def test_graphical_lasso_iris():
# Hard-coded solution from R glasso package for alpha=1.0
# (need to set penalize.diagonal to FALSE)
cov_R = np.array([
[0.68112222, 0.0000000, 0.265820, 0.02464314],
[0.00000000, 0.1887129, 0.000000, 0.00000000],
[0.26582000, 0.0000000, 3.095503, 0.28697200],
[0.02464314, 0.0000000, 0.286972, 0.57713289]
])
icov_R = np.array([
[1.5190747, 0.000000, -0.1304475, 0.0000000],
[0.0000000, 5.299055, 0.0000000, 0.0000000],
[-0.1304475, 0.000000, 0.3498624, -0.1683946],
[0.0000000, 0.000000, -0.1683946, 1.8164353]
])
X = datasets.load_iris().data
emp_cov = empirical_covariance(X)
for method in ('cd', 'lars'):
cov, icov = graphical_lasso(emp_cov, alpha=1.0, return_costs=False,
mode=method)
assert_array_almost_equal(cov, cov_R)
assert_array_almost_equal(icov, icov_R)
def test_skewed_chi2_sampler():
# test that RBFSampler approximates kernel on random data
# compute exact kernel
c = 0.03
# set on negative component but greater than c to ensure that the kernel
# approximation is valid on the group (-c; +\infty) endowed with the skewed
# multiplication.
Y[0, 0] = -c / 2.
# abbreviations for easier formula
X_c = (X + c)[:, np.newaxis, :]
Y_c = (Y + c)[np.newaxis, :, :]
# we do it in log-space in the hope that it's more stable
# this array is n_samples_x x n_samples_y big x n_features
log_kernel = ((np.log(X_c) / 2.) + (np.log(Y_c) / 2.) + np.log(2.) -
np.log(X_c + Y_c))
# reduce to n_samples_x x n_samples_y by summing over features in log-space
kernel = np.exp(log_kernel.sum(axis=2))
# approximate kernel mapping
transform = SkewedChi2Sampler(skewedness=c,
n_components=1000,
random_state=42)
X_trans = transform.fit_transform(X)
Y_trans = transform.transform(Y)
kernel_approx = np.dot(X_trans, Y_trans.T)
assert_array_almost_equal(kernel, kernel_approx, 1)
assert np.isfinite(kernel).all(), \
'NaNs found in the Gram matrix'
assert np.isfinite(kernel_approx).all(), \
'NaNs found in the approximate Gram matrix'
# test error is raised on when inputs contains values smaller than -c
Y_neg = Y.copy()
Y_neg[0, 0] = -c * 2.
assert_raises(ValueError, transform.transform, Y_neg)
def test_iforest_parallel_regression():
"""Check parallel regression."""
rng = check_random_state(0)
X_train, X_test, y_train, y_test = train_test_split(boston.data,
boston.target,
random_state=rng)
ensemble = IsolationForest(n_jobs=3,
random_state=0).fit(X_train)
ensemble.set_params(n_jobs=1)
y1 = ensemble.predict(X_test)
ensemble.set_params(n_jobs=2)
y2 = ensemble.predict(X_test)
assert_array_almost_equal(y1, y2)
ensemble = IsolationForest(n_jobs=1,
random_state=0).fit(X_train)
y3 = ensemble.predict(X_test)
assert_array_almost_equal(y1, y3)
def test_ridgecv_sample_weight():
rng = np.random.RandomState(0)
alphas = (0.1, 1.0, 10.0)
