def check_importances(name, X, y):
"""Check variable importances."""
ForestClassifier = FOREST_CLASSIFIERS[name]
for n_jobs in [1, 2]:
clf = ForestClassifier(n_estimators=10, n_jobs=n_jobs)
clf.fit(X, y)
importances = clf.feature_importances_
n_important = np.sum(importances > 0.1)
assert_equal(importances.shape[0], 10)
assert_equal(n_important, 3)
X_new = clf.transform(X, threshold="mean")
assert_less(0 < X_new.shape[1], X.shape[1])
# Check with sample weights
sample_weight = np.ones(y.shape)
sample_weight[y == 1] *= 100
clf = ForestClassifier(n_estimators=50, n_jobs=n_jobs, random_state=0)
clf.fit(X, y, sample_weight=sample_weight)
importances = clf.feature_importances_
assert_true(np.all(importances >= 0.0))
clf = ForestClassifier(n_estimators=50, n_jobs=n_jobs, random_state=0)
clf.fit(X, y, sample_weight=3 * sample_weight)
importances_bis = clf.feature_importances_
assert_almost_equal(importances, importances_bis)
def test_sparse_enet_not_as_toy_dataset():
n_samples, n_features, max_iter = 100, 100, 1000
n_informative = 10
X, y = make_sparse_data(n_samples, n_features, n_informative)
X_train, X_test = X[n_samples / 2:], X[:n_samples / 2]
y_train, y_test = y[n_samples / 2:], y[:n_samples / 2]
s_clf = SparseENet(alpha=0.1, rho=0.8, fit_intercept=False,
max_iter=max_iter, tol=1e-7)
s_clf.fit(X_train, y_train)
assert_almost_equal(s_clf.dual_gap_, 0, 4)
assert_greater(s_clf.score(X_test, y_test), 0.85)
# check the convergence is the same as the dense version
d_clf = DenseENet(alpha=0.1, rho=0.8, fit_intercept=False,
max_iter=max_iter, tol=1e-7)
d_clf.fit(X_train, y_train)
assert_almost_equal(d_clf.dual_gap_, 0, 4)
assert_greater(d_clf.score(X_test, y_test), 0.85)
assert_almost_equal(s_clf.coef_, d_clf.coef_, 5)
# check that the coefs are sparse
assert_less(np.sum(s_clf.coef_ != 0.0), 2 * n_informative)
def test_feature_importances_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_true(hasattr(transformer.estimator_, "coef_"))
X_new = transformer.transform(X)
assert_less(X_new.shape[1], X.shape[1])
# Manually check that the norm is correctly performed
est.fit(X, y)
importances = norm(est.coef_, axis=0, ord=order)
feature_mask = importances > func(importances)
assert_array_equal(X_new, X[:, feature_mask])
def check_boston(presort, loss, subsample):
# Check consistency on dataset boston house prices with least squares
# and least absolute deviation.
ones = np.ones(len(boston.target))
last_y_pred = None
for sample_weight in None, ones, 2 * ones:
clf = GradientBoostingRegressor(n_estimators=100,
loss=loss,
max_depth=4,
subsample=subsample,
min_samples_split=2,
random_state=1,
presort=presort)
assert_raises(ValueError, clf.predict, boston.data)
clf.fit(boston.data, boston.target,
sample_weight=sample_weight)
leaves = clf.apply(boston.data)
assert_equal(leaves.shape, (506, 100))
y_pred = clf.predict(boston.data)
mse = mean_squared_error(boston.target, y_pred)
assert_less(mse, 6.0)
if last_y_pred is not None:
assert_array_almost_equal(last_y_pred, y_pred)
last_y_pred = y_pred
def test_lars_drop_for_good():
# Create an ill-conditioned situation in which the LARS has to go
# far in the path to converge, and check that LARS and coordinate
# descent give the same answers
X = [[1e20, 1e20, 0],
[-1e-32, 0, 0],
[1, 1, 1]]
y = [10, 10, 1]
alpha = .0001
def objective_function(coef):
return (1. / (2. * len(X)) * linalg.norm(y - np.dot(X, coef)) ** 2
+ alpha * linalg.norm(coef, 1))
lars = linear_model.LassoLars(alpha=alpha, normalize=False)
assert_warns(ConvergenceWarning, lars.fit, X, y)
lars_coef_ = lars.coef_
lars_obj = objective_function(lars_coef_)
coord_descent = linear_model.Lasso(alpha=alpha, tol=1e-10, normalize=False)
with ignore_warnings():
cd_coef_ = coord_descent.fit(X, y).coef_
cd_obj = objective_function(cd_coef_)
assert_less(lars_obj, cd_obj * (1. + 1e-8))
def test_oob_score_classification():
# Check that oob prediction is a good estimation of the generalization
# error.
rng = check_random_state(0)
X_train, X_test, y_train, y_test = train_test_split(iris.data,
iris.target,
random_state=rng)
for base_estimator in [DecisionTreeClassifier(), SVC()]:
clf = BaggingClassifier(base_estimator=base_estimator,
n_estimators=100,
bootstrap=True,
oob_score=True,
random_state=rng).fit(X_train, y_train)
test_score = clf.score(X_test, y_test)
assert_less(abs(test_score - clf.oob_score_), 0.1)
# Test with few estimators
assert_warns(UserWarning,
BaggingClassifier(base_estimator=base_estimator,
n_estimators=1,
bootstrap=True,
oob_score=True,
random_state=rng).fit,
X_train,
y_train)
def test_dense_liblinear_intercept_handling(classifier=svm.LinearSVC):
# Test that dense liblinear honours intercept_scaling param
X = [[2, 1],
[3, 1],
[1, 3],
[2, 3]]
y = [0, 0, 1, 1]
clf = classifier(fit_intercept=True, penalty='l1', loss='squared_hinge',
dual=False, C=4, tol=1e-7, random_state=0)
assert_true(clf.intercept_scaling == 1, clf.intercept_scaling)
assert_true(clf.fit_intercept)
# when intercept_scaling is low the intercept value is highly "penalized"
# by regularization
clf.intercept_scaling = 1
clf.fit(X, y)
assert_almost_equal(clf.intercept_, 0, decimal=5)
# when intercept_scaling is sufficiently high, the intercept value
# is not affected by regularization
clf.intercept_scaling = 100
clf.fit(X, y)
intercept1 = clf.intercept_
assert_less(intercept1, -1)
# when intercept_scaling is sufficiently high, the intercept value
# doesn't depend on intercept_scaling value
clf.intercept_scaling = 1000
clf.fit(X, y)
intercept2 = clf.intercept_
assert_array_almost_equal(intercept1, intercept2, decimal=2)
def test_fitted_model(self):
# non centered, sparse centers to check the
centers = np.array([
[0.0, 5.0, 0.0, 0.0, 0.0],
[1.0, 1.0, 4.0, 0.0, 0.0],
[1.0, 0.0, 0.0, 5.0, 1.0],
])
n_samples = 100
n_clusters, n_features = centers.shape
X, true_labels = make_blobs(n_samples=n_samples, centers=centers,
cluster_std=1., random_state=42)
cbook = CoodeBook(n_words=3)
cbook = cbook.fit(X) # TODO: Is it neaded to reasign? or it can be just cbook.fit(X)
# check that the number of clusters centers and distinct labels match
# the expectation
centers = cbook.get_dictionary()
assert_equal(centers.shape, (n_clusters, n_features))
labels = cbook.predict(X)
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(cbook.cluster_core.inertia_, 0.0)
# check that the descriptor looks like the homogenous PDF used
# to create the original samples
cbook_hist = cbook.get_BoF_descriptor(X)
expected_value = float(1)/cbook.n_words
for bin_value in cbook_hist[0]:
assert_less(round(bin_value-expected_value,3), 0.01)
def test_oob_score_regression():
