peal.adaptors.group_distributionally_robust_optimization.GroupDROConfig¶

class peal.adaptors.group_distributionally_robust_optimization.GroupDROConfig[source]¶

Bases: AdaptorConfig

Config template for running the DRO adaptor.

Methods

Attributes

`adaptor_type`	The config template for an adaptor
`alpha`
`btl`
`gamma`
`generalization_adjustment`
`is_robust`	Parameters for the Gorup_DRO loss computer
`kwargs`	A dict containing all variables that could not be given with the current config structure
`minimum_variational_weight`
`model_path`	The path where the model is to be stored.
`predictor`	The config of the predictor to be adapted.
`replication_dataset`	Use paper's datasets.
`reset_weights`	Resets weights of the model if true.
`reweight_groups`	Reweights groups during training so the training set is balanced.
`robust_step_size`
`scheduler`	Use ReduceLROnPlateau scheduler if true.
`seed`	Sets the seed of the run.
`use_normalized_loss`

adaptor_type: str = 'GroupDRO'¶: The config template for an adaptor

predictor: PredictorConfig = None¶: The config of the predictor to be adapted.

is_robust: bool = False¶: Parameters for the Gorup_DRO loss computer

alpha: float = None¶

gamma: float = 0.1¶

generalization_adjustment: float = 0.0¶

minimum_variational_weight: float = 0.0¶

robust_step_size: float = 0.01¶

use_normalized_loss: bool = False¶

btl: bool = False¶

reweight_groups: bool = False¶: Reweights groups during training so the training set is balanced.

replication_dataset: str = None¶: Use paper’s datasets. Only accepts “celebA”, “waterbirds” and None, with None indicating to not use the paper’s datasets, but rather the one specified in predictor’s data config.

reset_weights: bool = True¶: Resets weights of the model if true.

scheduler: bool = False¶: Use ReduceLROnPlateau scheduler if true.

model_path: str = None¶: The path where the model is to be stored. Explicitly overwrites model_path in predictor.

seed: int = 0¶: Sets the seed of the run. Expliticly overwrites the seed parameter in predictor.

kwargs: dict = {}¶: A dict containing all variables that could not be given with the current config structure