README

I modified gibbs.c to not die when NaN encountered. Also, I modified sldea.em.R to handle case when a topic was not assigned any words. -awc

You can install R package from source with:
install.packages('/Users/awculott/Programs/lda', repos = NULL, type="source")


$ diff R/slda.em.R ~/Downloads/lda/R/slda.em.R 
13,14d12
< 	# to avoid topics with 0 counts. add small count to all docs for topic 1
< 	document.sums[,1] <- document.sums[,1] + 0.0000000001
20,21c18
<             library(splines, quietly=TRUE, verbose=FALSE)
<             library(penalized, quietly=TRUE, verbose=FALSE)
---
>             library(penalized)


$ diff src/gibbs.c ~/Downloads/lda/src/gibbs.c 
935,939d934
< 		if (REAL(beta)[kk] != REAL(beta)[kk]) {
< 			Rprintf("NaN! beta[0]=%g beta[1]=%g\n", kk, REAL(beta)[0], REAL(beta)[1]);
< 			REAL(beta)[kk] = 0.00001;
< 			Rprintf("\t now beta[%d]=%g nw=%d\n", kk, REAL(beta)[kk], nw);
< 		}
951,953d945
< 		  if (p[kk] != p[kk]) {
< 			Rprintf("encountered NaN with change=%g dv[dd]=%g var=%g beta[kk]=%g nw=%d\n", change, dv[dd], var, REAL(beta)[kk], nw);
< 		  }
967,968c959
< 		Rprintf("pr of topic %d is NaN. Assigning random probability.\n", kk);
< 		p[kk] = unif_rand();
---
> 		p[kk] = 0.0;
994c985
< 	  *z = 0;
---
> 	  *z = 0