File: gpc.py

package info (click to toggle)
scikit-learn 0.20.2%2Bdfsg-6
  • links: PTS, VCS
  • area: main
  • in suites: buster
  • size: 51,036 kB
  • sloc: python: 108,171; ansic: 8,722; cpp: 5,651; makefile: 192; sh: 40
file content (758 lines) | stat: -rw-r--r-- 32,820 bytes parent folder | download
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
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
"""Gaussian processes classification."""

# Authors: Jan Hendrik Metzen <jhm@informatik.uni-bremen.de>
#
# License: BSD 3 clause

import warnings
from operator import itemgetter

import numpy as np
from scipy.linalg import cholesky, cho_solve, solve
from scipy.optimize import fmin_l_bfgs_b
from scipy.special import erf, expit

from sklearn.base import BaseEstimator, ClassifierMixin, clone
from sklearn.gaussian_process.kernels \
    import RBF, CompoundKernel, ConstantKernel as C
from sklearn.utils.validation import check_X_y, check_is_fitted, check_array
from sklearn.utils import check_random_state
from sklearn.preprocessing import LabelEncoder
from sklearn.multiclass import OneVsRestClassifier, OneVsOneClassifier
from sklearn.exceptions import ConvergenceWarning


# Values required for approximating the logistic sigmoid by
# error functions. coefs are obtained via:
# x = np.array([0, 0.6, 2, 3.5, 4.5, np.inf])
# b = logistic(x)
# A = (erf(np.dot(x, self.lambdas)) + 1) / 2
# coefs = lstsq(A, b)[0]
LAMBDAS = np.array([0.41, 0.4, 0.37, 0.44, 0.39])[:, np.newaxis]
COEFS = np.array([-1854.8214151, 3516.89893646, 221.29346712,
                  128.12323805, -2010.49422654])[:, np.newaxis]


class _BinaryGaussianProcessClassifierLaplace(BaseEstimator):
    """Binary Gaussian process classification based on Laplace approximation.

    The implementation is based on Algorithm 3.1, 3.2, and 5.1 of
    ``Gaussian Processes for Machine Learning'' (GPML) by Rasmussen and
    Williams.

    Internally, the Laplace approximation is used for approximating the
    non-Gaussian posterior by a Gaussian.

    Currently, the implementation is restricted to using the logistic link
    function.

    .. versionadded:: 0.18

    Parameters
    ----------
    kernel : kernel object
        The kernel specifying the covariance function of the GP. If None is
        passed, the kernel "1.0 * RBF(1.0)" is used as default. Note that
        the kernel's hyperparameters are optimized during fitting.

    optimizer : string or callable, optional (default: "fmin_l_bfgs_b")
        Can either be one of the internally supported optimizers for optimizing
        the kernel's parameters, specified by a string, or an externally
        defined optimizer passed as a callable. If a callable is passed, it
        must have the  signature::

            def optimizer(obj_func, initial_theta, bounds):
                # * 'obj_func' is the objective function to be maximized, which
                #   takes the hyperparameters theta as parameter and an
                #   optional flag eval_gradient, which determines if the
                #   gradient is returned additionally to the function value
                # * 'initial_theta': the initial value for theta, which can be
                #   used by local optimizers
                # * 'bounds': the bounds on the values of theta
                ....
                # Returned are the best found hyperparameters theta and
                # the corresponding value of the target function.
                return theta_opt, func_min

        Per default, the 'fmin_l_bfgs_b' algorithm from scipy.optimize
        is used. If None is passed, the kernel's parameters are kept fixed.
        Available internal optimizers are::

            'fmin_l_bfgs_b'

    n_restarts_optimizer: int, optional (default: 0)
        The number of restarts of the optimizer for finding the kernel's
        parameters which maximize the log-marginal likelihood. The first run
        of the optimizer is performed from the kernel's initial parameters,
        the remaining ones (if any) from thetas sampled log-uniform randomly
        from the space of allowed theta-values. If greater than 0, all bounds
        must be finite. Note that n_restarts_optimizer=0 implies that one
        run is performed.

    max_iter_predict: int, optional (default: 100)
        The maximum number of iterations in Newton's method for approximating
        the posterior during predict. Smaller values will reduce computation
        time at the cost of worse results.