# There are different algorithms for n_samples > n_features
# and the opposite, so test them both.
for n_samples, n_features in ((6, 5), (5, 10)):
y = rng.randn(n_samples)
X = rng.randn(n_samples, n_features)
sample_weight = 1.0 + rng.rand(n_samples)
cv = KFold(5)
ridgecv = RidgeCV(alphas=alphas, cv=cv)
ridgecv.fit(X, y, sample_weight=sample_weight)
# Check using GridSearchCV directly
parameters = {'alpha': alphas}
gs = GridSearchCV(Ridge(), parameters, cv=cv)
gs.fit(X, y, sample_weight=sample_weight)
assert ridgecv.alpha_ == gs.best_estimator_.alpha
assert_array_almost_equal(ridgecv.coef_, gs.best_estimator_.coef_)
def test_compute_log_det_cholesky():
n_features = 2
rand_data = RandomData(np.random.RandomState(0))
for covar_type in COVARIANCE_TYPE:
covariance = rand_data.covariances[covar_type]
if covar_type == 'full':
predected_det = np.array([linalg.det(cov) for cov in covariance])
elif covar_type == 'tied':
predected_det = linalg.det(covariance)
elif covar_type == 'diag':
predected_det = np.array([np.prod(cov) for cov in covariance])
elif covar_type == 'spherical':
predected_det = covariance**n_features
# We compute the cholesky decomposition of the covariance matrix
expected_det = _compute_log_det_cholesky(_compute_precision_cholesky(
covariance, covar_type),
covar_type,
n_features=n_features)
assert_array_almost_equal(expected_det, -.5 * np.log(predected_det))
def _check_predict_proba(clf, X, y):
proba = clf.predict_proba(X)
# We know that we can have division by zero
log_proba = clf.predict_log_proba(X)
y = np.atleast_1d(y)
if y.ndim == 1:
y = np.reshape(y, (-1, 1))
n_outputs = y.shape[1]
n_samples = len(X)
if n_outputs == 1:
proba = [proba]
log_proba = [log_proba]
for k in range(n_outputs):
assert proba[k].shape[0] == n_samples
assert proba[k].shape[1] == len(np.unique(y[:, k]))
assert_array_almost_equal(proba[k].sum(axis=1), np.ones(len(X)))
# We know that we can have division by zero
assert_array_almost_equal(np.log(proba[k]), log_proba[k])
def test_2d_coef():
X, y = datasets.make_classification(
n_samples=1000, n_features=10, n_informative=3, n_redundant=0,
n_repeated=0, shuffle=False, random_state=0, n_classes=4)
est = LogisticRegression()
for threshold, func in zip(["mean", "median"], [np.mean, np.median]):
for order in [1, 2, np.inf]:
# Fit SelectFromModel a multi-class problem
transformer = SelectFromModel(estimator=LogisticRegression(),
threshold=threshold,
norm_order=order)
transformer.fit(X, y)
assert hasattr(transformer.estimator_, 'coef_')
X_new = transformer.transform(X)
assert X_new.shape[1] < X.shape[1]
# Manually check that the norm is correctly performed
est.fit(X, y)
importances = np.linalg.norm(est.coef_, axis=0, ord=order)
feature_mask = importances > func(importances)
assert_array_almost_equal(X_new, X[:, feature_mask])
def test_no_path_all_precomputed():
# Test that the ``return_path=False`` option with Gram and Xy remains
# correct
X, y = 3 * diabetes.data, diabetes.target
G = np.dot(X.T, X)
Xy = np.dot(X.T, y)
alphas_, _, coef_path_ = linear_model.lars_path(X,
y,
method='lasso',
Xy=Xy,
Gram=G,
alpha_min=0.9)
alpha_, _, coef = linear_model.lars_path(X,
y,
method='lasso',
Gram=G,
Xy=Xy,
alpha_min=0.9,
return_path=False)
assert_array_almost_equal(coef, coef_path_[:, -1])
assert alpha_ == alphas_[-1]
def test_isotonic_regression_with_ties_in_differently_sized_groups():
"""
Non-regression test to handle issue 9432:
https://github.com/scikit-learn/scikit-learn/issues/9432
Compare against output in R:
> library("isotone")
> x <- c(0, 1, 1, 2, 3, 4)
> y <- c(0, 0, 1, 0, 0, 1)
> res1 <- gpava(x, y, ties="secondary")
> res1$x
`isotone` version: 1.1-0, 2015-07-24
R version: R version 3.3.2 (2016-10-31)
"""
x = np.array([0, 1, 1, 2, 3, 4])
y = np.array([0, 0, 1, 0, 0, 1])
y_true = np.array([0., 0.25, 0.25, 0.25, 0.25, 1.])
