File: _linear_loss.py

package info (click to toggle)
scikit-learn 1.7.2%2Bdfsg-3
  • links: PTS, VCS
  • area: main
  • in suites: forky, sid
  • size: 25,752 kB
  • sloc: python: 219,120; cpp: 5,790; ansic: 846; makefile: 191; javascript: 110
file content (825 lines) | stat: -rw-r--r-- 34,113 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
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
"""
Loss functions for linear models with raw_prediction = X @ coef
"""

# Authors: The scikit-learn developers
# SPDX-License-Identifier: BSD-3-Clause

import numpy as np
from scipy import sparse

from ..utils.extmath import squared_norm


def sandwich_dot(X, W):
    """Compute the sandwich product X.T @ diag(W) @ X."""
    # TODO: This "sandwich product" is the main computational bottleneck for solvers
    # that use the full hessian matrix. Here, thread parallelism would pay-off the
    # most.
    # While a dedicated Cython routine could exploit the symmetry, it is very hard to
    # beat BLAS GEMM, even thought the latter cannot exploit the symmetry, unless one
    # pays the price of taking square roots and implements
    #    sqrtWX = sqrt(W)[: None] * X
    #    return sqrtWX.T @ sqrtWX
    # which (might) detect the symmetry and use BLAS SYRK under the hood.
    n_samples = X.shape[0]
    if sparse.issparse(X):
        return (
            X.T @ sparse.dia_matrix((W, 0), shape=(n_samples, n_samples)) @ X
        ).toarray()
    else:
        # np.einsum may use less memory but the following, using BLAS matrix
        # multiplication (gemm), is by far faster.
        WX = W[:, None] * X
        return X.T @ WX


class LinearModelLoss:
    """General class for loss functions with raw_prediction = X @ coef + intercept.

    Note that raw_prediction is also known as linear predictor.

    The loss is the average of per sample losses and includes a term for L2
    regularization::

        loss = 1 / s_sum * sum_i s_i loss(y_i, X_i @ coef + intercept)
               + 1/2 * l2_reg_strength * ||coef||_2^2

    with sample weights s_i=1 if sample_weight=None and s_sum=sum_i s_i.

    Gradient and hessian, for simplicity without intercept, are::

        gradient = 1 / s_sum * X.T @ loss.gradient + l2_reg_strength * coef
        hessian = 1 / s_sum * X.T @ diag(loss.hessian) @ X
                  + l2_reg_strength * identity

    Conventions:
        if fit_intercept:
            n_dof =  n_features + 1
        else:
            n_dof = n_features

        if base_loss.is_multiclass:
            coef.shape = (n_classes, n_dof) or ravelled (n_classes * n_dof,)
        else:
            coef.shape = (n_dof,)

        The intercept term is at the end of the coef array:
        if base_loss.is_multiclass:
            if coef.shape (n_classes, n_dof):
                intercept = coef[:, -1]
            if coef.shape (n_classes * n_dof,)
                intercept = coef[n_features::n_dof] = coef[(n_dof-1)::n_dof]
            intercept.shape = (n_classes,)
        else:
            intercept = coef[-1]

        Shape of gradient follows shape of coef.
        gradient.shape = coef.shape

        But hessian (to make our lives simpler) are always 2-d:
        if base_loss.is_multiclass:
            hessian.shape = (n_classes * n_dof, n_classes * n_dof)
        else:
            hessian.shape = (n_dof, n_dof)

    Note: If coef has shape (n_classes * n_dof,), the 2d-array can be reconstructed as

        coef.reshape((n_classes, -1), order="F")

    The option order="F" makes coef[:, i] contiguous. This, in turn, makes the
    coefficients without intercept, coef[:, :-1], contiguous and speeds up
    matrix-vector computations.

    Note: If the average loss per sample is wanted instead of the sum of the loss per
    sample, one can simply use a rescaled sample_weight such that
    sum(sample_weight) = 1.

