perl.ai

perl.ai http://www.nntp.perl.org/group/perl.ai/ ... Copyright 1998-2008 perl.org Wed, 08 Oct 2008 01:03:11 +0000 ask@perl.org Categorizer::Learner::SVM, scores of categories? by Jiuan-Ru Jennifer Lai Hi, I used the Categorizer::SVM library for large data classification (great tool); however, I'm having trouble analyzing the result from the SVM learner. - Categories have scores of either 0 or 1, with 1 being that this document belongs to this category, and 0 otherwise. Are there any scores representing probabilities or confidence level of belong to certain category other than these 0, 1 values? - Suppose this document could belong to 3 possible categories: cat1, cat2, and cat3. The best_category method simply picks the first category as the classification decision. If you call, $hypothesis->categories, the categories outputed don't seem to be in the order of probabilities or confidence level. They seem to be in the fixed order....and whatever listed first is favored. I hope someone can clear my confusion on the scores of categories in the SVM module. Thank you very much in advance, Jennifer http://www.nntp.perl.org/group/perl.ai/2008/03/msg576.html Tue, 18 Mar 2008 11:53:43 +0000 yawn cornerstone by Hester Newton Big News For SZSN! Shares Rocket! UP 37.5% Shandong Zhouyuan Seed and Nursery Co., Ltd (SZSN) $0.33 UP 37.5% SZSN new releases show huge expansion and Multi-Million dollar projects. Share prices rocket! Friday's trading was strong. Get On SZSN first thing Monday! Yes, it is that good. If they have work and family commitments, then perhaps it is easier to set aside a weekend to see a large number of bands in one go than go to regular gigs. He was detained by police in Middlesex on Saturday in connection with failing to attend a court hearing over alleged drugs offences in Glasgow. Closest to it in spirit is the Outsider Festival, held at Rothiemuchus, near Aviemore, with an emphasis on outdoor activities as well as music, and directly targeted at the older fan. With the likes of the Beastie Boys, Primal Scream and Bjork headlining, music is still to the fore, but more established groups dominate the line-up. Half Day FishingMaybe you guys can take it up on the northern Atlantic and Pacific coasts, but in the Southeast, a whole day of fishing in this unbearable summer heat can really wear a fellow out. She's very active in an Women help association, fighting for women rights, against excision. http://www.nntp.perl.org/group/perl.ai/2007/07/msg575.html Sun, 15 Jul 2007 10:04:53 +0000 Add documents to a learner? by Ignacio J. Ortega Lopera It's possible to add training to a learner? how? What i try is to reopen a state file, and add new documents to the training set without reading the entire corpus again.. It's seems that Algorithm::NativeBayes has a "purge" parameter that seems to help doing that, it permit add new instances, after a train.. Saludios, Ignacio J. Ortega http://www.nntp.perl.org/group/perl.ai/2007/06/msg574.html Thu, 07 Jun 2007 00:46:14 +0000 Re: Problems trying to predict by Ignacio J. Ortega Lopera Hola Ken: Many thanks, your advice, did the trick.. nad yes it was when reloading state from file.. 2007/6/7, Ken Williams <ken@mathforum.org>: > > Hi Ignacio, > > Is this when loading a pre-trained categorizer from a saved file? > This is a known problem, but I haven't settled on a good solution. > > A simple workaround is to just put: > > use Algorithm::NaiveBayes::Model::Frequency; > > in the script that's currently failing. > > -Ken > > > On May 30, 2007, at 1:47 PM, Ignacio J. Ortega Lopera wrote: > > > i'm getting this: > > > > Can't locate object method "predict" via package > > "Algorithm::NaiveBayes::Model:: > > Frequency" at > > /usr/lib/perl5/site_perl/5.8.0/AI/Categorizer/Learner/NaiveBayes.p > > m line 28. > > > > when trying to get hypoteses.. for a new doc.... > > > > anyone know if this is a silly one? > > http://www.nntp.perl.org/group/perl.ai/2007/06/msg573.html Thu, 07 Jun 2007 00:41:42 +0000 Re: Problems trying to predict by Ken Williams Hi Ignacio, Is this when loading a pre-trained categorizer from a saved file? This is a known problem, but I haven't settled on a good solution. A simple workaround is to just put: use Algorithm::NaiveBayes::Model::Frequency; in the script that's currently failing. -Ken On May 30, 2007, at 1:47 PM, Ignacio J. Ortega Lopera