logml.model_search.hp_optimizer

Functions

`create_fmin`(hpo_cfg)	Parses a given HPO config and returns a fmin with predefined params.
`create_hopt_space`(params_space_cfg[, ...])	Transforms a given params space config into a HyperOpt space Also supports embedded structures.

Classes

`HpoResult`(name[, best_params, trials, ...])	Result of HyperParamsOptimizerCore class.
`HyperParamsOptimizerCore`([hpo_algorithm, ...])	Applies HyperOpt to find the best hyperparams for a given model.
`ModelHpoResult`(name[, best_params, trials, ...])	Model HPO result
`TrialResults`()	Utility for retrieving hyperopt-compatible trial results.

Exceptions

HyperParamsOptimizerException

Exception raised when there is an Exception caught while hyper-parameters optimization.

class logml.model_search.hp_optimizer.HpoResult(name: str, best_params: Optional[dict] = None, trials: Optional[hyperopt.base.Trials] = None, all_trials_failed: bool = False)

Bases: object

Result of HyperParamsOptimizerCore class.

name: str

best_params: Optional[dict] = None

trials: hyperopt.base.Trials = None

all_trials_failed: bool = False

class logml.model_search.hp_optimizer.HyperParamsOptimizerCore(hpo_algorithm='tpe', max_evals=10, logger=None, show_progressbar=True, random_state=None)

Bases: object

Applies HyperOpt to find the best hyperparams for a given model. Class is stateless and can be reused to train different models without side effects.

By the principle of separation of core components and infrastructure: HyperParamsOptimizerCore can only calculate, while HyperParamsOptimizer wraps it, and can save-load data as needed.

search(model_cls: Type[logml.models.base.BaseModel], hpo_space: dict, dataset: logml.data.datasets.cv_dataset.ModelingDataset, default_params: Optional[dict] = None, random_state=None, test_default_params=True) → logml.model_search.hp_optimizer.HpoResult: Runs HyperOpt to find the best params for a given model.

class logml.model_search.hp_optimizer.ModelHpoResult(name: str, best_params: Optional[dict] = None, trials: Optional[List[dict]] = None, all_trials_failed: bool = False)

Bases: object

Model HPO result

name: str

best_params: Optional[dict] = None

trials: Optional[List[dict]] = None

all_trials_failed: bool = False

logml.model_search.hp_optimizer.create_hopt_space(params_space_cfg: Optional[dict], default_params=None)

Transforms a given params space config into a HyperOpt space Also supports embedded structures.

Returns hyperopt search space.

logml.model_search.hp_optimizer.create_fmin(hpo_cfg: logml.configuration.modeling.HPOSection): Parses a given HPO config and returns a fmin with predefined params.

exception logml.model_search.hp_optimizer.HyperParamsOptimizerException

Bases: Exception

Exception raised when there is an Exception caught while hyper-parameters optimization.

class logml.model_search.hp_optimizer.TrialResults

Bases: object

Utility for retrieving hyperopt-compatible trial results.

static get_cv_scores(model: logml.models.base.BaseModel, dataset: logml.data.datasets.cv_dataset.ModelingDataset): Returns evaluation results. Also casts np.float into float, for serialization.

static get_loss(model: logml.models.base.BaseModel, dataset: logml.data.datasets.cv_dataset.ModelingDataset) → float: Calculates the result loss based on CV scores.

static extract(model, dataset: logml.data.datasets.cv_dataset.ModelingDataset): Extracts trial results from a given model and dataset.

static create_failed_result(message): Returns the default placeholder for failed trials.