releso.base_parser.BaseParser

class releso.base_parser.BaseParser(*, save_location: Path, logger_name: str | None = None, verbosity: Verbosity = None, agent: PPOAgent | DDPGAgent | SACAgent | PretrainedAgent | DQNAgent | A2CAgent, environment: Environment, number_of_timesteps: ConstrainedIntValue, number_of_episodes: ConstrainedIntValue | None = None, validation: Validation | None = None, n_environments: ConstrainedIntValue | None = 1, normalize_training_values: bool = False, multi_env_sequential: bool = False, episode_log_update: ConstrainedIntValue = 100, export_step_log: bool = False, step_log_update: ConstrainedIntValue = 0, step_log_info: bool = False)

Bases: BaseModel

Class parses the experiment definition and conducts the training.

This class can be used to initialize the ReLeSO Framework from the command line by reading in a json representation of the RL based Shape Optimization problem which is to be solved via Reinforcement Learning.

__init__(**data: Any) → None

Constructor of the base parser.

Initializes the class correctly and also adds the correct logger to all subclasses.

Methods

`convert_to_pathlib_add_datetime`(v)	Add timestamp to save_location, of applicable.
`evaluate_model`([validation_env, ...])	Validate the current agent.
`get_logger`()	Gets the currently defined environment logger.
`learn`()	Starts the training that is specified in the loaded json file.
`save_model`([file_name])	Save the state of the agent.
`set_logger_name_recursively`(logger_name)	Set the logger_name variable for all child elements.

Attributes

`verbosity`	Defining the verbosity of the training process and environment loading
`agent`	Definition of the agent to be used during training and/or validation of the RL use case
`environment`	Definition of the environment which encodes the parameters of the RL use case
`number_of_timesteps`	Number of timesteps the training process should run for
`number_of_episodes`	Number of episodes the training process should run for.
`validation`	Definition of the validation .
`n_environments`	Number of environments to train in parallel.
`normalize_training_values`	but the training might be a little bit more unstable.
`multi_env_sequential`	Should the multi environment be run sequentially (True) or with multi processing (False).
`episode_log_update`	Number of episodes after which the episode_log is updated.
`export_step_log`	Flag indicating whether the step information (like actions, observations, ...) should be logged to file.
`step_log_update`	Number of steps after which the step_log is updated.
`step_log_info`	Flag indicating whether the step_log should also contain the information of the environment step.
`save_location`	Definition of the save location of the logs and validation results.
`logger_name`	name of the logger.

agent: PPOAgent | DDPGAgent | SACAgent | PretrainedAgent | DQNAgent | A2CAgent: Definition of the agent to be used during training and/or validation of the RL use case

environment: Environment: Definition of the environment which encodes the parameters of the RL use case

episode_log_update: ConstrainedIntValue: Number of episodes after which the episode_log is updated. It will be updated at the end of the training in any case. But making this number higher will lower the computational overhead. Defaults to 100.

evaluate_model(validation_env: None | Environment = None, throw_error_if_none: bool = False) → None

Validate the current agent.

Evaluate the model with the parameters defined in the validation variable. If an agent is already loaded use this agent, else get the agent from the agent variable. Validation will be done inside the

Parameters:

validation_env (Union[None, Environment], optional) – If validation
validation (environment already exists it will be used else a new)
None. (environment will be created. Defaults to)
throw_error_if_none (bool, optional) – If this is set and the
False. (validation variable is None an error is thrown. Defaults to)

Raises:

ValidationNotSet – Thrown if validation is absolutely needed.
If not absolutely needed the validation will not be done but no –
error will be thrown. –

export_step_log: bool: Flag indicating whether the step information (like actions, observations, …) should be logged to file. Defaults to False.

learn() → None

Starts the training that is specified in the loaded json file.

Successive calls to this function does not train the agent further

but reinitialize the agent.

multi_env_sequential: bool: Should the multi environment be run sequentially (True) or with multi processing (False). Defaults to False.

n_environments: ConstrainedIntValue | None: Number of environments to train in parallel. Defaults to None.

normalize_training_values: bool: but the training might be a little bit more unstable. Defaults to False.

number_of_episodes: ConstrainedIntValue | None: Number of episodes the training process should run for. If given both timesteps and max episodes can stop the trainings progress. Default: None

number_of_timesteps: ConstrainedIntValue: Number of timesteps the training process should run for

save_model(file_name: str | None = None) → str

Save the state of the agent.

Saves the current agent to the specified location or to a default location.

Parameters:

file_name (Optional[str]) – Path where the agent is saved to. If
will (None will see if json had a save location if also not given)
None. (use a default location. Defaults to)

Returns:

Path where the agent was saved to.

Return type:

str

step_log_info: bool: Flag indicating whether the step_log should also contain the information of the environment step. Defaults to False.

step_log_update: ConstrainedIntValue: Number of steps after which the step_log is updated. It will be updated at the end of the training in any case. But making this number higher will lower the computational overhead. Defaults to 0 which triggers the output after every episode.

validation: Validation | None: Definition of the validation . Defaults to None.

verbosity: Verbosity: Defining the verbosity of the training process and environment loading