def test_core_value_to_value_classifier(self):
core_node = decision_tree_pb2.Node()
core_node.classifier.distribution.counts[:] = [0.0, 8.0, 2.0]
core_node.classifier.distribution.sum = 10.0
self.assertEqual(
value_lib.core_value_to_value(core_node),
value_lib.ProbabilityValue(probability=[0.8, 0.2],
num_examples=10))
def test_core_value_to_value_regressor(self):
core_node = decision_tree_pb2.Node()
core_node.regressor.top_value = 1.0
core_node.regressor.distribution.sum = 10.0
core_node.regressor.distribution.sum_squares = 20.0
core_node.regressor.distribution.count = 10.0
self.assertEqual(
value_lib.core_value_to_value(core_node),
value_lib.RegressionValue(value=1.0,
num_examples=10,
standard_deviation=1.0))
def test_leaf(self):
node = node_lib.LeafNode(value=value_lib.RegressionValue(
value=5.0, num_examples=10, standard_deviation=1.0))
core_node = decision_tree_pb2.Node(
regressor=decision_tree_pb2.NodeRegressorOutput(top_value=5.0))
dist = core_node.regressor.distribution
dist.count = 10.0
dist.sum = 0
dist.sum_squares = 10.0
self.assertEqual(
node_lib.node_to_core_node(node,
data_spec_pb2.DataSpecification()),
core_node)
logging.info("node:\n%s", node)
def node_to_core_node(
node: AbstractNode,
dataspec: data_spec_pb2.DataSpecification) -> decision_tree_pb2.Node:
"""Converts a python node into a core node (proto format)."""
core_node = decision_tree_pb2.Node()
if isinstance(node, NonLeafNode):
condition_lib.set_core_node(node.condition, dataspec, core_node)
if node.value is not None:
value_lib.set_core_node(node.value, core_node)
elif isinstance(node, LeafNode):
value_lib.set_core_node(node.value, core_node)
else:
raise ValueError(
f"Expecting a LeafNode or a NonLeafNode. Got {node} instead")
return core_node
def build_toy_gbdt(path, num_classes):
"""Creates a toy GBDT model compatible with _build_toy_data_spec."""
logging.info("Create toy model in %s", path)
tf.io.gfile.makedirs(path)
with tf.io.gfile.GFile(os.path.join(path, "done"), "w") as f:
f.write("Something")
data_spec = build_toy_data_spec()
with tf.io.gfile.GFile(os.path.join(path, "data_spec.pb"), "w") as f:
f.write(data_spec.SerializeToString())
header = abstract_model_pb2.AbstractModel(
name="GRADIENT_BOOSTED_TREES",
task=abstract_model_pb2.CLASSIFICATION,
label_col_idx=4 if num_classes == 2 else 3,
input_features=[0, 1, 2])
with tf.io.gfile.GFile(os.path.join(path, "header.pb"), "w") as f:
f.write(header.SerializeToString())
num_iters = 2
num_trees_per_iter = 1 if num_classes == 2 else num_classes
rf_header = gradient_boosted_trees_pb2.Header(
num_node_shards=1,
num_trees=num_iters * num_trees_per_iter,
loss=gradient_boosted_trees_pb2.BINOMIAL_LOG_LIKELIHOOD if num_classes
== 2 else gradient_boosted_trees_pb2.MULTINOMIAL_LOG_LIKELIHOOD,
initial_predictions=[1.0] if num_classes == 2 else [0.0] * num_classes,
num_trees_per_iter=num_trees_per_iter,
node_format="BLOB_SEQUENCE")
with tf.io.gfile.GFile(
os.path.join(path, "gradient_boosted_trees_header.pb"), "w") as f:
f.write(rf_header.SerializeToString())
with blob_sequence.Writer(os.path.join(
path, "nodes-00000-of-00001")) as output_file:
for _ in range(num_iters):
for tree_in_iter_idx in range(num_trees_per_iter):
# [a > 1 ] // Node 0
# |-- [label = 1.0 + tree_in_iter_idx] // Node 1
# L-- [label = 5.0 + tree_in_iter_idx^2] // Node 2
#
# Two classes
# Case a<=1:
# logit = 1.0 + 1.0 * 2 = 3.0
# proba = [0.0474259, 0.9525741]
# Case a>1:
# logit = 1.0 + 5.0 * 2 = 11.0
# proba = [1.67e-05, 0.9999833]
#
# Three classes
# Case a<=1:
# logit = [1.0 * 2, 2.0 * 2, 3.0 * 2] = [2.0, 4.0, 6.0]
# proba = [0.01587624 0.11731043 0.86681333]
# Case a>1:
# logit = [5.0 * 2, 6.0 * 2, 9.0 * 2] = [10.0, 12.0, 18.0]
# proba = [0.01587624 0.11731043 0.86681333]
# Node 0
node = decision_tree_pb2.Node(
condition=decision_tree_pb2.NodeCondition(
na_value=False,
attribute=0,
condition=decision_tree_pb2.Condition(
higher_condition=decision_tree_pb2.Condition.
