mumbo.py 25 KB
Newer Older
Dominique Benielli's avatar
algo  
Dominique Benielli committed
1
# -*- coding: utf-8 -*-
Dominique Benielli's avatar
Dominique Benielli committed
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
# ######### COPYRIGHT #########
#
# Copyright(c) 2020
# -----------------
#
# * Université d'Aix Marseille (AMU) -
# * Centre National de la Recherche Scientifique (CNRS) -
# * Université de Toulon (UTLN).
# * Copyright © 2019-2020 AMU, CNRS, UTLN
#
# Contributors:
# ------------
#
# * Sokol Koço <sokol.koco_AT_lis-lab.fr>
# * Cécile Capponi <cecile.capponi_AT_univ-amu.fr>
# * Florent Jaillet <florent.jaillet_AT_math.cnrs.fr>
# * Dominique Benielli <dominique.benielli_AT_univ-amu.fr>
# * Riikka Huusari <rikka.huusari_AT_univ-amu.fr>
# * Baptiste Bauvin <baptiste.bauvin_AT_univ-amu.fr>
# * Hachem Kadri <hachem.kadri_AT_lis-lab.fr>
#
# Description:
# -----------
#
# The multimodal package implement classifiers multiview, 
Dominique Benielli's avatar
Dominique Benielli committed
27
# MumboClassifier class, MuComboClassifier class, MVML class, MKL class.
Dominique Benielli's avatar
Dominique Benielli committed
28
29
30
31
32
33
34
35
36
37
38
39
40
41
# compatible with sklearn
#
# Version:
# -------
#
# * multimodal version = 0.0.dev0
#
# Licence:
# -------
#
# License: New BSD License
#
#
# ######### COPYRIGHT #########
Dominique Benielli's avatar
algo  
Dominique Benielli committed
42
43
44
45
46
47
r"""Multimodal Boosting

This module contains a **Mu**\ lti\ **M**\ odal **Bo**\ osting (**MuMBo**)
estimator for classification implemented in the ``MumboClassifier`` class.
"""
import numpy as np
48

Dominique Benielli's avatar
algo  
Dominique Benielli committed
49
50
from sklearn.base import ClassifierMixin
from sklearn.ensemble import BaseEnsemble
Baptiste Bauvin's avatar
Baptiste Bauvin committed
51
from sklearn.ensemble._forest import BaseForest
Dominique Benielli's avatar
algo  
Dominique Benielli committed
52
53
from sklearn.metrics import accuracy_score
from sklearn.tree import DecisionTreeClassifier
Baptiste Bauvin's avatar
Baptiste Bauvin committed
54
from sklearn.tree import BaseDecisionTree
Dominique Benielli's avatar
algo  
Dominique Benielli committed
55
56
57
58
from sklearn.tree._tree import DTYPE
from sklearn.utils import check_array, check_X_y, check_random_state
from sklearn.utils.multiclass import check_classification_targets
from sklearn.utils.validation import check_is_fitted, has_fit_parameter
Dominique Benielli's avatar
Dominique Benielli committed
59
from .boost import UBoosting
Dominique Benielli's avatar
algo  
Dominique Benielli committed
60
61


Dominique Benielli's avatar
Dominique Benielli committed
62
class MumboClassifier(BaseEnsemble, ClassifierMixin, UBoosting):
Dominique Benielli's avatar
algo  
Dominique Benielli committed
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
    r"""It then iterates the process on the same dataset but where the weights of
    incorrectly classified instances are adjusted such that subsequent
    classifiers focus more on difficult cases.
    A MuMBo classifier.

    A MuMBo classifier is a meta-estimator that implements a multimodal
    (or multi-view) boosting algorithm:

    It fits a set of classifiers on the original dataset splitted into several
    views and retains the classifier obtained for the best view.

    This class implements the MuMBo algorithm [1]_.

