The result cannot be reproduced

[weihao20]

Thanks for your open-sourcing!
When I run this code, the results I got are not as good as those given in your paper. This is the result on flickr30k using VSRN+AOQ. My environment is python2.7 and pytorch0.4.1. What should I do to get the results on the paper?

rsum: 479.4
Average i2t Recall: 85.8
Image to text: 70.8 91.2 95.4 1.0 3.6
Average t2i Recall: 74.0
Text to image: 53.9 80.5 87.6 1.0 9.2