Accuracy differences between Niuzhiheng-Caffe and BLVC-Caffe?

[mgh1]

Has anyone done a comparison test that shows the Niuzhiheng-Caffe and BLVC-Caffe versions produce the same accuracy on a comprehensive task, say the ImageNet classification task?

It would be great to have some evidence that the porting to Windows did not introduce any accuracy issues. (I am asking this question because I did experience Windows ports of other neural network projects having accuracy differences.)

[niuzhiheng]

Can you please tell how much difference between them?

1. If the difference is very small, that may be caused by the stochastic process, e.g. random initialization of the CNN parameters. Even If you do it twice using the same code, the result would be different.
2. If the difference is very significant, it may caused by the code running differently cross platform caused by the third-party library or the code itself. There will be also a chance that a bug introduce during the code porting procedure.