reproduce 《Distilling the Knowledge in a Neural Network》
architecture : two hidden layer , 800 rectified linear unit and no regularization
performance ： accuracy is 0.9697 on test dataset
source code ： Student_NN.ipynb
architecture : ConvNets
performance ： accuracy is 0.9936 on test dataset
source code ： mnist_distilling.ipynb
weights file ： frozen_model.pb
architecture : the same to Student Network
performance ： accuracy is 0.9635 on test dataset
source code ： Student_Distilling.ipynb
weights file ： frozen_model.pb (teacher model)
super params : T = 1 （according to paper）， alpha = 0.1 is unkown
when net at a temperture of 20, the performace is very bad and can't converge !!!But I don't know why paper do this?
