The noise sampling and the Lz loss curve issue

[johnzlj]

Ln-44:
e=tf.random_normal((batch_size,z_size), mean=0, stddev=1) # Qsampler noise
I think it should be placed into the funcion sampleQ. Or else the inference will fail.

However, when I made such a modification, the Lz loss-curve will increase instead of decline, as shown here:

Why?