def test_fit(self, boston_X, boston_y):
forest = RangerForestRegressor()
with pytest.raises(NotFittedError):
check_is_fitted(forest)
forest.fit(boston_X, boston_y)
check_is_fitted(forest)
assert hasattr(forest, "ranger_forest_")
assert hasattr(forest, "n_features_in_")
def test_verbose(self, boston_X, boston_y, verbose, capfd):
forest = RangerForestRegressor(verbose=verbose)
forest.fit(boston_X, boston_y)
captured = capfd.readouterr()
if verbose:
assert len(captured.out) > 0
else:
assert len(captured.out) == 0
def test_sample_fraction_replace(self, boston_X, boston_y, replace):
forest = RangerForestRegressor(replace=replace)
forest.fit(boston_X, boston_y)
if replace:
assert forest.sample_fraction_ == [1.0]
else:
assert forest.sample_fraction_ == [0.632]
def test_sample_fraction(self, boston_X, boston_y):
forest = RangerForestRegressor(sample_fraction=0.69)
forest.fit(boston_X, boston_y)
assert forest.sample_fraction_ == [0.69]
# test with single record
boston_X_record = boston_X[0:1, :]
pred = forest.predict(boston_X_record)
assert len(pred) == 1
def test_serialize(self, boston_X, boston_y):
tf = tempfile.TemporaryFile()
forest = RangerForestRegressor()
forest.fit(boston_X, boston_y)
pickle.dump(forest, tf)
tf.seek(0)
new_forest = pickle.load(tf)
pred = new_forest.predict(boston_X)
assert len(pred) == boston_X.shape[0]
def test_shap_regressor(boston_X, boston_y):
from shap import TreeExplainer
forest = RangerForestRegressor(enable_tree_details=True)
forest.fit(boston_X, boston_y)
with shap_patch():
explainer = TreeExplainer(model=forest)
shap_values = explainer.shap_values(boston_X)
print(shap_values)
def test_plot():
from matplotlib import pyplot as plt
from sklearn.datasets import load_boston
from sklearn.tree import plot_tree
boston_X, boston_y = load_boston(return_X_y=True)
forest = RangerForestRegressor(enable_tree_details=True)
forest.fit(boston_X, boston_y)
estimator = forest.get_estimator(0)
plt.figure()
plot_tree(
estimator,
impurity=False, # impurity not yet implemented
)
plt.savefig(
"tree.svg",
bbox_inches="tight", # don't truncate
)
def test_estimators_(self, boston_X, boston_y):
forest = RangerForestRegressor(n_estimators=10)
with pytest.raises(AttributeError):
_ = forest.estimators_
forest.fit(boston_X, boston_y)
with pytest.raises(ValueError):
_ = forest.estimators_
forest = RangerForestRegressor(n_estimators=10, enable_tree_details=True)
forest.fit(boston_X, boston_y)
estimators = forest.estimators_
assert len(estimators) == 10
assert isinstance(estimators[0], RangerTreeRegressor)
check_is_fitted(estimators[0])
def test_importance(
self,
boston_X,
boston_y,
importance,
scale_permutation_importance,
local_importance,
):
forest = RangerForestRegressor(
importance=importance,
scale_permutation_importance=scale_permutation_importance,
local_importance=local_importance,
)
if importance not in ["none", "impurity", "impurity_corrected", "permutation"]:
with pytest.raises(ValueError):
forest.fit(boston_X, boston_y)
return
forest.fit(boston_X, boston_y)
if importance == "none":
assert forest.importance_mode_ == 0
elif importance == "impurity":
assert forest.importance_mode_ == 1
elif importance == "impurity_corrected":
assert forest.importance_mode_ == 5
elif importance == "permutation":
if local_importance:
assert forest.importance_mode_ == 6
elif scale_permutation_importance:
assert forest.importance_mode_ == 2
else:
assert forest.importance_mode_ == 3
