About the R50-MS4 36epoch training result

[cpsxhao]

Hi, thank you for your great work!

I was trying reproducing the results of 4-scale ResNet50 DINO model on COCO dataset. I trained in both FP16 and FP32, and the two experiments give the same result of 50.5 AP, which is a little bit lower than the reported 50.9 AP.

My initial LR is 1e-4 and LR drops at 30th epoch. The training log can be downloaded at https://drive.google.com/drive/folders/1yn7qOz0ViOSigyyj1JG3Pc2SISBk6bg7?usp=sharing

Could you help figure out the reason of the 0.4 AP gap? Looking forward to hearing from you!

[HaoZhang534]

We picked a good epoch to drop lr. Our checkpoint is 48.9 before lr drop. We can try to drop from the epoch with the highest AP, which is 48.8 in your log.