search

flyozcn / haines

Automatically exported from code.google.com/p/haines
0 stars 0 forks source link

The result model varies for each run #2

Open GoogleCodeExporter opened 8 years ago

GoogleCodeExporter commented 8 years ago 
Hi,

   I have little knowledge on DP. However, I used the code to estimate the density of some mixtures and try to get the mixture components, It seems that the algorithm produces different results each time. The self.v values differ each time.

   I am not sure whether this is a bug or it is the nature of the variational inference algorithm.

Original issue reported on code.google.com by caofan....@gmail.com on 17 Sep 2013 at 11:23

GoogleCodeExporter commented 8 years ago 
I am a little late - Google never notified me of this! Only noticed because it 
appeared when I transferred this project to github - sorry.

Anyway, variational methods are initialised with random numbers, and then 
converge to a local minima - different initialisations will lead to different 
results. So yes, getting different values each time is expected. You still 
expect to get something similar however - if there is a massive difference in 
the meaning of the output then there is problem, though note that mixture 
components can come out in an arbitrary order.

Original comment by thai...@gmail.com on 26 Mar 2015 at 10:51