Inefficient sampling of large categorical model

[junpenglao]

I have been trying to implement a large categorical model in pymc3 but without much progress. The model I am trying to implement is detailed in the last chapter of Lee and Wagenmakers' Bayesian Cognitive Modeling book.

In brief, each entry of the response data is generated by a different categorical pdf (constructed under some hierarchical constraints).Thus the probability vector of these categorical pdf is trial/subject specific. I have the following difficulties:

1, The model is very slow. I adapted the codes from the book (JAGS and STAN), both end up too many nodes for Theano. I hadn't manage to sample using NUTS - it took too long to compile.
2, I can use Metropoli as step method but the sampling of categorical parameter is incorrect (the trace is nearly uniform). I also try pymc2 but the problem remains.

Is there a working example of a similar problem? Or how should I build my model in this case (e.g., categorical response with a large sets of priors) to make it more efficient?

All my attempts could be found here: https://github.com/junpenglao/Bayesian-Cognitive-Modeling-in-Pymc3/blob/master/CaseStudies/NumberConceptDevelopment.ipynb

edit: this might relate to #624

[junpenglao]

Close as this problem is resolved using some matrix manipulations instead of creating large graph in theano.