lijiefan/advcl

[NeurIPS 2021] “When does Contrastive Learning Preserve Adversarial Robustness from Pretraining to Finetuning?”

advcl's Issues

About pretraining model

Could you please share us your pretraining model?
Actually, I am very interested about your work, but I met some difficulties.
So, I sincerely hope that the pretraining model can be shared. Thanks.

No support for CIFAR100 dataset

Hi Lijie, current code only supports CIFAR10, can you kindly share the code of the ClusterFit component or the pseudo-labels of CIFAR100? Thanks.

Details about the clustering process

Could you shed some light on the clustering process? In particular,

• Which latent represents did you use to train KMeans?
• If the encoder is being trained using AdvCL then which encoder was used to compute the latent representations that were used during KMeans training?
• What did you use to build the clustering index?

Doubt in the SupCon loss function

Hi there. In the end of SupCon loss fn, you multiply the final loss by temp / base_temp value. What's the purpose for that? Thanks

Training command

It seems that there is no argument --syncBN and --decay in pretraining_advCL.py.
syncBN is set by default in code, and maybe decay is --weight_decay.

Unable to reproduce the results

Dear Authors,

Thanks for your interesting paper and the code! I encountered the issue when I reproduced your results.

I conducted AdvCL on CIFAR-10 by using your pre-training code and your finetuning code to conduct SLF on CIFAR-10. I got the following results copied from the generated log of your finetuning code:

best accuracy: 10.43
best accuracy clean: 13.14

In addition, I repeated the experiments including the pre-training and the finetuning procedure again. The following results are copied from the generated log:

best accuracy: 10.92
best accuracy clean: 14.12

Compared with your reported results on CIFAR-10 (standard test accuracy is 80.85% and PGD-20 robust test accuracy is 50.45%), my reproduced results are significantly lower than what you reported.

Could you help me figure out the reasons and guide me to reproduce your results? Thank you!