Variational-Autoencoder | Tensorflow implementation of Variational Auto | Machine Learning library

I am running a VAE in Keras. the model compiles, and its summary is :

however, when I try to train the model I get the following error:

I want to train a VAE that had a huge dataset and decided to use a VAE code made for fashion MNIST and popular modifications for batch-loading using filenames that I found on github. My research collab notebook is here and a sample section of dataset.

But the way the VAE class is written it does not have a call function which should be there according to keras documentation. I am getting the error NotImplementedError: When subclassing the Model class, you should implement a call method.

I'm trying to train an auto-encoder in keras. In the end I would like to have a separate encoder and decoder models. I can do this for an ordinary AE like here:https://blog.keras.io/building-autoencoders-in-keras.html

However, I would like to train a conditional variant of the model where I pass conditional information to the encoder and the decoder. (https://www.vadimborisov.com/conditional-variational-autoencoder-cvae.html)

I can create the encoder and decoder fine:

This is a related question to this - Variational autoencoder and reconstruction Log Probability vs Reconstruction error

I'm trying to understand how variational autoencoders are optimized. I've read the math behind it and I think I understand the general concept of variational inference and the reparameterization trick used for the latent space.

I've seen some examples where the input and the output are compared to each other using cross entropy and KL divergence is on the latent variables. And then this loss is minimized.

On the other hand, there are other examples which uses the log probabilities and the KL divergence to generate the evidence lower bound (ELBO). Then the negative of the ELBO value is minimized.

In both, the latent space is partitioned based on the patterns of the inputs (numbers in MNIST for example). So I wonder if the ELBO is or contains information similar to the reconstruction loss.

I am adapting this implementation of VAE https://github.com/keras-team/keras/blob/master/examples/variational_autoencoder.py that i found here https://blog.keras.io/building-autoencoders-in-keras.html

This implementation does not use convolutional layers so everything happens in 1D so to speak. My goal is to implement 3D convolutional layers within this model.

However I run into a shape mismatch at the loss function when running the batches (which are of 128 samples):