map question

[lijain]

Thank you for sharing the code. I want to ask you how to set the epoch50 for your training? I verified that 12 and 24 are OK, but the results of 50 rounds of Deformable Detr are almost the same as those of 24 rounds. Is it not good to add more rounds later

[jozhang97]

Thanks for verifying the 12/24ep schedules are OK. For 50eps, we follow DeformDETR and drop learning rate at 40ep.

We also noticed that the model seems to converge at around 24ep and training for longer does not help.
The nice work in DINO seems to have similar findings. I copy their table below from here.