import pickle
from io import BytesIO
import numpy as np
import scipy.sparse
import pytest
from sklearn.datasets import load_digits, load_iris
from sklearn.model_selection import train_test_split
from sklearn.model_selection import cross_val_score
from sklearn.utils._testing import assert_almost_equal
from sklearn.utils._testing import assert_array_equal
from sklearn.utils._testing import assert_array_almost_equal
from sklearn.utils._testing import assert_raises
from sklearn.utils._testing import assert_raise_message
from sklearn.utils._testing import assert_warns
from sklearn.utils._testing import assert_no_warnings
from sklearn.naive_bayes import GaussianNB, BernoulliNB
from sklearn.naive_bayes import MultinomialNB, ComplementNB
from sklearn.naive_bayes import CategoricalNB
from sklearn.naive_bayes import BaseNB, BaseDiscreteNB
# Data is just 6 separable points in the plane
X = np.array([[-2, -1], [-1, -1], [-1, -2], [1, 1], [1, 2], [2, 1]])
y = np.array([1, 1, 1, 2, 2, 2])
# A bit more random tests
rng = np.random.RandomState(0)
X1 = rng.normal(size=(10, 3))
y1 = (rng.normal(size=(10)) > 0).astype(np.int)
# Data is 6 random integer points in a 100 dimensional space classified to
# three classes.
X2 = rng.randint(5, size=(6, 100))
y2 = np.array([1, 1, 2, 2, 3, 3])
def test_gnb():
# Gaussian Naive Bayes classification.
# This checks that GaussianNB implements fit and predict and returns
# correct values for a simple toy dataset.
clf = GaussianNB()
y_pred = clf.fit(X, y).predict(X)
assert_array_equal(y_pred, y)
y_pred_proba = clf.predict_proba(X)
y_pred_log_proba = clf.predict_log_proba(X)
assert_array_almost_equal(np.log(y_pred_proba), y_pred_log_proba, 8)
# Test whether label mismatch between target y and classes raises
# an Error
# FIXME Remove this test once the more general partial_fit tests are merged
assert_raises(ValueError, GaussianNB().partial_fit, X, y, classes=[0, 1])
def test_gnb_prior():
# Test whether class priors are properly set.
clf = GaussianNB().fit(X, y)
assert_array_almost_equal(np.array([3, 3]) / 6.0,
clf.class_prior_, 8)
clf.fit(X1, y1)
# Check that the class priors sum to 1
assert_array_almost_equal(clf.class_prior_.sum(), 1)
def test_gnb_sample_weight():
"""Test whether sample weights are properly used in GNB. """
# Sample weights all being 1 should not change results
sw = np.ones(6)
clf = GaussianNB().fit(X, y)
clf_sw = GaussianNB().fit(X, y, sw)
assert_array_almost_equal(clf.theta_, clf_sw.theta_)
assert_array_almost_equal(clf.sigma_, clf_sw.sigma_)
# Fitting twice with half sample-weights should result
# in same result as fitting once with full weights
sw = rng.rand(y.shape[0])
clf1 = GaussianNB().fit(X, y, sample_weight=sw)
clf2 = GaussianNB().partial_fit(X, y, classes=[1, 2], sample_weight=sw / 2)
clf2.partial_fit(X, y, sample_weight=sw / 2)
assert_array_almost_equal(clf1.theta_, clf2.theta_)
assert_array_almost_equal(clf1.sigma_, clf2.sigma_)
# Check that duplicate entries and correspondingly increased sample
# weights yield the same result
ind = rng.randint(0, X.shape[0], 20)
sample_weight = np.bincount(ind, minlength=X.shape[0])
clf_dupl = GaussianNB().fit(X[ind], y[ind])
clf_sw = GaussianNB().fit(X, y, sample_weight)
assert_array_almost_equal(clf_dupl.theta_, clf_sw.theta_)
assert_array_almost_equal(clf_dupl.sigma_, clf_sw.sigma_)
def test_gnb_neg_priors():
"""Test whether an error is raised in case of negative priors"""
clf = GaussianNB(priors=np.array([-1., 2.]))
