package yggdrasil_decision_forests.model.hyperparameters_optimizer_v2.proto

Get desktop application:
View/edit binary Protocol Buffers messages

Used in: HyperParametersOptimizerLearnerTrainingConfig

• optional metric.proto.MetricAccessor metric = 1

  Evaluation metric to optimize. If not set, uses the first available metric in the list: loss > auc (binary classification only) > accuracy > rmse ndcg > qini. Fails if none of those metrics are defined.

• optional bool maximize_metric = 2

  If true, maximize the metric value. If false, minimize the metric value. If not set, uses the metric definition e.g. maximize accuracy and minimize loss.

• oneof source

  • Evaluation.SelfEvaluation self_model_evaluation = 3

    Uses the self reported model evaluation e.g. validation score or OOB evaluation. Default.

Used in: Evaluation

(message has no fields)

• optional proto.TrainingConfig base_learner = 1

  Definition of the machine learning algorithm to tune as well as the value hyper parameters that are not tuned.

• optional Optimizer optimizer = 2

  Optimization method.

• optional Evaluation evaluation = 3

  Control how to evaluate a candidate set of hyper parameters.

• optional proto.HyperParameterSpace search_space = 4

  Manually define the space of hyper-parameters. Fields defined in "hyper_parameter_space" override the pre-defined space specified in "predefined_search_space" (if "predefined_hyper_parameter_space" is set).

• optional PredefinedHyperParameterSpace predefined_search_space = 8

  If set, configure automatically "search_space" with the pre-defined hyper-parameters of the learner (in "PredefinedHyperParameterSpace").

• optional proto.DeploymentConfig base_learner_deployment = 5

  Deployment configuration (i.e. computing resources) for the base learner.

• optional string serialized_dataset_format = 6

  Format used to serialize the dataset if the dataset is provided as a VerticalDataset (i.e. in memory dataset) and should be communicated to remote workers. The format IO library should be registered both to the learner and the workers.

• optional bool retrain_final_model = 7

  If true, the final model is re-trained using the best found hyper-parameters. If false, the best model found during optimization is returned. Model re-training is more expensive and can both improve or hurt the quality of the final model. This option has not impact if the training is deterministic.

Used in: HyperParametersOptimizerLearnerTrainingConfig

• optional string optimizer_key = 1

  Registered name of the optimizer.

• optional int32 parallel_trials = 2

  Number of trials to evaluate in parallel. If using distributed tuning, it is best for this parameter to be a multiple of the number of workers, since the training of an individual model takes "num_threads" threads (as configured in the learner). The total number of threads used by a worker is, on average, "num_threads * parallel_trials // num_workers".

Empty message

TODO: Make it possible to tune only a subset of paramters.

Used in: HyperParametersOptimizerLearnerTrainingConfig

(message has no fields)

• optional int64 num_trials = 1

  Number of randomly generate candidates.