def test_bin_mapper_idempotence(max_bins_small, max_bins_large):
assert max_bins_large >= max_bins_small
data = np.random.RandomState(42).normal(size=30000).reshape(-1, 1)
mapper_small = _BinMapper(n_bins=max_bins_small + 1)
mapper_large = _BinMapper(n_bins=max_bins_small + 1)
binned_small = mapper_small.fit_transform(data)
binned_large = mapper_large.fit_transform(binned_small)
assert_array_equal(binned_small, binned_large)
def test_subsample():
# Make sure bin thresholds are different when applying subsampling
mapper_no_subsample = _BinMapper(subsample=None, random_state=0).fit(DATA)
mapper_subsample = _BinMapper(subsample=256, random_state=0).fit(DATA)
for feature in range(DATA.shape[1]):
assert not np.allclose(mapper_no_subsample.bin_thresholds_[feature],
mapper_subsample.bin_thresholds_[feature],
rtol=1e-4)
def test_boston_dataset(n_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(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 = 8
max_leaf_nodes = 31
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(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_bin_mapper_random_data(max_bins):
n_samples, n_features = DATA.shape
expected_count_per_bin = n_samples // max_bins
tol = int(0.05 * expected_count_per_bin)
# max_bins is the number of bins for non-missing values
n_bins = max_bins + 1
mapper = _BinMapper(n_bins=n_bins, random_state=42).fit(DATA)
binned = mapper.transform(DATA)
assert binned.shape == (n_samples, n_features)
assert binned.dtype == np.uint8
assert_array_equal(binned.min(axis=0), np.array([0, 0]))
assert_array_equal(binned.max(axis=0),
np.array([max_bins - 1, max_bins - 1]))
assert len(mapper.bin_thresholds_) == n_features
for bin_thresholds_feature in mapper.bin_thresholds_:
assert bin_thresholds_feature.shape == (max_bins - 1, )
assert bin_thresholds_feature.dtype == DATA.dtype
assert np.all(mapper.n_bins_non_missing_ == max_bins)
# Check that the binned data is approximately balanced across bins.
for feature_idx in range(n_features):
for bin_idx in range(max_bins):
count = (binned[:, feature_idx] == bin_idx).sum()
assert abs(count - expected_count_per_bin) < tol
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(n_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,
n_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_same_predictions_regression(seed, min_samples_leaf, n_samples,
max_leaf_nodes):
# Make sure mrex has the same predictions as lightgbm for easy targets.
#
# In particular when the size of the trees are bound and the number of
# samples is large enough, the structure of the prediction trees found by
# LightGBM and mrex should be exactly identical.
#
# Notes:
# - Several candidate splits may have equal gains when the number of
# samples in a node is low (and because of float errors). Therefore the
# predictions on the test set might differ if the structure of the tree
# is not exactly the same. To avoid this issue we only compare the
# predictions on the test set when the number of samples is large enough
# and max_leaf_nodes is low enough.
# - To ignore discrepancies caused by small differences the binning
# strategy, data is pre-binned if n_samples > 255.
rng = np.random.RandomState(seed=seed)
n_samples = n_samples
max_iter = 1
max_bins = 255
X, y = make_regression(n_samples=n_samples,
n_features=5,
n_informative=5,
random_state=0)
if n_samples > 255:
# bin data and convert it to float32 so that the estimator doesn't
# treat it as pre-binned
X = _BinMapper(n_bins=max_bins + 1).fit_transform(X).astype(np.float32)
X_train, X_test, y_train, y_test = train_test_split(X, y, random_state=rng)
est_mrex = HistGradientBoostingRegressor(max_iter=max_iter,
max_bins=max_bins,
learning_rate=1,
n_iter_no_change=None,
min_samples_leaf=min_samples_leaf,
max_leaf_nodes=max_leaf_nodes)
est_lightgbm = get_equivalent_estimator(est_mrex, lib='lightgbm')
est_lightgbm.fit(X_train, y_train)
est_mrex.fit(X_train, y_train)
