ziyangwang007/CV-SSL-MIS

Get confuse on EMA operation

Opened this issue · 0 comments

Thank you for sharing the model. I have encountered some issues with the code.

In the paper, you mentioned that two projectors are modified by the other two projectors using the EMA operation. From my observation in 'train_Contrastive_Consistency_ViT_2D.py', projectors 1 and 2 are updated by EMA using projectors 3 and 4. However, I noticed that the parameters of projectors 3 and 4 are not included in optimizer 1 and 2.

This confuses me because if projectors 3 and 4 cannot be backpropagated by the optimizer, how can they influence the updates of projectors 1 and 2?

Is there something I misunderstood?