when i run your code about cyclegan svhn to mnist，i find the result is very pool，maybe the sample example is good，but，the test examples is pool. for example ,it can translate 8 to 5, 2 to 0,the correct rate of the generate mnist is about 25%,when the correct rate about train mnist classify is 99%.