import numpy as np
import scipy.sparse as sp

from sklearn.utils.testing import assert_array_equal
from sklearn.utils.testing import assert_equal
from sklearn.utils.testing import assert_almost_equal
from sklearn.utils.testing import assert_true
from sklearn.utils.testing import assert_false
from sklearn.utils.testing import assert_raises
from sklearn.utils.testing import assert_warns
from sklearn.utils.testing import assert_greater
from sklearn.utils.testing import assert_raise_message
from sklearn.multiclass import OneVsRestClassifier
from sklearn.multiclass import OneVsOneClassifier
from sklearn.multiclass import OutputCodeClassifier
from sklearn.utils.multiclass import check_classification_targets, type_of_target
from sklearn.utils import shuffle

from sklearn.metrics import precision_score
from sklearn.metrics import recall_score

from sklearn.svm import LinearSVC, SVC
from sklearn.naive_bayes import MultinomialNB
from sklearn.linear_model import (LinearRegression, Lasso, ElasticNet, Ridge,
                                  Perceptron, LogisticRegression)
from sklearn.tree import DecisionTreeClassifier, DecisionTreeRegressor
from sklearn.model_selection import GridSearchCV, cross_val_score
from sklearn.pipeline import Pipeline
from sklearn import svm
from sklearn import datasets
from sklearn.externals.six.moves import zip

iris = datasets.load_iris()
rng = np.random.RandomState(0)
perm = rng.permutation(iris.target.size)
iris.data = iris.data[perm]
iris.target = iris.target[perm]
n_classes = 3


def test_ovr_exceptions():
    ovr = OneVsRestClassifier(LinearSVC(random_state=0))
    assert_raises(ValueError, ovr.predict, [])

    # Fail on multioutput data
    assert_raises(ValueError, OneVsRestClassifier(MultinomialNB()).fit,
                  np.array([[1, 0], [0, 1]]),
                  np.array([[1, 2], [3, 1]]))
    assert_raises(ValueError, OneVsRestClassifier(MultinomialNB()).fit,
                  np.array([[1, 0], [0, 1]]),
                  np.array([[1.5, 2.4], [3.1, 0.8]]))


def test_check_classification_targets():
    # Test that check_classification_target return correct type. #5782
    y = np.array([0.0, 1.1, 2.0, 3.0])
    msg = type_of_target(y)
    assert_raise_message(ValueError, msg, check_classification_targets, y)


def test_ovr_fit_predict():
    # A classifier which implements decision_function.
    ovr = OneVsRestClassifier(LinearSVC(random_state=0))
    pred = ovr.fit(iris.data, iris.target).predict(iris.data)
    assert_equal(len(ovr.estimators_), n_classes)

    clf = LinearSVC(random_state=0)
    pred2 = clf.fit(iris.data, iris.target).predict(iris.data)
    assert_equal(np.mean(iris.target == pred), np.mean(iris.target == pred2))

    # A classifier which implements predict_proba.
    ovr = OneVsRestClassifier(MultinomialNB())
    pred = ovr.fit(iris.data, iris.target).predict(iris.data)
    assert_greater(np.mean(iris.target == pred), 0.65)


def test_ovr_partial_fit():
    # Test if partial_fit is working as intented
    X, y = shuffle(iris.data, iris.target, random_state=0)
    ovr = OneVsRestClassifier(MultinomialNB())
    ovr.partial_fit(X[:100], y[:100], np.unique(y))
    ovr.partial_fit(X[100:], y[100:])
    pred = ovr.predict(X)
    ovr2 = OneVsRestClassifier(MultinomialNB())
    pred2 = ovr2.fit(X, y).predict(X)

    assert_almost_equal(pred, pred2)
    assert_equal(len(ovr.estimators_), len(np.unique(y)))
    assert_greater(np.mean(y == pred), 0.65)