ir = IsotonicRegression()
ir.fit(x, y)
assert_array_almost_equal(ir.transform(x), y_true)
assert_array_almost_equal(ir.fit_transform(x, y), y_true)
def test_non_negative_factorization_consistency():
# Test that the function is called in the same way, either directly
# or through the NMF class
rng = np.random.mtrand.RandomState(42)
A = np.abs(rng.randn(10, 10))
A[:, 2 * np.arange(5)] = 0
for init in ['random', 'nndsvd']:
for solver in ('cd', 'mu'):
W_nmf, H, _ = non_negative_factorization(
A, init=init, solver=solver, random_state=1, tol=1e-2)
W_nmf_2, _, _ = non_negative_factorization(
A, H=H, update_H=False, init=init, solver=solver,
random_state=1, tol=1e-2)
model_class = NMF(init=init, solver=solver, random_state=1,
tol=1e-2)
W_cls = model_class.fit_transform(A)
W_cls_2 = model_class.transform(A)
assert_array_almost_equal(W_nmf, W_cls, decimal=10)
assert_array_almost_equal(W_nmf_2, W_cls_2, decimal=10)
def check_importances(name, criterion, dtype, tolerance):
# cast as dype
X = X_large.astype(dtype, copy=False)
y = y_large.astype(dtype, copy=False)
ForestEstimator = FOREST_ESTIMATORS[name]
est = ForestEstimator(n_estimators=10, criterion=criterion,
random_state=0)
est.fit(X, y)
importances = est.feature_importances_
# The forest estimator can detect that only the first 3 features of the
# dataset are informative:
n_important = np.sum(importances > 0.1)
assert importances.shape[0] == 10
assert n_important == 3
assert np.all(importances[:3] > 0.1)
# Check with parallel
importances = est.feature_importances_
est.set_params(n_jobs=2)
importances_parallel = est.feature_importances_
assert_array_almost_equal(importances, importances_parallel)
# Check with sample weights
sample_weight = check_random_state(0).randint(1, 10, len(X))
est = ForestEstimator(n_estimators=10, random_state=0, criterion=criterion)
est.fit(X, y, sample_weight=sample_weight)
importances = est.feature_importances_
assert np.all(importances >= 0.0)
for scale in [0.5, 100]:
est = ForestEstimator(n_estimators=10, random_state=0,
criterion=criterion)
est.fit(X, y, sample_weight=scale * sample_weight)
importances_bis = est.feature_importances_
assert np.abs(importances - importances_bis).mean() < tolerance
def test_kernel_pca():
rng = np.random.RandomState(0)
X_fit = rng.random_sample((5, 4))
X_pred = rng.random_sample((2, 4))
def histogram(x, y, **kwargs):
# Histogram kernel implemented as a callable.
assert kwargs == {} # no kernel_params that we didn't ask for
return np.minimum(x, y).sum()
for eigen_solver in ("auto", "dense", "arpack"):
for kernel in ("linear", "rbf", "poly", histogram):
# histogram kernel produces singular matrix inside linalg.solve
# XXX use a least-squares approximation?
inv = not callable(kernel)
# transform fit data
kpca = KernelPCA(4,
kernel=kernel,
eigen_solver=eigen_solver,
fit_inverse_transform=inv)
X_fit_transformed = kpca.fit_transform(X_fit)
X_fit_transformed2 = kpca.fit(X_fit).transform(X_fit)
assert_array_almost_equal(np.abs(X_fit_transformed),
np.abs(X_fit_transformed2))
# non-regression test: previously, gamma would be 0 by default,
# forcing all eigenvalues to 0 under the poly kernel
assert X_fit_transformed.size != 0
# transform new data
X_pred_transformed = kpca.transform(X_pred)
assert (X_pred_transformed.shape[1] == X_fit_transformed.shape[1])
# inverse transform
if inv:
X_pred2 = kpca.inverse_transform(X_pred_transformed)
assert X_pred2.shape == X_pred.shape
def test_newton_cg():
# Test that newton_cg gives same result as scipy's fmin_ncg
rng = np.random.RandomState(0)
A = rng.normal(size=(10, 10))
x0 = np.ones(10)
def func(x):
Ax = A.dot(x)
return .5 * (Ax).dot(Ax)
def grad(x):
return A.T.dot(A.dot(x))
def hess(x, p):
return p.dot(A.T.dot(A.dot(x.all())))
def grad_hess(x):
return grad(x), lambda x: A.T.dot(A.dot(x))
assert_array_almost_equal(
newton_cg(grad_hess, func, grad, x0, tol=1e-10)[0],
fmin_ncg(f=func, x0=x0, fprime=grad, fhess_p=hess))
def test_dict_learning_online_partial_fit():
n_components = 12
rng = np.random.RandomState(0)
V = rng.randn(n_components, n_features) # random init
V /= np.sum(V**2, axis=1)[:, np.newaxis]
dict1 = MiniBatchDictionaryLearning(n_components,
n_iter=10 * len(X),
batch_size=1,
alpha=1,
shuffle=False,
dict_init=V,
random_state=0).fit(X)
dict2 = MiniBatchDictionaryLearning(n_components,
alpha=1,
n_iter=1,
dict_init=V,
random_state=0)
for i in range(10):
for sample in X:
dict2.partial_fit(sample[np.newaxis, :])
