Divide by zero warning? about pygmmis HOT 2 CLOSED

Hi Jieyu,

I reproduced your problem. The root cause is in the first line of the traceback: it tells us that there's zero probability for some sample. This should be prevented by only evaluating the probability of each sample over the neighborhood H. In your case (the default configuration), this neighborhood comprises all samples, including some that may have very low probability under any of the GMM components. So, you've got a numerical underflow. I suspect that this happens because some components shrink very strongly.

Use the argument cutoff=5 in fit, which limits the evaluation to 5-sigma neighborhoods of each component (where each sigma refers to the covariance of the component). That prevents the underflow by ignoring samples that are too far away for any component and speeds up the process considerably. I just ran the full simu with K=10 in 5 seconds.

Another option is increasing the minimum component covariance w to something larger. 1e-6 doesn't help for the range your data cover.

from pygmmis.

Thanks for helping me out and for a quick reply!! Now it works fine for my model.

from pygmmis.