def test_clear(self):
"""
Tests KNN clearing (:func:`~fatf.utils.models.models.KNN.clear`).
"""
k = 2
clf = fumm.KNN(k=k)
self._test_unfitted_internals(clf, init_k=k, init_is_classifier=True)
# Clearing an unfitted model
with pytest.raises(UnfittedModelError) as exception_info:
clf.clear()
assert self.unfitted_model_error == str(exception_info.value)
# Clearing a fitted model
clf.fit(self.X, self.y)
self._test_fitted_internals(clf, False, self.X, self.y, self.X_n,
self.majority_label,
self.X_categorical_indices,
self.X_numerical_indices, self.unique_y,
self.unique_y_counts,
self.unique_y_probabilities)
clf.clear()
self._test_unfitted_internals(clf, init_k=k, init_is_classifier=True)
def test_predict_proba(self):
"""
Tests probas (:func:`~fatf.utils.models.models.KNN.predict_proba`).
"""
# pylint: disable=too-many-statements
# Regressor error
k = 3
clf = fumm.KNN(k=k, mode='r')
self._test_unfitted_internals(clf, init_k=k, init_is_classifier=False)
clf.fit(self.X, self.y)
self._test_fitted_internals(clf, False, self.X, self.y, self.X_n,
self.majority_label_regressor,
self.X_categorical_indices,
self.X_numerical_indices)
with pytest.raises(RuntimeError) as exception_info:
clf.predict_proba(self.X_test)
assert str(exception_info.value) == self.runtime_error
# Test other errors...
k = 3
clf = fumm.KNN(k=k)
self._test_unfitted_internals(clf, init_k=k, init_is_classifier=True)
# Unfitted model
with pytest.raises(UnfittedModelError) as exception_info:
clf.predict_proba(self.X_test)
assert self.unfitted_model_error == str(exception_info.value)
# ...
clf.fit(self.X, self.y)
self._test_fitted_internals(clf, False, self.X, self.y, self.X_n,
self.majority_label,
self.X_categorical_indices,
self.X_numerical_indices, self.unique_y,
self.unique_y_counts,
self.unique_y_probabilities)
# X is not 2D
with pytest.raises(IncorrectShapeError) as exception_info:
clf.predict_proba(self.X_3D)
assert self.incorrect_shape_error_singular == str(exception_info.value)
with pytest.raises(IncorrectShapeError) as exception_info:
clf.predict_proba(self.y)
assert self.incorrect_shape_error_singular == str(exception_info.value)
# dtype is not similar to the training data
with pytest.raises(ValueError) as exception_info:
clf.predict_proba(self.X_cat)
assert self.value_error_dtype == str(exception_info.value)
with pytest.raises(ValueError) as exception_info:
clf.predict_proba(self.X_cat_struct)
assert self.value_error_dtype == str(exception_info.value)
with pytest.raises(ValueError) as exception_info:
clf.predict_proba(self.X_struct)
assert self.value_error_dtype == str(exception_info.value)
# Predict 0 examples
with pytest.raises(IncorrectShapeError) as exception_info:
clf.predict_proba(np.ones((0, 2), dtype=int))
assert self.incorrect_shape_error_rows == str(exception_info.value)
# The number of features disagrees...
# ...unstructured
with pytest.raises(IncorrectShapeError) as exception_info:
clf.predict_proba(self.X_distances)
assert str(exception_info.value).startswith(
self.incorrect_shape_error_columns)
# ...structured
clf.clear()
self._test_unfitted_internals(clf, init_k=k, init_is_classifier=True)
clf.fit(self.X_struct, self.y)
self._test_fitted_internals(clf, True, self.X_struct, self.y, self.X_n,
self.majority_label,
self.X_struct_categorical_indices,
self.X_struct_numerical_indices,
self.unique_y, self.unique_y_counts,
self.unique_y_probabilities)
with pytest.raises(ValueError) as exception_info:
clf.predict_proba(self.X_test_struct[['a']])
assert self.value_error_dtype == str(exception_info.value)
# Numerical classifier on unstructured
# Sample smaller than k
k = 3
clf = fumm.KNN(k=k)
self._test_unfitted_internals(clf, init_k=k, init_is_classifier=True)
clf.fit(self.X, self.y)
self._test_fitted_internals(clf, False, self.X, self.y, self.X_n,
self.majority_label,
self.X_categorical_indices,
self.X_numerical_indices, self.unique_y,
self.unique_y_counts,
self.unique_y_probabilities)
y_hat = clf.predict_proba(self.X_test)
assert np.isclose(y_hat, self.y_test_3_proba, atol=1e-3).all()
# Sample bigger than k
k = 10
clf = fumm.KNN(k=k)
self._test_unfitted_internals(clf, init_k=k, init_is_classifier=True)
clf.fit(self.X, self.y)
self._test_fitted_internals(clf, False, self.X, self.y, self.X_n,
self.majority_label,
self.X_categorical_indices,
self.X_numerical_indices, self.unique_y,
self.unique_y_counts,
self.unique_y_probabilities)
y_hat = clf.predict_proba(self.X_test)
y_true = np.full(
(y_hat.shape[0], self.y_test_3_trainig_proba.shape[0]),
fill_value=self.y_test_3_trainig_proba)
assert np.isclose(y_hat, y_true, atol=1e-3).all()
# Numerical classifier on structured
# Sample smaller than k
k = 3
clf = fumm.KNN(k=k)
self._test_unfitted_internals(clf, init_k=k, init_is_classifier=True)