# Check that oob prediction is a good estimation of the generalization
# error.
rng = check_random_state(0)
X_train, X_test, y_train, y_test = train_test_split(boston.data,
boston.target,
random_state=rng)
clf = BaggingRegressor(base_estimator=DecisionTreeRegressor(),
n_estimators=50,
bootstrap=True,
oob_score=True,
random_state=rng).fit(X_train, y_train)
test_score = clf.score(X_test, y_test)
assert_less(abs(test_score - clf.oob_score_), 0.1)
# Test with few estimators
assert_warns(UserWarning,
BaggingRegressor(base_estimator=DecisionTreeRegressor(),
n_estimators=1,
bootstrap=True,
oob_score=True,
random_state=rng).fit,
X_train,
y_train)
def check_minimize(func, y_opt, bounds, acq_optimizer, acq_func,
margin, n_calls, n_random_starts=10):
r = gp_minimize(func, bounds, acq_optimizer=acq_optimizer,
acq_func=acq_func, n_random_starts=n_random_starts,
n_calls=n_calls, random_state=1,
noise=1e-10)
assert_less(r.fun, y_opt + margin)
def test_importances():
"""Check variable importances."""
X, y = datasets.make_classification(n_samples=2000,
n_features=10,
n_informative=3,
n_redundant=0,
n_repeated=0,
shuffle=False,
random_state=0)
for name, Tree in CLF_TREES.items():
clf = Tree(random_state=0)
clf.fit(X, y)
importances = clf.feature_importances_
n_important = np.sum(importances > 0.1)
assert_equal(importances.shape[0], 10, "Failed with {0}".format(name))
assert_equal(n_important, 3, "Failed with {0}".format(name))
X_new = clf.transform(X, threshold="mean")
assert_less(0, X_new.shape[1], "Failed with {0}".format(name))
assert_less(X_new.shape[1], X.shape[1], "Failed with {0}".format(name))
# Check on iris that importances are the same for all builders
clf = DecisionTreeClassifier(random_state=0)
clf.fit(iris.data, iris.target)
clf2 = DecisionTreeClassifier(random_state=0,
max_leaf_nodes=len(iris.data))
clf2.fit(iris.data, iris.target)
assert_array_equal(clf.feature_importances_,
clf2.feature_importances_)
def test_lasso_lars_vs_lasso_cd_ill_conditioned2():
# Create an ill-conditioned situation in which the LARS has to go
# far in the path to converge, and check that LARS and coordinate
# descent give the same answers
# Note it used to be the case that Lars had to use the drop for good
# strategy for this but this is no longer the case with the
# equality_tolerance checks
X = [[1e20, 1e20, 0],
[-1e-32, 0, 0],
[1, 1, 1]]
y = [10, 10, 1]
alpha = .0001
def objective_function(coef):
return (1. / (2. * len(X)) * linalg.norm(y - np.dot(X, coef)) ** 2
+ alpha * linalg.norm(coef, 1))
lars = linear_model.LassoLars(alpha=alpha, normalize=False)
assert_warns(ConvergenceWarning, lars.fit, X, y)
lars_coef_ = lars.coef_
lars_obj = objective_function(lars_coef_)
coord_descent = linear_model.Lasso(alpha=alpha, tol=1e-4, normalize=False)
cd_coef_ = coord_descent.fit(X, y).coef_
cd_obj = objective_function(cd_coef_)
assert_less(lars_obj, cd_obj * (1. + 1e-8))
def test_rank_deficient_design():
# consistency test that checks that LARS Lasso is handling rank
# deficient input data (with n_features < rank) in the same way
# as coordinate descent Lasso
y = [5, 0, 5]
for X in ([[5, 0],
[0, 5],
[10, 10]],
[[10, 10, 0],
[1e-32, 0, 0],
[0, 0, 1]],
):
# To be able to use the coefs to compute the objective function,
# we need to turn off normalization
lars = linear_model.LassoLars(.1, normalize=False)
coef_lars_ = lars.fit(X, y).coef_
obj_lars = (1. / (2. * 3.)
* linalg.norm(y - np.dot(X, coef_lars_)) ** 2
+ .1 * linalg.norm(coef_lars_, 1))
coord_descent = linear_model.Lasso(.1, tol=1e-6, normalize=False)
coef_cd_ = coord_descent.fit(X, y).coef_
obj_cd = ((1. / (2. * 3.)) * linalg.norm(y - np.dot(X, coef_cd_)) ** 2
+ .1 * linalg.norm(coef_cd_, 1))
assert_less(obj_lars, obj_cd * (1. + 1e-8))
def _test_sparse_enet_not_as_toy_dataset(alpha, fit_intercept, positive):
n_samples, n_features, max_iter = 100, 100, 1000
n_informative = 10
X, y = make_sparse_data(n_samples, n_features, n_informative,
positive=positive)
X_train, X_test = X[n_samples / 2:], X[:n_samples / 2]
y_train, y_test = y[n_samples / 2:], y[:n_samples / 2]
s_clf = ElasticNet(alpha=alpha, l1_ratio=0.8, fit_intercept=fit_intercept,
max_iter=max_iter, tol=1e-7, positive=positive,
warm_start=True)
s_clf.fit(X_train, y_train)
assert_almost_equal(s_clf.dual_gap_, 0, 4)
assert_greater(s_clf.score(X_test, y_test), 0.85)
# check the convergence is the same as the dense version
d_clf = ElasticNet(alpha=alpha, l1_ratio=0.8, fit_intercept=fit_intercept,
max_iter=max_iter, tol=1e-7, positive=positive,
warm_start=True)
d_clf.fit(X_train.todense(), y_train)
assert_almost_equal(d_clf.dual_gap_, 0, 4)
assert_greater(d_clf.score(X_test, y_test), 0.85)
assert_almost_equal(s_clf.coef_, d_clf.coef_, 5)
assert_almost_equal(s_clf.intercept_, d_clf.intercept_, 5)
# check that the coefs are sparse
assert_less(np.sum(s_clf.coef_ != 0.0), 2 * n_informative)
def test_nmf_fit_close(solver):
rng = np.random.mtrand.RandomState(42)
# Test that the fit is not too far away
pnmf = NMF(5, solver=solver, init='nndsvdar', random_state=0,
max_iter=600)
X = np.abs(rng.randn(6, 5))
assert_less(pnmf.fit(X).reconstruction_err_, 0.1)
def test_importances():
"""Check variable importances."""
