A Simplified Pytorch implementation of Using Self-Supervised Learning Can Improve Model Robustness and Uncertainty.
These results show that self-supervised auxiliary loss improves model robustness in terms of the Out-of-Distribution detection.
The code supports only Multi-class OOD Detection experiment(in-dist: CIFAR-10, out-of-dist: CIFAR-100/SVHN).
-
RotNet-OOD
python test.py --method=rot --ood_dataset=cifar100
-
baseline
python test.py --method=msp --ood_dataset=svhn