def test_select_kbest_classif():
# Test whether the relative univariate feature selection
# gets the correct items in a simple classification problem
# with the k best heuristic
X, y = make_classification(n_samples=200,
n_features=20,
n_informative=3,
n_redundant=2,
n_repeated=0,
n_classes=8,
n_clusters_per_class=1,
flip_y=0.0,
class_sep=10,
shuffle=False,
random_state=0)
univariate_filter = SelectKBest(f_classif, k=5)
X_r = univariate_filter.fit(X, y).transform(X)
X_r2 = GenericUnivariateSelect(f_classif, mode='k_best',
param=5).fit(X, y).transform(X)
assert_array_equal(X_r, X_r2)
support = univariate_filter.get_support()
gtruth = np.zeros(20)
gtruth[:5] = 1
assert_array_equal(support, gtruth)
def test_boundary_case_ch2():
# Test boundary case, and always aim to select 1 feature.
X = np.array([[10, 20], [20, 20], [20, 30]])
y = np.array([[1], [0], [0]])
scores, pvalues = chi2(X, y)
assert_array_almost_equal(scores, np.array([4., 0.71428571]))
assert_array_almost_equal(pvalues, np.array([0.04550026, 0.39802472]))
filter_fdr = SelectFdr(chi2, alpha=0.1)
filter_fdr.fit(X, y)
support_fdr = filter_fdr.get_support()
assert_array_equal(support_fdr, np.array([True, False]))
filter_kbest = SelectKBest(chi2, k=1)
filter_kbest.fit(X, y)
support_kbest = filter_kbest.get_support()
assert_array_equal(support_kbest, np.array([True, False]))
filter_percentile = SelectPercentile(chi2, percentile=50)
filter_percentile.fit(X, y)
support_percentile = filter_percentile.get_support()
assert_array_equal(support_percentile, np.array([True, False]))
filter_fpr = SelectFpr(chi2, alpha=0.1)
filter_fpr.fit(X, y)
support_fpr = filter_fpr.get_support()
assert_array_equal(support_fpr, np.array([True, False]))
filter_fwe = SelectFwe(chi2, alpha=0.1)
filter_fwe.fit(X, y)
support_fwe = filter_fwe.get_support()
assert_array_equal(support_fwe, np.array([True, False]))
def test_feature_union_weights():
# test feature union with transformer weights
iris = load_iris()
X = iris.data
y = iris.target
pca = PCA(n_components=2, svd_solver='randomized', random_state=0)
select = SelectKBest(k=1)
# test using fit followed by transform
fs = FeatureUnion([("pca", pca), ("select", select)],
transformer_weights={"pca": 10})
fs.fit(X, y)
X_transformed = fs.transform(X)
# test using fit_transform
fs = FeatureUnion([("pca", pca), ("select", select)],
transformer_weights={"pca": 10})
X_fit_transformed = fs.fit_transform(X, y)
# test it works with transformers missing fit_transform
fs = FeatureUnion([("mock", Transf()), ("pca", pca), ("select", select)],
transformer_weights={"mock": 10})
X_fit_transformed_wo_method = fs.fit_transform(X, y)
# check against expected result
# We use a different pca object to control the random_state stream
assert_array_almost_equal(X_transformed[:, :-1], 10 * pca.fit_transform(X))
assert_array_equal(X_transformed[:, -1],
select.fit_transform(X, y).ravel())
assert_array_almost_equal(X_fit_transformed[:, :-1],
10 * pca.fit_transform(X))
assert_array_equal(X_fit_transformed[:, -1],
select.fit_transform(X, y).ravel())
assert X_fit_transformed_wo_method.shape == (X.shape[0], 7)
def test_tied_scores():
# Test for stable sorting in k-best with tied scores.
X_train = np.array([[0, 0, 0], [1, 1, 1]])
y_train = [0, 1]
for n_features in [1, 2, 3]:
sel = SelectKBest(chi2, k=n_features).fit(X_train, y_train)
X_test = sel.transform([[0, 1, 2]])
assert_array_equal(X_test[0], np.arange(3)[-n_features:])
def test_select_kbest_all():
