Hyperparameter Search (HPS)

Algorithms

Asynchronous Model-Base Search (AMBS)

You can download the deephyper paper here

Environment variable to access the search on Theta: DH_AMBS

Asynchronous Model-Based Search.

Arguments of AMBS:

  • surrogate-model

    • RF : Random Forest (default)

    • ET : Extra Trees

    • GBRT : Gradient Boosting Regression Trees

    • DUMMY :

    • GP : Gaussian process

  • liar-strategy

    • cl_max : (default)

    • cl_min :

    • cl_mean :

  • acq-func : Acquisition function

    • LCB :

    • EI :

    • PI :

    • gp_hedge : (default)

class deephyper.search.hps.ambs.AMBS(problem, run, evaluator, surrogate_model='RF', acq_func='LCB', kappa=1.96, xi=0.001, liar_strategy='cl_max', n_jobs=32, **kwargs)[source]
get_surrogate_model(name: str, n_jobs: int = None)[source]

Get a surrogate model from Scikit-Optimize.

Parameters
  • name (str) – name of the surrogate model.

  • n_jobs (int) – number of parallel processes to distribute the computation of the surrogate model.

Raises

ValueError – when the name of the surrogate model is unknown.

deephyper.search.hps.ambs.isnan(x)bool[source]

Check if a value is NaN.

Genetic Algorithm (GA)

class deephyper.search.hps.ga.GA(problem, run, evaluator, **kwargs)[source]

Benchmarks

Available HPS benchmarks

Benchmarks are here for you to test the performance of different search algorithms and to help you to reproduce our results. They can also be used to test your installation of deephyper or to discover the many parameters of a search. If you want to see all available hypeparameter search benchmark please follow this link: Available HPS benchmarks. If you want to see how to create and run a new hyperparameter search problem please follow this link: Hyperparameter Search for Deep Learning (Basic).

Hyper Parameters Search Benchmarks deephyper.benchmark.hps

Name

Type

Description

mnistmlp

Classification

http://yann.lecun.com/exdb/mnist/

polynome2

Dummy

Problem

Use this class to define a hyperparameter search problem, see create-new-hps-problem for more details.

class deephyper.benchmark.problem.HpProblem(seed=None)[source]

Problem specification for Hyperparameter Search