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
//! Trainer for stochastic variational GPR.

use std::marker::PhantomData;

use crate::error::GprError;
use crate::policy::{JitterPolicy, KernelExp, with_kernel_exp};
use crate::sparse::{SparseCore, SparseSpec};
use crate::transform::{UnfittedTarget, UnfittedTransform};

use crate::kernel::{KernelSpec, ModelKernel, ModelKernelParts, PointKernel};
use crate::likelihood::GaussianLikelihood;
use crate::optimizer::{Adam, Fixed};
use crate::precision::{DoublePrecision, GpScalar};

use super::factor::{assemble_fitted, run_adam_fit};
use super::fitted::FittedSvgp;

/// Represents the trainer for stochastic variational GPR at an inducing set the caller chooses.
///
/// The type parameters are the optimizer `O` (default [`Fixed`]; [`Adam`]
/// for [`Svgp<Adam>::fit`]), the precision `P` (default
/// [`DoublePrecision`]), and the kernel `K` (default [`KernelSpec`]).
/// [`Svgp<Fixed>::factor`] prepares `K_mm` and a whitened prior `q(u)`
/// (`mean = 0`, `L = I`) at the current kernel and likelihood `θ`.
/// [`Svgp<Adam>::fit`] starts from that prior and runs mini-batch Adam.
/// [`Self::with_optimizer`] switches the type parameter; there is no fit
/// flag.
///
/// `K` selects how the inducing points are given. A [`KernelSpec`] model
/// takes their coordinates `Z` (column-major, `m` points by `d` features)
/// next to the training `X`; `Z` is not a parameter. A
/// [`crate::kernel::DistanceKernel`] model has no coordinates to place them
/// at: its inducing points are training points, named by their indices,
/// and the caller supplies per slot the `n × m` squared distances from the
/// training points to them (`K_mm` reads the inducing points' rows of that
/// block). A prediction then takes the `m × q` squared distances from the
/// inducing points to the queries.
///
/// # Examples
///
/// ```rust
/// use gprx::kernel::{KernelSpec, RbfKernel};
/// use gprx::{GaussianLikelihood, JitterPolicy, KernelExp, SinglePrecision, Svgp};
///
/// # fn main() -> Result<(), gprx::GprError> {
/// let kernel = KernelSpec::from(RbfKernel::new(1.0)?);
/// let likelihood = GaussianLikelihood::new(0.1)?;
/// let mut svgp = Svgp::new(kernel, likelihood).with_math(KernelExp::FastApprox);
/// assert_eq!(svgp.math(), KernelExp::FastApprox);
/// assert!(matches!(svgp.jitter_policy(), JitterPolicy::Adaptive(_)));
/// // θ is the kernel's log-lengthscale, then the likelihood's log σn²; `Z` and `q` are not in it.
/// assert_eq!(svgp.num_params(), svgp.kernel().num_params() + 1);
/// let mut theta = [0.0; 2];
/// svgp.get_params(&mut theta)?;
/// theta[0] = 0.5_f64.ln();
/// svgp.set_params(&theta)?;
/// assert!((svgp.likelihood().log_noise_variance() - theta[1]).abs() < 1e-12);
///
/// // Column-major `X` (n = 4, d = 1) and the inducing coordinates `Z` (m = 2).
/// let (x, y, z) = ([0.0, 1.0, 2.0, 3.0], [0.0, 1.0, 0.5, 0.25], [0.5, 2.5]);
/// let fitted = svgp.clone().factor(&x, 4, 1, &y, &z, 2).map_err(|(_, e)| e)?;
/// assert_eq!(fitted.n(), 4);
/// assert_eq!(fitted.m(), 2);
///
/// // The same model with `f32` storage.
/// let single = svgp
///     .with_precision::<SinglePrecision>()
///     .factor(&x, 4, 1, &y, &z, 2)
///     .map_err(|(_, e)| e)?;
/// assert_eq!(single.predict(&[1.5], 1, 1)?.mean.len(), 1);
/// # Ok(())
/// # }
/// ```
#[derive(Clone, Debug)]
pub struct Svgp<O = Fixed, P = DoublePrecision, K: ModelKernel = KernelSpec> {
    pub(super) spec: SparseSpec<K::Supply>,
    pub(super) optimizer: O,
    pub(super) _precision: PhantomData<P>,
    pub(super) _kernel: PhantomData<K>,
}

