Plans on approximate inference algorithms?

[Shuyang Sheng]

Hi guys,

While experimenting on a 2D binary pairwise CRF model, I found that the inference in pystruct is considerably slower than the Mean Field method provided by the Matlab package UGM.

I'd be happy to contribute some approximate inference methods like Gibbs sampling or Mean Field if that's part of the plan. Advises/references welcome.

[Andreas Mueller]

Hi Shuyang.

I'm a bit hesitant to include more inference methods. There is OpenGM, which is integrated into pystruct.
It comes with A LOT of inference methods. Have you tried using it?

http://hci.iwr.uni-heidelberg.de/opengm2/

I didn't make it a required dependency because it is a bit tricky to build.
We could probably try making a conda package and / or os X and windows wheels to ease installation.
I don't really have time for that at the moment, though.

Cheers,
Andy

--
You received this message because you are subscribed to the Google Groups "pystruct" group.
To unsubscribe from this group and stop receiving emails from it, send an email to pystruct+u...@googlegroups.com.
For more options, visit https://groups.google.com/d/optout.

[Shuyang Sheng]

Hi Andreas,

Oh I see, my apologies for the silly question - I thought OpenGM was only one inference method. Apparently I should've go through the documentation more carefully.

If I came across a situation that I'll need to play around with OpenGM, I'll try build some binaries and make conda packages etc. if I have extra time.

[Andreas Mueller]

Good luck!
And any feedback on your experience with OpenGM is welcome.