Code for training the ResNet18 w/o extra data? about score HOT 6 CLOSED

For Section 5.2 and Table 2, we follow the settings and code of Pang et al. 2021, and we substitute the KL divergence with SE.

from score.

Thank you so much for your prompt reply to my issue.

I implement the PGD-AT with SCORE by replacing the CE loss to LSE loss:

y_onehot = F.one_hot(y, num_classes=args.num_classes)
robust_output = model(adv_input)
logits_adv = F.softmax(robust_output, dim=1)
robust_loss = torch.sum((logits_adv - y_onehot) ** 2, dim=-1).mean()
opt.zero_grad()
robust_loss.backward()

I would also like to ask for your guidance on one more thing. Should the CE loss in PGD process be replaced as well? I noticed that in your code (trades.py/trades_loss_LSE), it is replaced.

Thank you again for your help. I appreciate your time and expertise.

from score.

Yes the CE loss in PGD should be replaced as well. This is because the objective of AT is a min-max problem, i.e., the minimization and maximization share the same objective function.

from score.

Yes the CE loss in PGD should be replaced as well. This is because the objective of AT is a min-max problem, i.e., the minimization and maximization share the same objective function.

I've tried for both,

the results are as follows:
Method Std Acc Robust Acc
PGD-AT 81.46 53.62
PGD-AT SCORE (Not replaced) 81.88 53.99
PGD-AT SCORE (replaced) 80.04 53.19

It seems that the CE loss in PGD should not be replaced?

I have been using BN mode as train() when using PGD. Could this be the reason why PGD-AT SCORE (replaced) is not performing as well as PGD-AT SCORE (not replaced)?

from score.

I tried to train PGD-AT & SCORE on CIFAR100, the results worse than vanilla PGD-AT.

Method Std Acc Robust Acc

PGD-AT 55.30 29.50
PGD-AT SCORE 29.28 22.32
PGD-AT SCORE w/o grad clip 36.46 25.20
PGD-AT SCORE (Not replaced) 42.96 27.01

It seems that the the CE loss in PGD should not be replaced as well?

Have you tried to run PGD-AT SCORE with CIFAR100? I wonder if I wrote the wrong code, but it seems not possible since for CIFAR10 it is right.

from score.

Hi, I find my scripts and upload them into https://github.com/P2333/SCORE/tree/main/without_DDPM, with running commands in README. You can check if there is any difference compared to your implementation.

from score.