optimizer state_dict not stable when resuming from checkpoint

[tenggyut]

Hi,

line 466 in hardnet.py iterates self.heads which is a dict. but the iterating order may differ in differ python command which will result in different parameter order in model.

when loading optimizer state dict from checkpoint, different parameter order would confusing the tensor id mapping and cause error....

simple solution would be iterate the sorted version of self.heads.keys().