wrote: > i'm getting this: > > Can't locate object method "predict" via package > "Algorithm::NaiveBayes::Model:: > Frequency" at > /usr/lib/perl5/site_perl/5.8.0/AI/Categorizer/Learner/NaiveBayes.p > m line 28. > > when trying to get hypoteses.. for a new doc.... > > anyone know if this is a silly one? http://www.nntp.perl.org/group/perl.ai/2007/06/msg572.html Wed, 06 Jun 2007 23:41:20 +0000 AI::Categorizer and Umlauts? by Robert Barta Hi, I seem to have problems with umlauts, such as in words Präsentation When a document is added with return new AI::Categorizer::Document(name => $filename, content => $content); to the collection, after loading and finish, the feature vector contains only fragments of these words, such as pr => 1 sentation => 1 Setting the locale on the shell or in Perl does not have any effect use locale; not even with turning on de_AT explicitly. -- Aaaaaah, lib/AI/Categorizer/Document.pm is NOT using locale and use locale is very, uhm, local %-) Patching the file does not seem to break the test cases. \rho http://www.nntp.perl.org/group/perl.ai/2007/06/msg571.html Mon, 04 Jun 2007 19:25:40 +0000 AI::Categorizer suggestion for repackaging by Robert Barta Hi, This is probably more relevant to the maintainer of AI::Categorizer: It would be a bit simpler to debianize the package if the dependency to the Weka system would be factored out to a separate Perl package. Otherwise I have not found a problem in making it a Debian package. \rho http://www.nntp.perl.org/group/perl.ai/2007/06/msg570.html Mon, 04 Jun 2007 19:25:25 +0000 Problems trying to predict by Ignacio J. Ortega Lopera i'm getting this: Can't locate object method "predict" via package "Algorithm::NaiveBayes::Model:: Frequency" at /usr/lib/perl5/site_perl/5.8.0/AI/Categorizer/Learner/NaiveBayes.p m line 28. when trying to get hypoteses.. for a new doc.... anyone know if this is a silly one? Thanks in advance Saludos, Ignacio J. Ortega ---------------------------------------------------------------- Technical manager http://www.derecho.com/ http://www.nntp.perl.org/group/perl.ai/2007/05/msg569.html Wed, 30 May 2007 11:47:45 +0000 package AI::Categorizer::Collection::DBI; by Ignacio J. Ortega Lopera http://www.nntp.perl.org/group/perl.ai/2007/05/msg568.html Wed, 30 May 2007 09:49:39 +0000 how to use the function of "feature selection" under AI::Categorizer by jhoon Hello, I¡¯d like to select more important features using AI::Categorizer, and so modified demo.pl as follows === FROM === my $k = AI::Categorizer::KnowledgeSet->new( verbose => 1 ); === TO === my $k = AI::Categorizer::KnowledgeSet->new( verbose => 1, feature_selector => new AI::Categorizer::FeatureSelector::DocFrequency( ¡¡ verbose => 1, ¡¡ features_kept => 1000 ¡¡ ) ); === END === I observed the performance according to change the value of features_kept, but the performance is always same. I¡¯d appreciate it if you tell me how to do the feature selection using AI::Categorizer? Thank you very much in advance. Jae-Hoon. http://www.nntp.perl.org/group/perl.ai/2007/05/msg567.html Fri, 25 May 2007 04:09:13 +0000 Re: how to do feature selection by Alan Gibson im not sure if this pertains to your problem exactly, but you probably want to specify the weighting method like my $k = AI::Categorizer::KnowledgeSet->new( verbose => 1 , features_kept = 5000, tfidf_weighting=>'nfc' ); the default weighting is 'xxx' which if i understand correctly doesnt actually do anything. alan http://www.nntp.perl.org/group/perl.ai/2007/05/msg566.html Wed, 23 May 2007 19:24:41 +0000 how to do feature selection by Jianmin WU hi, buddies, I am not sure if i am in the right place. :-) I am a fresh man to the perl and perl AI module. I am trying to do the NaiveBayes experiments with the help of code demo.pl in example of the module of AI::Categorizer. Now I am confused about how to do the feature selection. The documents say that KnowledgeSet::load( ) will do feature selection and read the corpus at the same time. So, I change the construction of KnowledgeSet in demo.pl from my $k = AI::Categorizer::KnowledgeSet->new( verbose => 1 ); $k->load( collection => $training ) to my $k = AI::Categorizer::KnowledgeSet->new( verbose => 1 , features_kept = 5000 ); $k->load( collection => $training ) Then I re-run the code with expection to keep the top 5000 features with high Document Frequency. But it seems that