How do you compare outputs from different augmentations for consistency training?

I’m trying to figure out how papers using consistency train on unsupervised data, and I’m stuck on how outputs from different augmentations are compared. Since the augmentations are transformations, I would figure that another transformation would have to be done to the outputs in order to compare them, but I can’t find (or recognize) any mention of this.

What method is used to compare the predictions from different augmentations of the same image?

One example of this would be UDA

Xie, Qizhe, et al. "Unsupervised data augmentation for consistency training." arXiv preprint arXiv:1904.12848 (2019).

Another would be FixMatch

Sohn, Kihyuk, et al. "Fixmatch: Simplifying semi-supervised learning with consistency and confidence." arXiv preprint arXiv:2001.07685 (2020).