leaplabthu / daprompt Goto Github PK

Thanks for your excellent work, but I'm curious about the Clip + FT experimental settings you posted. What data and which part of the parameters do you use for fine-tune ?

Hi John,

Fantastic work you have done!

I see you add an additional dataset experiment on this repo, that's the DomainNet which is not included in the paper.

May I have the Experimental Settings for the DomainNet? For example the training epoch, batch size, pseudo-labeling threshold and etc.

I want to run this code in VLCS dataset and i register this dataset as required, but it keeps showing that 'Object name "VLCS" does not exist in "DATASET" registry', what is the problem?

Does that also support text data? thanks

Thanks for this good work! Just a little wondering whether it also supports Transformer as backbone?

Thanks for your very interesting research.

I have some questions about the code of L402 in trainers/dapl.py

The original paper states that contrastive learning helps transfer learning.
But, to my understanding, in the above code, this operation does not perform contrastive learning over multi-domain and it seems to simply perform contrastive learning over single-domain.

To perform contrastive learning over multi-domain, I think I need to change the code in the following, right?
loss_x = F.cross_entropy(output_x[:, :-self.n_cls], label)