clf.fit(self.X_struct, self.y)
self._test_fitted_internals(clf, True, self.X_struct, self.y, self.X_n,
self.majority_label,
self.X_struct_categorical_indices,
self.X_struct_numerical_indices,
self.unique_y, self.unique_y_counts,
self.unique_y_probabilities)
y_hat = clf.predict_proba(self.X_test_struct)
assert np.isclose(y_hat, self.y_test_3_proba, atol=1e-3).all()
# Sample bigger than k
k = 10
clf = fumm.KNN(k=k)
self._test_unfitted_internals(clf, init_k=k, init_is_classifier=True)
clf.fit(self.X_struct, self.y)
self._test_fitted_internals(clf, True, self.X_struct, self.y, self.X_n,
self.majority_label,
self.X_struct_categorical_indices,
self.X_struct_numerical_indices,
self.unique_y, self.unique_y_counts,
self.unique_y_probabilities)
y_hat = clf.predict_proba(self.X_test_struct)
y_true = np.full(
(y_hat.shape[0], self.y_test_3_trainig_proba.shape[0]),
fill_value=self.y_test_3_trainig_proba)
assert np.isclose(y_hat, y_true, atol=1e-3).all()
# Categorical classifier on unstructured
# Sample smaller than k
k = 3
clf = fumm.KNN(k=k)
self._test_unfitted_internals(clf, init_k=k, init_is_classifier=True)
clf.fit(self.X, self.y_categorical)
self._test_fitted_internals(
clf, False, self.X, self.y_categorical, self.X_n,
self.majority_label_categorical, self.X_categorical_indices,
self.X_numerical_indices, self.unique_y_categorical,
self.unique_y_counts, self.unique_y_probabilities)
y_hat = clf.predict_proba(self.X_test)
assert np.isclose(y_hat, self.y_test_3_proba, atol=1e-3).all()
# Sample bigger than k
k = 10
clf = fumm.KNN(k=k)
self._test_unfitted_internals(clf, init_k=k, init_is_classifier=True)
clf.fit(self.X, self.y_categorical)
self._test_fitted_internals(
clf, False, self.X, self.y_categorical, self.X_n,
self.majority_label_categorical, self.X_categorical_indices,
self.X_numerical_indices, self.unique_y_categorical,
self.unique_y_counts, self.unique_y_probabilities)
y_hat = clf.predict_proba(self.X_test)
y_true = np.full(
(y_hat.shape[0], self.y_test_3_trainig_proba.shape[0]),
fill_value=self.y_test_3_trainig_proba)
assert np.isclose(y_hat, y_true, atol=1e-3).all()
# Categorical classifier on structured
# Sample smaller than k
k = 3
clf = fumm.KNN(k=k)
self._test_unfitted_internals(clf, init_k=k, init_is_classifier=True)
clf.fit(self.X_struct, self.y_categorical)
self._test_fitted_internals(
clf, True, self.X_struct, self.y_categorical, self.X_n,
self.majority_label_categorical, self.X_struct_categorical_indices,
self.X_struct_numerical_indices, self.unique_y_categorical,
self.unique_y_counts, self.unique_y_probabilities)
y_hat = clf.predict_proba(self.X_test_struct)
assert np.isclose(y_hat, self.y_test_3_proba, atol=1e-3).all()
# Sample bigger than k
k = 10
clf = fumm.KNN(k=k)
self._test_unfitted_internals(clf, init_k=k, init_is_classifier=True)
clf.fit(self.X_struct, self.y_categorical)
self._test_fitted_internals(
clf, True, self.X_struct, self.y_categorical, self.X_n,
self.majority_label_categorical, self.X_struct_categorical_indices,
self.X_struct_numerical_indices, self.unique_y_categorical,
self.unique_y_counts, self.unique_y_probabilities)
y_hat = clf.predict_proba(self.X_test_struct)
y_true = np.full(
(y_hat.shape[0], self.y_test_3_trainig_proba.shape[0]),
fill_value=self.y_test_3_trainig_proba)
assert np.isclose(y_hat, y_true, atol=1e-3).all()
def test_predict(self):
"""
Tests KNN predictions (:func:`~fatf.utils.models.models.KNN.predict`).
"""
# pylint: disable=too-many-statements
k = 2
clf = fumm.KNN(k=k)
self._test_unfitted_internals(clf, init_k=k, init_is_classifier=True)
# Unfitted model
with pytest.raises(UnfittedModelError) as exception_info:
clf.predict(self.X_test)
assert self.unfitted_model_error == str(exception_info.value)
# X is not 2D
clf.fit(self.X, self.y)
with pytest.raises(IncorrectShapeError) as exception_info:
clf.predict(self.X_3D)
assert self.incorrect_shape_error_singular == str(exception_info.value)
with pytest.raises(IncorrectShapeError) as exception_info:
clf.predict(self.y)
assert self.incorrect_shape_error_singular == str(exception_info.value)
# dtype is not similar to the training data
with pytest.raises(ValueError) as exception_info:
clf.predict(self.X_cat)
assert self.value_error_dtype == str(exception_info.value)
with pytest.raises(ValueError) as exception_info:
clf.predict(self.X_cat_struct)
assert self.value_error_dtype == str(exception_info.value)
with pytest.raises(ValueError) as exception_info:
clf.predict(self.X_struct)
assert self.value_error_dtype == str(exception_info.value)
# Predict 0 examples
with pytest.raises(IncorrectShapeError) as exception_info:
clf.predict(np.ones((0, 2), dtype=int))
assert self.incorrect_shape_error_rows == str(exception_info.value)