    Parameters
    ----------
    base_estimator : object, optional (default=DecisionTreeClassifier)
        Base estimator from which the boosted ensemble is built.
        Support for sample weighting is required, as well as proper `classes_`
        and `n_classes_` attributes. The default is a DecisionTreeClassifier
        with parameter ``max_depth=1``.

    n_estimators : integer, optional (default=50)
        Maximum number of estimators at which boosting is terminated.

    random_state : int, RandomState instance or None, optional (default=None)
        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`.

    best_view_mode : {"edge", "error"}, optional (default="edge")
        Mode used to select the best view at each iteration:

        - if ``best_view_mode == "edge"``, the best view is the view maximizing
          the edge value (variable δ (*delta*) in [1]_),
        - if ``best_view_mode == "error"``, the best view is the view
          minimizing the classification error.

    Attributes
    ----------
    estimators\_ : list of classifiers
        Collection of fitted sub-estimators.

    classes\_ : numpy.ndarray, shape = (n_classes,)
        Classes labels.

    n_classes\_ : int
        Number of classes.

    estimator_weights\_ : numpy.ndarray of floats, shape = (len(estimators\_),)
        Weights for each estimator in the boosted ensemble.

    estimator_errors_ : array of floats
        Empirical loss for each iteration.


    best_views\_ : numpy.ndarray of integers, shape = (len(estimators\_),)
        Indices of the best view for each estimator in the boosted ensemble.

    Examples
    --------
Dominique Benielli's avatar
Dominique Benielli committed
124
    >>> from multimodal.boosting.mumbo import MumboClassifier
Dominique Benielli's avatar
algo  
Dominique Benielli committed
125
126
127
128
129
    >>> from sklearn.datasets import load_iris
    >>> X, y = load_iris(return_X_y=True)
    >>> views_ind = [0, 2, 4]  # view 0: sepal data, view 1: petal data
    >>> clf = MumboClassifier(random_state=0)
    >>> clf.fit(X, y, views_ind)  # doctest: +NORMALIZE_WHITESPACE
130
    MumboClassifier(random_state=0)
Dominique Benielli's avatar
algo  
Dominique Benielli committed
131
132
133
134
135
    >>> print(clf.predict([[ 5.,  3.,  1.,  1.]]))
    [1]
    >>> views_ind = [[0, 2], [1, 3]]  # view 0: length data, view 1: width data
    >>> clf = MumboClassifier(random_state=0)
    >>> clf.fit(X, y, views_ind)  # doctest: +NORMALIZE_WHITESPACE
136
    MumboClassifier(random_state=0)
Dominique Benielli's avatar
algo  
Dominique Benielli committed
137
138
139
140
141
142
143
    >>> print(clf.predict([[ 5.,  3.,  1.,  1.]]))
    [1]

    >>> from sklearn.tree import DecisionTreeClassifier
    >>> base_estimator = DecisionTreeClassifier(max_depth=2)
    >>> clf = MumboClassifier(base_estimator=base_estimator, random_state=0)
    >>> clf.fit(X, y, views_ind)  # doctest: +NORMALIZE_WHITESPACE
144
145
    MumboClassifier(base_estimator=DecisionTreeClassifier(max_depth=2),
                    random_state=0)
Dominique Benielli's avatar
algo  
Dominique Benielli committed
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
    >>> print(clf.predict([[ 5.,  3.,  1.,  1.]]))
    [1]

    See also
    --------
    sklearn.ensemble.AdaBoostClassifier,
    sklearn.ensemble.GradientBoostingClassifier,
    sklearn.tree.DecisionTreeClassifier