    warm_start : bool, optional (default: False)
        If warm-starts are enabled, the solution of the last Newton iteration
        on the Laplace approximation of the posterior mode is used as
        initialization for the next call of _posterior_mode(). This can speed
        up convergence when _posterior_mode is called several times on similar
        problems as in hyperparameter optimization. See :term:`the Glossary
        <warm_start>`.

    copy_X_train : bool, optional (default: True)
        If True, a persistent copy of the training data is stored in the
        object. Otherwise, just a reference to the training data is stored,
        which might cause predictions to change if the data is modified
        externally.

    random_state : int, RandomState instance or None, optional (default: None)
        The generator used to initialize the centers. If int, random_state is
        the seed used by the random number generator; If RandomState instance,
        random_state is the random number generator; If None, the random number
        generator is the RandomState instance used by `np.random`.

    Attributes
    ----------
    X_train_ : array-like, shape = (n_samples, n_features)
        Feature values in training data (also required for prediction)

    y_train_ : array-like, shape = (n_samples,)
        Target values in training data (also required for prediction)

    classes_ : array-like, shape = (n_classes,)
        Unique class labels.

    kernel_ : kernel object
        The kernel used for prediction. The structure of the kernel is the
        same as the one passed as parameter but with optimized hyperparameters

    L_ : array-like, shape = (n_samples, n_samples)
        Lower-triangular Cholesky decomposition of the kernel in X_train_

    pi_ : array-like, shape = (n_samples,)
        The probabilities of the positive class for the training points
        X_train_

    W_sr_ : array-like, shape = (n_samples,)
        Square root of W, the Hessian of log-likelihood of the latent function
        values for the observed labels. Since W is diagonal, only the diagonal
        of sqrt(W) is stored.

    log_marginal_likelihood_value_ : float
        The log-marginal-likelihood of ``self.kernel_.theta``

    """
    def __init__(self, kernel=None, optimizer="fmin_l_bfgs_b",
                 n_restarts_optimizer=0, max_iter_predict=100,
                 warm_start=False, copy_X_train=True, random_state=None):
        self.kernel = kernel
        self.optimizer = optimizer
        self.n_restarts_optimizer = n_restarts_optimizer
        self.max_iter_predict = max_iter_predict
        self.warm_start = warm_start
        self.copy_X_train = copy_X_train
        self.random_state = random_state

    def fit(self, X, y):
        """Fit Gaussian process classification model

        Parameters
        ----------
        X : array-like, shape = (n_samples, n_features)
            Training data

        y : array-like, shape = (n_samples,)
            Target values, must be binary

        Returns
        -------
        self : returns an instance of self.
        """
        if self.kernel is None:  # Use an RBF kernel as default
            self.kernel_ = C(1.0, constant_value_bounds="fixed") \
                * RBF(1.0, length_scale_bounds="fixed")
        else:
            self.kernel_ = clone(self.kernel)

        self.rng = check_random_state(self.random_state)

        self.X_train_ = np.copy(X) if self.copy_X_train else X

        # Encode class labels and check that it is a binary classification
        # problem
        label_encoder = LabelEncoder()
        self.y_train_ = label_encoder.fit_transform(y)
        self.classes_ = label_encoder.classes_
        if self.classes_.size > 2:
            raise ValueError("%s supports only binary classification. "
                             "y contains classes %s"
                             % (self.__class__.__name__, self.classes_))
        elif self.classes_.size == 1:
            raise ValueError("{0:s} requires 2 classes; got {1:d} class"
                             .format(self.__class__.__name__,
                                     self.classes_.size))

        if self.optimizer is not None and self.kernel_.n_dims > 0:
            # Choose hyperparameters based on maximizing the log-marginal
            # likelihood (potentially starting from several initial values)
            def obj_func(theta, eval_gradient=True):
                if eval_gradient:
                    lml, grad = self.log_marginal_likelihood(
                        theta, eval_gradient=True)
                    return -lml, -grad
                else:
                    return -self.log_marginal_likelihood(theta)

            # First optimize starting from theta specified in kernel
            optima = [self._constrained_optimization(obj_func,
                                                     self.kernel_.theta,
                                                     self.kernel_.bounds)]