# The number of features disagrees...
# ...unstructured
with pytest.raises(IncorrectShapeError) as exception_info:
clf.predict(self.X_distances)
assert str(exception_info.value).startswith(
self.incorrect_shape_error_columns)
# ...structured
clf.clear()
self._test_unfitted_internals(clf, init_k=k, init_is_classifier=True)
clf.fit(self.X_struct, self.y)
self._test_fitted_internals(clf, True, self.X_struct, self.y, self.X_n,
self.majority_label,
self.X_struct_categorical_indices,
self.X_struct_numerical_indices,
self.unique_y, self.unique_y_counts,
self.unique_y_probabilities)
with pytest.raises(ValueError) as exception_info:
clf.predict(self.X_test_struct[['a']])
assert self.value_error_dtype == str(exception_info.value)
# Regressor on unstructured
# Sample smaller than k
k = 3
clf = fumm.KNN(k=k, mode='r')
self._test_unfitted_internals(clf, init_k=k, init_is_classifier=False)
clf.fit(self.X, self.y)
self._test_fitted_internals(clf, False, self.X, self.y, self.X_n,
self.majority_label_regressor,
self.X_categorical_indices,
self.X_numerical_indices)
y_hat = clf.predict(self.X_test)
assert np.isclose(y_hat, self.y_test_3_regression, atol=1e-3).all()
# Sample bigger than k
k = 10
clf = fumm.KNN(k=k, mode='r')
self._test_unfitted_internals(clf, init_k=k, init_is_classifier=False)
clf.fit(self.X, self.y)
self._test_fitted_internals(clf, False, self.X, self.y, self.X_n,
self.majority_label_regressor,
self.X_categorical_indices,
self.X_numerical_indices)
y_hat = clf.predict(self.X_test)
y_true = np.array(y_hat.shape[0] * [self.majority_label_regressor])
assert np.isclose(y_hat, y_true, atol=1e-3).all()
# Regressor on structured
# Sample smaller than k
k = 3
clf = fumm.KNN(k=k, mode='r')
self._test_unfitted_internals(clf, init_k=k, init_is_classifier=False)
clf.fit(self.X_struct, self.y)
self._test_fitted_internals(clf, True, self.X_struct, self.y, self.X_n,
self.majority_label_regressor,
self.X_struct_categorical_indices,
self.X_struct_numerical_indices)
y_hat = clf.predict(self.X_test_struct)
assert np.isclose(y_hat, self.y_test_3_regression, atol=1e-3).all()
# Sample bigger than k
k = 10
clf = fumm.KNN(k=k, mode='r')
self._test_unfitted_internals(clf, init_k=k, init_is_classifier=False)
clf.fit(self.X_struct, self.y)
self._test_fitted_internals(clf, True, self.X_struct, self.y, self.X_n,
self.majority_label_regressor,
self.X_struct_categorical_indices,
self.X_struct_numerical_indices)
y_hat = clf.predict(self.X_test_struct)
y_true = np.array(y_hat.shape[0] * [self.majority_label_regressor])
assert np.isclose(y_hat, y_true, atol=1e-3).all()
# Numerical classifier on unstructured
# Sample smaller than k
k = 3
clf = fumm.KNN(k=k)
self._test_unfitted_internals(clf, init_k=k, init_is_classifier=True)
clf.fit(self.X, self.y)
self._test_fitted_internals(clf, False, self.X, self.y, self.X_n,
self.majority_label,
self.X_categorical_indices,
self.X_numerical_indices, self.unique_y,
self.unique_y_counts,
self.unique_y_probabilities)
y_hat = clf.predict(self.X_test)
assert np.isclose(y_hat, self.y_test_3_classification, atol=1e-3).all()
# Sample bigger than k
k = 10
clf = fumm.KNN(k=k)
self._test_unfitted_internals(clf, init_k=k, init_is_classifier=True)
clf.fit(self.X, self.y)
self._test_fitted_internals(clf, False, self.X, self.y, self.X_n,
self.majority_label,
self.X_categorical_indices,
self.X_numerical_indices, self.unique_y,
self.unique_y_counts,
self.unique_y_probabilities)
y_hat = clf.predict(self.X_test)
y_true = np.array(y_hat.shape[0] * [self.majority_label])
assert np.isclose(y_hat, y_true, atol=1e-3).all()
# Numerical classifier on structured
# Sample smaller than k
k = 3
clf = fumm.KNN(k=k)
self._test_unfitted_internals(clf, init_k=k, init_is_classifier=True)
clf.fit(self.X_struct, self.y)
self._test_fitted_internals(clf, True, self.X_struct, self.y, self.X_n,
self.majority_label,
self.X_struct_categorical_indices,
self.X_struct_numerical_indices,
self.unique_y, self.unique_y_counts,
self.unique_y_probabilities)
y_hat = clf.predict(self.X_test_struct)
assert np.isclose(y_hat, self.y_test_3_classification, atol=1e-3).all()
# Sample bigger than k
k = 10
clf = fumm.KNN(k=k)
self._test_unfitted_internals(clf, init_k=k, init_is_classifier=True)
clf.fit(self.X_struct, self.y)
self._test_fitted_internals(clf, True, self.X_struct, self.y, self.X_n,
self.majority_label,
self.X_struct_categorical_indices,
self.X_struct_numerical_indices,
self.unique_y, self.unique_y_counts,
self.unique_y_probabilities)
y_hat = clf.predict(self.X_test_struct)
y_true = np.array(y_hat.shape[0] * [self.majority_label])
assert np.isclose(y_hat, y_true, atol=1e-3).all()
# Categorical classifier on unstructured
# Sample smaller than k
k = 3
clf = fumm.KNN(k=k)
self._test_unfitted_internals(clf, init_k=k, init_is_classifier=True)
clf.fit(self.X, self.y_categorical)
self._test_fitted_internals(
clf, False, self.X, self.y_categorical, self.X_n,
self.majority_label_categorical, self.X_categorical_indices,
self.X_numerical_indices, self.unique_y_categorical,