# We need X to be treated an numerical data, not pre-binned data.
X_train, X_test = X_train.astype(np.float32), X_test.astype(np.float32)
pred_lightgbm = est_lightgbm.predict(X_train)
pred_mrex = est_mrex.predict(X_train)
# less than 1% of the predictions are different up to the 3rd decimal
assert np.mean(abs(pred_lightgbm - pred_mrex) > 1e-3) < .011
if max_leaf_nodes < 10 and n_samples >= 1000:
pred_lightgbm = est_lightgbm.predict(X_test)
pred_mrex = est_mrex.predict(X_test)
# less than 1% of the predictions are different up to the 4th decimal
assert np.mean(abs(pred_lightgbm - pred_mrex) > 1e-4) < .01
def test_same_predictions_classification(seed, min_samples_leaf, n_samples,
max_leaf_nodes):
# Same as test_same_predictions_regression but for classification
rng = np.random.RandomState(seed=seed)
n_samples = n_samples
max_iter = 1
max_bins = 255
X, y = make_classification(n_samples=n_samples,
n_classes=2,
n_features=5,
n_informative=5,
n_redundant=0,
random_state=0)
if n_samples > 255:
# bin data and convert it to float32 so that the estimator doesn't
# treat it as pre-binned
X = _BinMapper(n_bins=max_bins + 1).fit_transform(X).astype(np.float32)
X_train, X_test, y_train, y_test = train_test_split(X, y, random_state=rng)
est_mrex = HistGradientBoostingClassifier(
loss='binary_crossentropy',
max_iter=max_iter,
max_bins=max_bins,
learning_rate=1,
n_iter_no_change=None,
min_samples_leaf=min_samples_leaf,
max_leaf_nodes=max_leaf_nodes)
est_lightgbm = get_equivalent_estimator(est_mrex, lib='lightgbm')
est_lightgbm.fit(X_train, y_train)
est_mrex.fit(X_train, y_train)
# We need X to be treated an numerical data, not pre-binned data.
X_train, X_test = X_train.astype(np.float32), X_test.astype(np.float32)
pred_lightgbm = est_lightgbm.predict(X_train)
pred_mrex = est_mrex.predict(X_train)
assert np.mean(pred_mrex == pred_lightgbm) > .89
acc_lightgbm = accuracy_score(y_train, pred_lightgbm)
acc_mrex = accuracy_score(y_train, pred_mrex)
np.testing.assert_almost_equal(acc_lightgbm, acc_mrex)
if max_leaf_nodes < 10 and n_samples >= 1000:
pred_lightgbm = est_lightgbm.predict(X_test)
pred_mrex = est_mrex.predict(X_test)
assert np.mean(pred_mrex == pred_lightgbm) > .89
acc_lightgbm = accuracy_score(y_test, pred_lightgbm)
acc_mrex = accuracy_score(y_test, pred_mrex)
np.testing.assert_almost_equal(acc_lightgbm, acc_mrex, decimal=2)
def test_n_bins_non_missing(n_bins, diff):
# Check that n_bins_non_missing is n_unique_values when
# there are not a lot of unique values, else n_bins - 1.
n_unique_values = n_bins + diff
X = list(range(n_unique_values)) * 2
X = np.array(X).reshape(-1, 1)
mapper = _BinMapper(n_bins=n_bins).fit(X)
assert np.all(mapper.n_bins_non_missing_ == min(n_bins -
1, n_unique_values))
def test_infinite_values():
# Make sure infinite values are properly handled.
bin_mapper = _BinMapper()
X = np.array([-np.inf, 0, 1, np.inf]).reshape(-1, 1)
bin_mapper.fit(X)
assert_allclose(bin_mapper.bin_thresholds_[0], [-np.inf, .5, ALMOST_INF])
assert bin_mapper.n_bins_non_missing_ == [4]
expected_binned_X = np.array([0, 1, 2, 3]).reshape(-1, 1)
assert_array_equal(bin_mapper.transform(X), expected_binned_X)
def test_bin_mapper_small_random_data(n_samples, max_bins):
data = np.random.RandomState(42).normal(size=n_samples).reshape(-1, 1)
assert len(np.unique(data)) == n_samples
# max_bins is the number of bins for non-missing values
n_bins = max_bins + 1
mapper = _BinMapper(n_bins=n_bins, random_state=42)
binned = mapper.fit_transform(data)
assert binned.shape == data.shape
assert binned.dtype == np.uint8
assert_array_equal(binned.ravel()[np.argsort(data.ravel())],
np.arange(n_samples))
def test_bin_mapper_repeated_values_invariance(n_distinct):
rng = np.random.RandomState(42)
distinct_values = rng.normal(size=n_distinct)
assert len(np.unique(distinct_values)) == n_distinct
repeated_indices = rng.randint(low=0, high=n_distinct, size=1000)
data = distinct_values[repeated_indices]
rng.shuffle(data)
assert_array_equal(np.unique(data), np.sort(distinct_values))
data = data.reshape(-1, 1)
mapper_1 = _BinMapper(n_bins=n_distinct + 1)
binned_1 = mapper_1.fit_transform(data)
assert_array_equal(np.unique(binned_1[:, 0]), np.arange(n_distinct))
# Adding more bins to the mapper yields the same results (same thresholds)
mapper_2 = _BinMapper(n_bins=min(256, n_distinct * 3) + 1)
binned_2 = mapper_2.fit_transform(data)
assert_allclose(mapper_1.bin_thresholds_[0], mapper_2.bin_thresholds_[0])
assert_array_equal(binned_1, binned_2)
def test_max_depth(max_depth):
# Make sure max_depth parameter works as expected
rng = np.random.RandomState(seed=0)
