Shared weights transfer

[yfji]

Hi @starimeL Thanks to the excellent work! I have a question about the shared weights in PyTorch. For example, RetinaNet has shared classification weights and regression weights which process the 5 output branches of FPN. The shared weights are replicated 5 times after saved as a caffemodel, which causes the much bigger size of the model file. Is there a way to prohibit the replication?I also wonder if the shared weights are copied during forward in PyTorch, even if they are only saved one time in the model file.
Wish for your reply!