Minor improvement in AP during training

[WUyinwei-hah]

Hi, very great work!
I'm trying to reproduce your training results.
I trained with batchsize of 124 for 2W steps. But I found the AP on COCO is pretty low, as shown in figure below:

The AP is only 0.041, so I wonder is it normal that the metrics improve slowly during the training?
If not, could you provide me with any suggestions toward training?
Thanks!

[WUyinwei-hah]

I further checked the samples generated from function save_ckpt_and_result, and found it looks pretty good.
So is the above phenomenon caused by the poor generalization from LAION to COCO?

[sweetDream6609]

Hello, does the generated image instance in your training results match the position of the input control？