Higher(threshold=1.0)),
))
output_file.write(node.SerializeToString())
# Node 1
node = decision_tree_pb2.Node(
regressor=decision_tree_pb2.NodeRegressorOutput(
top_value=1.0 + tree_in_iter_idx))
output_file.write(node.SerializeToString())
# Node 2
node = decision_tree_pb2.Node(
regressor=decision_tree_pb2.NodeRegressorOutput(
top_value=5.0 + tree_in_iter_idx * tree_in_iter_idx))
output_file.write(node.SerializeToString())
def build_toy_random_forest(path,
winner_take_all_inference,
add_boolean_features=False,
has_catset=False,
num_trees=2):
"""Creates a toy Random Forest model compatible with _build_toy_data_spec."""
logging.info("Create toy model in %s", path)
tf.io.gfile.makedirs(path)
with tf.io.gfile.GFile(os.path.join(path, "done"), "w") as f:
f.write("Something")
data_spec = build_toy_data_spec(add_boolean_features=add_boolean_features,
has_catset=has_catset)
with tf.io.gfile.GFile(os.path.join(path, "data_spec.pb"), "w") as f:
f.write(data_spec.SerializeToString())
header = abstract_model_pb2.AbstractModel(
name="RANDOM_FOREST",
task=abstract_model_pb2.CLASSIFICATION,
label_col_idx=3,
input_features=[0, 1, 2] + ([5] if add_boolean_features else []) +
([5, 6] if has_catset else []))
with tf.io.gfile.GFile(os.path.join(path, "header.pb"), "w") as f:
f.write(header.SerializeToString())
rf_header = random_forest_pb2.Header(
num_node_shards=1,
num_trees=num_trees,
winner_take_all_inference=winner_take_all_inference,
node_format="BLOB_SEQUENCE")
with tf.io.gfile.GFile(os.path.join(path, "random_forest_header.pb"),
"w") as f:
f.write(rf_header.SerializeToString())
with blob_sequence.Writer(os.path.join(
path, "nodes-00000-of-00001")) as output_file:
for _ in range(rf_header.num_trees):
# [a > 1 ] // Node 0
# |-- [b in ["x,"y"] ] // Node 1
# | |-- [label = 80%;10%;10%] // Node 2
# | L-- [label = 10%;80%;10%] // Node 3
# L-- [c in [1, 3] ] // Node 4
# |-- [label = 50%;50%;0%] // Node 5
# L-- [label = 0%;50%;50%] // Node 6
#
# If add_boolean_features is True, Node 6 is repurposed as follows:
#
# ['bool' is True] // Node 6
# | -- [label = 0%;20%;80%] // Node 7
# L -- [label = 0%;80%;20%] // Node 8
#
# If has_catset is True, Node 4 condition is replaced by:
# [ d \intersect [1,3] != \emptyset
# Node 0
node = decision_tree_pb2.Node(
condition=decision_tree_pb2.NodeCondition(
na_value=False,
attribute=0,
condition=decision_tree_pb2.Condition(
higher_condition=decision_tree_pb2.Condition.Higher(
threshold=1.0)),
))
output_file.write(node.SerializeToString())
# Node 1
node = decision_tree_pb2.Node(
condition=decision_tree_pb2.NodeCondition(
na_value=False,
attribute=1,
condition=decision_tree_pb2.Condition(
contains_bitmap_condition=decision_tree_pb2.Condition.
ContainsBitmap(elements_bitmap=b"\x06")), # [1,2]
))
output_file.write(node.SerializeToString())
# Node 2
node = decision_tree_pb2.Node(
classifier=decision_tree_pb2.NodeClassifierOutput(
top_value=1,
distribution=distribution_pb2.IntegerDistributionDouble(
counts=[0, 0.8, 0.1, 0.1], sum=1)))
output_file.write(node.SerializeToString())
# Node 3
node = decision_tree_pb2.Node(
classifier=decision_tree_pb2.NodeClassifierOutput(
top_value=2,
distribution=distribution_pb2.IntegerDistributionDouble(
counts=[0, 0.1, 0.8, 0.1], sum=1)))
output_file.write(node.SerializeToString())
# Node 4
node = decision_tree_pb2.Node(
condition=decision_tree_pb2.NodeCondition(
na_value=False,
attribute=5 if has_catset else 2,
condition=decision_tree_pb2.Condition(
contains_condition=decision_tree_pb2.Condition.
ContainsVector(elements=[1, 3]))))
output_file.write(node.SerializeToString())
# Node 5
node = decision_tree_pb2.Node(
classifier=decision_tree_pb2.NodeClassifierOutput(
top_value=1,
distribution=distribution_pb2.IntegerDistributionDouble(
counts=[0, 1, 1, 0], sum=2)))
output_file.write(node.SerializeToString())
if not add_boolean_features:
# Node 6
node = decision_tree_pb2.Node(
classifier=decision_tree_pb2.NodeClassifierOutput(
top_value=2,
distribution=distribution_pb2.
IntegerDistributionDouble(counts=[0, 0, 1, 1], sum=2)))
output_file.write(node.SerializeToString())
else:
# Node 6
node = decision_tree_pb2.Node(
condition=decision_tree_pb2.NodeCondition(
na_value=False,
attribute=5,
condition=decision_tree_pb2.Condition(
true_value_condition=decision_tree_pb2.Condition.
TrueValue())))
output_file.write(node.SerializeToString())
# Node 7
node = decision_tree_pb2.Node(
classifier=decision_tree_pb2.NodeClassifierOutput(
top_value=3,
distribution=distribution_pb2.
IntegerDistributionDouble(counts=[0, 0, 0.2, 0.8],
sum=1)))
output_file.write(node.SerializeToString())
# Node 8
node = decision_tree_pb2.Node(
classifier=decision_tree_pb2.NodeClassifierOutput(
top_value=2,
distribution=distribution_pb2.
IntegerDistributionDouble(counts=[0, 0, 0.8, 0.2],
sum=1)))
output_file.write(node.SerializeToString())