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)
clf = RandomForestClassifier(n_estimators=10)
clf.fit(X, y)
importances = clf.feature_importances_
n_important = sum(importances > 0.1)
assert_equal(importances.shape[0], 10)
assert_equal(n_important, 3)
X_new = clf.transform(X, threshold="mean")
assert_less(0 < X_new.shape[1], X.shape[1])
# Check with sample weights
sample_weight = np.ones(y.shape)
sample_weight[y == 1] *= 100
clf = RandomForestClassifier(n_estimators=50, random_state=0)
clf.fit(X, y, sample_weight=sample_weight)
importances = clf.feature_importances_
assert np.all(importances >= 0.0)
clf = RandomForestClassifier(n_estimators=50, random_state=0)
clf.fit(X, y, sample_weight=3*sample_weight)
importances_bis = clf.feature_importances_
assert_almost_equal(importances, importances_bis)
def test_lasso_lars_vs_lasso_cd_early_stopping(verbose=False):
# Test that LassoLars and Lasso using coordinate descent give the
# same results when early stopping is used.
# (test : before, in the middle, and in the last part of the path)
alphas_min = [10, 0.9, 1e-4]
for alpha_min in alphas_min:
alphas, _, lasso_path = linear_model.lars_path(X, y, method='lasso',
alpha_min=alpha_min)
lasso_cd = linear_model.Lasso(fit_intercept=False, tol=1e-8)
lasso_cd.alpha = alphas[-1]
lasso_cd.fit(X, y)
error = linalg.norm(lasso_path[:, -1] - lasso_cd.coef_)
assert_less(error, 0.01)
# same test, with normalization
for alpha_min in alphas_min:
alphas, _, lasso_path = linear_model.lars_path(X, y, method='lasso',
alpha_min=alpha_min)
lasso_cd = linear_model.Lasso(fit_intercept=True, normalize=True,
tol=1e-8)
lasso_cd.alpha = alphas[-1]
lasso_cd.fit(X, y)
error = linalg.norm(lasso_path[:, -1] - lasso_cd.coef_)
assert_less(error, 0.01)
def check_minimize(minimizer, func, y_opt, dimensions, margin,
n_calls, n_random_starts=10, x0=None):
for n in range(3):
r = minimizer(
func, dimensions, n_calls=n_calls, random_state=n,
n_random_starts=n_random_starts, x0=x0)
assert_less(r.fun, y_opt + margin)
def test_sparse_encode_error():
n_atoms = 12
V = rng.randn(n_atoms, n_features) # random init
V /= np.sum(V ** 2, axis=1)[:, np.newaxis]
code = sparse_encode(X, V, alpha=0.001)
assert_true(not np.all(code == 0))
assert_less(np.sqrt(np.sum((np.dot(code, V) - X) ** 2)), 0.1)
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(X).sum())
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))
with warnings.catch_warnings(record=True) as w:
warnings.simplefilter('always', ConvergenceWarning)
fa1.max_iter = 1
fa1.verbose = True
fa1.fit(X)
assert_true(w[-1].category == ConvergenceWarning)
warnings.simplefilter('always', DeprecationWarning)
FactorAnalysis(verbose=1)
assert_true(w[-1].category == DeprecationWarning)
def test_random_projection_embedding_quality():
data, _ = make_sparse_random_data(8, 5000, 15000)
eps = 0.2
original_distances = euclidean_distances(data, squared=True)
original_distances = original_distances.ravel()
non_identical = original_distances != 0.0
# remove 0 distances to avoid division by 0
original_distances = original_distances[non_identical]
for RandomProjection in all_RandomProjection:
rp = RandomProjection(n_components='auto', eps=eps, random_state=0)
projected = rp.fit_transform(data)
projected_distances = euclidean_distances(projected, squared=True)
projected_distances = projected_distances.ravel()
# remove 0 distances to avoid division by 0
projected_distances = projected_distances[non_identical]
distances_ratio = projected_distances / original_distances
# check that the automatically tuned values for the density respect the
# contract for eps: pairwise distances are preserved according to the
# Johnson-Lindenstrauss lemma
assert_less(distances_ratio.max(), 1 + eps)
assert_less(1 - eps, distances_ratio.min())
def check_importances(X, y, name, criterion):
ForestEstimator = FOREST_ESTIMATORS[name]
est = ForestEstimator(n_estimators=20, criterion=criterion,
random_state=0)
est.fit(X, y)
importances = est.feature_importances_
n_important = np.sum(importances > 0.1)
assert_equal(importances.shape[0], 10)
assert_equal(n_important, 3)
# XXX: Remove this test in 0.19 after transform support to estimators
# is removed.
X_new = assert_warns(
DeprecationWarning, est.transform, X, threshold="mean")
assert_less(0 < X_new.shape[1], X.shape[1])
# Check with parallel
importances = est.feature_importances_
est.set_params(n_jobs=2)
importances_parrallel = est.feature_importances_
assert_array_almost_equal(importances, importances_parrallel)
# Check with sample weights
sample_weight = check_random_state(0).randint(1, 10, len(X))
est = ForestEstimator(n_estimators=20, random_state=0, criterion=criterion)
est.fit(X, y, sample_weight=sample_weight)
importances = est.feature_importances_
assert_true(np.all(importances >= 0.0))
for scale in [0.5, 10, 100]:
est = ForestEstimator(n_estimators=20, random_state=0, criterion=criterion)
est.fit(X, y, sample_weight=scale * sample_weight)
importances_bis = est.feature_importances_
assert_less(np.abs(importances - importances_bis).mean(), 0.001)
def check_importances(name, criterion, X, y):
ForestEstimator = FOREST_ESTIMATORS[name]
est = ForestEstimator(n_estimators=20, criterion=criterion, random_state=0)
est.fit(X, y)
importances = est.feature_importances_
n_important = np.sum(importances > 0.1)
assert_equal(importances.shape[0], 10)
assert_equal(n_important, 3)
# Check with parallel
importances = est.feature_importances_
est.set_params(n_jobs=2)
importances_parrallel = est.feature_importances_
assert_array_almost_equal(importances, importances_parrallel)
# Check with sample weights
sample_weight = check_random_state(0).randint(1, 10, len(X))
est = ForestEstimator(n_estimators=20, random_state=0, criterion=criterion)
est.fit(X, y, sample_weight=sample_weight)
importances = est.feature_importances_
assert_true(np.all(importances >= 0.0))
for scale in [0.5, 10, 100]:
est = ForestEstimator(n_estimators=20, random_state=0, criterion=criterion)
est.fit(X, y, sample_weight=scale * sample_weight)
importances_bis = est.feature_importances_
assert_less(np.abs(importances - importances_bis).mean(), 0.001)
def check_tree_models(name,
use_feature_hashing=False,
use_rescaling=False):
# create a FeatureSet object with the data we want to use
if use_feature_hashing:
train_fs, test_fs, _ = make_regression_data(num_examples=5000,
num_features=10,
use_feature_hashing=True,
feature_bins=5)
else:
train_fs, test_fs, _ = make_regression_data(num_examples=2000,
num_features=3)
# create the learner
if use_rescaling:
name = 'Rescaled' + name
learner = Learner(name)
# train it with the training feature set we created
# make sure to set the grid objective to pearson
learner.train(train_fs, grid_objective='pearson')
