improved mAP on my run

[spokV]

I've trained the AOPG using aopg_r50_fpn_1x_dior config on DIOR-R and received 0.66 mAP on the test set.
The only change I've done is the use of CUDA 11 and the migration of the cuda ext (mmdet/ops/) from the THC to ATen (in order it to run on cuda 11.
This is an increase from the 64.41 mAP published on your paper.
I've trained on trainval set and tested on test set.
Does this improvement in mAP make sense?