def test_ohe_equivalence(min_samples_leaf, n_unique_categories, target):
# Make sure that native categorical splits are equivalent to using a OHE,
# when given enough depth
rng = np.random.RandomState(0)
n_samples = 10_000
X_binned = rng.randint(0,
n_unique_categories,
size=(n_samples, 1),
dtype=np.uint8)
X_ohe = OneHotEncoder(sparse=False).fit_transform(X_binned)
X_ohe = np.asfortranarray(X_ohe).astype(np.uint8)
if target == "equal":
gradients = X_binned.reshape(-1)
elif target == "binary":
gradients = (X_binned % 2).reshape(-1)
else:
gradients = rng.randn(n_samples)
gradients = gradients.astype(G_H_DTYPE)
hessians = np.ones(shape=1, dtype=G_H_DTYPE)
grower_params = {
"min_samples_leaf": min_samples_leaf,
"max_depth": None,
"max_leaf_nodes": None,
}
grower = TreeGrower(X_binned,
gradients,
hessians,
is_categorical=[True],
**grower_params)
grower.grow()
# we pass undefined bin_thresholds because we won't use predict()
predictor = grower.make_predictor(
binning_thresholds=np.zeros((1, n_unique_categories)))
preds = predictor.predict_binned(X_binned,
missing_values_bin_idx=255,
n_threads=n_threads)
grower_ohe = TreeGrower(X_ohe, gradients, hessians, **grower_params)
grower_ohe.grow()
predictor_ohe = grower_ohe.make_predictor(
binning_thresholds=np.zeros((X_ohe.shape[1], n_unique_categories)))
preds_ohe = predictor_ohe.predict_binned(X_ohe,
missing_values_bin_idx=255,
n_threads=n_threads)
assert predictor.get_max_depth() <= predictor_ohe.get_max_depth()
if target == "binary" and n_unique_categories > 2:
# OHE needs more splits to achieve the same predictions
assert predictor.get_max_depth() < predictor_ohe.get_max_depth()
np.testing.assert_allclose(preds, preds_ohe)
def test_min_samples_leaf(n_samples, min_samples_leaf, n_bins,
constant_hessian, noise):
rng = np.random.RandomState(seed=0)
# data = linear target, 3 features, 1 irrelevant.
X = rng.normal(size=(n_samples, 3))
y = X[:, 0] - X[:, 1]
if noise:
y_scale = y.std()
y += rng.normal(scale=noise, size=n_samples) * y_scale
mapper = _BinMapper(max_bins=n_bins)
X = mapper.fit_transform(X)
all_gradients = y.astype(G_H_DTYPE)
shape_hessian = 1 if constant_hessian else all_gradients.shape
all_hessians = np.ones(shape=shape_hessian, dtype=G_H_DTYPE)
grower = TreeGrower(X,
all_gradients,
all_hessians,
max_bins=n_bins,
shrinkage=1.,
min_samples_leaf=min_samples_leaf,
max_leaf_nodes=n_samples)
grower.grow()
predictor = grower.make_predictor(bin_thresholds=mapper.bin_thresholds_)
if n_samples >= min_samples_leaf:
for node in predictor.nodes:
if node['is_leaf']:
assert node['count'] >= min_samples_leaf
else:
assert predictor.nodes.shape[0] == 1
assert predictor.nodes[0]['is_leaf']
assert predictor.nodes[0]['count'] == n_samples
def test_boston_dataset(max_bins):
X, y = load_boston(return_X_y=True)
X_train, X_test, y_train, y_test = train_test_split(X, y, random_state=42)
mapper = _BinMapper(max_bins=max_bins, random_state=42)
X_train_binned = mapper.fit_transform(X_train)
# Init gradients and hessians to that of least squares loss
gradients = -y_train.astype(G_H_DTYPE)
hessians = np.ones(1, dtype=G_H_DTYPE)
min_samples_leaf = 8
max_leaf_nodes = 31
grower = TreeGrower(X_train_binned,
gradients,
hessians,
min_samples_leaf=min_samples_leaf,
max_leaf_nodes=max_leaf_nodes,
max_bins=max_bins,
actual_n_bins=mapper.actual_n_bins_)
grower.grow()
predictor = grower.make_predictor(bin_thresholds=mapper.bin_thresholds_)
assert r2_score(y_train, predictor.predict(X_train)) > 0.85
assert r2_score(y_test, predictor.predict(X_test)) > 0.70
def test_regression_dataset(n_bins):
X, y = make_regression(n_samples=500, n_features=10, n_informative=5,
random_state=42)
X_train, X_test, y_train, y_test = train_test_split(
X, y, random_state=42)
mapper = _BinMapper(n_bins=n_bins, random_state=42)
X_train_binned = mapper.fit_transform(X_train)
# Init gradients and hessians to that of least squares loss
gradients = -y_train.astype(G_H_DTYPE)
hessians = np.ones(1, dtype=G_H_DTYPE)
min_samples_leaf = 10
max_leaf_nodes = 30
grower = TreeGrower(X_train_binned, gradients, hessians,
min_samples_leaf=min_samples_leaf,
max_leaf_nodes=max_leaf_nodes, n_bins=n_bins,
n_bins_non_missing=mapper.n_bins_non_missing_)
grower.grow()
predictor = grower.make_predictor(
binning_thresholds=mapper.bin_thresholds_)
known_cat_bitsets = np.zeros((0, 8), dtype=X_BITSET_INNER_DTYPE)
f_idx_map = np.zeros(0, dtype=np.uint32)
y_pred_train = predictor.predict(X_train, known_cat_bitsets, f_idx_map)
assert r2_score(y_train, y_pred_train) > 0.82
y_pred_test = predictor.predict(X_test, known_cat_bitsets, f_idx_map)
assert r2_score(y_test, y_pred_test) > 0.67
def test_missing_value_predict_only():