    References
    ----------
    .. [1] Sokol Koço,
           "Tackling the uneven views problem with cooperation based ensemble
           learning methods",
           PhD Thesis, Aix-Marseille Université, 2013,
           http://www.theses.fr/en/2013AIXM4101.
    """

    def __init__(self,
                 base_estimator=None,
                 n_estimators=50,
                 random_state=None,
                 best_view_mode="edge"):

        super(MumboClassifier, self).__init__(
            base_estimator=base_estimator,
            n_estimators=n_estimators)

        self.random_state = random_state
        self.best_view_mode = self._validate_best_view_mode(best_view_mode)

    def _validate_estimator(self):
        """Check the estimator and set the base_estimator_ attribute."""
        super(MumboClassifier, self)._validate_estimator(
            default=DecisionTreeClassifier(max_depth=1))
        if not has_fit_parameter(self.base_estimator_, "sample_weight"):
            raise ValueError("%s doesn't support sample_weight."
                             % self.base_estimator_.__class__.__name__)

    def _validate_best_view_mode(self, best_view_mode):
        """Ensure that best_view_mode has a proper value."""
        if best_view_mode not in ("edge", "error"):
            raise ValueError('best_view_mode value must be either "edge" '
                             + 'or "error"')
        return best_view_mode

    def _init_var(self, n_views, y):
        "Create and initialize the variables used by the MuMBo algorithm."
        n_classes = self.n_classes_
        n_samples = y.shape[0]

        cost = np.ones((n_views, n_samples, n_classes))
        cost[:, np.arange(n_samples), y] = -(n_classes-1)

        cost_global = np.ones((n_samples, n_classes))
        cost_global[np.arange(n_samples), y] = -(n_classes-1)

        label_score = np.zeros((n_views, n_samples, n_classes))

        label_score_global = np.zeros((n_samples, n_classes))

        predicted_classes = np.empty((n_views, n_samples), dtype=np.int64)

        return (cost, cost_global, label_score, label_score_global,
                predicted_classes)

    def _compute_edge_global(self, cost_global, predicted_classes, y):
        """Compute edge values for the global cost matrix."""
        n_samples = y.shape[0]
        edge_global = - np.sum(
            cost_global[np.arange(n_samples), predicted_classes], axis=1) \
            / (np.sum(cost_global)
               - np.sum(cost_global[np.arange(n_samples), y]))
        return edge_global

    def _compute_dist(self, cost, y):
        """Compute the sample distribution (i.e. the weights to use)."""
        n_samples = y.shape[0]
        # dist is forced to be c-contiguous so that sub-arrays of dist used
        # as weights for the weak classifiers are also c-contiguous, which is
        # required by some scikit-learn classifiers (for example
        # sklearn.svm.SVC)
        dist = np.empty(cost.shape[:2], dtype=cost.dtype, order="C")
        # NOTE: In Sokol's PhD thesis, the formula for dist is mistakenly given
        # with a minus sign in section 2.2.2 page 31
        dist[:, :] = cost[:, np.arange(n_samples), y] \
            / np.sum(cost[:, np.arange(n_samples), y], axis=1)[:, np.newaxis]
        return dist

    def _compute_coop_coef(self, predicted_classes, y):
        """Compute the cooperation coefficients."""
        coop_coef = np.zeros(predicted_classes.shape)
        coop_coef[predicted_classes == y] = 1.
        coop_coef[:, np.logical_not(coop_coef.any(axis=0))] = 1.
        return coop_coef

    def _compute_edges(self, cost, predicted_classes, y):
        """Compute edge values for the cost matrices for all the views."""
        n_views = predicted_classes.shape[0]
        n_samples = y.shape[0]
        edges = - np.sum(
            cost[np.arange(n_views)[:, np.newaxis],
                 np.arange(n_samples)[np.newaxis, :],
                 predicted_classes[np.arange(n_views), :]], axis=1) \
            / (np.sum(cost, axis=(1, 2))
               - np.sum(cost[:, np.arange(n_samples), y], axis=1))
        return edges

    def _compute_alphas(self, edges):
        """Compute values of confidence rate alpha given edge values."""
Dominique Benielli's avatar
Dominique Benielli committed
256
257
        np.where(edges > 1.0, edges, 1.0)
        alphas = 0.5 * np.log((1. + edges) / (1. - edges))
Dominique Benielli's avatar
Dominique Benielli committed
258
        if np.any(np.isinf(alphas)):
259
260
261
262
            if isinstance(alphas, float):
                alphas = 1.0
            else:
                alphas[np.where(np.isinf(alphas))[0]] = 1.0
Dominique Benielli's avatar
Dominique Benielli committed
263
        if np.any(np.isnan(alphas)):
264
265
266
267
            if isinstance(alphas, float):
                alphas = 1.0
            else:
                alphas[np.where(np.isnan(alphas))[0]] = 1.0
Dominique Benielli's avatar
algo  
Dominique Benielli committed
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
        return alphas