there is no difference as before. do i misunderstand any point ? And also, is there any smoothing method implemented in AI::Categorizer::Learner::NaiveBayes ? Thanks for your attention Jianmin http://www.nntp.perl.org/group/perl.ai/2007/05/msg565.html Sat, 19 May 2007 05:43:12 +0000 Re: [ANNOUNCE] AI-Categorizer 0.08 -> CPAN by mentifex > Hi, > > After almost 4 years since the previous release, > I've uploaded a new AI::Categorizer to CPAN. > It's a minor set of changes with only a > couple bug fixes and additions: > [...] > -Ken Glad to hear the news of progress. Arthur -- http://mind.sourceforge.net/perl.html http://mind.sourceforge.net/Mind.html http://www.nntp.perl.org/group/perl.ai/2007/03/msg564.html Wed, 21 Mar 2007 04:34:39 +0000 [ANNOUNCE] AI-Categorizer 0.08 -> CPAN by Ken Williams Hi, After almost 4 years since the previous release, I've uploaded a new AI::Categorizer to CPAN. It's a minor set of changes with only a couple bug fixes and additions: 0.08 - Tue Mar 20 19:39:41 2007 - Added a ChiSquared feature selection class. [Francois Paradis] - Changed the web locations of the reuters-21578 corpus that eg/demo.pl uses, since the location it referenced previously has gone away. - The building & installing process now uses Module::Build rather than ExtUtils::MakeMaker. - When the features_kept mechanism was used to explicitly state the features to use, and the scan_first parameter was left as its default value, the features_kept mechanism would silently fail to do anything. This has now been fixed. [Spotted by Arnaud Gaudinat] - Recent versions of Weka have changed the name of the SVM class, so I've updated it in our test (t/03-weka.t) of the Weka wrapper too. [Sebastien Aperghis-Tramoni] -Ken http://www.nntp.perl.org/group/perl.ai/2007/03/msg563.html Wed, 21 Mar 2007 01:07:05 +0000 Probabilities with SVM by Alan Gibson im trying to use expectation-maximization to bootstrap an svm classifier. for this to work, the classifier needs to return better than random probabilities for its classification decisions. so not being one to repeat work, i thought i would see if anyone is setting on an implementation of AI::Categorizer::Learner::SVM that returns the probabilities produced by libsvm. any code or criticisms would be greatly appreciated. thanks, alan gibson http://www.nntp.perl.org/group/perl.ai/2007/03/msg562.html Sun, 11 Mar 2007 17:51:06 +0000 Re: text categorization with SVM and NaiveBayes by Ken Williams On Jan 8, 2007, at 10:51 AM, Tom Fawcett wrote: > Just to add a note here: Ken is correct -- both NB and SVMs are > known to be rather poor at providing accurate probabilities. Their > scores tend to be too extreme. Producing good probabilities from > these scores is called calibrating the classifier, and it's more > complex than just taking a root of the score. There are several > methods for calibrating scores. The good news is that there's an > effective one called isotonic regression (or Pool Adjacent > Violators) which is pretty easy and fast. The bad news is that > there's no plug-in (ie, CPAN-ready) perl implementation of it (I've > got a simple implementation which I should convert and contribute > someday). > > If you want to read about classifier calibration, google one of > these titles: > > "Transforming classifier scores into accurate multiclass > probability estimates" > by Bianca Zadrozny and Charles Elkan > > "Predicting Good Probabilities With Supervised Learning" > by A. Niculescu-Mizil and R. Caruana Cool, thanks for the references. It might be nice to add somesuch scheme to Algorithm::NaiveBayes (and friends), so that the user has a choice of several normalization schemes, including "none". If I get a surplus of tuits I'll add it, or if you feel like contributing your stuff that would be great too. -Ken http://www.nntp.perl.org/group/perl.ai/2007/01/msg561.html Tue, 09 Jan 2007 04:52:25 +0000 Re: text categorization with SVM and NaiveBayes by Tom Fawcett On Jan 7, 2007, at 9:23 PM, Ken Williams wrote: >> I would happily ignore all this and use NB, but it has one major >> flaw. >> "The winner takes it all", the first result returned is way too far >> (as in distance :)) from the others, which isn't exactly accurate if >> one cares of a balanced results pool. I don't know whether this is an >> implementation problem - I poked around the rescale() function in >> Util.pm with no