releso.validation.Validation

class releso.validation.Validation(*, save_location: Path, logger_name: str | None = None, validation_freq: ConstrainedIntValue | None = None, validation_values: ConstrainedListValue[float], save_best_agent: bool, validate_on_training_end: bool, max_timesteps_in_episode: ConstrainedIntValue | None = None, end_episode_on_geometry_not_changed: bool = False, reward_on_geometry_not_changed: float | None = None, reward_on_episode_exceeds_max_timesteps: float | None = None)

Bases: BaseModel

Parser class to define the validation to be performed during training.

This class is used for the configuration of how validation is to be performed during training.

__init__(**data: Any) → None: Constructor for the ReLeSO basemodel object.

Methods

`convert_to_pathlib_add_datetime`(v)	Add timestamp to save_location, of applicable.
`end_validation`(agent, environment)	Function is called at the end of a validation.
`get_callback`(eval_environment[, ...])	Creates the EvalCallback with the values given in this object.
`get_environment_validation_parameters`()	Gather the validation arguments used to initialize validator.
`get_logger`()	Gets the currently defined environment logger.
`set_logger_name_recursively`(logger_name)	Set the logger_name variable for all child elements.
`should_add_callback`()	Bool function whether or not a validation callback is needed.

Attributes

`validation_freq`	How many timesteps should pass by learning between validation runs
`validation_values`	List of validation items.
`save_best_agent`	Whether or not to save the best agent.
`max_timesteps_in_episode`	after how many timesteps inside a single episode should the episode be terminated.
`end_episode_on_geometry_not_changed`	Should the episode be terminated if the geometry representation has not changed between timesteps?
`reward_on_geometry_not_changed`	What reward should be added to the step reward if the geometry was not changed for the defined number of steps.
`reward_on_episode_exceeds_max_timesteps`	What reward should be added to the step reward if the maximal timesteps per episode is exceeded.
`save_location`	Definition of the save location of the logs and validation results.
`logger_name`	name of the logger.

end_episode_on_geometry_not_changed: bool: Should the episode be terminated if the geometry representation has not changed between timesteps?

end_validation(agent: BaseAlgorithm, environment: gymnasium.Env | VecEnv) → Tuple[float, float]

Function is called at the end of a validation.

All clean up and last evaluation is going in here.

Parameters:

agent (BaseAlgorithm) – Agent which is to be used to validate.
environment (GymEnv) – Validation environment

Returns:

See ‘funct’evaluate_policy() for definition

Return type:

Tuple[float, float]

get_callback(eval_environment: gymnasium.Env | VecEnv, save_location: Path | None = None, normalizer_divisor: int = 1) → EvalCallback

Creates the EvalCallback with the values given in this object.

Parameters:

eval_environment (GymEnv) – Evaluation environment. Should be the same as the normal training environment only that here the goal values should be set and not random.
save_location (Optional[pathlib.Path]) – Path to where the best models should be save to.
normalizer_divisor (int, optional) – Divisor for the eval_freq. Defaults to 1.

Returns:

Validation callback parametrized by this object.

Return type:

EvalCallback

get_environment_validation_parameters() → Dict[str, Any]

Gather the validation arguments used to initialize validator.

Gets the validation parameters that need to be send to the environment if it gets converted to be a validation environment.

Returns:: dict with all the necessary parameters. Should mirror the parameters in parser_environment.Environment.set_validation
Return type:: Dict[str, Any]

max_timesteps_in_episode: ConstrainedIntValue | None: after how many timesteps inside a single episode should the episode be terminated.

reward_on_episode_exceeds_max_timesteps: float | None: What reward should be added to the step reward if the maximal timesteps per episode is exceeded.

reward_on_geometry_not_changed: float | None: What reward should be added to the step reward if the geometry was not changed for the defined number of steps.

save_best_agent: bool: Whether or not to save the best agent. If agent is not saved only results will be reported by the agent which produced them will not be saved.

should_add_callback() → bool

Bool function whether or not a validation callback is needed.

Returns:: Return True if a validation callback is needed else False.
Return type:: bool

validation_freq: ConstrainedIntValue | None: How many timesteps should pass by learning between validation runs

validation_values: ConstrainedListValue[float]: List of validation items. This will be revised later on #TODO