Change source code to save MCMC model to text file... However, weights in model file are large, not able to inference, eg. for classification task, prediction result is sigmoid(x), where x = 1000

[Jiayin Lei]

Wonder what is wrong with the MCMC model? train and test process looks fine

cmd: ../bin/libFM -task c -train ../data/spcz_train.txt -test ../data/spcz_test.txt -dim '1,1,8' -verbosity 1 -out ../data/spcz.predict -save_model ../model/spcz.mdl

#Iter= 96    Train=0.933577    Test=0.882    Test(ll)=0.14724
#Iter= 97    Train=0.933921    Test=0.882    Test(ll)=0.146699
#Iter= 98    Train=0.93274    Test=0.882    Test(ll)=0.146124
#Iter= 99    Train=0.932643    Test=0.883    Test(ll)=0.145564

model file looks like this:

#global bias W0

-6.38154

#unary interactions Wj

0.416005

-0.175864

-1.92583

-0.111393

0.00136478

3.04672

0.731291

2.621

-0.739681

-2.36962

-6.19163

-1.01834

0.161435

-5.00533

2.09333

3.73539

0.616305

0.203633

-8.45758

-3.0175

-2.10704

0.284601

...

When inference using below code written in Scala, I get variable pred equals to 1000 or more, which sigmoid(pred)=1.0

  def predict(testData: Vector): Double = {

    require(testData.size == numFeatures)

    var pred = intercept

    if (weightVector.isDefined) {

      testData.foreachActive {

        case (i, v) =>

          pred += weightVector.get(i) * v

      }

    }

    for (f <- 0 until numFactors) {

      var sum = 0.0

      var sumSqr = 0.0

      testData.foreachActive {

        case (i, v) =>

          val d = factorMatrix(f, i) * v

          sum += d

          sumSqr += d * d

      }

      pred += (sum * sum - sumSqr) * 0.5

    }

    task match {

      case 0 =>

        Math.min(Math.max(pred, min), max)

      case 1 =>

        1.0 / (1.0 + Math.exp(-pred))

    }

  }

BTW, spcz.predict is good. Scores are distributed between [0, 1]

Please can you advise what is wrong with my procedure...

Thanks in advance!!!