I ran directly in the author code and only achieved 94.95 accuracy. I wonder how wrn-10's 3.89% accuracy is achieved