Jointly optimizing autoencoder and fully connected network for classification

Question

I have a large set of unlabeled data and a smaller set of labeled data. Thus, I would like to first train a variational autoencoder on the unlabeled data and then use the encode