def test_categorical_features(self, boston_X, boston_y,
respect_categorical_features):
# add a categorical feature
categorical_col = np.atleast_2d(
np.array([random.choice([0, 1])
for _ in range(boston_X.shape[0])]))
boston_X_c = np.hstack((boston_X, categorical_col.transpose()))
categorical_features = [boston_X.shape[1]]
rfr = RangerForestRegressor(
respect_categorical_features=respect_categorical_features,
categorical_features=categorical_features,
)
if respect_categorical_features not in [
"partition", "ignore", "order"
]:
with pytest.raises(ValueError):
rfr.fit(boston_X_c, boston_y)
return
rfr.fit(boston_X_c, boston_y)
if respect_categorical_features in ("ignore", "order"):
assert rfr.categorical_features_ == []
else:
assert rfr.categorical_features_ == [
str(c).encode() for c in categorical_features
]
def test_mtry(self, boston_X, boston_y, mtry):
forest = RangerForestRegressor(mtry=mtry)
if callable(mtry) and mtry(5) > 5:
with pytest.raises(ValueError):
forest.fit(boston_X, boston_y)
return
elif not callable(mtry) and (mtry < 0 or mtry > boston_X.shape[0]):
with pytest.raises(ValueError):
forest.fit(boston_X, boston_y)
return
forest.fit(boston_X, boston_y)
if callable(mtry):
assert forest.mtry_ == mtry(boston_X.shape[1])
else:
assert forest.mtry_ == mtry
def test_predict(self, boston_X, boston_y):
forest = RangerForestRegressor()
forest.fit(boston_X, boston_y)
pred = forest.predict(boston_X)
assert len(pred) == boston_X.shape[0]
# test with single record
boston_X_record = boston_X[0:1, :]
pred = forest.predict(boston_X_record)
assert len(pred) == 1
def test_categorical_features(
self, boston_X, boston_y, respect_categorical_features
):
# add a categorical feature
categorical_col = np.atleast_2d(
np.array([random.choice([0, 1]) for _ in range(boston_X.shape[0])])
)
boston_X_c = np.hstack((boston_X, categorical_col.transpose()))
categorical_features = [boston_X.shape[1]]
forest = RangerForestRegressor(
respect_categorical_features=respect_categorical_features,
categorical_features=categorical_features,
)
if respect_categorical_features not in ["partition", "ignore", "order"]:
with pytest.raises(ValueError):
forest.fit(boston_X_c, boston_y)
return
forest.fit(boston_X_c, boston_y)
forest.predict(boston_X_c)
def test_feature_importances_(self, boston_X, boston_y, importance,
local_importance):
rfr = RangerForestRegressor(importance=importance,
local_importance=local_importance)
with pytest.raises(AttributeError):
_ = rfr.feature_importances_
if importance == "INVALID":
with pytest.raises(ValueError):
rfr.fit(boston_X, boston_y)
return
rfr.fit(boston_X, boston_y)
if importance == "none":
with pytest.raises(ValueError):
_ = rfr.feature_importances_
else:
assert len(rfr.feature_importances_) == boston_X.shape[1]
def test_inbag(self, boston_X, boston_y):
inbag = [[1, 2, 3], [2, 3, 4]]
forest = RangerForestRegressor(n_estimators=2, inbag=inbag)
forest.fit(boston_X, boston_y)
# inbag list different length from n_estimators
forest = RangerForestRegressor(n_estimators=1, inbag=inbag)
with pytest.raises(ValueError):
forest.fit(boston_X, boston_y)
# can't use inbag with sample weight
forest = RangerForestRegressor(inbag=inbag)
with pytest.raises(ValueError):
forest.fit(boston_X, boston_y, sample_weight=[1] * len(boston_y))
# can't use class sampling and inbag
forest = RangerForestRegressor(inbag=inbag, sample_fraction=[1, 1])
with pytest.raises(ValueError):
