Mismatch of the range between generator's outputs and teacher's inputs

[sharinka0715]

Thanks for sharing your work! I have noticed that the generator's output is in range [-1, 1] (since a tanh function follows the last layer), but it does not match the teacher's input range. In dataloader.py, every different dataset is normalized, e.g. CIFAR10:

transforms.Normalize((0.4914, 0.4822, 0.4465), (0.2023, 0.1994, 0.2010)),

which indicates that the input range of teacher is in around [-2.5, 2.5]. The range of the generator's output is a subset of the teacher's input.

More importantly, I found that fixing this mismatch would affect the final accuracy of student. Could you please tell the reason of this mismatch or handle this problem? Thanks!