def test_ovr_partial_fit():
# Test if partial_fit is working as intended
X, y = shuffle(iris.data, iris.target, random_state=0)
ovr = OneVsRestClassifier(MultinomialNB())
ovr.partial_fit(X[:100], y[:100], np.unique(y))
ovr.partial_fit(X[100:], y[100:])
pred = ovr.predict(X)
ovr2 = OneVsRestClassifier(MultinomialNB())
pred2 = ovr2.fit(X, y).predict(X)
assert_almost_equal(pred, pred2)
assert_equal(len(ovr.estimators_), len(np.unique(y)))
assert_greater(np.mean(y == pred), 0.65)
# Test when mini batches doesn't have all classes
# with SGDClassifier
X = np.abs(np.random.randn(14, 2))
y = [1, 1, 1, 1, 2, 3, 3, 0, 0, 2, 3, 1, 2, 3]
ovr = OneVsRestClassifier(SGDClassifier(max_iter=1, tol=None,
shuffle=False, random_state=0))
ovr.partial_fit(X[:7], y[:7], np.unique(y))
ovr.partial_fit(X[7:], y[7:])
pred = ovr.predict(X)
ovr1 = OneVsRestClassifier(SGDClassifier(max_iter=1, tol=None,
shuffle=False, random_state=0))
pred1 = ovr1.fit(X, y).predict(X)
assert_equal(np.mean(pred == y), np.mean(pred1 == y))
# test partial_fit only exists if estimator has it:
ovr = OneVsRestClassifier(SVC())
assert_false(hasattr(ovr, "partial_fit"))
def test_k_means_function():
# test calling the k_means function directly
# catch output
old_stdout = sys.stdout
sys.stdout = StringIO()
try:
cluster_centers, labels, inertia = k_means(X, n_clusters=n_clusters,
verbose=True)
finally:
sys.stdout = old_stdout
centers = cluster_centers
assert_equal(centers.shape, (n_clusters, n_features))
labels = labels
assert_equal(np.unique(labels).shape[0], n_clusters)
# check that the labels assignment are perfect (up to a permutation)
assert_equal(v_measure_score(true_labels, labels), 1.0)
assert_greater(inertia, 0.0)
# check warning when centers are passed
assert_warns(RuntimeWarning, k_means, X, n_clusters=n_clusters,
init=centers)
# to many clusters desired
assert_raises(ValueError, k_means, X, n_clusters=X.shape[0] + 1)
def check_clustering(name, Alg):
X, y = make_blobs(n_samples=50, random_state=1)
X, y = shuffle(X, y, random_state=7)
X = StandardScaler().fit_transform(X)
n_samples, n_features = X.shape
# catch deprecation and neighbors warnings
with warnings.catch_warnings(record=True):
alg = Alg()
set_fast_parameters(alg)
if hasattr(alg, "n_clusters"):
alg.set_params(n_clusters=3)
set_random_state(alg)
if name == 'AffinityPropagation':
alg.set_params(preference=-100)
alg.set_params(max_iter=100)
# fit
alg.fit(X)
# with lists
alg.fit(X.tolist())
assert_equal(alg.labels_.shape, (n_samples,))
pred = alg.labels_
assert_greater(adjusted_rand_score(pred, y), 0.4)
# fit another time with ``fit_predict`` and compare results
if name is 'SpectralClustering':
# there is no way to make Spectral clustering deterministic :(
return
set_random_state(alg)
with warnings.catch_warnings(record=True):
pred2 = alg.fit_predict(X)
assert_array_equal(pred, pred2)
def test_euclidean_distances():
# Check the pairwise Euclidean distances computation
X = [[0]]
Y = [[1], [2]]
D = euclidean_distances(X, Y)
assert_array_almost_equal(D, [[1., 2.]])
X = csr_matrix(X)
Y = csr_matrix(Y)
D = euclidean_distances(X, Y)
assert_array_almost_equal(D, [[1., 2.]])
rng = np.random.RandomState(0)
X = rng.random_sample((10, 4))
Y = rng.random_sample((20, 4))
X_norm_sq = (X ** 2).sum(axis=1).reshape(1, -1)
Y_norm_sq = (Y ** 2).sum(axis=1).reshape(1, -1)
# check that we still get the right answers with {X,Y}_norm_squared
D1 = euclidean_distances(X, Y)
D2 = euclidean_distances(X, Y, X_norm_squared=X_norm_sq)
D3 = euclidean_distances(X, Y, Y_norm_squared=Y_norm_sq)
D4 = euclidean_distances(X, Y, X_norm_squared=X_norm_sq,
Y_norm_squared=Y_norm_sq)
assert_array_almost_equal(D2, D1)
assert_array_almost_equal(D3, D1)
assert_array_almost_equal(D4, D1)
# check we get the wrong answer with wrong {X,Y}_norm_squared
X_norm_sq *= 0.5
Y_norm_sq *= 0.5
wrong_D = euclidean_distances(X, Y,
X_norm_squared=np.zeros_like(X_norm_sq),
Y_norm_squared=np.zeros_like(Y_norm_sq))
assert_greater(np.max(np.abs(wrong_D - D1)), .01)
def check_regressors_train(name, Regressor):
X, y = _boston_subset()
y = StandardScaler().fit_transform(y) # X is already scaled
y = multioutput_estimator_convert_y_2d(name, y)
rnd = np.random.RandomState(0)
# catch deprecation warnings
with warnings.catch_warnings(record=True):
regressor = Regressor()
set_fast_parameters(regressor)
if not hasattr(regressor, 'alphas') and hasattr(regressor, 'alpha'):
# linear regressors need to set alpha, but not generalized CV ones
regressor.alpha = 0.01
if name == 'PassiveAggressiveRegressor':
regressor.C = 0.01
# raises error on malformed input for fit
assert_raises(ValueError, regressor.fit, X, y[:-1])
# fit
if name in CROSS_DECOMPOSITION:
y_ = np.vstack([y, 2 * y + rnd.randint(2, size=len(y))])
y_ = y_.T
else:
y_ = y
set_random_state(regressor)
regressor.fit(X, y_)
regressor.fit(X.tolist(), y_.tolist())
regressor.predict(X)
# TODO: find out why PLS and CCA fail. RANSAC is random
# and furthermore assumes the presence of outliers, hence
# skipped
if name not in ('PLSCanonical', 'CCA', 'RANSACRegressor'):
print(regressor)
assert_greater(regressor.score(X, y_), 0.5)
def check_min_samples_leaf(name):
X, y = hastie_X, hastie_y
# Test if leaves contain more than leaf_count training examples
ForestEstimator = FOREST_ESTIMATORS[name]
# test boundary value
assert_raises(ValueError,
ForestEstimator(min_samples_leaf=-1).fit, X, y)
assert_raises(ValueError,
ForestEstimator(min_samples_leaf=0).fit, X, y)
est = ForestEstimator(min_samples_leaf=5, n_estimators=1, random_state=0)
est.fit(X, y)
out = est.estimators_[0].tree_.apply(X)
node_counts = np.bincount(out)
# drop inner nodes
leaf_count = node_counts[node_counts != 0]
assert_greater(np.min(leaf_count), 4,
"Failed with {0}".format(name))
est = ForestEstimator(min_samples_leaf=0.25, n_estimators=1,
random_state=0)
est.fit(X, y)
out = est.estimators_[0].tree_.apply(X)
node_counts = np.bincount(out)
# drop inner nodes
leaf_count = node_counts[node_counts != 0]
assert_greater(np.min(leaf_count), len(X) * 0.25 - 1,
"Failed with {0}".format(name))
def check_class_weight_classifiers(name, Classifier):
if name == "NuSVC":
# the sparse version has a parameter that doesn't do anything
raise SkipTest
if name.endswith("NB"):
# NaiveBayes classifiers have a somewhat different interface.
# FIXME SOON!
raise SkipTest
for n_centers in [2, 3]:
# create a very noisy dataset
X, y = make_blobs(centers=n_centers, random_state=0, cluster_std=20)
X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=.5,
random_state=0)
n_centers = len(np.unique(y_train))
if n_centers == 2:
class_weight = {0: 1000, 1: 0.0001}
else:
class_weight = {0: 1000, 1: 0.0001, 2: 0.0001}
with warnings.catch_warnings(record=True):
classifier = Classifier(class_weight=class_weight)
if hasattr(classifier, "n_iter"):
classifier.set_params(n_iter=100)
if hasattr(classifier, "min_weight_fraction_leaf"):
classifier.set_params(min_weight_fraction_leaf=0.01)
set_random_state(classifier)
classifier.fit(X_train, y_train)
y_pred = classifier.predict(X_test)
assert_greater(np.mean(y_pred == 0), 0.89)
def test_oneclass_decision_function():
# Test OneClassSVM decision function
clf = svm.OneClassSVM()
rnd = check_random_state(2)
# Generate train data
X = 0.3 * rnd.randn(100, 2)
X_train = np.r_[X + 2, X - 2]
# Generate some regular novel observations
X = 0.3 * rnd.randn(20, 2)
X_test = np.r_[X + 2, X - 2]
# Generate some abnormal novel observations
X_outliers = rnd.uniform(low=-4, high=4, size=(20, 2))
# fit the model
clf = svm.OneClassSVM(nu=0.1, kernel="rbf", gamma=0.1)
clf.fit(X_train)
# predict things
y_pred_test = clf.predict(X_test)
assert_greater(np.mean(y_pred_test == 1), .9)
y_pred_outliers = clf.predict(X_outliers)
assert_greater(np.mean(y_pred_outliers == -1), .9)
dec_func_test = clf.decision_function(X_test)
assert_array_equal((dec_func_test > 0).ravel(), y_pred_test == 1)
dec_func_outliers = clf.decision_function(X_outliers)
assert_array_equal((dec_func_outliers > 0).ravel(), y_pred_outliers == 1)
def check_min_samples_split(name):
X, y = hastie_X, hastie_y
ForestEstimator = FOREST_ESTIMATORS[name]
# test boundary value
assert_raises(ValueError,
ForestEstimator(min_samples_split=-1).fit, X, y)
assert_raises(ValueError,
ForestEstimator(min_samples_split=0).fit, X, y)
assert_raises(ValueError,
ForestEstimator(min_samples_split=1.1).fit, X, y)
est = ForestEstimator(min_samples_split=10, n_estimators=1, random_state=0)
est.fit(X, y)
node_idx = est.estimators_[0].tree_.children_left != -1
node_samples = est.estimators_[0].tree_.n_node_samples[node_idx]
assert_greater(np.min(node_samples), len(X) * 0.5 - 1,
"Failed with {0}".format(name))
est = ForestEstimator(min_samples_split=0.5, n_estimators=1, random_state=0)
est.fit(X, y)
node_idx = est.estimators_[0].tree_.children_left != -1
node_samples = est.estimators_[0].tree_.n_node_samples[node_idx]
assert_greater(np.min(node_samples), len(X) * 0.5 - 1,
"Failed with {0}".format(name))
def test_randomized_svd_low_rank_with_noise():
"""Check that extmath.randomized_svd can handle noisy matrices"""
n_samples = 100
n_features = 500
rank = 5
k = 10
# generate a matrix X wity structure approximate rank `rank` and an
# important noisy component
X = make_low_rank_matrix(n_samples=n_samples, n_features=n_features,
effective_rank=rank, tail_strength=0.5,
random_state=0)
assert_equal(X.shape, (n_samples, n_features))
# compute the singular values of X using the slow exact method
_, s, _ = linalg.svd(X, full_matrices=False)
# compute the singular values of X using the fast approximate method
# without the iterated power method
_, sa, _ = randomized_svd(X, k, n_iter=0)
# the approximation does not tolerate the noise:
assert_greater(np.abs(s[:k] - sa).max(), 0.05)
# compute the singular values of X using the fast approximate method with
# iterated power method
_, sap, _ = randomized_svd(X, k, n_iter=5)
# the iterated power method is helping getting rid of the noise:
assert_almost_equal(s[:k], sap, decimal=3)
def test_randomized_svd_infinite_rank():
"""Check that extmath.randomized_svd can handle noisy matrices"""
n_samples = 100
n_features = 500
rank = 5
k = 10
# let us try again without 'low_rank component': just regularly but slowly
# decreasing singular values: the rank of the data matrix is infinite
X = make_low_rank_matrix(n_samples=n_samples, n_features=n_features,
effective_rank=rank, tail_strength=1.0,
random_state=0)
assert_equal(X.shape, (n_samples, n_features))
# compute the singular values of X using the slow exact method
_, s, _ = linalg.svd(X, full_matrices=False)
# compute the singular values of X using the fast approximate method
# without the iterated power method
_, sa, _ = randomized_svd(X, k, n_iter=0)
# the approximation does not tolerate the noise:
assert_greater(np.abs(s[:k] - sa).max(), 0.1)
# compute the singular values of X using the fast approximate method with
# iterated power method
_, sap, _ = randomized_svd(X, k, n_iter=5)
# the iterated power method is still managing to get most of the structure
# at the requested rank
assert_almost_equal(s[:k], sap, decimal=3)
def test_predict_iris():
# Test logistic regression with the iris dataset
n_samples, n_features = iris.data.shape
target = iris.target_names[iris.target]