impl<K: ModelKernel> Svgp<Fixed, DoublePrecision, K> {
    /// Builds a trainer with the current kernel `θ` and [`Fixed`].
    ///
    /// Inducing coordinates are an argument of [`Svgp<Fixed>::factor`], not
    /// of this constructor. The whitened prior is allocated at `factor` from
    /// the inducing count.
    ///
    /// See the example on [`Svgp`].
    pub fn new(kernel: K, likelihood: GaussianLikelihood) -> Self {
        Self {
            spec: SparseSpec::new(<K as ModelKernelParts>::into_spec(kernel), likelihood),
            optimizer: Fixed,
            _precision: PhantomData,
            _kernel: PhantomData,
        }
    }
}

impl<O, P, K: ModelKernel> Svgp<O, P, K> {
    /// The same settings under new type parameters, with `map` applied to
    /// the optimizer.
    fn retype<O2, P2>(self, map: impl FnOnce(O) -> O2) -> Svgp<O2, P2, K> {
        Svgp {
            spec: self.spec,
            optimizer: map(self.optimizer),
            _precision: PhantomData,
            _kernel: PhantomData,
        }
    }

    /// Replaces the optimizer type parameter.
    ///
    /// [`Fixed`] keeps [`Svgp<Fixed>::factor`]. [`Adam`] enables
    /// [`Svgp<Adam>::fit`]. [`Adam`] is not an [`crate::Optimizer`].
    ///
    /// # Examples
    ///
    /// ```rust
    /// use gprx::kernel::{KernelSpec, RbfKernel};
    /// use gprx::{Adam, GaussianLikelihood, Svgp};
    ///
    /// # fn main() -> Result<(), gprx::GprError> {
    /// let kernel = KernelSpec::from(RbfKernel::new(1.0)?);
    /// let likelihood = GaussianLikelihood::new(0.1)?;
    /// let fitted = Svgp::new(kernel, likelihood)
    ///     .with_optimizer(Adam::new())
    ///     .fit(&[0.0, 1.0, 2.0, 3.0], 4, 1, &[0.0, 1.0, 0.5, 0.25], &[0.5, 2.5], 2)
    ///     .map_err(|(_, e)| e)?;
    /// assert!(fitted.neg_elbo()?.is_finite());
    /// # Ok(())
    /// # }
    /// ```
    pub fn with_optimizer<O2>(self, optimizer: O2) -> Svgp<O2, P, K> {
        self.retype(|_| optimizer)
    }

    /// Selects the storage precision.
    ///
    /// Omitting it leaves [`DoublePrecision`].
    ///
    /// See the example on [`Svgp`].
    pub fn with_precision<P2: GpScalar>(self) -> Svgp<O, P2, K> {
        self.retype(|optimizer| optimizer)
    }

    /// Selects the kernel `exp`.
    ///
    /// Omitting it leaves [`KernelExp::Accurate`].
    ///
    /// `factor`, `fit`, and predict use the same `exp`. Hyperparameter
    /// `exp(θ)` is unchanged.
    ///
    /// See the example on [`Svgp`].
    pub fn with_math(mut self, math: KernelExp) -> Self {
        self.spec.math = math;
        self
    }

    /// Returns the kernel `exp` mode.
    ///
    /// See the example on [`Svgp`].
    pub fn math(&self) -> KernelExp {
        self.spec.math
    }

    /// Replaces the jitter retries for factoring `K_mm = k(Z, Z)`.
    ///
    /// The default is `JitterPolicy::adaptive(1e-8, 10.0, 5, 1e-3)`, not
    /// the [`JitterPolicy::default`] (no retry) of [`crate::Gpr`]: the
    /// Exact system `K + σn² I` carries the observation noise, `K_mm`
    /// does not, so close inducing points leave it singular in floating
    /// point. The policy applies wherever `K_mm` is factored: fit, factor,
    /// `set_params`, and predict.
    ///
    /// # Examples
    ///
    /// ```rust
    /// use gprx::kernel::{KernelSpec, RbfKernel};
    /// use gprx::{GaussianLikelihood, JitterPolicy, Svgp};
    ///
    /// # fn main() -> Result<(), gprx::GprError> {
    /// let policy = JitterPolicy::fixed(1e-6)?;
    /// let fitted = Svgp::new(
    ///     KernelSpec::from(RbfKernel::new(1.0)?),
    ///     GaussianLikelihood::new(0.1)?,
    /// )
    /// .with_jitter_policy(policy)
    /// .factor(&[0.0, 1.0, 2.0, 3.0], 4, 1, &[0.0, 1.0, 0.5, 0.25], &[0.5, 2.5], 2)
    /// .map_err(|(_, e)| e)?;
    /// assert_eq!(fitted.jitter_policy(), policy);
    /// # Ok(())
    /// # }
    /// ```
    pub fn with_jitter_policy(mut self, policy: JitterPolicy) -> Self {
        self.spec.jitter = policy;
        self
    }

