hritikbansal / generative-robustness Goto Github PK

Hi,

I am having some issues downloading the dataset from google drive - since the dataset is quite large / contains many files, the server doesn't seem to zip the files properly for download and so the entire dataset isn't downloaded. Would it be possible to .zip or .tar the entire folder and then upload to drive as one file? It would make it a lot easier to download.

Thanks!

Hey, thanks so much for releasing the ImageNet-G-1K datasets (both base-generated and finetuned-generated), this is very useful!
While going through the synthetic dataset, I noticed one anomaly which I thought might be worth flagging. There are two redundant sets of class-names "Sunglasses" and "Missile". These also occur in the original class-names used by the OpenAI folks: https://github.com/openai/CLIP/blob/main/notebooks/Prompt_Engineering_for_ImageNet.ipynb
However, I think the right class-names to synset mappings to use would be sunglasses: n04356056, sunglass: n04355933, missile: n03773504 and projectile: n04008634 (for the full synset-classname mapping, you could check this).
This might therefore lead to some redundancy (although debatable how much of a difference this would make for sunglass vs sunglasses) in both the train and val sets that are generated. It might be worth regenerating the data for the two classes, and see if this improves performance slightly (my hunch is it wouldn't matter too much).