            # Additional runs are performed from log-uniform chosen initial
            # theta
            if self.n_restarts_optimizer > 0:
                if not np.isfinite(self.kernel_.bounds).all():
                    raise ValueError(
                        "Multiple optimizer restarts (n_restarts_optimizer>0) "
                        "requires that all bounds are finite.")
                bounds = self.kernel_.bounds
                for iteration in range(self.n_restarts_optimizer):
                    theta_initial = np.exp(self.rng.uniform(bounds[:, 0],
                                                            bounds[:, 1]))
                    optima.append(
                        self._constrained_optimization(obj_func, theta_initial,
                                                       bounds))
            # Select result from run with minimal (negative) log-marginal
            # likelihood
            lml_values = list(map(itemgetter(1), optima))
            self.kernel_.theta = optima[np.argmin(lml_values)][0]
            self.log_marginal_likelihood_value_ = -np.min(lml_values)
        else:
            self.log_marginal_likelihood_value_ = \
                self.log_marginal_likelihood(self.kernel_.theta)

        # Precompute quantities required for predictions which are independent
        # of actual query points
        K = self.kernel_(self.X_train_)

        _, (self.pi_, self.W_sr_, self.L_, _, _) = \
            self._posterior_mode(K, return_temporaries=True)

        return self

    def predict(self, X):
        """Perform classification on an array of test vectors X.

        Parameters
        ----------
        X : array-like, shape = (n_samples, n_features)

        Returns
        -------
        C : array, shape = (n_samples,)
            Predicted target values for X, values are from ``classes_``
        """
        check_is_fitted(self, ["X_train_", "y_train_", "pi_", "W_sr_", "L_"])

        # As discussed on Section 3.4.2 of GPML, for making hard binary
        # decisions, it is enough to compute the MAP of the posterior and
        # pass it through the link function
        K_star = self.kernel_(self.X_train_, X)  # K_star =k(x_star)
        f_star = K_star.T.dot(self.y_train_ - self.pi_)  # Algorithm 3.2,Line 4

        return np.where(f_star > 0, self.classes_[1], self.classes_[0])

    def predict_proba(self, X):
        """Return probability estimates for the test vector X.

        Parameters
        ----------
        X : array-like, shape = (n_samples, n_features)

        Returns
        -------
        C : array-like, shape = (n_samples, n_classes)
            Returns the probability of the samples for each class in
            the model. The columns correspond to the classes in sorted
            order, as they appear in the attribute ``classes_``.
        """
        check_is_fitted(self, ["X_train_", "y_train_", "pi_", "W_sr_", "L_"])

        # Based on Algorithm 3.2 of GPML
        K_star = self.kernel_(self.X_train_, X)  # K_star =k(x_star)
        f_star = K_star.T.dot(self.y_train_ - self.pi_)  # Line 4
        v = solve(self.L_, self.W_sr_[:, np.newaxis] * K_star)  # Line 5
        # Line 6 (compute np.diag(v.T.dot(v)) via einsum)
        var_f_star = self.kernel_.diag(X) - np.einsum("ij,ij->j", v, v)

        # Line 7:
        # Approximate \int log(z) * N(z | f_star, var_f_star)
        # Approximation is due to Williams & Barber, "Bayesian Classification
        # with Gaussian Processes", Appendix A: Approximate the logistic
        # sigmoid by a linear combination of 5 error functions.
        # For information on how this integral can be computed see
        # blitiri.blogspot.de/2012/11/gaussian-integral-of-error-function.html
        alpha = 1 / (2 * var_f_star)
        gamma = LAMBDAS * f_star
        integrals = np.sqrt(np.pi / alpha) \
            * erf(gamma * np.sqrt(alpha / (alpha + LAMBDAS**2))) \
            / (2 * np.sqrt(var_f_star * 2 * np.pi))
        pi_star = (COEFS * integrals).sum(axis=0) + .5 * COEFS.sum()

        return np.vstack((1 - pi_star, pi_star)).T

    def log_marginal_likelihood(self, theta=None, eval_gradient=False):
        """Returns log-marginal likelihood of theta for training data.