    /// Returns the jitter retries for factoring `K_mm`.
    ///
    /// See the example on [`Svgp`].
    pub fn jitter_policy(&self) -> JitterPolicy {
        self.spec.jitter
    }

    /// Replaces the target (`y`) transform.
    ///
    /// Omitting it leaves identity.
    ///
    /// The map is fitted on training `y`. Predictions are mapped back to the
    /// original scale. The negative ELBO and its gradient are those of the
    /// transformed `y`. A single map or a
    /// [`crate::transform::TargetPipeline`].
    ///
    /// # Examples
    ///
    /// ```rust
    /// use gprx::kernel::{KernelSpec, RbfKernel};
    /// use gprx::transform::StandardizeTarget;
    /// use gprx::{GaussianLikelihood, Svgp};
    ///
    /// # fn main() -> Result<(), gprx::GprError> {
    /// let fitted = Svgp::new(
    ///     KernelSpec::from(RbfKernel::new(1.0)?),
    ///     GaussianLikelihood::new(0.1)?,
    /// )
    /// .with_target_transform(StandardizeTarget::new())
    /// .factor(&[0.0, 1.0, 2.0, 3.0], 4, 1, &[100.0, 101.0, 100.5, 100.25], &[0.5, 2.5], 2)
    /// .map_err(|(_, e)| e)?;
    /// let pred = fitted.predict(&[1.0], 1, 1)?;
    /// assert!(pred.mean[0] > 99.0);
    /// # Ok(())
    /// # }
    /// ```
    pub fn with_target_transform(mut self, transform: impl UnfittedTarget + 'static) -> Self {
        self.spec.y_transform = Box::new(transform);
        self
    }

    /// Returns the observation-noise model.
    ///
    /// See the example on [`Svgp`].
    pub fn likelihood(&self) -> &GaussianLikelihood {
        &self.spec.likelihood
    }

    /// Returns the concatenated kernel and likelihood parameter count.
    ///
    /// Inducing coordinates and the variational posterior are not counted.
    ///
    /// See the example on [`Svgp`].
    pub fn num_params(&self) -> usize {
        self.spec.theta_len()
    }

    /// Writes kernel `θ` then likelihood `θ` into `out`.
    ///
    /// # Errors
    ///
    /// Returns [`GprError::LengthMismatch`] if `out` is the wrong length
    /// or a custom leaf rejects the write.
    ///
    /// See the example on [`Svgp`].
    pub fn get_params(&self, out: &mut [f64]) -> Result<(), GprError> {
        self.spec.read_theta(out)
    }

    /// Sets kernel then likelihood `θ` without forming the SVGP system.
    ///
    /// `params` is kernel parameters followed by the likelihood parameter,
    /// matching [`Self::get_params`]. Inducing coordinates are not in this
    /// slice.
    ///
    /// # Errors
    ///
    /// Returns [`GprError::LengthMismatch`] if `params` is the wrong
    /// length, [`GprError::InvalidHyperparameter`] or
    /// [`GprError::InvalidNoiseVariance`] if a kernel or likelihood `θ` is
    /// not finite or its `exp` is not a positive finite value, or
    /// [`GprError::InvalidInterval`] if a value is outside the interval of
    /// its parameter (see [`crate::BoundedParam`] and
    /// [`GaussianLikelihood::bounds`]). A custom kernel leaf returns its own
    /// error. Kernel and likelihood `θ` are committed together only after
    /// both writes succeed.
    ///
    /// See the example on [`Svgp`].
    pub fn set_params(&mut self, params: &[f64]) -> Result<(), GprError> {
        self.spec.write_theta(params)
    }
}

