CGCL-codes / AdvEncoder

The implementation of our ICCV 2023 paper "Downstream-agnostic Adversarial Examples"
https://arxiv.org/pdf/2307.12280.pdf
MIT License
68 stars 4 forks source link

where is the data loader and victim loader of ImageNet? #1

Open ChnanChan opened 10 months ago

ChnanChan commented 10 months ago

In the supplement of your paper, you randomly select 100 classes from ImageNet to build the dataset, but there is no code about ImageNet in utils/load_data.py and utils/load_victim.py Can you let me know? Thank you

RozerFun commented 8 months ago

In the supplement of your paper, you randomly select 100 classes from ImageNet to build the dataset, but there is no code about ImageNet in utils/load_data.py and utils/load_victim.py Can you let me know? Thank you

I found the construction code for imagenet-100 from the sole-learn working code repository, which can be referred to at https://github.com/vturrisi/solo-learn/blob/b69b4bd27472593919956d9ac58902a301537a4d/scripts/utils/make_imagenet100.py However, I still encountered the problem of low training accuracy of downstream models (between 30% and 40% on training accuracy) when reproducing it.

linghuchong111da commented 6 months ago

请问这2个代码有了吗,utils/load_data.py 和 utils/load_victim.py