Closed night3759 closed 2 months ago
The topology head is trained, but its update would not influence the detection when using "self.is_detach". From my experiment, it doesn't matter if you use "self.is_detach".
The topology head is trained, but its update would not influence the detection when using "self.is_detach". From my experiment, it doesn't matter if you use "self.is_detach".
does "self.is_detach" option prevent the gradients from bp to the upstream module?
yep.
yep.
Thanks .
If the "self.is_detach" is True, the MLP of topo_ll_head and topo_lt_head may not be trained. Do I understand right? Why detach?