Cannot reproduce the results of the paper

[kinredon]

Hi, thanks for your excellent work and the open-source.

I am trying to reproduce the results reported in the paper. Currently, I conduct the experiment of 1% COCO. I only obtained 18.5 AP, but the paper reported 19.64. Also, I use the large-scale jitter, setting the scale range to 400-1200. Still, I only obtained 20.7 AP (22.38 in the paper).

|   AP   |  AP50  |  AP75  |  APs  |  APm   |  APl   |
|:------:|:------:|:------:|:-----:|:------:|:------:|
| 20.679 | 35.090 | 21.424 | 9.386 | 22.590 | 28.191 |

I train the model on 8 3090 GPUs, and I pick the highest AP model of the teacher model.

Here is the full log when I use the large-scale jitter log.txt.

[ZRandomize]

The result on 1% data seems unstable, please check the performance not only on the last checkpoint. Besides, check whether you are testing the student model, we reported the performance of teacher model in the paper.

[Lyndon-wong]

Maybe the weight of unsupervised loss should be set as 4 ? (as the Table 7(b) in the paper said)

[PlutoQyl]

Have you tried centerness branch？