It seems that I am unable to download a version of Joone… in my Eclipse.
I installed a copy of the libraries & then I update the source from CVS but one of the library takes ages (if not centuries to download) : ./lib/ferra_oat.jar seems impossible to retrieve.
I will try again this night but is it not the first time I tried … hence :(
Tx for any suggestion
\T,
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
You must be working on the dev version. I added the oat library - check it out on sf. I never received a reply from the oat admin, so ended up building my own version of the oat library. This to integrate oat and joone, really to allow for the optimisation of recurrent nets using oat.
Why it takes so long I don't know, but you can ignore it safely and get all the vanilla functionality of joone going.
Regards
MG
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
It is the first time I have had a such issue …
Now with further attempts, my Mac mamanged to download it all … so it is probably somehting with the W2K3 I am using at the moment !
Personnally, I would have like the project to migrate under SVN so I could clean up the repository a bit … to have something like
Glad you got it working. There were prior attempts to clean up things….. I've given up and developed my own little library, which can do forward and recurrent networks and supports backprop (standard, rprop, qprop, irprop, jrprop) and for RNN it does RTRL, Levenberg-Marquardt, EKF and BFGS optimisations.
Regards
MG
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
BTW, oat did not help at all….. Neither did all those others. For all the effort, forward networks training using standard backprop still rules in the stuff I've tried it for….. I've tried so hard to get RNN to be productive but have yet to find a single instance…..
Regards
MG
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
What's the purpose of "oat" ?
May be you could throw a few lines of the problems (unused / not clean / … libraries).
We have to do some data cleansing and to suggest merge, I am considering NN or Lucene or … as long as I could embedded a piece of clean code … I would not matter.
\T,
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
OAT is a general optimisation library. Has many techniques to toy with. Joone did not really support RNN, so the idea with OAT was to plug it in and use it to train RNN networks. Never really worked. So then I did the RTRL, EKF etc stuff. Worked much much better from an optimisation point of view, but in practise still not really as it should.
Since I plug the stuff I do into various platforms, it can export a completed network as a single class without any dependencies that you can then use in anything out there. The class is autogenerated and gets big (in relation to the network) and I first ran into java's 'code too big' error when working on it, so now it slices the code into a whole lot of members…..
Joone has an interrogate function (as part of the JooneUtils class if I remember correctly) but it still has the overhead of the full Joone. This is where you should probably start looking.
If you can prepare input and output files, maybe I can give it a go and give you a single, optimised network to play with?
Regards
MG
If you would like to refer to this comment somewhere else in this project, copy and paste the following link:
It seems that I am unable to download a version of Joone… in my Eclipse.
I installed a copy of the libraries & then I update the source from CVS but one of the library takes ages (if not centuries to download) : ./lib/ferra_oat.jar seems impossible to retrieve.
I will try again this night but is it not the first time I tried … hence :(
Tx for any suggestion
\T,
Hi there,
You must be working on the dev version. I added the oat library - check it out on sf. I never received a reply from the oat admin, so ended up building my own version of the oat library. This to integrate oat and joone, really to allow for the optimisation of recurrent nets using oat.
Why it takes so long I don't know, but you can ignore it safely and get all the vanilla functionality of joone going.
Regards
MG
It is the first time I have had a such issue …
Now with further attempts, my Mac mamanged to download it all … so it is probably somehting with the W2K3 I am using at the moment !
Personnally, I would have like the project to migrate under SVN so I could clean up the repository a bit … to have something like
main/java/src
main/java/test
core/java/src
core/javatest
gui/java/src
gui/javatest
docs/img
docs/
also remove the hard coded path in the ANT file
\T,
Linux rules….
Glad you got it working. There were prior attempts to clean up things….. I've given up and developed my own little library, which can do forward and recurrent networks and supports backprop (standard, rprop, qprop, irprop, jrprop) and for RNN it does RTRL, Levenberg-Marquardt, EKF and BFGS optimisations.
Regards
MG
BTW, oat did not help at all….. Neither did all those others. For all the effort, forward networks training using standard backprop still rules in the stuff I've tried it for….. I've tried so hard to get RNN to be productive but have yet to find a single instance…..
Regards
MG
What's the purpose of "oat" ?
May be you could throw a few lines of the problems (unused / not clean / … libraries).
We have to do some data cleansing and to suggest merge, I am considering NN or Lucene or … as long as I could embedded a piece of clean code … I would not matter.
\T,
OAT is a general optimisation library. Has many techniques to toy with. Joone did not really support RNN, so the idea with OAT was to plug it in and use it to train RNN networks. Never really worked. So then I did the RTRL, EKF etc stuff. Worked much much better from an optimisation point of view, but in practise still not really as it should.
Since I plug the stuff I do into various platforms, it can export a completed network as a single class without any dependencies that you can then use in anything out there. The class is autogenerated and gets big (in relation to the network) and I first ran into java's 'code too big' error when working on it, so now it slices the code into a whole lot of members…..
Joone has an interrogate function (as part of the JooneUtils class if I remember correctly) but it still has the overhead of the full Joone. This is where you should probably start looking.
If you can prepare input and output files, maybe I can give it a go and give you a single, optimised network to play with?
Regards
MG