Search Algorithms (tune.suggest)

Tune’s Search Algorithms are wrappers around open-source optimization libraries for efficient hyperparameter selection. Each library has a specific way of defining the search space - please refer to their documentation for more details.

You can utilize these search algorithms as follows:

from ray.tune.suggest.hyperopt import HyperOptSearch
tune.run(my_function, search_alg=HyperOptSearch(...))

Summary

SearchAlgorithm

Summary

Website

Code Example

AxSearch

Bayesian/Bandit Optimization

[Ax]

Link

DragonflySearch

Scalable Bayesian Optimization

[Dragonfly]

Link

SkoptSearch

Bayesian Optimization

[Scikit-Optimize]

Link

HyperOptSearch

Tree-Parzen Estimators

[HyperOpt]

Link

BayesOptSearch

Bayesian Optimization

[BayesianOptimization]

Link

TuneBOHB

Bayesian Opt/HyperBand

[BOHB]

Link

NevergradSearch

Gradient-free Optimization

[Nevergrad]

Link

ZOOptSearch

Zeroth-order Optimization

[ZOOpt]

Link

SigOptSearch

Closed source

[SigOpt]

Link

Note

Unlike Tune’s Trial Schedulers, Tune SearchAlgorithms cannot affect or stop training processes. However, you can use them together to early stop the evaluation of bad trials.

Want to use your own algorithm? The interface is easy to implement. Read instructions here.

Tune also provides helpful utilities to use with Search Algorithms:

Ax (tune.suggest.ax.AxSearch)

class ray.tune.suggest.ax.AxSearch(ax_client, mode='max', use_early_stopped_trials=None, max_concurrent=None)[source]

Uses Ax to optimize hyperparameters.

Ax is a platform for understanding, managing, deploying, and automating adaptive experiments. Ax provides an easy to use interface with BoTorch, a flexible, modern library for Bayesian optimization in PyTorch. More information can be found in https://ax.dev/.

To use this search algorithm, you must install Ax and sqlalchemy:

$ pip install ax-platform sqlalchemy
Parameters
  • parameters (list[dict]) – Parameters in the experiment search space. Required elements in the dictionaries are: “name” (name of this parameter, string), “type” (type of the parameter: “range”, “fixed”, or “choice”, string), “bounds” for range parameters (list of two values, lower bound first), “values” for choice parameters (list of values), and “value” for fixed parameters (single value).

  • objective_name (str) – Name of the metric used as objective in this experiment. This metric must be present in raw_data argument to log_data. This metric must also be present in the dict reported/returned by the Trainable.

  • mode (str) – One of {min, max}. Determines whether objective is minimizing or maximizing the metric attribute. Defaults to “max”.

  • parameter_constraints (list[str]) – Parameter constraints, such as “x3 >= x4” or “x3 + x4 >= 2”.

  • outcome_constraints (list[str]) – Outcome constraints of form “metric_name >= bound”, like “m1 <= 3.”

  • max_concurrent (int) – Deprecated.

  • use_early_stopped_trials – Deprecated.

from ax.service.ax_client import AxClient
from ray import tune
from ray.tune.suggest.ax import AxSearch

parameters = [
    {"name": "x1", "type": "range", "bounds": [0.0, 1.0]},
    {"name": "x2", "type": "range", "bounds": [0.0, 1.0]},
]

def easy_objective(config):
    for i in range(100):
        intermediate_result = config["x1"] + config["x2"] * i
        tune.track.log(score=intermediate_result)

client = AxClient(enforce_sequential_optimization=False)
client.create_experiment(parameters=parameters, objective_name="score")
algo = AxSearch(client)
tune.run(easy_objective, search_alg=algo)

Bayesian Optimization (tune.suggest.bayesopt.BayesOptSearch)

class ray.tune.suggest.bayesopt.BayesOptSearch(space, metric, mode='max', utility_kwargs=None, random_state=42, random_search_steps=10, verbose=0, analysis=None, max_concurrent=None, use_early_stopped_trials=None)[source]

Uses fmfn/BayesianOptimization to optimize hyperparameters.

fmfn/BayesianOptimization is a library for Bayesian Optimization. More info can be found here: https://github.com/fmfn/BayesianOptimization.

