Why is the result different from the article?

[zengkaipeng]

The result printed by the code on USPTO-50K, class unknown is
top 10 accuracy: 0.5145 0.6519 0.7220 0.7595 0.7831 0.8366

while that reported in the article is
top1: 54.1 top 3: 74.1 top 5: 81.2 top10: 86.7

The result in the article is much better than reported. Can you explain why?