Feature Request: Implementation of SSDA like data loaders

[shreejalt]

Hi,
Is there any plan to introduce the semi-supervised domain adaptation data loaders in the code? There are UDA and SSL loaders, but for SSDA, we might need different target data loaders during the training, which, as far as I know, cannot be directly used from the codebase.

If the implementation of ssda loaders will help, I can make a PR as I have worked on it

[KaiyangZhou]

train_loader_x can take both (labeled) source and target data (this could be instead implemented in the dataset class)

while train_loader_u still takes unlabeled target data

so seems like a new loader for ssda isn't a must?

[shreejalt]

Hi @KaiyangZhou
Yes correct, but there are some algorithms that do forward pass differently on labeled source and labeled target.

But yes, it can be implemented in the dataset class. I did it.

But for the former use case, it might be helpful.

I thought that it can be useful if we implement different DA papers.