Contents

Scikit-Learn API (tune.sklearn)

TuneGridSearchCV

class ray.tune.sklearn.TuneGridSearchCV(estimator, param_grid, early_stopping=None, scoring=None, n_jobs=None, cv=5, refit=True, verbose=0, error_score='raise', return_train_score=False, local_dir='~/ray_results', name=None, max_iters=1, use_gpu=False, loggers=None, pipeline_auto_early_stop=True, stopper=None, time_budget_s=None, sk_n_jobs=None, mode=None)[source]

Exhaustive search over specified parameter values for an estimator.

Important members are fit, predict.

GridSearchCV implements a “fit” and a “score” method. It also implements “predict”, “predict_proba”, “decision_function”, “transform” and “inverse_transform” if they are implemented in the estimator used.

The parameters of the estimator used to apply these methods are optimized by cross-validated grid-search over a parameter grid.

Parameters

  • estimator (estimator) – Object that implements the scikit-learn estimator interface. Either estimator needs to provide a score function, or scoring must be passed.

  • param_grid (dict or list of dict) – Dictionary with parameters names (string) as keys and lists or tune.grid_search outputs of parameter settings to try as values, or a list of such dictionaries, in which case the grids spanned by each dictionary in the list are explored. This enables searching over any sequence of parameter settings.

  • early_stopping (bool, str or TrialScheduler, optional) –

    Option to stop fitting to a hyperparameter configuration if it performs poorly. Possible inputs are:

    • If True, defaults to ASHAScheduler.

    • A string corresponding to the name of a Tune Trial Scheduler (i.e., “ASHAScheduler”). To specify parameters of the scheduler, pass in a scheduler object instead of a string.

    • Scheduler for executing fit with early stopping. Only a subset of schedulers are currently supported. The scheduler will only be used if the estimator supports partial fitting

    • If None or False, early stopping will not be used.

  • scoring (str, list/tuple, dict, or None) – A single string or a callable to evaluate the predictions on the test set. See https://scikit-learn.org/stable/modules/model_evaluation.html #scoring-parameter for all options. For evaluating multiple metrics, either give a list/tuple of (unique) strings or a dict with names as keys and callables as values. If None, the estimator’s score method is used. Defaults to None.

  • n_jobs (int) – Number of jobs to run in parallel. None or -1 means using all processors. Defaults to None. If set to 1, jobs will be run using Ray’s ‘local mode’. This can lead to significant speedups if the model takes < 10 seconds to fit due to removing inter-process communication overheads.

  • cv (int, cross-validation generator or iterable) –

    Determines the cross-validation splitting strategy. Possible inputs for cv are:

    • None, to use the default 5-fold cross validation,

    • integer, to specify the number of folds in a (Stratified)KFold,

    • An iterable yielding (train, test) splits as arrays of indices.

    For integer/None inputs, if the estimator is a classifier and y is either binary or multiclass, StratifiedKFold is used. In all other cases, KFold is used. Defaults to None.

  • refit (bool or str) – Refit an estimator using the best found parameters on the whole dataset. For multiple metric evaluation, this needs to be a string denoting the scorer that would be used to find the best parameters for refitting the estimator at the end. The refitted estimator is made available at the best_estimator_ attribute and permits using predict directly on this GridSearchCV instance. Also for multiple metric evaluation, the attributes best_index_, best_score_ and best_params_ will only be available if refit is set and all of them will be determined w.r.t this specific scorer. If refit not needed, set to False. See scoring parameter to know more about multiple metric evaluation. Defaults to True.

  • verbose (int) – Controls the verbosity: 0 = silent, 1 = only status updates, 2 = status and trial results. Defaults to 0.

  • error_score ('raise' or int or float) – Value to assign to the score if an error occurs in estimator fitting. If set to ‘raise’, the error is raised. If a numeric value is given, FitFailedWarning is raised. This parameter does not affect the refit step, which will always raise the error. Defaults to np.nan.

  • return_train_score (bool) – If False, the cv_results_ attribute will not include training scores. Defaults to False. Computing training scores is used to get insights on how different parameter settings impact the overfitting/underfitting trade-off. However computing the scores on the training set can be computationally expensive and is not strictly required to select the parameters that yield the best generalization performance.

  • local_dir (str) – A string that defines where checkpoints will be stored. Defaults to “~/ray_results”.

  • name (str) – Name of experiment (for Ray Tune) –

  • max_iters (int) – Indicates the maximum number of epochs to run for each hyperparameter configuration sampled. This parameter is used for early stopping. Defaults to 1. Depending on the classifier type provided, a resource parameter (resource_param = max_iter or n_estimators) will be detected. The value of resource_param will be treated as a “max resource value”, and all classifiers will be initialized with max resource value // max_iters, where max_iters is this defined parameter. On each epoch, resource_param (max_iter or n_estimators) is incremented by max resource value // max_iters.

  • use_gpu (bool) – Indicates whether to use gpu for fitting. Defaults to False. If True, training will start processes with the proper CUDA VISIBLE DEVICE settings set. If a Ray cluster has been initialized, all available GPUs will be used.

