Yang Nie

## reproduction all results trained on 1 gpu (V100) with total batch size 16 r18vd with amp ```python Average Precision (AP) @[ IoU=0.50:0.95 | area= all | maxDets=100 ] =...

> @flytocc Thank you very much. I would like to confirm why a previous pull request (PR) could not align the precision. Was something incorrect there? There are many differences...

- The training (w. amp) AP of `r50vd` arch fluctuates between `52.9` and `53.1`. - Random interpolations has almost no effect on AP.

> > * The training AP of `r50vd` arch fluctuates between `52.9` and `53.1`. > > * Random interpolations has almost no effect on AP. > > 您好，我试用了下您写的rt-detr，发现在我的数据集中存在不收敛的现象，就是训练到20个epoch左右，map突然变成0，但我在rt-detr官方代码中并没有这个问题，都是用的4卡4batch 目前只测试过COCO数据集。你试着可以检查一下数据增强

@ychensu 要不你到 [flytocc/mmdetection](https://github.com/flytocc/mmdetection) 提一个issue

try https://github.com/pytorch/pytorch/issues/121222

fix: ``` active_ids = [ id for id in self.ids if id not in ids and self.tracks[id].frame_ids[-1] == frame_id - 1 ] if self.with_reid else self.ids ```

> 你好，我使用`rtdetv2_r18vd_120e_coco.yaml`训练，最终结果为47.6，与论文中的48.1存在一定差距，请问有可能是什么原因。 我把rtdetrv2移植到mmdet上试了一下[repo](https://github.com/flytocc/mmdetection/tree/rtdetr-v2)，精度47.5 (torch 2.2.1)