    # Test when mini batches doesn't have all classes
    ovr = OneVsRestClassifier(MultinomialNB())
    ovr.partial_fit(iris.data[:60], iris.target[:60], np.unique(iris.target))
    ovr.partial_fit(iris.data[60:], iris.target[60:])
    pred = ovr.predict(iris.data)
    ovr2 = OneVsRestClassifier(MultinomialNB())
    pred2 = ovr2.fit(iris.data, iris.target).predict(iris.data)

    assert_almost_equal(pred, pred2)
    assert_equal(len(ovr.estimators_), len(np.unique(iris.target)))
    assert_greater(np.mean(iris.target == pred), 0.65)


def test_ovr_ovo_regressor():
    # test that ovr and ovo work on regressors which don't have a decision_function
    ovr = OneVsRestClassifier(DecisionTreeRegressor())
    pred = ovr.fit(iris.data, iris.target).predict(iris.data)
    assert_equal(len(ovr.estimators_), n_classes)
    assert_array_equal(np.unique(pred), [0, 1, 2])
    # we are doing something sensible
    assert_greater(np.mean(pred == iris.target), .9)

    ovr = OneVsOneClassifier(DecisionTreeRegressor())
    pred = ovr.fit(iris.data, iris.target).predict(iris.data)
    assert_equal(len(ovr.estimators_), n_classes * (n_classes - 1) / 2)
    assert_array_equal(np.unique(pred), [0, 1, 2])
    # we are doing something sensible
    assert_greater(np.mean(pred == iris.target), .9)


def test_ovr_fit_predict_sparse():
    for sparse in [sp.csr_matrix, sp.csc_matrix, sp.coo_matrix, sp.dok_matrix,
                   sp.lil_matrix]:
        base_clf = MultinomialNB(alpha=1)

        X, Y = datasets.make_multilabel_classification(n_samples=100,
                                                       n_features=20,
                                                       n_classes=5,
                                                       n_labels=3,
                                                       length=50,
                                                       allow_unlabeled=True,
                                                       random_state=0)

        X_train, Y_train = X[:80], Y[:80]
        X_test = X[80:]

        clf = OneVsRestClassifier(base_clf).fit(X_train, Y_train)
        Y_pred = clf.predict(X_test)

        clf_sprs = OneVsRestClassifier(base_clf).fit(X_train, sparse(Y_train))
        Y_pred_sprs = clf_sprs.predict(X_test)

        assert_true(clf.multilabel_)
        assert_true(sp.issparse(Y_pred_sprs))
        assert_array_equal(Y_pred_sprs.toarray(), Y_pred)

        # Test predict_proba
        Y_proba = clf_sprs.predict_proba(X_test)

        # predict assigns a label if the probability that the
        # sample has the label is greater than 0.5.
        pred = Y_proba > .5
        assert_array_equal(pred, Y_pred_sprs.toarray())

        # Test decision_function
        clf_sprs = OneVsRestClassifier(svm.SVC()).fit(X_train, sparse(Y_train))
        dec_pred = (clf_sprs.decision_function(X_test) > 0).astype(int)
        assert_array_equal(dec_pred, clf_sprs.predict(X_test).toarray())


def test_ovr_always_present():
    # Test that ovr works with classes that are always present or absent.
    # Note: tests is the case where _ConstantPredictor is utilised
    X = np.ones((10, 2))
    X[:5, :] = 0

    # Build an indicator matrix where two features are always on.
    # As list of lists, it would be: [[int(i >= 5), 2, 3] for i in range(10)]
    y = np.zeros((10, 3))
    y[5:, 0] = 1
    y[:, 1] = 1
    y[:, 2] = 1

    ovr = OneVsRestClassifier(LogisticRegression())
    assert_warns(UserWarning, ovr.fit, X, y)
    y_pred = ovr.predict(X)
    assert_array_equal(np.array(y_pred), np.array(y))
    y_pred = ovr.decision_function(X)
    assert_equal(np.unique(y_pred[:, -2:]), 1)
    y_pred = ovr.predict_proba(X)
    assert_array_equal(y_pred[:, -1], np.ones(X.shape[0]))

