TTA-ImageNet-ResNet50

Mirror of the standard ImageNet-1K ResNet-50 source model used for ImageNet-C test-time adaptation baselines.

This checkpoint is generated from torchvision ResNet50_Weights.IMAGENET1K_V1. It matches the model definition used by RobustBench's ImageNet-C Standard_R50 entry: torchvision.models.resnet50(pretrained=True) with ImageNet mean/std normalization applied outside the backbone.

Why This Model

• TENT's official example stack depends on RobustBench for datasets and pre-trained models.
• EATA reports ImageNet-C severity-5 results with ResNet-50.
• SAR evaluates ImageNet-C with ResNet-50 and ViT variants; the classic ResNet-50 backbone remains the baseline anchor.

For more recent "wild" or batch-size-1 settings, papers often add ResNet-50-GN and ViT-B/LN variants. This repo uses the BN ResNet-50 as the canonical ImageNet-C source checkpoint because it matches the original TENT/EATA-style setting and exposes BN affine parameters for TENT.

Model Details

• Upstream equivalent: RobustBench Standard_R50 for ImageNet corruptions
• Torchvision weights: ResNet50_Weights.IMAGENET1K_V1
• Arch: resnet50
• Params: 25,557,032
• Clean ImageNet val accuracy: not evaluated
• Input: RGB image, resized/cropped to 224, then normalized with mean [0.485, 0.456, 0.406] and std [0.229, 0.224, 0.225].

Usage

from huggingface_hub import hf_hub_download
from safetensors.torch import load_file
from torchvision.models import resnet50

path = hf_hub_download("WNJXYK/TTA-ImageNet-ResNet50", "model.safetensors", revision="v1.0")
model = resnet50(weights=None, num_classes=1000)
model.load_state_dict(load_file(path))

Inside TTA-Evaluation-Harness:

# configs/source_models/resnet50_imagenet.yaml
framework: torchvision_hf
arch:      resnet50
hf_repo:   WNJXYK/TTA-ImageNet-ResNet50
revision:  v1.0

Citations

@inproceedings{he2016deep,
  title={Deep Residual Learning for Image Recognition},
  author={He, Kaiming and Zhang, Xiangyu and Ren, Shaoqing and Sun, Jian},
  booktitle={CVPR}, year={2016}
}

@inproceedings{hendrycks2019benchmarking,
  title={Benchmarking Neural Network Robustness to Common Corruptions and Perturbations},
  author={Hendrycks, Dan and Dietterich, Thomas},
  booktitle={ICLR}, year={2019}
}

@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}
}