Collapsed Gibbs Sampling of HDP

[dionman]

Is there an implementation of collapsed gibbs sampling in hierarchical dirichlet process?

[dtrckd]

Hello,

I have implemented a CGS inference for HDP for MMSB (which is a LDA kind model for networks). (Which was inspired from the present kzhai code implementation).

I have integrated it in a open platform that I develop for experiment/research reproducibility, accessible here : https://github.com/dtrckd/pymake

After installing pymake, you can run the experiment as follows :

pymake default_expe -i 30 -c BA -n 100 -m immsb_cgs -x fit -w

which means:

• immsb_cgs: use the (infinite) mmsb model with CGS inference
• -i 30 : 30 iterations
• -c BA : use a Barabasi-Albert random generated graph as input data
• -n 100: 100 nodes in the graph
• -x fit : run the fit script
• -w: save the results
  Then you can plot the perplexity convergence with the following command :

pymake default_expe -i 30 -c BA -n 100 -m immsb_cgs -x plot

I hope it'll help.