Question about cross_training.

In the forward function of CrosSCLR, the key parameters in joint, bone, motion view are all momentum updated. However, in cross_training function, only key parameters in motion view are momentum updated.
Can you explain the reason of this?
Thanks.