        Parameters
        ----------
        theta : array-like, shape = (n_kernel_params,) or None
            Kernel hyperparameters for which the log-marginal likelihood is
            evaluated. If None, the precomputed log_marginal_likelihood
            of ``self.kernel_.theta`` is returned.

        eval_gradient : bool, default: False
            If True, the gradient of the log-marginal likelihood with respect
            to the kernel hyperparameters at position theta is returned
            additionally. If True, theta must not be None.

        Returns
        -------
        log_likelihood : float
            Log-marginal likelihood of theta for training data.

        log_likelihood_gradient : array, shape = (n_kernel_params,), optional
            Gradient of the log-marginal likelihood with respect to the kernel
            hyperparameters at position theta.
            Only returned when eval_gradient is True.
        """
        if theta is None:
            if eval_gradient:
                raise ValueError(
                    "Gradient can only be evaluated for theta!=None")
            return self.log_marginal_likelihood_value_

        kernel = self.kernel_.clone_with_theta(theta)

        if eval_gradient:
            K, K_gradient = kernel(self.X_train_, eval_gradient=True)
        else:
            K = kernel(self.X_train_)

        # Compute log-marginal-likelihood Z and also store some temporaries
        # which can be reused for computing Z's gradient
        Z, (pi, W_sr, L, b, a) = \
            self._posterior_mode(K, return_temporaries=True)

        if not eval_gradient:
            return Z

        # Compute gradient based on Algorithm 5.1 of GPML
        d_Z = np.empty(theta.shape[0])
        # XXX: Get rid of the np.diag() in the next line
        R = W_sr[:, np.newaxis] * cho_solve((L, True), np.diag(W_sr))  # Line 7
        C = solve(L, W_sr[:, np.newaxis] * K)  # Line 8
        # Line 9: (use einsum to compute np.diag(C.T.dot(C))))
        s_2 = -0.5 * (np.diag(K) - np.einsum('ij, ij -> j', C, C)) \
            * (pi * (1 - pi) * (1 - 2 * pi))  # third derivative

        for j in range(d_Z.shape[0]):
            C = K_gradient[:, :, j]   # Line 11
            # Line 12: (R.T.ravel().dot(C.ravel()) = np.trace(R.dot(C)))
            s_1 = .5 * a.T.dot(C).dot(a) - .5 * R.T.ravel().dot(C.ravel())

            b = C.dot(self.y_train_ - pi)  # Line 13
            s_3 = b - K.dot(R.dot(b))  # Line 14

            d_Z[j] = s_1 + s_2.T.dot(s_3)  # Line 15

        return Z, d_Z

    def _posterior_mode(self, K, return_temporaries=False):
        """Mode-finding for binary Laplace GPC and fixed kernel.

        This approximates the posterior of the latent function values for given
        inputs and target observations with a Gaussian approximation and uses
        Newton's iteration to find the mode of this approximation.
        """
        # Based on Algorithm 3.1 of GPML

        # If warm_start are enabled, we reuse the last solution for the
        # posterior mode as initialization; otherwise, we initialize with 0
        if self.warm_start and hasattr(self, "f_cached") \
           and self.f_cached.shape == self.y_train_.shape:
            f = self.f_cached
        else:
            f = np.zeros_like(self.y_train_, dtype=np.float64)

        # Use Newton's iteration method to find mode of Laplace approximation
        log_marginal_likelihood = -np.inf
        for _ in range(self.max_iter_predict):
            # Line 4
            pi = expit(f)
            W = pi * (1 - pi)
            # Line 5
            W_sr = np.sqrt(W)
            W_sr_K = W_sr[:, np.newaxis] * K
            B = np.eye(W.shape[0]) + W_sr_K * W_sr
            L = cholesky(B, lower=True)
            # Line 6
            b = W * f + (self.y_train_ - pi)
            # Line 7
            a = b - W_sr * cho_solve((L, True), W_sr_K.dot(b))
            # Line 8
            f = K.dot(a)