    Parameters
    ----------
    base_loss : instance of class BaseLoss from sklearn._loss.
    fit_intercept : bool
    """

    def __init__(self, base_loss, fit_intercept):
        self.base_loss = base_loss
        self.fit_intercept = fit_intercept

    def init_zero_coef(self, X, dtype=None):
        """Allocate coef of correct shape with zeros.

        Parameters:
        -----------
        X : {array-like, sparse matrix} of shape (n_samples, n_features)
            Training data.
        dtype : data-type, default=None
            Overrides the data type of coef. With dtype=None, coef will have the same
            dtype as X.

        Returns
        -------
        coef : ndarray of shape (n_dof,) or (n_classes, n_dof)
            Coefficients of a linear model.
        """
        n_features = X.shape[1]
        n_classes = self.base_loss.n_classes
        if self.fit_intercept:
            n_dof = n_features + 1
        else:
            n_dof = n_features
        if self.base_loss.is_multiclass:
            coef = np.zeros_like(X, shape=(n_classes, n_dof), dtype=dtype, order="F")
        else:
            coef = np.zeros_like(X, shape=n_dof, dtype=dtype)
        return coef

    def weight_intercept(self, coef):
        """Helper function to get coefficients and intercept.

        Parameters
        ----------
        coef : ndarray of shape (n_dof,), (n_classes, n_dof) or (n_classes * n_dof,)
            Coefficients of a linear model.
            If shape (n_classes * n_dof,), the classes of one feature are contiguous,
            i.e. one reconstructs the 2d-array via
            coef.reshape((n_classes, -1), order="F").

        Returns
        -------
        weights : ndarray of shape (n_features,) or (n_classes, n_features)
            Coefficients without intercept term.
        intercept : float or ndarray of shape (n_classes,)
            Intercept terms.
        """
        if not self.base_loss.is_multiclass:
            if self.fit_intercept:
                intercept = coef[-1]
                weights = coef[:-1]
            else:
                intercept = 0.0
                weights = coef
        else:
            # reshape to (n_classes, n_dof)
            if coef.ndim == 1:
                weights = coef.reshape((self.base_loss.n_classes, -1), order="F")
            else:
                weights = coef
            if self.fit_intercept:
                intercept = weights[:, -1]
                weights = weights[:, :-1]
            else:
                intercept = 0.0

        return weights, intercept

    def weight_intercept_raw(self, coef, X):
        """Helper function to get coefficients, intercept and raw_prediction.

        Parameters
        ----------
        coef : ndarray of shape (n_dof,), (n_classes, n_dof) or (n_classes * n_dof,)
            Coefficients of a linear model.
            If shape (n_classes * n_dof,), the classes of one feature are contiguous,
            i.e. one reconstructs the 2d-array via
            coef.reshape((n_classes, -1), order="F").
        X : {array-like, sparse matrix} of shape (n_samples, n_features)
            Training data.

        Returns
        -------
        weights : ndarray of shape (n_features,) or (n_classes, n_features)
            Coefficients without intercept term.
        intercept : float or ndarray of shape (n_classes,)
            Intercept terms.
        raw_prediction : ndarray of shape (n_samples,) or \
            (n_samples, n_classes)
        """
        weights, intercept = self.weight_intercept(coef)

        if not self.base_loss.is_multiclass:
            raw_prediction = X @ weights + intercept
        else:
            # weights has shape (n_classes, n_dof)
            raw_prediction = X @ weights.T + intercept  # ndarray, likely C-contiguous

        return weights, intercept, raw_prediction

    def l2_penalty(self, weights, l2_reg_strength):
        """Compute L2 penalty term l2_reg_strength/2 *||w||_2^2."""
        norm2_w = weights @ weights if weights.ndim == 1 else squared_norm(weights)
        return 0.5 * l2_reg_strength * norm2_w

    def loss(
        self,
        coef,
        X,
        y,
        sample_weight=None,
        l2_reg_strength=0.0,
        n_threads=1,
        raw_prediction=None,
    ):
        """Compute the loss as weighted average over point-wise losses.