    # y has a constantly absent label
    y = np.zeros((10, 2))
    y[5:, 0] = 1  # variable label
    ovr = OneVsRestClassifier(LogisticRegression())
    assert_warns(UserWarning, ovr.fit, X, y)
    y_pred = ovr.predict_proba(X)
    assert_array_equal(y_pred[:, -1], np.zeros(X.shape[0]))


def test_ovr_multiclass():
    # Toy dataset where features correspond directly to labels.
    X = np.array([[0, 0, 5], [0, 5, 0], [3, 0, 0], [0, 0, 6], [6, 0, 0]])
    y = ["eggs", "spam", "ham", "eggs", "ham"]
    Y = np.array([[0, 0, 1],
                  [0, 1, 0],
                  [1, 0, 0],
                  [0, 0, 1],
                  [1, 0, 0]])

    classes = set("ham eggs spam".split())

    for base_clf in (MultinomialNB(), LinearSVC(random_state=0),
                     LinearRegression(), Ridge(),
                     ElasticNet()):

        clf = OneVsRestClassifier(base_clf).fit(X, y)
        assert_equal(set(clf.classes_), classes)
        y_pred = clf.predict(np.array([[0, 0, 4]]))[0]
        assert_equal(set(y_pred), set("eggs"))

        # test input as label indicator matrix
        clf = OneVsRestClassifier(base_clf).fit(X, Y)
        y_pred = clf.predict([[0, 0, 4]])[0]
        assert_array_equal(y_pred, [0, 0, 1])


def test_ovr_binary():
    # Toy dataset where features correspond directly to labels.
    X = np.array([[0, 0, 5], [0, 5, 0], [3, 0, 0], [0, 0, 6], [6, 0, 0]])
    y = ["eggs", "spam", "spam", "eggs", "spam"]
    Y = np.array([[0, 1, 1, 0, 1]]).T

    classes = set("eggs spam".split())

    def conduct_test(base_clf, test_predict_proba=False):
        clf = OneVsRestClassifier(base_clf).fit(X, y)
        assert_equal(set(clf.classes_), classes)
        y_pred = clf.predict(np.array([[0, 0, 4]]))[0]
        assert_equal(set(y_pred), set("eggs"))

        if test_predict_proba:
            X_test = np.array([[0, 0, 4]])
            probabilities = clf.predict_proba(X_test)
            assert_equal(2, len(probabilities[0]))
            assert_equal(clf.classes_[np.argmax(probabilities, axis=1)],
                         clf.predict(X_test))

        # test input as label indicator matrix
        clf = OneVsRestClassifier(base_clf).fit(X, Y)
        y_pred = clf.predict([[3, 0, 0]])[0]
        assert_equal(y_pred, 1)

    for base_clf in (LinearSVC(random_state=0), LinearRegression(),
                     Ridge(), ElasticNet()):
        conduct_test(base_clf)

    for base_clf in (MultinomialNB(), SVC(probability=True),
                     LogisticRegression()):
        conduct_test(base_clf, test_predict_proba=True)


def test_ovr_multilabel():
    # Toy dataset where features correspond directly to labels.
    X = np.array([[0, 4, 5], [0, 5, 0], [3, 3, 3], [4, 0, 6], [6, 0, 0]])
    y = np.array([[0, 1, 1],
                  [0, 1, 0],
                  [1, 1, 1],
                  [1, 0, 1],
                  [1, 0, 0]])

    for base_clf in (MultinomialNB(), LinearSVC(random_state=0),
                     LinearRegression(), Ridge(),
                     ElasticNet(), Lasso(alpha=0.5)):
        clf = OneVsRestClassifier(base_clf).fit(X, y)
        y_pred = clf.predict([[0, 4, 4]])[0]
        assert_array_equal(y_pred, [0, 1, 1])
        assert_true(clf.multilabel_)


def test_ovr_fit_predict_svc():
    ovr = OneVsRestClassifier(svm.SVC())
    ovr.fit(iris.data, iris.target)
    assert_equal(len(ovr.estimators_), 3)
    assert_greater(ovr.score(iris.data, iris.target), .9)


