A serious problem: the target domain label is used in the training phase.

[WUT-YYN]

From the code "def preprocess_labels(self, source_loader, target_loader):
trg_y = copy.deepcopy(target_loader.dataset.y_data)
src_y = source_loader.dataset.y_data
pri_c = np.setdiff1d(trg_y, src_y) # Get the private class in the target domain
mask = np.isin(trg_y, pri_c) # Mark the private sample in the target domain
trg_y[mask] = -1 # Mark the label of the private sample as -1
return trg_y, pri_c" we can see that

1. The true label trg_y of the target domain sample is directly extracted from target_loader.dataset.y_data, and the private class label set in the target domain is clearly obtained.
2. trg_y[mask] = -1, all "unknown classes" are marked as "-1", but there is more than one "unknown class" in the target domain, how to solve this problem?
3. The "unknown class" in the target domain is only detected by the bimodal, but how can the model recognize it? There is no relevant content for learning "unknown classes" in the code.