Do you have any regularization for learning with L_consistency to avoid pushing Z very small?

[seedlingfl]

L_con tends to make Z_i and Z_j very small, as the loss will decrease this way. But in your paper, the visualized Z_i and Z_j histogram is ranged well within [0, 1], so I wonder if there are any techniques you may not mention in the paper?