n_bins = 256
n_samples = 1000
# data = linear target, 3 features, 1 irrelevant.
X = rng.normal(size=(n_samples, 3))
y = X[:, 0] - X[:, 1]
mapper = _BinMapper(n_bins=n_bins)
X = mapper.fit_transform(X)
all_gradients = y.astype(G_H_DTYPE)
all_hessians = np.ones(shape=1, dtype=G_H_DTYPE)
grower = TreeGrower(X, all_gradients, all_hessians, max_depth=max_depth)
grower.grow()
depth = max(leaf.depth for leaf in grower.finalized_leaves)
assert depth == max_depth
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)
grower.grow()
predictor = grower.make_predictor(
bin_thresholds=bin_mapper.bin_thresholds_
)
# sanity check: this was a split on nan
assert predictor.nodes[0]['threshold'] == np.inf
assert predictor.nodes[0]['bin_threshold'] == n_bins_non_missing - 1
# 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)
predictions_binned = predictor.predict_binned(
X_binned, missing_values_bin_idx=bin_mapper.missing_values_bin_idx_)
assert np.all(predictions == -gradients)
assert np.all(predictions_binned == -gradients)
def test_missing_values_support(n_bins, n_bins_non_missing, X_trans_expected):
# check for missing values: make sure nans are mapped to the last bin
# and that the _BinMapper attributes are correct
X = [[1, 1, 0], [np.NaN, np.NaN, 0], [2, 1, 0], [np.NaN, 2, 1], [3, 2, 1],
[4, 1, 0]]
X = np.array(X)
mapper = _BinMapper(n_bins=n_bins)
mapper.fit(X)
assert_array_equal(mapper.n_bins_non_missing_, n_bins_non_missing)
for feature_idx in range(X.shape[1]):
assert len(mapper.bin_thresholds_[feature_idx]) == \
n_bins_non_missing[feature_idx] - 1
assert mapper.missing_values_bin_idx_ == n_bins - 1
X_trans = mapper.transform(X)
assert_array_equal(X_trans, X_trans_expected)
def test_binning_train_validation_are_separated():
# Make sure training and validation data are binned separately.
# See issue 13926
rng = np.random.RandomState(0)
validation_fraction = .2
gb = HistGradientBoostingClassifier(
n_iter_no_change=5,
validation_fraction=validation_fraction,
random_state=rng
)
gb.fit(X_classification, y_classification)
mapper_training_data = gb.bin_mapper_
# Note that since the data is small there is no subsampling and the
# random_state doesn't matter
mapper_whole_data = _BinMapper(random_state=0)
mapper_whole_data.fit(X_classification)
n_samples = X_classification.shape[0]
assert np.all(mapper_training_data.n_bins_non_missing_ ==
int((1 - validation_fraction) * n_samples))
assert np.all(mapper_training_data.n_bins_non_missing_ !=
mapper_whole_data.n_bins_non_missing_)
def test_min_samples_leaf_root(n_samples, min_samples_leaf):
# Make sure root node isn't split if n_samples is not at least twice
# min_samples_leaf
rng = np.random.RandomState(seed=0)
n_bins = 256
# data = linear target, 3 features, 1 irrelevant.
X = rng.normal(size=(n_samples, 3))
y = X[:, 0] - X[:, 1]
mapper = _BinMapper(n_bins=n_bins)
X = mapper.fit_transform(X)
all_gradients = y.astype(G_H_DTYPE)
all_hessians = np.ones(shape=1, dtype=G_H_DTYPE)
grower = TreeGrower(X, all_gradients, all_hessians,
n_bins=n_bins, shrinkage=1.,
min_samples_leaf=min_samples_leaf,
max_leaf_nodes=n_samples)
grower.grow()
if n_samples >= min_samples_leaf * 2:
assert len(grower.finalized_leaves) >= 2
else:
assert len(grower.finalized_leaves) == 1
def test_invalid_n_bins(n_bins):
err_msg = ('n_bins={} should be no smaller than 3 and no larger than 256'.
format(n_bins))
with pytest.raises(ValueError, match=err_msg):
_BinMapper(n_bins=n_bins).fit(DATA)
def test_bin_mapper_n_features_transform():
mapper = _BinMapper(n_bins=42, random_state=42).fit(DATA)
err_msg = 'This estimator was fitted with 2 features but 4 got passed'
with pytest.raises(ValueError, match=err_msg):
mapper.transform(np.repeat(DATA, 2, axis=1))
def test_bin_mapper_identity_repeated_values(max_bins, n_distinct, multiplier):
data = np.array(list(range(n_distinct)) * multiplier).reshape(-1, 1)
# max_bins is the number of bins for non-missing values
n_bins = max_bins + 1
binned = _BinMapper(n_bins=n_bins).fit_transform(data)
assert_array_equal(data, binned)
def test_bin_mapper_identity_small(max_bins, scale, offset):
data = np.arange(max_bins).reshape(-1, 1) * scale + offset
# max_bins is the number of bins for non-missing values
n_bins = max_bins + 1
binned = _BinMapper(n_bins=n_bins).fit_transform(data)
assert_array_equal(binned, np.arange(max_bins).reshape(-1, 1))
