cvlab-columbia / zsrobust4foundationmodel Goto Github PK

Sorry, I think the y the equation 4 and 5， the y =1 or y=0
but in your code, the y is the label.

https://github.com/cvlab-columbia/ZSRobust4FoundationModel/blob/1243db6600f213851c491c6290f1ec756db98fa4/finetuning.py#L952C1-L952C1

Hi! Thanks for sharing the code for your work.

I wanted to know if the pretrained weights of the fine tuning experiments will be released anytime soon?
Further, the from the documentation, it appears the finetuning script is only available for VPT, not for the best results achieved by finetuning of visual encoder.

Running the fine tuning script as mentioned in the readme file (with batch size of 1)
python finetuning.py --batch_size 1 --dataset ImageNet --name feimogu --train_eps 1 --train_numsteps 2 --train_stepsize 1

RuntimeError: CUDA out of memory. Tried to allocate 302.00 MiB (GPU 0; 15.75 GiB total capacity; 11.37 GiB already allocated; 216.19 MiB free; 11.78 GiB reserved in total by PyTorch) If reserved memory is >> allocated memory try setting max_split_size_mb to avoid fragmentation. See documentation for Memory Management and PYTORCH_CUDA_ALLOC_CONF

Hi! Thanks for sharing the code for your work.
Where can I get the file for this path?
/local/*/datasets/ImageNet-clean

I read your paper with interest! Currently, I have a plan to develop your research.
I have some question in your paper. Please answering below three questions

(1) In paper, I confirm that you evaluated 100 PGD attack on test data. Then, setting test_numsteps=100 when evaluating is same with this?

(2) In finetuning.py, I found attack_CW (not ```attach_pgd``). Can you explain what this attack means? and What is paper result correspoding this attack?

(3) Do you have full result l2-norm bound pgd attack? In your paper, l_infinity-norm bound is used.
I am curious about l2-norm result!