# Test that both multinomial and OvR solvers handle
# multiclass data correctly and give good accuracy
# score (>0.95) for the training data.
for clf in [LogisticRegression(C=len(iris.data)),
LogisticRegression(C=len(iris.data), solver='lbfgs',
multi_class='multinomial'),
LogisticRegression(C=len(iris.data), solver='newton-cg',
multi_class='multinomial'),
LogisticRegression(C=len(iris.data), solver='sag', tol=1e-2,
multi_class='ovr', random_state=42),
LogisticRegression(C=len(iris.data), solver='saga', tol=1e-2,
multi_class='ovr', random_state=42)
]:
clf.fit(iris.data, target)
assert_array_equal(np.unique(target), clf.classes_)
pred = clf.predict(iris.data)
assert_greater(np.mean(pred == target), .95)
probabilities = clf.predict_proba(iris.data)
assert_array_almost_equal(probabilities.sum(axis=1),
np.ones(n_samples))
pred = iris.target_names[probabilities.argmax(axis=1)]
assert_greater(np.mean(pred == target), .95)
def test_spectral_amg_mode():
# Test the amg mode of SpectralClustering
centers = np.array([
[0., 0., 0.],
[10., 10., 10.],
[20., 20., 20.],
])
X, true_labels = make_blobs(n_samples=100, centers=centers,
cluster_std=1., random_state=42)
D = pairwise_distances(X) # Distance matrix
S = np.max(D) - D # Similarity matrix
S = sparse.coo_matrix(S)
try:
from pyamg import smoothed_aggregation_solver
amg_loaded = True
except ImportError:
amg_loaded = False
if amg_loaded:
labels = spectral_clustering(S, n_clusters=len(centers),
random_state=0, mode="amg")
# We don't care too much that it's good, just that it *worked*.
# There does have to be some lower limit on the performance though.
assert_greater(np.mean(labels == true_labels), .3)
else:
assert_raises(ValueError, spectral_embedding, S,
n_components=len(centers), random_state=0, mode="amg")
def test_ovo_ties():
# test that ties are broken using the decision function, not defaulting to
# the smallest label
X = np.array([[1, 2], [2, 1], [-2, 1], [-2, -1]])
y = np.array([2, 0, 1, 2])
multi_clf = OneVsOneClassifier(Perceptron())
ovo_prediction = multi_clf.fit(X, y).predict(X)
# recalculate votes to make sure we have a tie
predictions = np.vstack([clf.predict(X) for clf in multi_clf.estimators_])
scores = np.vstack([clf.decision_function(X)
for clf in multi_clf.estimators_])
# classifiers are in order 0-1, 0-2, 1-2
# aggregate votes:
votes = np.zeros((4, 3))
votes[np.arange(4), predictions[0]] += 1
votes[np.arange(4), 2 * predictions[1]] += 1
votes[np.arange(4), 1 + predictions[2]] += 1
# for the first point, there is one vote per class
assert_array_equal(votes[0, :], 1)
# for the rest, there is no tie and the prediction is the argmax
assert_array_equal(np.argmax(votes[1:], axis=1), ovo_prediction[1:])
# for the tie, the prediction is the class with the highest score
assert_equal(ovo_prediction[0], 1)
# score for one is greater than score for zero
assert_greater(scores[2, 0] - scores[0, 0], scores[0, 0] + scores[1, 0])
# score for one is greater than score for two
assert_greater(scores[2, 0] - scores[0, 0], -scores[1, 0] - scores[2, 0])
def test_warm_start(solver, warm_start, fit_intercept, multi_class):
# A 1-iteration second fit on same data should give almost same result
# with warm starting, and quite different result without warm starting.
# Warm starting does not work with liblinear solver.
X, y = iris.data, iris.target
clf = LogisticRegression(tol=1e-4, multi_class=multi_class,
warm_start=warm_start,
solver=solver,
random_state=42, max_iter=100,
fit_intercept=fit_intercept)
with ignore_warnings(category=ConvergenceWarning):
clf.fit(X, y)
coef_1 = clf.coef_
clf.max_iter = 1
clf.fit(X, y)
cum_diff = np.sum(np.abs(coef_1 - clf.coef_))
msg = ("Warm starting issue with %s solver in %s mode "
"with fit_intercept=%s and warm_start=%s"
% (solver, multi_class, str(fit_intercept),
str(warm_start)))
if warm_start:
assert_greater(2.0, cum_diff, msg)
else:
assert_greater(cum_diff, 2.0, msg)
def test_enet_path():
# We use a large number of samples and of informative features so that
# the l1_ratio selected is more toward ridge than lasso
X, y, X_test, y_test = build_dataset(n_samples=200, n_features=100, n_informative_features=100)
max_iter = 150
with warnings.catch_warnings():
# Here we have a small number of iterations, and thus the
# ElasticNet might not converge. This is to speed up tests
warnings.simplefilter("ignore", UserWarning)
clf = ElasticNetCV(n_alphas=5, eps=2e-3, l1_ratio=[0.5, 0.7], cv=3, max_iter=max_iter)
clf.fit(X, y)
# Well-conditionned settings, we should have selected our
# smallest penalty
assert_almost_equal(clf.alpha_, min(clf.alphas_))
# Non-sparse ground truth: we should have seleted an elastic-net
# that is closer to ridge than to lasso
assert_equal(clf.l1_ratio_, min(clf.l1_ratio))
clf = ElasticNetCV(n_alphas=5, eps=2e-3, l1_ratio=[0.5, 0.7], cv=3, max_iter=max_iter, precompute=True)
clf.fit(X, y)
# Well-conditionned settings, we should have selected our
# smallest penalty
assert_almost_equal(clf.alpha_, min(clf.alphas_))
# Non-sparse ground truth: we should have seleted an elastic-net
# that is closer to ridge than to lasso
assert_equal(clf.l1_ratio_, min(clf.l1_ratio))
# We are in well-conditionned settings with low noise: we should
# have a good test-set performance
assert_greater(clf.score(X_test, y_test), 0.99)
def test_rfe_estimator_tags():
rfe = RFE(SVC(kernel='linear'))
assert_equal(rfe._estimator_type, "classifier")
# make sure that cross-validation is stratified
iris = load_iris()
score = cross_val_score(rfe, iris.data, iris.target)
assert_greater(score.min(), .7)
def check_regressors_train(name, Regressor, X, y):
if name == 'OrthogonalMatchingPursuitCV':
