Our project for UCLA CS 260D in 2023fall is Protecting against Corruption and Poisoning with SAS in Contrastive Learning. This repo implements the protection against corruption with SAS, which is one part of our project.
Download CIFAR100 and place it in data/cifar100-original/.
Run data/cifar100/cifar-100-python/gauss.py to generate the corrupted CIFAR100 dataset and place it in data/cifar100/cifar-100-python/.
Run cifar100gau_subset_creation_new.py to generate the subset with different fractions. We have selected some subsets with different fractions and saved them in the sas_subset directory.
Run simclr_sassub.py or run_sassub.sh to do the contrastive learning on subsets selected by SAS. Run simclr_ransub.py or run_ransub.sh to do the contrastive learning on randomly selected subsets.