--- license: mit --- This repository contains the robust ImageNet models used in our paper ["Do adversarially robust imagenet models transfer better?"](https://arxiv.org/abs/2007.08489). See our papers's [GitHub repository](https://github.com/microsoft/robust-models-transfer) for more details! ### Standard Accuracy of L2-Robust ImageNet Models |Model|ε=0|ε=0.01|ε=0.03|ε=0.05|ε=0.1|ε=0.25|ε=0.5|ε=1.0|ε=3.0|ε=5.0| |---|:---:|:---:|:---:|:---:|:---:|:---:|:---:|:---:|:---:|:---:| |ResNet-18 |[69.79](https://huggingface.co/madrylab/robust-imagenet-models/resolve/main/resnet18_l2_eps0.ckpt) | [69.90](https://huggingface.co/madrylab/robust-imagenet-models/resolve/main/resnet18_l2_eps0.01.ckpt) | [69.24](https://huggingface.co/madrylab/robust-imagenet-models/resolve/main/resnet18_l2_eps0.03.ckpt) | [69.15](https://huggingface.co/madrylab/robust-imagenet-models/resolve/main/resnet18_l2_eps0.05.ckpt) | [68.77](https://huggingface.co/madrylab/robust-imagenet-models/resolve/main/resnet18_l2_eps0.1.ckpt) | [67.43](https://huggingface.co/madrylab/robust-imagenet-models/resolve/main/resnet18_l2_eps0.25.ckpt) | [65.49](https://huggingface.co/madrylab/robust-imagenet-models/resolve/main/resnet18_l2_eps0.5.ckpt) | [62.32](https://huggingface.co/madrylab/robust-imagenet-models/resolve/main/resnet18_l2_eps1.ckpt) | [53.12](https://huggingface.co/madrylab/robust-imagenet-models/resolve/main/resnet18_l2_eps3.ckpt) | [45.59](https://huggingface.co/madrylab/robust-imagenet-models/resolve/main/resnet18_l2_eps5.ckpt) ResNet-50|[75.80](https://huggingface.co/madrylab/robust-imagenet-models/resolve/main/resnet50_l2_eps0.ckpt) | [75.68](https://huggingface.co/madrylab/robust-imagenet-models/resolve/main/resnet50_l2_eps0.01.ckpt) | [75.76](https://huggingface.co/madrylab/robust-imagenet-models/resolve/main/resnet50_l2_eps0.03.ckpt) | [75.59](https://huggingface.co/madrylab/robust-imagenet-models/resolve/main/resnet50_l2_eps0.05.ckpt) | [74.78](https://huggingface.co/madrylab/robust-imagenet-models/resolve/main/resnet50_l2_eps0.1.ckpt) | [74.14](https://huggingface.co/madrylab/robust-imagenet-models/resolve/main/resnet50_l2_eps0.25.ckpt) | [73.16](https://huggingface.co/madrylab/robust-imagenet-models/resolve/main/resnet50_l2_eps0.5.ckpt) | [70.43](https://huggingface.co/madrylab/robust-imagenet-models/resolve/main/resnet50_l2_eps1.ckpt) | [62.83](https://huggingface.co/madrylab/robust-imagenet-models/resolve/main/resnet50_l2_eps3.ckpt) | [56.13](https://huggingface.co/madrylab/robust-imagenet-models/resolve/main/resnet50_l2_eps5.ckpt) Wide-ResNet-50-2|[76.97](https://huggingface.co/madrylab/robust-imagenet-models/resolve/main/wide_resnet50_2_l2_eps0.ckpt) | [77.25](https://huggingface.co/madrylab/robust-imagenet-models/resolve/main/wide_resnet50_2_l2_eps0.01.ckpt) | [77.26](https://huggingface.co/madrylab/robust-imagenet-models/resolve/main/wide_resnet50_2_l2_eps0.03.ckpt) | [77.17](https://huggingface.co/madrylab/robust-imagenet-models/resolve/main/wide_resnet50_2_l2_eps0.05.ckpt) | [76.74](https://huggingface.co/madrylab/robust-imagenet-models/resolve/main/wide_resnet50_2_l2_eps0.1.ckpt) | [76.21](https://huggingface.co/madrylab/robust-imagenet-models/resolve/main/wide_resnet50_2_l2_eps0.25.ckpt) | [75.11](https://huggingface.co/madrylab/robust-imagenet-models/resolve/main/wide_resnet50_2_l2_eps0.5.ckpt) | [73.41](https://huggingface.co/madrylab/robust-imagenet-models/resolve/main/wide_resnet50_2_l2_eps1.ckpt) | [66.90](https://huggingface.co/madrylab/robust-imagenet-models/resolve/main/wide_resnet50_2_l2_eps3.ckpt) | [60.94](https://huggingface.co/madrylab/robust-imagenet-models/resolve/main/wide_resnet50_2_l2_eps5.ckpt) Wide-ResNet-50-4|[77.91](https://huggingface.co/madrylab/robust-imagenet-models/resolve/main/wide_resnet50_4_l2_eps0.ckpt) |[78.02](https://huggingface.co/madrylab/robust-imagenet-models/resolve/main/wide_resnet50_4_l2_eps0.01.ckpt)|[77.87](https://huggingface.co/madrylab/