search

chokkan / crfsuite

CRFsuite: a fast implementation of Conditional Random Fields (CRFs)
http://www.chokkan.org/software/crfsuite/
Other
641 stars 208 forks source link

Question about CRF model #87

Open TompsonL opened 7 years ago

TompsonL commented 7 years ago

Hi ,

Can I use the crfsuite to implement the model in the attached image? Thanks a lot.

crf

usptact commented 7 years ago

If you can extract a "sequence" from your image, and expect sequence labels + one global label, I suggest you to look at TriCRF instead. https://github.com/minwoo/TriCRF

The implemented model by TriCRF is:

       y
  /    |   \
h  -   h  -  h
|      |     |
x      x     x

The sequence labels as well as the global image label must be discrete.

TompsonL commented 7 years ago

Thanks for your response. Looks the h nodes in TriCRF are in a chain, Can I use a full connected ?

usptact commented 7 years ago

@TompsonL In principle yes, in practice inference in such graphs would be ahem difficult. Even this triangular chain CRF is fairly expensive (the way it is implemented). Unfortunately I am not familiar well enough with graphical models to suggest or make bigger claims.

usptact commented 7 years ago

@TompsonL I would try out a neural architecture - sequence model as for NER for "h" and classification on top of that for "y".

TompsonL commented 7 years ago

Thanks for your effort.

On Wed, Aug 2, 2017 at 1:08 PM, Vladislavs Dovgalecs < notifications@github.com> wrote:

@TompsonL https://github.com/tompsonl I would try out a neural architecture - sequence model as for NER for "h" and classification on top of that for "y".

— You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub https://github.com/chokkan/crfsuite/issues/87#issuecomment-319568808, or mute the thread https://github.com/notifications/unsubscribe-auth/AdJ8GJ_yLMltBGAELEObwCAzThRRNDKOks5sUARSgaJpZM4Ono8S .