Pretraining on ImageNet.

[tarun005]

I noticed that the backbone architecture is pretrained (pretrained=True).

visda-2019-public/semisupervised/model/resnet.py

Line 229 in e1d06e1

def resnet50(pretrained=True):

I am curious to know which dataset was the backbone resnet pretrained on? If it is Imagenet, there is a lot of overlap in the unlabeled target classes from visDA and some labels from Imagenet. Then wouldn't you violate the assumption that some target classes are unlabeled or few shot labeled, since the model would have already looked at some labeled examples in the pretraining phase?