CIFAR-10 robustness study: seed-0 final models
The final-epoch models (seed 0) of the 6 configurations of the CIFAR-10 robustness study: a Simple CNN and a CIFAR ResNet-18, each trained with standard augmentation, AugMix, or corruption-aware training. They are used by the study's demo app.
| File | Model | Strategy | Clean test accuracy |
|---|---|---|---|
cnn-standard-s0.pt |
simple_cnn | standard | 84.04% |
cnn-augmix-s0.pt |
simple_cnn | augmix | 83.68% |
cnn-corruption_aware-s0.pt |
simple_cnn | corruption_aware | 82.02% |
resnet18-standard-s0.pt |
resnet18 | standard | 93.64% |
resnet18-augmix-s0.pt |
resnet18 | augmix | 93.72% |
resnet18-corruption_aware-s0.pt |
resnet18 | corruption_aware | 92.36% |
Each .pt file is a PyTorch state dict for src.models.build_model(model) of the study's code.
Inputs are 32x32 RGB images, converted with ToTensor and normalised with the training-set
mean and std in splits/cifar10_stats.json. Classes, in output order: airplane, automobile, bird, cat, deer, dog, frog, horse, ship, truck.
models.json lists every file with its sha256.
Trained on CIFAR-10 (Krizhevsky, 2009, Learning Multiple Layers of Features from Tiny Images) and evaluated on CIFAR-10-C (Hendrycks & Dietterich, 2019, ICLR; CC BY 4.0). The code is under the MIT licence; see the study's repository and report for the training setup and results.