Batch normalization shared between real and fake examples?

Hi,
It seems that in model.py when discriminator() is called twice, variables of the batch norm layers are also shared between the real and fake example batches. Is this design intentional? But in https://github.com/soumith/ganhacks it suggests that batch norm be separate for the two kinds of examples. Which design is more reasonable?
Thank you!!

Start by reading model.py and tracing the two discriminator() calls for the real and fake example batches, focusing on how TensorFlow batch-normalization variables are scoped and reused. Compare that behavior with the separate-batch guidance in the linked ganhacks resource. Done requires a decided, documented design and corresponding implementation change, if needed.

Written by the indexing model from the issue text.