# FIXME: This test is unstable on Travis, see issue #3190.
check_skip_travis()
rnd = np.random.RandomState(0)
# catch deprecation warnings
with warnings.catch_warnings(record=True):
regressor = Regressor()
if not hasattr(regressor, 'alphas') and hasattr(regressor, 'alpha'):
# linear regressors need to set alpha, but not generalized CV ones
regressor.alpha = 0.01
# raises error on malformed input for fit
assert_raises(ValueError, regressor.fit, X, y[:-1])
# fit
if name in ('PLSCanonical', 'PLSRegression', 'CCA'):
y_ = np.vstack([y, 2 * y + rnd.randint(2, size=len(y))])
y_ = y_.T
else:
y_ = y
set_random_state(regressor)
regressor.fit(X, y_)
regressor.predict(X)
# TODO: find out why PLS and CCA fail. RANSAC is random
# and furthermore assumes the presence of outliers, hence
# skipped
if name not in ('PLSCanonical', 'CCA', 'RANSACRegressor'):
assert_greater(regressor.score(X, y_), 0.5)
def test_k_means_function():
# test calling the k_means function directly
# catch output
old_stdout = sys.stdout
sys.stdout = StringIO()
try:
cluster_centers, labels, inertia = k_means(X, n_clusters=n_clusters,
sample_weight=None,
verbose=True)
finally:
sys.stdout = old_stdout
centers = cluster_centers
assert_equal(centers.shape, (n_clusters, n_features))
labels = labels
assert_equal(np.unique(labels).shape[0], n_clusters)
# check that the labels assignment are perfect (up to a permutation)
assert_equal(v_measure_score(true_labels, labels), 1.0)
assert_greater(inertia, 0.0)
# check warning when centers are passed
assert_warns(RuntimeWarning, k_means, X, n_clusters=n_clusters,
sample_weight=None, init=centers)
# to many clusters desired
assert_raises(ValueError, k_means, X, n_clusters=X.shape[0] + 1,
sample_weight=None)
# kmeans for algorithm='elkan' raises TypeError on sparse matrix
assert_raise_message(TypeError, "algorithm='elkan' not supported for "
"sparse input X", k_means, X=X_csr, n_clusters=2,
sample_weight=None, algorithm="elkan")
def test_lasso_cv():
X, y, X_test, y_test = build_dataset()
max_iter = 150
clf = LassoCV(n_alphas=10, eps=1e-3, max_iter=max_iter).fit(X, y)
assert_almost_equal(clf.alpha_, 0.056, 2)
clf = LassoCV(n_alphas=10, eps=1e-3, max_iter=max_iter, precompute=True)
clf.fit(X, y)
assert_almost_equal(clf.alpha_, 0.056, 2)
# Check that the lars and the coordinate descent implementation
# select a similar alpha
lars = LassoLarsCV(normalize=False, max_iter=30).fit(X, y)
# for this we check that they don't fall in the grid of
# clf.alphas further than 1
assert_true(np.abs(
np.searchsorted(clf.alphas_[::-1], lars.alpha_) -
np.searchsorted(clf.alphas_[::-1], clf.alpha_)) <= 1)
# check that they also give a similar MSE
mse_lars = interpolate.interp1d(lars.cv_alphas_, lars.cv_mse_path_.T)
np.testing.assert_approx_equal(mse_lars(clf.alphas_[5]).mean(),
clf.mse_path_[5].mean(), significant=2)
# test set
assert_greater(clf.score(X_test, y_test), 0.99)
def check_classifiers_classes(name, Classifier, X, y, y_names):
if name in ["LabelPropagation", "LabelSpreading"]:
# TODO some complication with -1 label
y_ = y
else:
y_ = y_names
classes = np.unique(y_)
# catch deprecation warnings
with warnings.catch_warnings(record=True):
classifier = Classifier()
# fit
try:
classifier.fit(X, y_)
except Exception as e:
print(e)
y_pred = classifier.predict(X)
# training set performance
assert_array_equal(np.unique(y_), np.unique(y_pred))
accuracy = accuracy_score(y_, y_pred)
assert_greater(accuracy, 0.78,
"accuracy %f of %s not greater than 0.78"
% (accuracy, name))
#assert_array_equal(
#clf.classes_, classes,
#"Unexpected classes_ attribute for %r" % clf)
if np.any(classifier.classes_ != classes):
print("Unexpected classes_ attribute for %r: "
"expected %s, got %s" %
(classifier, classes, classifier.classes_))
def test_nmf_decreasing():
# test that the objective function is decreasing at each iteration
n_samples = 20
n_features = 15
n_components = 10
alpha = 0.1
l1_ratio = 0.5
tol = 0.
# initialization
rng = np.random.mtrand.RandomState(42)
X = rng.randn(n_samples, n_features)
np.abs(X, 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):
for solver in ('cd', 'mu'):
if solver != 'mu' and beta_loss != 2:
# not implemented
continue
W, H = W0.copy(), H0.copy()
previous_loss = None
for _ in range(30):
# one more iteration starting from the previous results
W, H, _ = non_negative_factorization(
X, W, H, beta_loss=beta_loss, init='custom',
n_components=n_components, max_iter=1, alpha=alpha,
solver=solver, tol=tol, l1_ratio=l1_ratio, verbose=0,
regularization='both', random_state=0, update_H=True)
loss = nmf._beta_divergence(X, W, H, beta_loss)
if previous_loss is not None:
assert_greater(previous_loss, loss)
previous_loss = loss
def test_class_weight_classifiers():
# test that class_weight works and that the semantics are consistent
classifiers = all_estimators(type_filter="classifier")
with warnings.catch_warnings(record=True):
classifiers = [c for c in classifiers if "class_weight" in c[1]().get_params().keys()]
for n_centers in [2, 3]:
# create a very noisy dataset
X, y = make_blobs(centers=n_centers, random_state=0, cluster_std=20)
X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.5, random_state=0)
for name, Classifier in classifiers:
if name == "NuSVC":
# the sparse version has a parameter that doesn't do anything
continue
if name.endswith("NB"):
# NaiveBayes classifiers have a somewhat different interface.
# FIXME SOON!
continue
if n_centers == 2:
class_weight = {0: 1000, 1: 0.0001}
else:
class_weight = {0: 1000, 1: 0.0001, 2: 0.0001}
with warnings.catch_warnings(record=True):
classifier = Classifier(class_weight=class_weight)
if hasattr(classifier, "n_iter"):
classifier.set_params(n_iter=100)
set_random_state(classifier)
classifier.fit(X_train, y_train)
y_pred = classifier.predict(X_test)
assert_greater(np.mean(y_pred == 0), 0.9)
def test_ovo_partial_fit_predict():
X, y = shuffle(iris.data, iris.target)
ovo1 = OneVsOneClassifier(MultinomialNB())
ovo1.partial_fit(X[:100], y[:100], np.unique(y))
ovo1.partial_fit(X[100:], y[100:])
pred1 = ovo1.predict(X)
ovo2 = OneVsOneClassifier(MultinomialNB())
ovo2.fit(X, y)
pred2 = ovo2.predict(X)
assert_equal(len(ovo1.estimators_), n_classes * (n_classes - 1) / 2)
assert_greater(np.mean(y == pred1), 0.65)
assert_almost_equal(pred1, pred2)
# Test when mini-batches don't have all target classes
ovo1 = OneVsOneClassifier(MultinomialNB())
ovo1.partial_fit(iris.data[:60], iris.target[:60], np.unique(iris.target))
ovo1.partial_fit(iris.data[60:], iris.target[60:])
pred1 = ovo1.predict(iris.data)
ovo2 = OneVsOneClassifier(MultinomialNB())
pred2 = ovo2.fit(iris.data, iris.target).predict(iris.data)
assert_almost_equal(pred1, pred2)
assert_equal(len(ovo1.estimators_), len(np.unique(iris.target)))
assert_greater(np.mean(iris.target == pred1), 0.65)
def test_ovr_partial_fit():
# Test if partial_fit is working as intented
X, y = shuffle(iris.data, iris.target, random_state=0)
ovr = OneVsRestClassifier(MultinomialNB())
ovr.partial_fit(X[:100], y[:100], np.unique(y))
ovr.partial_fit(X[100:], y[100:])
pred = ovr.predict(X)
ovr2 = OneVsRestClassifier(MultinomialNB())
pred2 = ovr2.fit(X, y).predict(X)
assert_almost_equal(pred, pred2)
assert_equal(len(ovr.estimators_), len(np.unique(y)))
assert_greater(np.mean(y == pred), 0.65)
# Test when mini batches doesn't have all classes
ovr = OneVsRestClassifier(MultinomialNB())
ovr.partial_fit(iris.data[:60], iris.target[:60], np.unique(iris.target))
ovr.partial_fit(iris.data[60:], iris.target[60:])
pred = ovr.predict(iris.data)
ovr2 = OneVsRestClassifier(MultinomialNB())
pred2 = ovr2.fit(iris.data, iris.target).predict(iris.data)
assert_almost_equal(pred, pred2)
assert_equal(len(ovr.estimators_), len(np.unique(iris.target)))
assert_greater(np.mean(iris.target == pred), 0.65)
def test_classifiers_classes():
# test if classifiers can cope with non-consecutive classes
classifiers = all_estimators(type_filter='classifier')
X, y = make_blobs(random_state=12345)
X, y = shuffle(X, y, random_state=7)
X = StandardScaler().fit_transform(X)
y = 2 * y + 1
classes = np.unique(y)
# TODO: make work with next line :)
#y = y.astype(np.str)
for name, Clf in classifiers:
if Clf in dont_test:
continue
if Clf in [MultinomialNB, BernoulliNB]:
# TODO also test these!
continue
# catch deprecation warnings
with warnings.catch_warnings(record=True):
clf = Clf()
# fit
clf.fit(X, y)
y_pred = clf.predict(X)
# training set performance
assert_array_equal(np.unique(y), np.unique(y_pred))
assert_greater(zero_one_score(y, y_pred), 0.78,
"accuracy of %s not greater than 0.78" % str(Clf))
assert_array_equal(
clf.classes_, classes,
"Unexpected classes_ attribute for %r" % clf)
def test_warm_start():
# A 1-iteration second fit on same data should give almost same result
# with warm starting, and quite different result without warm starting.
# Warm starting does not work with liblinear solver.
X, y = iris.data, iris.target
solvers = ['newton-cg', 'sag']
# old scipy doesn't have maxiter
if sp_version >= (0, 12):
solvers.append('lbfgs')
for warm_start in [True, False]:
for fit_intercept in [True, False]:
for solver in solvers:
for multi_class in ['ovr', 'multinomial']:
clf = LogisticRegression(tol=1e-4, multi_class=multi_class,
warm_start=warm_start,
solver=solver,
random_state=42, max_iter=100,
fit_intercept=fit_intercept)
clf.fit(X, y)
coef_1 = clf.coef_
clf.max_iter = 1
with ignore_warnings():
clf.fit(X, y)
cum_diff = np.sum(np.abs(coef_1 - clf.coef_))
msg = ("Warm starting issue with %s solver in %s mode "
"with fit_intercept=%s and warm_start=%s"
% (solver, multi_class, str(fit_intercept),
str(warm_start)))
if warm_start:
assert_greater(2.0, cum_diff, msg)
else:
assert_greater(cum_diff, 2.0, msg)
def test_classifiers_classes():
# test if classifiers can cope with non-consecutive classes
estimators = all_estimators()
classifiers = [(name, E) for name, E in estimators if issubclass(E,
ClassifierMixin)]
iris = load_iris()
X, y = iris.data, iris.target
X, y = shuffle(X, y, random_state=7)
X = StandardScaler().fit_transform(X)
y = 2 * y + 1
# TODO: make work with next line :)
#y = y.astype(np.str)
for name, Clf in classifiers:
if Clf in dont_test or Clf in meta_estimators:
continue
if Clf in [MultinomialNB, BernoulliNB]:
# TODO also test these!
continue
# catch deprecation warnings
with warnings.catch_warnings(record=True):
clf = Clf()
# fit
clf.fit(X, y)
y_pred = clf.predict(X)
# training set performance
assert_array_equal(np.unique(y), np.unique(y_pred))
assert_greater(zero_one_score(y, y_pred), 0.78)
def test_lml_improving():
""" Test that hyperparameter-tuning improves log-marginal likelihood. """
for kernel in kernels:
if kernel == fixed_kernel: continue
gpr = GaussianProcessRegressor(kernel=kernel).fit(X, y)
assert_greater(gpr.log_marginal_likelihood(gpr.kernel_.theta),
gpr.log_marginal_likelihood(kernel.theta))
def test_fit_linear_multi():
for data in (mult_dense, mult_sparse):
clf = LinearSVC(random_state=0)
clf.fit(data, mult_target)
y_pred = clf.predict(data)
acc = np.mean(y_pred == mult_target)
assert_greater(acc, 0.85)
def test_enet_path():
# We use a large number of samples and of informative features so that
# the l1_ratio selected is more toward ridge than lasso
X, y, X_test, y_test = build_dataset(n_samples=200,
n_features=100,
n_informative_features=100)
max_iter = 150
# Here we have a small number of iterations, and thus the
# ElasticNet might not converge. This is to speed up tests
clf = ElasticNetCV(alphas=[0.01, 0.05, 0.1],
eps=2e-3,
l1_ratio=[0.5, 0.7],
cv=3,
max_iter=max_iter)
ignore_warnings(clf.fit)(X, y)
# Well-conditioned settings, we should have selected our
# smallest penalty
assert_almost_equal(clf.alpha_, min(clf.alphas_))
# Non-sparse ground truth: we should have seleted an elastic-net
# that is closer to ridge than to lasso
assert_equal(clf.l1_ratio_, min(clf.l1_ratio))
clf = ElasticNetCV(alphas=[0.01, 0.05, 0.1],
eps=2e-3,
l1_ratio=[0.5, 0.7],
cv=3,
max_iter=max_iter,
precompute=True)
ignore_warnings(clf.fit)(X, y)
# Well-conditioned settings, we should have selected our
# smallest penalty
assert_almost_equal(clf.alpha_, min(clf.alphas_))
# Non-sparse ground truth: we should have seleted an elastic-net
# that is closer to ridge than to lasso
assert_equal(clf.l1_ratio_, min(clf.l1_ratio))
# We are in well-conditioned settings with low noise: we should
# have a good test-set performance
assert_greater(clf.score(X_test, y_test), 0.99)
# Multi-output/target case
X, y, X_test, y_test = build_dataset(n_features=10, n_targets=3)
clf = MultiTaskElasticNetCV(n_alphas=5,
eps=2e-3,
l1_ratio=[0.5, 0.7],
cv=3,
max_iter=max_iter)
ignore_warnings(clf.fit)(X, y)
# We are in well-conditioned settings with low noise: we should
# have a good test-set performance
assert_greater(clf.score(X_test, y_test), 0.99)
assert_equal(clf.coef_.shape, (3, 10))
# Mono-output should have same cross-validated alpha_ and l1_ratio_
# in both cases.
X, y, _, _ = build_dataset(n_features=10)
clf1 = ElasticNetCV(n_alphas=5, eps=2e-3, l1_ratio=[0.5, 0.7])
clf1.fit(X, y)
clf2 = MultiTaskElasticNetCV(n_alphas=5, eps=2e-3, l1_ratio=[0.5, 0.7])
clf2.fit(X, y[:, np.newaxis])
assert_almost_equal(clf1.l1_ratio_, clf2.l1_ratio_)
assert_almost_equal(clf1.alpha_, clf2.alpha_)
def test_kfold_can_detect_dependent_samples_on_digits(): # see #2372
# The digits samples are dependent: they are apparently grouped by authors
# although we don't have any information on the groups segment locations
# for this data. We can highlight this fact be computing k-fold cross-
# validation with and without shuffling: we observe that the shuffling case
# wrongly makes the IID assumption and is therefore too optimistic: it
# estimates a much higher accuracy (around 0.96) than the non
# shuffling variant (around 0.86).
digits = load_digits()
X, y = digits.data[:800], digits.target[:800]
model = SVC(C=10, gamma=0.005)
n = len(y)
cv = cval.KFold(n, 5, shuffle=False)
mean_score = cval.cross_val_score(model, X, y, cv=cv).mean()
assert_greater(0.88, mean_score)
assert_greater(mean_score, 0.85)
# Shuffling the data artificially breaks the dependency and hides the
# overfitting of the model with regards to the writing style of the authors
# by yielding a seriously overestimated score:
cv = cval.KFold(n, 5, shuffle=True, random_state=0)
mean_score = cval.cross_val_score(model, X, y, cv=cv).mean()
assert_greater(mean_score, 0.95)
cv = cval.KFold(n, 5, shuffle=True, random_state=1)
mean_score = cval.cross_val_score(model, X, y, cv=cv).mean()
assert_greater(mean_score, 0.95)
# Similarly, StratifiedKFold should try to shuffle the data as little
# as possible (while respecting the balanced class constraints)
# and thus be able to detect the dependency by not overestimating
# the CV score either. As the digits dataset is approximately balanced
# the estimated mean score is close to the score measured with
# non-shuffled KFold
cv = cval.StratifiedKFold(y, 5)
mean_score = cval.cross_val_score(model, X, y, cv=cv).mean()
assert_greater(0.88, mean_score)
assert_greater(mean_score, 0.85)
def test_ovr_fit_predict_svc():
ovr = OneVsRestClassifier(svm.SVC())
ovr.fit(iris.data, iris.target)
assert_equal(len(ovr.estimators_), 3)
assert_greater(ovr.score(iris.data, iris.target), .9)
def test_permutation_score():
iris = load_iris()
X = iris.data
X_sparse = coo_matrix(X)
y = iris.target
svm = SVC(kernel='linear')
cv = StratifiedKFold(2)
score, scores, pvalue = permutation_test_score(svm,
X,
y,
n_permutations=30,
cv=cv,
scoring="accuracy")
assert_greater(score, 0.9)
assert_almost_equal(pvalue, 0.0, 1)
score_group, _, pvalue_group = permutation_test_score(svm,
X,
y,
n_permutations=30,
cv=cv,
scoring="accuracy",
groups=np.ones(
y.size),
random_state=0)
assert_true(score_group == score)
assert_true(pvalue_group == pvalue)
# check that we obtain the same results with a sparse representation
svm_sparse = SVC(kernel='linear')
cv_sparse = StratifiedKFold(2)
score_group, _, pvalue_group = permutation_test_score(svm_sparse,
X_sparse,
y,
n_permutations=30,
cv=cv_sparse,
scoring="accuracy",
groups=np.ones(
y.size),
random_state=0)
assert_true(score_group == score)
assert_true(pvalue_group == pvalue)
# test with custom scoring object
def custom_score(y_true, y_pred):
return (((y_true == y_pred).sum() - (y_true != y_pred).sum()) /
y_true.shape[0])
scorer = make_scorer(custom_score)
score, _, pvalue = permutation_test_score(svm,
X,
y,
n_permutations=100,
scoring=scorer,
cv=cv,
random_state=0)
assert_almost_equal(score, .93, 2)
assert_almost_equal(pvalue, 0.01, 3)
# set random y
y = np.mod(np.arange(len(y)), 3)
score, scores, pvalue = permutation_test_score(svm,
X,
y,
n_permutations=30,
cv=cv,
scoring="accuracy")
assert_less(score, 0.5)
assert_greater(pvalue, 0.2)
def test_incremental_variance_numerical_stability():
# Test Youngs and Cramer incremental variance formulas.
def np_var(A):
return A.var(axis=0)
# Naive one pass variance computation - not numerically stable
# https://en.wikipedia.org/wiki/Algorithms_for_calculating_variance
def one_pass_var(X):
n = X.shape[0]
exp_x2 = (X ** 2).sum(axis=0) / n
expx_2 = (X.sum(axis=0) / n) ** 2
return exp_x2 - expx_2
# Two-pass algorithm, stable.
# We use it as a benchmark. It is not an online algorithm
# https://en.wikipedia.org/wiki/Algorithms_for_calculating_variance#Two-pass_algorithm
def two_pass_var(X):
mean = X.mean(axis=0)
Y = X.copy()
return np.mean((Y - mean)**2, axis=0)
# Naive online implementation
# https://en.wikipedia.org/wiki/Algorithms_for_calculating_variance#Online_algorithm
# This works only for chunks for size 1
def naive_mean_variance_update(x, last_mean, last_variance,
last_sample_count):
updated_sample_count = (last_sample_count + 1)
samples_ratio = last_sample_count / float(updated_sample_count)