    def _compute_cost_global(self, label_score_global, best_predicted_classes,
                             y, alpha):
        """Update label_score_global and compute the global cost matrix."""
        n_samples = y.shape[0]
        label_score_global[np.arange(n_samples), best_predicted_classes] \
            += alpha
        cost_global = np.exp(
            label_score_global
            - label_score_global[np.arange(n_samples), y][:, np.newaxis])
        cost_global[np.arange(n_samples), y] -= np.sum(cost_global, axis=1)
        return (cost_global, label_score_global)

    def _compute_cost(self, label_score, predicted_classes, y, alphas,
                      use_coop_coef=True):
        """Update label_score and compute the cost matrices for all views."""
        # use_coop_coef is a boolean parameter used to choose if the
        # cooperation coefficients are computed and taken into account when
        # updating the cost matrices.
        # It is introduced here for future explorations.
        n_views = predicted_classes.shape[0]
        n_samples = y.shape[0]
        if use_coop_coef:
            coop_coef = self._compute_coop_coef(predicted_classes, y)
            increment = alphas[:, np.newaxis] * coop_coef
        else:
            increment = alphas[:, np.newaxis]
        label_score[np.arange(n_views)[:, np.newaxis],
                    np.arange(n_samples)[np.newaxis, :],
                    predicted_classes[np.arange(n_views), :]] += increment
        cost = np.exp(
            label_score
            - label_score[:, np.arange(n_samples), y][:, :, np.newaxis])
        cost[:, np.arange(n_samples), y] -= np.sum(cost, axis=2)
        return (cost, label_score)

Dominique Benielli's avatar
Dominique Benielli committed
305
306
307
308
309
310
311
312
313
314
315
    def _compute_predictions(self, X):
        """Compute predictions for all the stored estimators on the data X."""
        n_samples = X.shape[0]
        n_estimators = len(self.estimators_)
        predictions = np.zeros((n_samples, n_estimators), dtype=np.int64)
        for ind_estimator, estimator in enumerate(self.estimators_):
            ind_view = self.best_views_[ind_estimator]
            predictions[:, ind_estimator] \
                = estimator.predict(X._extract_view(ind_view))
        return predictions

Dominique Benielli's avatar
algo  
Dominique Benielli committed
316
317
318
319
320
    def fit(self, X, y, views_ind=None):
        """Build a multimodal boosted classifier from the training set (X, y).

        Parameters
        ----------
Dominique Benielli's avatar
Dominique Benielli committed
321
322
323
324
325
        X : dict dictionary with all views
            or
            `MultiModalData` ,  `MultiModalArray`, `MultiModalSparseArray`
            or
            {array-like, sparse matrix}, shape = (n_samples, n_features)
Dominique Benielli's avatar
algo  
Dominique Benielli committed
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
            Training multi-view input samples.
            Sparse matrix can be CSC, CSR, COO, DOK, or LIL.
            COO, DOK and LIL are converted to CSR.

        y : array-like, shape = (n_samples,)
            Target values (class labels).

        views_ind : array-like (default=[0, n_features//2, n_features])
            Paramater specifying how to extract the data views from X:

            - If views_ind is a 1-D array of sorted integers, the entries
              indicate the limits of the slices used to extract the views,
              where view ``n`` is given by
              ``X[:, views_ind[n]:views_ind[n+1]]``.

              With this convention each view is therefore a view (in the NumPy
              sense) of X and no copy of the data is done.

            - If views_ind is an array of arrays of integers, then each array
              of integers ``views_ind[n]`` specifies the indices of the view
              ``n``, which is then given by ``X[:, views_ind[n]]``.

