Tensorflow probablity layer with mixed distribution families per-component?

Question

I want to train a neural net where output is tensorflow-probability distributions, similar to https://www.tensorflow.org/probability/examples/Probabilistic_Layers_VAE. However u