# make sure that the feature importances are as expected.
if name.endswith('DecisionTreeRegressor'):
expected_feature_importances = ([0.37331461,
0.08572699,
0.2543484,
0.1841172,
0.1024928] if use_feature_hashing else
[0.08931994,
0.15545093,
0.75522913])
expected_cor_range = [0.5, 0.6] if use_feature_hashing else [0.9, 1.0]
else:
if use_feature_hashing:
expected_feature_importances = [0.40195655,
0.06702161,
0.25814858,
0.18183947,
0.09103379]
else:
expected_feature_importances = [0.07975691, 0.16122862, 0.75901447]
expected_cor_range = [0.7, 0.8] if use_feature_hashing else [0.9, 1.0]
feature_importances = learner.model.feature_importances_
assert_allclose(feature_importances, expected_feature_importances,
rtol=1e-2)
# now generate the predictions on the test FeatureSet
predictions = learner.predict(test_fs)
# now make sure that the predictions are close to
# the actual test FeatureSet labels that we generated
# using make_regression_data. To do this, we just
# make sure that they are correlated with pearson > 0.95
cor, _ = pearsonr(predictions, test_fs.labels)
assert_greater(cor, expected_cor_range[0])
assert_less(cor, expected_cor_range[1])
def check_tree_models(name,
use_feature_hashing=False,
use_rescaling=False):
# create a FeatureSet object with the data we want to use
if use_feature_hashing:
train_fs, test_fs, _ = make_regression_data(num_examples=5000,
num_features=10,
use_feature_hashing=True,
feature_bins=5)
else:
train_fs, test_fs, _ = make_regression_data(num_examples=2000,
num_features=3)
# create the learner
if use_rescaling:
name = 'Rescaled' + name
learner = Learner(name)
# train it with the training feature set we created
# make sure to set the grid objective to pearson
learner.train(train_fs, grid_objective='pearson')
# make sure that the feature importances are as expected.
if name.endswith('DecisionTreeRegressor'):
expected_feature_importances = ([0.37483895,
0.08816508,
0.25379838,
0.18337128,
0.09982631] if use_feature_hashing else
[0.08926899,
0.15585068,
0.75488033])
expected_cor_range = [0.5, 0.6] if use_feature_hashing else [0.9, 1.0]
else:
expected_feature_importances = ([0.40195798,
0.06702903,
0.25816559,
0.18185518,
0.09099222] if use_feature_hashing else
[0.07974267,
0.16121895,
0.75903838])
expected_cor_range = [0.7, 0.8] if use_feature_hashing else [0.9, 1.0]
feature_importances = learner.model.feature_importances_
assert_allclose(feature_importances, expected_feature_importances,
atol=1e-2, rtol=0)
# now generate the predictions on the test FeatureSet
predictions = learner.predict(test_fs)
# now make sure that the predictions are close to
# the actual test FeatureSet labels that we generated
# using make_regression_data. To do this, we just
# make sure that they are correlated with pearson > 0.95
cor, _ = pearsonr(predictions, test_fs.labels)
assert_greater(cor, expected_cor_range[0])
assert_less(cor, expected_cor_range[1])
def test_oob_score_classification():
"""Check that oob prediction is a good estimation of the generalization
error."""
clf = RandomForestClassifier(oob_score=True, random_state=rng)
n_samples = iris.data.shape[0]
clf.fit(iris.data[: n_samples / 2, :], iris.target[: n_samples / 2])
test_score = clf.score(iris.data[n_samples / 2 :, :], iris.target[n_samples / 2 :])
assert_less(abs(test_score - clf.oob_score_), 0.05)
def test_sparse_coder_estimator():
n_atoms = 12
V = rng.randn(n_atoms, n_features) # random init
V /= np.sum(V ** 2, axis=1)[:, np.newaxis]
code = SparseCoder(dictionary=V, transform_algorithm='lasso_lars',
transform_alpha=0.001).transform(X)
assert_true(not np.all(code == 0))
assert_less(np.sqrt(np.sum((np.dot(code, V) - X) ** 2)), 0.1)
def test_sparse_encode_error():
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]
code = sparse_encode(X, V, alpha=0.001)
assert not np.all(code == 0)
assert_less(np.sqrt(np.sum((np.dot(code, V) - X) ** 2)), 0.1)
def check_ensemble_models(name,
use_feature_hashing=False,
use_rescaling=False):
# create a FeatureSet object with the data we want to use
if use_feature_hashing:
train_fs, test_fs, _ = make_regression_data(num_examples=5000,
num_features=10,
use_feature_hashing=True,
feature_bins=5)
else:
train_fs, test_fs, _ = make_regression_data(num_examples=2000,
num_features=3)
# create the learner
if use_rescaling:
name = 'Rescaled' + name
learner = Learner(name)
# train it with the training feature set we created
# make sure to set the grid objective to pearson
learner.train(train_fs, grid_objective='pearson')