forest.fit(boston_X, boston_y)
def test_clone(self, boston_X, boston_y):
forest = RangerForestRegressor()
forest.fit(boston_X, boston_y)
clone(forest)
def test_check_estimator(self):
check_estimator(RangerForestRegressor())
def test_get_estimator(self, boston_X, boston_y):
forest = RangerForestRegressor(n_estimators=10)
with pytest.raises(NotFittedError):
_ = forest.get_estimator(idx=0)
forest.fit(boston_X, boston_y)
with pytest.raises(ValueError):
_ = forest.get_estimator(0)
forest = RangerForestRegressor(n_estimators=10, enable_tree_details=True)
forest.fit(boston_X, boston_y)
estimator = forest.get_estimator(0)
estimator.predict(boston_X)
assert isinstance(estimator, RangerTreeRegressor)
with pytest.raises(IndexError):
_ = forest.get_estimator(idx=20)
def test_init(self):
_ = RangerForestRegressor()
def test_quantile_regression(self, boston_X, boston_y):
X_train, X_test, y_train, y_test = train_test_split(boston_X, boston_y)
forest = RangerForestRegressor(quantiles=False)
forest.fit(X_train, y_train)
assert not hasattr(forest, "random_node_values_")
with pytest.raises(ValueError):
forest.predict_quantiles(X_test, quantiles=[0.2, 0.5, 0.8])
forest = RangerForestRegressor(quantiles=True)
forest.fit(X_train, y_train)
assert hasattr(forest, "random_node_values_")
quantiles_lower = forest.predict_quantiles(X_test, quantiles=[0.1])
quantiles_upper = forest.predict_quantiles(X_test, quantiles=[0.9])
assert np.less(quantiles_lower, quantiles_upper).all()
assert quantiles_upper.ndim == 1
quantiles = forest.predict_quantiles(X_test, quantiles=[0.1, 0.9])
assert quantiles.shape == (X_test.shape[0], 2)
assert np.sum(np.isnan(quantiles_lower)) == 0
assert np.sum(np.isnan(quantiles_upper)) == 0
# test predict method
pred = forest.predict(X_test, quantiles=[0.2, 0.5])
assert pred.shape == (X_test.shape[0], 2)
assert np.sum(np.isnan(pred)) == 0
pred = forest.predict(X_test, quantiles=[0.2])
assert pred.ndim == 1
assert np.sum(np.isnan(pred)) == 0
# test with single record
boston_X_record = boston_X[0:1, :]
pred = forest.predict(boston_X_record, quantiles=[0.2, 0.5])
assert pred.shape == (1, 2)
assert np.sum(np.isnan(pred)) == 0
def test_always_split_features(self, boston_X, boston_y):
forest = RangerForestRegressor(always_split_features=[0])
forest.fit(boston_X, boston_y)
# feature 0 is in every tree split
for tree in forest.ranger_forest_["forest"]["split_var_ids"]:
assert 0 in tree
def test_clone(self, boston_X, boston_y):
rfr = RangerForestRegressor()
rfr.fit(boston_X, boston_y)
clone(rfr)
def test_regularization(self, boston_X, boston_y):
forest = RangerForestRegressor()
forest.fit(boston_X, boston_y)
assert forest.regularization_factor_ == []
assert not forest.use_regularization_factor_
# vector must be between 0 and 1 and length matching feature num
for r in [[1.1], [-0.1], [1, 1]]:
forest = RangerForestRegressor(regularization_factor=r)
with pytest.raises(ValueError):
forest.fit(boston_X, boston_y)
# vector of ones isn't applied
forest = RangerForestRegressor(regularization_factor=[1] * boston_X.shape[1])
forest.fit(boston_X, boston_y)
assert forest.regularization_factor_ == []
assert not forest.use_regularization_factor_
# regularization vector is used
reg = [0.5]
forest = RangerForestRegressor(regularization_factor=reg, n_jobs=2)
# warns if n_jobs is not one since parallelization can't be used
with pytest.warns(Warning):
forest.fit(boston_X, boston_y)
