README.md

1. To generate the CIFAR-10-C data, please refer to the official github code https://github.com/bethgelab/imagecorruptions
2. You should download the original CIFAR-10 dataset for training the SSL model. The standard model will be downloaded at the first time you run the training command. 

3. To run our code for training, please run following command:
>> CUDA_VISIBLE_DEVICES=0 python3 ssl_reversed_cifar10c.py --batch-size 32 --test_batch 16
--data-dir ../cifar-data --md_path ./ckpt/cifar10_standard.pth --output_dir output/train_cifar10c_ssl --epoch 100 
4. The standard model will store with the path ./ckpt/cifar10_standard.pth
4. After training the SSL backbone, you can find the checkpoint in the directory [output/train_cifar10c_ssl]
5. To evaluate CVP on CIFAR-10-C corruption, please refer to our bash script run_ours_script.sh
6. This is just the source code for CIFAR-C. The official code will be published later.
