ZOC

This repository is an implementation of AAAI2022 paper "Zero-Shot Out-of-Distribution Detection Based on the Pre-trained Model CLIP"

Datasets

cifar10, cifar100 and cifarplus are available in torchvision datasets. Tinyimagenet can be downloaded by running

tinyimagenet.sh

Training

1. Training of the Decoder_text is done once. It is used later for evaluation of all datasets. To train, please run:

python train_decoder.py

Decoder checkpoint

The fine-tuned decoder weights can be downloaded from this Google Drive link

Evaluation

Every evaluation script loads the fine-tuned decoder from training step, please run one of the following for test results :

python cifar10_eval.py

python cifar100_eval.py

python cifarplus_eval.py

python tinyimagenet_eval.py