GOOD.ood_algorithms.algorithms.GroupDRO

Implementation of the GroupDRO algorithm from “Distributionally Robust Neural Networks for Group Shifts: On the Importance of Regularization for Worst-Case Generalization” paper

Classes

GroupDRO(config)

Implementation of the GroupDRO algorithm from "Distributionally Robust Neural Networks for Group Shifts: On the Importance of Regularization for Worst-Case Generalization" paper
class GOOD.ood_algorithms.algorithms.GroupDRO.GroupDRO(config: Union[CommonArgs, Munch])[source]

Bases: BaseOODAlg

Implementation of the GroupDRO algorithm from “Distributionally Robust Neural Networks for Group Shifts: On the Importance of Regularization for Worst-Case Generalization” paper

Args:

config (Union[CommonArgs, Munch]): munchified dictionary of args (config.device, config.dataset.num_envs, config.ood.ood_param)

loss_postprocess(loss: Tensor, data: Batch, mask: Tensor, config: Union[CommonArgs, Munch], **kwargs) → Tensor[source]

Process loss based on GroupDRO algorithm

Parameters

  • loss (Tensor) – base loss between model predictions and input labels

  • data (Batch) – input data

  • mask (Tensor) – NAN masks for data formats

  • config (Union[CommonArgs, Munch]) – munchified dictionary of args (config.device, config.dataset.num_envs, config.ood.ood_param)

config = munchify({device: torch.device('cuda'),
                       dataset: {num_envs: int(10)},
                       ood: {ood_param: float(0.1)}
                       })
Returns (Tensor):

loss based on GroupDRO algorithm