[Topic-models] Need help with fast LDA gibbs sampling

Dave Newman newman at uci.edu
Fri Mar 4 09:56:34 EST 2011

Chi --

There is some additional overhead with doing the FastLDA computation.
But this overhead becomes small compared to the savings due to doing
<< T loops inside the Gibbs sampler.  The savings will increase for
increasing T and decreasing \alpha.  So you should eventually see a
speedup (I'm not surprised you are not seeing a speedup for T=15).
Code for FastLDA can be found here:



On Fri, Mar 4, 2011 at 3:22 AM, chi yuan <carrier24sg at yahoo.com.sg> wrote:
> I've look at my code for a couple of hours before seeking help here.
> In response to my previous thread, I created a python script to test FastLDA's
> sampling method to see it works.
> I compare collapsed gibbs sampling from "finding scientific topics" and "Fast
> LDA" by Ian Porteous to see if the latter really speed up, if not comparable.
> I ran them over a small sample of 100 documents , 15 topics beta = 0.01 and
> alpha = 50/T
> The test result shows that the latter is much slower than the former! Can anyone
> please tell me  it is more possible a implementation/logic problem than using a
> wrong setting here? I'm willing to send my script if anyone offers to help.
> _______________________________________________
> Topic-models mailing list
> Topic-models at lists.cs.princeton.edu
> https://lists.cs.princeton.edu/mailman/listinfo/topic-models

David Newman
Research Faculty
Department of Computer Science
University of California, Irvine

More information about the Topic-models mailing list