        Parameters
        ----------
        coef : ndarray of shape (n_dof,), (n_classes, n_dof) or (n_classes * n_dof,)
            Coefficients of a linear model.
            If shape (n_classes * n_dof,), the classes of one feature are contiguous,
            i.e. one reconstructs the 2d-array via
            coef.reshape((n_classes, -1), order="F").
        X : {array-like, sparse matrix} of shape (n_samples, n_features)
            Training data.
        y : contiguous array of shape (n_samples,)
            Observed, true target values.
        sample_weight : None or contiguous array of shape (n_samples,), default=None
            Sample weights.
        l2_reg_strength : float, default=0.0
            L2 regularization strength
        n_threads : int, default=1
            Number of OpenMP threads to use.
        raw_prediction : C-contiguous array of shape (n_samples,) or array of \
            shape (n_samples, n_classes)
            Raw prediction values (in link space). If provided, these are used. If
            None, then raw_prediction = X @ coef + intercept is calculated.

        Returns
        -------
        loss : float
            Weighted average of losses per sample, plus penalty.
        """
        if raw_prediction is None:
            weights, intercept, raw_prediction = self.weight_intercept_raw(coef, X)
        else:
            weights, intercept = self.weight_intercept(coef)

        loss = self.base_loss.loss(
            y_true=y,
            raw_prediction=raw_prediction,
            sample_weight=None,
            n_threads=n_threads,
        )
        loss = np.average(loss, weights=sample_weight)

        return loss + self.l2_penalty(weights, l2_reg_strength)

    def loss_gradient(
        self,
        coef,
        X,
        y,
        sample_weight=None,
        l2_reg_strength=0.0,
        n_threads=1,
        raw_prediction=None,
    ):
        """Computes the sum of loss and gradient w.r.t. coef.

        Parameters
        ----------
        coef : ndarray of shape (n_dof,), (n_classes, n_dof) or (n_classes * n_dof,)
            Coefficients of a linear model.
            If shape (n_classes * n_dof,), the classes of one feature are contiguous,
            i.e. one reconstructs the 2d-array via
            coef.reshape((n_classes, -1), order="F").
        X : {array-like, sparse matrix} of shape (n_samples, n_features)
            Training data.
        y : contiguous array of shape (n_samples,)
            Observed, true target values.
        sample_weight : None or contiguous array of shape (n_samples,), default=None
            Sample weights.
        l2_reg_strength : float, default=0.0
            L2 regularization strength
        n_threads : int, default=1
            Number of OpenMP threads to use.
        raw_prediction : C-contiguous array of shape (n_samples,) or array of \
            shape (n_samples, n_classes)
            Raw prediction values (in link space). If provided, these are used. If
            None, then raw_prediction = X @ coef + intercept is calculated.

        Returns
        -------
        loss : float
            Weighted average of losses per sample, plus penalty.

        gradient : ndarray of shape coef.shape
             The gradient of the loss.
        """
        (n_samples, n_features), n_classes = X.shape, self.base_loss.n_classes
        n_dof = n_features + int(self.fit_intercept)

        if raw_prediction is None:
            weights, intercept, raw_prediction = self.weight_intercept_raw(coef, X)
        else:
            weights, intercept = self.weight_intercept(coef)

        loss, grad_pointwise = self.base_loss.loss_gradient(
            y_true=y,
            raw_prediction=raw_prediction,
            sample_weight=sample_weight,
            n_threads=n_threads,
        )
        sw_sum = n_samples if sample_weight is None else np.sum(sample_weight)
        loss = loss.sum() / sw_sum
        loss += self.l2_penalty(weights, l2_reg_strength)

        grad_pointwise /= sw_sum

        if not self.base_loss.is_multiclass:
            grad = np.empty_like(coef, dtype=weights.dtype)
            grad[:n_features] = X.T @ grad_pointwise + l2_reg_strength * weights
            if self.fit_intercept:
                grad[-1] = grad_pointwise.sum()
        else:
            grad = np.empty((n_classes, n_dof), dtype=weights.dtype, order="F")
            # grad_pointwise.shape = (n_samples, n_classes)
            grad[:, :n_features] = grad_pointwise.T @ X + l2_reg_strength * weights
            if self.fit_intercept:
                grad[:, -1] = grad_pointwise.sum(axis=0)
            if coef.ndim == 1:
                grad = grad.ravel(order="F")

        return loss, grad

    def gradient(
        self,
        coef,
        X,
        y,
        sample_weight=None,
        l2_reg_strength=0.0,
        n_threads=1,
        raw_prediction=None,
    ):
        """Computes the gradient w.r.t. coef.