# Test whether k="all" correctly returns all features.
X, y = make_classification(n_samples=20,
n_features=10,
shuffle=False,
random_state=0)
univariate_filter = SelectKBest(f_classif, k='all')
X_r = univariate_filter.fit(X, y).transform(X)
assert_array_equal(X, X_r)
def test_invalid_k():
X = [[0, 1, 0], [0, -1, -1], [0, .5, .5]]
y = [1, 0, 1]
with pytest.raises(ValueError):
SelectKBest(k=-1).fit(X, y)
with pytest.raises(ValueError):
SelectKBest(k=4).fit(X, y)
with pytest.raises(ValueError):
GenericUnivariateSelect(mode='k_best', param=-1).fit(X, y)
with pytest.raises(ValueError):
GenericUnivariateSelect(mode='k_best', param=4).fit(X, y)
def test_classes_property():
iris = load_iris()
X = iris.data
y = iris.target
reg = make_pipeline(SelectKBest(k=1), LinearRegression())
reg.fit(X, y)
assert_raises(AttributeError, getattr, reg, "classes_")
clf = make_pipeline(SelectKBest(k=1), LogisticRegression(random_state=0))
assert_raises(AttributeError, getattr, clf, "classes_")
clf.fit(X, y)
assert_array_equal(clf.classes_, np.unique(y))
def test_select_kbest_zero():
# Test whether k=0 correctly returns no features.
X, y = make_classification(n_samples=20,
n_features=10,
shuffle=False,
random_state=0)
univariate_filter = SelectKBest(f_classif, k=0)
univariate_filter.fit(X, y)
support = univariate_filter.get_support()
gtruth = np.zeros(10, dtype=bool)
assert_array_equal(support, gtruth)
X_selected = assert_warns_message(UserWarning, 'No features were selected',
univariate_filter.transform, X)
assert X_selected.shape == (20, 0)
def test_feature_union():
# basic sanity check for feature union
iris = load_iris()
X = iris.data
X -= X.mean(axis=0)
y = iris.target
svd = TruncatedSVD(n_components=2, random_state=0)
select = SelectKBest(k=1)
fs = FeatureUnion([("svd", svd), ("select", select)])
fs.fit(X, y)
X_transformed = fs.transform(X)
assert X_transformed.shape == (X.shape[0], 3)
# check if it does the expected thing
assert_array_almost_equal(X_transformed[:, :-1], svd.fit_transform(X))
assert_array_equal(X_transformed[:, -1],
select.fit_transform(X, y).ravel())
# test if it also works for sparse input
# We use a different svd object to control the random_state stream
fs = FeatureUnion([("svd", svd), ("select", select)])
X_sp = sparse.csr_matrix(X)
X_sp_transformed = fs.fit_transform(X_sp, y)
assert_array_almost_equal(X_transformed, X_sp_transformed.toarray())
# Test clone
fs2 = assert_no_warnings(clone, fs)
assert fs.transformer_list[0][1] is not fs2.transformer_list[0][1]
# test setting parameters
fs.set_params(select__k=2)
assert fs.fit_transform(X, y).shape == (X.shape[0], 4)
# test it works with transformers missing fit_transform
fs = FeatureUnion([("mock", Transf()), ("svd", svd), ("select", select)])
X_transformed = fs.fit_transform(X, y)
assert X_transformed.shape == (X.shape[0], 8)
# test error if some elements do not support transform
assert_raises_regex(TypeError,
'All estimators should implement fit and '
'transform.*\\bNoTrans\\b',
FeatureUnion,
[("transform", Transf()), ("no_transform", NoTrans())])
# test that init accepts tuples
fs = FeatureUnion((("svd", svd), ("select", select)))
fs.fit(X, y)
def test_selectkbest_tiebreaking():