# make sure that the feature importances are as expected.
if name.endswith('AdaBoostRegressor'):
if use_feature_hashing:
expected_feature_importances = [0.33718443,
0.07810721,
0.25621769,
0.19489766,
0.13359301]
else:
expected_feature_importances = [0.10266744, 0.18681777, 0.71051479]
else:
expected_feature_importances = ([0.204,
0.172,
0.178,
0.212,
0.234] if use_feature_hashing else
[0.262,
0.288,
0.45])
feature_importances = learner.model.feature_importances_
assert_allclose(feature_importances, expected_feature_importances,
atol=1e-2, rtol=0)
# now generate the predictions on the test FeatureSet
predictions = learner.predict(test_fs)
# now make sure that the predictions are close to
# the actual test FeatureSet labels that we generated
# using make_regression_data. To do this, we just
# make sure that they are correlated with pearson > 0.95
cor, _ = pearsonr(predictions, test_fs.labels)
expected_cor_range = [0.7, 0.8] if use_feature_hashing else [0.9, 1.0]
assert_greater(cor, expected_cor_range[0])
assert_less(cor, expected_cor_range[1])
def test_sublinear_tf():
X = [[1], [2], [3]]
tr = TfidfTransformer(sublinear_tf=True, use_idf=False, norm=None)
tfidf = tr.fit_transform(X).toarray()
assert_equal(tfidf[0], 1)
assert_greater(tfidf[1], tfidf[0])
assert_greater(tfidf[2], tfidf[1])
assert_less(tfidf[1], 2)
assert_less(tfidf[2], 3)
def test_regressor_mse():
y_bin = y.copy()
y_bin[y != 1] = -1
for data in (X, X_csr):
for fit_intercept in (True, False):
for average in (False, True):
reg = PassiveAggressiveRegressor(C=1.0,
fit_intercept=fit_intercept,
random_state=0,
average=average,
max_iter=5)
reg.fit(data, y_bin)
pred = reg.predict(data)
assert_less(np.mean((pred - y_bin)**2), 1.7)
if average:
assert_true(hasattr(reg, 'average_coef_'))
assert_true(hasattr(reg, 'average_intercept_'))
assert_true(hasattr(reg, 'standard_intercept_'))
assert_true(hasattr(reg, 'standard_coef_'))
def test_select_fwe_regression():
# Test whether the relative univariate feature selection
# gets the correct items in a simple regression problem
# with the fwe heuristic
X, y = make_regression(n_samples=200,
n_features=20,
n_informative=5,
shuffle=False,
random_state=0)
univariate_filter = SelectFwe(f_regression, alpha=0.01)
X_r = univariate_filter.fit(X, y).transform(X)
X_r2 = GenericUnivariateSelect(f_regression, mode='fwe',
param=0.01).fit(X, y).transform(X)
assert_array_equal(X_r, X_r2)
support = univariate_filter.get_support()
gtruth = np.zeros(20)
gtruth[:5] = 1
assert_array_equal(support[:5], np.ones((5, ), dtype=np.bool))
assert_less(np.sum(support[5:] == 1), 2)
def test_regressor_partial_fit():
y_bin = y.copy()
y_bin[y != 1] = -1
for data in (X, X_csr):
for average in (False, True):
reg = PassiveAggressiveRegressor(C=1.0,
fit_intercept=True,
random_state=0,
average=average,
max_iter=100)
for t in range(50):
reg.partial_fit(data, y_bin)
pred = reg.predict(data)
assert_less(np.mean((pred - y_bin)**2), 1.7)
if average:
assert hasattr(reg, 'average_coef_')
assert hasattr(reg, 'average_intercept_')
assert hasattr(reg, 'standard_intercept_')
assert hasattr(reg, 'standard_coef_')
def test_transform_linear_model():
for clf in (LogisticRegression(C=0.1),
LinearSVC(C=0.01, dual=False),
SGDClassifier(alpha=0.001, n_iter=50, shuffle=True,
random_state=0)):
for thresh in (None, ".09*mean", "1e-5 * median"):
for func in (np.array, sp.csr_matrix):
X = func(data)
clf.set_params(penalty="l1")
clf.fit(X, y)
X_new = assert_warns(
DeprecationWarning, clf.transform, X, thresh)
if isinstance(clf, SGDClassifier):
assert_true(X_new.shape[1] <= X.shape[1])
else:
assert_less(X_new.shape[1], X.shape[1])
clf.set_params(penalty="l2")
clf.fit(X_new, y)
pred = clf.predict(X_new)
assert_greater(np.mean(pred == y), 0.7)
def test_importances():
"""Check variable importances."""
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)
clf = RandomForestClassifier(n_estimators=10)
clf.fit(X, y)
importances = clf.feature_importances_
n_important = sum(importances > 0.1)
assert_equal(importances.shape[0], 10)
assert_equal(n_important, 3)
X_new = clf.transform(X, threshold="mean")
assert_less(0 < X_new.shape[1], X.shape[1])
def test_collinearity():
"""Check that lars_path is robust to collinearity in input"""
X = np.array([[3., 3., 1.],
[2., 2., 0.],
[1., 1., 0]])
y = np.array([1., 0., 0])
_, _, coef_path_ = linear_model.lars_path(X, y, alpha_min=0.01)
assert_true(not np.isnan(coef_path_).any())
residual = np.dot(X, coef_path_[:, -1]) - y
assert_less((residual ** 2).sum(), 1.) # just make sure it's bounded
n_samples = 10
X = np.random.rand(n_samples, 5)
y = np.zeros(n_samples)
_, _, coef_path_ = linear_model.lars_path(X, y, Gram='auto', copy_X=False,
copy_Gram=False, alpha_min=0.,
method='lasso', verbose=0,
max_iter=500)
assert_array_almost_equal(coef_path_, np.zeros_like(coef_path_))
def test_ransac_max_trials():
base_estimator = LinearRegression()
ransac_estimator = RANSACRegressor(base_estimator,
min_samples=2,
residual_threshold=5,
max_trials=0,
random_state=0)
assert_raises(ValueError, ransac_estimator.fit, X, y)
# there is a 1e-9 chance it will take these many trials. No good reason
# 1e-2 isn't enough, can still happen
# 2 is the what ransac defines as min_samples = X.shape[1] + 1
max_trials = _dynamic_max_trials(
len(X) - len(outliers), X.shape[0], 2, 1 - 1e-9)
ransac_estimator = RANSACRegressor(base_estimator, min_samples=2)
for i in range(50):
ransac_estimator.set_params(min_samples=2, random_state=i)
ransac_estimator.fit(X, y)
assert_less(ransac_estimator.n_trials_, max_trials + 1)
def test_KL_projection():
rng = np.random.RandomState(0)
for poly in (
UnitCube(),
ProbabilitySimplex(),
Knapsack(max_labels=4),
Knapsack(max_labels=4, min_labels=2),
Birkhoff(max_iter=3000, tol=1e-5),
CartesianProduct(ProbabilitySimplex()),
#Permutahedron(), # not implemented
#OrderSimplex(), # not implemented
):
for func in (
rng.randn,
rng.rand,
):
if isinstance(poly, (Birkhoff, CartesianProduct)):
theta = func(100) * 5
else:
theta = func(10) * 5
sol = project_fw(
theta,
poly.argmax,
#variant="pairwise",
projection_type="KL",
max_iter=1000,
tol=1e-9)
# Test correctness.
sol_proj = poly.KL_project(theta)
error = np.mean((sol - sol_proj)**2)
assert_less(error, 1e-3)
# Test vectorization.
sol_proj2 = poly.KL_project([theta, theta])
assert_equal(len(sol_proj2), 2)
assert_array_almost_equal(sol_proj2[0], sol_proj)
assert_array_almost_equal(sol_proj2[1], sol_proj)
def test_lasso_lars_ic():