              With this convention each view creates therefore a partial copy
              of the data in X. This convention is thus more flexible but less
              efficient than the previous one.

        Returns
        -------
        self : object
            Returns self.
        """
        self.best_view_mode = self._validate_best_view_mode(
            self.best_view_mode)
        if (self.base_estimator is None or
                isinstance(self.base_estimator, (BaseDecisionTree,
                                                 BaseForest))):
            dtype = DTYPE
            accept_sparse = 'csc'
        else:
            dtype = None
            accept_sparse = ['csr', 'csc']
Dominique Benielli's avatar
Dominique Benielli committed
367

368
369
370
        self.X_ = self._global_X_transform(X, views_ind=views_ind)
        views_ind_, n_views = self.X_._validate_views_ind(self.X_.views_ind,
                                                          self.X_.shape[1])
Dominique Benielli's avatar
Dominique Benielli committed
371
        check_X_y(self.X_, y, accept_sparse=accept_sparse, dtype=dtype)
Baptiste Bauvin's avatar
Baptiste Bauvin committed
372
373
        if not isinstance(y, np.ndarray):
            y = np.asarray(y)
Dominique Benielli's avatar
Dominique Benielli committed
374
375
        check_classification_targets(y)
        self._validate_estimator()
Dominique Benielli's avatar
algo  
Dominique Benielli committed
376
377
378

        self.classes_, y = np.unique(y, return_inverse=True)
        self.n_classes_ = len(self.classes_)
379
        self.n_features_ = self.X_.shape[1]
Dominique Benielli's avatar
Dominique Benielli committed
380
        self.n_features_in_ = self.n_features_ 
Dominique Benielli's avatar
algo  
Dominique Benielli committed
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
        if self.n_classes_ == 1:
            # This case would lead to division by 0 when computing the cost
            # matrix so it needs special handling (but it is an obvious case as
            # there is only one single class in the data).
            self.estimators_ = []
            self.estimator_weights_ = np.array([], dtype=np.float64)
            self.estimator_errors_ = np.array([], dtype=np.float64)
            self.best_views_ = np.array([], dtype=np.int64)
            return

        self.estimators_ = []
        self.estimator_weights_ = np.zeros(self.n_estimators, dtype=np.float64)
        self.estimator_errors_ = np.zeros(self.n_estimators, dtype=np.float64)
        self.best_views_ = - np.ones(self.n_estimators, dtype=np.int64)

        random_state = check_random_state(self.random_state)
        (cost, cost_global, label_score, label_score_global,
         predicted_classes) = self._init_var(n_views, y)

        current_iteration = 0
        while True:
            estimators = []
            dist = self._compute_dist(cost, y)
            for ind_view in range(n_views):
                estimator = self._make_estimator(append=False,
                                                 random_state=random_state)
Dominique Benielli's avatar
Dominique Benielli committed
407
                estimator.fit(self.X_._extract_view(ind_view), y,
Dominique Benielli's avatar
algo  
Dominique Benielli committed
408
409
410
                              sample_weight=dist[ind_view, :])
                estimators.append(estimator)
                predicted_classes[ind_view, :] = estimator.predict(
Dominique Benielli's avatar
Dominique Benielli committed
411
                    self.X_._extract_view(ind_view))
Dominique Benielli's avatar
algo  
Dominique Benielli committed
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

            edges = self._compute_edge_global(
                    cost_global, predicted_classes, y)
            if self.best_view_mode == "edge":
                best_view = np.argmax(edges)
            else:  # self.best_view_mode == "error"
                n_errors = np.sum(predicted_classes != y, axis=1)
                best_view = np.argmin(n_errors)

            edge = edges[best_view]

            if (edge == 1.):
                self.estimator_weights_[0] = 1.
                self.estimator_weights_.resize((1, ))
                self.best_views_[0] = best_view
                self.best_views_.resize((1, ))
                self.estimators_ = [estimators[best_view]]
                self.estimator_errors_[0] = 0.
                self.estimator_errors_.resize((1, ))
                break

