Training problem

Help
2012-11-06
2013-01-04
  • Daniel Dante

    Daniel Dante - 2012-11-06

    Hi all. I'm trying to train a dataset, recovered from the UCI Machine, with name "Bank Marketing". I normalized the data, and transformed the atributes into 30 inputs and 2 outputs. Using the Neuroph Studio, and using a low number of rows, like 1500 - 3000 rows (the total data is something about 45k), I did the 0.01 network error, but, during the tests, the YES OUTPUT, that correspond to my 1 0 binary, never comes! I don't know what I'm doing wrong, but I've try almost everything that envolves architeture and training set variation, but all tests returns 0 1 (NO) . Can anyone help me? Thanks!

     
  • Daniel Dante

    Daniel Dante - 2012-11-13

    There's support no more? Neuroph isn't the best choice for java NN framework?

     
  • Anonymous - 2012-11-19

    HI :d

    nice to get your message….

    Maby i can give you a hint. The output neuron maby wont give you 1.0, it comes up with 0.99998 or something.
    So look at the output >0.9, for true :d

    If you are using a YES/NO output, i even dont know what you mean.

    Why you are using 1500 rows????
    I made a small robot :d, and tried 8 rows, more rows, and only three rows. Less are sometimes better.

    Try to train the net with the same input several times, but take care! too often and your output gets overdriven and bring up only fault values.

    Hope i can help, tell us :d

    Lg Andy

     
  • Zoran Sevarac

    Zoran Sevarac - 2012-11-21

    Hi, sorry for delay in replying, we've been busy with new release…
    Try also using single  output where 1 corresponds to YES and 0 to NO
    If it doesent work you can send me data set sample to try it and I'll let you know

    Zoran

     
  • Daniel Dante

    Daniel Dante - 2012-11-26

    Sev, thanks for the reply, I imagine how hard are the release of new versions of Neuroph.

    The problem is: I study almost all the sample projects avaiable on your site. I normalized more than 4 datasets from the UCI Machine learning site, from different cases, and nothing! I also copied your solution, with your dataset normalized, like the Habermans Survival sample, but without success. I don't know what I'm doing wrong.. I think just in case of Zoo I had success copying your solution and Architecture. First I tought that I was normalizing the data wrong, but with your dataset normalized, I cannot did too! There's a hint of what I'm doing wrong?

    Thanks in advance!

     
  • Zoran Sevarac

    Zoran Sevarac - 2012-11-26

    Ok. Let me know which version of Neuroph you are using.

    All the examples are done in 2.6 and you should be able to download complete Neuroph Studio project for most of them and run it.

    Also lets focus on some specific which you can choose, compare results  and figure out what wrong.

    Zoran

     
  • Daniel Dante

    Daniel Dante - 2012-11-26

    Ok Sev. I use the 2.6 version, but now I just downloaded the full pack of version 2.7.
    I have some doubts, like which differences  between the "Total Network Error" and the "Total Mean Square Error"?
    I'll show you my try on the "Breast Tissue" sample.
    I start creating a Neural Network project on Neuroph Studio. After, I created a Training Set and loaded the file "breasttissue-normalized.txt". Then I created a Neural Network with 9 inputs, 20 Hidden neurons and 6 outputs. Now, I clicked on the "Train" button, choosing, like the best architecture from your example, the learning rate 0.2 and the momentum to 0.4. In your sample, on 66 iterations the Neural Network comes to 0.0058 Total Mean Square Error. But mine Neural Network just got 30K iterations and come to 0.0528, too far from your experiment. The Total Network Error, comes to 0.09, a number that I judge to be high. Choosing a random rows to test, I realize more errors than I can assume. The situation isn't good for me! LOL!
    Thanks again!

     
  • Musa Yusuf

    Musa Yusuf - 2013-01-04

    I am experimenting Optical Recognition of Handwriting Digit with dataset from UCI Machine Learning , The network i created has 64 input neurons, 44 hidden neurons and 1 output neurons. I was able to train the network with the TRAIN DATA SET but i don't know how to test the neural network with the TEST DATA SET. Is there any idea of the way forward please?

     
    Last edit: Musa Yusuf 2013-01-04

Get latest updates about Open Source Projects, Conferences and News.

Sign up for the SourceForge newsletter:





No, thanks