self.unique_y_counts, self.unique_y_probabilities)
y_hat = clf.predict(self.X_test)
assert np.array_equal(y_hat, self.y_test_3_classification_categorical)
# Sample bigger than k
k = 10
clf = fumm.KNN(k=k)
self._test_unfitted_internals(clf, init_k=k, init_is_classifier=True)
clf.fit(self.X, self.y_categorical)
self._test_fitted_internals(
clf, False, self.X, self.y_categorical, self.X_n,
self.majority_label_categorical, self.X_categorical_indices,
self.X_numerical_indices, self.unique_y_categorical,
self.unique_y_counts, self.unique_y_probabilities)
y_hat = clf.predict(self.X_test)
y_true = np.array(y_hat.shape[0] * [self.majority_label_categorical])
assert np.array_equal(y_hat, y_true)
# Categorical classifier on structured
# Sample smaller than k
k = 3
clf = fumm.KNN(k=k)
self._test_unfitted_internals(clf, init_k=k, init_is_classifier=True)
clf.fit(self.X_struct, self.y_categorical)
self._test_fitted_internals(
clf, True, self.X_struct, self.y_categorical, self.X_n,
self.majority_label_categorical, self.X_struct_categorical_indices,
self.X_struct_numerical_indices, self.unique_y_categorical,
self.unique_y_counts, self.unique_y_probabilities)
y_hat = clf.predict(self.X_test_struct)
assert np.array_equal(y_hat, self.y_test_3_classification_categorical)
# Sample bigger than k
k = 10
clf = fumm.KNN(k=k)
self._test_unfitted_internals(clf, init_k=k, init_is_classifier=True)
clf.fit(self.X_struct, self.y_categorical)
self._test_fitted_internals(
clf, True, self.X_struct, self.y_categorical, self.X_n,
self.majority_label_categorical, self.X_struct_categorical_indices,
self.X_struct_numerical_indices, self.unique_y_categorical,
self.unique_y_counts, self.unique_y_probabilities)
y_hat = clf.predict(self.X_test_struct)
y_true = np.array(y_hat.shape[0] * [self.majority_label_categorical])
assert np.array_equal(y_hat, y_true)
# Test when the majority class is ambiguous -- sample smaller than k
y = np.array([0, 1, 0, 1, 0, 1]) # pylint: disable=invalid-name
majority_label = 0
unique_y = np.array([0, 1])
unique_y_counts = np.array([3, 3])
unique_y_probabilities = np.array([.5, .5])
X_test = np.array([[0, 0], [2, 0]]) # pylint: disable=invalid-name
y_test = np.array([0, 0])
#
k = 4
clf = fumm.KNN(k=k)
self._test_unfitted_internals(clf, init_k=k, init_is_classifier=True)
clf.fit(self.X, y)
self._test_fitted_internals(clf, False, self.X, y, self.X_n,
majority_label, self.X_categorical_indices,
self.X_numerical_indices, unique_y,
unique_y_counts, unique_y_probabilities)
y_hat = clf.predict(X_test)
assert np.array_equal(y_hat, y_test)
def test_get_distances(self):
"""
Tests distances (:func:`~fatf.utils.models.models.KNN._get_distances`).
"""
k = 2
clf = fumm.KNN(k=k)
def is_unfitted():
return self._test_unfitted_internals(clf,
init_k=2,
init_is_classifier=True)
is_unfitted()
# Numerical distances on unstructured
clf.fit(self.X, self.y)
self._test_fitted_internals(clf, False, self.X, self.y, self.X_n,
self.majority_label,
self.X_categorical_indices,
self.X_numerical_indices, self.unique_y,
self.unique_y_counts,
self.unique_y_probabilities)
dist = clf._get_distances(self.X_test)
assert np.isclose(dist, self.X_distances, atol=1e-3).all()
# Categorical distances on unstructured
clf.clear()
is_unfitted()
clf.fit(self.X_cat, self.y)
self._test_fitted_internals(clf, False, self.X_cat, self.y, self.X_n,
self.majority_label,
self.X_cat_categorical_indices,
self.X_cat_numerical_indices,
self.unique_y, self.unique_y_counts,
self.unique_y_probabilities)
dist = clf._get_distances(self.X_cat_test)
assert np.isclose(dist, self.X_cat_distances, atol=1e-3).all()
# Numerical distances on structured
clf.clear()
is_unfitted()
clf.fit(self.X_struct, self.y)
self._test_fitted_internals(clf, True, self.X_struct, self.y, self.X_n,
self.majority_label,
self.X_struct_categorical_indices,
self.X_struct_numerical_indices,
self.unique_y, self.unique_y_counts,
self.unique_y_probabilities)
dist = clf._get_distances(self.X_test_struct)
assert np.isclose(dist, self.X_distances, atol=1e-3).all()
# Categorical distances on structured
clf.clear()
is_unfitted()
clf.fit(self.X_cat_struct, self.y)
self._test_fitted_internals(clf, True, self.X_cat_struct, self.y,
self.X_n, self.majority_label,
self.X_cat_struct_categorical_indices,
self.X_cat_struct_numerical_indices,
self.unique_y, self.unique_y_counts,
self.unique_y_probabilities)
dist = clf._get_distances(self.X_cat_struct_test)
assert np.isclose(dist, self.X_cat_distances, atol=1e-3).all()
# Numerical-categorical distances on structured
clf.clear()
is_unfitted()
clf.fit(self.X_mix, self.y)
self._test_fitted_internals(clf, True, self.X_mix, self.y, self.X_n,
self.majority_label,
self.X_mix_categorical_indices,
self.X_mix_numerical_indices,
self.unique_y, self.unique_y_counts,
self.unique_y_probabilities)
dist = clf._get_distances(self.X_test_mix)
assert np.isclose(dist, self.X_mix_distances, atol=1e-3).all()
def test_fit(self):
"""
Tests KNN fitting (:func:`~fatf.utils.models.models.KNN.fit`).
