Efficient Adversarial Contrastive Learning via Robustness-aware Coreset Selection

This repository provides codes for NeurIPS 2023 (Spotlight) paper: Efficient Adversarial Contrastive Learning via Robustness-aware Coreset Selection (https://arxiv.org/pdf/2302.03857.pdf)
Xilie Xu* (NUS), Jingfeng Zhang* (RIKEN-AIP/The University of Auckland), Feng Liu (The University of Melbourne), Masashi Sugiyama (RIKEN-AIP/The University of Toyko), Mohan Kankanhalli (NUS).

Robustness-aware coreset selection (RCS) can significantly speed up both robust self-supervised pre-training and supervised pre-training while maintaining the transferability of the pre-trained models.

RCS outputs a coreset which is an informative training subset. The model can achieve the minimized representational divergence between natural data and their adversarial counterparts after training on the coreset selected by RCS. In this way, RCS helps maintain the adversarial robustness of representations while accelerating robust pre-training.

Script

In our repo, we provide the codes and scripts of RCS for speeding up robust pre-training as follows:

• Efficient adversarial contrastive learning (ACL) via on unlabeled datasets (CIFAR-10/CIFAR-100/STL-10/ImageNet-1K);
• Efficient standard adversarial training (SAT) via RCS on labelled datasets (ImageNet-1K).

RCS for Efficient ACL on Small-Scale datasets (CIFAR-10/CIFAR-100/STL-10)

cd ./ACL_RCS/small_scale_datasets
python DynACL_RCS.py exp_dir --ACL_DS --dataset dataset --fraction 0.2

RCS for Efficient ACL on Large-Scale ImageNet-1K of $32 \times 32$ Resolution

cd ./ACL_RCS/ImageNet_32
python ACL_RCS.py exp_dir --ACL_DS --fraction 0.05

As for preparing ImageNet-1K of $32 \times 32$ resolution, we use the following scripts:

wget https://image-net.org/data/downsample/Imagenet32_train.zip
wget https://image-net.org/data/downsample/Imagenet32_val.zip

RCS for Efficient SAT on Large-Scale ImageNet-1K of $32 \times 32$ Resolution

cd ./SAT_RCS/ImageNet_32
python SAT_RCS.py --out_dir exp_dir --fraction 0.2

RCS for Efficient SAT on Large-Scale ImageNet-1K of $224 \times 224$ Resolution

cd ./SAT_RCS/ImageNet_224
python SAT_RCS.py --out_dir exp_dir --fraction 0.2

As for preparing ImageNet-1K of $224\times 224$ resolution, we use the following scripts:

wget https://image-net.org/data/ILSVRC/2012/ILSVRC2012_img_train.tar
wget https://image-net.org/data/ILSVRC/2012/ILSVRC2012_img_val.tar
wget https://image-net.org/data/ILSVRC/2012/ILSVRC2012_devkit_t12.tar.gz

Pre-Trained Checkpoints

We provide the following pre-trained checkpoints:

• Efficient ACL via RCS on CIFAR-10/CIFAR-100/STL-10 using ResNet-18 in Model Zoo of RobustSSL Benchmark;
• Efficient ACL via RCS on ImageNet-1K of $32 \times 32$ resolution using WRN-28-10 in Dropbox;
• Efficient SAT via RCS on ImageNet-1K of $32 \times 32$ resolution using WRN-28-10 in Dropbox;
• Efficient SAT via RCS on ImageNet-1K of $224 \times 224$ resolution using ResNet-50 in Dropbox.

Package

• Python 3.8
• Pytorch 1.13
• CUDA 11.6

BibTeX

@inproceedings{
xu2023efficient,
title={Efficient Adversarial Contrastive Learning via Robustness-Aware Coreset Selection},
author={Xilie Xu and Jingfeng Zhang and Feng Liu and Masashi Sugiyama and Mohan Kankanhalli},
booktitle={Thirty-seventh Conference on Neural Information Processing Systems},
year={2023},
url={https://openreview.net/forum?id=fpzA8uRA95}
}

Contact

Please drop an e-mail to [email protected] and [email protected] if you have any issue.