def test_ovr_multilabel_dataset():
    base_clf = MultinomialNB(alpha=1)
    for au, prec, recall in zip((True, False), (0.51, 0.66), (0.51, 0.80)):
        X, Y = datasets.make_multilabel_classification(n_samples=100,
                                                       n_features=20,
                                                       n_classes=5,
                                                       n_labels=2,
                                                       length=50,
                                                       allow_unlabeled=au,
                                                       random_state=0)
        X_train, Y_train = X[:80], Y[:80]
        X_test, Y_test = X[80:], Y[80:]
        clf = OneVsRestClassifier(base_clf).fit(X_train, Y_train)
        Y_pred = clf.predict(X_test)

        assert_true(clf.multilabel_)
        assert_almost_equal(precision_score(Y_test, Y_pred, average="micro"),
                            prec,
                            decimal=2)
        assert_almost_equal(recall_score(Y_test, Y_pred, average="micro"),
                            recall,
                            decimal=2)


def test_ovr_multilabel_predict_proba():
    base_clf = MultinomialNB(alpha=1)
    for au in (False, True):
        X, Y = datasets.make_multilabel_classification(n_samples=100,
                                                       n_features=20,
                                                       n_classes=5,
                                                       n_labels=3,
                                                       length=50,
                                                       allow_unlabeled=au,
                                                       random_state=0)
        X_train, Y_train = X[:80], Y[:80]
        X_test = X[80:]
        clf = OneVsRestClassifier(base_clf).fit(X_train, Y_train)

        # decision function only estimator. Fails in current implementation.
        decision_only = OneVsRestClassifier(svm.SVR()).fit(X_train, Y_train)
        assert_raises(AttributeError, decision_only.predict_proba, X_test)

        # Estimator with predict_proba disabled, depending on parameters.
        decision_only = OneVsRestClassifier(svm.SVC(probability=False))
        decision_only.fit(X_train, Y_train)
        assert_raises(AttributeError, decision_only.predict_proba, X_test)

        Y_pred = clf.predict(X_test)
        Y_proba = clf.predict_proba(X_test)

        # predict assigns a label if the probability that the
        # sample has the label is greater than 0.5.
        pred = Y_proba > .5
        assert_array_equal(pred, Y_pred)


def test_ovr_single_label_predict_proba():
    base_clf = MultinomialNB(alpha=1)
    X, Y = iris.data, iris.target
    X_train, Y_train = X[:80], Y[:80]
    X_test = X[80:]
    clf = OneVsRestClassifier(base_clf).fit(X_train, Y_train)

    # decision function only estimator. Fails in current implementation.
    decision_only = OneVsRestClassifier(svm.SVR()).fit(X_train, Y_train)
    assert_raises(AttributeError, decision_only.predict_proba, X_test)

    Y_pred = clf.predict(X_test)
    Y_proba = clf.predict_proba(X_test)

    assert_almost_equal(Y_proba.sum(axis=1), 1.0)
    # predict assigns a label if the probability that the
    # sample has the label is greater than 0.5.
    pred = np.array([l.argmax() for l in Y_proba])
    assert_false((pred - Y_pred).any())


def test_ovr_multilabel_decision_function():
    X, Y = datasets.make_multilabel_classification(n_samples=100,
                                                   n_features=20,
                                                   n_classes=5,
                                                   n_labels=3,
                                                   length=50,
                                                   allow_unlabeled=True,
                                                   random_state=0)
    X_train, Y_train = X[:80], Y[:80]
    X_test = X[80:]
    clf = OneVsRestClassifier(svm.SVC()).fit(X_train, Y_train)
    assert_array_equal((clf.decision_function(X_test) > 0).astype(int),
                       clf.predict(X_test))


