The topic model takes too long to converge: is there a max no. of iterations?

[kkdey]

If you use the Github version of the package maptpx

library(devtools)
install_github("kkdey/maptpx")

then you can choose the tmax as an option under control.

The default is tmax=10000

out <- FitGoM(ex.counts, K=4, tol=100, control=list(tmax=100))

[jhsiao999]

10,000 iterations take a long time. Are there any other ways around this?

On Thu, Sep 22, 2016 at 10:30 AM Kushal K Dey notifications@github.com wrote:

Closed #25 https://github.com/kkdey/CountClust/issues/25.

— You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub https://github.com/kkdey/CountClust/issues/25#event-799042077, or mute the thread https://github.com/notifications/unsubscribe-auth/AGwfei0VJadsg-XkUN_0Tb7qNcYuJk6Wks5qsp8VgaJpZM4KEB9M .

[kkdey]

The default is 10,000 ....but you can set it to any value you want, see the example above for instance

[jhsiao999]

I understand. How about tolerance? Makes more sense to constraint on the likelihood change that on the number of iterations.

On Thu, Sep 22, 2016 at 10:34 AM Kushal K Dey notifications@github.com wrote:

The default is 10,000 ....but you can set it to any value you want, see the example above for instance

— You are receiving this because you commented.

Reply to this email directly, view it on GitHub https://github.com/kkdey/CountClust/issues/25#issuecomment-248939603, or mute the thread https://github.com/notifications/unsubscribe-auth/AGwfeilpNXbWnnooUTcDb7muye7OGOe1ks5qsqAKgaJpZM4KEB9M .

[kkdey]

There is already a tol function for that, default to tol=0.1, meaning it stops when the log likelihood change between successive iterations is less than 0.1, you can change that too if you want to stop early

[jhsiao999]

What's the range of tol? Also does this define the percent decrease in likelihood? For example, .1 is 10 percent decrease in likelihood.

On Thu, Sep 22, 2016 at 10:43 AM Kushal K Dey notifications@github.com wrote:

There is already a tol function for that, default to tol=0.1, you can change that too if you want to stop early

— You are receiving this because you commented.

Reply to this email directly, view it on GitHub https://github.com/kkdey/CountClust/issues/25#issuecomment-248942190, or mute the thread https://github.com/notifications/unsubscribe-auth/AGwfes-aUlgWHs0uYn_HSDg1x-HwBiY9ks5qsqIWgaJpZM4KEB9M .

[kkdey]

no it is stronger, it is the absolute change, not relative change in loglikelihood, the smaller the tol, the better it is, but also would take more time to converge.

[jhsiao999]

Okay let's add the definition to the function description. Thanks!

On Thu, Sep 22, 2016 at 10:53 AM Kushal K Dey notifications@github.com wrote:

no it is stronger, it is the absolute change, not relative change in loglikelihood, the smaller the tol, the better it is, but also would take more time to converge.

— You are receiving this because you commented.

Reply to this email directly, view it on GitHub https://github.com/kkdey/CountClust/issues/25#issuecomment-248945234, or mute the thread https://github.com/notifications/unsubscribe-auth/AGwfeufIpE02Z1AuU1HjbPt1hSo8s5CIks5qsqSEgaJpZM4KEB9M .