You will need to install fmfn/BayesianOptimization via the following:

pip install bayesian-optimization

This algorithm requires setting a search space using the BayesianOptimization search space specification.

Parameters
  • space (dict) – Continuous search space. Parameters will be sampled from this space which will be used to run trials.

  • metric (str) – The training result objective value attribute.

  • mode (str) – One of {min, max}. Determines whether objective is minimizing or maximizing the metric attribute.

  • utility_kwargs (dict) – Parameters to define the utility function. The default value is a dictionary with three keys: - kind: ucb (Upper Confidence Bound) - kappa: 2.576 - xi: 0.0

  • random_state (int) – Used to initialize BayesOpt.

  • random_search_steps (int) – Number of initial random searches. This is necessary to avoid initial local overfitting of the Bayesian process.

  • analysis (ExperimentAnalysis) – Optionally, the previous analysis to integrate.

  • verbose (int) – Sets verbosity level for BayesOpt packages.

  • max_concurrent – Deprecated.

  • use_early_stopped_trials – Deprecated.

from ray import tune
from ray.tune.suggest.bayesopt import BayesOptSearch

space = {
    'width': (0, 20),
    'height': (-100, 100),
}
algo = BayesOptSearch(space, metric="mean_loss", mode="min")
tune.run(my_func, algo=algo)

BOHB (tune.suggest.bohb.TuneBOHB)

BOHB (Bayesian Optimization HyperBand) is an algorithm that both terminates bad trials and also uses Bayesian Optimization to improve the hyperparameter search. It is backed by the HpBandSter library.

Importantly, BOHB is intended to be paired with a specific scheduler class: HyperBandForBOHB.

This algorithm requires using the ConfigSpace search space specification. In order to use this search algorithm, you will need to install HpBandSter and ConfigSpace:

$ pip install hpbandster ConfigSpace

See the BOHB paper for more details.

class ray.tune.suggest.bohb.TuneBOHB(space, bohb_config=None, max_concurrent=10, metric='neg_mean_loss', mode='max')[source]

BOHB suggestion component.

Requires HpBandSter and ConfigSpace to be installed. You can install HpBandSter and ConfigSpace with: pip install hpbandster ConfigSpace.

This should be used in conjunction with HyperBandForBOHB.

Parameters
  • space (ConfigurationSpace) – Continuous ConfigSpace search space. Parameters will be sampled from this space which will be used to run trials.

  • bohb_config (dict) – configuration for HpBandSter BOHB algorithm

  • max_concurrent (int) – Number of maximum concurrent trials. Defaults to 10.

  • metric (str) – The training result objective value attribute.

  • mode (str) – One of {min, max}. Determines whether objective is minimizing or maximizing the metric attribute.

Example:

import ConfigSpace as CS

config_space = CS.ConfigurationSpace()
config_space.add_hyperparameter(
    CS.UniformFloatHyperparameter('width', lower=0, upper=20))
config_space.add_hyperparameter(
    CS.UniformFloatHyperparameter('height', lower=-100, upper=100))
config_space.add_hyperparameter(
    CS.CategoricalHyperparameter(
        name='activation', choices=['relu', 'tanh']))

algo = TuneBOHB(
    config_space, max_concurrent=4, metric='mean_loss', mode='min')
bohb = HyperBandForBOHB(
    time_attr='training_iteration',
    metric='mean_loss',
    mode='min',
    max_t=100)
run(MyTrainableClass, scheduler=bohb, search_alg=algo)

Dragonfly (tune.suggest.dragonfly.DragonflySearch)

class ray.tune.suggest.dragonfly.DragonflySearch(optimizer, metric='episode_reward_mean', mode='max', points_to_evaluate=None, evaluated_rewards=None, **kwargs)[source]

Uses Dragonfly to optimize hyperparameters.

