Constant variance for the generating network of autoencoder

[dugarsumit]

Why are we using constant variance for the generating network of autoencoder instead of learning it like mean from the network itself. What advantage does it have over the learnable variance? This is done in the models.py file at line numbers 473 and 681.

mean_x = T.clip(output+.5, 0+1/512., 1-1/512.)
logsd_x = 0*mean_x + w['logsd_x']