Closed DrRyanHuang closed 1 year ago
Hello, we just follow the existing initialization as in Deformable DETR. I suppose these initializations would make little difference. With zero-init, the model will output exactly a bounding box around the reference point before training.
https://github.com/CASIA-IVA-Lab/Obj2Seq/blob/a81484ef61d692e8d2b3f88e58c6df8344a50131/models/predictors/unified_seq_head.py#L106-L114
Hi, Thanks for your helpful code! Could you please tell me why the parameters of MLP are set to 0? Why not Normal init or others?