The experimental reproduction has a large variance.

[xlHuang0719]

Hi, dear author. Recently I've been working on a MapTRv1 reproduction. I experimented on 4 GPUs and with a batch size of 4 per GPU. I found that the jitter range of the reproduced mAP sometimes exceeds one point, e.g. from 49.9 to 51.2. I guess this is not normal. Is there any solution for this?