"""
# pylint: disable=too-many-statements
k = 2
clf = fumm.KNN(k=k)
self._test_unfitted_internals(clf, init_k=k)
clf.fit(self.X, self.y)
self._test_fitted_internals(clf, False, self.X, self.y, self.X_n,
self.majority_label,
self.X_categorical_indices,
self.X_numerical_indices, self.unique_y,
self.unique_y_counts,
self.unique_y_probabilities)
# Fitting a pre-fitted model
with pytest.raises(PrefittedModelError) as exception_info:
clf.fit(self.X, self.y)
assert self.prefitted_model_error == str(exception_info.value)
clf = fumm.KNN(k=k)
self._test_unfitted_internals(clf, init_k=k)
# X is not 2D
with pytest.raises(IncorrectShapeError) as exception_info:
clf.fit(self.X_3D, self.y)
assert self.incorrect_shape_error_2d == str(exception_info.value)
with pytest.raises(IncorrectShapeError) as exception_info:
clf.fit(self.X_3D, self.X)
assert self.incorrect_shape_error_2d == str(exception_info.value)
# y is not 1D
with pytest.raises(IncorrectShapeError) as exception_info:
clf.fit(self.X, self.X_3D)
assert self.incorrect_shape_error_1d == str(exception_info.value)
# 0 examples
with pytest.raises(IncorrectShapeError) as exception_info:
clf.fit(np.ndarray((0, 5)), self.y)
assert self.incorrect_shape_error_X0 == str(exception_info.value)
with pytest.raises(IncorrectShapeError) as exception_info:
clf.fit(np.ndarray((0, ), dtype=[('a', str), ('b', int)]), self.y)
assert self.incorrect_shape_error_X0 == str(exception_info.value)
# 0 features
with pytest.raises(IncorrectShapeError) as exception_info:
clf.fit(np.ndarray((5, 0)), self.y)
assert self.incorrect_shape_error_X1 == str(exception_info.value)
# Test whether the shape of X agrees with the shape of y
with pytest.raises(IncorrectShapeError) as exception_info:
clf.fit(self.X, self.X_numerical_indices)
assert self.incorrect_shape_error_Xy == str(exception_info.value)
# Fitting regressor to a categorical label vector
clf = fumm.KNN(k=k, mode='r')
self._test_unfitted_internals(clf, init_k=k, init_is_classifier=False)
with pytest.raises(TypeError) as exception_info:
y_pred = np.array(self.X.shape[0] * ['a'])
clf.fit(self.X, y_pred)
assert self.type_error_regressor == str(exception_info.value)
# Fitting to a structured numerical array
clf = fumm.KNN(k=k)
self._test_unfitted_internals(clf, init_k=k, init_is_classifier=True)
clf.fit(self.X_struct, self.y)
self._test_fitted_internals(clf, True, self.X_struct, self.y, self.X_n,
self.majority_label,
self.X_struct_categorical_indices,
self.X_struct_numerical_indices,
self.unique_y, self.unique_y_counts,
self.unique_y_probabilities)
# Fitting to a structured mixed numerical-categorical array
clf = fumm.KNN(k=k)
self._test_unfitted_internals(clf, init_k=k, init_is_classifier=True)
clf.fit(self.X_mix, self.y)
self._test_fitted_internals(clf, True, self.X_mix, self.y, self.X_n,
self.majority_label,
self.X_mix_categorical_indices,
self.X_mix_numerical_indices,
self.unique_y, self.unique_y_counts,
self.unique_y_probabilities)
# Fit a regressor to a numerical data and check internal parameters
clf = fumm.KNN(k=k, mode='regressor')
self._test_unfitted_internals(clf, init_k=k, init_is_classifier=False)
clf.fit(self.X_short, self.y_short_numerical)
self._test_fitted_internals(
clf, False, self.X_short, self.y_short_numerical, self.X_short_n,
self.short_numerical_majority_label_regressor,
self.X_short_categorical_indices, self.X_short_numerical_indices)
# Fit a classifier to a numerical data and check internal parameters
clf = fumm.KNN(k=k, mode='classifier')
self._test_unfitted_internals(clf, init_k=k, init_is_classifier=True)
clf.fit(self.X_short, self.y_short_numerical)
self._test_fitted_internals(
clf, False, self.X_short, self.y_short_numerical, self.X_short_n,
self.short_numerical_majority_label_classifier,
self.X_short_categorical_indices, self.X_short_numerical_indices,
self.short_numerical_unique_y,
self.short_numerical_unique_y_counts,
self.short_numerical_unique_y_probabilities)
# Fit a classifier to a categorical data and check internal parameters
clf = fumm.KNN(k=k, mode='classifier')
self._test_unfitted_internals(clf, init_k=k, init_is_classifier=True)
clf.fit(self.X_short, self.y_short)
self._test_fitted_internals(
clf, False, self.X_short, self.y_short, self.X_short_n,
self.short_majority_label, self.X_short_categorical_indices,
self.X_short_numerical_indices, self.short_unique_y,
self.short_unique_y_counts, self.short_unique_y_probabilities)
def test_knn(self):
"""
Tests KNN initialisation.
