Views
No views yet
Hendrycks2020AugMix_WRN entry (CIFAR-100 corruptions threat
model). This is the CIFAR-100 AugMix source checkpoint used by common
TTA baselines — CoTTA, EATA, SAR, and friends — for CIFAR-100-C evaluation.Hendrycks2020AugMix_WRN, cifar100, corruptions)(x - 0.5)/0.5 normalization)[0, 1] (normalization is baked into the model's
mu/sigma buffers — do not pre-normalize with CIFAR mean/std).1from huggingface_hub import hf_hub_download
2from safetensors.torch import load_file
3
4path = hf_hub_download("WNJXYK/TTA-CIFAR-100-AugMix-WRN40", "model.safetensors", revision="v1.0")
5state_dict = load_file(path)
6# Plug into the matching WRN-40-2 arch (see `arch` above) and
7# `load_state_dict(state_dict)`. The `mu`/`sigma` buffers at the head of the
8# state_dict must be present in your arch.1# configs/source_models/augmix_wrn40_cifar100.yaml
2framework: torchvision_hf
3arch: wrn_40_2_augmix
4hf_repo: WNJXYK/TTA-CIFAR-100-AugMix-WRN40
5revision: v1.01@inproceedings{hendrycks2020augmix,
2 title={AugMix: A Simple Data Processing Method to Improve Robustness and Uncertainty},
3 author={Hendrycks, Dan and Mu, Norman and Cubuk, Ekin D. and Zoph, Barret
4 and Gilmer, Justin and Lakshminarayanan, Balaji},
5 booktitle={ICLR}, year={2020}
6}
7@inproceedings{croce2021robustbench,
8 title={RobustBench: a standardized adversarial robustness benchmark},
9 author={Croce, Francesco and Andriushchenko, Maksym and Sehwag, Vikash
10 and Debenedetti, Edoardo and Flammarion, Nicolas and Chiang, Mung
11 and Mittal, Prateek and Hein, Matthias},
12 booktitle={NeurIPS Datasets and Benchmarks Track}, year={2021}
13}