assert not np.all(sparse_encode(X, dict1.components_, alpha=1) == 0)
assert_array_almost_equal(dict1.components_, dict2.components_, decimal=2)
def test_nmf_multiplicative_update_sparse():
# Compare sparse and dense input in multiplicative update NMF
# Also test continuity of the results with respect to beta_loss parameter
n_samples = 20
n_features = 10
n_components = 5
alpha = 0.1
l1_ratio = 0.5
n_iter = 20
# initialization
rng = np.random.mtrand.RandomState(1337)
X = rng.randn(n_samples, n_features)
X = np.abs(X)
X_csr = sp.csr_matrix(X)
W0, H0 = nmf._initialize_nmf(X, n_components, init='random',
random_state=42)
for beta_loss in (-1.2, 0, 0.2, 1., 2., 2.5):
# Reference with dense array X
W, H = W0.copy(), H0.copy()
W1, H1, _ = non_negative_factorization(
X, W, H, n_components, init='custom', update_H=True,
solver='mu', beta_loss=beta_loss, max_iter=n_iter, alpha=alpha,
l1_ratio=l1_ratio, regularization='both', random_state=42)
# Compare with sparse X
W, H = W0.copy(), H0.copy()
W2, H2, _ = non_negative_factorization(
X_csr, W, H, n_components, init='custom', update_H=True,
solver='mu', beta_loss=beta_loss, max_iter=n_iter, alpha=alpha,
l1_ratio=l1_ratio, regularization='both', random_state=42)
assert_array_almost_equal(W1, W2, decimal=7)
assert_array_almost_equal(H1, H2, decimal=7)
# Compare with almost same beta_loss, since some values have a specific
# behavior, but the results should be continuous w.r.t beta_loss
beta_loss -= 1.e-5
W, H = W0.copy(), H0.copy()
W3, H3, _ = non_negative_factorization(
X_csr, W, H, n_components, init='custom', update_H=True,
solver='mu', beta_loss=beta_loss, max_iter=n_iter, alpha=alpha,
l1_ratio=l1_ratio, regularization='both', random_state=42)
assert_array_almost_equal(W1, W3, decimal=4)
assert_array_almost_equal(H1, H3, decimal=4)
def test_lda_predict():
# Test LDA classification.
# This checks that LDA implements fit and predict and returns correct
# values for simple toy data.
for test_case in solver_shrinkage:
solver, shrinkage = test_case
clf = LinearDiscriminantAnalysis(solver=solver, shrinkage=shrinkage)
y_pred = clf.fit(X, y).predict(X)
assert_array_equal(y_pred, y, 'solver %s' % solver)
# Assert that it works with 1D data
y_pred1 = clf.fit(X1, y).predict(X1)
assert_array_equal(y_pred1, y, 'solver %s' % solver)
# Test probability estimates
y_proba_pred1 = clf.predict_proba(X1)
assert_array_equal((y_proba_pred1[:, 1] > 0.5) + 1, y,
'solver %s' % solver)
y_log_proba_pred1 = clf.predict_log_proba(X1)
assert_array_almost_equal(np.exp(y_log_proba_pred1), y_proba_pred1, 8,
'solver %s' % solver)
# Primarily test for commit 2f34950 -- "reuse" of priors
y_pred3 = clf.fit(X, y3).predict(X)
# LDA shouldn't be able to separate those
assert np.any(y_pred3 != y3), 'solver %s' % solver
# Test invalid shrinkages
clf = LinearDiscriminantAnalysis(solver="lsqr", shrinkage=-0.2231)
assert_raises(ValueError, clf.fit, X, y)
clf = LinearDiscriminantAnalysis(solver="eigen", shrinkage="dummy")
assert_raises(ValueError, clf.fit, X, y)
clf = LinearDiscriminantAnalysis(solver="svd", shrinkage="auto")
assert_raises(NotImplementedError, clf.fit, X, y)
# Test unknown solver
clf = LinearDiscriminantAnalysis(solver="dummy")
assert_raises(ValueError, clf.fit, X, y)
def test_prefit():
# Test all possible combinations of the prefit parameter.
# Passing a prefit parameter with the selected model
# and fitting a unfit model with prefit=False should give same results.
clf = SGDClassifier(alpha=0.1, max_iter=10, shuffle=True,
random_state=0, tol=None)
model = SelectFromModel(clf)
model.fit(data, y)
X_transform = model.transform(data)
clf.fit(data, y)
model = SelectFromModel(clf, prefit=True)
assert_array_almost_equal(model.transform(data), X_transform)
# Check that the model is rewritten if prefit=False and a fitted model is
# passed
model = SelectFromModel(clf, prefit=False)
model.fit(data, y)
assert_array_almost_equal(model.transform(data), X_transform)
# Check that prefit=True and calling fit raises a ValueError
model = SelectFromModel(clf, prefit=True)
with pytest.raises(ValueError):
model.fit(data, y)
def test_ridge_regression_sample_weights():
rng = np.random.RandomState(0)
for solver in ("cholesky", ):
for n_samples, n_features in ((6, 5), (5, 10)):
for alpha in (1.0, 1e-2):
y = rng.randn(n_samples)
X = rng.randn(n_samples, n_features)
sample_weight = 1.0 + rng.rand(n_samples)
coefs = ridge_regression(X,
y,
alpha=alpha,
sample_weight=sample_weight,
solver=solver)
