cyclic graph for CRF

[Wenya Wang]

Hi all,

I would like to test with pystruct with more general CRF graph structures. Currently, my graph includes a linear chain structure, as well as tree structure. (you can think of it as a skip-chain CRF). Hence, it is possible to have cyclic components.

Is it possible to train with GraphCRF? And how about the inference method? Is max-product able to deal with this cyclic graph? I checked GRMM in mallet and they use tree reparameterization (TRP) for inference. Is any method in pystruct similar to that?

Thanks!

Wenya

[Andreas Mueller]

Hi Wenya.
Sorry for the slow reply over the holidays.
PyStruct can work with a large array of inference methods. max-product might not be the best choice here.
I recommend installing OpenGM, which comes with many different algorithms. I'm not sure what TRP is referring to,
is that the same as TRW?

You can use GraphCRF for training for you graph. Just be aware that all edges in GraphCRF have "the same meaning", and there is a single n_states x n_states transition matrix.

Best,
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.

[Wenya Wang]

Hi Andreas,

Thanks for your reply! 

I am not quite familiar with all the inference method. TRP is the one I found in mallet, which represent tree-based reparameterization method for tree-CRF inference. 

I will try with graphCRF. In that case, it should work. Maybe I also need to get more familiar with tree-based inference methods. 

Thanks for the explanation, it helps!

Best regards,

Wenya