about offset loss, difficult to convergence?

[jianzfb]

Hi, I am reproducing your work. I found that the offset loss is very difficult to convergence. The offset value is about 1000~2000. I think it's very abnormal. I don't know why this happen, could you help me?

Thank you very much.

[Gengzigang]

I have not met this problem. I do not know if you had modified the code and you can try to run the code without any modification.

[chenrxi]

what gpu nums and batch size should be set to reproduce the result in paper @Gengzigang

[Gengzigang]

I used 4GPUs for HRNetw32, and 8GPUs for HRNetw48. The total batch size is 40.