Add DCN about fgd

[yzd-v]

It seems the grad explodes. Maybe tha gap between them is too largre. If the loss keeps too large, you can try to decrease the grad_norm to avoid explosion.

[ywaner]

It seems the grad explodes. Maybe tha gap between them is too largre. If the loss keeps too large, you can try to decrease the grad_norm to avoid explosion.

Thanks for your reply! But it still grad explodes when I try to make the teacher and student model the same, that is make the model distill itself. I think this case may not be due to the gap between the teacher and student. I also try to distill the model itself without adding DCN, it trains well.

[ywaner]

I also try on your further fantastic work of MGD, it also has the same situation when I add DCN.

[yzd-v]

Fine, I don't know the reason either. However, RepPoints can be trained with DCN in the congig

[ywaner]

OK, thanks a lot! I will do more experiments later.