Original Paper code

[TaWald]

Hello,

I would like to integrate VoCo into my pretraining framework and have a few questions.
Currently the repo seems to diverge from the method described in the paper as it does a teacher-student approach that is never mentioned in the paper.
Q1: Where can I find the original implementation of the code from the paper? Is there an old branch I can checkout that holds this?
Q2: How does the teacher-student setting affect overall performance? Is it needed to avoid feature collapse? Or why was it added?
Q3: You seem to do additional augmentations on the subcrops independently of each other. Do I need to flip/rotate these subcrops independently? Did you experiment how on how important this independent augmentation is?

Cheers,
Tassilo

[Luffy03]

Hi, many thanks for your attention to our work and sorry for my late reply.
A similar issue was raised in #15. Teacher-student is widely adopted in pre-training and is slightly better than our original implementation. It will be introduced in our new version of paper.
In our augmentation, random filp/rotate is performed independently on each subcrop. I have not yet evaluated its importance, since we consider it as a basic setting. In general, random augmentation is also important in pre-training.

[Luffy03]

Dear researchers, our work is now available at Large-Scale-Medical, if you are still interested in this topic. Thank you very much for your attention to our work, it does encourage me a lot!