Dragonfly provides an array of tools to scale up Bayesian optimisation to expensive large scale problems, including high dimensional optimisation. parallel evaluations in synchronous or asynchronous settings, multi-fidelity optimisation (using cheap approximations to speed up the optimisation process), and multi-objective optimisation. For more info:

To use this search algorithm, install Dragonfly:

$ pip install dragonfly-opt

This interface requires using FunctionCallers and optimizers provided by Dragonfly.

from ray import tune
from dragonfly.opt.gp_bandit import EuclideanGPBandit
from dragonfly.exd.experiment_caller import EuclideanFunctionCaller
from dragonfly import load_config

domain_vars = [{
    "name": "LiNO3_vol",
    "type": "float",
    "min": 0,
    "max": 7
}, {
    "name": "Li2SO4_vol",
    "type": "float",
    "min": 0,
    "max": 7
}, {
    "name": "NaClO4_vol",
    "type": "float",
    "min": 0,
    "max": 7
}]

domain_config = load_config({"domain": domain_vars})
func_caller = EuclideanFunctionCaller(None,
    domain_config.domain.list_of_domains[0])
optimizer = EuclideanGPBandit(func_caller, ask_tell_mode=True)

algo = DragonflySearch(optimizer, metric="objective", mode="max")

tune.run(my_func, algo=algo)
Parameters
  • optimizer (dragonfly.opt.BlackboxOptimiser) – Optimizer provided from dragonfly. Choose an optimiser that extends BlackboxOptimiser.

  • metric (str) – The training result objective value attribute.

  • mode (str) – One of {min, max}. Determines whether objective is minimizing or maximizing the metric attribute.

  • points_to_evaluate (list of lists) – A list of points you’d like to run first before sampling from the optimiser, e.g. these could be parameter configurations you already know work well to help the optimiser select good values. Each point is a list of the parameters using the order definition given by parameter_names.

  • evaluated_rewards (list) – If you have previously evaluated the parameters passed in as points_to_evaluate you can avoid re-running those trials by passing in the reward attributes as a list so the optimiser can be told the results without needing to re-compute the trial. Must be the same length as points_to_evaluate.

HyperOpt (tune.suggest.hyperopt.HyperOptSearch)

class ray.tune.suggest.hyperopt.HyperOptSearch(space, metric='episode_reward_mean', mode='max', points_to_evaluate=None, n_initial_points=20, random_state_seed=None, gamma=0.25, max_concurrent=None, use_early_stopped_trials=None)[source]

A wrapper around HyperOpt to provide trial suggestions.

HyperOpt a Python library for serial and parallel optimization over awkward search spaces, which may include real-valued, discrete, and conditional dimensions. More info can be found at http://hyperopt.github.io/hyperopt.

HyperOptSearch uses the Tree-structured Parzen Estimators algorithm, though it can be trivially extended to support any algorithm HyperOpt supports.

To use this search algorithm, you will need to install HyperOpt:

pip install -U hyperopt

You will not be able to leverage Tune’s default grid_search and random search primitives when using HyperOptSearch. You need to use the HyperOpt search space specification.

space = {
    'width': hp.uniform('width', 0, 20),
    'height': hp.uniform('height', -100, 100),
    'activation': hp.choice("activation", ["relu", "tanh"])
}
current_best_params = [{
    'width': 10,
    'height': 0,
    'activation': 0, # The index of "relu"
}]
algo = HyperOptSearch(
    space, metric="mean_loss", mode="min",
    points_to_evaluate=current_best_params)
Parameters
  • space (dict) – HyperOpt configuration. Parameters will be sampled from this configuration and will be used to override parameters generated in the variant generation process.

  • metric (str) – The training result objective value attribute.

  • mode (str) – One of {min, max}. Determines whether objective is minimizing or maximizing the metric attribute.

  • points_to_evaluate (list) – Initial parameter suggestions to be run first. This is for when you already have some good parameters you want hyperopt to run first to help the TPE algorithm make better suggestions for future parameters. Needs to be a list of dict of hyperopt-named variables. Choice variables should be indicated by their index in the list (see example)

  • n_initial_points (int) – number of random evaluations of the objective function before starting to aproximate it with tree parzen estimators. Defaults to 20.

  • random_state_seed (int, array_like, None) – seed for reproducible results. Defaults to None.

  • gamma (float in range (0,1)) – parameter governing the tree parzen estimators suggestion algorithm. Defaults to 0.25.

  • max_concurrent – Deprecated.

  • use_early_stopped_trials – Deprecated.

Nevergrad (tune.suggest.nevergrad.NevergradSearch)

class ray.tune.suggest.nevergrad.NevergradSearch(optimizer, parameter_names, metric='episode_reward_mean', mode='max', max_concurrent=None, **kwargs)[source]

Uses Nevergrad to optimize hyperparameters.