updated_mean = x / updated_sample_count + last_mean * samples_ratio
updated_variance = last_variance * samples_ratio + \
(x - last_mean) * (x - updated_mean) / updated_sample_count
return updated_mean, updated_variance, updated_sample_count
# We want to show a case when one_pass_var has error > 1e-3 while
# _batch_mean_variance_update has less.
tol = 200
n_features = 2
n_samples = 10000
x1 = np.array(1e8, dtype=np.float64)
x2 = np.log(1e-5, dtype=np.float64)
A0 = np.full((n_samples // 2, n_features), x1, dtype=np.float64)
A1 = np.full((n_samples // 2, n_features), x2, dtype=np.float64)
A = np.vstack((A0, A1))
# Older versions of numpy have different precision
# In some old version, np.var is not stable
if np.abs(np_var(A) - two_pass_var(A)).max() < 1e-6:
stable_var = np_var
else:
stable_var = two_pass_var
# Naive one pass var: >tol (=1063)
assert_greater(np.abs(stable_var(A) - one_pass_var(A)).max(), tol)
# Starting point for online algorithms: after A0
# Naive implementation: >tol (436)
mean, var, n = A0[0, :], np.zeros(n_features), n_samples // 2
for i in range(A1.shape[0]):
mean, var, n = \
naive_mean_variance_update(A1[i, :], mean, var, n)
assert_equal(n, A.shape[0])
# the mean is also slightly unstable
assert_greater(np.abs(A.mean(axis=0) - mean).max(), 1e-6)
assert_greater(np.abs(stable_var(A) - var).max(), tol)
# Robust implementation: <tol (177)
mean, var = A0[0, :], np.zeros(n_features)
n = np.full(n_features, n_samples // 2, dtype=np.int32)
for i in range(A1.shape[0]):
mean, var, n = \
_incremental_mean_and_var(A1[i, :].reshape((1, A1.shape[1])),
mean, var, n)
assert_array_equal(n, A.shape[0])
assert_array_almost_equal(A.mean(axis=0), mean)
assert_greater(tol, np.abs(stable_var(A) - var).max())
def test_minibatch_update_consistency():
"""Check that dense and sparse minibatch update give the same results"""
rng = np.random.RandomState(42)
old_centers = centers + rng.normal(size=centers.shape)
new_centers = old_centers.copy()
new_centers_csr = old_centers.copy()
counts = np.zeros(new_centers.shape[0], dtype=np.int32)
counts_csr = np.zeros(new_centers.shape[0], dtype=np.int32)
x_squared_norms = (X**2).sum(axis=1)
x_squared_norms_csr = csr_row_norm_l2(X_csr, squared=True)
buffer = np.zeros(centers.shape[1], dtype=np.double)
buffer_csr = np.zeros(centers.shape[1], dtype=np.double)
# extract a small minibatch
X_mb = X[:10]
X_mb_csr = X_csr[:10]
x_mb_squared_norms = x_squared_norms[:10]
x_mb_squared_norms_csr = x_squared_norms_csr[:10]
# step 1: compute the dense minibatch update
old_inertia, incremental_diff = _mini_batch_step(X_mb, x_mb_squared_norms,
new_centers, counts,
buffer, 1)
assert_greater(old_inertia, 0.0)
# compute the new inertia on the same batch to check that it decreased
labels, new_inertia = _labels_inertia(X_mb, x_mb_squared_norms,
new_centers)
assert_greater(new_inertia, 0.0)
assert_less(new_inertia, old_inertia)
# check that the incremental difference computation is matching the
# final observed value
effective_diff = np.sum((new_centers - old_centers)**2)
assert_almost_equal(incremental_diff, effective_diff)
# step 2: compute the sparse minibatch update
old_inertia_csr, incremental_diff_csr = _mini_batch_step(
X_mb_csr, x_mb_squared_norms_csr, new_centers_csr, counts_csr,
buffer_csr, 1)
assert_greater(old_inertia_csr, 0.0)
# compute the new inertia on the same batch to check that it decreased
labels_csr, new_inertia_csr = _labels_inertia(X_mb_csr,
x_mb_squared_norms_csr,
new_centers_csr)
assert_greater(new_inertia_csr, 0.0)
assert_less(new_inertia_csr, old_inertia_csr)
# check that the incremental difference computation is matching the
# final observed value
effective_diff = np.sum((new_centers_csr - old_centers)**2)
assert_almost_equal(incremental_diff_csr, effective_diff)
# step 3: check that sparse and dense updates lead to the same results
assert_array_equal(labels, labels_csr)
assert_array_almost_equal(new_centers, new_centers_csr)
assert_almost_equal(incremental_diff, incremental_diff_csr)
assert_almost_equal(old_inertia, old_inertia_csr)
assert_almost_equal(new_inertia, new_inertia_csr)
def test_score():
km1 = KMeans(n_clusters=n_clusters, max_iter=1, random_state=42)
s1 = km1.fit(X).score(X)
km2 = KMeans(n_clusters=n_clusters, max_iter=10, random_state=42)
s2 = km2.fit(X).score(X)
assert_greater(s2, s1)
def test_assert_greater():
assert_greater(1, 0)
assert_raises(AssertionError, assert_greater, 0, 1)
def test_binnedstratifiedkfold_has_more_stable_distribution_moments_between_folds(
):
"""check if BinnedStratifiedKFold performs on average better than KFold in terms of
lower between-fold variance of fold mean(y_test) and fold std(y_test)
"""
binned_has_more_stable_std_list = []
binned_has_more_stable_mean_list = []
for trial in range(100):
n_folds = 2 + int(10 * np.random.rand())
y = np.random.randn(30)
np.random.shuffle(y)
ymeans_binned = []
ystds_binned = []
cv_bs = BinnedStratifiedKFold(n_folds=n_folds,
shuffle=False,
random_state=None)
bskf = cv_bs.split(y)
cv = KFold(n_folds=n_folds, shuffle=True, random_state=None)
kf = cv.split(y)
#bins = np.percentile(y, np.arange(n_folds))
bins = np.array([np.percentile(y, q) for q in range(n_folds)])
for train_index, test_index in bskf:
y_test = y[test_index]
ymeans_binned.append(y_test.mean())
ystds_binned.append(y_test.std())
hist_, _ = np.histogram(y[test_index], bins=bins)
assert_true(all(abs(hist_ - np.mean(hist_)) <= 1),
msg="too ragged bins")
ymeans_regular = []
ystds_regular = []
for train_index_reg, test_index_reg in kf:
ymeans_regular.append(y[test_index_reg].mean())
ystds_regular.append(y[test_index_reg].std())
binned_has_more_stable_std = np.std(ystds_regular) > np.std(
ystds_binned)
binned_has_more_stable_std_list.append(binned_has_more_stable_std)
binned_has_more_stable_mean = np.std(ymeans_regular) > np.std(
ymeans_binned)
binned_has_more_stable_mean_list.append(binned_has_more_stable_mean)
binned_has_more_stable_std_fraction = np.mean(
binned_has_more_stable_std_list)
binned_has_more_stable_mean_fraction = np.mean(
binned_has_more_stable_mean_list)
assert_greater(binned_has_more_stable_std_fraction, 0.5)
assert_greater(binned_has_more_stable_mean_fraction, 0.5)
print(" std(y_test) of BinnedStratifiedKFold was more stable than "
"one of KFold in\t%.2f%% cases" % \
(100.0*binned_has_more_stable_std_fraction))
print("mean(y_test) of BinnedStratifiedKFold was more stable than "
"one of KFold in\t%.2f%% cases" % \
(100.0*binned_has_more_stable_mean_fraction))
def test_permutation_score():
iris = load_iris()
X = iris.data
X_sparse = coo_matrix(X)
y = iris.target
svm = SVC(kernel='linear')
cv = cval.StratifiedKFold(y, 2)
score, scores, pvalue = cval.permutation_test_score(svm,
X,
y,
cv=cv,
scoring="accuracy")
assert_greater(score, 0.9)
assert_almost_equal(pvalue, 0.0, 1)
score_label, _, pvalue_label = cval.permutation_test_score(
svm,
X,
y,
cv=cv,
scoring="accuracy",
labels=np.ones(y.size),
random_state=0)
assert_true(score_label == score)
assert_true(pvalue_label == pvalue)
# test with custom scoring object
scorer = make_scorer(fbeta_score, beta=2)
score_label, _, pvalue_label = cval.permutation_test_score(svm,
X,
y,
scoring=scorer,
cv=cv,
labels=np.ones(
y.size),
random_state=0)
assert_almost_equal(score_label, .97, 2)
assert_almost_equal(pvalue_label, 0.01, 3)
# check that we obtain the same results with a sparse representation
svm_sparse = SVC(kernel='linear')
cv_sparse = cval.StratifiedKFold(y, 2)
score_label, _, pvalue_label = cval.permutation_test_score(
svm_sparse,
X_sparse,
y,
cv=cv_sparse,
scoring="accuracy",
labels=np.ones(y.size),
random_state=0)
assert_true(score_label == score)
assert_true(pvalue_label == pvalue)
# set random y
y = np.mod(np.arange(len(y)), 3)
score, scores, pvalue = cval.permutation_test_score(svm,
X,
y,
cv=cv,
scoring="accuracy")
assert_less(score, 0.5)
assert_greater(pvalue, 0.2)
# test with deprecated interface
with warnings.catch_warnings(record=True):
score, scores, pvalue = cval.permutation_test_score(
svm, X, y, score_func=accuracy_score, cv=cv)
assert_less(score, 0.5)
assert_greater(pvalue, 0.2)
def test_all_estimators():
estimators = all_estimators(include_meta_estimators=True)
# Meta sanity-check to make sure that the estimator introspection runs
# properly
assert_greater(len(estimators), 0)
def test_radius_neighbors():
# Checks whether Returned distances are less than `radius`
# At least one point should be returned when the `radius` is set
# to mean distance from the considering point to other points in
# the database.
# Moreover, this test compares the radius neighbors of LSHForest
# with the `sklearn.neighbors.NearestNeighbors`.
n_samples = 12
n_features = 2
n_iter = 10
rng = np.random.RandomState(42)
X = rng.rand(n_samples, n_features)
lshf = ignore_warnings(LSHForest, category=DeprecationWarning)()
# Test unfitted estimator
assert_raises(ValueError, lshf.radius_neighbors, X[0])
ignore_warnings(lshf.fit)(X)
for i in range(n_iter):
# Select a random point in the dataset as the query
query = X[rng.randint(0, n_samples)].reshape(1, -1)
# At least one neighbor should be returned when the radius is the
# mean distance from the query to the points of the dataset.
mean_dist = np.mean(pairwise_distances(query, X, metric='cosine'))
neighbors = lshf.radius_neighbors(query, radius=mean_dist,
return_distance=False)
assert_equal(neighbors.shape, (1,))
assert_equal(neighbors.dtype, object)
assert_greater(neighbors[0].shape[0], 0)
# All distances to points in the results of the radius query should
# be less than mean_dist
distances, neighbors = lshf.radius_neighbors(query,
radius=mean_dist,
return_distance=True)
assert_array_less(distances[0], mean_dist)
# Multiple points
n_queries = 5
queries = X[rng.randint(0, n_samples, n_queries)]
distances, neighbors = lshf.radius_neighbors(queries,
return_distance=True)
# dists and inds should not be 1D arrays or arrays of variable lengths
# hence the use of the object dtype.
assert_equal(distances.shape, (n_queries,))
assert_equal(distances.dtype, object)
assert_equal(neighbors.shape, (n_queries,))
assert_equal(neighbors.dtype, object)
# Compare with exact neighbor search
query = X[rng.randint(0, n_samples)].reshape(1, -1)
mean_dist = np.mean(pairwise_distances(query, X, metric='cosine'))
nbrs = NearestNeighbors(algorithm='brute', metric='cosine').fit(X)
distances_exact, _ = nbrs.radius_neighbors(query, radius=mean_dist)
distances_approx, _ = lshf.radius_neighbors(query, radius=mean_dist)
# Radius-based queries do not sort the result points and the order
# depends on the method, the random_state and the dataset order. Therefore
# we need to sort the results ourselves before performing any comparison.
sorted_dists_exact = np.sort(distances_exact[0])
sorted_dists_approx = np.sort(distances_approx[0])
