def test_error_gbt_wrong_number_of_trees(self):
builder = builder_lib.GradientBoostedTreeBuilder(
path=os.path.join(tmp_path(), "model"),
objective=py_tree.objective.ClassificationObjective(
"label", classes=["red", "blue", "green"]))
builder.add_tree(Tree(LeafNode(RegressionValue(1, num_examples=10))))
self.assertRaises(ValueError, builder.close)
def test_multi_class_classification_gbt(self, file_prefix):
model_path = os.path.join(tmp_path(), "multi_class_classification_gbt")
logging.info("Create model in %s", model_path)
builder = builder_lib.GradientBoostedTreeBuilder(
path=model_path,
model_format=builder_lib.ModelFormat.TENSORFLOW_SAVED_MODEL,
objective=py_tree.objective.ClassificationObjective(
label="color", classes=["red", "blue", "green"]),
file_prefix=file_prefix)
# f1>=1.5
# ├─(pos)─ +1.0 (toward "red")
# └─(neg)─ -1.0 (toward "red")
# f1>=2.5
# ├─(pos)─ +1.0 (toward "blue")
# └─(neg)─ -1.0 (toward "blue")
# f1>=3.5
# ├─(pos)─ +1.0 (toward "green")
# └─(neg)─ -1.0 (toward "green")
for threshold in [1.5, 2.5, 3.5]:
builder.add_tree(
Tree(
NonLeafNode(
condition=NumericalHigherThanCondition(
feature=SimpleColumnSpec(
name="f1",
type=py_tree.dataspec.ColumnType.NUMERICAL),
threshold=threshold,
missing_evaluation=False),
pos_child=LeafNode(
value=RegressionValue(value=+1, num_examples=30)),
neg_child=LeafNode(value=RegressionValue(
value=-1, num_examples=30)))))
builder.close()
if file_prefix is not None:
self.assertEqual(
inspector_lib.detect_model_file_prefix(
os.path.join(model_path, "assets")), file_prefix)
logging.info("Loading model")
loaded_model = tf.keras.models.load_model(model_path)
logging.info("Make predictions")
tf_dataset = tf.data.Dataset.from_tensor_slices({
"f1": [1.0, 2.0],
}).batch(2)
predictions = loaded_model.predict(tf_dataset)
soft_max_sum = np.sum(np.exp([+1, -1, -1]))
self.assertAllClose(predictions, [[1.0 / 3.0, 1.0 / 3.0, 1.0 / 3.0],
[
math.exp(+1) / soft_max_sum,
math.exp(-1) / soft_max_sum,
math.exp(-1) / soft_max_sum
]])
def test_error_gbt_with_class_tree(self):
builder = builder_lib.GradientBoostedTreeBuilder(
path=os.path.join(tmp_path(), "model"),
objective=py_tree.objective.ClassificationObjective(
"label", classes=["red", "blue", "green"]))
self.assertRaises(
ValueError, lambda: builder.add_tree(
Tree(
LeafNode(
ProbabilityValue(probability=[0.8, 0.1, 0.1],
num_examples=10)))))
def test_binary_classification_gbt(self):
model_path = os.path.join(tmp_path(), "binary_classification_gbt")
logging.info("Create model in %s", model_path)
builder = builder_lib.GradientBoostedTreeBuilder(
path=model_path,
model_format=builder_lib.ModelFormat.TENSORFLOW_SAVED_MODEL,
bias=1.0,
objective=py_tree.objective.ClassificationObjective(
label="color", classes=["red", "blue"]))
# bias: 1.0 (toward "blue")
# f1>=1.5
# ├─(pos)─ +1.0 (toward "blue")
# └─(neg)─ -1.0 (toward "blue")
builder.add_tree(
Tree(
NonLeafNode(
condition=NumericalHigherThanCondition(
feature=SimpleColumnSpec(
name="f1",
type=py_tree.dataspec.ColumnType.NUMERICAL),
threshold=1.5,
missing_evaluation=False),
pos_child=LeafNode(
value=RegressionValue(value=+1, num_examples=30)),
neg_child=LeafNode(
value=RegressionValue(value=-1, num_examples=30)))))
builder.close()
logging.info("Loading model")
loaded_model = tf.keras.models.load_model(model_path)
logging.info("Make predictions")
tf_dataset = tf.data.Dataset.from_tensor_slices({
"f1": [1.0, 2.0],
}).batch(2)
predictions = loaded_model.predict(tf_dataset)
self.assertAllClose(
predictions,
[[1.0 / (1.0 + math.exp(0.0))], [1.0 / (1.0 + math.exp(-2.0))]])
def test_fast_serving_with_custom_numerical_default_evaluation(self):
model_path = os.path.join(tmp_path(), "regression_gbt")
logging.info("Create model in %s", model_path)
builder = builder_lib.GradientBoostedTreeBuilder(
path=model_path,
bias=0.0,
model_format=builder_lib.ModelFormat.TENSORFLOW_SAVED_MODEL,
objective=py_tree.objective.RegressionObjective(label="label"))
# f1>=-1.0 (default: false)
# │
# ├─f1>=2.0 (default: false)
# │ │
# │ ├─1
# │ └─2
# └─f2>=-3.0 (default: true)
# │
# ├─f2>=4.0 (default: false)
# │ │
# │ ├─3
# │ └─4
# └─5
def condition(feature, threshold, missing_evaluation, pos, neg):
return NonLeafNode(condition=NumericalHigherThanCondition(
feature=SimpleColumnSpec(
name=feature, type=py_tree.dataspec.ColumnType.NUMERICAL),
threshold=threshold,
missing_evaluation=missing_evaluation),
pos_child=pos,
neg_child=neg)
def leaf(value):
return LeafNode(RegressionValue(value=value, num_examples=1))
builder.add_tree(
Tree(
condition(
"f1", -1.0, False,
condition("f1", 2.0, False, leaf(1), leaf(2)),
condition(
"f2",
-3.0,
True,
condition("f2", 4.0, False, leaf(3), leaf(4)),
leaf(5),
))))
builder.close()
logging.info("Loading model")
# There is no easy way to assert that an optimized inference engine was
# chosen. If checking manually, make sure the "Use fast generic engine"
# string is present (instead of the "Use slow generic engine" string).
#
# TODO(gbm):: Add API to check which inference engine is used.
loaded_model = tf.keras.models.load_model(model_path)
logging.info("Make predictions")
tf_dataset = tf.data.Dataset.from_tensor_slices({
"f1": [math.nan, 1.0, -2.0],
"f2": [-4.0, -4.0, math.nan],
}).batch(2)
predictions = loaded_model.predict(tf_dataset)
self.assertAllClose(predictions, [[5.0], [2.0], [4.0]])
inspector = inspector_lib.make_inspector(
os.path.join(model_path, "assets"))
self.assertEquals(inspector.dataspec.columns[1].numerical.mean,
-1.0 - 0.5)
self.assertEquals(inspector.dataspec.columns[2].numerical.mean,
(4.0 - 3.0) / 2.0)