Nevergrad is an open source tool from Facebook for derivative free optimization. More info can be found at: https://github.com/facebookresearch/nevergrad.

You will need to install Nevergrad via the following command:

$ pip install nevergrad

This algorithm requires using an optimizer provided by Nevergrad, of which there are many options. A good rundown can be found on the Nevergrad README’s Optimization section.

from nevergrad.optimization import optimizerlib

instrumentation = 1
optimizer = optimizerlib.OnePlusOne(instrumentation, budget=100)
algo = NevergradSearch(
    optimizer, ["lr"], metric="mean_loss", mode="min")
Parameters
  • optimizer (nevergrad.optimization.Optimizer) – Optimizer provided from Nevergrad.

  • parameter_names (list) – List of parameter names. Should match the dimension of the optimizer output. Alternatively, set to None if the optimizer is already instrumented with kwargs (see nevergrad v0.2.0+).

  • metric (str) – The training result objective value attribute.

  • mode (str) – One of {min, max}. Determines whether objective is minimizing or maximizing the metric attribute.

  • use_early_stopped_trials – Deprecated.

  • max_concurrent – Deprecated.

Note

In nevergrad v0.2.0+, optimizers can be instrumented. For instance, the following will specifies searching for “lr” from 1 to 2.

>>> from nevergrad.optimization import optimizerlib
>>> from nevergrad import instrumentation as inst
>>> lr = inst.var.Array(1).bounded(1, 2).asfloat()
>>> instrumentation = inst.Instrumentation(lr=lr)
>>> optimizer = optimizerlib.OnePlusOne(instrumentation, budget=100)
>>> algo = NevergradSearch(
        optimizer, None, metric="mean_loss", mode="min")

SigOpt (tune.suggest.sigopt.SigOptSearch)

You will need to use the SigOpt experiment and space specification to specify your search space.

class ray.tune.suggest.sigopt.SigOptSearch(space, name='Default Tune Experiment', max_concurrent=1, reward_attr=None, metric='episode_reward_mean', mode='max', **kwargs)[source]

A wrapper around SigOpt to provide trial suggestions.

You must install SigOpt and have a SigOpt API key to use this module. Store the API token as an environment variable SIGOPT_KEY as follows:

pip install -U sigopt
export SIGOPT_KEY= ...

You will need to use the SigOpt experiment and space specification.

This module manages its own concurrency.

Parameters
  • space (list of dict) – SigOpt configuration. Parameters will be sampled from this configuration and will be used to override parameters generated in the variant generation process.

  • name (str) – Name of experiment. Required by SigOpt.

  • max_concurrent (int) – Number of maximum concurrent trials supported based on the user’s SigOpt plan. Defaults to 1.

  • metric (str) – The training result objective value attribute.

  • mode (str) – One of {min, max}. Determines whether objective is minimizing or maximizing the metric attribute.

Example:

space = [
    {
        'name': 'width',
        'type': 'int',
        'bounds': {
            'min': 0,
            'max': 20
        },
    },
    {
        'name': 'height',
        'type': 'int',
        'bounds': {
            'min': -100,
            'max': 100
        },
    },
]
algo = SigOptSearch(
    space, name="SigOpt Example Experiment",
    max_concurrent=1, metric="mean_loss", mode="min")

Scikit-Optimize (tune.suggest.skopt.SkOptSearch)

class ray.tune.suggest.skopt.SkOptSearch(optimizer, parameter_names, metric='episode_reward_mean', mode='max', points_to_evaluate=None, evaluated_rewards=None, max_concurrent=None, use_early_stopped_trials=None)[source]

Uses Scikit Optimize (skopt) to optimize hyperparameters.

Scikit-optimize is a black-box optimization library. Read more here: https://scikit-optimize.github.io.

You will need to install Scikit-Optimize to use this module.

pip install scikit-optimize

This Search Algorithm requires you to pass in a skopt Optimizer object.

Parameters
  • optimizer (skopt.optimizer.Optimizer) – Optimizer provided from skopt.

  • parameter_names (list) – List of parameter names. Should match the dimension of the optimizer output.

  • metric (str) – The training result objective value attribute.