# Distances to exact neighbors are less than or equal to approximate
# counterparts as the approximate radius query might have missed some
# closer neighbors.
assert_true(np.all(np.less_equal(sorted_dists_exact,
sorted_dists_approx)))
def check_classifiers_train(name, Classifier):
X_m, y_m = make_blobs(random_state=0)
X_m, y_m = shuffle(X_m, y_m, random_state=7)
X_m = StandardScaler().fit_transform(X_m)
# generate binary problem from multi-class one
y_b = y_m[y_m != 2]
X_b = X_m[y_m != 2]
for (X, y) in [(X_m, y_m), (X_b, y_b)]:
# catch deprecation warnings
classes = np.unique(y)
n_classes = len(classes)
n_samples, n_features = X.shape
with warnings.catch_warnings(record=True):
classifier = Classifier()
if name in ['BernoulliNB', 'MultinomialNB']:
X -= X.min()
set_fast_parameters(classifier)
# raises error on malformed input for fit
assert_raises(ValueError, classifier.fit, X, y[:-1])
# fit
classifier.fit(X, y)
# with lists
classifier.fit(X.tolist(), y.tolist())
assert_true(hasattr(classifier, "classes_"))
y_pred = classifier.predict(X)
assert_equal(y_pred.shape, (n_samples,))
# training set performance
if name not in ['BernoulliNB', 'MultinomialNB']:
assert_greater(accuracy_score(y, y_pred), 0.85)
# raises error on malformed input for predict
assert_raises(ValueError, classifier.predict, X.T)
if hasattr(classifier, "decision_function"):
try:
# decision_function agrees with predict:
decision = classifier.decision_function(X)
if n_classes is 2:
assert_equal(decision.shape, (n_samples,))
dec_pred = (decision.ravel() > 0).astype(np.int)
assert_array_equal(dec_pred, y_pred)
if (n_classes is 3
and not isinstance(classifier, BaseLibSVM)):
# 1on1 of LibSVM works differently
assert_equal(decision.shape, (n_samples, n_classes))
assert_array_equal(np.argmax(decision, axis=1), y_pred)
# raises error on malformed input
assert_raises(ValueError,
classifier.decision_function, X.T)
# raises error on malformed input for decision_function
assert_raises(ValueError,
classifier.decision_function, X.T)
except NotImplementedError:
pass
if hasattr(classifier, "predict_proba"):
# predict_proba agrees with predict:
y_prob = classifier.predict_proba(X)
assert_equal(y_prob.shape, (n_samples, n_classes))
assert_array_equal(np.argmax(y_prob, axis=1), y_pred)
# check that probas for all classes sum to one
assert_array_almost_equal(np.sum(y_prob, axis=1),
np.ones(n_samples))
# raises error on malformed input
assert_raises(ValueError, classifier.predict_proba, X.T)
# raises error on malformed input for predict_proba
assert_raises(ValueError, classifier.predict_proba, X.T)
def test_classifiers_train():
# test if classifiers do something sensible on training set
# also test all shapes / shape errors
estimators = all_estimators()
classifiers = [(name, E) for name, E in estimators
if issubclass(E, ClassifierMixin)]
iris = load_iris()
X_m, y_m = iris.data, iris.target
X_m, y_m = shuffle(X_m, y_m, random_state=7)
X_m = Scaler().fit_transform(X_m)
# generate binary problem from multi-class one
y_b = y_m[y_m != 2]
X_b = X_m[y_m != 2]
for (X, y) in [(X_m, y_m), (X_b, y_b)]:
# do it once with binary, once with multiclass
n_labels = len(np.unique(y))
n_samples, n_features = X.shape
for name, Clf in classifiers:
if Clf in dont_test or Clf in meta_estimators:
continue
if Clf in [MultinomialNB, BernoulliNB]:
# TODO also test these!
continue
# catch deprecation warnings
with warnings.catch_warnings(record=True):
clf = Clf()
# raises error on malformed input for fit
assert_raises(ValueError, clf.fit, X, y[:-1])
# fit
clf.fit(X, y)
y_pred = clf.predict(X)
assert_equal(y_pred.shape, (n_samples, ))
# training set performance
assert_greater(zero_one_score(y, y_pred), 0.78)
# raises error on malformed input for predict
assert_raises(ValueError, clf.predict, X.T)
if hasattr(clf, "decision_function"):
try:
# decision_function agrees with predict:
decision = clf.decision_function(X)
if n_labels is 2:
assert_equal(decision.ravel().shape, (n_samples, ))
dec_pred = (decision.ravel() > 0).astype(np.int)
assert_array_equal(dec_pred, y_pred)
if n_labels is 3 and not isinstance(clf, BaseLibSVM):
# 1on1 of LibSVM works differently
assert_equal(decision.shape, (n_samples, n_labels))
assert_array_equal(np.argmax(decision, axis=1), y_pred)
# raises error on malformed input
assert_raises(ValueError, clf.decision_function, X.T)
# raises error on malformed input for decision_function
assert_raises(ValueError, clf.decision_function, X.T)
except NotImplementedError:
pass
if hasattr(clf, "predict_proba"):
try:
# predict_proba agrees with predict:
y_prob = clf.predict_proba(X)
assert_equal(y_prob.shape, (n_samples, n_labels))
# raises error on malformed input
assert_raises(ValueError, clf.predict_proba, X.T)
assert_array_equal(np.argmax(y_prob, axis=1), y_pred)
# raises error on malformed input for predict_proba
assert_raises(ValueError, clf.predict_proba, X.T)
except NotImplementedError:
pass
def test_factor_analysis():
# Test FactorAnalysis ability to recover the data covariance structure
rng = np.random.RandomState(0)
n_samples, n_features, n_components = 20, 5, 3
# Some random settings for the generative model
W = rng.randn(n_components, n_features)
# latent variable of dim 3, 20 of it
h = rng.randn(n_samples, n_components)
# using gamma to model different noise variance
# per component
noise = rng.gamma(1, size=n_features) * rng.randn(n_samples, n_features)
# generate observations
# wlog, mean is 0
X = np.dot(h, W) + noise
assert_raises(ValueError, FactorAnalysis, svd_method='foo')
fa_fail = FactorAnalysis()
fa_fail.svd_method = 'foo'
assert_raises(ValueError, fa_fail.fit, X)
fas = []
for method in ['randomized', 'lapack']:
fa = FactorAnalysis(n_components=n_components, svd_method=method)
fa.fit(X)
fas.append(fa)
X_t = fa.transform(X)
assert_equal(X_t.shape, (n_samples, n_components))
assert_almost_equal(fa.loglike_[-1], fa.score_samples(X).sum())
assert_almost_equal(fa.score_samples(X).mean(), fa.score(X))
diff = np.all(np.diff(fa.loglike_))
assert_greater(diff, 0., 'Log likelihood dif not increase')
# Sample Covariance
scov = np.cov(X, rowvar=0., bias=1.)
# Model Covariance
mcov = fa.get_covariance()
diff = np.sum(np.abs(scov - mcov)) / W.size
assert_less(diff, 0.1, "Mean absolute difference is %f" % diff)
fa = FactorAnalysis(n_components=n_components,
noise_variance_init=np.ones(n_features))
assert_raises(ValueError, fa.fit, X[:, :2])
f = lambda x, y: np.abs(getattr(x, y)) # sign will not be equal
fa1, fa2 = fas
for attr in ['loglike_', 'components_', 'noise_variance_']:
assert_almost_equal(f(fa1, attr), f(fa2, attr))
fa1.max_iter = 1
fa1.verbose = True
assert_warns(ConvergenceWarning, fa1.fit, X)
# Test get_covariance and get_precision with n_components == n_features
# with n_components < n_features and with n_components == 0
for n_components in [0, 2, X.shape[1]]:
fa.n_components = n_components
fa.fit(X)
cov = fa.get_covariance()
precision = fa.get_precision()
assert_array_almost_equal(np.dot(cov, precision), np.eye(X.shape[1]),
12)
def test_calibration():
"""Test calibration objects with isotonic and sigmoid"""
n_samples = 100
X, y = make_classification(n_samples=2 * n_samples,
n_features=6,
random_state=42)
sample_weight = np.random.RandomState(seed=42).uniform(size=y.size)
X -= X.min() # MultinomialNB only allows positive X
# split train and test
X_train, y_train, sw_train = \
X[:n_samples], y[:n_samples], sample_weight[:n_samples]
X_test, y_test = X[n_samples:], y[n_samples:]
# Naive-Bayes
clf = MultinomialNB().fit(X_train, y_train, sample_weight=sw_train)
prob_pos_clf = clf.predict_proba(X_test)[:, 1]
pc_clf = CalibratedClassifierCV(clf, cv=y.size + 1)
assert_raises(ValueError, pc_clf.fit, X, y)
# Naive Bayes with calibration
for this_X_train, this_X_test in [(X_train, X_test),
(sparse.csr_matrix(X_train),
sparse.csr_matrix(X_test))]:
for method in ['isotonic', 'sigmoid']:
pc_clf = CalibratedClassifierCV(clf, method=method, cv=2)
# Note that this fit overwrites the fit on the entire training
# set
pc_clf.fit(this_X_train, y_train, sample_weight=sw_train)
prob_pos_pc_clf = pc_clf.predict_proba(this_X_test)[:, 1]
# Check that brier score has improved after calibration
assert_greater(brier_score_loss(y_test, prob_pos_clf),
brier_score_loss(y_test, prob_pos_pc_clf))
# Check invariance against relabeling [0, 1] -> [1, 2]
pc_clf.fit(this_X_train, y_train + 1, sample_weight=sw_train)
prob_pos_pc_clf_relabeled = pc_clf.predict_proba(this_X_test)[:, 1]
assert_array_almost_equal(prob_pos_pc_clf,
prob_pos_pc_clf_relabeled)
# Check invariance against relabeling [0, 1] -> [-1, 1]
pc_clf.fit(this_X_train, 2 * y_train - 1, sample_weight=sw_train)
prob_pos_pc_clf_relabeled = pc_clf.predict_proba(this_X_test)[:, 1]
assert_array_almost_equal(prob_pos_pc_clf,
prob_pos_pc_clf_relabeled)
# Check invariance against relabeling [0, 1] -> [1, 0]
pc_clf.fit(this_X_train, (y_train + 1) % 2, sample_weight=sw_train)
prob_pos_pc_clf_relabeled = \
pc_clf.predict_proba(this_X_test)[:, 1]
if method == "sigmoid":
assert_array_almost_equal(prob_pos_pc_clf,
1 - prob_pos_pc_clf_relabeled)
else:
# Isotonic calibration is not invariant against relabeling
# but should improve in both cases
assert_greater(
brier_score_loss(y_test, prob_pos_clf),
