training with multiple GPUs

[Rody1911641]

I follow issues 9,but Manager().dict() is too slowly. I train with two GPUs.but it's as slow as one GPU

[jihanyang]

Sorry, as I haven't meet the problem of #9, I cannot provide some useful information for fix it in another way. However, in my opionion, such issue is more likely to be caused by the IO bottlecneck.