        Parameters
        ----------
        coef : ndarray of shape (n_dof,), (n_classes, n_dof) or (n_classes * n_dof,)
            Coefficients of a linear model.
            If shape (n_classes * n_dof,), the classes of one feature are contiguous,
            i.e. one reconstructs the 2d-array via
            coef.reshape((n_classes, -1), order="F").
        X : {array-like, sparse matrix} of shape (n_samples, n_features)
            Training data.
        y : contiguous array of shape (n_samples,)
            Observed, true target values.
        sample_weight : None or contiguous array of shape (n_samples,), default=None
            Sample weights.
        l2_reg_strength : float, default=0.0
            L2 regularization strength
        n_threads : int, default=1
            Number of OpenMP threads to use.
        raw_prediction : C-contiguous array of shape (n_samples,) or array of \
            shape (n_samples, n_classes)
            Raw prediction values (in link space). If provided, these are used. If
            None, then raw_prediction = X @ coef + intercept is calculated.

        Returns
        -------
        gradient : ndarray of shape coef.shape
             The gradient of the loss.
        """
        (n_samples, n_features), n_classes = X.shape, self.base_loss.n_classes
        n_dof = n_features + int(self.fit_intercept)

        if raw_prediction is None:
            weights, intercept, raw_prediction = self.weight_intercept_raw(coef, X)
        else:
            weights, intercept = self.weight_intercept(coef)

        grad_pointwise = self.base_loss.gradient(
            y_true=y,
            raw_prediction=raw_prediction,
            sample_weight=sample_weight,
            n_threads=n_threads,
        )
        sw_sum = n_samples if sample_weight is None else np.sum(sample_weight)
        grad_pointwise /= sw_sum

        if not self.base_loss.is_multiclass:
            grad = np.empty_like(coef, dtype=weights.dtype)
            grad[:n_features] = X.T @ grad_pointwise + l2_reg_strength * weights
            if self.fit_intercept:
                grad[-1] = grad_pointwise.sum()
            return grad
        else:
            grad = np.empty((n_classes, n_dof), dtype=weights.dtype, order="F")
            # gradient.shape = (n_samples, n_classes)
            grad[:, :n_features] = grad_pointwise.T @ X + l2_reg_strength * weights
            if self.fit_intercept:
                grad[:, -1] = grad_pointwise.sum(axis=0)
            if coef.ndim == 1:
                return grad.ravel(order="F")
            else:
                return grad

    def gradient_hessian(
        self,
        coef,
        X,
        y,
        sample_weight=None,
        l2_reg_strength=0.0,
        n_threads=1,
        gradient_out=None,
        hessian_out=None,
        raw_prediction=None,
    ):
        """Computes gradient and hessian w.r.t. coef.

        Parameters
        ----------
        coef : ndarray of shape (n_dof,), (n_classes, n_dof) or (n_classes * n_dof,)
            Coefficients of a linear model.
            If shape (n_classes * n_dof,), the classes of one feature are contiguous,
            i.e. one reconstructs the 2d-array via
            coef.reshape((n_classes, -1), order="F").
        X : {array-like, sparse matrix} of shape (n_samples, n_features)
            Training data.
        y : contiguous array of shape (n_samples,)
            Observed, true target values.
        sample_weight : None or contiguous array of shape (n_samples,), default=None
            Sample weights.
        l2_reg_strength : float, default=0.0
            L2 regularization strength
        n_threads : int, default=1
            Number of OpenMP threads to use.
        gradient_out : None or ndarray of shape coef.shape
            A location into which the gradient is stored. If None, a new array
            might be created.
        hessian_out : None or ndarray of shape (n_dof, n_dof) or \
            (n_classes * n_dof, n_classes * n_dof)
            A location into which the hessian is stored. If None, a new array
            might be created.
        raw_prediction : C-contiguous array of shape (n_samples,) or array of \
            shape (n_samples, n_classes)
            Raw prediction values (in link space). If provided, these are used. If
            None, then raw_prediction = X @ coef + intercept is calculated.

