Didn't use Karpathy's split

[jamiechoi1995]

The result of the paper was reported using Karpathy's split(http://cs.stanford.edu/people/karpathy/
deepimagesent/),

while this repo didn't do so,

just note that for your information.

[weicheng51]

Hello, the result of my training is far from the one published in the paper, is it because of the different split？