# Test whether SelectKBest actually selects k features in case of ties.
# Prior to 0.11, SelectKBest would return more features than requested.
Xs = [[0, 1, 1], [0, 0, 1], [1, 0, 0], [1, 1, 0]]
y = [1]
dummy_score = lambda X, y: (X[0], X[0])
for X in Xs:
sel = SelectKBest(dummy_score, k=1)
X1 = ignore_warnings(sel.fit_transform)([X], y)
assert X1.shape[1] == 1
assert_best_scores_kept(sel)
sel = SelectKBest(dummy_score, k=2)
X2 = ignore_warnings(sel.fit_transform)([X], y)
assert X2.shape[1] == 2
assert_best_scores_kept(sel)
def test_mutual_info_classif():
X, y = make_classification(n_samples=100,
n_features=5,
n_informative=1,
n_redundant=1,
n_repeated=0,
n_classes=2,
n_clusters_per_class=1,
flip_y=0.0,
class_sep=10,
shuffle=False,
random_state=0)
# Test in KBest mode.
univariate_filter = SelectKBest(mutual_info_classif, k=2)
X_r = univariate_filter.fit(X, y).transform(X)
X_r2 = GenericUnivariateSelect(mutual_info_classif, mode='k_best',
param=2).fit(X, y).transform(X)
assert_array_equal(X_r, X_r2)
support = univariate_filter.get_support()
gtruth = np.zeros(5)
gtruth[:2] = 1
assert_array_equal(support, gtruth)
# Test in Percentile mode.
univariate_filter = SelectPercentile(mutual_info_classif, percentile=40)
X_r = univariate_filter.fit(X, y).transform(X)
X_r2 = GenericUnivariateSelect(mutual_info_classif,
mode='percentile',
param=40).fit(X, y).transform(X)
assert_array_equal(X_r, X_r2)
support = univariate_filter.get_support()
gtruth = np.zeros(5)
gtruth[:2] = 1
assert_array_equal(support, gtruth)
def test_mutual_info_regression():
X, y = make_regression(n_samples=100,
n_features=10,
n_informative=2,
shuffle=False,
random_state=0,
noise=10)
# Test in KBest mode.
univariate_filter = SelectKBest(mutual_info_regression, k=2)
X_r = univariate_filter.fit(X, y).transform(X)
assert_best_scores_kept(univariate_filter)
X_r2 = GenericUnivariateSelect(mutual_info_regression,
mode='k_best',
param=2).fit(X, y).transform(X)
assert_array_equal(X_r, X_r2)
support = univariate_filter.get_support()
gtruth = np.zeros(10)
gtruth[:2] = 1
assert_array_equal(support, gtruth)
# Test in Percentile mode.
univariate_filter = SelectPercentile(mutual_info_regression, percentile=20)
X_r = univariate_filter.fit(X, y).transform(X)
X_r2 = GenericUnivariateSelect(mutual_info_regression,
mode='percentile',
param=20).fit(X, y).transform(X)
assert_array_equal(X_r, X_r2)
support = univariate_filter.get_support()
gtruth = np.zeros(10)
gtruth[:2] = 1
assert_array_equal(support, gtruth)
def test_nans():
# Assert that SelectKBest and SelectPercentile can handle NaNs.
# First feature has zero variance to confuse f_classif (ANOVA) and
# make it return a NaN.
X = [[0, 1, 0], [0, -1, -1], [0, .5, .5]]
y = [1, 0, 1]
for select in (SelectKBest(f_classif,
2), SelectPercentile(f_classif, percentile=67)):
ignore_warnings(select.fit)(X, y)
assert_array_equal(select.get_support(indices=True), np.array([1, 2]))
def test_select_kbest_regression():
# Test whether the relative univariate feature selection
# gets the correct items in a simple regression problem
# with the k best heuristic
X, y = make_regression(n_samples=200,
n_features=20,
n_informative=5,
shuffle=False,
random_state=0,
noise=10)
univariate_filter = SelectKBest(f_regression, k=5)
X_r = univariate_filter.fit(X, y).transform(X)
assert_best_scores_kept(univariate_filter)
X_r2 = GenericUnivariateSelect(f_regression, mode='k_best',
param=5).fit(X, y).transform(X)
assert_array_equal(X_r, X_r2)
support = univariate_filter.get_support()
gtruth = np.zeros(20)
gtruth[:5] = 1
assert_array_equal(support, gtruth)
def test_scorefunc_multilabel():