assert_raises(ValueError, clf.fit, X, y)
def test_gnb_priors():
"""Test whether the class prior override is properly used"""
clf = GaussianNB(priors=np.array([0.3, 0.7])).fit(X, y)
assert_array_almost_equal(clf.predict_proba([[-0.1, -0.1]]),
np.array([[0.825303662161683,
0.174696337838317]]), 8)
assert_array_almost_equal(clf.class_prior_, np.array([0.3, 0.7]))
def test_gnb_priors_sum_isclose():
# test whether the class prior sum is properly tested"""
X = np.array([[-1, -1], [-2, -1], [-3, -2], [-4, -5], [-5, -4],
[1, 1], [2, 1], [3, 2], [4, 4], [5, 5]])
priors = np.array([0.08, 0.14, 0.03, 0.16, 0.11, 0.16, 0.07, 0.14,
0.11, 0.0])
Y = np.array([1, 2, 3, 4, 5, 6, 7, 8, 9, 10])
clf = GaussianNB(priors=priors)
# smoke test for issue #9633
clf.fit(X, Y)
def test_gnb_wrong_nb_priors():
""" Test whether an error is raised if the number of prior is different
from the number of class"""
clf = GaussianNB(priors=np.array([.25, .25, .25, .25]))
assert_raises(ValueError, clf.fit, X, y)
def test_gnb_prior_greater_one():
"""Test if an error is raised if the sum of prior greater than one"""
clf = GaussianNB(priors=np.array([2., 1.]))
assert_raises(ValueError, clf.fit, X, y)
def test_gnb_prior_large_bias():
"""Test if good prediction when class prior favor largely one class"""
clf = GaussianNB(priors=np.array([0.01, 0.99]))
clf.fit(X, y)
assert clf.predict([[-0.1, -0.1]]) == np.array([2])
def test_gnb_check_update_with_no_data():
""" Test when the partial fit is called without any data"""
# Create an empty array
prev_points = 100
mean = 0.
var = 1.
x_empty = np.empty((0, X.shape[1]))
tmean, tvar = GaussianNB._update_mean_variance(prev_points, mean,
var, x_empty)
assert tmean == mean
assert tvar == var
def test_gnb_pfit_wrong_nb_features():
"""Test whether an error is raised when the number of feature changes
between two partial fit"""
clf = GaussianNB()
# Fit for the first time the GNB
clf.fit(X, y)
# Partial fit a second time with an incoherent X
assert_raises(ValueError, clf.partial_fit, np.hstack((X, X)), y)
def test_gnb_partial_fit():
clf = GaussianNB().fit(X, y)
clf_pf = GaussianNB().partial_fit(X, y, np.unique(y))
assert_array_almost_equal(clf.theta_, clf_pf.theta_)
assert_array_almost_equal(clf.sigma_, clf_pf.sigma_)
assert_array_almost_equal(clf.class_prior_, clf_pf.class_prior_)
clf_pf2 = GaussianNB().partial_fit(X[0::2, :], y[0::2], np.unique(y))
clf_pf2.partial_fit(X[1::2], y[1::2])
assert_array_almost_equal(clf.theta_, clf_pf2.theta_)
assert_array_almost_equal(clf.sigma_, clf_pf2.sigma_)
assert_array_almost_equal(clf.class_prior_, clf_pf2.class_prior_)
def test_gnb_naive_bayes_scale_invariance():
# Scaling the data should not change the prediction results
iris = load_iris()
X, y = iris.data, iris.target
labels = [GaussianNB().fit(f * X, y).predict(f * X)
for f in [1E-10, 1, 1E10]]
assert_array_equal(labels[0], labels[1])
assert_array_equal(labels[1], labels[2])
@pytest.mark.parametrize("cls", [MultinomialNB, BernoulliNB, CategoricalNB])
def test_discretenb_prior(cls):
# Test whether class priors are properly set.
clf = cls().fit(X2, y2)
assert_array_almost_equal(np.log(np.array([2, 2, 2]) / 6.0),
clf.class_log_prior_, 8)
@pytest.mark.parametrize("cls", [MultinomialNB, BernoulliNB, CategoricalNB])
def test_discretenb_partial_fit(cls):
clf1 = cls()
clf1.fit([[0, 1], [1, 0], [1, 1]], [0, 1, 1])
clf2 = cls()
clf2.partial_fit([[0, 1], [1, 0], [1, 1]], [0, 1, 1], classes=[0, 1])
assert_array_equal(clf1.class_count_, clf2.class_count_)
if cls is CategoricalNB:
for i in range(len(clf1.category_count_)):
assert_array_equal(clf1.category_count_[i],
clf2.category_count_[i])
else:
assert_array_equal(clf1.feature_count_, clf2.feature_count_)
clf3 = cls()
# all categories have to appear in the first partial fit
clf3.partial_fit([[0, 1]], [0], classes=[0, 1])
clf3.partial_fit([[1, 0]], [1])
clf3.partial_fit([[1, 1]], [1])
assert_array_equal(clf1.class_count_, clf3.class_count_)
if cls is CategoricalNB:
# the categories for each feature of CategoricalNB are mapped to an
# index chronologically with each call of partial fit and therefore
# the category_count matrices cannot be compared for equality
for i in range(len(clf1.category_count_)):
