why the test accuracy are different?

[zxl11]

I use the generate_RML2016.10a.py to generate the data set, which is the same number as the data set published by the author. And I use the same network to train the two data set, but the test accuracy are different. The accuracy of the data set generated by the code is 4-5 percentage points lower. Why? Are the parameters in the code provided by the author different from those in the public dataset?