        Returns
        -------
        gradient : ndarray of shape coef.shape
             The gradient of the loss.

        hessian : ndarray of shape (n_dof, n_dof) or \
            (n_classes, n_dof, n_dof, n_classes)
            Hessian matrix.

        hessian_warning : bool
            True if pointwise hessian has more than 25% of its elements non-positive.
        """
        (n_samples, n_features), n_classes = X.shape, self.base_loss.n_classes
        n_dof = n_features + int(self.fit_intercept)
        if raw_prediction is None:
            weights, intercept, raw_prediction = self.weight_intercept_raw(coef, X)
        else:
            weights, intercept = self.weight_intercept(coef)
        sw_sum = n_samples if sample_weight is None else np.sum(sample_weight)

        # Allocate gradient.
        if gradient_out is None:
            grad = np.empty_like(coef, dtype=weights.dtype, order="F")
        elif gradient_out.shape != coef.shape:
            raise ValueError(
                f"gradient_out is required to have shape coef.shape = {coef.shape}; "
                f"got {gradient_out.shape}."
            )
        elif self.base_loss.is_multiclass and not gradient_out.flags.f_contiguous:
            raise ValueError("gradient_out must be F-contiguous.")
        else:
            grad = gradient_out
        # Allocate hessian.
        n = coef.size  # for multinomial this equals n_dof * n_classes
        if hessian_out is None:
            hess = np.empty((n, n), dtype=weights.dtype)
        elif hessian_out.shape != (n, n):
            raise ValueError(
                f"hessian_out is required to have shape ({n, n}); got "
                f"{hessian_out.shape=}."
            )
        elif self.base_loss.is_multiclass and (
            not hessian_out.flags.c_contiguous and not hessian_out.flags.f_contiguous
        ):
            raise ValueError("hessian_out must be contiguous.")
        else:
            hess = hessian_out

        if not self.base_loss.is_multiclass:
            grad_pointwise, hess_pointwise = self.base_loss.gradient_hessian(
                y_true=y,
                raw_prediction=raw_prediction,
                sample_weight=sample_weight,
                n_threads=n_threads,
            )
            grad_pointwise /= sw_sum
            hess_pointwise /= sw_sum

            # For non-canonical link functions and far away from the optimum, the
            # pointwise hessian can be negative. We take care that 75% of the hessian
            # entries are positive.
            hessian_warning = (
                np.average(hess_pointwise <= 0, weights=sample_weight) > 0.25
            )
            hess_pointwise = np.abs(hess_pointwise)

            grad[:n_features] = X.T @ grad_pointwise + l2_reg_strength * weights
            if self.fit_intercept:
                grad[-1] = grad_pointwise.sum()

            if hessian_warning:
                # Exit early without computing the hessian.
                return grad, hess, hessian_warning

            hess[:n_features, :n_features] = sandwich_dot(X, hess_pointwise)

            if l2_reg_strength > 0:
                # The L2 penalty enters the Hessian on the diagonal only. To add those
                # terms, we use a flattened view of the array.
                order = "C" if hess.flags.c_contiguous else "F"
                hess.reshape(-1, order=order)[: (n_features * n_dof) : (n_dof + 1)] += (
                    l2_reg_strength
                )

