When Training HDP model perplexity and LL per word return Nan

[LiuTaiShi]

Hi @bab2min,

I'm now training a HDP model and report perplexity and loglikelihood for every 200 iters.
But when the number of iteration reaches 36600(as presented in pic), the both metrics retrun Nan.
I noted that this bug has been fixed. So what's wrong with the training?
By the way, I'm wondering what's the apt standard to choose model from perplexity(or LL per word)?
Thank you!