brier_score_loss((y_test + 1) % 2,
prob_pos_pc_clf_relabeled))
# check that calibration can also deal with regressors that have
# a decision_function
clf_base_regressor = CalibratedClassifierCV(Ridge())
clf_base_regressor.fit(X_train, y_train)
clf_base_regressor.predict(X_test)
# Check failure cases:
# only "isotonic" and "sigmoid" should be accepted as methods
clf_invalid_method = CalibratedClassifierCV(clf, method="foo")
assert_raises(ValueError, clf_invalid_method.fit, X_train, y_train)
# base-estimators should provide either decision_function or
# predict_proba (most regressors, for instance, should fail)
clf_base_regressor = \
CalibratedClassifierCV(RandomForestRegressor(), method="sigmoid")
assert_raises(RuntimeError, clf_base_regressor.fit, X_train, y_train)
def test_check_accuracy_on_digits():
# Non regression test to make sure that any further refactoring / optim
# of the NB models do not harm the performance on a slightly non-linearly
# separable dataset
digits = load_digits()
X, y = digits.data, digits.target
binary_3v8 = np.logical_or(digits.target == 3, digits.target == 8)
X_3v8, y_3v8 = X[binary_3v8], y[binary_3v8]
# Multinomial NB
scores = cross_val_score(MultinomialNB(alpha=10), X, y, cv=10)
assert_greater(scores.mean(), 0.86)
scores = cross_val_score(MultinomialNB(alpha=10), X_3v8, y_3v8, cv=10)
assert_greater(scores.mean(), 0.94)
# Bernoulli NB
scores = cross_val_score(BernoulliNB(alpha=10), X > 4, y, cv=10)
assert_greater(scores.mean(), 0.83)
scores = cross_val_score(BernoulliNB(alpha=10), X_3v8 > 4, y_3v8, cv=10)
assert_greater(scores.mean(), 0.92)
# Gaussian NB
scores = cross_val_score(GaussianNB(), X, y, cv=10)
assert_greater(scores.mean(), 0.77)
scores = cross_val_score(GaussianNB(), X_3v8, y_3v8, cv=10)
assert_greater(scores.mean(), 0.86)
print("Timimg exact rbf: \t\t", exact_spent_time)
rbf_transform = Fastfood(
sigma=sigma,
n_components=number_of_features_to_generate,
tradeoff_mem_accuracy="mem",
random_state=42,
)
_ = rbf_transform.fit(X)
fastfood_fast_vec_start = datetime.datetime.utcnow()
# Fastfood: approximate kernel mapping
_ = rbf_transform.transform(X)
_ = rbf_transform.transform(Y)
fastfood_fast_vec_end = datetime.datetime.utcnow()
fastfood_fast_vec_spent_time = fastfood_fast_vec_end - fastfood_fast_vec_start
print("Timimg fastfood fast vectorized: \t\t", fastfood_fast_vec_spent_time)
rks_rbf_transform = RBFSampler(
gamma=gamma, n_components=number_of_features_to_generate, random_state=42
)
_ = rks_rbf_transform.fit(X)
rks_start = datetime.datetime.utcnow()
# Random Kitchens Sinks: approximate kernel mapping
_ = rks_rbf_transform.transform(X)
_ = rks_rbf_transform.transform(Y)
rks_end = datetime.datetime.utcnow()
rks_spent_time = rks_end - rks_start
print("Timimg rks: \t\t\t", rks_spent_time)
assert_greater(rks_spent_time, fastfood_fast_vec_spent_time)
def test_classifiers_train():
# test if classifiers do something sensible on training set
# also test all shapes / shape errors
classifiers = all_estimators(type_filter='classifier')
X_m, y_m = make_blobs(random_state=0)
X_m, y_m = shuffle(X_m, y_m, random_state=7)
X_m = StandardScaler().fit_transform(X_m)
# generate binary problem from multi-class one
y_b = y_m[y_m != 2]
X_b = X_m[y_m != 2]
for (X, y) in [(X_m, y_m), (X_b, y_b)]:
# do it once with binary, once with multiclass
classes = np.unique(y)
n_classes = len(classes)
n_samples, n_features = X.shape
for name, Classifier in classifiers:
if name in dont_test:
continue
if name in ['MultinomialNB', 'BernoulliNB']:
# TODO also test these!
continue
# catch deprecation warnings
with warnings.catch_warnings(record=True):
classifier = Classifier()
# raises error on malformed input for fit
assert_raises(ValueError, classifier.fit, X, y[:-1])
# fit
classifier.fit(X, y)
assert_true(hasattr(classifier, "classes_"))
y_pred = classifier.predict(X)
assert_equal(y_pred.shape, (n_samples, ))
# training set performance
assert_greater(accuracy_score(y, y_pred), 0.85)
# raises error on malformed input for predict
assert_raises(ValueError, classifier.predict, X.T)
if hasattr(classifier, "decision_function"):
try:
# decision_function agrees with predict:
decision = classifier.decision_function(X)
if n_classes is 2:
assert_equal(decision.ravel().shape, (n_samples, ))
dec_pred = (decision.ravel() > 0).astype(np.int)
assert_array_equal(dec_pred, y_pred)
if (n_classes is 3
and not isinstance(classifier, BaseLibSVM)):
# 1on1 of LibSVM works differently
assert_equal(decision.shape, (n_samples, n_classes))
assert_array_equal(np.argmax(decision, axis=1), y_pred)
# raises error on malformed input
assert_raises(ValueError, classifier.decision_function,
X.T)
# raises error on malformed input for decision_function
assert_raises(ValueError, classifier.decision_function,
X.T)
except NotImplementedError:
pass
if hasattr(classifier, "predict_proba"):
try:
# predict_proba agrees with predict:
y_prob = classifier.predict_proba(X)
assert_equal(y_prob.shape, (n_samples, n_classes))
assert_array_equal(np.argmax(y_prob, axis=1), y_pred)
# check that probas for all classes sum to one
assert_array_almost_equal(np.sum(y_prob, axis=1),
np.ones(n_samples))
# raises error on malformed input
assert_raises(ValueError, classifier.predict_proba, X.T)
# raises error on malformed input for predict_proba
assert_raises(ValueError, classifier.predict_proba, X.T)
except NotImplementedError:
pass
def test_explained_variance():
# Test sparse data
svd_a_10_sp = TruncatedSVD(10, algorithm="arpack")
svd_r_10_sp = TruncatedSVD(10, algorithm="randomized", random_state=42)
svd_a_20_sp = TruncatedSVD(20, algorithm="arpack")
svd_r_20_sp = TruncatedSVD(20, algorithm="randomized", random_state=42)
X_trans_a_10_sp = svd_a_10_sp.fit_transform(X)
X_trans_r_10_sp = svd_r_10_sp.fit_transform(X)
X_trans_a_20_sp = svd_a_20_sp.fit_transform(X)
X_trans_r_20_sp = svd_r_20_sp.fit_transform(X)
# Test dense data
svd_a_10_de = TruncatedSVD(10, algorithm="arpack")
svd_r_10_de = TruncatedSVD(10, algorithm="randomized", random_state=42)
svd_a_20_de = TruncatedSVD(20, algorithm="arpack")
svd_r_20_de = TruncatedSVD(20, algorithm="randomized", random_state=42)
X_trans_a_10_de = svd_a_10_de.fit_transform(X.toarray())
X_trans_r_10_de = svd_r_10_de.fit_transform(X.toarray())
X_trans_a_20_de = svd_a_20_de.fit_transform(X.toarray())
X_trans_r_20_de = svd_r_20_de.fit_transform(X.toarray())
# helper arrays for tests below
svds = (svd_a_10_sp, svd_r_10_sp, svd_a_20_sp, svd_r_20_sp, svd_a_10_de,
svd_r_10_de, svd_a_20_de, svd_r_20_de)
svds_trans = (
(svd_a_10_sp, X_trans_a_10_sp),
(svd_r_10_sp, X_trans_r_10_sp),
(svd_a_20_sp, X_trans_a_20_sp),
(svd_r_20_sp, X_trans_r_20_sp),
(svd_a_10_de, X_trans_a_10_de),
(svd_r_10_de, X_trans_r_10_de),
(svd_a_20_de, X_trans_a_20_de),
(svd_r_20_de, X_trans_r_20_de),
)
svds_10_v_20 = (
(svd_a_10_sp, svd_a_20_sp),
(svd_r_10_sp, svd_r_20_sp),
(svd_a_10_de, svd_a_20_de),
(svd_r_10_de, svd_r_20_de),
)
svds_sparse_v_dense = (
(svd_a_10_sp, svd_a_10_de),
(svd_a_20_sp, svd_a_20_de),
(svd_r_10_sp, svd_r_10_de),
(svd_r_20_sp, svd_r_20_de),
)
# Assert the 1st component is equal
for svd_10, svd_20 in svds_10_v_20:
assert_array_almost_equal(
svd_10.explained_variance_ratio_,
svd_20.explained_variance_ratio_[:10],
decimal=5,
)
# Assert that 20 components has higher explained variance than 10
for svd_10, svd_20 in svds_10_v_20:
assert_greater(
svd_20.explained_variance_ratio_.sum(),
svd_10.explained_variance_ratio_.sum(),
)
# Assert that all the values are greater than 0
for svd in svds:
assert_array_less(0.0, svd.explained_variance_ratio_)
# Assert that total explained variance is less than 1
for svd in svds:
assert_array_less(svd.explained_variance_ratio_.sum(), 1.0)
# Compare sparse vs. dense
for svd_sparse, svd_dense in svds_sparse_v_dense:
assert_array_almost_equal(svd_sparse.explained_variance_ratio_,
svd_dense.explained_variance_ratio_)
# Test that explained_variance is correct
for svd, transformed in svds_trans:
total_variance = np.var(X.toarray(), axis=0).sum()
variances = np.var(transformed, axis=0)
true_explained_variance_ratio = variances / total_variance
assert_array_almost_equal(
svd.explained_variance_ratio_,
true_explained_variance_ratio,
)
def test_linear_svr_fit_intercept():
reg = LinearSVR(random_state=0, fit_intercept=True)
reg.fit(reg_dense, reg_target)
assert_greater(reg.score(reg_dense, reg_target), 0.99)
def test_prediction_scores(self):
pred_scores = self.clf.decision_function(self.X_test)
# check score shapes
assert_equal(pred_scores.shape[0], self.X_test.shape[0])
# check performance
assert_greater(roc_auc_score(self.y_test, pred_scores), self.roc_floor)
def check_explicit_sparse_zeros(tree, max_depth=3, n_features=10):
TreeEstimator = ALL_TREES[tree]
# n_samples set n_feature to ease construction of a simultaneous
# construction of a csr and csc matrix
n_samples = n_features
samples = np.arange(n_samples)
# Generate X, y
random_state = check_random_state(0)
indices = []
data = []
offset = 0
indptr = [offset]
for i in range(n_features):
n_nonzero_i = random_state.binomial(n_samples, 0.5)
indices_i = random_state.permutation(samples)[:n_nonzero_i]
indices.append(indices_i)
data_i = random_state.binomial(3, 0.5, size=(n_nonzero_i, )) - 1
data.append(data_i)
offset += n_nonzero_i