# Test whether k-best and percentiles works with multilabels with chi2.
X = np.array([[10000, 9999, 0], [100, 9999, 0], [1000, 99, 0]])
y = [[1, 1], [0, 1], [1, 0]]
Xt = SelectKBest(chi2, k=2).fit_transform(X, y)
assert Xt.shape == (3, 2)
assert 0 not in Xt
Xt = SelectPercentile(chi2, percentile=67).fit_transform(X, y)
assert Xt.shape == (3, 2)
assert 0 not in Xt
def test_pipeline_methods_anova():
# Test the various methods of the pipeline (anova).
iris = load_iris()
X = iris.data
y = iris.target
# Test with Anova + LogisticRegression
clf = LogisticRegression()
filter1 = SelectKBest(f_classif, k=2)
pipe = Pipeline([('anova', filter1), ('logistic', clf)])
pipe.fit(X, y)
pipe.predict(X)
pipe.predict_proba(X)
pipe.predict_log_proba(X)
pipe.score(X, y)
def test_tied_pvalues():
# Test whether k-best and percentiles work with tied pvalues from chi2.
# chi2 will return the same p-values for the following features, but it
# will return different scores.
X0 = np.array([[10000, 9999, 9998], [1, 1, 1]])
y = [0, 1]
for perm in itertools.permutations((0, 1, 2)):
X = X0[:, perm]
Xt = SelectKBest(chi2, k=2).fit_transform(X, y)
assert Xt.shape == (2, 2)
assert 9998 not in Xt
Xt = SelectPercentile(chi2, percentile=67).fit_transform(X, y)
assert Xt.shape == (2, 2)
assert 9998 not in Xt
def test_no_feature_selected():
rng = np.random.RandomState(0)
# Generate random uncorrelated data: a strict univariate test should
# rejects all the features
X = rng.rand(40, 10)
y = rng.randint(0, 4, size=40)
strict_selectors = [
SelectFwe(alpha=0.01).fit(X, y),
SelectFdr(alpha=0.01).fit(X, y),
SelectFpr(alpha=0.01).fit(X, y),
SelectPercentile(percentile=0).fit(X, y),
SelectKBest(k=0).fit(X, y),
]
for selector in strict_selectors:
assert_array_equal(selector.get_support(), np.zeros(10))
X_selected = assert_warns_message(UserWarning,
'No features were selected',
selector.transform, X)
assert X_selected.shape == (40, 0)
from mrex.pipeline import Pipeline, FeatureUnion
from mrex.model_selection import GridSearchCV
from mrex.svm import SVC
from mrex.datasets import load_iris
from mrex.decomposition import PCA
from mrex.feature_selection import SelectKBest
iris = load_iris()
X, y = iris.data, iris.target
# This dataset is way too high-dimensional. Better do PCA:
pca = PCA(n_components=2)
