taddylab / maptpx Goto Github PK

we're running maptpx on some larger datasets. When we do this we get infinite L sometimes in this line:

if(is.infinite(L)){ L <- sum( (log(Q0)*col_sums(X))[Q0>0] )}

but the problem is that the sum() returns NA.
This then causes a crash later at
if(QNup$L < L){
because L is NA.

We tried to fix this by:
i) comment out the if(is.infinite(L)) line
ii) comment out reldif # reldif <- dif/L
iii) change tolerance check to dif instead of reldif: if(abs(dif) < tol){

This seemed to work except that it crashed (presumably after 1000 iterations) at
if(((iter+1)%%1000)==0){
cat(sprintf("p %d iter %d diff %g\n",
nrow(theta), iter+1,round(diff)))

due to the bug round(diff) should be round(dif).

We've fixed that and are now re-running - currently waiting for results.

What do you think? I think that the tolerance check should be on the difference in loglikelihood anyway - the relative change in loglikelihood is nonsense (at least so I always tell my students!)
But maybe there is also something to understand about why that sum() produces NA?

We can prepare a formal pull request with these changes if you like.

Hello,

I would like to know if there is a way of using the function "topics" from the "maptpx" package anymore considering that the package is not usable now on R version 4.0.0 because of lack of updates. Could you re orientate me towards other solutions ?
Thanks for your kind help.
Best regards,
Pierre