assert forest.n_jobs_ == 1
assert forest.regularization_factor_ == reg
assert forest.use_regularization_factor_
def test_split_select_weights(self, boston_X, boston_y):
n_trees = 10
weights = [0.1] * boston_X.shape[1]
forest = RangerForestRegressor(n_estimators=n_trees)
forest.fit(boston_X, boston_y, split_select_weights=weights)
weights = [0.1] * (boston_X.shape[1] - 1)
forest = RangerForestRegressor(n_estimators=n_trees)
with pytest.raises(RuntimeError):
forest.fit(boston_X, boston_y, split_select_weights=weights)
weights = [[0.1] * (boston_X.shape[1])] * n_trees
forest = RangerForestRegressor(n_estimators=n_trees)
forest.fit(boston_X, boston_y, split_select_weights=weights)
weights = [[0.1] * (boston_X.shape[1])] * (n_trees + 1)
forest = RangerForestRegressor(n_estimators=n_trees)
with pytest.raises(RuntimeError):
forest.fit(boston_X, boston_y, split_select_weights=weights)
def test_quantile_regression(self, boston_X, boston_y):
X_train, X_test, y_train, y_test = train_test_split(boston_X, boston_y)
rfr = RangerForestRegressor(quantiles=False)
rfr.fit(X_train, y_train)
assert not hasattr(rfr, "random_node_values_")
with pytest.raises(ValueError):
rfr.predict_quantiles(X_test)
rfr = RangerForestRegressor(quantiles=True)
rfr.fit(X_train, y_train)
assert hasattr(rfr, "random_node_values_")
quantiles_lower = rfr.predict_quantiles(X_test, quantiles=[0.1])
quantiles_upper = rfr.predict_quantiles(X_test, quantiles=[0.9])
assert np.less(quantiles_lower, quantiles_upper).all()
assert quantiles_upper.ndim == 1
quantiles = rfr.predict_quantiles(X_test, quantiles=[0.1, 0.9])
assert quantiles.ndim == 2
def test_importance_pvalues(self, boston_X_mod, boston_y, importance, mod):
rfc = RangerForestRegressor(importance=importance)
np.random.seed(42)
if importance not in ["none", "impurity", "impurity_corrected", "permutation"]:
with pytest.raises(ValueError):
rfc.fit(boston_X_mod, boston_y)
return
if not importance == "impurity_corrected":
rfc.fit(boston_X_mod, boston_y)
with pytest.raises(ValueError):
rfc.get_importance_pvalues()
return
# Test error for no non-negative importance values
if mod == "none":
rfc.fit(boston_X_mod, boston_y)
with pytest.raises(ValueError):
rfc.get_importance_pvalues()
return
rfc.fit(boston_X_mod, boston_y)
assert len(rfc.get_importance_pvalues()) == boston_X_mod.shape[1]
def test_split_rule(self, boston_X, boston_y, split_rule):
forest = RangerForestRegressor(split_rule=split_rule)
assert forest.criterion == split_rule
if split_rule not in ["variance", "extratrees", "maxstat", "beta"]:
with pytest.raises(ValueError):
forest.fit(boston_X, boston_y)
return
# beta can only be used with targets between 0 and 1
if split_rule == "beta":
with pytest.raises(ValueError):
forest.fit(boston_X, boston_y)
boston_01 = [0.5 for _ in boston_y]
forest.fit(boston_X, boston_01)
if split_rule == "variance":
assert forest.split_rule_ == 1
elif split_rule == "extratrees":
assert forest.split_rule_ == 5
elif split_rule == "maxstat":
assert forest.split_rule_ == 4
elif split_rule == "beta":
assert forest.split_rule_ == 6
if split_rule == "extratrees":
forest = RangerForestRegressor(
split_rule=split_rule,
respect_categorical_features="partition",
save_memory=True,
)
with pytest.raises(ValueError):
forest.fit(boston_X, boston_y)
else:
forest = RangerForestRegressor(split_rule=split_rule, num_random_splits=2)
with pytest.raises(ValueError):
forest.fit(boston_X, boston_y)