# Make sure that missing values are supported at predict time even if they
# were not encountered in the training data: the missing values are
# assigned to whichever child has the most samples.
rng = np.random.RandomState(0)
n_samples = 100
X_binned = rng.randint(0, 256, size=(n_samples, 1), dtype=np.uint8)
X_binned = np.asfortranarray(X_binned)
gradients = rng.normal(size=n_samples).astype(G_H_DTYPE)
hessians = np.ones(shape=1, dtype=G_H_DTYPE)
grower = TreeGrower(X_binned, gradients, hessians, min_samples_leaf=5,
has_missing_values=False)
grower.grow()
predictor = grower.make_predictor()
# go from root to a leaf, always following node with the most samples.
# That's the path nans are supposed to take
node = predictor.nodes[0]
while not node['is_leaf']:
left = predictor.nodes[node['left']]
right = predictor.nodes[node['right']]
node = left if left['count'] > right['count'] else right
prediction_main_path = node['value']
# now build X_test with only nans, and make sure all predictions are equal
# to prediction_main_path
all_nans = np.full(shape=(n_samples, 1), fill_value=np.nan)
assert np.all(predictor.predict(all_nans) == prediction_main_path)
def test_regression_dataset(n_bins):
X, y = make_regression(n_samples=500, n_features=10, n_informative=5,
random_state=42)
X_train, X_test, y_train, y_test = train_test_split(
X, y, random_state=42)
mapper = _BinMapper(n_bins=n_bins, random_state=42)
X_train_binned = mapper.fit_transform(X_train)
# Init gradients and hessians to that of least squares loss
gradients = -y_train.astype(G_H_DTYPE)
hessians = np.ones(1, dtype=G_H_DTYPE)
min_samples_leaf = 10
max_leaf_nodes = 30
grower = TreeGrower(X_train_binned, gradients, hessians,
min_samples_leaf=min_samples_leaf,
max_leaf_nodes=max_leaf_nodes, n_bins=n_bins,
n_bins_non_missing=mapper.n_bins_non_missing_)
grower.grow()
predictor = grower.make_predictor(num_thresholds=mapper.bin_thresholds_)
assert r2_score(y_train, predictor.predict(X_train)) > 0.82
assert r2_score(y_test, predictor.predict(X_test)) > 0.67
def test_nodes_values(monotonic_cst, seed):
# Build a single tree with only one feature, and make sure the nodes
# values respect the monotonic constraints.
# Considering the following tree with a monotonic POS constraint, we
# should have:
#
# root
# / \
# 5 10 # middle = 7.5
# / \ / \
# a b c d
#
# a <= b and c <= d (assert_children_values_monotonic)
# a, b <= middle <= c, d (assert_children_values_bounded)
# a <= b <= c <= d (assert_leaves_values_monotonic)
#
# The last one is a consequence of the others, but can't hurt to check
rng = np.random.RandomState(seed)
n_samples = 1000
n_features = 1
X_binned = rng.randint(0,
255,
size=(n_samples, n_features),
dtype=np.uint8)
X_binned = np.asfortranarray(X_binned)
gradients = rng.normal(size=n_samples).astype(G_H_DTYPE)
hessians = np.ones(shape=1, dtype=G_H_DTYPE)
grower = TreeGrower(X_binned,
gradients,
hessians,
monotonic_cst=[monotonic_cst],
shrinkage=.1)
grower.grow()
