API Reference

This page documents the Python API for YDF. Users can also train models using the C++ and CLI APIs.

Learners

A Learner trains models and can be cross-validated.

• GradientBoostedTreesLearner
• RandomForestLearner
• CartLearner
• DecisionTreeLearner: Alias to CartLearner.
• DistributedGradientBoostedTreesLearner

All learners derive from GenericLearner.

Models

A Model makes predictions and can be evaluated.

Note: Models (e.g., GradientBoostedTreesModel) do not contain training capabilities. To train a model, you need to create a learner (e.g., GradientBoostedTreesLearner). Training hyperparameters are constructor arguments of learner classes.

• GradientBoostedTreesModel
• RandomForestModel
• CARTModel: Alias to RandomForestModel.

All models derive from GenericModel.

Tuners

A Tuner finds the optimal set of hyper-parameters using repeated training and evaluation.

• RandomSearchTuner
• VizierTuner (currently, for Googlers only)

Utilities

• verbose: Control the amount of logging.
• load_model: Load a model from disk.
• Feature: Input feature specific hyper-parameters e.g. semantic, constraints.
• Column: Alias for Feature.
• Task: Specify the task solved by the model e.g. classification.
• Semantic: How an input feature in interpreted e.g. numerical, categorical.
• start_worker: Start a worker for distributed training.
• strict: Show more logs.

Advanced Utilities

• ModelIOOptions: Options to save a model to disk.
• create_vertical_dataset: Load a dataset in memory.
• ModelMetadata: Meta-data about the model e.g. training date, uid.
• from_tensorflow_decision_forests: Load a TensorFlow Decision Forests model from disk.

Custom Loss

• RegressionLoss: Custom loss for regression tasks.
• BinaryClassificationLoss: Custom loss for binary classification tasks.
• MultiClassificationLoss: Custom loss for multi-class classification tasks.
• Activation: Collection of activation (aka linkage) functions for custom losses.

Tree

The ydf.tree.* classes provides programmatic read and write access to the tree structure, leaves, and values.

• tree.Tree: A decision tree as returned and consumed by model.get_tree(...) and model.set_tree(...)..

Conditions

• tree.AbstractCondition: Base condition class.
• tree.NumericalHigherThanCondition: Condition of the form attribute >= threshold.
• tree.CategoricalIsInCondition: Condition of the form attribute in mask.
• tree.CategoricalSetContainsCondition: Condition of the form attribute intersect mask != empty.
• tree.DiscretizedNumericalHigherThanCondition: Condition of the form attribute >= bounds[threshold].
• tree.IsMissingInCondition: Condition of the form attribute is missing.
• tree.IsTrueCondition: Condition of the form attribute is true.
• tree.NumericalSparseObliqueCondition: Condition of the form sum(attributes[i] * weights[i]) >= threshold.

Nodes

• tree.AbstractNode: Base node class.
• tree.Leaf: A leaf node containing a value.
• tree.NonLeaf: A non-leaf node containing a condition.

Values

• tree.AbstractValue: Base value class.
• tree.ProbabilityValue: A probability distribution value.
• tree.Leaf: The regression value of a regressive tree.
• tree.UpliftValue: The uplift value of a classification or regression uplift tree.