"""
# k is not an integer
with pytest.raises(TypeError) as exception_info:
clf = fumm.KNN(k=None)
assert str(exception_info.value) == self.type_error_k
with pytest.raises(TypeError) as exception_info:
clf = fumm.KNN(k='k')
assert str(exception_info.value) == self.type_error_k
with pytest.raises(TypeError) as exception_info:
clf = fumm.KNN(k=-5.5)
assert str(exception_info.value) == self.type_error_k
# k is a negative integer
with pytest.raises(ValueError) as exception_info:
clf = fumm.KNN(k=-5)
assert str(exception_info.value) == self.value_error_k
# mode specifier is wrong
with pytest.raises(ValueError) as exception_info:
clf = fumm.KNN(k=5, mode=object())
assert str(exception_info.value) == self.value_error_mode
with pytest.raises(ValueError) as exception_info:
clf = fumm.KNN(k=5, mode=3)
assert str(exception_info.value) == self.value_error_mode
with pytest.raises(ValueError) as exception_info:
clf = fumm.KNN(k=5, mode='C')
assert str(exception_info.value) == self.value_error_mode
clf = fumm.KNN()
self._test_unfitted_internals(clf,
init_k=self.k,
init_is_classifier=True)
k = 8
clf = fumm.KNN(k=k)
self._test_unfitted_internals(clf, init_k=k, init_is_classifier=True)
clf = fumm.KNN(k=k, mode=None)
self._test_unfitted_internals(clf, init_k=k, init_is_classifier=True)
clf = fumm.KNN(k=k, mode='c')
self._test_unfitted_internals(clf, init_k=k, init_is_classifier=True)
clf = fumm.KNN(k=k, mode='classifier')
self._test_unfitted_internals(clf, init_k=k, init_is_classifier=True)
clf = fumm.KNN(k=k, mode='r')
self._test_unfitted_internals(clf, init_k=k, init_is_classifier=False)
clf = fumm.KNN(k=k, mode='regressor')
self._test_unfitted_internals(clf, init_k=k, init_is_classifier=False)
def test_validate_input_local_fidelity():
"""
Tests the ``_validate_input_local_fidelity`` function.
This function tests the :func:`fatf.utils.transparency.\
surrogate_evaluation._validate_input_local_fidelity` function.
"""
incorrect_shape_dataset = ('The input dataset must be a 2-dimensional '
'numpy array.')
type_error_dataset = ('The input dataset must be of a base type -- '
'numbers and/or strings.')
incorrect_shape_datarow = ('The data_row must either be a 1-dimensional '
'numpy array or a numpy void object for '
'structured data rows.')
incorrect_dtype_data = ('The dtype of the data_row is too different from '
'the dtype of the dataset array.')
datarow_features_error = ('The data_row must contain the same number of '
'features as the dataset.')
global_model_incompatible = ('The global predictive function must have '
'exactly *one* required parameter to work '
'with this metric.')
global_model_type = ('The global_predictive_function should be a Python '
'callable, e.g., a Python function.')
local_model_incompatible = ('The local predictive function must have '
'exactly *one* required parameter to work '
'with this metric.')
local_model_type = ('The local_predictive_function should be a Python '
'callable, e.g., a Python function.')
metric_param_error = ('The metric_function must take exactly *two* '
'required parameters.')
metric_type_error = ('The metric_function should be a Python callable, '
'e.g., a Python function.')
explained_class_value_error = ('The explained_class_index parameter is '
'negative or larger than the number of '
'classes output by the global '
'probabilistic model.')
explained_class_type_error = ('For probabilistic global models, i.e., '
'global predictive functions, the '
'explained_class_index parameter has to be '
'an integer or None.')
explained_class_warning = ('The explained_class_index parameter is not '
'None and will be ignored since the global '
'model is not probabilistic.')
features_index_error = ('The following column indices are invalid for '
'the input dataset: {}.')
features_type_error = ('The explained_feature_indices parameter must be '
'a Python list or None.')
fidelity_radius_type_error = ('The fidelity_radius_percentage must be an '
'integer between 1 and 100.')
fidelity_radius_value_error = ('The fidelity_radius_percentage must be an '
'integer between 1 and 100.')
samples_number_value_error = ('The samples_number must be a positive '
'integer.')
samples_number_type_error = 'The samples_number must be an integer.'
with pytest.raises(IncorrectShapeError) as exin:
futs._validate_input_local_fidelity(NUMERICAL_NP_ARRAY[0], None, None,
None, None, None, None, None, None)
assert str(exin.value) == incorrect_shape_dataset
with pytest.raises(TypeError) as exin:
futs._validate_input_local_fidelity(np.array([[None]]), None, None,
None, None, None, None, None, None)
assert str(exin.value) == type_error_dataset
with pytest.raises(IncorrectShapeError) as exin:
futs._validate_input_local_fidelity(NUMERICAL_NP_ARRAY,
NUMERICAL_NP_ARRAY, None, None,
None, None, None, None, None)
assert str(exin.value) == incorrect_shape_datarow
with pytest.raises(TypeError) as exin:
futs._validate_input_local_fidelity(
NUMERICAL_NP_ARRAY, np.array(['0']), None, None, None, None, None,
None, None) # yapf: disable
assert str(exin.value) == incorrect_dtype_data
with pytest.raises(IncorrectShapeError) as exin:
futs._validate_input_local_fidelity(NUMERICAL_NP_ARRAY,
NUMERICAL_NP_ARRAY[0][0:2], None,
None, None, None, None, None, None)
assert str(exin.value) == datarow_features_error
def predict(x):
return np.ones(x.shape[0])
def predict_invalid(x_1, x_2):
pass # pragma: no cover
def predict_proba(x):
return np.ones((x.shape[0], 3))