# grow() will shrink the leaves values at the very end. For our comparison
# tests, we need to revert the shrinkage of the leaves, else we would
# compare the value of a leaf (shrunk) with a node (not shrunk) and the
# test would not be correct.
for leave in grower.finalized_leaves:
leave.value /= grower.shrinkage
# We pass undefined binning_thresholds because we won't use predict anyway
predictor = grower.make_predictor(
binning_thresholds=np.zeros((X_binned.shape[1], X_binned.max() + 1)))
# The consistency of the bounds can only be checked on the tree grower
# as the node bounds are not copied into the predictor tree. The
# consistency checks on the values of node children and leaves can be
# done either on the grower tree or on the predictor tree. We only
# do those checks on the predictor tree as the latter is derived from
# the former.
assert_children_values_monotonic(predictor, monotonic_cst)
assert_children_values_bounded(grower, monotonic_cst)
assert_leaves_values_monotonic(predictor, monotonic_cst)
def test_grow_tree_categories():
# Check that the grower produces the right predictor tree when a split is
# categorical
X_binned = np.array([[0, 1] * 11 + [1]], dtype=X_BINNED_DTYPE).T
X_binned = np.asfortranarray(X_binned)
all_gradients = np.array([10, 1] * 11 + [1], dtype=G_H_DTYPE)
all_hessians = np.ones(1, dtype=G_H_DTYPE)
is_categorical = np.ones(1, dtype=np.uint8)
grower = TreeGrower(X_binned, all_gradients, all_hessians,
n_bins=4, shrinkage=1.0, min_samples_leaf=1,
is_categorical=is_categorical)
grower.grow()
assert grower.n_nodes == 3
categories = [np.array([4, 9], dtype=X_DTYPE)]
predictor = grower.make_predictor(binning_thresholds=categories)
root = predictor.nodes[0]
assert root['count'] == 23
assert root['depth'] == 0
assert root['is_categorical']
left, right = predictor.nodes[root['left']], predictor.nodes[root['right']]
# arbitrary validation, but this means ones go to the left.
assert left['count'] >= right['count']
# check binned category value (1)
expected_binned_cat_bitset = [2**1] + [0] * 7
binned_cat_bitset = predictor.binned_left_cat_bitsets
assert_array_equal(binned_cat_bitset[0], expected_binned_cat_bitset)
# check raw category value (9)
expected_raw_cat_bitsets = [2**9] + [0] * 7
raw_cat_bitsets = predictor.raw_left_cat_bitsets
assert_array_equal(raw_cat_bitsets[0], expected_raw_cat_bitsets)
# Note that since there was no missing values during training, the missing
# values aren't part of the bitsets. However, we expect the missing values
# to go to the biggest child (i.e. the left one).
# The left child has a value of -1 = negative gradient.
assert root['missing_go_to_left']
# make sure binned missing values are mapped to the left child during
# prediction
prediction_binned = predictor.predict_binned(
np.asarray([[6]]).astype(X_BINNED_DTYPE), missing_values_bin_idx=6)
assert_allclose(prediction_binned, [-1]) # negative gradient
# make sure raw missing values are mapped to the left child during
# prediction
known_cat_bitsets = np.zeros((1, 8), dtype=np.uint32) # ignored anyway
f_idx_map = np.array([0], dtype=np.uint32)
prediction = predictor.predict(np.array([[np.nan]]), known_cat_bitsets,
f_idx_map)
assert_allclose(prediction, [-1])
def test_predictor_from_grower():
# Build a tree on the toy 3-leaf dataset to extract the predictor.
n_bins = 256
X_binned, all_gradients, all_hessians = _make_training_data(n_bins=n_bins)
grower = TreeGrower(
X_binned,
all_gradients,
all_hessians,
n_bins=n_bins,
shrinkage=1.0,
max_leaf_nodes=3,
min_samples_leaf=5,
)
grower.grow()
assert grower.n_nodes == 5 # (2 decision nodes + 3 leaves)
# Check that the node structure can be converted into a predictor
# object to perform predictions at scale
# We pass undefined binning_thresholds because we won't use predict anyway
predictor = grower.make_predictor(
binning_thresholds=np.zeros((X_binned.shape[1], n_bins))
)
assert predictor.nodes.shape[0] == 5
assert predictor.nodes["is_leaf"].sum() == 3
# Probe some predictions for each leaf of the tree
# each group of 3 samples corresponds to a condition in _make_training_data
input_data = np.array(
[
[0, 0],
[42, 99],
[128, 254],
[129, 0],
[129, 85],
[254, 85],
[129, 86],
[129, 254],
[242, 100],
],
dtype=np.uint8,
)
missing_values_bin_idx = n_bins - 1
predictions = predictor.predict_binned(
input_data, missing_values_bin_idx, n_threads
)
expected_targets = [1, 1, 1, 1, 1, 1, -1, -1, -1]
assert np.allclose(predictions, expected_targets)
# Check that training set can be recovered exactly:
predictions = predictor.predict_binned(X_binned, missing_values_bin_idx, n_threads)
assert np.allclose(predictions, -all_gradients)
def test_split_on_nan_with_infinite_values():
