Hi, thanks for your code.
I tested moon_tiny_iter_1000000.caffemodel on my computer(use test_moon_euclidean.ipynb) and got average error 12.69%. In README.txt, you noted that the accuracy is 89+%, and in paper, it says the average error is 9.06%. I'm not sure if the above caffemodel is the final weights.