            # Line 10: Compute log marginal likelihood in loop and use as
            #          convergence criterion
            lml = -0.5 * a.T.dot(f) \
                - np.log1p(np.exp(-(self.y_train_ * 2 - 1) * f)).sum() \
                - np.log(np.diag(L)).sum()
            # Check if we have converged (log marginal likelihood does
            # not decrease)
            # XXX: more complex convergence criterion
            if lml - log_marginal_likelihood < 1e-10:
                break
            log_marginal_likelihood = lml

        self.f_cached = f  # Remember solution for later warm-starts
        if return_temporaries:
            return log_marginal_likelihood, (pi, W_sr, L, b, a)
        else:
            return log_marginal_likelihood

    def _constrained_optimization(self, obj_func, initial_theta, bounds):
        if self.optimizer == "fmin_l_bfgs_b":
            theta_opt, func_min, convergence_dict = \
                fmin_l_bfgs_b(obj_func, initial_theta, bounds=bounds)
            if convergence_dict["warnflag"] != 0:
                warnings.warn("fmin_l_bfgs_b terminated abnormally with the "
                              " state: %s" % convergence_dict,
                              ConvergenceWarning)
        elif callable(self.optimizer):
            theta_opt, func_min = \
                self.optimizer(obj_func, initial_theta, bounds=bounds)
        else:
            raise ValueError("Unknown optimizer %s." % self.optimizer)

        return theta_opt, func_min


class GaussianProcessClassifier(BaseEstimator, ClassifierMixin):
    """Gaussian process classification (GPC) based on Laplace approximation.

    The implementation is based on Algorithm 3.1, 3.2, and 5.1 of
    Gaussian Processes for Machine Learning (GPML) by Rasmussen and
    Williams.

    Internally, the Laplace approximation is used for approximating the
    non-Gaussian posterior by a Gaussian.

    Currently, the implementation is restricted to using the logistic link
    function. For multi-class classification, several binary one-versus rest
    classifiers are fitted. Note that this class thus does not implement
    a true multi-class Laplace approximation.

    Parameters
    ----------
    kernel : kernel object
        The kernel specifying the covariance function of the GP. If None is
        passed, the kernel "1.0 * RBF(1.0)" is used as default. Note that
        the kernel's hyperparameters are optimized during fitting.

    optimizer : string or callable, optional (default: "fmin_l_bfgs_b")
        Can either be one of the internally supported optimizers for optimizing
        the kernel's parameters, specified by a string, or an externally
        defined optimizer passed as a callable. If a callable is passed, it
        must have the  signature::

            def optimizer(obj_func, initial_theta, bounds):
                # * 'obj_func' is the objective function to be maximized, which
                #   takes the hyperparameters theta as parameter and an
                #   optional flag eval_gradient, which determines if the
                #   gradient is returned additionally to the function value
                # * 'initial_theta': the initial value for theta, which can be
                #   used by local optimizers
                # * 'bounds': the bounds on the values of theta
                ....
                # Returned are the best found hyperparameters theta and
                # the corresponding value of the target function.
                return theta_opt, func_min

        Per default, the 'fmin_l_bfgs_b' algorithm from scipy.optimize
        is used. If None is passed, the kernel's parameters are kept fixed.
        Available internal optimizers are::

            'fmin_l_bfgs_b'

    n_restarts_optimizer : int, optional (default: 0)
        The number of restarts of the optimizer for finding the kernel's
        parameters which maximize the log-marginal likelihood. The first run
        of the optimizer is performed from the kernel's initial parameters,
        the remaining ones (if any) from thetas sampled log-uniform randomly
        from the space of allowed theta-values. If greater than 0, all bounds
        must be finite. Note that n_restarts_optimizer=0 implies that one
        run is performed.

    max_iter_predict : int, optional (default: 100)
        The maximum number of iterations in Newton's method for approximating
        the posterior during predict. Smaller values will reduce computation
        time at the cost of worse results.

    warm_start : bool, optional (default: False)
        If warm-starts are enabled, the solution of the last Newton iteration
        on the Laplace approximation of the posterior mode is used as
        initialization for the next call of _posterior_mode(). This can speed
        up convergence when _posterior_mode is called several times on similar
        problems as in hyperparameter optimization. See :term:`the Glossary
        <warm_start>`.

    copy_X_train : bool, optional (default: True)
        If True, a persistent copy of the training data is stored in the
        object. Otherwise, just a reference to the training data is stored,
        which might cause predictions to change if the data is modified
        externally.

    random_state : int, RandomState instance or None, optional (default: None)
        The generator used to initialize the centers.
        If int, random_state is the seed used by the random number generator;
        If RandomState instance, random_state is the random number generator;
        If None, the random number generator is the RandomState instance used
        by `np.random`.