assert_array_equal(clf1.category_count_[i].shape,
clf3.category_count_[i].shape)
assert_array_equal(np.sum(clf1.category_count_[i], axis=1),
np.sum(clf3.category_count_[i], axis=1))
# assert category 0 occurs 1x in the first class and 0x in the 2nd
# class
assert_array_equal(clf1.category_count_[0][0], np.array([1, 0]))
# assert category 1 occurs 0x in the first class and 2x in the 2nd
# class
assert_array_equal(clf1.category_count_[0][1], np.array([0, 2]))
# assert category 0 occurs 0x in the first class and 1x in the 2nd
# class
assert_array_equal(clf1.category_count_[1][0], np.array([0, 1]))
# assert category 1 occurs 1x in the first class and 1x in the 2nd
# class
assert_array_equal(clf1.category_count_[1][1], np.array([1, 1]))
else:
assert_array_equal(clf1.feature_count_, clf3.feature_count_)
@pytest.mark.parametrize('cls', [BernoulliNB, MultinomialNB, GaussianNB,
CategoricalNB])
def test_discretenb_pickle(cls):
# Test picklability of discrete naive Bayes classifiers
clf = cls().fit(X2, y2)
y_pred = clf.predict(X2)
store = BytesIO()
pickle.dump(clf, store)
clf = pickle.load(BytesIO(store.getvalue()))
assert_array_equal(y_pred, clf.predict(X2))
# Test pickling of estimator trained with partial_fit
clf2 = cls().partial_fit(X2[:3], y2[:3], classes=np.unique(y2))
clf2.partial_fit(X2[3:], y2[3:])
store = BytesIO()
pickle.dump(clf2, store)
clf2 = pickle.load(BytesIO(store.getvalue()))
assert_array_equal(y_pred, clf2.predict(X2))
@pytest.mark.parametrize('cls', [BernoulliNB, MultinomialNB, GaussianNB,
CategoricalNB])
def test_discretenb_input_check_fit(cls):
# Test input checks for the fit method
# check shape consistency for number of samples at fit time
assert_raises(ValueError, cls().fit, X2, y2[:-1])
# check shape consistency for number of input features at predict time
clf = cls().fit(X2, y2)
assert_raises(ValueError, clf.predict, X2[:, :-1])
@pytest.mark.parametrize('cls', [BernoulliNB, MultinomialNB, CategoricalNB])
def test_discretenb_input_check_partial_fit(cls):
# check shape consistency
assert_raises(ValueError, cls().partial_fit, X2, y2[:-1],
classes=np.unique(y2))
# classes is required for first call to partial fit
assert_raises(ValueError, cls().partial_fit, X2, y2)
# check consistency of consecutive classes values
clf = cls()
clf.partial_fit(X2, y2, classes=np.unique(y2))
assert_raises(ValueError, clf.partial_fit, X2, y2,
classes=np.arange(42))
# check consistency of input shape for partial_fit
assert_raises(ValueError, clf.partial_fit, X2[:, :-1], y2)
# check consistency of input shape for predict
assert_raises(ValueError, clf.predict, X2[:, :-1])
def test_discretenb_predict_proba():
# Test discrete NB classes' probability scores
# The 100s below distinguish Bernoulli from multinomial.
# FIXME: write a test to show this.
X_bernoulli = [[1, 100, 0], [0, 1, 0], [0, 100, 1]]
X_multinomial = [[0, 1], [1, 3], [4, 0]]
# test binary case (1-d output)
y = [0, 0, 2] # 2 is regression test for binary case, 02e673
for cls, X in zip([BernoulliNB, MultinomialNB],
[X_bernoulli, X_multinomial]):
clf = cls().fit(X, y)
assert clf.predict(X[-1:]) == 2
assert clf.predict_proba([X[0]]).shape == (1, 2)
assert_array_almost_equal(clf.predict_proba(X[:2]).sum(axis=1),
np.array([1., 1.]), 6)
# test multiclass case (2-d output, must sum to one)
y = [0, 1, 2]
for cls, X in zip([BernoulliNB, MultinomialNB],
[X_bernoulli, X_multinomial]):
clf = cls().fit(X, y)
assert clf.predict_proba(X[0:1]).shape == (1, 3)
assert clf.predict_proba(X[:2]).shape == (2, 3)
assert_almost_equal(np.sum(clf.predict_proba([X[1]])), 1)
assert_almost_equal(np.sum(clf.predict_proba([X[-1]])), 1)
assert_almost_equal(np.sum(np.exp(clf.class_log_prior_)), 1)
assert_almost_equal(np.sum(np.exp(clf.intercept_)), 1)
@pytest.mark.parametrize('cls', [BernoulliNB, MultinomialNB, CategoricalNB])
def test_discretenb_uniform_prior(cls):
# Test whether discrete NB classes fit a uniform prior
# when fit_prior=False and class_prior=None
Loading ...