zhengzangw/doprompt

Official implementation of PCS in essay "Prompt Vision Transformer for Domain Generalization"

Hello there! This is the Github page of Zangwei Zheng (Alex).

• 📖 Pursuing PhD in CS at the National University of Singapore.
• 🎓 Conducting research on Efficient Maching Learning, ML Optimization.
• 💡 Have a wide interest on applications of NLP, CV, Rec, RL, ...
• 👀 Open to collaborating on new projects.
• 💬 Visit my homepage for more information.

doprompt's Issues

About hyperparameters

Hi, thanks for sharing your great works.

I'm trying to reproduce the result of ERM(3rd row in Table.1 from your paper), but I got a bit lower performance on OfficeHome dataset. I ran your source code with hyperparameters described in your paper(lr=1e-05, weight_decay=1e-02, last dropout=0.1), and I got 72.767% of target accuracy averaged on 3 different random seeds. I think I'm missing some other hyperparameters such as classifier LR multiplier or attention dropout probability in vision transformer which are not clarified in the paper.

Could you kindly share the searched hyperparameters on each dataset?

Is the dimension of adapted_prompt same to that of the domain prompt?

adapted_prompt is a combination of domain prompts

Table3

Hello, your explanation in Table 3 states that the upper right corner is the distance between the domains, that is, the distance between the centroids of the domains. However, according to the data in the table, there is also a distance between the centroids of the same domain. For example, in the VIT-base table, the centroid distance between P and P is 23.5.

So I have a question: should the order of A P R C on the vertical axis be the same as the order of A C P R on the horizontal axis?

reproduce result

I try to use your recommend training command (3rd and 4th) to reproduce the result of ERM and DoPrompt in PACS, but the performance of DoPrompt is lower than ERM (i.e., in A out of PACS, ERM is 90 while DoPrompt is 89.2.). Can you check the training commend again and suggest me what I am wrong? Thanks!

Is the pretrained model frozen in the training phase?

Hi, it seems that all the parameters in the model are updated during training, including the pretrained model. However, I find it that they are usually kept frozen in the field of prompt learning (e.g. in Visual Prompt Tuning, the backbone is frozen, while only the prompt tokens are allowed to be trained). I want to know whether it's my misunderstanding or intentional design. Thanks.