            if self.fit_intercept:
                # With intercept included as added column to X, the hessian becomes
                # hess = (X, 1)' @ diag(h) @ (X, 1)
                #      = (X' @ diag(h) @ X, X' @ h)
                #        (           h @ X, sum(h))
                # The left upper part has already been filled, it remains to compute
                # the last row and the last column.
                Xh = X.T @ hess_pointwise
                hess[:-1, -1] = Xh
                hess[-1, :-1] = Xh
                hess[-1, -1] = hess_pointwise.sum()
        else:
            # Here we may safely assume HalfMultinomialLoss aka categorical
            # cross-entropy.
            # HalfMultinomialLoss computes only the diagonal part of the hessian, i.e.
            # diagonal in the classes. Here, we want the full hessian. Therefore, we
            # call gradient_proba.
            grad_pointwise, proba = self.base_loss.gradient_proba(
                y_true=y,
                raw_prediction=raw_prediction,
                sample_weight=sample_weight,
                n_threads=n_threads,
            )
            grad_pointwise /= sw_sum
            grad = grad.reshape((n_classes, n_dof), order="F")
            grad[:, :n_features] = grad_pointwise.T @ X + l2_reg_strength * weights
            if self.fit_intercept:
                grad[:, -1] = grad_pointwise.sum(axis=0)
            if coef.ndim == 1:
                grad = grad.ravel(order="F")

            # The full hessian matrix, i.e. not only the diagonal part, dropping most
            # indices, is given by:
            #
            #   hess = X' @ h @ X
            #
            # Here, h is a priori a 4-dimensional matrix of shape
            # (n_samples, n_samples, n_classes, n_classes). It is diagonal its first
            # two dimensions (the ones with n_samples), i.e. it is
            # effectively a 3-dimensional matrix (n_samples, n_classes, n_classes).
            #
            #   h = diag(p) - p' p
            #
            # or with indices k and l for classes
            #
            #   h_kl = p_k * delta_kl - p_k * p_l
            #
            # with p_k the (predicted) probability for class k. Only the dimension in
            # n_samples multiplies with X.
            # For 3 classes and n_samples = 1, this looks like ("@" is a bit misused
            # here):
            #
            #   hess = X' @ (h00 h10 h20) @ X
            #               (h10 h11 h12)
            #               (h20 h12 h22)
            #        = (X' @ diag(h00) @ X, X' @ diag(h10), X' @ diag(h20))
            #          (X' @ diag(h10) @ X, X' @ diag(h11), X' @ diag(h12))
            #          (X' @ diag(h20) @ X, X' @ diag(h12), X' @ diag(h22))
            #
            # Now coef of shape (n_classes * n_dof) is contiguous in n_classes.
            # Therefore, we want the hessian to follow this convention, too, i.e.
            #     hess[:n_classes, :n_classes] = (x0' @ h00 @ x0, x0' @ h10 @ x0, ..)
            #                                    (x0' @ h10 @ x0, x0' @ h11 @ x0, ..)
            #                                    (x0' @ h20 @ x0, x0' @ h12 @ x0, ..)
            # is the first feature, x0, for all classes. In our implementation, we
            # still want to take advantage of BLAS "X.T @ X". Therefore, we have some
            # index/slicing battle to fight.
            if sample_weight is not None:
                sw = sample_weight / sw_sum
            else:
                sw = 1.0 / sw_sum

            for k in range(n_classes):
                # Diagonal terms (in classes) hess_kk.
                # Note that this also writes to some of the lower triangular part.
                h = proba[:, k] * (1 - proba[:, k]) * sw
                hess[
                    k : n_classes * n_features : n_classes,
                    k : n_classes * n_features : n_classes,
                ] = sandwich_dot(X, h)
                if self.fit_intercept:
                    # See above in the non multiclass case.
                    Xh = X.T @ h
                    hess[
                        k : n_classes * n_features : n_classes,
                        n_classes * n_features + k,
                    ] = Xh
                    hess[
                        n_classes * n_features + k,
                        k : n_classes * n_features : n_classes,
                    ] = Xh
                    hess[n_classes * n_features + k, n_classes * n_features + k] = (
                        h.sum()
                    )
                # Off diagonal terms (in classes) hess_kl.
                for l in range(k + 1, n_classes):
                    # Upper triangle (in classes).
                    h = -proba[:, k] * proba[:, l] * sw
                    hess[
                        k : n_classes * n_features : n_classes,
                        l : n_classes * n_features : n_classes,
                    ] = sandwich_dot(X, h)
                    if self.fit_intercept:
                        Xh = X.T @ h
                        hess[
                            k : n_classes * n_features : n_classes,
                            n_classes * n_features + l,
                        ] = Xh
                        hess[
                            n_classes * n_features + k,
                            l : n_classes * n_features : n_classes,
                        ] = Xh
                        hess[n_classes * n_features + k, n_classes * n_features + l] = (
                            h.sum()
                        )
                    # Fill lower triangle (in classes).
                    hess[l::n_classes, k::n_classes] = hess[k::n_classes, l::n_classes]

