|
Is there a software package out there that supports arbitrary structures of graphs for applying CRF model on them ? Even if it only support cliques of size 1 and 2, it would be useful to me. Mallet and most other CRF packages support only linear chain CRFs. GRMM supports only Dynamic CRFs (from what I can understand from its code, even though it claims to support general structures). I am implementing CRFs for my own research, but was wondering if I could save time by using any such package or at least compare the performance of my implementation against it. |
|
Do you know of any group outside UMass that has used Factorie ?
(Sep 03 '10 at 15:52)
Aman
@Aman: In addition to UMass, FACTORIE is being used at Xerox Research Centre Europe. It has been run by folks at MIT, John Hopkins and UMD, but I think it was early for them to be making heavy use of it. Probably best for folks outside UMass to wait for version 0.9 before trying to do more than simple experimentation.
(Sep 07 '10 at 15:16)
Andrew McCallum
@Andrew. Ok, Thanks. I am eagerly awaiting its release. Meanwhile I am writing my own package for this.
(Sep 07 '10 at 15:53)
Aman
|
|
The following graphical-model toolkits support what you want: LibDAI doesn't have much code for training, but you can use the provided inference algorithms to get feature counts and then do feature weight updates yourself (using simple SDG etc.). (I agree that GRMM in Mallet is not general enough and can only do certain structures easily.) What limitation do you see in GRMM?
(Aug 25 '10 at 08:42)
yura
The limitations are that you apparently cannot easily create your own graph in GRMM and run inference in it. It has to be unrollable according to some template. See also a comment to this answer: http://metaoptimize.com/qa/questions/2086/experience-with-factorie-for-crfs-with-arbitrary-structure#2089
(Aug 26 '10 at 21:48)
Frank
|
|
This Mark Schmidt software has support for conditional and unconditional training of undirected graphical models. GRMM, a part of mallet, also supports general CRFs. GRMM doesn't support general CRFs. Although it says so on their main page, a detailed inspection of the code doesn't reveal so. Also, see what @Frank has to say.
(Aug 25 '10 at 13:29)
Aman
|
|
Mark Schmidt has a matlab implementation of CRFs that supports general graph structures. Also, Kevin Murphy had an older implementation of CRFs that works for general graphs. |
