tim-learn / ba3us Goto Github PK

Great work!
I'm looking at the experiments in the paper and I'm a bit confused. As far as I understand, your source consists of examples from the source distribution+(fewer) examples from the target distribution. If that's the case, does the ResNet-50 baseline also includes examples from the target distribution or only examples from the source distribution?

1. If choose the samples which belong to the missing class in target domain, what will happen?
2. Since the samples of the source domain are randomly selected, will it cause the problem of inconsistency of the categories?

Hi @tim-learn

Can you please share the implementation details you used to obtain the results with the VGG-16 backbone. Like, which layers did you train, and which layers did you freeze? And did you use the BatchNorm variant of VGG or the normal one?

Thanks

Hi Tim Team, I'm a bit confused about the results of the vanilla domain adaptation (Table 4). How could I get the results for the w/BAA and BA3US model? And could you also explain a little on your implementation of BA3US + CDAN model (last line). Thanks!