# Test the LassoLarsIC object by checking that
# - some good features are selected.
# - alpha_bic > alpha_aic
# - n_nonzero_bic < n_nonzero_aic
lars_bic = linear_model.LassoLarsIC('bic')
lars_aic = linear_model.LassoLarsIC('aic')
rng = np.random.RandomState(42)
X = diabetes.data
X = np.c_[X, rng.randn(X.shape[0], 5)] # add 5 bad features
lars_bic.fit(X, y)
lars_aic.fit(X, y)
nonzero_bic = np.where(lars_bic.coef_)[0]
nonzero_aic = np.where(lars_aic.coef_)[0]
assert_greater(lars_bic.alpha_, lars_aic.alpha_)
assert_less(len(nonzero_bic), len(nonzero_aic))
assert_less(np.max(nonzero_bic), diabetes.data.shape[1])
# test error on unknown IC
lars_broken = linear_model.LassoLarsIC('<unknown>')
assert_raises(ValueError, lars_broken.fit, X, y)
def test_make_matrix_words(self):
"""Test words in each doc are in the same topic
"""
n_topics = self.rand.randint(100, 200)
words_per_topic = 30
words_per_doc = self.rand.randint(10, 20)
params = {
'n_topics': n_topics,
'words_per_topic': words_per_topic,
'docs_per_topic': 100,
'words_per_doc': words_per_doc
}
matrix = make_doc_word_matrix(**params)
dense = matrix.toarray()
for i in xrange(dense.shape[0]):
col_idx = np.where(dense[i, :] > 0)[0]
max_idx = np.max(col_idx)
min_idx = np.min(col_idx)
assert_less(max_idx - min_idx, words_per_topic)
def test_feature_importances():
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)
est = RandomForestClassifier(n_estimators=50, random_state=0)
for threshold, func in zip(["mean", "median"], [np.mean, np.median]):
transformer = SelectFromModel(estimator=est, threshold=threshold)
transformer.fit(X, y)
assert_true(hasattr(transformer.estimator_, 'feature_importances_'))
X_new = transformer.transform(X)
assert_less(X_new.shape[1], X.shape[1])
importances = transformer.estimator_.feature_importances_
feature_mask = np.abs(importances) > func(importances)
assert_array_almost_equal(X_new, X[:, feature_mask])
def test_stacked_regressor(self):
bclf = LinearRegression()
clfs = [RandomForestRegressor(n_estimators=50, random_state=1),
GradientBoostingRegressor(n_estimators=25, random_state=1),
Ridge(random_state=1)]
# Friedman1
X, y = datasets.make_friedman1(n_samples=1200,
random_state=1,
noise=1.0)
X_train, y_train = X[:200], y[:200]
X_test, y_test = X[200:], y[200:]
sr = StackedRegressor(bclf,
clfs,
n_folds=3,
verbose=0,
oob_score_flag=True)
sr.fit(X_train, y_train)
mse = mean_squared_error(y_test, sr.predict(X_test))
assert_less(mse, 6.0)
def test_nested_circles():
# Test the linear separability of the first 2D KPCA transform
X, y = make_circles(n_samples=400, factor=.3, noise=.05,
random_state=0)
# 2D nested circles are not linearly separable
train_score = Perceptron().fit(X, y).score(X, y)
assert_less(train_score, 0.8)
# Project the circles data into the first 2 components of a RBF Kernel
# PCA model.
# Note that the gamma value is data dependent. If this test breaks
# and the gamma value has to be updated, the Kernel PCA example will
# have to be updated too.
kpca = KernelPCA(kernel="rbf", n_components=2,
fit_inverse_transform=True, gamma=2.)
X_kpca = kpca.fit_transform(X)
# The data is perfectly linearly separable in that space
train_score = Perceptron().fit(X_kpca, y).score(X_kpca, y)
assert_equal(train_score, 1.0)
def check_oob_score(name, X, y, n_estimators=20):
# Check that oob prediction is a good estimation of the generalization
# error.
# Proper behavior
est = FOREST_ESTIMATORS[name](oob_score=True, random_state=0,
n_estimators=n_estimators, bootstrap=True)
n_samples = X.shape[0]
est.fit(X[:n_samples // 2, :], y[:n_samples // 2])
test_score = est.score(X[n_samples // 2:, :], y[n_samples // 2:])
if name in FOREST_CLASSIFIERS:
assert_less(abs(test_score - est.oob_score_), 0.1)
else:
assert_greater(test_score, est.oob_score_)
assert_greater(est.oob_score_, .8)
# Check warning if not enough estimators
with np.errstate(divide="ignore", invalid="ignore"):
est = FOREST_ESTIMATORS[name](oob_score=True, random_state=0,
n_estimators=1, bootstrap=True)
assert_warns(UserWarning, est.fit, X, y)
def test_rank_deficient_design():
# consistency test that checks that LARS Lasso is handling rank
# deficient input data (with n_features < rank) in the same way
# as coordinate descent Lasso
y = [5, 0, 5]
for X in (
[[5, 0], [0, 5], [10, 10]],
[[10, 10, 0], [1e-32, 0, 0], [0, 0, 1]],
):
# To be able to use the coefs to compute the objective function,
# we need to turn off normalization
lars = linear_model.LassoLars(.1, normalize=False)
coef_lars_ = lars.fit(X, y).coef_
obj_lars = (1. /
(2. * 3.) * linalg.norm(y - np.dot(X, coef_lars_))**2 +
.1 * linalg.norm(coef_lars_, 1))
coord_descent = linear_model.Lasso(.1, tol=1e-6, normalize=False)
coef_cd_ = coord_descent.fit(X, y).coef_
obj_cd = ((1. / (2. * 3.)) * linalg.norm(y - np.dot(X, coef_cd_))**2 +
.1 * linalg.norm(coef_cd_, 1))
assert_less(obj_lars, obj_cd * (1. + 1e-8))
def test_lasso_lars_vs_lasso_cd(verbose=False):
