Concern on the details of the comparison results in Table-2 about wss HOT 4 CLOSED

Hi @PkuRainBow

1. Yes, your understanding is correct. We here use the same number of iterations for all the data splits, this is because we need to iterate through the unlabeled set enough times (if you count the number of epochs based on the unlabeled set, then they are the same).
2. Yes, those 92 images are also in the unlabeled set. I follow the common practice in SSL classification here.

BTW, we sample those 92 images so that the number of pixels for each class is roughly balanced. You might not always get a good result if you pick arbitrary 92 images (see Appendix C).

from wss.

@Yuliang-Zou Great thanks for your explanation. We still have a small concern about your experimental settings.

According to your explanation, in fact, your method will train over the 92 images (labeled set) for 20869 epochs, which might cause serious overfitting problems on the supervised loss training part. We also find that the authors of CutMix face the same challenge and we paste the discussion here: Britefury/cutmix-semisup-seg#5 (comment)

So we are really interested in how your experimental setting can address the overfitting problem? Hope for your explanation!

from wss.

I don't have a clear answer yet. But I guess it could be related to the training schedule. In the beginning, the supervised loss dominates the optimization; as we train for more and more iterations, the unsupervised loss starts to take effects and gradually dominates the loss. Just for your reference, FixMatch (semi-supervised classification) has an experiment, training cifar10 on 10 images only, but it works quite well.

from wss.

@Yuliang-Zou Thanks for your reply. The balance between the supervised loss and the unsupervised loss might be a good point to avoid this problem. If my understanding is correct, it is very important to ensure the unsupervised loss to dominate in the late stage. However, there seem no explicit mechanisms to ensure such a scheme, therefore, we guess that an explicit re-weighting scheme might address this problem.

from wss.