real success - or a general algorithm problem. My >> goal >> is to have an implementation that can say: this text is 60% cat X, >> 20% >> cat Y, 18% cat Z and 2% other cats. Is this feasible ? If so, what >> approach would you recommend (which algorithm, which >> implementation or >> what path for implementing it ) ? > > Unfortunately, neither NB nor SVMs can really tell you that. SVMs > are purely discriminative, so all they can tell you is "I think > this new example is more like class A than class B in my training > data". There's no probability involved at all. That said, I > believe there has been some research into how to translate SVM > output scores into probabilities or confidence scores, but I'm not > really familiar with it. > > NB on the surface would seem to be a better option since it's > directly based on probabilities, but again the algorithm was > designed only to discriminate, so all those denominators that are > thrown away (the "P(words)" terms in the A::NB documentation) mean > that the notion of probabilities is lost. The rescale() function > is basically just a hack to return scores that are a little more > convenient to work with than the raw output of the algorithm. As > you've seen, it tends to be a little arrogant, greatly exaggerating > the score for the first category and giving tiny scores to the > rest. I'm sure there are better algorithms that could be used > there, but in many cases either one doesn't really care about the > actual scores, or one (*ahem*) does something ad hoc like taking > the square root of all the scores, or the fifth root, or whatever, > just to get some numbers that look better to end users. Just to add a note here: Ken is correct -- both NB and SVMs are known to be rather poor at providing accurate probabilities. Their scores tend to be too extreme. Producing good probabilities from these scores is called calibrating the classifier, and it's more complex than just taking a root of the score. There are several methods for calibrating scores. The good news is that there's an effective one called isotonic regression (or Pool Adjacent Violators) which is pretty easy and fast. The bad news is that there's no plug-in (ie, CPAN-ready) perl implementation of it (I've got a simple implementation which I should convert and contribute someday). If you want to read about classifier calibration, google one of these titles: "Transforming classifier scores into accurate multiclass probability estimates" by Bianca Zadrozny and Charles Elkan "Predicting Good Probabilities With Supervised Learning" by A. Niculescu-Mizil and R. Caruana Regards, -Tom http://www.nntp.perl.org/group/perl.ai/2007/01/msg560.html Mon, 08 Jan 2007 12:23:57 +0000 Re: text categorization with SVM and NaiveBayes by Ken Williams On Jan 5, 2007, at 7:10 AM, zgrim wrote: > So, back to my dilemmas. :) The results are puzzling, as many of the > research papers on the subject I've consulted say that SVM is > supposedly the best algorithm for this task. The radial kernel should > give the best results, for empirical-found values of gamma and C. This may be an issue with your corpus - I quite often find that when I don't have enough training data for the SVM to pick up on the "truth" patterns, or (somewhat equivalently) when there's a lot of noise in the data, a linear kernel will outperform a radial (RBF). I tend to think that's because the RBF is more expressive, and it's overfitting the noise in the training set. > Ignoring the fact that SVM is much, much slower to train than NB, it > still has worse accuracy. What am I doing wrong ? That may be an accident of your corpus too. Are you using cross- validation for these experiments? If so, you should be able to get some error bars to tell whether the difference is statistically significant or not. I'm guessing a 2% advantage may not be, in this case. > I would happily ignore all this and use NB, but it has one major flaw. > "The winner takes it all", the first result returned is way too far > (as in distance :)) from the others, which isn't exactly accurate if > one cares of a balanced results pool. I don't know whether this is an > implementation problem - I poked around the rescale() function in > Util.pm with no real success - or a general algorithm problem. My goal > is to have an implementation that can say: this text is 60% cat X, 20% > cat Y, 