# Test that LassoLars and Lasso using coordinate descent give the
# same results.
X = 3 * diabetes.data
alphas, _, lasso_path = linear_model.lars_path(X, y, method='lasso')
lasso_cd = linear_model.Lasso(fit_intercept=False, tol=1e-8)
for c, a in zip(lasso_path.T, alphas):
if a == 0:
continue
lasso_cd.alpha = a
lasso_cd.fit(X, y)
error = linalg.norm(c - lasso_cd.coef_)
assert_less(error, 0.01)
# similar test, with the classifiers
for alpha in np.linspace(1e-2, 1 - 1e-2, 20):
clf1 = linear_model.LassoLars(alpha=alpha, normalize=False).fit(X, y)
clf2 = linear_model.Lasso(alpha=alpha, tol=1e-8,
normalize=False).fit(X, y)
err = linalg.norm(clf1.coef_ - clf2.coef_)
assert_less(err, 1e-3)
# same test, with normalized data
X = diabetes.data
alphas, _, lasso_path = linear_model.lars_path(X, y, method='lasso')
lasso_cd = linear_model.Lasso(fit_intercept=False, normalize=True,
tol=1e-8)
for c, a in zip(lasso_path.T, alphas):
if a == 0:
continue
lasso_cd.alpha = a
lasso_cd.fit(X, y)
error = linalg.norm(c - lasso_cd.coef_)
assert_less(error, 0.01)
def test_feature_importances():
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)
est = RandomForestClassifier(n_estimators=50, random_state=0)
for threshold, func in zip(["mean", "median"], [np.mean, np.median]):
transformer = SelectFromModel(estimator=est, threshold=threshold)
transformer.fit(X, y)
assert_true(hasattr(transformer.estimator_, 'feature_importances_'))
X_new = transformer.transform(X)
assert_less(X_new.shape[1], X.shape[1])
importances = transformer.estimator_.feature_importances_
feature_mask = np.abs(importances) > func(importances)
assert_array_almost_equal(X_new, X[:, feature_mask])
# Check with sample weights
sample_weight = np.ones(y.shape)
sample_weight[y == 1] *= 100
est = RandomForestClassifier(n_estimators=50, random_state=0)
transformer = SelectFromModel(estimator=est)
transformer.fit(X, y, sample_weight=sample_weight)
importances = transformer.estimator_.feature_importances_
transformer.fit(X, y, sample_weight=3 * sample_weight)
importances_bis = transformer.estimator_.feature_importances_
assert_almost_equal(importances, importances_bis)
# For the Lasso and related models, the threshold defaults to 1e-5
transformer = SelectFromModel(estimator=Lasso(alpha=0.1))
transformer.fit(X, y)
X_new = transformer.transform(X)
mask = np.abs(transformer.estimator_.coef_) > 1e-5
assert_array_equal(X_new, X[:, mask])
def test_lle_manifold():
rng = np.random.RandomState(0)
# similar test on a slightly more complex manifold
X = np.array(list(product(range(20), repeat=2)))
X = np.c_[X, X[:, 0]**2 / 20]
X = X + 1e-10 * rng.uniform(size=X.shape)
n_components = 2
clf = manifold.LocallyLinearEmbedding(n_neighbors=5,
n_components=n_components,
random_state=0)
tol = 1.5
N = barycenter_kneighbors_graph(X, clf.n_neighbors).toarray()
reconstruction_error = np.linalg.norm(np.dot(N, X) - X)
assert_less(reconstruction_error, tol)
for solver in eigen_solvers:
clf.set_params(eigen_solver=solver)
clf.fit(X)
assert_true(clf.embedding_.shape[1] == n_components)
reconstruction_error = np.linalg.norm(
np.dot(N, clf.embedding_) - clf.embedding_, 'fro')**2
details = "solver: " + solver
assert_less(reconstruction_error, tol, msg=details)
assert_less(np.abs(clf.reconstruction_error_ - reconstruction_error),
tol * reconstruction_error,
msg=details)
def test_lle_simple_grid():
rng = np.random.RandomState(0)
# grid of equidistant points in 2D, n_components = n_dim
X = np.array(list(product(range(5), repeat=2)))
X = X + 1e-10 * rng.uniform(size=X.shape)
n_components = 2
clf = manifold.LocallyLinearEmbedding(n_neighbors=5,
n_components=n_components)
tol = .1
N = barycenter_kneighbors_graph(X, clf.n_neighbors).todense()
reconstruction_error = np.linalg.norm(np.dot(N, X) - X, 'fro')
assert_less(reconstruction_error, tol)
for solver in eigen_solvers:
clf.set_params(eigen_solver=solver)
clf.fit(X)
assert_true(clf.embedding_.shape[1] == n_components)
reconstruction_error = np.linalg.norm(
np.dot(N, clf.embedding_) - clf.embedding_, 'fro')**2
# FIXME: ARPACK fails this test ...
if solver != 'arpack':
assert_less(reconstruction_error, tol)
assert_almost_equal(clf.reconstruction_error_,
reconstruction_error,
decimal=4)
# re-embed a noisy version of X using the transform method
noise = rng.randn(*X.shape) / 100
X_reembedded = clf.transform(X + noise)
assert_less(np.linalg.norm(X_reembedded - clf.embedding_), tol)
def test_lle_simple_grid():
# note: ARPACK is numerically unstable, so this test will fail for
# some random seeds. We choose 2 because the tests pass.
rng = np.random.RandomState(2)
tol = 0.1
# grid of equidistant points in 2D, n_components = n_dim
X = np.array(list(product(range(5), repeat=2)))
X = X + 1e-10 * rng.uniform(size=X.shape)
n_components = 2
clf = manifold.LocallyLinearEmbedding(n_neighbors=5,
n_components=n_components,
random_state=rng)
tol = 0.1
N = barycenter_kneighbors_graph(X, clf.n_neighbors).todense()
reconstruction_error = linalg.norm(np.dot(N, X) - X, 'fro')
assert_less(reconstruction_error, tol)
for solver in eigen_solvers:
clf.set_params(eigen_solver=solver)
clf.fit(X)
assert_true(clf.embedding_.shape[1] == n_components)
reconstruction_error = linalg.norm(
np.dot(N, clf.embedding_) - clf.embedding_, 'fro')**2
assert_less(reconstruction_error, tol)
assert_almost_equal(clf.reconstruction_error_,
reconstruction_error,
decimal=1)
# re-embed a noisy version of X using the transform method
noise = rng.randn(*X.shape) / 100
X_reembedded = clf.transform(X + noise)
assert_less(linalg.norm(X_reembedded - clf.embedding_), tol)
def test_goodness_of_fit():
"""Check the goodness of fit computation."""
for name, TreeEstimator in UNSCLF_TREES.items():
X = np.asarray(iris.data, dtype=np.float64)
Xl = X[:X.shape[0] // 2]
Xr = X[X.shape[0] // 2:]
y = iris.target
est1 = TreeEstimator(random_state=0)
est2 = TreeEstimator(random_state=0)
# check exceptions
assert_raises(NotFittedError, est1.goodness_of_fit, X,
"Failed to raise NotFittedError with {0}".format(name))
assert_raises(ValueError, est1.goodness_of_fit, y,
"Failed to raise ValueError with {0}".format(name))
est1.fit(X)
est2.fit(Xl)
# check that using more training samples leads to a better fit
assert_less(est2.goodness_of_fit(Xr, 'maximum'),
est1.goodness_of_fit(Xr, 'maximum'))
#"Failed `kolmogorov_smirnov` with {0}".format(name))
assert_less(est2.goodness_of_fit(Xr, 'mean_squared_error'),
est1.goodness_of_fit(Xr, 'mean_squared_error'))
#"Failed `mean_squared_error` with {0}".format(name))
assert_less(est2.goodness_of_fit(Xr, 'mean_squared_error_weighted'),
est1.goodness_of_fit(Xr, 'mean_squared_error_weighted'))
def test_lle_manifold():
rng = np.random.RandomState(0)
# similar test on a slightly more complex manifold
X = np.array(list(product(np.arange(18), repeat=2)))
X = np.c_[X, X[:, 0]**2 / 18]
X = X + 1e-10 * rng.uniform(size=X.shape)
n_components = 2
for method in ["standard", "hessian", "modified", "ltsa"]:
clf = manifold.LocallyLinearEmbedding(n_neighbors=6,
n_components=n_components,
method=method,
random_state=0)
tol = 1.5 if method == "standard" else 3
N = barycenter_kneighbors_graph(X, clf.n_neighbors).toarray()
reconstruction_error = linalg.norm(np.dot(N, X) - X)
assert_less(reconstruction_error, tol)
for solver in eigen_solvers:
clf.set_params(eigen_solver=solver)
clf.fit(X)
assert_true(clf.embedding_.shape[1] == n_components)
reconstruction_error = linalg.norm(
np.dot(N, clf.embedding_) - clf.embedding_, 'fro')**2
details = ("solver: %s, method: %s" % (solver, method))
assert_less(reconstruction_error, tol, msg=details)
assert_less(np.abs(clf.reconstruction_error_ -
reconstruction_error),
tol * reconstruction_error,
msg=details)
def check_non_linear_models(name,
use_feature_hashing=False,
use_rescaling=False):
# create a FeatureSet object with the data we want to use
if use_feature_hashing:
train_fs, test_fs, weightdict = make_regression_data(
num_examples=5000,
num_features=10,
use_feature_hashing=True,
feature_bins=5)
else:
train_fs, test_fs, weightdict = make_regression_data(num_examples=2000,
num_features=3)
# create the learner
if use_rescaling:
name = 'Rescaled' + name
learner = Learner(name)
# train it with the training feature set we created
# make sure to set the grid objective to pearson
learner.train(train_fs, grid_objective='pearson')