    multi_class : string, default : "one_vs_rest"
        Specifies how multi-class classification problems are handled.
        Supported are "one_vs_rest" and "one_vs_one". In "one_vs_rest",
        one binary Gaussian process classifier is fitted for each class, which
        is trained to separate this class from the rest. In "one_vs_one", one
        binary Gaussian process classifier is fitted for each pair of classes,
        which is trained to separate these two classes. The predictions of
        these binary predictors are combined into multi-class predictions.
        Note that "one_vs_one" does not support predicting probability
        estimates.

    n_jobs : int or None, optional (default=None)
        The number of jobs to use for the computation.
        ``None`` means 1 unless in a :obj:`joblib.parallel_backend` context.
        ``-1`` means using all processors. See :term:`Glossary <n_jobs>`
        for more details.

    Attributes
    ----------
    kernel_ : kernel object
        The kernel used for prediction. In case of binary classification,
        the structure of the kernel is the same as the one passed as parameter
        but with optimized hyperparameters. In case of multi-class
        classification, a CompoundKernel is returned which consists of the
        different kernels used in the one-versus-rest classifiers.

    log_marginal_likelihood_value_ : float
        The log-marginal-likelihood of ``self.kernel_.theta``

    classes_ : array-like, shape = (n_classes,)
        Unique class labels.

    n_classes_ : int
        The number of classes in the training data

    Examples
    --------
    >>> from sklearn.datasets import load_iris
    >>> from sklearn.gaussian_process import GaussianProcessClassifier
    >>> from sklearn.gaussian_process.kernels import RBF
    >>> X, y = load_iris(return_X_y=True)
    >>> kernel = 1.0 * RBF(1.0)
    >>> gpc = GaussianProcessClassifier(kernel=kernel,
    ...         random_state=0).fit(X, y)
    >>> gpc.score(X, y) # doctest: +ELLIPSIS
    0.9866...
    >>> gpc.predict_proba(X[:2,:])
    array([[0.83548752, 0.03228706, 0.13222543],
           [0.79064206, 0.06525643, 0.14410151]])

    .. versionadded:: 0.18
    """
    def __init__(self, kernel=None, optimizer="fmin_l_bfgs_b",
                 n_restarts_optimizer=0, max_iter_predict=100,
                 warm_start=False, copy_X_train=True, random_state=None,
                 multi_class="one_vs_rest", n_jobs=None):
        self.kernel = kernel
        self.optimizer = optimizer
        self.n_restarts_optimizer = n_restarts_optimizer
        self.max_iter_predict = max_iter_predict
        self.warm_start = warm_start
        self.copy_X_train = copy_X_train
        self.random_state = random_state
        self.multi_class = multi_class
        self.n_jobs = n_jobs

    def fit(self, X, y):
        """Fit Gaussian process classification model

        Parameters
        ----------
        X : array-like, shape = (n_samples, n_features)
            Training data

        y : array-like, shape = (n_samples,)
            Target values, must be binary

        Returns
        -------
        self : returns an instance of self.
        """
        X, y = check_X_y(X, y, multi_output=False)

        self.base_estimator_ = _BinaryGaussianProcessClassifierLaplace(
            self.kernel, self.optimizer, self.n_restarts_optimizer,
            self.max_iter_predict, self.warm_start, self.copy_X_train,
            self.random_state)

        self.classes_ = np.unique(y)
        self.n_classes_ = self.classes_.size
        if self.n_classes_ == 1:
            raise ValueError("GaussianProcessClassifier requires 2 or more "
                             "distinct classes; got %d class (only class %s "
                             "is present)"
                             % (self.n_classes_, self.classes_[0]))
        if self.n_classes_ > 2:
            if self.multi_class == "one_vs_rest":
                self.base_estimator_ = \
                    OneVsRestClassifier(self.base_estimator_,
                                        n_jobs=self.n_jobs)
            elif self.multi_class == "one_vs_one":
                self.base_estimator_ = \
                    OneVsOneClassifier(self.base_estimator_,
                                       n_jobs=self.n_jobs)
            else:
                raise ValueError("Unknown multi-class mode %s"
                                 % self.multi_class)

        self.base_estimator_.fit(X, y)

        if self.n_classes_ > 2:
            self.log_marginal_likelihood_value_ = np.mean(
                [estimator.log_marginal_likelihood()
                 for estimator in self.base_estimator_.estimators_])
        else:
            self.log_marginal_likelihood_value_ = \
                self.base_estimator_.log_marginal_likelihood()

        return self

    def predict(self, X):
        """Perform classification on an array of test vectors X.