  • loggers (list) – A list of the names of the Tune loggers as strings to be used to log results. Possible values are “tensorboard”, “csv”, “mlflow”, and “json”

  • pipeline_auto_early_stop (bool) – Only relevant if estimator is Pipeline object and early_stopping is enabled/True. If True, early stopping will be performed on the last stage of the pipeline (which must support early stopping). If False, early stopping will be determined by ‘Pipeline.warm_start’ or ‘Pipeline.partial_fit’ capabilities, which are by default not supported by standard SKlearn. Defaults to True.

  • stopper (ray.tune.stopper.Stopper) – Stopper objects passed to tune.run().

  • time_budget_s (int|float|datetime.timedelta) – Global time budget in seconds after which all trials are stopped. Can also be a datetime.timedelta object.

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

TuneSearchCV

class ray.tune.sklearn.TuneSearchCV(estimator, param_distributions, early_stopping=None, n_trials=10, scoring=None, n_jobs=None, refit=True, cv=None, verbose=0, random_state=None, error_score=nan, return_train_score=False, local_dir='~/ray_results', name=None, max_iters=1, search_optimization='random', use_gpu=False, loggers=None, pipeline_auto_early_stop=True, stopper=None, time_budget_s=None, sk_n_jobs=None, mode=None, search_kwargs=None, **kwargs)[source]

Generic, non-grid search on hyper parameters.

Randomized search is invoked with search_optimization set to "random" and behaves like scikit-learn’s RandomizedSearchCV.

Bayesian search can be invoked with several values of search_optimization.

Tree-Parzen Estimators search is invoked with search_optimization set to "hyperopt", using HyperOpt - http://hyperopt.github.io/hyperopt

All types of search aside from Randomized search require parent libraries to be installed.

TuneSearchCV implements a “fit” and a “score” method. It also implements “predict”, “predict_proba”, “decision_function”, “transform” and “inverse_transform” if they are implemented in the estimator used.

The parameters of the estimator used to apply these methods are optimized by cross-validated search over parameter settings.

In contrast to GridSearchCV, not all parameter values are tried out, but rather a fixed number of parameter settings is sampled from the specified distributions. The number of parameter settings that are tried is given by n_trials.

Parameters

  • estimator (estimator) – This is assumed to implement the scikit-learn estimator interface. Either estimator needs to provide a score function, or scoring must be passed.

  • param_distributions (dict or list or ConfigurationSpace) –

    Serves as the param_distributions parameter in scikit-learn’s RandomizedSearchCV or as the search_space parameter in BayesSearchCV. For randomized search: dictionary with parameters names (string) as keys and distributions or lists of parameter settings to try for randomized search. Distributions must provide a rvs method for sampling (such as those from scipy.stats.distributions). Ray Tune search spaces are also supported. If a list is given, it is sampled uniformly. If a list of dicts is given, first a dict is sampled uniformly, and then a parameter is sampled using that dict as above. For other types of search: dictionary with parameter names (string) as keys. Values can be

    • a (lower_bound, upper_bound) tuple (for Real or Integer params)

    • a (lower_bound, upper_bound, “prior”) tuple (for Real params)

    • as a list of categories (for Categorical dimensions)

    • Ray Tune search space (eg. tune.uniform)

    "bayesian" (scikit-optimize) also accepts

    • skopt.space.Dimension instance (Real, Integer or Categorical).

    "hyperopt" (HyperOpt) also accepts

    • an instance of a hyperopt.pyll.base.Apply object.

    "bohb" (HpBandSter) also accepts

    • ConfigSpace.hyperparameters.Hyperparameter instance.

    "optuna" (Optuna) also accepts

    • an instance of a optuna.distributions.BaseDistribution object.

    For "bohb" (HpBandSter) it is also possible to pass a ConfigSpace.ConfigurationSpace object instead of dict or a list.

    https://scikit-optimize.github.io/stable/modules/ classes.html#module-skopt.space.space

  • early_stopping (bool, str or TrialScheduler, optional) –

    Option to stop fitting to a hyperparameter configuration if it performs poorly. Possible inputs are:

    • If True, defaults to ASHAScheduler.

    • A string corresponding to the name of a Tune Trial Scheduler (i.e., “ASHAScheduler”). To specify parameters of the scheduler, pass in a scheduler object instead of a string.

    • Scheduler for executing fit with early stopping. Only a subset of schedulers are currently supported. The scheduler will only be used if the estimator supports partial fitting

    • If None or False, early stopping will not be used.

    Unless a HyperBandForBOHB object is passed, this parameter is ignored for "bohb", as it requires HyperBandForBOHB.

  • n_trials (int) – Number of parameter settings that are sampled. n_trials trades off runtime vs quality of the solution. Defaults to 10.

  • scoring (str, callable, list/tuple, dict, or None) – A single string or a callable to evaluate the predictions on the test set. See https://scikit-learn.org/stable/modules/model_evaluation.html #scoring-parameter for all options. For evaluating multiple metrics, either give a list/tuple of (unique) strings or a dict with names as keys and callables as values. If None, the estimator’s score method is used. Defaults to None.

