albert0147 / bait_sfuda Goto Github PK

I have noticed that most of the UDA experiments on Office 31 or visda provide a train_list.txt and val_list.txt during the dataloader creation stage.
Doesn't providing the list for both source and target domain defeat the purpose of unsupervised domain adaptation where target domain has no information about the target labels.

For example the val_list.txt included in the data folder includes labels for all the classes in target domain,and then this file is used by the dataloader to create the target dataloaders.,thereby allowing the model to have acccess to both source and target labels.

Of course,if there is something which I am missing here please correct me

bootleneck -> bottleneck