search

FightingFighting / GPS

This is the repository for paper: Gradient-based Parameter Selection for Efficient Fine-Tuning
14 stars 0 forks source link

We hope to reproduce the performance of GPS on VTAB #4

Open dafeizhong opened 1 month ago

dafeizhong commented 1 month ago

Can you provide a fine-tuning script for VTAB?

FightingFighting commented 1 month ago

we will release the code shortly.

FightingFighting commented 1 month ago

the scripts already released

xjiangmed commented 1 month ago

@FightingFighting I have run the provided training scripts for the VTAB-1K benchmark, and the performance on three datasets (Cifar100, flower102, and pets) is still far from the results reported in the paper even after tweaking the hyperparameters. Notably, on the flower102 and pets datasets, the SSF method similarly yielded results of 87.70 and 82.36. Could you give me some suggestions to reproduce the results of GPS? Thank you in advance for your guidance. image

FightingFighting commented 1 month ago

Hi @xjiangmed , I just update all scripts of VTAB for the hyperparameters. As for cifar100, in order to compare with VPT, we run GPS on their code. For the flower, is there something wrong with your code? I just rerun flowers (using the old scripts) and I got the Acc results: 99.5.

FightingFighting commented 1 month ago

and you can also try to change some hyperparameters for searching, such as lr, bz and times_para. Hope this could help.

xjiangmed commented 1 month ago

@FightingFighting Thank you for your help. I really appreciate it!