Mask R-CNN: Cannot Reproduce Reported Performance (mAP Drop)

[chamely]

I am trying to reproduce the official performance of Mask R-CNN with Swin-T using the config provided in MMDetection.
However, the reproduced box mAP and mask mAP are significantly lower than the results reported in the model zoo.

tools/train.py configs/swin/mask-rcnn_swin-t-p4-w7_fpn_1x_coco.py

Mask R-CNN + Swin-T
Official bbox mAP 42.2
My bbox mAP 37.8