Want to get information of Standard model

RobustBench / robustbench

RobustBench: a standardized adversarial robustness benchmark [NeurIPS 2021 Benchmarks and Datasets Track]

https://robustbench.github.io

Other

671 stars 99 forks source link

Want to get information of Standard model #160

Closed maengjemo closed 10 months ago

maengjemo commented 1 year ago

Hi!

I wonder train configuration of the Standard model which got 94.79 Clean score at the paper.

Which augmentation, optimizer, learning-rate are used?

Thanks for your answer in advance and have a nice day :)

fra31 commented 1 year ago

Hi,

please find some information in https://github.com/RobustBench/robustbench/issues/128.

inesvalentim commented 11 months ago

Hi @fra31!

I'm trying to train the standard model from scratch following the configuration described in #128, but I'm only reaching ~90% clean accuracy. I was wondering if you could provide some additional details:

Did you use Nesterov momentum in the SGD optimizer?
For the cosine annealing scheduler, how did you define the maximum number of iterations (T_max, assuming the torch.optim.lr_scheduler.CosineAnnealingLR was used)? And did you keep the default value (i.e., zero) for eta_min?
From what I understood, you didn't use any sort of data augmentation (e.g., random cropping or flipping of the images) while training the model. Is that correct?

Thanks!

fra31 commented 11 months ago

Hi,

as mentioned in https://github.com/RobustBench/robustbench/issues/128, I didn't train that model myself, but I'd assume that basic augmentation (cropping and flipping) has been used since it's the standard for CIFAR-10, and this might already get a significant improvement in accuracy. I'd also think momentum is used in SGD, not sure about the scheduler parameters, but I don't think they were tuned to train that model.

Hope this helps!

fra31 commented 10 months ago

Closing this for now, feel free to reopen it if you have more questions.