Offline Training

offline_training ¶

Pretraining of a surrogate model on pseudo-random parameters.

Classes:

Name	Description
`OfflineTrainingSettings`	Configuration of the offline training run.
`OfflineTrainer`	Class for pretraining of surrogate models.
`OfflineTrainingLogger`	Logger for information during pretraining.

surrogate.offline_training.OfflineTrainingSettings `dataclass` ¶

Configuration of the offline training run.

Attributes:

Name	Type	Description
`num_offline_training_points`	`int`	Number of parameter samples to generate for training through Latin Hypercube Sampling.
`num_threads`	`int`	Number of parallel threads to use for pretraining. Only makes sense if the calls to the simulation model are dispatched to an actually parallel setup
`offline_model_config`	`dict`	Configuration of UMBridge calls to the simulation model server.
`lhs_bounds`	`list`	Dimension-wise bounds for the Latin Hypercube Sampling.
`lhs_seed`	`list`	Seed for the Latin Hypercube Sampling.
`checkpoint_save_name`	`Path`	Name of the checkpoint file to save the surrogate model and data to.

surrogate.offline_training.OfflineTrainer ¶

Class for pretraining of surrogate models.

Implements simple pretraining without the asynchronous server. Input parameters are generated via Latin Hypercube Sampling on the domain of interest. Outputs are obtained from calls to a simulation model server.

Methods:

Name	Description
`run`	Execute the pretraining.

init ¶

__init__(
    training_settings: OfflineTrainingSettings,
    logger_settings: utilities.LoggerSettings,
    surrogate_model: surrogate_model.BaseSurrogateModel,
    simulation_model: Callable,
) -> None

Constructor.

Parameters:

Name	Type	Description	Default
`training_settings`	`OfflineTrainingSettings`	Configuration of the offline training run.	required
`logger_settings`	`utilities.LoggerSettings`	Configuration of the logger	required
`surrogate_model`	`surrogate_model.BaseSurrogateModel`	Surrogate model to train	required
`simulation_model`	`Callable`	Simulation model to request evaluations from to generate training data	required

run ¶

run() -> None

Execute pretraining, based on LHS exploration of the parameter space.

surrogate.offline_training.OfflineTrainingLogger ¶

Bases: utilities.BaseLogger

Logger for information during pretraining.

The logger records to events: 1. Generation of training data sample (input and output) 2. Fitting of the surrogate

init ¶

__init__(logger_settings: utilities.LoggerSettings) -> None

Constructor.

Parameters:

Name	Type	Description	Default
`logger_settings`	`utilities.LoggerSettings`	Configuration of the logger	required

log_simulation_run ¶

log_simulation_run(parameter: float | Iterable, result: float) -> None

Log information on generation of a training sample.

Parameters:

Name	Type	Description	Default
`parameter`	`float \| Iterable`	Input parameter	required
`result`	`float`	Simulation model result	required

log_surrogate_fit ¶

log_surrogate_fit(scale: float | Iterable, correlation_length: float | Iterable) -> None

Log information on the fitting of the surrogate.

Parameters:

Name	Type	Description	Default
`scale`	`float`	Log scale parameter of the surrogate model (for GPs)	required
`correlation_length`	`float \| Iterable`	Log correlation length per dimension of the surrogate model (for GPs)	required

Offline Training