def predict_proba_invalid():
pass # pragma: no cover
with pytest.raises(TypeError) as exin:
futs._validate_input_local_fidelity(NUMERICAL_NP_ARRAY,
NUMERICAL_NP_ARRAY[0], None, None,
None, None, None, None, None)
assert str(exin.value) == global_model_type
with pytest.raises(IncompatibleModelError) as exin:
futs._validate_input_local_fidelity(NUMERICAL_NP_ARRAY,
NUMERICAL_NP_ARRAY[0],
predict_invalid, None, None, None,
None, None, None)
assert str(exin.value) == global_model_incompatible
with pytest.raises(TypeError) as exin:
futs._validate_input_local_fidelity(NUMERICAL_NP_ARRAY,
NUMERICAL_NP_ARRAY[0], predict,
None, None, None, None, None, None)
assert str(exin.value) == local_model_type
with pytest.raises(IncompatibleModelError) as exin:
futs._validate_input_local_fidelity(NUMERICAL_NP_ARRAY,
NUMERICAL_NP_ARRAY[0],
predict_proba,
predict_proba_invalid, None, None,
None, None, None)
assert str(exin.value) == local_model_incompatible
def invalid_metric(x):
pass # pragma: no cover
def metric(x_1, x_2):
pass # pragma: no cover
with pytest.raises(TypeError) as exin:
futs._validate_input_local_fidelity(NUMERICAL_NP_ARRAY,
NUMERICAL_NP_ARRAY[0], predict,
predict, None, None, None, None,
None)
assert str(exin.value) == metric_type_error
with pytest.raises(TypeError) as exin:
futs._validate_input_local_fidelity(NUMERICAL_NP_ARRAY,
NUMERICAL_NP_ARRAY[0], predict,
predict, invalid_metric, None,
None, None, None)
assert str(exin.value) == metric_param_error
with pytest.raises(TypeError) as exin:
futs._validate_input_local_fidelity(NUMERICAL_NP_ARRAY,
NUMERICAL_NP_ARRAY[0],
predict_proba, predict, metric,
'1', None, None, None)
assert str(exin.value) == explained_class_type_error
with pytest.raises(ValueError) as exin:
futs._validate_input_local_fidelity(NUMERICAL_NP_ARRAY,
NUMERICAL_NP_ARRAY[0],
predict_proba, predict, metric, -1,
None, None, None)
assert str(exin.value) == explained_class_value_error
with pytest.raises(ValueError) as exin:
futs._validate_input_local_fidelity(NUMERICAL_NP_ARRAY,
NUMERICAL_NP_ARRAY[0],
predict_proba, predict, metric, 3,
None, None, None)
assert str(exin.value) == explained_class_value_error
#
with pytest.warns(UserWarning) as w:
futs._validate_input_local_fidelity(NUMERICAL_NP_ARRAY,
NUMERICAL_NP_ARRAY[0], predict,
predict, metric, 3, None, 1, 1)
assert len(w) == 1
assert str(w[0].message) == explained_class_warning
with pytest.raises(TypeError) as exin:
futs._validate_input_local_fidelity(NUMERICAL_NP_ARRAY,
NUMERICAL_NP_ARRAY[0],
predict_proba,
predict, metric, None,
np.array([10, 11]), None, None)
assert str(exin.value) == features_type_error
with pytest.raises(IndexError) as exin:
futs._validate_input_local_fidelity(NUMERICAL_NP_ARRAY,
NUMERICAL_NP_ARRAY[0], predict,
predict, metric, None, [10, 11],
None, None)
assert str(exin.value) == features_index_error.format(np.array([10, 11]))
with pytest.raises(TypeError) as exin:
futs._validate_input_local_fidelity(NUMERICAL_NP_ARRAY,
NUMERICAL_NP_ARRAY[0],
predict_proba, predict, metric, 1,
[1, 2], 'a', None)
assert str(exin.value) == fidelity_radius_type_error
with pytest.raises(TypeError) as exin:
futs._validate_input_local_fidelity(NUMERICAL_NP_ARRAY,
NUMERICAL_NP_ARRAY[0],
predict_proba, predict, metric, 1,
[1, 2], None, None)
assert str(exin.value) == fidelity_radius_type_error
with pytest.raises(TypeError) as exin:
futs._validate_input_local_fidelity(NUMERICAL_NP_ARRAY,
NUMERICAL_NP_ARRAY[0],
predict_proba, predict, metric, 1,
[1, 2], 55.0, None)
assert str(exin.value) == fidelity_radius_type_error
#
with pytest.raises(ValueError) as exin:
futs._validate_input_local_fidelity(NUMERICAL_NP_ARRAY,
NUMERICAL_NP_ARRAY[0], predict,
predict, metric, None, [1, 2], 0,
None)
assert str(exin.value) == fidelity_radius_value_error
with pytest.raises(ValueError) as exin:
futs._validate_input_local_fidelity(NUMERICAL_NP_ARRAY,
NUMERICAL_NP_ARRAY[0], predict,
predict, metric, None, [1, 2], 101,
None)
assert str(exin.value) == fidelity_radius_value_error
with pytest.raises(TypeError) as exin:
futs._validate_input_local_fidelity(NUMERICAL_NP_ARRAY,
NUMERICAL_NP_ARRAY[0], predict,
predict, metric, None, None, 100,
None)
assert str(exin.value) == samples_number_type_error
with pytest.raises(TypeError) as exin:
futs._validate_input_local_fidelity(NUMERICAL_NP_ARRAY,
NUMERICAL_NP_ARRAY[0], predict,
predict, metric, None, None, 100,
55.0)
assert str(exin.value) == samples_number_type_error
#
with pytest.raises(ValueError) as exin:
futs._validate_input_local_fidelity(NUMERICAL_NP_ARRAY,
NUMERICAL_NP_ARRAY[0], predict,
predict, metric, None, None, 100,
0)
assert str(exin.value) == samples_number_value_error
with pytest.raises(ValueError) as exin:
futs._validate_input_local_fidelity(NUMERICAL_NP_ARRAY,
NUMERICAL_NP_ARRAY[0], predict,
predict, metric, None, None, 100,
-42)
assert str(exin.value) == samples_number_value_error
clf = fumm.KNN(k=3)
clf.fit(NUMERICAL_NP_ARRAY, NUMERICAL_NP_ARRAY_TARGET)
with pytest.raises(ValueError) as exin:
futs._validate_input_local_fidelity(NUMERICAL_NP_ARRAY,
NUMERICAL_NP_ARRAY[0],
clf.predict_proba, predict, metric,
10, None, 10, 1)
assert str(exin.value) == explained_class_value_error
# All OK
assert futs._validate_input_local_fidelity(NUMERICAL_NP_ARRAY,
NUMERICAL_NP_ARRAY[0],
clf.predict_proba, predict,
metric, 1, [0, 1], 10, 1)
def test_local_fidelity_score():
"""
Tests the ``local_fidelity_score`` function.