# Make sure the split on nan situations are respected even when there are
# samples with +inf values (we set the threshold to +inf when we have a
# split on nan so this test makes sure this does not introduce edge-case
# bugs). We need to use the private API so that we can also test
# predict_binned().
X = np.array([0, 1, np.inf, np.nan, np.nan]).reshape(-1, 1)
# the gradient values will force a split on nan situation
gradients = np.array([0, 0, 0, 100, 100], dtype=G_H_DTYPE)
hessians = np.ones(shape=1, dtype=G_H_DTYPE)
bin_mapper = _BinMapper()
X_binned = bin_mapper.fit_transform(X)
n_bins_non_missing = 3
has_missing_values = True
grower = TreeGrower(
X_binned,
gradients,
hessians,
n_bins_non_missing=n_bins_non_missing,
has_missing_values=has_missing_values,
min_samples_leaf=1,
n_threads=n_threads,
)
grower.grow()
predictor = grower.make_predictor(
binning_thresholds=bin_mapper.bin_thresholds_)
# sanity check: this was a split on nan
assert predictor.nodes[0]["num_threshold"] == np.inf
assert predictor.nodes[0]["bin_threshold"] == n_bins_non_missing - 1
known_cat_bitsets, f_idx_map = bin_mapper.make_known_categories_bitsets()
# Make sure in particular that the +inf sample is mapped to the left child
# Note that lightgbm "fails" here and will assign the inf sample to the
# right child, even though it's a "split on nan" situation.
predictions = predictor.predict(X, known_cat_bitsets, f_idx_map, n_threads)
predictions_binned = predictor.predict_binned(
X_binned,
missing_values_bin_idx=bin_mapper.missing_values_bin_idx_,
n_threads=n_threads,
)
np.testing.assert_allclose(predictions, -gradients)
np.testing.assert_allclose(predictions_binned, -gradients)
def test_missing_value_predict_only():
# Make sure that missing values are supported at predict time even if they
# were not encountered in the training data: the missing values are
# assigned to whichever child has the most samples.
rng = np.random.RandomState(0)
n_samples = 100
X_binned = rng.randint(0, 256, size=(n_samples, 1), dtype=np.uint8)
X_binned = np.asfortranarray(X_binned)
gradients = rng.normal(size=n_samples).astype(G_H_DTYPE)
hessians = np.ones(shape=1, dtype=G_H_DTYPE)
grower = TreeGrower(X_binned,
gradients,
hessians,
min_samples_leaf=5,
has_missing_values=False)
grower.grow()
# We pass undefined binning_thresholds because we won't use predict anyway
predictor = grower.make_predictor(
binning_thresholds=np.zeros((X_binned.shape[1], X_binned.max() + 1)))
# go from root to a leaf, always following node with the most samples.
# That's the path nans are supposed to take
node = predictor.nodes[0]
while not node["is_leaf"]:
left = predictor.nodes[node["left"]]
right = predictor.nodes[node["right"]]
node = left if left["count"] > right["count"] else right
prediction_main_path = node["value"]
# now build X_test with only nans, and make sure all predictions are equal
# to prediction_main_path
all_nans = np.full(shape=(n_samples, 1), fill_value=np.nan)
known_cat_bitsets = np.zeros((0, 8), dtype=X_BITSET_INNER_DTYPE)
f_idx_map = np.zeros(0, dtype=np.uint32)
y_pred = predictor.predict(all_nans, known_cat_bitsets, f_idx_map,
n_threads)
assert np.all(y_pred == prediction_main_path)
def test_predictor_from_grower():
# Build a tree on the toy 3-leaf dataset to extract the predictor.
n_bins = 256
X_binned, all_gradients, all_hessians = _make_training_data(n_bins=n_bins)
grower = TreeGrower(X_binned,
all_gradients,
all_hessians,
max_bins=n_bins,
shrinkage=1.,
max_leaf_nodes=3,
min_samples_leaf=5)
grower.grow()
assert grower.n_nodes == 5 # (2 decision nodes + 3 leaves)
# Check that the node structure can be converted into a predictor
# object to perform predictions at scale
predictor = grower.make_predictor()
assert predictor.nodes.shape[0] == 5
assert predictor.nodes['is_leaf'].sum() == 3
# Probe some predictions for each leaf of the tree
# each group of 3 samples corresponds to a condition in _make_training_data
input_data = np.array([
[0, 0],
[42, 99],
[128, 255],
[129, 0],
[129, 85],
[255, 85],
[129, 86],
[129, 255],
[242, 100],
],
dtype=np.uint8)
predictions = predictor.predict_binned(input_data)
expected_targets = [1, 1, 1, 1, 1, 1, -1, -1, -1]
assert np.allclose(predictions, expected_targets)
# Check that training set can be recovered exactly:
predictions = predictor.predict_binned(X_binned)
assert np.allclose(predictions, -all_gradients)