def test_ovr_single_label_decision_function():
    X, Y = datasets.make_classification(n_samples=100,
                                        n_features=20,
                                        random_state=0)
    X_train, Y_train = X[:80], Y[:80]
    X_test = X[80:]
    clf = OneVsRestClassifier(svm.SVC()).fit(X_train, Y_train)
    assert_array_equal(clf.decision_function(X_test).ravel() > 0,
                       clf.predict(X_test))


def test_ovr_gridsearch():
    ovr = OneVsRestClassifier(LinearSVC(random_state=0))
    Cs = [0.1, 0.5, 0.8]
    cv = GridSearchCV(ovr, {'estimator__C': Cs})
    cv.fit(iris.data, iris.target)
    best_C = cv.best_estimator_.estimators_[0].C
    assert_true(best_C in Cs)


def test_ovr_pipeline():
    # Test with pipeline of length one
    # This test is needed because the multiclass estimators may fail to detect
    # the presence of predict_proba or decision_function.
    clf = Pipeline([("tree", DecisionTreeClassifier())])
    ovr_pipe = OneVsRestClassifier(clf)
    ovr_pipe.fit(iris.data, iris.target)
    ovr = OneVsRestClassifier(DecisionTreeClassifier())
    ovr.fit(iris.data, iris.target)
    assert_array_equal(ovr.predict(iris.data), ovr_pipe.predict(iris.data))


def test_ovr_coef_():
    for base_classifier in [SVC(kernel='linear', random_state=0), LinearSVC(random_state=0)]:
        # SVC has sparse coef with sparse input data

        ovr = OneVsRestClassifier(base_classifier)
        for X in [iris.data, sp.csr_matrix(iris.data)]:
            # test with dense and sparse coef
            ovr.fit(X, iris.target)
            shape = ovr.coef_.shape
            assert_equal(shape[0], n_classes)
            assert_equal(shape[1], iris.data.shape[1])
            # don't densify sparse coefficients
            assert_equal(sp.issparse(ovr.estimators_[0].coef_), sp.issparse(ovr.coef_))


def test_ovr_coef_exceptions():
    # Not fitted exception!
    ovr = OneVsRestClassifier(LinearSVC(random_state=0))
    # lambda is needed because we don't want coef_ to be evaluated right away
    assert_raises(ValueError, lambda x: ovr.coef_, None)

    # Doesn't have coef_ exception!
    ovr = OneVsRestClassifier(DecisionTreeClassifier())
    ovr.fit(iris.data, iris.target)
    assert_raises(AttributeError, lambda x: ovr.coef_, None)


def test_ovo_exceptions():
    ovo = OneVsOneClassifier(LinearSVC(random_state=0))
    assert_raises(ValueError, ovo.predict, [])


def test_ovo_fit_on_list():
    # Test that OneVsOne fitting works with a list of targets and yields the
    # same output as predict from an array
    ovo = OneVsOneClassifier(LinearSVC(random_state=0))
    prediction_from_array = ovo.fit(iris.data, iris.target).predict(iris.data)
    iris_data_list = [list(a) for a in iris.data]
    prediction_from_list = ovo.fit(iris_data_list,
                                   list(iris.target)).predict(iris_data_list)
    assert_array_equal(prediction_from_array, prediction_from_list)


def test_ovo_fit_predict():
    # A classifier which implements decision_function.
    ovo = OneVsOneClassifier(LinearSVC(random_state=0))
    ovo.fit(iris.data, iris.target).predict(iris.data)
    assert_equal(len(ovo.estimators_), n_classes * (n_classes - 1) / 2)

    # A classifier which implements predict_proba.
    ovo = OneVsOneClassifier(MultinomialNB())
    ovo.fit(iris.data, iris.target).predict(iris.data)
    assert_equal(len(ovo.estimators_), n_classes * (n_classes - 1) / 2)


def test_ovo_partial_fit_predict():
    X, y = shuffle(iris.data, iris.target)
    ovo1 = OneVsOneClassifier(MultinomialNB())
    ovo1.partial_fit(X[:100], y[:100], np.unique(y))
    ovo1.partial_fit(X[100:], y[100:])
    pred1 = ovo1.predict(X)

    ovo2 = OneVsOneClassifier(MultinomialNB())
    ovo2.fit(X, y)
    pred2 = ovo2.predict(X)
    assert_equal(len(ovo1.estimators_), n_classes * (n_classes - 1) / 2)
    assert_greater(np.mean(y == pred1), 0.65)
    assert_almost_equal(pred1, pred2)