            self.estimator_errors_[current_iteration] = (
                np.average(cost_global[np.arange(y.shape[0]), y])
                * (-1. / (self.n_classes_-1)))

            alpha = self._compute_alphas(edge)
            self.estimator_weights_[current_iteration] = alpha
            self.best_views_[current_iteration] = best_view
            self.estimators_.append(estimators[best_view])

            if current_iteration == self.n_estimators-1:
                break

            cost_global, label_score_global = self._compute_cost_global(
                label_score_global, predicted_classes[best_view, :], y, alpha)

            edges = self._compute_edges(cost, predicted_classes, y)
            alphas = self._compute_alphas(edges)
            cost, label_score = self._compute_cost(
                label_score, predicted_classes, y, alphas)

            current_iteration += 1

        return self

    def decision_function(self, X):
        """Compute the decision function of X.

        Parameters
        ----------
Dominique Benielli's avatar
Dominique Benielli committed
462
463
        X : { array-like, sparse matrix},
            shape = (n_samples, n_views * n_features)
Dominique Benielli's avatar
algo  
Dominique Benielli committed
464
465
466
            Multi-view input samples.
            Sparse matrix can be CSC, CSR, COO, DOK, or LIL.
            COO, DOK and LIL are converted to CSR.
Dominique Benielli's avatar
Dominique Benielli committed
467
            maybe also MultimodalData
Dominique Benielli's avatar
algo  
Dominique Benielli committed
468
469
470
471
472
473
474
475
476
477
478
479
480
481

        Returns
        -------
        dec_fun : numpy.ndarray, shape = (n_samples, k)
            Decision function of the input samples.
            The order of outputs is the same of that of the `classes_`
            attribute.
            Binary classification is a special cases with ``k == 1``,
            otherwise ``k == n_classes``. For binary classification,
            values <=0 mean classification in the first class in ``classes_``
            and values >0 mean classification in the second class in
            ``classes_``.
        """
        check_is_fitted(self, ("estimators_", "estimator_weights_",
482
483
                               "best_views_", "n_classes_", "X_"))
        X = self._global_X_transform(X, views_ind=self.X_.views_ind)
Dominique Benielli's avatar
algo  
Dominique Benielli committed
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
        X = self._validate_X_predict(X)

        n_samples = X.shape[0]
        n_estimators = len(self.estimators_)
        n_classes = self.n_classes_

        predictions = self._compute_predictions(X)

        dec_func = np.zeros((n_samples, n_classes))
        for ind_estimator in range(n_estimators):
            dec_func[np.arange(n_samples), predictions[:, ind_estimator]] \
                += self.estimator_weights_[ind_estimator]

        if n_classes == 2:
            dec_func[:, 0] *= -1
            return np.sum(dec_func, axis=1)

        return dec_func

    def staged_decision_function(self, X):
        """Compute decision function of X for each boosting iteration.

        This method allows monitoring (i.e. determine error on testing set)
        after each boosting iteration.

        Parameters
        ----------
        X : {array-like, sparse matrix}, shape = (n_samples, n_features)
            Multi-view input samples.
            Sparse matrix can be CSC, CSR, COO, DOK, or LIL.
            COO, DOK and LIL are converted to CSR.
Dominique Benielli's avatar
Dominique Benielli committed
515
            maybe also MultimodalData
Dominique Benielli's avatar
algo  
Dominique Benielli committed
516
517
518
519
520
521
522
523
524
525
526
527
528
529

