reparameterization trick implementation

[Yan107351111]

It's unclear to me how the authors of the paper went from the 256 features before the random sampling to the parameters of the embedded distribution.
I saw you used additional two layer to compute 128 features each for use as parameters of the random distribution.
Since what I would have figured from the paper is that the authors may have split the 256 features to two sets of 128 and used them as distribution parameters.
This may be inconsequential, but from my personal experience deep architectures that have over 8 non-skippable layers are very difficult to optimize. I generally try to limit my nets to about 5 non-skippable layers. I'm mentioning this because these two additional linear layers bump you up from 8 to 9 non-skippable layers after the embedded space features of the encoder.
Please, let me know if you think I'm mistaken.