    # Test when mini-batches don't have all target classes
    ovo1 = OneVsOneClassifier(MultinomialNB())
    ovo1.partial_fit(iris.data[:60], iris.target[:60], np.unique(iris.target))
    ovo1.partial_fit(iris.data[60:], iris.target[60:])
    pred1 = ovo1.predict(iris.data)
    ovo2 = OneVsOneClassifier(MultinomialNB())
    pred2 = ovo2.fit(iris.data, iris.target).predict(iris.data)

    assert_almost_equal(pred1, pred2)
    assert_equal(len(ovo1.estimators_), len(np.unique(iris.target)))
    assert_greater(np.mean(iris.target == pred1), 0.65)


def test_ovo_decision_function():
    n_samples = iris.data.shape[0]

    ovo_clf = OneVsOneClassifier(LinearSVC(random_state=0))
    ovo_clf.fit(iris.data, iris.target)
    decisions = ovo_clf.decision_function(iris.data)

    assert_equal(decisions.shape, (n_samples, n_classes))
    assert_array_equal(decisions.argmax(axis=1), ovo_clf.predict(iris.data))

    # Compute the votes
    votes = np.zeros((n_samples, n_classes))

    k = 0
    for i in range(n_classes):
        for j in range(i + 1, n_classes):
            pred = ovo_clf.estimators_[k].predict(iris.data)
            votes[pred == 0, i] += 1
            votes[pred == 1, j] += 1
            k += 1

    # Extract votes and verify
    assert_array_equal(votes, np.round(decisions))

    for class_idx in range(n_classes):
        # For each sample and each class, there only 3 possible vote levels
        # because they are only 3 distinct class pairs thus 3 distinct
        # binary classifiers.
        # Therefore, sorting predictions based on votes would yield
        # mostly tied predictions:
        assert_true(set(votes[:, class_idx]).issubset(set([0., 1., 2.])))

        # The OVO decision function on the other hand is able to resolve
        # most of the ties on this data as it combines both the vote counts
        # and the aggregated confidence levels of the binary classifiers
        # to compute the aggregate decision function. The iris dataset
        # has 150 samples with a couple of duplicates. The OvO decisions
        # can resolve most of the ties:
        assert_greater(len(np.unique(decisions[:, class_idx])), 146)


def test_ovo_gridsearch():
    ovo = OneVsOneClassifier(LinearSVC(random_state=0))
    Cs = [0.1, 0.5, 0.8]
    cv = GridSearchCV(ovo, {'estimator__C': Cs})
    cv.fit(iris.data, iris.target)
    best_C = cv.best_estimator_.estimators_[0].C
    assert_true(best_C in Cs)


def test_ovo_ties():
    # Test that ties are broken using the decision function,
    # not defaulting to the smallest label
    X = np.array([[1, 2], [2, 1], [-2, 1], [-2, -1]])
    y = np.array([2, 0, 1, 2])
    multi_clf = OneVsOneClassifier(Perceptron(shuffle=False))
    ovo_prediction = multi_clf.fit(X, y).predict(X)
    ovo_decision = multi_clf.decision_function(X)

    # Classifiers are in order 0-1, 0-2, 1-2
    # Use decision_function to compute the votes and the normalized
    # sum_of_confidences, which is used to disambiguate when there is a tie in
    # votes.
    votes = np.round(ovo_decision)
    normalized_confidences = ovo_decision - votes