18% cat Z and 2% other cats. Is this feasible ? If so, what > approach would you recommend (which algorithm, which implementation or > what path for implementing it ) ? Unfortunately, neither NB nor SVMs can really tell you that. SVMs are purely discriminative, so all they can tell you is "I think this new example is more like class A than class B in my training data". There's no probability involved at all. That said, I believe there has been some research into how to translate SVM output scores into probabilities or confidence scores, but I'm not really familiar with it. NB on the surface would seem to be a better option since it's directly based on probabilities, but again the algorithm was designed only to discriminate, so all those denominators that are thrown away (the "P(words)" terms in the A::NB documentation) mean that the notion of probabilities is lost. The rescale() function is basically just a hack to return scores that are a little more convenient to work with than the raw output of the algorithm. As you've seen, it tends to be a little arrogant, greatly exaggerating the score for the first category and giving tiny scores to the rest. I'm sure there are better algorithms that could be used there, but in many cases either one doesn't really care about the actual scores, or one (*ahem*) does something ad hoc like taking the square root of all the scores, or the fifth root, or whatever, just to get some numbers that look better to end users. As for a better alternative, I'm not familiar with any that will be as accessible from a perl world, but you might want to look at some language modeling papers - I really like the LDA papers from Michael Jordan (no, not that Michael Jordan, this one: http:// citeseer.ist.psu.edu/541352.html), which are by no means straightforward, but they will indeed let you describe each document as generated by a mixture of categories. -Ken http://www.nntp.perl.org/group/perl.ai/2007/01/msg559.html Mon, 08 Jan 2007 04:20:12 +0000 text categorization with SVM and NaiveBayes by zgrim Hello, I am "playing" with the task of automated text categorization and inevitably hit a few dilemmas. I have tried different combinations of SVM and NaiveBayes, here are some results: - algorithm::svm (single world, through AI::Categorizer) ~ 92% accuracy (with the linear kernel, the radial one has bellow 10% with all sorts of values tried for gamma and c) - algorithm::svmlight ( nr. of categories worlds - each trained against the others ) ~ 62% in ranking mode - algorithm::naivebayes (one world, through AI::Categorizer) ~ 94% - algorithm::naivebayes (each against all other) ~ 73% These are on the same corpus ( which isn't perfect at all, but that a negligible information for now :) ). By accuracy I mean tested accuracy on a single category, which is, if the first category returned (highest score) is the supposed one, it's a hit, else, a miss. By single world I mean all categories build a single model, against tests are run. By multiple worlds (each against all other) I mean each category builds a model in which the tokens from that category are positive and the tokens from all other categories are negative. So, back to my dilemmas. :) The results are puzzling, as many of the research papers on the subject I've consulted say that SVM is supposedly the best algorithm for this task. The radial kernel should give the best results, for empirical-found values of gamma and C. Ignoring the fact that SVM is much, much slower to train than NB, it still has worse accuracy. What am I doing wrong ? I would happily ignore all this and use NB, but it has one major flaw. "The winner takes it all", the first result returned is way too far (as in distance :)) from the others, which isn't exactly accurate if one cares of a balanced results pool. I don't know whether this is an implementation problem - I poked around the rescale() function in Util.pm with no real success - or a general algorithm problem. My goal is to have an implementation that can say: this text is 60% cat X, 20% cat Y, 18% cat Z and 2% other cats. Is this feasible ? If so, what approach would you recommend (which algorithm, which implementation or what path for implementing it ) ? TIA -- perl -MLWP::Simple -e'print$_[rand(split(q|%%\n|, get(q=http://cpan.org/misc/japh=)))]' http://www.nntp.perl.org/group/perl.ai/2007/01/msg558.html Fri, 05 