about multi-gpu

[DavieHR]

hi.
I found that the code in multi-gpu training may be wrong.
In trainer.py, when you refresh the network, in multi-gpu, you should re-start parralle-data again. otherwise, your renew model can not replicate to other gpu.