    # For the first point, there is one vote per class
    assert_array_equal(votes[0, :], 1)
    # For the rest, there is no tie and the prediction is the argmax
    assert_array_equal(np.argmax(votes[1:], axis=1), ovo_prediction[1:])
    # For the tie, the prediction is the class with the highest score
    assert_equal(ovo_prediction[0], normalized_confidences[0].argmax())


def test_ovo_ties2():
    # test that ties can not only be won by the first two labels
    X = np.array([[1, 2], [2, 1], [-2, 1], [-2, -1]])
    y_ref = np.array([2, 0, 1, 2])

    # cycle through labels so that each label wins once
    for i in range(3):
        y = (y_ref + i) % 3
        multi_clf = OneVsOneClassifier(Perceptron(shuffle=False))
        ovo_prediction = multi_clf.fit(X, y).predict(X)
        assert_equal(ovo_prediction[0], i % 3)


def test_ovo_string_y():
    # Test that the OvO doesn't mess up the encoding of string labels
    X = np.eye(4)
    y = np.array(['a', 'b', 'c', 'd'])

    ovo = OneVsOneClassifier(LinearSVC())
    ovo.fit(X, y)
    assert_array_equal(y, ovo.predict(X))


def test_ecoc_exceptions():
    ecoc = OutputCodeClassifier(LinearSVC(random_state=0))
    assert_raises(ValueError, ecoc.predict, [])


def test_ecoc_fit_predict():
    # A classifier which implements decision_function.
    ecoc = OutputCodeClassifier(LinearSVC(random_state=0),
                                code_size=2, random_state=0)
    ecoc.fit(iris.data, iris.target).predict(iris.data)
    assert_equal(len(ecoc.estimators_), n_classes * 2)

    # A classifier which implements predict_proba.
    ecoc = OutputCodeClassifier(MultinomialNB(), code_size=2, random_state=0)
    ecoc.fit(iris.data, iris.target).predict(iris.data)
    assert_equal(len(ecoc.estimators_), n_classes * 2)


def test_ecoc_gridsearch():
    ecoc = OutputCodeClassifier(LinearSVC(random_state=0),
                                random_state=0)
    Cs = [0.1, 0.5, 0.8]
    cv = GridSearchCV(ecoc, {'estimator__C': Cs})
    cv.fit(iris.data, iris.target)
    best_C = cv.best_estimator_.estimators_[0].C
    assert_true(best_C in Cs)


def test_pairwise_indices():
    clf_precomputed = svm.SVC(kernel='precomputed')
    X, y = iris.data, iris.target

    ovr_false = OneVsOneClassifier(clf_precomputed)
    linear_kernel = np.dot(X, X.T)
    ovr_false.fit(linear_kernel, y)

    n_estimators = len(ovr_false.estimators_)
    precomputed_indices = ovr_false.pairwise_indices_

    for idx in precomputed_indices:
        assert_equal(idx.shape[0] * n_estimators / (n_estimators - 1),
                     linear_kernel.shape[0])


def test_pairwise_attribute():
    clf_precomputed = svm.SVC(kernel='precomputed')
    clf_notprecomputed = svm.SVC()

    for MultiClassClassifier in [OneVsRestClassifier, OneVsOneClassifier]:
        ovr_false = MultiClassClassifier(clf_notprecomputed)
        assert_false(ovr_false._pairwise)

        ovr_true = MultiClassClassifier(clf_precomputed)
        assert_true(ovr_true._pairwise)


def test_pairwise_cross_val_score():
    clf_precomputed = svm.SVC(kernel='precomputed')
    clf_notprecomputed = svm.SVC(kernel='linear')

    X, y = iris.data, iris.target

    for MultiClassClassifier in [OneVsRestClassifier, OneVsOneClassifier]:
        ovr_false = MultiClassClassifier(clf_notprecomputed)
        ovr_true = MultiClassClassifier(clf_precomputed)

        linear_kernel = np.dot(X, X.T)
        score_precomputed = cross_val_score(ovr_true, linear_kernel, y)
        score_linear = cross_val_score(ovr_false, X, y)
        assert_array_equal(score_precomputed, score_linear)