impl<O, P, K: PointKernel> Svgp<O, P, K> {
    /// Replaces the input (`X`) transform.
    ///
    /// Omitting it leaves identity.
    ///
    /// The map is fitted on training `X`. `X`, the inducing points `Z`, and
    /// every later query go through it, so `Z` is passed in the same
    /// coordinates as `X`; the fitted model reports `Z` in the original
    /// ones. A single map, a [`crate::transform::Pipeline`], or
    /// [`crate::transform::ColumnwiseInput`].
    ///
    /// # Examples
    ///
    /// ```rust
    /// use gprx::kernel::{KernelSpec, RbfKernel};
    /// use gprx::transform::StandardizeInput;
    /// use gprx::{GaussianLikelihood, Svgp};
    ///
    /// # fn main() -> Result<(), gprx::GprError> {
    /// let fitted = Svgp::new(
    ///     KernelSpec::from(RbfKernel::new(1.0)?),
    ///     GaussianLikelihood::new(0.1)?,
    /// )
    /// .with_input_transform(StandardizeInput::new())
    /// .factor(&[0.0, 10.0, 20.0, 30.0], 4, 1, &[0.0, 1.0, 0.5, 0.25], &[5.0, 25.0], 2)
    /// .map_err(|(_, e)| e)?;
    /// assert_eq!(fitted.z(), &[5.0, 25.0]);
    /// # Ok(())
    /// # }
    /// ```
    pub fn with_input_transform(mut self, transform: impl UnfittedTransform + 'static) -> Self {
        self.spec.x_transform = Box::new(transform);
        self
    }
}

impl<O, P> Svgp<O, P> {
    /// Returns the kernel whose hyperparameters this trainer owns.
    ///
    /// See the example on [`Svgp`].
    pub fn kernel(&self) -> &KernelSpec {
        &self.spec.kernel
    }
}

impl<O, P, C: crate::kernel::PointUse> Svgp<O, P, crate::kernel::DistanceKernel<C>> {
    /// Returns a copy of the kernel whose hyperparameters this trainer owns.
    ///
    /// # Examples
    ///
    /// ```rust
    /// use gprx::kernel::{RbfKernel, ScalarDistance};
    /// use gprx::{GaussianLikelihood, Svgp};
    ///
    /// # fn main() -> Result<(), gprx::GprError> {
    /// let image = ScalarDistance::new();
    /// let svgp = Svgp::new(image.kernel(RbfKernel::new(1.0)?), GaussianLikelihood::new(0.1)?);
    /// assert_eq!(svgp.to_kernel().num_params(), 1);
    /// # Ok(())
    /// # }
    /// ```
    pub fn to_kernel(&self) -> crate::kernel::DistanceKernel<C> {
        <crate::kernel::DistanceKernel<C> as ModelKernelParts>::from_spec(self.spec.kernel.clone())
    }
}

impl<P> Svgp<Fixed, P>
where
    P: GpScalar,
{
    /// Factors `K_mm` and installs a whitened prior `q(u)` at the current `θ`.
    ///
    /// `x` and `z` are column-major (`n` / `m` points by `d` features). `Z`
    /// is supplied by the caller and is not a parameter. Likelihood noise is
    /// not added to `K_mm`. The variational mean is zero and the whitened
    /// Cholesky factor is the identity of order `m`.
    ///
    /// # Errors
    ///
    /// Returns [`GprError::EmptyInput`] when `n`, `d`, or `m` is zero.
    /// Returns [`GprError::DimensionMismatch`] when `z` is packed with a
    /// different feature count than `x` (the same `n_cols` is required).
    /// Returns [`GprError::LengthMismatch`] when `x`, `y`, or `z` has the
    /// wrong length, and [`GprError::NonFiniteInput`] when a value is `NaN`
    /// or `Inf`. [`GprError::CholeskyFailed`] when `K_mm` cannot be
    /// factored.
    ///
    /// # Examples
    ///
    /// ```rust
    /// use gprx::kernel::{KernelSpec, RbfKernel};
    /// use gprx::{GaussianLikelihood, Svgp};
    ///
    /// # fn main() -> Result<(), gprx::GprError> {
    /// let kernel = KernelSpec::from(RbfKernel::new(1.0)?);
    /// let likelihood = GaussianLikelihood::new(0.1)?;
    /// let fitted = Svgp::new(kernel, likelihood)
    ///     .factor(&[0.0, 1.0, 2.0, 3.0], 4, 1, &[0.0, 1.0, 0.5, 0.25], &[0.5, 2.5], 2)
    ///     .map_err(|(_, e)| e)?;
    /// assert_eq!(fitted.num_params(), 7);
    /// # Ok(())
    /// # }
    /// ```
    #[allow(clippy::result_large_err)] // failure returns the trainer so the caller can retry
    pub fn factor(
        self,
        x: &[f64],
        n_rows: usize,
        n_cols: usize,
        y: &[f64],
        z: &[f64],
        n_inducing: usize,
    ) -> Result<FittedSvgp<P>, (Self, GprError)> {
        let core = match SparseCore::prepare(&self.spec, x, n_rows, n_cols, y, z, n_inducing) {
            Ok(core) => core,
            Err(err) => return Err((self, err)),
        };
        match with_kernel_exp!(self.spec.math, M => assemble_fitted::<M, _, KernelSpec>(core, None))
        {
            Ok(fitted) => Ok(fitted),
            Err(err) => Err((self, err)),
        }
    }
}