  • mode (str) – One of {min, max}. Determines whether objective is minimizing or maximizing the metric attribute.

  • points_to_evaluate (list of lists) – A list of points you’d like to run first before sampling from the optimiser, e.g. these could be parameter configurations you already know work well to help the optimiser select good values. Each point is a list of the parameters using the order definition given by parameter_names.

  • evaluated_rewards (list) – If you have previously evaluated the parameters passed in as points_to_evaluate you can avoid re-running those trials by passing in the reward attributes as a list so the optimiser can be told the results without needing to re-compute the trial. Must be the same length as points_to_evaluate. (See tune/examples/skopt_example.py)

  • max_concurrent – Deprecated.

  • use_early_stopped_trials – Deprecated.

Example:

from skopt import Optimizer
optimizer = Optimizer([(0,20),(-100,100)])
current_best_params = [[10, 0], [15, -20]]

algo = SkOptSearch(optimizer,
    ["width", "height"],
    metric="mean_loss",
    mode="min",
    points_to_evaluate=current_best_params)

ZOOpt (tune.suggest.zoopt.ZOOptSearch)

class ray.tune.suggest.zoopt.ZOOptSearch(algo='asracos', budget=None, dim_dict=None, metric='episode_reward_mean', mode='min', **kwargs)[source]

A wrapper around ZOOpt to provide trial suggestions.

ZOOptSearch is a library for derivative-free optimization. It is backed by the ZOOpt package. Currently, Asynchronous Sequential RAndomized COordinate Shrinking (ASRacos) is implemented in Tune.

To use ZOOptSearch, install zoopt (>=0.4.0): pip install -U zoopt.

from ray.tune import run
from ray.tune.suggest.zoopt import ZOOptSearch
from zoopt import ValueType

dim_dict = {
    "height": (ValueType.CONTINUOUS, [-10, 10], 1e-2),
    "width": (ValueType.DISCRETE, [-10, 10], False)
}

config = {
    "num_samples": 200,
    "config": {
        "iterations": 10,  # evaluation times
    },
    "stop": {
        "timesteps_total": 10  # cumstom stop rules
    }
}

zoopt_search = ZOOptSearch(
    algo="Asracos",  # only support Asracos currently
    budget=config["num_samples"],
    dim_dict=dim_dict,
    metric="mean_loss",
    mode="min")

run(my_objective,
    search_alg=zoopt_search,
    name="zoopt_search",
    **config)
Parameters
  • algo (str) – To specify an algorithm in zoopt you want to use. Only support ASRacos currently.

  • budget (int) – Number of samples.

  • dim_dict (dict) – Dimension dictionary. For continuous dimensions: (continuous, search_range, precision); For discrete dimensions: (discrete, search_range, has_order). More details can be found in zoopt package.

  • metric (str) – The training result objective value attribute. Defaults to “episode_reward_mean”.

  • mode (str) – One of {min, max}. Determines whether objective is minimizing or maximizing the metric attribute. Defaults to “min”.

Repeated Evaluations (tune.suggest.Repeater)

Use ray.tune.suggest.Repeater to average over multiple evaluations of the same hyperparameter configurations. This is useful in cases where the evaluated training procedure has high variance (i.e., in reinforcement learning).

By default, Repeater will take in a repeat parameter and a search_alg. The search_alg will suggest new configurations to try, and the Repeater will run repeat trials of the configuration. It will then average the search_alg.metric from the final results of each repeated trial.

Warning

It is recommended to not use Repeater with a TrialScheduler. Early termination can negatively affect the average reported metric.

class ray.tune.suggest.Repeater(searcher, repeat=1, set_index=True)[source]

A wrapper algorithm for repeating trials of same parameters.

Set tune.run(num_samples=…) to be a multiple of repeat. For example, set num_samples=15 if you intend to obtain 3 search algorithm suggestions and repeat each suggestion 5 times. Any leftover trials (num_samples mod repeat) will be ignored.

It is recommended that you do not run an early-stopping TrialScheduler simultaneously.

Parameters
  • searcher (Searcher) – Searcher object that the Repeater will optimize. Note that the Searcher will only see 1 trial among multiple repeated trials. The result/metric passed to the Searcher upon trial completion will be averaged among all repeats.

