The vae theory is not much intuitive.
Although there are a couple of blogs explaining vae.
What does the z_mean and z_var mean to z?
It seems each number has own z_mean and z_var.
The epsilon is just adding noise to z_mean.
Then the z space gets simulated to Normal(0,1).
How is it possible?

When running the training_m2.py, there is ValueError.
It happens on the part of fit.

training.fit([z1_train]+[y_train]+[z1_train]+y_u_train, <---here
z1_train,
batch_size=100,
nb_epoch=nb_epoch,
validation_data=([z1_test]+[y_test]+[z1_test]+y_u_test, z1_test), <---here
callbacks=[EarlyStopping(patience=10)],
shuffle=True)

How to use inputs with Gaussian distribution instead Bernoulli?

In keras 2 environment, running the training1 causes error.
It says, two parameters from custombatch normalization are not references, so immutable to update.
By the way, I changed the initilization to initializers since keras 2 provides no more support of initialization class.

Guess this class change somehow modified the weights of the custom layer to be immutable.
Any help is welcom.