* The reference points in deformable detr are detached with [Link](https://github.com/IDEA-Research/DAB-DETR/blob/main/models/dab_deformable_detr/deformable_transformer.py#L425). I want to know the reason. * The topk proposals in deformable detr with two stage is detached [Link](https://github.com/IDEA-Research/DAB-DETR/blob/main/models/dab_deformable_detr/deformable_transformer.py#L186). I want to know the reason. * The weight dict does not include enc loss. If using two stage, the encoder is not trained directly?