About the nasm implementation.

Is the implementation correct?

The computation of KL Divergence in your code is

```
# Kullback Leibler divergence
self.e_loss = -0.5 * tf.reduce_sum(1 + self.log_sigma_sq - tf.square(self.mu) - tf.exp(self.log_sigma_sq))
```
However, this is the KL Divergence of a gaussian and a standard gaussion. According to the paper, the KL Divergence is about two non-standard gaussian.