Some problems about this work

[hxi667]

Sorry, I know this work has been finished for a long time, but I still want to ask you some questions.

1. I used tensorboard to record some information during training and found that the "betweenLoss" (similarity loss for comparing teacher and student model outputs) is almost oscillating without convergence. I want to know why.
2. Why take the outputs of the five middle layers of the model as inputs to the discriminator， parameter "out_dims=[5000,1000,500,200,10]", why 5000,1000,500,200,10?

These questions are really important to me and I would appreciate it if you could reply!