# Sample weight can be implemented via a simple rescaling
# for the square loss.
coefs2 = ridge_regression(
X * np.sqrt(sample_weight)[:, np.newaxis],
y * np.sqrt(sample_weight),
alpha=alpha,
solver=solver)
assert_array_almost_equal(coefs, coefs2)
def test_lda_explained_variance_ratio():
# Test if the sum of the normalized eigen vectors values equals 1,
# Also tests whether the explained_variance_ratio_ formed by the
# eigen solver is the same as the explained_variance_ratio_ formed
# by the svd solver
state = np.random.RandomState(0)
X = state.normal(loc=0, scale=100, size=(40, 20))
y = state.randint(0, 3, size=(40, ))
clf_lda_eigen = LinearDiscriminantAnalysis(solver="eigen")
clf_lda_eigen.fit(X, y)
assert_almost_equal(clf_lda_eigen.explained_variance_ratio_.sum(), 1.0, 3)
assert clf_lda_eigen.explained_variance_ratio_.shape == (2, ), (
"Unexpected length for explained_variance_ratio_")
clf_lda_svd = LinearDiscriminantAnalysis(solver="svd")
clf_lda_svd.fit(X, y)
assert_almost_equal(clf_lda_svd.explained_variance_ratio_.sum(), 1.0, 3)
assert clf_lda_svd.explained_variance_ratio_.shape == (2, ), (
"Unexpected length for explained_variance_ratio_")
assert_array_almost_equal(clf_lda_svd.explained_variance_ratio_,
clf_lda_eigen.explained_variance_ratio_)
def test_fit_predict_on_pipeline():
# test that the fit_predict method is implemented on a pipeline
# test that the fit_predict on pipeline yields same results as applying
# transform and clustering steps separately
iris = load_iris()
scaler = StandardScaler()
km = KMeans(random_state=0)
# As pipeline doesn't clone estimators on construction,
# it must have its own estimators
scaler_for_pipeline = StandardScaler()
km_for_pipeline = KMeans(random_state=0)
# first compute the transform and clustering step separately
scaled = scaler.fit_transform(iris.data)
separate_pred = km.fit_predict(scaled)
# use a pipeline to do the transform and clustering in one step
pipe = Pipeline([
('scaler', scaler_for_pipeline),
('Kmeans', km_for_pipeline)
])
pipeline_pred = pipe.fit_predict(iris.data)
assert_array_almost_equal(pipeline_pred, separate_pred)
def test_qda():
# QDA classification.
# This checks that QDA implements fit and predict and returns
# correct values for a simple toy dataset.
clf = QuadraticDiscriminantAnalysis()
y_pred = clf.fit(X6, y6).predict(X6)
assert_array_equal(y_pred, y6)
# Assure that it works with 1D data
y_pred1 = clf.fit(X7, y6).predict(X7)
assert_array_equal(y_pred1, y6)
# Test probas estimates
y_proba_pred1 = clf.predict_proba(X7)
assert_array_equal((y_proba_pred1[:, 1] > 0.5) + 1, y6)
y_log_proba_pred1 = clf.predict_log_proba(X7)
assert_array_almost_equal(np.exp(y_log_proba_pred1), y_proba_pred1, 8)
y_pred3 = clf.fit(X6, y7).predict(X6)
# QDA shouldn't be able to separate those
assert np.any(y_pred3 != y7)
# Classes should have at least 2 elements
assert_raises(ValueError, clf.fit, X6, y4)
def test_dump_concise():
one = 1
two = 2.1
three = 3.01
exact = 1.000000000000001
# loses the last decimal place
almost = 1.0000000000000001
X = [[one, two, three, exact, almost], [1e9, 2e18, 3e27, 0, 0],
[0, 0, 0, 0, 0], [0, 0, 0, 0, 0], [0, 0, 0, 0, 0]]
y = [one, two, three, exact, almost]
f = BytesIO()
dump_svmlight_file(X, y, f)
f.seek(0)
# make sure it's using the most concise format possible
assert (f.readline() == b"1 0:1 1:2.1 2:3.01 3:1.000000000000001 4:1\n")
assert f.readline() == b"2.1 0:1000000000 1:2e+18 2:3e+27\n"
assert f.readline() == b"3.01 \n"
assert f.readline() == b"1.000000000000001 \n"
assert f.readline() == b"1 \n"
f.seek(0)
# make sure it's correct too :)
X2, y2 = load_svmlight_file(f)
assert_array_almost_equal(X, X2.toarray())
assert_array_almost_equal(y, y2)
def test_load_with_offsets(sparsity, n_samples, n_features):
rng = np.random.RandomState(0)