impl<P> Svgp<Adam, P>
where
    P: GpScalar,
{
    /// Factors a whitened prior `q` and runs mini-batch Adam on `θ` and `q`.
    ///
    /// Starts from the same prior as [`Svgp<Fixed>::factor`]. Kernel and
    /// likelihood `θ` move in the existing [`crate::Interval`] logit.
    /// The diagonal of `L` moves in log space without that interval. The
    /// variational mean and the off-diagonal of `L` stay in user units.
    /// Each epoch shuffles the training indices. A remainder batch is kept.
    /// When `batch_size ≥ n` the epoch is one full-data step. Public
    /// [`FittedSvgp::value_and_gradient_into`] stays a full-data sum; this
    /// loop scales the data term by `n / b_actual` and leaves the KL full.
    ///
    /// # Errors
    ///
    /// Same input errors as `factor` of [`Svgp<Fixed>`].
    /// [`GprError::CholeskyFailed`] when `K_mm` cannot be factored.
    /// [`GprError::InvalidHyperparameter`] / [`GprError::NonFiniteInput`]
    /// when a trial `θ` or `L` is rejected.
    ///
    /// # Examples
    ///
    /// ```rust
    /// use gprx::kernel::{KernelSpec, RbfKernel};
    /// use gprx::{Adam, GaussianLikelihood, Svgp};
    ///
    /// # fn main() -> Result<(), gprx::GprError> {
    /// let kernel = KernelSpec::from(RbfKernel::new(1.0)?);
    /// let likelihood = GaussianLikelihood::new(0.1)?;
    /// let fitted = Svgp::new(kernel, likelihood)
    ///     .with_optimizer(Adam::new())
    ///     .fit(&[0.0, 1.0, 2.0, 3.0], 4, 1, &[0.0, 1.0, 0.5, 0.25], &[0.5, 2.5], 2)
    ///     .map_err(|(_, e)| e)?;
    /// assert_eq!(fitted.n(), 4);
    /// # Ok(())
    /// # }
    /// ```
    #[allow(clippy::result_large_err)]
    pub fn fit(
        self,
        x: &[f64],
        n_rows: usize,
        n_cols: usize,
        y: &[f64],
        z: &[f64],
        n_inducing: usize,
    ) -> Result<FittedSvgp<P>, (Self, GprError)> {
        let core = match SparseCore::prepare(&self.spec, x, n_rows, n_cols, y, z, n_inducing) {
            Ok(core) => core,
            Err(err) => return Err((self, err)),
        };
        match with_kernel_exp!(self.spec.math, M => assemble_fitted::<M, _, KernelSpec>(core, None))
        {
            Ok(mut fitted) => match with_kernel_exp!(
                self.spec.math,
                M => run_adam_fit::<M, _, KernelSpec>(&mut fitted, &self.optimizer)
            ) {
                Ok(()) => Ok(fitted),
                Err(err) => Err((self, err)),
            },
            Err(err) => Err((self, err)),
        }
    }
}