peal.adaptors.group_distributionally_robust_optimization¶
Group Distributionally Robust Optimization (GroupDRO) as a PEAL adaptor.
Baseline repair method (Sagawa et al., 2020) that minimises the worst-group
loss over the (target, confounder) groups instead of the average loss. The
adaptor either trains on a PEAL dataset with enable_groups() (labels come
as (y, confounder) pairs) or on the paper’s CelebA/Waterbirds replication
data from peal.dependencies.group_dro. Training uses the vendored
LossComputer and writes per-epoch checkpoints plus the best validation
model to <model_path>/checkpoints.
Classes
Adaptor that (re)trains a predictor with the GroupDRO objective. |
|
Config template for running the DRO adaptor. |