Sheng Su 2015-10-12

four GPU training: --

• having tried to change learning rate, mini-batch size and the gap, still diverge.
• having tried to use asynchronous way to update, still diverge.
• keep going to find the reason of divergency, and going to use some other methods to try.