# Note that we cannot check the feature weights here
# since `model_params()` is not defined for non-linear
# kernels.
# now generate the predictions on the test FeatureSet
predictions = learner.predict(test_fs)
# now make sure that the predictions are close to
# the actual test FeatureSet labels that we generated
# using make_regression_data. To do this, we just
# make sure that they are correlated with pearson > 0.95
cor, _ = pearsonr(predictions, test_fs.labels)
expected_cor_range = [0.7, 0.8] if use_feature_hashing else [0.9, 1.0]
assert_greater(cor, expected_cor_range[0])
assert_less(cor, expected_cor_range[1])
def test_randomized_pca_inverse():
# Test that randomized PCA is inversible on dense data
rng = np.random.RandomState(0)
n, p = 50, 3
X = rng.randn(n, p) # spherical data
X[:, 1] *= .00001 # make middle component relatively small
X += [5, 4, 3] # make a large mean
# same check that we can find the original data from the transformed signal
# (since the data is almost of rank n_components)
pca = PCA(n_components=2, svd_solver='randomized', random_state=0).fit(X)
Y = pca.transform(X)
Y_inverse = pca.inverse_transform(Y)
assert_almost_equal(X, Y_inverse, decimal=2)
# same as above with whitening (approximate reconstruction)
pca = PCA(n_components=2, whiten=True, svd_solver='randomized',
random_state=0).fit(X)
Y = pca.transform(X)
Y_inverse = pca.inverse_transform(Y)
relative_max_delta = (np.abs(X - Y_inverse) / np.abs(X).mean()).max()
assert_less(relative_max_delta, 1e-5)
def test_importances():
"""Check variable importances."""
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)
for name, Tree in CLF_TREES.items():
clf = Tree(random_state=0)
clf.fit(X, y)
importances = clf.feature_importances_
n_important = np.sum(importances > 0.1)
assert_equal(importances.shape[0], 10, "Failed with {0}".format(name))
assert_equal(n_important, 3, "Failed with {0}".format(name))
X_new = clf.transform(X, threshold="mean")
assert_less(0, X_new.shape[1], "Failed with {0}".format(name))
assert_less(X_new.shape[1], X.shape[1], "Failed with {0}".format(name))
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_true(hasattr(transformer.estimator_, 'coef_'))
X_new = transformer.transform(X)
assert_less(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_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, zero_one_score, cv)
assert_greater(score, 0.9)
np.testing.assert_almost_equal(pvalue, 0.0, 1)
score_label, _, pvalue_label = cval.permutation_test_score(
svm, X, y, zero_one_score, cv, labels=np.ones(y.size), random_state=0)
assert_true(score_label == score)
assert_true(pvalue_label == pvalue)
# check that we obtain the same results with a sparse representation
svm_sparse = SVC(kernel='linear')
cv_sparse = cval.StratifiedKFold(y, 2, indices=True)
score_label, _, pvalue_label = cval.permutation_test_score(
svm_sparse, X_sparse, y, zero_one_score, cv_sparse,
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,
zero_one_score, cv)
assert_less(score, 0.5)
assert_greater(pvalue, 0.4)
def check_importances(name, criterion, X, y):
ForestEstimator = FOREST_ESTIMATORS[name]
est = ForestEstimator(n_estimators=20, criterion=criterion, random_state=0)
est.fit(X, y)
importances = est.feature_importances_
n_important = np.sum(importances > 0.1)
assert_equal(importances.shape[0], 10)
assert_equal(n_important, 3)
# XXX: Remove this test in 0.19 after transform support to estimators
# is removed.
X_new = assert_warns(DeprecationWarning,
est.transform,
X,
threshold="mean")
assert_less(0 < X_new.shape[1], X.shape[1])
# Check with parallel
importances = est.feature_importances_
est.set_params(n_jobs=2)
importances_parrallel = est.feature_importances_
assert_array_almost_equal(importances, importances_parrallel)
# Check with sample weights
sample_weight = check_random_state(0).randint(1, 10, len(X))
est = ForestEstimator(n_estimators=20, random_state=0, criterion=criterion)
est.fit(X, y, sample_weight=sample_weight)
importances = est.feature_importances_
assert_true(np.all(importances >= 0.0))
for scale in [0.5, 10, 100]:
est = ForestEstimator(n_estimators=20,
random_state=0,
criterion=criterion)
est.fit(X, y, sample_weight=scale * sample_weight)
importances_bis = est.feature_importances_
assert_less(np.abs(importances - importances_bis).mean(), 0.001)
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_equal(importances.shape[0], 10)
assert_equal(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_less(np.abs(importances - importances_bis).mean(), tolerance)
def check_classification_synthetic(presort, loss):
# Test GradientBoostingClassifier on synthetic dataset used by
# Hastie et al. in ESLII Example 12.7.
X, y = datasets.make_hastie_10_2(n_samples=12000, random_state=1)
X_train, X_test = X[:2000], X[2000:]
y_train, y_test = y[:2000], y[2000:]
gbrt = GradientBoostingClassifier(n_estimators=100, min_samples_split=2,
max_depth=1, loss=loss,
learning_rate=1.0, random_state=0)
gbrt.fit(X_train, y_train)
error_rate = (1.0 - gbrt.score(X_test, y_test))
assert_less(error_rate, 0.09)
gbrt = GradientBoostingClassifier(n_estimators=200, min_samples_split=2,
max_depth=1, loss=loss,
learning_rate=1.0, subsample=0.5,
random_state=0,
presort=presort)
gbrt.fit(X_train, y_train)
error_rate = (1.0 - gbrt.score(X_test, y_test))
assert_less(error_rate, 0.08)