out.fill(0.0);
///         Ok(())
///     }
/// }
///
/// impl TwiceDifferentiable for Zero {
///     fn hessian_into(
///         &mut self,
///         _params: &[f64],
///         out: &mut [f64],
///     ) -> Result<(), gprx::GprError> {
///         out.fill(0.0);
///         Ok(())
///     }
/// }
///
/// # fn main() -> Result<(), gprx::GprError> {
/// let mut objective = Zero;
/// let mut hess = [1.0];
/// objective.hessian_into(&[0.0], &mut hess)?;
/// assert_eq!(hess, [0.0]);
/// // The default joint call runs the value, gradient, and Hessian in turn.
/// let mut grad = [1.0];
/// let value = objective.value_gradient_hessian_into(&[0.0], &mut grad, &mut hess)?;
/// assert_eq!((value, grad, hess), (0.0, [0.0], [0.0]));
/// # Ok(())
/// # }
/// ```
pub trait TwiceDifferentiable: Differentiable {
    /// Writes the Hessian (row-major `n×n`) into `out`.
    ///
    /// # Errors
    ///
    /// Same as [`Objective::value`]. The exact GPR fit objective also
    /// returns [`GprError::LengthMismatch`] when `out` is not `n²` long.
    ///
    /// See the example on [`TwiceDifferentiable`].
    fn hessian_into(&mut self, params: &[f64], out: &mut [f64]) -> Result<(), GprError>;

    /// Returns the value at `params` and writes the gradient into `grad` and the Hessian (row-major `n×n`) into `hess`.
    ///
    /// A second-order solver evaluates all three at each candidate. The
    /// default calls [`Differentiable::value_and_gradient_into`] then
    /// [`Self::hessian_into`]. Override it when the three share work, as
    /// the GPR fit objective does: one factorization serves all three.
    ///
    /// # Errors
    ///
    /// Same as [`Differentiable::value_and_gradient_into`] and
    /// [`Self::hessian_into`].
    ///
    /// See the example on [`TwiceDifferentiable`].
    fn value_gradient_hessian_into(
        &mut self,
        params: &[f64],
        grad: &mut [f64],
        hess: &mut [f64],
    ) -> Result<f64, GprError> {
        let value = self.value_and_gradient_into(params, grad)?;
        self.hessian_into(params, hess)?;
        Ok(value)
    }
}

/// Represents the partial kernel rebuild from changed parameter indices.
///
/// `indices` is every flat `θ` position that changed since the previous
/// evaluation on this objective (see [`Objective::value_at_changes`]). Only
/// those positions decide which leaves are rebuilt; changes are never
/// inferred from the numbers. Empty, duplicate, or out-of-range indices, and
/// an unlisted change, are a [`GprError`] at this boundary. Full rebuilds use
/// [`Objective::value`]. The GPR fit objective implements this for every
/// optimizer and buffer policy.
///
/// # Examples
///
/// ```rust
/// use gprx::{IncrementalObjective, Objective};
///
/// struct Zero;
///
/// impl Objective for Zero {
///     fn num_params(&self) -> usize {
///         1
///     }
///
///     fn value(&mut self, _params: &[f64]) -> Result<f64, gprx::GprError> {
///         Ok(0.0)
///     }
/// }
///
/// impl IncrementalObjective for Zero {
///     fn value_with_changes(
///         &mut self,
///         _params: &[f64],
///         _indices: &[usize],
///     ) -> Result<f64, gprx::GprError> {
///         Ok(0.0)
///     }
/// }
///
/// # fn main() -> Result<(), gprx::GprError> {
/// let mut objective = Zero;
/// assert_eq!(objective.value_with_changes(&[0.0], &[0])?, 0.0);
/// # Ok(())
/// # }
/// ```
pub trait IncrementalObjective: Objective {
    /// Returns the objective after rebuilding only the leaves that `indices` touch.
    ///
    /// # Errors
    ///
    /// Same as [`Objective::value`]. The exact GPR fit objective also
    /// returns [`GprError::IndexOutOfRange`] when `indices` is empty,
    /// contains a duplicate or an `i >= n_params`, or leaves out a coordinate
    /// that changed since the previous evaluation.
    ///
    /// See the example on [`IncrementalObjective`].
    fn value_with_changes(&mut self, params: &[f64], indices: &[usize]) -> Result<f64, GprError>;
}

#[cfg(test)]
mod interval_tests {
    use super::Objective;
    use crate::error::GprError;
    use crate::param::Interval;

    struct Plain;

    impl Objective for Plain {
        fn num_params(&self) -> usize {
            2
        }

        fn value(&mut self, params: &[f64]) -> Result<f64, GprError> {
            Ok(params.iter().map(|p| p * p).sum())
        }
    }

    #[test]
    fn default_intervals_are_the_default_positive_range() {
        let mut out = [Interval::new(1.0, 2.0).expect("valid"); 2];
        Plain.fill_intervals(&mut out).expect("len 2");
        assert_eq!(out, [Interval::DEFAULT_POSITIVE; 2]);
        let mut short = [Interval::DEFAULT_POSITIVE; 1];
        assert!(matches!(
            Plain.fill_intervals(&mut short),
            Err(GprError::LengthMismatch { .. })
        ));
    }
}