X = rng.uniform(low=0.0, high=1.0, size=(n_samples, n_features))
if sparsity:
X[X < sparsity] = 0.0
X = sp.csr_matrix(X)
y = rng.randint(low=0, high=2, size=n_samples)
f = BytesIO()
dump_svmlight_file(X, y, f)
f.seek(0)
size = len(f.getvalue())
# put some marks that are likely to happen anywhere in a row
mark_0 = 0
mark_1 = size // 3
length_0 = mark_1 - mark_0
mark_2 = 4 * size // 5
length_1 = mark_2 - mark_1
# load the original sparse matrix into 3 independent CSR matrices
X_0, y_0 = load_svmlight_file(f,
n_features=n_features,
offset=mark_0,
length=length_0)
X_1, y_1 = load_svmlight_file(f,
n_features=n_features,
offset=mark_1,
length=length_1)
X_2, y_2 = load_svmlight_file(f, n_features=n_features, offset=mark_2)
y_concat = np.concatenate([y_0, y_1, y_2])
X_concat = sp.vstack([X_0, X_1, X_2])
assert_array_almost_equal(y, y_concat)
assert_array_almost_equal(X.toarray(), X_concat.toarray())
def test_lasso_path_return_models_vs_new_return_gives_same_coefficients():
# Test that lasso_path with lars_path style output gives the
# same result
# Some toy data
X = np.array([[1, 2, 3.1], [2.3, 5.4, 4.3]]).T
y = np.array([1, 2, 3.1])
alphas = [5., 1., .5]
# Use lars_path and lasso_path(new output) with 1D linear interpolation
# to compute the same path
alphas_lars, _, coef_path_lars = lars_path(X, y, method='lasso')
coef_path_cont_lars = interpolate.interp1d(alphas_lars[::-1],
coef_path_lars[:, ::-1])
alphas_lasso2, coef_path_lasso2, _ = lasso_path(X,
y,
alphas=alphas,
return_models=False)
coef_path_cont_lasso = interpolate.interp1d(alphas_lasso2[::-1],
coef_path_lasso2[:, ::-1])
assert_array_almost_equal(coef_path_cont_lasso(alphas),
coef_path_cont_lars(alphas),
decimal=1)
def test_f_classif():
# Test whether the F test yields meaningful results
# on a simple simulated classification problem
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)
F, pv = f_classif(X, y)
F_sparse, pv_sparse = f_classif(sparse.csr_matrix(X), y)
assert (F > 0).all()
assert (pv > 0).all()
assert (pv < 1).all()
assert (pv[:5] < 0.05).all()
assert (pv[5:] > 1.e-4).all()
assert_array_almost_equal(F_sparse, F)
assert_array_almost_equal(pv_sparse, pv)
def test_covariance():
# Tests Covariance module on a simple dataset.
# test covariance fit from data
cov = EmpiricalCovariance()
cov.fit(X)
emp_cov = empirical_covariance(X)
assert_array_almost_equal(emp_cov, cov.covariance_, 4)
assert_almost_equal(cov.error_norm(emp_cov), 0)
assert_almost_equal(cov.error_norm(emp_cov, norm='spectral'), 0)
assert_almost_equal(cov.error_norm(emp_cov, norm='frobenius'), 0)
assert_almost_equal(cov.error_norm(emp_cov, scaling=False), 0)
assert_almost_equal(cov.error_norm(emp_cov, squared=False), 0)
with pytest.raises(NotImplementedError):
cov.error_norm(emp_cov, norm='foo')
# Mahalanobis distances computation test
mahal_dist = cov.mahalanobis(X)
assert np.amin(mahal_dist) > 0
# test with n_features = 1
X_1d = X[:, 0].reshape((-1, 1))
cov = EmpiricalCovariance()
cov.fit(X_1d)
assert_array_almost_equal(empirical_covariance(X_1d), cov.covariance_, 4)
assert_almost_equal(cov.error_norm(empirical_covariance(X_1d)), 0)
assert_almost_equal(
cov.error_norm(empirical_covariance(X_1d), norm='spectral'), 0)
# test with one sample
# Create X with 1 sample and 5 features
X_1sample = np.arange(5).reshape(1, 5)
cov = EmpiricalCovariance()
assert_warns(UserWarning, cov.fit, X_1sample)
assert_array_almost_equal(cov.covariance_,
np.zeros(shape=(5, 5), dtype=np.float64))
# test integer type
X_integer = np.asarray([[0, 1], [1, 0]])
result = np.asarray([[0.25, -0.25], [-0.25, 0.25]])
assert_array_almost_equal(empirical_covariance(X_integer), result)
# test centered case
cov = EmpiricalCovariance(assume_centered=True)
cov.fit(X)
assert_array_equal(cov.location_, np.zeros(X.shape[1]))
def test_random_descent():
