noise problem

Lucille
2008-11-03
2013-04-16
  • Lucille
    Lucille
    2008-11-03

    Hello,

    I`m having a problem trying to add noise to a net. The messege i get is the following:

    org.joone.engine.Synapse - [] Cannot add noise because weight array is null.

    However i'm having results different from zero:

    Epoch;RMSE;d(index of agreement)
    0;0.0;1.7976931348623157E308
    20;0.17598680425089797;0.7350177377372829
    40;0.17303501365454285;0.7504335194970658
    60;0.1706777430149117;0.7608232244424271
    80;0.16901938792275684;1.7976931348623157E308

    I was looking for into the souce code of joone engine, and i didn't find it. I downloaded the source code availeble from the joone page, exits any chance that the library code isn't the same from the source engine code?

    thanks

    ps: i would like to join to development team, plese tell me what i could do.

     
    • ferra4models
      ferra4models
      2008-11-03

      You need to first connect the synapse to both an input and output layer. This will then set the internals of the synapse, which, in this case, would be the dimensions of the weight array. Once that is done it should work.

      Note that the weight array is randomised by default, so you don't have to do it manually actually.

      Regards
      MG
      http://www.ferra4models.com

       
    • Lucille
      Lucille
      2008-11-03

      Hello,

      The code used to build the net:

              LinearLayer input = new LinearLayer("input");
              TanhLayer hidden = new TanhLayer("hidden");
              TanhLayer output = new TanhLayer("output");

              FullSynapse input_hidden = new FullSynapse();
              FullSynapse hidden_output = new FullSynapse();

              NeuralNet red = new NeuralNet();
             

              // <editor-fold desc="arquitectura" defaultstate="collapsed">

              //configurando dimiensiones
              input.setRows(9);
              hidden.setRows(14);
              output.setRows(1);

              //conectando sinapsis de input -> hidden
              input.addOutputSynapse(input_hidden);
              hidden.addInputSynapse(input_hidden);

              //conectando sinapsis de hidden -> output
              hidden.addOutputSynapse(hidden_output);
              output.addInputSynapse(hidden_output);
              // </editor-fold>

              the code continues... file input & output configuration in the net.

      The net is save in a serialized object, and then restore. I'm doing early stopping, cross-validation.

      I'm having a very very sirious problem with overfitting and also the net is learning quickly... so during the validation when the difference between this epoch and the previous one is less than 0.03 i want to add some noise, and see what happens there. but i get the msg i wrote before.

      thanks

       
    • Xharze
      Xharze
      2008-11-07

      I've sent you an email via your SourceForge about joining the development team.