Request for Pre-trained AdvIVLP Weights

[msds24001]

Hi @xinwong,

Thank you for the excellent work on TAPT! I'm trying to replicate the results from your paper.

I noticed that the TAPT evaluation scripts require pre-trained weights from Multimodal-Adversarial-Prompt-Tuning (specifically the AdvIVLP weights trained on ImageNet with 16-shot):

WEIGHTSPATH='/path/to/Multimodal-Adversarial-Prompt-Tuning/output/train/imagenet/AdvIVLP/vit_b16_c2_ep100_batch32_2+2ctx_9depth_16shots'
The statistics/ folder with mean/variance files is already included, but the trained prompt weights (VLPromptLearner/model.pth.tar-100) are not available in the repository.

Could you please share the pre-trained AdvIVLP (and optionally AdvMaPLe) weights? This would greatly help researchers who want to replicate your results without having to train on ImageNet from scratch.

Thank you for your time!

[xinwong]

Hi, thank you for your interest in TAPT and for the detailed description of the issue!

I appreciate you pointing this out. I plan to release all the pre-trained weights (including AdvIVLP, AdvMaPLe, and AdvVPT) as well as the adversarial dataset subsets for quick reproduction in approximately 1 months.

I'll update the repository and post an announcement here once everything is ready.

Thanks again for your patience and support!

[IanKimROK]

Hi @xinwong, I’m also very interested in this project and am looking forward to the release of the pre-trained weights for my research as well. Any updates on the timeline would be greatly appreciated. Thank you!