  • n_jobs (int) – Number of jobs to run in parallel. None or -1 means using all processors. Defaults to None. If set to 1, jobs will be run using Ray’s ‘local mode’. This can lead to significant speedups if the model takes < 10 seconds to fit due to removing inter-process communication overheads.

  • refit (bool, str, or callable) – Refit an estimator using the best found parameters on the whole dataset. For multiple metric evaluation, this needs to be a string denoting the scorer that would be used to find the best parameters for refitting the estimator at the end. The refitted estimator is made available at the best_estimator_ attribute and permits using predict directly on this GridSearchCV instance. Also for multiple metric evaluation, the attributes best_index_, best_score_ and best_params_ will only be available if refit is set and all of them will be determined w.r.t this specific scorer. If refit not needed, set to False. See scoring parameter to know more about multiple metric evaluation. Defaults to True.

  • cv (int, cross-validation generator or iterable) –

    Determines the cross-validation splitting strategy. Possible inputs for cv are:

    • None, to use the default 5-fold cross validation,

    • integer, to specify the number of folds in a (Stratified)KFold,

    • An iterable yielding (train, test) splits as arrays of indices.

    For integer/None inputs, if the estimator is a classifier and y is either binary or multiclass, StratifiedKFold is used. In all other cases, KFold is used. Defaults to None.

  • verbose (int) – Controls the verbosity: 0 = silent, 1 = only status updates, 2 = status and trial results. Defaults to 0.

  • random_state (int or RandomState) – Pseudo random number generator state used for random uniform sampling from lists of possible values instead of scipy.stats distributions. If int, random_state is the seed used by the random number generator; If RandomState instance, a seed is sampled from random_state; If None, the random number generator is the RandomState instance used by np.random and no seed is provided. Defaults to None. Ignored when using BOHB.

  • error_score ('raise' or int or float) – Value to assign to the score if an error occurs in estimator fitting. If set to ‘raise’, the error is raised. If a numeric value is given, FitFailedWarning is raised. This parameter does not affect the refit step, which will always raise the error. Defaults to np.nan.

  • return_train_score (bool) – If False, the cv_results_ attribute will not include training scores. Defaults to False. Computing training scores is used to get insights on how different parameter settings impact the overfitting/underfitting trade-off. However computing the scores on the training set can be computationally expensive and is not strictly required to select the parameters that yield the best generalization performance.

  • local_dir (str) – A string that defines where checkpoints and logs will be stored. Defaults to “~/ray_results”

  • name (str) – Name of experiment (for Ray Tune) –

  • max_iters (int) – Indicates the maximum number of epochs to run for each hyperparameter configuration sampled (specified by n_trials). This parameter is used for early stopping. Defaults to 1. Depending on the classifier type provided, a resource parameter (resource_param = max_iter or n_estimators) will be detected. The value of resource_param will be treated as a “max resource value”, and all classifiers will be initialized with max resource value // max_iters, where max_iters is this defined parameter. On each epoch, resource_param (max_iter or n_estimators) is incremented by max resource value // max_iters.

  • ("random" or "bayesian" or "bohb" or "hyperopt" (search_optimization) –

    or “optuna” or ray.tune.suggest.Searcher instance): Randomized search is invoked with search_optimization set to "random" and behaves like scikit-learn’s RandomizedSearchCV.

    Bayesian search can be invoked with several values of search_optimization.

    Tree-Parzen Estimators search is invoked with search_optimization set to "hyperopt" via HyperOpt: http://hyperopt.github.io/hyperopt

    All types of search aside from Randomized search require parent libraries to be installed.

    Alternatively, instead of a string, a Ray Tune Searcher instance can be used, which will be passed to tune.run().

  • use_gpu (bool) – Indicates whether to use gpu for fitting. Defaults to False. If True, training will start processes with the proper CUDA VISIBLE DEVICE settings set. If a Ray cluster has been initialized, all available GPUs will be used.

  • loggers (list) – A list of the names of the Tune loggers as strings to be used to log results. Possible values are “tensorboard”, “csv”, “mlflow”, and “json”

  • pipeline_auto_early_stop (bool) – Only relevant if estimator is Pipeline object and early_stopping is enabled/True. If True, early stopping will be performed on the last stage of the pipeline (which must support early stopping). If False, early stopping will be determined by ‘Pipeline.warm_start’ or ‘Pipeline.partial_fit’ capabilities, which are by default not supported by standard SKlearn. Defaults to True.

  • stopper (ray.tune.stopper.Stopper) – Stopper objects passed to tune.run().

  • time_budget_s (int|float|datetime.timedelta) – Global time budget in seconds after which all trials are stopped. Can also be a datetime.timedelta object. The stopping condition is checked after receiving a result, i.e. after each training iteration.

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

  • search_kwargs (dict) – Additional arguments to pass to the SearchAlgorithms (tune.suggest) objects.

Trial Schedulers (tune.schedulers) Stopping mechanisms (tune.stopper)