Further Neuroph Studio 2.6 Problem Notes

Help
2012-01-30
2012-12-24
  • Arthur Younger

    Arthur Younger - 2012-01-30

    Further ot my previous note, I did a full uninstall of Neuroph Studio, verified Netbeans 7.1 install, and reinstalled Neuroph Studio 2.6.
    Then stepped through the very simple MLP exercise as documented.  Here are my observations:

    1.  The basic part of the experiment went fine (up to middle of page 7 of pdf document.  That is: could create the new network, build AND train on the "hand entered" XOR training set.  Get similar results that seem to be documented correctly in the "SupervisedTrainingMonitorFrame" window.

    Beyond that, things pretty much go off the rails. 

    2.  Built a second network in the same project to test Backprop + momentum. It won't train.

    3.  Won't refresh and update "SupervisedTrainingMonitorFrame" window at all, so you don't really know what is going on.

    4.  Therefore set up new project for the new network and reconstructed training set.  Still won't train and "SupervisedTrainingMonitorFrame" window data won't update properly.

    5.  Considered that there might be a behavioural difference between hand-built test sets and those loaded from files.  Closed all open projects.  Set up new project to re-test Time Series example.  Set up NN, loaded test set.  It still won't train past iteration 1 (took screen shot to prove it to myself)..

    6.  Is there any useful information in the Help system.  Hard to tell, because in fact the help system, after a page or two, locks up - claims is is using 50% of 4 gig of machine memory.

    Broadly speaking then, from the perspective of a would-be research user … the thing doesn't work.

    I suppose that, if I wanted to learn something about java, this would be an interesting test case.  But I don't.
    Used to own and manage a software company.  Have no further interest in that.  Want to work on ANNs, not on fixing their code.  A shame, because, of the things I have looked at so far, Neuroph has some very compelling features - one of which is the interface … if it worked.

    Hope someone has a good idea.  Perhaps I'm just doing something horribly foolish.  But it shouldn't be this hard.

    best
    ay

     
  • Zoran Sevarac

    Zoran Sevarac - 2012-01-31

    Thanks for this detailed description, it is very valuable and helpfull for us, since we want Neuroph to make bugfree and easy to use for everyone. We'll try to help you, and if you give us additional feedback, there are bigger chances that we'll resolve the issues you're experiencing. Please also let us know the OS you're using and try some things again.
    We confirmed that examples Time Serises Tutorial have some issues, however some things you outlined shoud work:

    2. Built a second network in the same project to test Backprop + momentum. It won't train.

    This works for us. Please try again, you can also click to select desired training set again (altough that should not be neccesary step).

    3. Won't refresh and update "SupervisedTrainingMonitorFrame" window at all, so you don't really know what is going on.

    Probably related to previous issue, when we make that work, this should work also. Please try 2. again and let us know.

    4. Therefore set up new project for the new network and reconstructed training set. Still won't train and "SupervisedTrainingMonitorFrame" window data won't update properly.

    Most likely some hidden exception broke the application, just for testing purporse you may try to restart the app and try again the same thing.

    6. Is there any useful information in the Help system. Hard to tell, because in fact the help system, after a page or two, locks up - claims is is using 50% of 4 gig of machine memory.

    Maybe some memory leak in JavaHelp component, however we're not experiencing the issue.
    Are you sure that only help (not the whole app)  system is consuming that amount of memory, looks pretty strange.
    Maybe assigning more heap memory to app could help - see https://sourceforge.net/projects/neuroph/forums/forum/862858/topic/4795768

    Please give us a feedback, and hopefully we'll make this work. Although TimeSeriesTutorial has issues, other things shoud work…

    Zoran

     
  • Arthur Younger

    Arthur Younger - 2012-01-31

    Zoran,

    Thanks for the comments! 
    First, the OS is a fully updated Win XP on a dual core Dell with 1.75 Tbytes of drive space
    JDK 1.07; Netbeans 7.1

    Here are the results of testing your ideas today ….

    Curiously, now I can create multiple different networks in a project.
    They appear to train correctly always on the same data set (I do have to clilck to select it … but that is good, in fact).

    The Error Graph display seems to update correctly for different training runs.

    The “SupervisedTrainingMonitorFrame”  does NOT update itself after the first training run in a session.

    I have confirmed this up, down and sideways by repeatedly closing and restarting Nph Studio, testing any one of the 3 networks first, then some other one.  Every time, the graph updates but teh "Superviserd…" window stays identical to the initial training attempt (even though I close it each time to make sure it's not just a "display" effect).

    Regarding the "Help" problem, I agree that it must be a massive memory leak.  I have tested it twice while running Process Explorer.  For the application itself PE reports:
     
       Virtual Size: 420,128
       Working Set: 125,656

    There is almost no CPU load until Help is activated. 

    If you activate Help, after viewing a few pages, CPU load goes to 50%, the application freezes and can only be shutdown via “Kill Process Tree”

    You do not say what the problems might be in the Time Series Tutorial.  The only thing I could think of related to loading an external data file. 

    I may try another tutorial because, unfortunately, I too am puzzled by what is going on.

    best
    ay

     
  • Zoran Sevarac

    Zoran Sevarac - 2012-02-02

    You're welcome.  I'm glad that some of the things you mentioned now work for you, we'll try to provide you the solution for TimeSeries in a reasonable time. I think that training stops from some unknown reason, some strange conditin is met for specificly that data set, but we'll see.

    Regarding "SupervisedTrainingMonitorFrame" issue, my guss is messy thread synchronization at the moment, which doesnt  work well inder Java 7. Maybe you could try to run it with some Java 1.6 to make it work but, not sure. We'll deal with that one also.

    I'm aware of the fact that NeurophStudio needs polishing, and for the next release the focus will be exactly  that.

    Best,
    Zoran

     
  • Zoran Sevarac

    Zoran Sevarac - 2012-02-12

    Ok, I think I've figured out what both of us were doing wrong :)
    Just make sure you use select Tanh as transfer function in the wizard for neural network and it will work.
    The settings I used is MomentumBackpropagation with learning rate 0.2 and momentum 0.5
    Tanh must be used since training set has negative values, and Sigmoid is not able to produce them at output

    Zoran

     
  • Jaxx

    Jaxx - 2012-03-06

    Hi Zoran,

    I appear to be having similar issues as the previous user.
    Mainly, Im not sure if the training is working as I only get a few iterations (sometimes just a single dot on the error graph).
    I did run the AND Tutorial successfully, but there are too few inputs in the training file, so Im not sure that this is working.

    Do you perhaps have a sample project / load file that is confirmed to work - would like to use it to test ?
    Im also a research user, though a novice on at that.

     
  • Zoran Sevarac

    Zoran Sevarac - 2012-03-13

    Hi,

    It seems thats this is graph related issue, due the drawing and training threads sync. Thanks for notifying me, I'm working on this…
    If you try to train Perceptron for AND from Java code it works perfectly…

    Zoran

     

Log in to post a comment.

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

Sign up for the SourceForge newsletter:

JavaScript is required for this form.





No, thanks