梁俊杰

Hello my friend, I've encountered the same issue as well. Have you managed to resolve it?

Thank you author, I have solved this problem, but I am now running your code and not achieving the results corresponding to your paper. The refcocog only reaches 68.6% in...

What I mean is that during my reproduction process, I fixed the random seed, yet the results were different each time. Moreover, after removing the token merging code, the issue...

The random seed has already been fixed at the beginning of the main function. Below is my random seed code, yet the issue still occurs. if seed is not None:...

Namespace(IS_not_position_VPT=False, aa='rand-m9-mstd0.5-inc1', amp=False, batch_size=32, cfg='experiments/LoRA/ViT-B_prompt_lora_12.yaml', change_qkv=False, clip_grad=None, color_jitter=0.4, cooldown_epochs=10, cutmix=0.0, cutmix_minmax=None, data_path='/data/dataset/liangjunjie/vtab-1k/caltech101', data_set='caltech101', decay_epochs=30, decay_rate=0.1, device='cuda', direct_resize=True, dist_eval=False, dist_url='env://', distributed=False, drop=0.0, drop_block=None, drop_path=0.1, drop_rate_LoRA=0.1, drop_rate_adapter=0.1, drop_rate_prompt=0.1, epochs=100, eval=False, few_shot_seed=3072, few_shot_shot=2,...

I have tried it, and experiments show that the performance instability comes from token merging.