            if l2_reg_strength > 0:
                # See above in the non multiclass case.
                order = "C" if hess.flags.c_contiguous else "F"
                hess.reshape(-1, order=order)[
                    : (n_classes**2 * n_features * n_dof) : (n_classes * n_dof + 1)
                ] += l2_reg_strength

            # The pointwise hessian is always non-negative for the multinomial loss.
            hessian_warning = False

        return grad, hess, hessian_warning

    def gradient_hessian_product(
        self, coef, X, y, sample_weight=None, l2_reg_strength=0.0, n_threads=1
    ):
        """Computes gradient and hessp (hessian product function) w.r.t. coef.

        Parameters
        ----------
        coef : ndarray of shape (n_dof,), (n_classes, n_dof) or (n_classes * n_dof,)
            Coefficients of a linear model.
            If shape (n_classes * n_dof,), the classes of one feature are contiguous,
            i.e. one reconstructs the 2d-array via
            coef.reshape((n_classes, -1), order="F").
        X : {array-like, sparse matrix} of shape (n_samples, n_features)
            Training data.
        y : contiguous array of shape (n_samples,)
            Observed, true target values.
        sample_weight : None or contiguous array of shape (n_samples,), default=None
            Sample weights.
        l2_reg_strength : float, default=0.0
            L2 regularization strength
        n_threads : int, default=1
            Number of OpenMP threads to use.

        Returns
        -------
        gradient : ndarray of shape coef.shape
             The gradient of the loss.

        hessp : callable
            Function that takes in a vector input of shape of gradient and
            and returns matrix-vector product with hessian.
        """
        (n_samples, n_features), n_classes = X.shape, self.base_loss.n_classes
        n_dof = n_features + int(self.fit_intercept)
        weights, intercept, raw_prediction = self.weight_intercept_raw(coef, X)
        sw_sum = n_samples if sample_weight is None else np.sum(sample_weight)

        if not self.base_loss.is_multiclass:
            grad_pointwise, hess_pointwise = self.base_loss.gradient_hessian(
                y_true=y,
                raw_prediction=raw_prediction,
                sample_weight=sample_weight,
                n_threads=n_threads,
            )
            grad_pointwise /= sw_sum
            hess_pointwise /= sw_sum
            grad = np.empty_like(coef, dtype=weights.dtype)
            grad[:n_features] = X.T @ grad_pointwise + l2_reg_strength * weights
            if self.fit_intercept:
                grad[-1] = grad_pointwise.sum()

            # Precompute as much as possible: hX, hX_sum and hessian_sum
            hessian_sum = hess_pointwise.sum()
            if sparse.issparse(X):
                hX = (
                    sparse.dia_matrix((hess_pointwise, 0), shape=(n_samples, n_samples))
                    @ X
                )
            else:
                hX = hess_pointwise[:, np.newaxis] * X

            if self.fit_intercept:
                # Calculate the double derivative with respect to intercept.
                # Note: In case hX is sparse, hX.sum is a matrix object.
                hX_sum = np.squeeze(np.asarray(hX.sum(axis=0)))
                # prevent squeezing to zero-dim array if n_features == 1
                hX_sum = np.atleast_1d(hX_sum)

