MRF and Infer.NET (Migrated from community.research.microsoft.com)

All replies

• 

  

  0

  Sign in to vote

  minka replied on 10-17-2009 11:10 AM

  You can specify clique potentials using constraints or soft constraints.  For example, if you want a potential which is 0.2 when x1==x2 and 0.8 otherwise, you can say:
    Variable.ConstrainEqualRandom(x1==x2, new Bernoulli(0.2));
  which is equivalent to:
    Variable.ConstrainEqual(x1==x2, Variable.Bernoulli(0.2));
  The latter form is preferable if you want 0.2 to be a parameter of the model.  For example:
    Variable<double> p = Variable.Beta(1,1);
    Variable.ConstrainEqual(x1==x2, Variable.Bernoulli(p));
    Variable.ConstrainEqual(x2==x3, Variable.Bernoulli(p));
    etc.

  Friday, June 3, 2011 6:03 PM
• 

  

  0

  Sign in to vote

  jwinn replied on 10-18-2009 6:16 PM

  Whilst you can construct this model in Infer.NET, as Tom describes, you should be aware that Infer.NET does not yet handle large grid models efficiently - such as would be needed to segment a normal sized image.  This is a high priority feature and we intend to support it in a forthcoming release. 

  Best regards,

  John W.

  Friday, June 3, 2011 6:03 PM
• 

  

  0

  Sign in to vote

  smh0816 replied on 09-24-2010 3:49 AM

  I still have no idea how to model MRF in general cases. What it the compatibility function gives different values when (x1 == 1 && x2 == 1) and (x1 == 2 && x2 == 2)? What it we have multiple labels? I think discrete 2D distribution is needed for the compatibility function.

  Friday, June 3, 2011 6:03 PM
• 

  

  0

  Sign in to vote

  minka replied on 09-24-2010 7:56 AM

  You can use multiple constraints to build up arbitrarily complex compatibility functions.  See for example these threads: http://community.research.microsoft.com/forums/t/4542.aspx
  http://community.research.microsoft.com/forums/t/4448.aspx

  • Marked as answer by Microsoft Research Friday, June 3, 2011 6:04 PM

  Friday, June 3, 2011 6:04 PM