# Test that both random and cyclic selection give the same results.
# Ensure that the test models fully converge and check a wide
# range of conditions.
# This uses the coordinate descent algo using the gram trick.
X, y, _, _ = build_dataset(n_samples=50, n_features=20)
clf_cyclic = ElasticNet(selection='cyclic', tol=1e-8)
clf_cyclic.fit(X, y)
clf_random = ElasticNet(selection='random', tol=1e-8, random_state=42)
clf_random.fit(X, y)
assert_array_almost_equal(clf_cyclic.coef_, clf_random.coef_)
assert_almost_equal(clf_cyclic.intercept_, clf_random.intercept_)
# This uses the descent algo without the gram trick
clf_cyclic = ElasticNet(selection='cyclic', tol=1e-8)
clf_cyclic.fit(X.T, y[:20])
clf_random = ElasticNet(selection='random', tol=1e-8, random_state=42)
clf_random.fit(X.T, y[:20])
assert_array_almost_equal(clf_cyclic.coef_, clf_random.coef_)
assert_almost_equal(clf_cyclic.intercept_, clf_random.intercept_)
# Sparse Case
clf_cyclic = ElasticNet(selection='cyclic', tol=1e-8)
clf_cyclic.fit(sparse.csr_matrix(X), y)
clf_random = ElasticNet(selection='random', tol=1e-8, random_state=42)
clf_random.fit(sparse.csr_matrix(X), y)
assert_array_almost_equal(clf_cyclic.coef_, clf_random.coef_)
assert_almost_equal(clf_cyclic.intercept_, clf_random.intercept_)
# Multioutput case.
new_y = np.hstack((y[:, np.newaxis], y[:, np.newaxis]))
clf_cyclic = MultiTaskElasticNet(selection='cyclic', tol=1e-8)
clf_cyclic.fit(X, new_y)
clf_random = MultiTaskElasticNet(selection='random',
tol=1e-8,
random_state=42)
clf_random.fit(X, new_y)
assert_array_almost_equal(clf_cyclic.coef_, clf_random.coef_)
assert_almost_equal(clf_cyclic.intercept_, clf_random.intercept_)
# Raise error when selection is not in cyclic or random.
clf_random = ElasticNet(selection='invalid')
assert_raises(ValueError, clf_random.fit, X, y)
def test_gaussian_mixture_log_probabilities():
from mrex.mixture.gaussian_mixture import _estimate_log_gaussian_prob
# test against with _naive_lmvnpdf_diag
rng = np.random.RandomState(0)
rand_data = RandomData(rng)
n_samples = 500
n_features = rand_data.n_features
n_components = rand_data.n_components
means = rand_data.means
covars_diag = rng.rand(n_components, n_features)
X = rng.rand(n_samples, n_features)
log_prob_naive = _naive_lmvnpdf_diag(X, means, covars_diag)
# full covariances
precs_full = np.array([np.diag(1. / np.sqrt(x)) for x in covars_diag])
log_prob = _estimate_log_gaussian_prob(X, means, precs_full, 'full')
assert_array_almost_equal(log_prob, log_prob_naive)
# diag covariances
precs_chol_diag = 1. / np.sqrt(covars_diag)
log_prob = _estimate_log_gaussian_prob(X, means, precs_chol_diag, 'diag')
assert_array_almost_equal(log_prob, log_prob_naive)
# tied
covars_tied = np.array([x for x in covars_diag]).mean(axis=0)
precs_tied = np.diag(np.sqrt(1. / covars_tied))
log_prob_naive = _naive_lmvnpdf_diag(X, means,
[covars_tied] * n_components)
log_prob = _estimate_log_gaussian_prob(X, means, precs_tied, 'tied')
assert_array_almost_equal(log_prob, log_prob_naive)
# spherical
covars_spherical = covars_diag.mean(axis=1)
precs_spherical = 1. / np.sqrt(covars_diag.mean(axis=1))
log_prob_naive = _naive_lmvnpdf_diag(X, means, [[k] * n_features
for k in covars_spherical])
log_prob = _estimate_log_gaussian_prob(X, means, precs_spherical,
'spherical')
assert_array_almost_equal(log_prob, log_prob_naive)
def test_scale_and_stability():
# We test scale=True parameter
# This allows to check numerical stability over platforms as well
d = load_linnerud()
X1 = d.data
Y1 = d.target
# causes X[:, -1].std() to be zero
X1[:, -1] = 1.0
# From bug #2821
# Test with X2, T2 s.t. clf.x_score[:, 1] == 0, clf.y_score[:, 1] == 0
# This test robustness of algorithm when dealing with value close to 0
X2 = np.array([[0., 0., 1.], [1., 0., 0.], [2., 2., 2.], [3., 5., 4.]])