indptr.append(offset)
indices = np.concatenate(indices)
data = np.array(np.concatenate(data), dtype=np.float32)
X_sparse = csc_matrix((data, indices, indptr),
shape=(n_samples, n_features))
X = X_sparse.toarray()
X_sparse_test = csr_matrix((data, indices, indptr),
shape=(n_samples, n_features))
X_test = X_sparse_test.toarray()
y = random_state.randint(0, 3, size=(n_samples, ))
# Ensure that X_sparse_test owns its data, indices and indptr array
X_sparse_test = X_sparse_test.copy()
# Ensure that we have explicit zeros
assert_greater((X_sparse.data == 0.).sum(), 0)
assert_greater((X_sparse_test.data == 0.).sum(), 0)
# Perform the comparison
d = TreeEstimator(random_state=0, max_depth=max_depth).fit(X, y)
s = TreeEstimator(random_state=0, max_depth=max_depth).fit(X_sparse, y)
assert_tree_equal(
d.tree_, s.tree_, "{0} with dense and sparse format gave different "
"trees".format(tree))
Xs = (X_test, X_sparse_test)
for X1, X2 in product(Xs, Xs):
assert_array_almost_equal(s.tree_.apply(X1), d.tree_.apply(X2))
assert_array_almost_equal(s.apply(X1), d.apply(X2))
assert_array_almost_equal(s.apply(X1), s.tree_.apply(X1))
assert_array_almost_equal(
s.tree_.decision_path(X1).toarray(),
d.tree_.decision_path(X2).toarray())
assert_array_almost_equal(
s.decision_path(X1).toarray(),
d.decision_path(X2).toarray())
assert_array_almost_equal(
s.decision_path(X1).toarray(),
s.tree_.decision_path(X1).toarray())
assert_array_almost_equal(s.predict(X1), d.predict(X2))
if tree in CLF_TREES:
assert_array_almost_equal(s.predict_proba(X1), d.predict_proba(X2))
def test_multiclass_sgd():
clf = SGDClassifier(random_state=0)
clf.fit(mult_dense, mult_target)
assert_greater(clf.score(mult_dense, mult_target), 0.80)
assert_equal(list(clf.classes_), [0, 1, 2])
def test_lgmlvq_iris():
check_estimator(LgmlvqModel)
model = LgmlvqModel()
model.fit(iris.data, iris.target)
assert_greater(model.score(iris.data, iris.target), 0.95)
assert_raise_message(ValueError, 'regularization must be a positive float',
LgmlvqModel(regularization=-1.0).fit, iris.data,
iris.target)
assert_raise_message(ValueError,
'length of regularization'
' must be number of prototypes',
LgmlvqModel(regularization=[-1.0]).fit, iris.data,
iris.target)
assert_raise_message(ValueError,
'length of regularization must be number of classes',
LgmlvqModel(regularization=[-1.0],
classwise=True).fit, iris.data,
iris.target)
assert_raise_message(ValueError, 'initial matrices must be a list',
LgmlvqModel(initial_matrices=np.array(
[[1, 2], [3, 4], [5, 6]])).fit, iris.data,
iris.target)
assert_raise_message(ValueError, 'length of matrices wrong',
LgmlvqModel(
initial_matrices=[[[1, 2], [3, 4], [5, 6]]]).fit,
iris.data, iris.target)
assert_raise_message(ValueError, 'each matrix should have',
LgmlvqModel(
initial_matrices=[[[1]], [[1]], [[1]]]).fit,
iris.data, iris.target)
assert_raise_message(ValueError, 'length of matrices wrong',
LgmlvqModel(initial_matrices=[[[1, 2, 3]]],
classwise=True).fit, iris.data,
iris.target)
assert_raise_message(ValueError, 'each matrix should have',
LgmlvqModel(initial_matrices=[[[1]], [[1]], [[1]]],
classwise=True).fit, iris.data,
iris.target)
assert_raise_message(ValueError, 'classwise must be a boolean',
LgmlvqModel(classwise="a").fit, iris.data,
iris.target)
assert_raise_message(ValueError, 'dim must be a list of positive ints',
LgmlvqModel(dim=[-1]).fit, iris.data, iris.target)
assert_raise_message(ValueError, 'dim length must be number of prototypes',
LgmlvqModel(dim=[1, 1]).fit, iris.data, iris.target)
assert_raise_message(ValueError, 'dim length must be number of classes',
LgmlvqModel(dim=[1, 1], classwise=True).fit,
iris.data, iris.target)
LgmlvqModel(classwise=True, dim=[1], prototypes_per_class=2).fit(
iris.data, iris.target)
model = LgmlvqModel(regularization=0.1)
model.fit(iris.data, iris.target)
model = LgmlvqModel(initial_prototypes=[[0, 2, 1], [1, 6, 2]],
initial_matrices=[np.ones([2, 2]), np.ones([2, 2])],
dim=[2, 2])
x = np.array([[0, 0], [0, 4], [1, 4], [1, 8]])
y = np.array([1, 1, 2, 2])
model.fit(x, y)
def test_linear_svr():
reg = LinearSVR(random_state=0)
reg.fit(reg_dense, reg_target)
assert_greater(reg.score(reg_dense, reg_target), 0.99)
def test_perceptron_accuracy():
for data in (X, X_csr):
clf = Perceptron(max_iter=100, tol=None, shuffle=False)
clf.fit(data, y)
score = clf.score(data, y)
assert_greater(score, 0.7)
def test_lml_improving(kernel):
# Test that hyperparameter-tuning improves log-marginal likelihood.
gpr = GaussianProcessRegressor(kernel=kernel).fit(X, y)
assert_greater(gpr.log_marginal_likelihood(gpr.kernel_.theta),
gpr.log_marginal_likelihood(kernel.theta))
def check_classifiers_train(name, classifier_orig, readonly_memmap=False):
X_m, y_m = make_blobs(n_samples=300, random_state=0)
X_m, y_m = shuffle(X_m, y_m, random_state=7)
X_m = StandardScaler().fit_transform(X_m)
# generate binary problem from multi-class one
y_b = y_m[y_m != 2]
X_b = X_m[y_m != 2]
if readonly_memmap:
X_b, y_b = create_memmap_backed_data([X_b, y_b])
for (X, y) in [(X_b, y_b)]:
classes = np.unique(y)
n_classes = len(classes)
n_samples, _ = X.shape
classifier = clone(classifier_orig)
X = pairwise_estimator_convert_X(X, classifier)
y = multioutput_estimator_convert_y_2d(classifier, y)
set_random_state(classifier)
# raises error on malformed input for fit
with assert_raises(ValueError,
msg="The classifier {} does not "
"raise an error when incorrect/malformed input "
"data for fit is passed. The number of training "
"examples is not the same as the number of labels. "
"Perhaps use check_X_y in fit.".format(name)):
classifier.fit(X, y[:-1])
# fit
classifier.fit(X, y)
# with lists
classifier.fit(X.tolist(), y.tolist())
assert hasattr(classifier, "classes_")
y_pred = classifier.predict(X)
assert_equal(y_pred.shape, (n_samples, ))
# training set performance
assert_greater(accuracy_score(y, y_pred), 0.83)
# raises error on malformed input for predict
msg = ("The classifier {} does not raise an error when the number of "
"features in {} is different from the number of features in "
"fit.")
with assert_raises(ValueError, msg=msg.format(name, "predict")):
classifier.predict(X.T)
if hasattr(classifier, "decision_function"):
try:
# decision_function agrees with predict
decision = classifier.decision_function(X)
if n_classes == 2:
assert_equal(decision.shape, (n_samples, 1))
dec_pred = (decision.ravel() > 0).astype(np.int)
assert_array_equal(dec_pred, y_pred)
else:
assert_equal(decision.shape, (n_samples, n_classes))
assert_array_equal(np.argmax(decision, axis=1), y_pred)
# raises error on malformed input for decision_function
with assert_raises(ValueError,
msg=msg.format(name, "decision_function")):
classifier.decision_function(X.T)
except NotImplementedError:
pass
if hasattr(classifier, "predict_proba"):
# predict_proba agrees with predict
y_prob = classifier.predict_proba(X)
assert_equal(y_prob.shape, (n_samples, n_classes))
assert_array_equal(np.argmax(y_prob, axis=1), y_pred)
# check that probas for all classes sum to one
assert_array_almost_equal(np.sum(y_prob, axis=1),
np.ones(n_samples))
# raises error on malformed input for predict_proba
with assert_raises(ValueError,
msg=msg.format(name, "predict_proba")):
classifier.predict_proba(X.T)
if hasattr(classifier, "predict_log_proba"):
# predict_log_proba is a transformation of predict_proba
y_log_prob = classifier.predict_log_proba(X)
assert_allclose(y_log_prob, np.log(y_prob), 8, atol=1e-9)
assert_array_equal(np.argsort(y_log_prob), np.argsort(y_prob))
def test_whitening():
# Check that PCA output has unit-variance
rng = np.random.RandomState(0)
n_samples = 100
n_features = 80
n_components = 30
rank = 50
# some low rank data with correlated features
X = np.dot(
rng.randn(n_samples, rank),
np.dot(np.diag(np.linspace(10.0, 1.0, rank)),
rng.randn(rank, n_features)),
)
# the component-wise variance of the first 50 features is 3 times the
# mean component-wise variance of the remaining 30 features
X[:, :50] *= 3
assert_equal(X.shape, (n_samples, n_features))
# the component-wise variance is thus highly varying:
assert_greater(X.std(axis=0).std(), 43.8)
dX = da.from_array(X, chunks=(50, n_features))
for solver, copy in product(solver_list, (True, False)):
# whiten the data while projecting to the lower dim subspace
X_ = dX.copy() # make sure we keep an original across iterations.
pca = dd.PCA(
n_components=n_components,
whiten=True,
copy=copy,
svd_solver=solver,
random_state=0,
iterated_power=4,
)
# test fit_transform
X_whitened = pca.fit_transform(X_.copy())
assert_equal(X_whitened.shape, (n_samples, n_components))
# X_whitened2 = pca.transform(X_)
# XXX: These differ for randomized.
# assert_eq(X_whitened.compute(), X_whitened2.compute(),
# atol=tol, rtol=tol)
assert_almost_equal(X_whitened.std(ddof=1, axis=0),
np.ones(n_components),
decimal=6)
assert_almost_equal(X_whitened.mean(axis=0), np.zeros(n_components))
X_ = dX.copy()
pca = dd.PCA(
n_components=n_components,
whiten=False,
copy=copy,
svd_solver=solver,
random_state=0,
).fit(X_)
X_unwhitened = pca.transform(X_)
assert_equal(X_unwhitened.shape, (n_samples, n_components))
# in that case the output components still have varying variances
assert_almost_equal(X_unwhitened.std(axis=0).std(), 74.1, 1)