how BNs in the teacher model are handled

[zhd2rng]

Hi, while the teacher model is frozen, how BNs in the teacher model are handled:

1. BNs use the data batch statistics? i.e., training mode but with no grad
2. BNs use the running statistics? i.e., eval mode
3. BNs in the backbone and head, are they treated the same as the BNs in the heads (which also process the student features).

thanks!