Description
(Elham Azizi asks about DP mixture models for use in some of her work with Dana Pe’er in computational biology.)
part of my reply below:
DPs in their full infiniteness aren’t possible in Edward for similar reasons as Stan. It requires lazy evaluation to bake in a generative process with an infinite number of latent variables but whose output only uses a finite number of latent variables. We’re still working on that regard.
Truncated DP mixture models are definitely possible, e.g., setting the maximum number of stick breaks a priori. And it seems like that’s what your work uses. I imagine Edward is the fastest for inferring it in practice because all other systems out there use MCMC or SMC. And Edward will handle data subsampling and things like GPU support or distributed training if you want to take advantage of them.
Would be great if someone gets a truncated DP mixture model up and running!