DGGD PACS final checkpoints
Code, environment, protocol and results: https://github.com/MrHeatcliff/DGGD
360 completed final checkpoints: 48 tuned ERM runs (4 backbones × 4 domains × 3 seeds), plus 312 runs from 26 other algorithms (ResNet50 AugMix, default hyperparameters). All are selected on source-domain validation, not target oracle. No incomplete or smoke runs, intermediate Optuna trial weights, training data or optimizer-resume guarantee.
See manifest.json for paths, SHA256, sizes, source validation, target accuracy and selected update. Domain order: art_painting, cartoon, photo, sketch. Seeds: 100, 101, 102.
Loading
Use the matching DGGD code and package environment; these are PyTorch state dictionaries, not Transformers models. The backbone initially requires cached pretrained weights and DINOv2 code where applicable. best.pt contains model_dict, hparams, and task metadata. Build domainbed.algorithms.ERM for ERM, or scripts.pacs_algorithms.compat.algorithm_class(name) for other algorithms, then load model_dict and call eval(). Fishr requires the compatibility class. Only load pickle checkpoints from a trusted source.
from huggingface_hub import hf_hub_download
import torch
path = hf_hub_download(repo_id="datTrantien17/DGGD-checkpoints", filename="PATH_FROM_MANIFEST")
checkpoint = torch.load(path, map_location="cpu", weights_only=False)
These checkpoints inherit the applicable terms of their original pretrained backbones and code; see DINOv2, torchvision and timm model documentation. This upload does not claim a new blanket license over third-party weights. Evaluation uses the custom protocol documented in the linked repository; ERM and other algorithms have different tuning budgets and numeric precision.