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_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_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, max_iter=4, tol=None))
ovo_prediction = multi_clf.fit(X, y).predict(X)
assert ovo_prediction[0] == i % 3
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 (idx.shape[0] * n_estimators /
(n_estimators - 1) == linear_kernel.shape[0])
def test_ovo_float_y():
# Test that the OvO errors on float targets
X = iris.data
y = iris.data[:, 0]
ovo = OneVsOneClassifier(LinearSVC())
assert_raise_message(ValueError, "Unknown label type", ovo.fit, X, y)
def test_ovo_one_class():
# Test error for OvO with one class
X = np.eye(4)
y = np.array(['a'] * 4)
ovo = OneVsOneClassifier(LinearSVC())
assert_raise_message(ValueError, "when only one class", ovo.fit, X, y)
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 best_C in Cs
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 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 len(ovo.estimators_) == n_classes * (n_classes - 1) / 2
def test_ovo_decision_function():
n_samples = iris.data.shape[0]
ovo_clf = OneVsOneClassifier(LinearSVC(random_state=0))
# first binary
ovo_clf.fit(iris.data, iris.target == 0)
decisions = ovo_clf.decision_function(iris.data)
assert decisions.shape == (n_samples, )
# then multi-class
ovo_clf.fit(iris.data, iris.target)
decisions = ovo_clf.decision_function(iris.data)
assert 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 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 len(np.unique(decisions[:, class_idx])) > 146
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, max_iter=4, tol=None))
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 ovo_prediction[0] == normalized_confidences[0].argmax()
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 len(ovr.estimators_) == n_classes
assert_array_equal(np.unique(pred), [0, 1, 2])
# we are doing something sensible
assert np.mean(pred == iris.target) > .9
ovr = OneVsOneClassifier(DecisionTreeRegressor())
pred = ovr.fit(iris.data, iris.target).predict(iris.data)
assert len(ovr.estimators_) == n_classes * (n_classes - 1) / 2
assert_array_equal(np.unique(pred), [0, 1, 2])
# we are doing something sensible
assert np.mean(pred == iris.target) > .9
def test_ovo_partial_fit_predict():
temp = datasets.load_iris()
X, y = temp.data, temp.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 len(ovo1.estimators_) == n_classes * (n_classes - 1) / 2
assert np.mean(y == pred1) > 0.65
assert_almost_equal(pred1, pred2)
# Test when mini-batches have binary target classes
ovo1 = OneVsOneClassifier(MultinomialNB())
ovo1.partial_fit(X[:60], y[:60], np.unique(y))
ovo1.partial_fit(X[60:], y[60:])
pred1 = ovo1.predict(X)
ovo2 = OneVsOneClassifier(MultinomialNB())
pred2 = ovo2.fit(X, y).predict(X)
assert_almost_equal(pred1, pred2)
assert len(ovo1.estimators_) == len(np.unique(y))
assert np.mean(y == pred1) > 0.65
ovo = OneVsOneClassifier(MultinomialNB())
X = np.random.rand(14, 2)
y = [1, 1, 2, 3, 3, 0, 0, 4, 4, 4, 4, 4, 2, 2]
ovo.partial_fit(X[:7], y[:7], [0, 1, 2, 3, 4])
ovo.partial_fit(X[7:], y[7:])
pred = ovo.predict(X)
ovo2 = OneVsOneClassifier(MultinomialNB())
pred2 = ovo2.fit(X, y).predict(X)
assert_almost_equal(pred, pred2)
# raises error when mini-batch does not have classes from all_classes
ovo = OneVsOneClassifier(MultinomialNB())
error_y = [0, 1, 2, 3, 4, 5, 2]
message_re = escape("Mini-batch contains {0} while "
"it must be subset of {1}".format(
np.unique(error_y), np.unique(y)))
assert_raises_regexp(ValueError, message_re, ovo.partial_fit, X[:7],
error_y, np.unique(y))
# test partial_fit only exists if estimator has it:
ovr = OneVsOneClassifier(SVC())
assert not hasattr(ovr, "partial_fit")
def test_ovo_exceptions():
ovo = OneVsOneClassifier(LinearSVC(random_state=0))
assert_raises(ValueError, ovo.predict, [])