# Maybe some original features where good, too?
selection = SelectKBest(k=1)
# Build estimator from PCA and Univariate selection:
combined_features = FeatureUnion([("pca", pca), ("univ_select", selection)])
# Use combined features to transform dataset:
X_features = combined_features.fit(X, y).transform(X)
print("Combined space has", X_features.shape[1], "features")
svm = SVC(kernel="linear")
# Do grid search over k, n_components and C:
pipeline = Pipeline([("features", combined_features), ("svm", svm)])
from mrex.datasets import samples_generator
from mrex.feature_selection import SelectKBest, f_regression
from mrex.pipeline import make_pipeline
from mrex.model_selection import train_test_split
from mrex.metrics import classification_report
print(__doc__)
# import some data to play with
X, y = samples_generator.make_classification(n_features=20,
n_informative=3,
n_redundant=0,
n_classes=4,
n_clusters_per_class=2)
X_train, X_test, y_train, y_test = train_test_split(X, y, random_state=42)
# ANOVA SVM-C
# 1) anova filter, take 3 best ranked features
anova_filter = SelectKBest(f_regression, k=3)
# 2) svm
clf = svm.LinearSVC()
anova_svm = make_pipeline(anova_filter, clf)
anova_svm.fit(X_train, y_train)
y_pred = anova_svm.predict(X_test)
print(classification_report(y_test, y_pred))
coef = anova_svm[:-1].inverse_transform(anova_svm['linearsvc'].coef_)
print(coef)
def test_pipeline_init():
# Test the various init parameters of the pipeline.
assert_raises(TypeError, Pipeline)
# Check that we can't instantiate pipelines with objects without fit
# method
assert_raises_regex(TypeError,
'Last step of Pipeline should implement fit '
'or be the string \'passthrough\''
'.*NoFit.*',
Pipeline, [('clf', NoFit())])
# Smoke test with only an estimator
clf = NoTrans()
pipe = Pipeline([('svc', clf)])
assert (pipe.get_params(deep=True) ==
dict(svc__a=None, svc__b=None, svc=clf,
**pipe.get_params(deep=False)))
# Check that params are set
pipe.set_params(svc__a=0.1)
assert clf.a == 0.1
assert clf.b is None
# Smoke test the repr:
repr(pipe)
# Test with two objects
clf = SVC()
filter1 = SelectKBest(f_classif)
pipe = Pipeline([('anova', filter1), ('svc', clf)])
# Check that estimators are not cloned on pipeline construction
assert pipe.named_steps['anova'] is filter1
assert pipe.named_steps['svc'] is clf
# Check that we can't instantiate with non-transformers on the way
# Note that NoTrans implements fit, but not transform
assert_raises_regex(TypeError,
'All intermediate steps should be transformers'
'.*\\bNoTrans\\b.*',
Pipeline, [('t', NoTrans()), ('svc', clf)])
# Check that params are set
pipe.set_params(svc__C=0.1)
assert clf.C == 0.1
# Smoke test the repr:
repr(pipe)
# Check that params are not set when naming them wrong
assert_raises(ValueError, pipe.set_params, anova__C=0.1)
# Test clone
pipe2 = assert_no_warnings(clone, pipe)
assert not pipe.named_steps['svc'] is pipe2.named_steps['svc']
# Check that apart from estimators, the parameters are the same
params = pipe.get_params(deep=True)
params2 = pipe2.get_params(deep=True)
for x in pipe.get_params(deep=False):
params.pop(x)
for x in pipe2.get_params(deep=False):
params2.pop(x)
# Remove estimators that where copied
params.pop('svc')
params.pop('anova')
params2.pop('svc')
params2.pop('anova')
assert params == params2
pipe = Pipeline([
# the reduce_dim stage is populated by the param_grid
('reduce_dim', 'passthrough'),
('classify', LinearSVC(dual=False, max_iter=10000))
])
N_FEATURES_OPTIONS = [2, 4, 8]
C_OPTIONS = [1, 10, 100, 1000]
param_grid = [
{
'reduce_dim': [PCA(iterated_power=7), NMF()],