        Returns
        -------
        dec_fun : generator of numpy.ndarrays, shape = (n_samples, k)
            Decision function of the input samples.
            The order of outputs is the same of that of the `classes_`
            attribute.
            Binary classification is a special cases with ``k == 1``,
            otherwise ``k==n_classes``. For binary classification,
            values <=0 mean classification in the first class in ``classes_``
            and values >0 mean classification in the second class in
            ``classes_``.
        """
        check_is_fitted(self, ("estimators_", "estimator_weights_",
530
531
                               "n_classes_", "X_"))
        X = self._global_X_transform(X, views_ind=self.X_.views_ind)
Dominique Benielli's avatar
algo  
Dominique Benielli committed
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
        X = self._validate_X_predict(X)

        n_samples = X.shape[0]
        n_estimators = len(self.estimators_)
        n_classes = self.n_classes_

        predictions = self._compute_predictions(X)

        dec_func = np.zeros((n_samples, n_classes))
        for ind_estimator in range(n_estimators):
            dec_func[np.arange(n_samples), predictions[:, ind_estimator]] \
                += self.estimator_weights_[ind_estimator]
            if n_classes == 2:
                tmp_dec_func = np.array(dec_func)
                tmp_dec_func[:, 0] *= -1
                yield tmp_dec_func.sum(axis=1)
            else:
                yield np.array(dec_func)

    def predict(self, X):
        """Predict classes for X.

        The predicted class of an input sample is computed as the weighted mean
        prediction of the classifiers in the ensemble.

        Parameters
        ----------
        X : {array-like, sparse matrix}, shape = (n_samples, n_features)
            Multi-view input samples.
            Sparse matrix can be CSC, CSR, COO, DOK, or LIL.
            COO, DOK and LIL are converted to CSR.

        Returns
        -------
        y : numpy.ndarray, shape = (n_samples,)
            Predicted classes.
        """
569

Dominique Benielli's avatar
algo  
Dominique Benielli committed
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
        pred = self.decision_function(X)

        if self.n_classes_ == 2:
            return self.classes_.take(pred > 0, axis=0)

        return self.classes_.take(np.argmax(pred, axis=1), axis=0)

    def staged_predict(self, X):
        """Return staged predictions for X.

        The predicted class of an input sample is computed as the weighted mean
        prediction of the classifiers in the ensemble.

        This generator method yields the ensemble prediction after each
        iteration of boosting and therefore allows monitoring, such as to
        determine the prediction on a test set after each boost.

        Parameters
        ----------
        X : {array-like, sparse matrix} of shape = (n_samples, n_features)
            Multi-view input samples.
            Sparse matrix can be CSC, CSR, COO, DOK, or LIL.
            COO, DOK and LIL are converted to CSR.

        Returns
        -------
        y : generator of numpy.ndarrays, shape = (n_samples,)
            Predicted classes.
        """
        n_classes = self.n_classes_
        classes = self.classes_

        if n_classes == 2:
            for pred in self.staged_decision_function(X):
                yield np.array(classes.take(pred > 0, axis=0))
        else:
            for pred in self.staged_decision_function(X):
                yield np.array(classes.take(np.argmax(pred, axis=1), axis=0))

    def score(self, X, y):
        """Return the mean accuracy on the given test data and labels.

        Parameters
        ----------
        X : {array-like, sparse matrix} of shape = (n_samples, n_features)
            Multi-view test samples.
616
            Sparse matrix can be CSC, CSR
Dominique Benielli's avatar
algo  
Dominique Benielli committed
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
        y : array-like, shape = (n_samples,)
            True labels for X.

        Returns
        -------
        score : float
            Mean accuracy of self.predict(X) wrt. y.
        """
        return super(MumboClassifier, self).score(X, y)

    def staged_score(self, X, y):
        """Return staged mean accuracy on the given test data and labels.

        This generator method yields the ensemble score after each iteration of
        boosting and therefore allows monitoring, such as to determine the
        score on a test set after each boost.

        Parameters
        ----------
        X : {array-like, sparse matrix} of shape = (n_samples, n_features)
            Multi-view test samples.
            Sparse matrix can be CSC, CSR, COO, DOK, or LIL.
            COO, DOK and LIL are converted to CSR.
        y : array-like, shape = (n_samples,)
            True labels for X.

        Returns
        -------
        score : generator of floats
            Mean accuracy of self.staged_predict(X) wrt. y.
        """
        for y_pred in self.staged_predict(X):
            yield accuracy_score(y, y_pred)