peal.adaptors.group_dro_v2.GroupDROv2Config

class peal.adaptors.group_dro_v2.GroupDROv2Config

Bases: AdaptorConfig

Config for GroupDROv2.

Parameters:

• model_path (str) – torch.load-able classifier to repair.

• base_dir (str) – Run directory; an existing one is renamed with a timestamp suffix.

• data (DataConfig) – Training/validation data, which must provide group labels.

• unpoisoned_data (DataConfig, optional) – Second dataset whose test split is used for the unbiased evaluation.

• training (TrainingConfig) – Supplies learning_rate, max_epochs, test_batch_size and early_stopping_goal ("worst_group_accuracy" or "average_group_accuracy").

• task (TaskConfig) – Task of the classifier, passed on to the test dataloader.

• alpha (float, optional) – Tail fraction of the BTL/greedy weighting. Default 0.2.

• generalization_adjustment (float or list of float, optional) – Per-group adjustment C_g; a scalar is broadcast to all 4 groups.

• automatic_adjustment (bool, optional) – Re-estimate the adjustment after every epoch from the train/validation generalization gap.

• robust_step_size (float, optional) – Step size of the adversarial weight update. Default 0.01.

• use_normalized_loss – Forwarded to GroupDroLossComputer.

• btl – Forwarded to GroupDroLossComputer.

• gamma – Forwarded to GroupDroLossComputer.

• minimum_variational_weight – Forwarded to GroupDroLossComputer.

• weight_decay (float, optional) – L2 penalty of the SGD optimizer. Default 5e-5.

• track_test_acc (int, optional) – Evaluate on the unpoisoned test split every n epochs; None disables the intermediate test passes.

• save_intermediate (bool, optional) – Write checkpoints/model_epoch<n>.cpl after every epoch.

Methods

Attributes

alpha
automatic_adjustment
btl
gamma
generalization_adjustment
minimum_variational_weight
robust_step_size
save_intermediate
track_test_acc
unpoisoned_data
use_normalized_loss
weight_decay
model_path
base_dir
data
training
task
adaptor_type	The type of adaptor that shall be used.

model_path: str

base_dir: str

data: DataConfig

unpoisoned_data: DataConfig = None

training: TrainingConfig

task: TaskConfig

alpha: float = 0.2

generalization_adjustment: float | list[float] = 0.0

automatic_adjustment: bool = False

robust_step_size: float = 0.01

use_normalized_loss: bool = False

btl: bool = False

weight_decay: float = 5e-05

gamma: float = 0.1

minimum_variational_weight: float = 0.0

track_test_acc: int = None

save_intermediate: bool = True