I am new to pyro and generative modelling.
Recently, I have been interested in models that combine the interpretability of a discrete structured PGM with NNET likelihood functions. In particular, a model such as, Nonparametric Variational Auto-encoders for Hierarchical Representation Learning. The model uses an alternating optimisation strategy, alternating between optimising the nested CRP and the VAE. The generative model consists of the Nested CRP that generates z and a NNET decoder mapping z to x.
I was excited about pyro because I believed that pyro’s flexibility would let us design models as above. Building a simple VAE is quite straightforward in other toolkits as well, such as PyTorch. Although, the DMM and AIR examples show that how powerful a PPL such as pyro could be, building models with discrete structures is still not clear to me. For e.g. in a language acquisition domain, you might want to infer the underlying linguistic structure (discrete), given speech etc.
Is pyro in the current form not suitable for models such as the one referenced above? or models that combine complicated discrete structures with NNETs.
A noob to this field. Your thoughts and comments would help me a lot.