the comment in vae.py is totally wrong !

[wangpeng138375]

`# ELBO
marginal_likelihood = tf.reduce_sum(x * tf.log(x_) + (1 - x) * tf.log(1 - x_), 1)
KL_divergence = 0.5 * tf.reduce_sum(tf.square(mu) + tf.square(sigma) - tf.log(1e-8 + tf.square(sigma)) - 1, 1)

marginal_likelihood = tf.reduce_mean(marginal_likelihood)
KL_divergence = tf.reduce_mean(KL_divergence)

ELBO = marginal_likelihood - KL_divergence

# minimize loss instead of maximizing ELBO
loss = -ELBO`

The result going right is just lucky.