            # With intercept included and l2_reg_strength = 0, hessp returns
            # res = (X, 1)' @ diag(h) @ (X, 1) @ s
            #     = (X, 1)' @ (hX @ s[:n_features], sum(h) * s[-1])
            # res[:n_features] = X' @ hX @ s[:n_features] + sum(h) * s[-1]
            # res[-1] = 1' @ hX @ s[:n_features] + sum(h) * s[-1]
            def hessp(s):
                ret = np.empty_like(s)
                if sparse.issparse(X):
                    ret[:n_features] = X.T @ (hX @ s[:n_features])
                else:
                    ret[:n_features] = np.linalg.multi_dot([X.T, hX, s[:n_features]])
                ret[:n_features] += l2_reg_strength * s[:n_features]

                if self.fit_intercept:
                    ret[:n_features] += s[-1] * hX_sum
                    ret[-1] = hX_sum @ s[:n_features] + hessian_sum * s[-1]
                return ret

        else:
            # Here we may safely assume HalfMultinomialLoss aka categorical
            # cross-entropy.
            # HalfMultinomialLoss computes only the diagonal part of the hessian, i.e.
            # diagonal in the classes. Here, we want the matrix-vector product of the
            # full hessian. Therefore, we call gradient_proba.
            grad_pointwise, proba = self.base_loss.gradient_proba(
                y_true=y,
                raw_prediction=raw_prediction,
                sample_weight=sample_weight,
                n_threads=n_threads,
            )
            grad_pointwise /= sw_sum
            grad = np.empty((n_classes, n_dof), dtype=weights.dtype, order="F")
            grad[:, :n_features] = grad_pointwise.T @ X + l2_reg_strength * weights
            if self.fit_intercept:
                grad[:, -1] = grad_pointwise.sum(axis=0)

            # Full hessian-vector product, i.e. not only the diagonal part of the
            # hessian. Derivation with some index battle for input vector s:
            #   - sample index i
            #   - feature indices j, m
            #   - class indices k, l
            #   - 1_{k=l} is one if k=l else 0
            #   - p_i_k is the (predicted) probability that sample i belongs to class k
            #     for all i: sum_k p_i_k = 1
            #   - s_l_m is input vector for class l and feature m
            #   - X' = X transposed
            #
            # Note: Hessian with dropping most indices is just:
            #       X' @ p_k (1(k=l) - p_l) @ X
            #
            # result_{k j} = sum_{i, l, m} Hessian_{i, k j, m l} * s_l_m
            #   = sum_{i, l, m} (X')_{ji} * p_i_k * (1_{k=l} - p_i_l)
            #                   * X_{im} s_l_m
            #   = sum_{i, m} (X')_{ji} * p_i_k
            #                * (X_{im} * s_k_m - sum_l p_i_l * X_{im} * s_l_m)
            #
            # See also https://github.com/scikit-learn/scikit-learn/pull/3646#discussion_r17461411
            def hessp(s):
                s = s.reshape((n_classes, -1), order="F")  # shape = (n_classes, n_dof)
                if self.fit_intercept:
                    s_intercept = s[:, -1]
                    s = s[:, :-1]  # shape = (n_classes, n_features)
                else:
                    s_intercept = 0
                tmp = X @ s.T + s_intercept  # X_{im} * s_k_m
                tmp += (-proba * tmp).sum(axis=1)[:, np.newaxis]  # - sum_l ..
                tmp *= proba  # * p_i_k
                if sample_weight is not None:
                    tmp *= sample_weight[:, np.newaxis]
                # hess_prod = empty_like(grad), but we ravel grad below and this
                # function is run after that.
                hess_prod = np.empty((n_classes, n_dof), dtype=weights.dtype, order="F")
                hess_prod[:, :n_features] = (tmp.T @ X) / sw_sum + l2_reg_strength * s
                if self.fit_intercept:
                    hess_prod[:, -1] = tmp.sum(axis=0) / sw_sum
                if coef.ndim == 1:
                    return hess_prod.ravel(order="F")
                else:
                    return hess_prod

            if coef.ndim == 1:
                return grad.ravel(order="F"), hessp

        return grad, hessp