Y2 = np.array([[0.1, -0.2], [0.9, 1.1], [6.2, 5.9], [11.9, 12.3]])
for (X, Y) in [(X1, Y1), (X2, Y2)]:
X_std = X.std(axis=0, ddof=1)
X_std[X_std == 0] = 1
Y_std = Y.std(axis=0, ddof=1)
Y_std[Y_std == 0] = 1
X_s = (X - X.mean(axis=0)) / X_std
Y_s = (Y - Y.mean(axis=0)) / Y_std
for clf in [
CCA(),
pls_.PLSCanonical(),
pls_.PLSRegression(),
pls_.PLSSVD()
]:
clf.set_params(scale=True)
X_score, Y_score = clf.fit_transform(X, Y)
clf.set_params(scale=False)
X_s_score, Y_s_score = clf.fit_transform(X_s, Y_s)
assert_array_almost_equal(X_s_score, X_score)
assert_array_almost_equal(Y_s_score, Y_score)
# Scaling should be idempotent
clf.set_params(scale=True)
X_score, Y_score = clf.fit_transform(X_s, Y_s)
assert_array_almost_equal(X_s_score, X_score)
assert_array_almost_equal(Y_s_score, Y_score)
def test_set_estimator_none(drop):
"""VotingClassifier set_params should be able to set estimators as None or
drop"""
# Test predict
clf1 = LogisticRegression(random_state=123)
clf2 = RandomForestClassifier(n_estimators=10, random_state=123)
clf3 = GaussianNB()
eclf1 = VotingClassifier(estimators=[('lr', clf1), ('rf', clf2),
('nb', clf3)],
voting='hard', weights=[1, 0, 0.5]).fit(X, y)
eclf2 = VotingClassifier(estimators=[('lr', clf1), ('rf', clf2),
('nb', clf3)],
voting='hard', weights=[1, 1, 0.5])
eclf2.set_params(rf=drop).fit(X, y)
assert_array_equal(eclf1.predict(X), eclf2.predict(X))
assert dict(eclf2.estimators)["rf"] is drop
assert len(eclf2.estimators_) == 2
assert all(isinstance(est, (LogisticRegression, GaussianNB))
for est in eclf2.estimators_)
assert eclf2.get_params()["rf"] is drop
eclf1.set_params(voting='soft').fit(X, y)
eclf2.set_params(voting='soft').fit(X, y)
assert_array_equal(eclf1.predict(X), eclf2.predict(X))
assert_array_almost_equal(eclf1.predict_proba(X), eclf2.predict_proba(X))
msg = 'All estimators are None or "drop". At least one is required!'
assert_raise_message(
ValueError, msg, eclf2.set_params(lr=drop, rf=drop, nb=drop).fit, X, y)
# Test soft voting transform
X1 = np.array([[1], [2]])
y1 = np.array([1, 2])
eclf1 = VotingClassifier(estimators=[('rf', clf2), ('nb', clf3)],
voting='soft', weights=[0, 0.5],
flatten_transform=False).fit(X1, y1)
eclf2 = VotingClassifier(estimators=[('rf', clf2), ('nb', clf3)],
voting='soft', weights=[1, 0.5],
flatten_transform=False)
eclf2.set_params(rf=drop).fit(X1, y1)
assert_array_almost_equal(eclf1.transform(X1),
np.array([[[0.7, 0.3], [0.3, 0.7]],
[[1., 0.], [0., 1.]]]))
assert_array_almost_equal(eclf2.transform(X1),
np.array([[[1., 0.],
[0., 1.]]]))
eclf1.set_params(voting='hard')
eclf2.set_params(voting='hard')
assert_array_equal(eclf1.transform(X1), np.array([[0, 0], [1, 1]]))
assert_array_equal(eclf2.transform(X1), np.array([[0], [1]]))