'reduce_dim__n_components': N_FEATURES_OPTIONS,
'classify__C': C_OPTIONS
},
{
'reduce_dim': [SelectKBest(chi2)],
'reduce_dim__k': N_FEATURES_OPTIONS,
'classify__C': C_OPTIONS
},
]
reducer_labels = ['PCA', 'NMF', 'KBest(chi2)']
grid = GridSearchCV(pipe, n_jobs=1, param_grid=param_grid)
X, y = load_digits(return_X_y=True)
grid.fit(X, y)
mean_scores = np.array(grid.cv_results_['mean_test_score'])
# scores are in the order of param_grid iteration, which is alphabetical
mean_scores = mean_scores.reshape(len(C_OPTIONS), -1, len(N_FEATURES_OPTIONS))
# select score for best C
mean_scores = mean_scores.max(axis=0)
# Split dataset to select feature and evaluate the classifier
X_train, X_test, y_train, y_test = train_test_split(X,
y,
stratify=y,
random_state=0)
plt.figure(1)
plt.clf()
X_indices = np.arange(X.shape[-1])
# #############################################################################
# Univariate feature selection with F-test for feature scoring
# We use the default selection function to select the four
# most significant features
selector = SelectKBest(f_classif, k=4)
selector.fit(X_train, y_train)
scores = -np.log10(selector.pvalues_)
scores /= scores.max()
plt.bar(X_indices - .45,
scores,
width=.2,
label=r'Univariate score ($-Log(p_{value})$)',
color='darkorange',
edgecolor='black')
# #############################################################################
# Compare to the weights of an SVM
clf = make_pipeline(MinMaxScaler(), LinearSVC())
clf.fit(X_train, y_train)
print('Classification accuracy without selecting features: {:.3f}'.format(
def test_gridsearch_pipeline():
# render a pipeline inside a gridsearch
pp = _EstimatorPrettyPrinter(compact=True, indent=1, indent_at_name=True)
pipeline = Pipeline([
('reduce_dim', PCA()),
('classify', SVC())
])
N_FEATURES_OPTIONS = [2, 4, 8]
C_OPTIONS = [1, 10, 100, 1000]
param_grid = [
{
'reduce_dim': [PCA(iterated_power=7), NMF()],
'reduce_dim__n_components': N_FEATURES_OPTIONS,
'classify__C': C_OPTIONS
},
{
'reduce_dim': [SelectKBest(chi2)],
'reduce_dim__k': N_FEATURES_OPTIONS,
'classify__C': C_OPTIONS
}
]
gspipline = GridSearchCV(pipeline, cv=3, n_jobs=1, param_grid=param_grid)
expected = """
GridSearchCV(cv=3, error_score='raise-deprecating',
estimator=Pipeline(memory=None,
steps=[('reduce_dim',
PCA(copy=True, iterated_power='auto',
n_components=None,
random_state=None,
svd_solver='auto', tol=0.0,
whiten=False)),
('classify',
SVC(C=1.0, cache_size=200,
class_weight=None, coef0=0.0,
decision_function_shape='ovr',
degree=3, gamma='auto_deprecated',
kernel='rbf', max_iter=-1,
probability=False,
random_state=None, shrinking=True,
tol=0.001, verbose=False))]),
iid='warn', n_jobs=1,
param_grid=[{'classify__C': [1, 10, 100, 1000],
'reduce_dim': [PCA(copy=True, iterated_power=7,
n_components=None,
random_state=None,
svd_solver='auto', tol=0.0,
whiten=False),
NMF(alpha=0.0, beta_loss='frobenius',
init=None, l1_ratio=0.0,
max_iter=200, n_components=None,
random_state=None, shuffle=False,
solver='cd', tol=0.0001,
verbose=0)],
'reduce_dim__n_components': [2, 4, 8]},
{'classify__C': [1, 10, 100, 1000],
'reduce_dim': [SelectKBest(k=10,
score_func=<function chi2 at some_address>)],
'reduce_dim__k': [2, 4, 8]}],
pre_dispatch='2*n_jobs', refit=True, return_train_score=False,
scoring=None, verbose=0)"""
expected = expected[1:] # remove first \n
repr_ = pp.pformat(gspipline)
# Remove address of '<function chi2 at 0x.....>' for reproducibility
repr_ = re.sub('function chi2 at 0x.*>',
'function chi2 at some_address>', repr_)
assert repr_ == expected
def mkchi2(k):
"""Make k-best chi2 selector"""
return SelectKBest(chi2, k=k)