Sign in
Microsoft.com
 
Microsoft
 
 
 

locked
How to perform sth like AllTrue or Sum Factor in Gibbs Sampling? RRS feed

 > 

  • Question

  • Question
    Sign in to vote
    0
    Sign in to vote

    Hi, Infer.Net People,

              I know Factor "AllTrue" or "Sum" is not supported in Gibbs Sampling,  but I do want some way to summarize a VariableArray to a single value using Gibbs Sampling.  Because Gibbs sampling shows much better and robust performance in my model,  switching to EP or VMP is not a option.

    I tried to implement the Sum Operation using similar way as

    public Variable<double> Sum(VariableArray<double> array)
{
  Range n = array.Range;
  var SumUpTo = Variable.Array<double>(n).Named("SumUpTo");
  using (var fb = Variable.ForEach(n))
  {
    var i = fb.Index;
    using (Variable.Case(i, 0))
    {
      SumUpTo[i] = Variable.Copy(array[i]);
    }
    using (Variable.If(i > 0))
    {
      SumUpTo[i] = SumUpTo[i - 1]+ array[i];
    }
  }
  var s= Variable.Copy(SumUpTo[((Variable<int>)n.Size) - 1]);
  return s;
}

    But a error "Internal: schedule splits group 169 " pop out.

    Thank you for any help in advance!

     

    Sunday, February 1, 2015 4:11 PM

All replies

  • Question
    Sign in to vote
    0
    Sign in to vote


           I finally found a approximate way to implement a Fake AllTrue factor which can be used in Gibbs Sampling.

      public static Variable<bool> FakeAllTrue(VariableArray<bool> arr)
        {
            Variable<bool> ret = Variable.Bernoulli(0.5);
            var index = arr.Range;
            using (Variable.ForEach(index))
            {
                using (Variable.If(arr[index]))
                {
                    using (Variable.If(Variable.Bernoulli(1.0 / Variable.Double(index.Size))))
                    Variable.ConstrainTrue(ret);
                }
                using (Variable.IfNot(arr[index]))
                {                    
                        Variable.ConstrainFalse(ret);
                }
            }
            return ret;
        }


        public static void TestAllTrue()
        {
            int n = 4;
            var index = new Range(n);
            for (int i = 0; i < 9; i++)
            {
                double probT = 0.1 * (i + 1);
                VariableArray<bool> arr = Variable.Array<bool>(index);
                arr[index] = Variable.Bernoulli(probT).ForEach(index);
                var p1 = Variable.AllTrue(arr);
                InferenceEngine engine = new InferenceEngine();
                engine.ShowProgress = false;
                Console.WriteLine("probT: " + probT);
                Console.WriteLine("true: " + engine.Infer(p1));
                var p2 = FakeAllTrue(arr);
                Console.WriteLine("approx: " + engine.Infer(p2));
            }
            Console.ReadKey();
        }

    Although it is super accurate, it is better than none.

    Cheers!

    Sunday, February 8, 2015 3:36 AM