TTA-CIFAR-10-AugMix-WRN40

Mirror of the WideResNet-40-2 CIFAR-10 source model trained with AugMix (Hendrycks et al., ICLR 2020), as indexed by RobustBench under the Hendrycks2020AugMix_WRN entry (CIFAR-10 corruptions threat model). This is the CIFAR-10 AugMix source checkpoint used by common TTA baselines — CoTTA, EATA, SAR, and friends — for CIFAR-10-C evaluation.

TENT (Wang et al., ICLR 2021) Table 3 used a different (ResNet-26) AugMix variant released separately via Google Drive. Subsequent TTA work converged on this WRN-40-2 entry because it is the one pinned in RobustBench.

• Upstream: RobustBench model zoo (Hendrycks2020AugMix_WRN)
• License: Apache-2.0 (matches AugMix reference code)
• Arch: wrn_40_2_augmix (WideResNet-40-2; BasicBlock; internal (x - 0.5)/0.5 normalization)
• Params: 2,243,546
• Clean CIFAR-10 accuracy: 95.08%
• Input: 32×32×3 RGB in [0, 1] (normalization is baked into the model's mu/sigma buffers — do not pre-normalize with CIFAR mean/std).

Usage

from huggingface_hub import hf_hub_download
from safetensors.torch import load_file

path = hf_hub_download("WNJXYK/TTA-CIFAR-10-AugMix-WRN40", "model.safetensors", revision="v1.0")
state_dict = load_file(path)
# Plug into the matching WRN-40-2 arch (see `arch` above) and
# `load_state_dict(state_dict)`. The `mu`/`sigma` buffers at the head of the
# state_dict must be present in your arch.

Inside the TTA-Evaluation-Harness, the wiring is done via:

# configs/source_models/augmix_wrn40_cifar10.yaml
framework: torchvision_hf
arch:      wrn_40_2_augmix
hf_repo:   WNJXYK/TTA-CIFAR-10-AugMix-WRN40
revision:  v1.0

Citation

@inproceedings{hendrycks2020augmix,
  title={AugMix: A Simple Data Processing Method to Improve Robustness and Uncertainty},
  author={Hendrycks, Dan and Mu, Norman and Cubuk, Ekin D. and Zoph, Barret
           and Gilmer, Justin and Lakshminarayanan, Balaji},
  booktitle={ICLR}, year={2020}
}
@inproceedings{croce2021robustbench,
  title={RobustBench: a standardized adversarial robustness benchmark},
  author={Croce, Francesco and Andriushchenko, Maksym and Sehwag, Vikash
           and Debenedetti, Edoardo and Flammarion, Nicolas and Chiang, Mung
           and Mittal, Prateek and Hein, Matthias},
  booktitle={NeurIPS Datasets and Benchmarks Track}, year={2021}
}