Test1

[zzx-zhangxian]

when taking multi gpus, the last mini batch may return an error just like: #14
so drop the last imcomplete minibatch.

FFB*(s/4) -> FB(s/4) as the channels of feature in all feature outside the cell, in cell is B*(s/4), it now take about 9G-10G gpu memory when take 320*320 input

when using multi gpus, the architecture.step will only use gpu 0, because it is model.module(input) not model(input) or maybe the definition of architecture is before the dataparallel