This function tests the
:func:`fatf.utils.transparency.surrogate_evaluation.local_fidelity_score`
function.
"""
accuracy_warning = ('Some of the given labels are not present in either '
'of the input arrays: {}.')
fatf.setup_random_seed()
def accuracy(global_predictions, local_predictions):
global_predictions[global_predictions >= 0.5] = 1
global_predictions[global_predictions < 0.5] = 0
local_predictions[local_predictions >= 0.5] = 1
local_predictions[local_predictions < 0.5] = 0
confusion_matrix = fumt.get_confusion_matrix(global_predictions,
local_predictions,
labels=[0, 1])
accuracy = fummet.accuracy(confusion_matrix)
return accuracy
def accuracy_prob(global_predictions,
local_predictions,
global_proba=True,
local_proba=True):
if global_proba:
global_predictions = np.argmax(global_predictions, axis=1)
if local_proba:
local_predictions = np.argmax(local_predictions, axis=1)
confusion_matrix = fumt.get_confusion_matrix(global_predictions,
local_predictions,
labels=[0, 1, 2])
accuracy = fummet.accuracy(confusion_matrix)
return accuracy
def accuracy_proba_np(global_predictions, local_predictions):
return accuracy_prob(global_predictions,
local_predictions,
global_proba=False,
local_proba=True)
def accuracy_proba_nn(global_predictions, local_predictions):
return accuracy_prob(global_predictions,
local_predictions,
global_proba=False,
local_proba=False)
def reg_dist(global_predictions, local_predictions):
return (global_predictions - local_predictions).sum()
predictor = fumm.KNN(k=3)
predictor.fit(NUMERICAL_NP_ARRAY, NUMERICAL_NP_ARRAY_TARGET)
regressor = fumm.KNN(k=3, mode='regressor')
regressor.fit(NUMERICAL_NP_ARRAY_LOCAL, NUMERICAL_NP_ARRAY_LOCAL_TARGET)
regressor_23 = fumm.KNN(k=3, mode='regressor')
regressor_23.fit(NUMERICAL_NP_ARRAY_LOCAL[:, [2, 3]],
NUMERICAL_NP_ARRAY_LOCAL_TARGET)
# Structured array
predictor_struct = fumm.KNN(k=3)
predictor_struct.fit(NUMERICAL_STRUCT_ARRAY, NUMERICAL_NP_ARRAY_TARGET)
#
regressor_struct_cd = fumm.KNN(k=3, mode='regressor')
regressor_struct_cd.fit(NUMERICAL_STRUCT_ARRAY_LOCAL[['c', 'd']],
NUMERICAL_NP_ARRAY_LOCAL_TARGET)
# Global: probabilistic...
# ...local: regressor
comparison = futs.local_fidelity_score(NUMERICAL_NP_ARRAY,
NUMERICAL_NP_ARRAY[0],
predictor.predict_proba,
regressor.predict, accuracy, 2)
assert np.isclose(comparison, 0.26)
# ...local: classifier
comparison = futs.local_fidelity_score(NUMERICAL_NP_ARRAY,
NUMERICAL_NP_ARRAY[0],
predictor.predict_proba,
predictor.predict, accuracy, 2)
assert np.isclose(comparison, 1.0)
# ...local: probabilistic
with pytest.warns(UserWarning) as w:
comparison = futs.local_fidelity_score(NUMERICAL_NP_ARRAY,
NUMERICAL_NP_ARRAY[0],
predictor.predict_proba,
predictor.predict_proba,
accuracy_prob)
assert len(w) == 1
assert str(w[0].message) == accuracy_warning.format(set([1]))
assert np.isclose(comparison, 1.0)
# Global: classifier...
# ...local: probabilistic
with pytest.warns(UserWarning) as w:
comparison = futs.local_fidelity_score(NUMERICAL_NP_ARRAY,
NUMERICAL_NP_ARRAY[0],
predictor.predict,
predictor.predict_proba,
accuracy_proba_np)
assert len(w) == 1
assert str(w[0].message) == accuracy_warning.format(set([1]))
assert np.isclose(comparison, 1.0)
# ...local: classifier
with pytest.warns(UserWarning) as w:
comparison = futs.local_fidelity_score(NUMERICAL_NP_ARRAY,
NUMERICAL_NP_ARRAY[0],
predictor.predict,
predictor.predict,
accuracy_proba_nn)
assert len(w) == 1
assert str(w[0].message) == accuracy_warning.format(set([1]))
assert np.isclose(comparison, 1.0)
# Global: regressor...
# ...local: regressor
comparison = futs.local_fidelity_score(NUMERICAL_NP_ARRAY,
NUMERICAL_NP_ARRAY[0],
regressor.predict,
regressor_23.predict,
reg_dist,
explained_feature_indices=[2, 3])
assert np.isclose(comparison, 0)
# Structured array
# Global: probabilistic...
# ...local: regressor
comparison = futs.local_fidelity_score(
NUMERICAL_STRUCT_ARRAY,
NUMERICAL_STRUCT_ARRAY[0],
predictor_struct.predict_proba,
regressor_struct_cd.predict,
accuracy,
0,
explained_feature_indices=['c', 'd'])
assert np.isclose(comparison, 0.94)