package tensorflow.boosted_trees

Get desktop application:
View/edit binary Protocol Buffers messages

Used in: Node

• int32 feature_id = 1

  Float feature column and split threshold describing the rule feature <= threshold.

• int32 threshold = 2

• int32 dimension_id = 5

  If feature column is multivalent, this holds the index of the dimension for the split. Defaults to 0.

• DefaultDirection default_direction = 6

  default direction for missing values.

• int32 left_id = 3

  Node children indexing into a contiguous vector of nodes starting from the root.

• int32 right_id = 4

Used in: Node

• int32 feature_id = 1

  Categorical feature column and split describing the rule feature value == value.

• int32 value = 2

• int32 dimension_id = 5

  If feature column is multivalent, this holds the index of the dimension for the split. Defaults to 0.

• int32 left_id = 3

  Node children indexing into a contiguous vector of nodes starting from the root.

• int32 right_id = 4

DebugOutput contains outputs useful for debugging/model interpretation, at the individual example-level. Debug outputs that are available to the user are: 1) Directional feature contributions (DFCs) 2) Node IDs for ensemble prediction path 3) Leaf node IDs.

• repeated int32 feature_ids = 1

  Return the logits and associated feature splits across prediction paths for each tree, for every example, at predict time. We will use these values to compute DFCs in Python, by subtracting each child prediction from its parent prediction and associating this change with its respective feature id.

• repeated float logits_path = 2

Used in: BucketizedSplit

• DEFAULT_LEFT = 0

  Left is the default direction.

• DEFAULT_RIGHT = 1

TODO(nponomareva): move out of boosted_trees and rename to trees.proto

Used in: Node

• int32 feature_id = 1

  Float feature column and split threshold describing the rule feature <= threshold.

• float threshold = 2

• int32 left_id = 3

  Node children indexing into a contiguous vector of nodes starting from the root.

• int32 right_id = 4

Used in: TreeEnsemble

• int64 num_trees_attempted = 1

  Number of trees that we have attempted to build. After pruning, these trees might have been removed.

• int64 num_layers_attempted = 2

  Number of layers that we have attempted to build. After pruning, these layers might have been removed.

• int32 last_layer_node_start = 3

  The start (inclusive) and end (exclusive) ids of the nodes in the latest layer of the latest tree.

• int32 last_layer_node_end = 4

Leaves can either hold dense or sparse information.

Used in: Node, NodeMetadata

• oneof leaf

  • Vector vector = 1

    See third_party/tensorflow/contrib/decision_trees/ proto/generic_tree_model.proto for a description of how vector and sparse_vector might be used.

  • SparseVector sparse_vector = 2

• float scalar = 3

Node describes a node in a tree.

Used in: Tree

• oneof node

  • Leaf leaf = 1

  • BucketizedSplit bucketized_split = 2

  • CategoricalSplit categorical_split = 3

  • DenseSplit dense_split = 4

• optional NodeMetadata metadata = 777

NodeMetadata encodes metadata associated with each node in a tree.

Used in: Node

• float gain = 1

  The gain associated with this node.

• optional Leaf original_leaf = 2

  The original leaf node before this node was split.

Used in: Leaf

• repeated int32 index = 1

• repeated float value = 2

• INEQUALITY_DEFAULT_LEFT = 0

• INEQUALITY_DEFAULT_RIGHT = 1

• EQUALITY_DEFAULT_RIGHT = 3

Tree describes a list of connected nodes. Node 0 must be the root and can carry any payload including a leaf in the case of representing the bias. Note that each node id is implicitly its index in the list of nodes.

Used in: TreeEnsemble

• repeated Node nodes = 1

TreeEnsemble describes an ensemble of decision trees.

• repeated Tree trees = 1

• repeated float tree_weights = 2

• repeated TreeMetadata tree_metadata = 3

• optional GrowingMetadata growing_metadata = 4

  Metadata that is used during the training.

Used in: TreeEnsemble

• int32 num_layers_grown = 2

  Number of layers grown for this tree.

• bool is_finalized = 3

  Whether the tree is finalized in that no more layers can be grown.

• repeated TreeMetadata.PostPruneNodeUpdate post_pruned_nodes_meta = 4

  If tree was finalized and post pruning happened, it is possible that cache still refers to some nodes that were deleted or that the node ids changed (e.g. node id 5 became node id 2 due to pruning of the other branch). The mapping below allows us to understand where the old ids now map to and how the values should be adjusted due to post-pruning. The size of the list should be equal to the number of nodes in the tree before post-pruning happened. If the node was pruned, it will have new_node_id equal to the id of a node that this node was collapsed into. For a node that didn't get pruned, it is possible that its id still changed, so new_node_id will have the corresponding id in the pruned tree. If post-pruning didn't happen, or it did and it had no effect (e.g. no nodes got pruned), this list will be empty.

Used in: TreeMetadata

• int32 new_node_id = 1

• repeated float logit_change = 2

Used in: Leaf

• repeated float value = 1