Compared with the original version, is there any change in faster-rnn?

[HuDi2018]

I use the original faster-rcnn repo:https://github.com/jwyang/faster-rcnn.pytorch, and train res50 backbone on coco, get around 32mAP(fix input size to 800x800), and try detectron2 get 34 mAP(fix input size to 800x800).
So I wonder is any difference in training details of faster-rcnn