  • repeat (int) – Number of times to generate a trial with a repeated configuration. Defaults to 1.

  • set_index (bool) – Sets a tune.suggest.repeater.TRIAL_INDEX in Trainable/Function config which corresponds to the index of the repeated trial. This can be used for seeds. Defaults to True.

Example:

from ray.tune.suggest import Repeater

search_alg = BayesOptSearch(...)
re_search_alg = Repeater(search_alg, repeat=10)

# Repeat 2 samples 10 times each.
tune.run(trainable, num_samples=20, search_alg=re_search_alg)

ConcurrencyLimiter (tune.suggest.ConcurrencyLimiter)

Use ray.tune.suggest.ConcurrencyLimiter to limit the amount of concurrency when using a search algorithm. This is useful when a given optimization algorithm does not parallelize very well (like a naive Bayesian Optimization).

class ray.tune.suggest.ConcurrencyLimiter(searcher, max_concurrent)[source]

A wrapper algorithm for limiting the number of concurrent trials.

Parameters

searcher (Searcher) – Searcher object that the ConcurrencyLimiter will manage.

Example:

from ray.tune.suggest import ConcurrencyLimiter
search_alg = HyperOptSearch(metric="accuracy")
search_alg = ConcurrencyLimiter(search_alg, max_concurrent=2)
tune.run(trainable, search_alg=search_alg)

Implementing your own Search Algorithm

If you are interested in implementing or contributing a new Search Algorithm, provide the following interface:

class ray.tune.suggest.Searcher(metric='episode_reward_mean', mode='max', max_concurrent=None, use_early_stopped_trials=None)[source]

Bases: object

Abstract class for wrapping suggesting algorithms.

Custom algorithms can extend this class easily by overriding the suggest method provide generated parameters for the trials.

Any subclass that implements __init__ must also call the constructor of this class: super(Subclass, self).__init__(...).

To track suggestions and their corresponding evaluations, the method suggest will be passed a trial_id, which will be used in subsequent notifications.

Parameters
  • metric (str) – The training result objective value attribute.

  • mode (str) – One of {min, max}. Determines whether objective is minimizing or maximizing the metric attribute.

class ExampleSearch(Searcher):
    def __init__(self, metric="mean_loss", mode="min", **kwargs):
        super(ExampleSearch, self).__init__(
            metric=metric, mode=mode, **kwargs)
        self.optimizer = Optimizer()
        self.configurations = {}

    def suggest(self, trial_id):
        configuration = self.optimizer.query()
        self.configurations[trial_id] = configuration

    def on_trial_complete(self, trial_id, result, **kwargs):
        configuration = self.configurations[trial_id]
        if result and self.metric in result:
            self.optimizer.update(configuration, result[self.metric])

tune.run(trainable_function, search_alg=ExampleSearch())
on_trial_result(trial_id, result)[source]

Optional notification for result during training.

Note that by default, the result dict may include NaNs or may not include the optimization metric. It is up to the subclass implementation to preprocess the result to avoid breaking the optimization process.

Parameters
  • trial_id (str) – A unique string ID for the trial.

  • result (dict) – Dictionary of metrics for current training progress. Note that the result dict may include NaNs or may not include the optimization metric. It is up to the subclass implementation to preprocess the result to avoid breaking the optimization process.

on_trial_complete(trial_id, result=None, error=False)[source]

Notification for the completion of trial.

Typically, this method is used for notifying the underlying optimizer of the result.

Parameters
  • trial_id (str) – A unique string ID for the trial.

  • result (dict) – Dictionary of metrics for current training progress. Note that the result dict may include NaNs or may not include the optimization metric. It is up to the subclass implementation to preprocess the result to avoid breaking the optimization process. Upon errors, this may also be None.

  • error (bool) – True if the training process raised an error.

suggest(trial_id)[source]

Queries the algorithm to retrieve the next set of parameters.

Parameters

trial_id (str) – Trial ID used for subsequent notifications.

Returns

Configuration for a trial, if possible.

Return type

dict|None

save(checkpoint_dir)[source]

Save function for this object.

restore(checkpoint_dir)[source]

Restore function for this object.

property metric

The training result objective value attribute.

property mode

Specifies if minimizing or maximizing the metric.