Feature Clipping

Pretrained models

All logits and features and extracted from the following models:

• CIFAR10 (from focal loss calibration)
  • Resnet-50
  • Resnet-110
  • DenseNet-121
• CIFAR100 (from focal loss calibration)
  • Resnet-50
  • Resnet-110
  • DenseNet-121
• IMAGENET (from pytorch's torchvision.models)
  • Resnet-50: torchvision.models.resnet50(weights=torchvision.models.ResNet50_Weights.IMAGENET1K_V1)
  • DenseNet-121: torchvision.models.densenet121(weights=torchvision.models.DenseNet121_Weights.IMAGENET1K_V1)
  • Wide-Resnet-50: torchvision.models.wide_resnet50_2(weights=torchvision.models.Wide_ResNet50_2_Weights.IMAGENET1K_V1)
  • MobileNet-V2: torchvision.models.mobilenet_v2(weights=torchvision.models.MobileNet_V2_Weights.IMAGENET1K_V1)
  • ViT-L-16: torchvision.models.vit_l_16(weights=torchvision.models.ViT_L_16_Weights.IMAGENET1K_V1)

Dependencies

conda create -n feature-clipping python=3.10

python -m pip install -r requirements.txt

Evalutation

run bash evaluate_scripts_post_hoc.sh to evaluate post hoc methods

run bash evaluate_scripts_train-time.sh to evaluate train time methods