chs20 / robustvlm Goto Github PK

Hi,
Thanks for the great work!
The paper mentions that "evaluate robustness on all datasets at resolution 224x224, except for CIFAR10, CIFAR100 and STL-10", however I find that in the code : https://github.com/chs20/RobustVLM/blob/main/CLIP_eval/clip_robustbench.py#L114, a resize also occurs for CIFAR10. I was wondering what would be the right implementation to consider?

Thanks,
Agneet

Hi! Thank you for your work and congratulations on ICML acceptance!
I am not sure if the issues is appropriate place but I wanted to share some of my findings regarding image generation with guidance from the gradients of Robust CLIP. Here I ran some experiments, and the results seem to be decent, which indicates that the gradients are perceptually-aligned. I thought you might find this interesting.

Hello, while going through your repository, I noticed in your adversarial_training_clip.py you've only utilized the experiment_name argument for file saving purposes. However, when I attempted to run the file with other experiment_name (such as TECOA4), it ran the same functions as the (FLARE4). Does this repository include all the required modules and functions to execute the unsupervised adversarial training method (i.e., FLARE4)?

Hi, currently, the code throws a NotImplementedError for LLaVA, but I believe the paper demonstrates zero-shot classification on LLaVA. When will the code be updated to include this feature? Alternatively, could you point out the main parts that would need significant changes to incorporate LLaVA?

Thank you.

I have to say it's an amazing job。

I noticed a parameter when executing llava_eval.sh
--coco_karpathy_json_path /mnt/datasets/coco/annotations/karpathy_coco.json

I want to know karpathy_coco.json is how to get, Whether the same as in https://github.com/mehdidc/retrieval_annotations/releases