        Parameters
        ----------
        X : array-like, shape = (n_samples, n_features)

        Returns
        -------
        C : array, shape = (n_samples,)
            Predicted target values for X, values are from ``classes_``
        """
        check_is_fitted(self, ["classes_", "n_classes_"])
        X = check_array(X)
        return self.base_estimator_.predict(X)

    def predict_proba(self, X):
        """Return probability estimates for the test vector X.

        Parameters
        ----------
        X : array-like, shape = (n_samples, n_features)

        Returns
        -------
        C : array-like, shape = (n_samples, n_classes)
            Returns the probability of the samples for each class in
            the model. The columns correspond to the classes in sorted
            order, as they appear in the attribute `classes_`.
        """
        check_is_fitted(self, ["classes_", "n_classes_"])
        if self.n_classes_ > 2 and self.multi_class == "one_vs_one":
            raise ValueError("one_vs_one multi-class mode does not support "
                             "predicting probability estimates. Use "
                             "one_vs_rest mode instead.")
        X = check_array(X)
        return self.base_estimator_.predict_proba(X)

    @property
    def kernel_(self):
        if self.n_classes_ == 2:
            return self.base_estimator_.kernel_
        else:
            return CompoundKernel(
                [estimator.kernel_
                 for estimator in self.base_estimator_.estimators_])

    def log_marginal_likelihood(self, theta=None, eval_gradient=False):
        """Returns log-marginal likelihood of theta for training data.

        In the case of multi-class classification, the mean log-marginal
        likelihood of the one-versus-rest classifiers are returned.

        Parameters
        ----------
        theta : array-like, shape = (n_kernel_params,) or none
            Kernel hyperparameters for which the log-marginal likelihood is
            evaluated. In the case of multi-class classification, theta may
            be the  hyperparameters of the compound kernel or of an individual
            kernel. In the latter case, all individual kernel get assigned the
            same theta values. If None, the precomputed log_marginal_likelihood
            of ``self.kernel_.theta`` is returned.

        eval_gradient : bool, default: False
            If True, the gradient of the log-marginal likelihood with respect
            to the kernel hyperparameters at position theta is returned
            additionally. Note that gradient computation is not supported
            for non-binary classification. If True, theta must not be None.

        Returns
        -------
        log_likelihood : float
            Log-marginal likelihood of theta for training data.

        log_likelihood_gradient : array, shape = (n_kernel_params,), optional
            Gradient of the log-marginal likelihood with respect to the kernel
            hyperparameters at position theta.
            Only returned when eval_gradient is True.
        """
        check_is_fitted(self, ["classes_", "n_classes_"])

        if theta is None:
            if eval_gradient:
                raise ValueError(
                    "Gradient can only be evaluated for theta!=None")
            return self.log_marginal_likelihood_value_

        theta = np.asarray(theta)
        if self.n_classes_ == 2:
            return self.base_estimator_.log_marginal_likelihood(
                theta, eval_gradient)
        else:
            if eval_gradient:
                raise NotImplementedError(
                    "Gradient of log-marginal-likelihood not implemented for "
                    "multi-class GPC.")
            estimators = self.base_estimator_.estimators_
            n_dims = estimators[0].kernel_.n_dims
            if theta.shape[0] == n_dims:  # use same theta for all sub-kernels
                return np.mean(
                    [estimator.log_marginal_likelihood(theta)
                     for i, estimator in enumerate(estimators)])
            elif theta.shape[0] == n_dims * self.classes_.shape[0]:
                # theta for compound kernel
                return np.mean(
                    [estimator.log_marginal_likelihood(
                        theta[n_dims * i:n_dims * (i + 1)])
                     for i, estimator in enumerate(estimators)])
            else:
                raise ValueError("Shape of theta must be either %d or %d. "
                                 "Obtained theta with shape %d."
                                 % (n_dims, n_dims * self.classes_.shape[0],
                                    theta.shape[0]))