Why is H-deformable-detr's mAP(49.1) higher than official implementation(48.7)?

[PhoenixZ810]

In official implementation, R50-based 12epoch model's mAP is 48.7, but model in detrex achieves mAP 49.1, are there any changes or improvements in detrex?

[rentainhe]

We evaluate with the top-300 predictions in detrex, which may brain 0.2 AP gain in the final results. Another possible reason might be the jitter introduced by the random seed.