robust-imagenet-models/resolve/main/wide_resnet50_4_l2_eps0.03.ckpt)|[77.77](https://huggingface.co/madrylab/robust-imagenet-models/resolve/main/wide_resnet50_4_l2_eps0.05.ckpt)|[77.64](https://huggingface.co/madrylab/robust-imagenet-models/resolve/main/wide_resnet50_4_l2_eps0.1.ckpt)|[77.10](https://huggingface.co/madrylab/robust-imagenet-models/resolve/main/wide_resnet50_4_l2_eps0.25.ckpt)|[76.52](https://huggingface.co/madrylab/robust-imagenet-models/resolve/main/wide_resnet50_4_l2_eps0.5.ckpt)| [75.51](https://huggingface.co/madrylab/robust-imagenet-models/resolve/main/wide_resnet50_4_l2_eps1.ckpt) | [69.67](https://huggingface.co/madrylab/robust-imagenet-models/resolve/main/wide_resnet50_4_l2_eps3.ckpt)|[65.20](https://huggingface.co/madrylab/robust-imagenet-models/resolve/main/wide_resnet50_4_l2_eps5.ckpt) |Model | ε=0|ε=3| |:-----:|:-----:|:-----:| DenseNet |[77.37](https://huggingface.co/madrylab/robust-imagenet-models/resolve/main/densenet_l2_eps0.ckpt) | [66.98](https://huggingface.co/madrylab/robust-imagenet-models/resolve/main/densenet_l2_eps3.ckpt) MNASNET|[60.97](https://huggingface.co/madrylab/robust-imagenet-models/resolve/main/mnasnet_l2_eps0.ckpt) | [41.83](https://huggingface.co/madrylab/robust-imagenet-models/resolve/main/mnasnet_l2_eps3.ckpt) MobileNet-v2|[65.26](https://huggingface.co/madrylab/robust-imagenet-models/resolve/main/mobilenet_l2_eps0.ckpt) | [50.40](https://huggingface.co/madrylab/robust-imagenet-models/resolve/main/mobilenet_l2_eps3.ckpt) ResNeXt50_32x4d|[77.38](https://huggingface.co/madrylab/robust-imagenet-models/resolve/main/resnext50_32x4d_l2_eps0.ckpt) | [66.25](https://huggingface.co/madrylab/robust-imagenet-models/resolve/main/resnext50_32x4d_l2_eps3.ckpt) ShuffleNet|[64.25](https://huggingface.co/madrylab/robust-imagenet-models/resolve/main/shufflenet_l2_eps0.ckpt) | [43.32](https://huggingface.co/madrylab/robust-imagenet-models/resolve/main/shufflenet_l2_eps3.ckpt) VGG16_bn|[73.66](https://huggingface.co/madrylab/robust-imagenet-models/resolve/main/vgg16_bn_l2_eps0.ckpt) | [57.19](https://huggingface.co/madrylab/robust-imagenet-models/resolve/main/vgg16_bn_l2_eps3.ckpt) ### Standard Accuracy of Linf-Robust ImageNet Models |Model|ε=0.5/255|ε=1/255|ε=2/255|ε=4/255|ε=8/255| |---|:---:|:---:|:---:|:---:|:---:| |ResNet-18|[66.13](https://huggingface.co/madrylab/robust-imagenet-models/resolve/main/resnet18_linf_eps0.5.ckpt) | [63.46](https://huggingface.co/madrylab/robust-imagenet-models/resolve/main/resnet18_linf_eps1.0.ckpt) | [59.63](https://huggingface.co/madrylab/robust-imagenet-models/resolve/main/resnet18_linf_eps2.0.ckpt) | [52.49](https://huggingface.co/madrylab/robust-imagenet-models/resolve/main/resnet18_linf_eps4.0.ckpt) | [42.11](https://huggingface.co/madrylab/robust-imagenet-models/resolve/main/resnet18_linf_eps8.0.ckpt) ResNet-50 |[73.73](https://huggingface.co/madrylab/robust-imagenet-models/resolve/main/resnet50_linf_eps0.5.ckpt) | [72.05](https://huggingface.co/madrylab/robust-imagenet-models/resolve/main/resnet50_linf_eps1.0.ckpt) | [69.10](https://huggingface.co/madrylab/robust-imagenet-models/resolve/main/resnet50_linf_eps2.0.ckpt) | [63.86](https://huggingface.co/madrylab/robust-imagenet-models/resolve/main/resnet50_linf_eps4.0.ckpt) | [54.53](https://huggingface.co/madrylab/robust-imagenet-models/resolve/main/resnet50_linf_eps8.0.ckpt) Wide-ResNet-50-2 |[75.82](https://huggingface.co/madrylab/robust-imagenet-models/resolve/main/wide_resnet50_2_linf_eps0.5.ckpt) | [74.65](https://huggingface.co/madrylab/robust-imagenet-models/resolve/main/wide_resnet50_2_linf_eps1.0.ckpt) | [72.35](https://huggingface.co/madrylab/robust-imagenet-models/resolve/main/wide_resnet50_2_linf_eps2.0.ckpt) | [68.41](https://huggingface.co/madrylab/robust-imagenet-models/resolve/main/wide_resnet50_2_linf_eps4.0.ckpt) | [60.82](https://huggingface.co/madrylab/robust-imagenet-models/resolve/main/wide_resnet50_2_linf_eps8.0.ckpt)