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
We use cookies for essential functionality and analytics. You can accept or reject analytics cookies.Cookie policy