is Pyro any helpful in performing the correspondence analysis for the categorical data?
Also, I don’t yet understand how good is Pyro for creating models almost from scratch (“zero-conf”). It looks to me like structure of the model should be quite well defined forehand by the human. In particular, the probabilistic distributions family, the dependencies between input/latent/output variables should be quite explicitly defined to do anything based on that model. Or am I wrong and the model structure could be inferred as well?
P.S. I’ve invested only a few hours in learning Pyro, so pardon me, if my questions are naive.
how good is Pyro for creating models almost from scratch
Pyro is mainly a good way of expressing domain knowledge as a probabilistic program (a generalization of probabilistic graphical model). If you have little or no domain knowledge about the structure of your data or is distributions, then probably a black-box ML method is more appropriate than Pyro. Currently we have no structure learning inference in Pyro.
Pyro any helpful in performing the correspondence analysis
Pyro has no built-in inference for correspondence analysis. You could probably implement that by hand in Pyro, but I would only recommend doing so if your correspondence analysis were part of a larger structured probabilistic model. Probably not the easiest way to start using Pyro
