hyperopt

Type Members

1. trait BayesianParamOptimizer extends AnyRef

   

   Used to perform hyperparameters search via sampling.

   Used to perform hyperparameters search via sampling. Sampling is performed from the [0,1]^^N hypercube and mapping from [0,1] to particular values are made by the parameter domains. Availible implementations of Bayessian optimizers are largely based on the LinkedIn Photon-ML project.

2. trait BayesianParamOptimizerFactory extends AnyRef

   

   Creates optimizer given the problem description.

3. case class CategorialParam[T](values: Array[T]) extends ParamDomain[T] with Product with Serializable

   

   Models parameter having limited set of values

4. case class ConfigHolder(number: Int, config: ParamMap) extends Product with Serializable

   

5. case class ConfigNumber(number: Int, config: ParamMap) extends Product with Serializable

   

6. case class DoubleRangeDomain(lower: Double, upper: Double) extends ParamDomain[Double] with Product with Serializable

   

   Models a simple real valued parameter from the range [lower,upper]

7. class GaussianProcessOptimizer extends GaussianProcessSearch[Double] with BayesianParamOptimizer

   

   Advanced sampler modeling the efficiency function as a family of Gaussian processes (with integrated kernel parameters) and sampling from it trying to maximize Expected Improvement.

8. class GridSearch[ModelIn <: ModelWithSummary[ModelIn]] extends ForkedEstimator[ModelIn, ConfigHolder, ModelIn] with HyperparametersOptimizer[ModelIn]

   

   Provides ability to search through multiple configurations in parallel mode, collecting all the stat and metrics.

   Provides ability to search through multiple configurations in parallel mode, collecting all the stat and metrics.

   Supports persisting temporary models in order to restore after failures, but only when used with StableOrderParamGridBuilder.

9. class GroupedSearch[ModelIn <: ModelWithSummary[ModelIn]] extends ForkedEstimator[ModelIn, OptimizerStage, ModelIn] with HasConfigurations

   

   Utility used to perform stepwise search for hyper-parameters.

   Utility used to perform stepwise search for hyper-parameters. Usefull in case if there are certain groups of parameters which do not influence each other and can be optimized separatelly. Pass sequence of optimizers to the grouped search to apply sequential optimization.

   NB: Grouped search must itself be configured to use single thread, but nested optimizers allowed to use as many threads as they need.

10. trait HasConfigurations extends Params with MetricsExtractor

    

    Common summary block to store history of the hyperparameters search.

11. trait HyperparametersOptimizer[M <: ModelWithSummary[M]] extends Estimator[M] with SummarizableEstimator[M] with HasConfigurations with HasNumThreads

    

    Common trait to all the hyper-parameter optimizers.

12. case class IntRangeDomain(lower: Int, upper: Int) extends ParamDomain[Int] with Product with Serializable

    

    Models a ordinal valued parameter from the sequence {lower, lower + 1, ...

    Models a ordinal valued parameter from the sequence {lower, lower + 1, ... , upper}

13. case class OptimizerStage(stage: String, accumulated: ParamMap) extends Product with Serializable

    

14. trait ParamDomain[T] extends AnyRef

    

    Parameters domain is used to map from the [0,1] value sampled from the optimizer to the actual parameter value.

15. case class ParamDomainPair[T](param: Param[T], domain: ParamDomain[T]) extends Product with Serializable

    

    Holds the actual SparkML param and its domain.

    Holds the actual SparkML param and its domain. Support type-safe methods for moving data between optimizer, data frame and SparkML estimator.

16. class RandomOptimizer extends RandomSearch[Double] with BayesianParamOptimizer

    

    Simple sampler from the Sobol points sequence

17. class StableOrderParamGridBuilder extends AnyRef

    

    Builder for a param grid used in grid search-based model selection.

    Builder for a param grid used in grid search-based model selection. This builder provdes stable order and thus might be reliablly used with persist temp models feature

18. class StochasticHyperopt[ModelIn <: ModelWithSummary[ModelIn]] extends ForkedEstimator[ModelIn, ConfigNumber, ModelIn] with HyperparametersOptimizer[ModelIn] with HasMaxIter with HasTol with HasSeed

    

    Searches for optimal parameters using Bayesian approach.

    Searches for optimal parameters using Bayesian approach. Important difference of this searcher compared to other forked estimators is the need to get previous evaluation to know where to sample next params.