gaopengcuhk / clip-adapter Goto Github PK

I think when using coop's style, all I need is to put the clip_adapter file into the trainer's directory. Doesn't using @TRAINER_REGISTRY.register() automatically detect the name?
So why would I get this error KeyError: 'Object name "CLIP_Adapter" does not exist in "TRAINER" registry'

I have tried both conda and pip methods, but they are both incorrect.

hello, you said in your paper that you use adapter in both visual and text stream, but in your code i just find the visual one, which one is correct? Thanks a lot.

Thanks for your great job!
When I run this code with COOP, I find the result of trained on oxford_flowers is great,
=> result
=> result

• total: 2,463
• correct: 2,332
• accuracy: 94.7%
• error: 5.3%
• macro_f1: 94.6%
  Elapsed: 0:15:09
  But if I run it again(as your code show, it will use the model I trained last time, which got good results), it gets a bad result as follows:
  => result
• total: 2,463
• correct: 24
• accuracy: 1.0%
• error: 99.0%
• macro_f1: 0.1%
  Elapsed: 0:00:11
  I am not sure why it produces a bad result， can you give me some advice.

Hi author,
Could you please release the pre-trained weights for visual and text adapter?
I'm trying to introduce the CLIP adapter to my task. Since I don't have enough sample pairs for further finetuning. I hope the pre-trained weights can help.

Thank you.

Thank you for your great work!
But I have a simple question, how do you find the best residual ratio of different datasets?
Is that only based on experiment?
You also mention that α and β can be set learnable, could you tell me how you learned these ratios?
Thank you very much!

Hi, could you please let me know if/when you are planning to release the code for CLIP-Adapter on Imagenet?

Can you release the code for tSNE visualization?

Hi! Thank you for release this interesting work. Just to check my understanding: during few-shot training, is the label made by pseudo labeling with the original CLIP feature and classification weights (before the adapter layers)? Does the few-shot training set contain ground truth lables?

Hi authors, I'm a researcher at UCSD and I'm currently conducting the same research as you guys. I want to compare my model with the results from your model. Seems like there is no appendix or tables specifically showing the numbers of few shot (1,2,4,8,16) accuracy scores for 11 datasets. Is there a better way I can have these numbers than just trying to estimate them from your graphs?

Hello authors,

Any chance that you would like to release the codes?