Jan 2007 05:10:40 +0000 Creating Collection of uncategorized data by Alan Gibson Hello, First post to this list. Im beginning a project that will use automated text classification to classify congressional bills and AI::Categorizer looks like the best framework to use. However, Im hitting a snag on what should be a simple operation. I train an svm classifier on 1000 documents; this operation goes fine. I then try to create an instance of AI::Categorizer::Collection::Files containing 5 unclassified documents. I supply only the path because the 5 documents are not yet categorized: my $c = new AI::Categorizer::Collection::Files( path => "$path"); while (my $document = $c->next) { my $hypothesis = $nb->categorize($document); print "Best assigned category: ", $hypothesis->best_category, "\n"; print "All assigned categories: ", join(', ', $hypothesis->categories), "\n"; } This produces the error No category information about '5-508' at /usr/local/share/perl/5.8.7/AI/Categorizer/Collection/Files.pm line 44. Mandatory parameter 'all_categories' missing in call to AI::Categorizer::Hypothesis->new() To get around this error I could just supply the categories of the 5 unknown test documents, but in our real world application we will have a constant stream of unclassified documents coming in that will recieve human attention only long after they have been automatically classified. Is the design intent to only allow test documents that already are categorized (eg for creating confidence statistics)? If so, does anyone have any suggestions on the preffered way to classifiy unknown documents with AI::Categorizer? Thanks, Alan http://www.nntp.perl.org/group/perl.ai/2007/01/msg557.html Thu, 04 Jan 2007 19:25:49 +0000 Re: ai::categorize samples by Russell Foltz-Smith Thanks, that's helpful. It's very difficult to find this corpus, especially in most of the existing documentation on AI::Categorizer. Thanks, Dr. Math. Russ Ken Williams wrote: > On Jan 2, 2007, at 8:53 PM, Russell Foltz-Smith wrote: > >> Does someone have an examples category text file that works with the >> demo.pl? > > Yup, you can download it from > http://campstaff.com/~ken/reuters-21578.tar.gz . > > >> Also, does anyone know of an online/web service implementation for web >> pages/urls to categorize them into dmoz categories? > > Not I. > > -Ken http://www.nntp.perl.org/group/perl.ai/2007/01/msg556.html Thu, 04 Jan 2007 06:39:30 +0000 Re: ai::categorize samples by Ken Williams On Jan 2, 2007, at 8:53 PM, Russell Foltz-Smith wrote: > Does someone have an examples category text file that works with the > demo.pl? Yup, you can download it from http://campstaff.com/~ken/ reuters-21578.tar.gz . > Also, does anyone know of an online/web service implementation for web > pages/urls to categorize them into dmoz categories? Not I. -Ken http://www.nntp.perl.org/group/perl.ai/2007/01/msg555.html Thu, 04 Jan 2007 04:16:05 +0000 ai::categorize samples by Russell Foltz-Smith Does someone have an examples category text file that works with the demo.pl? Also, does anyone know of an online/web service implementation for web pages/urls to categorize them into dmoz categories? Russ Smith http://www.nntp.perl.org/group/perl.ai/2007/01/msg554.html Wed, 03 Jan 2007 03:23:38 +0000 ai::categorize samples by Russell Foltz-Smith Does someone have an examples category text file that works with the demo.pl? Also, does anyone know of an online/web service implementation for web pages/urls to categorize them into dmoz categories? Russ Smith http://www.nntp.perl.org/group/perl.ai/2007/01/msg553.html Tue, 02 Jan 2007 18:53:20 +0000 Re: AI::Genetic by Gregg Allen Awesome! That helped a lot. I'm looking forward to your new and improved module. Thanks for creating that module, in the first place, also. When I first discovered the module about a year ago, I ran about a dozen random optimization problems from my graduate level operations research textbook from 25 years ago. I didn't find a single one it couldn't solve in less than a few minutes. (Mere seconds in most cases.) Thanks! Gregg Allen Cerebra, Inc. On Dec 8, 2006, at 8:56 AM, Ala Qumsieh wrote: > > --- Benjamin Tucker <ben@greenriver.org> wrote: > >> I don't actually have any experience with >> AI::Genetic, but >> Storable.pm is probably your best bet. Take a look >> at how >> AI::Categorizer interfaces with it: >> > http://search.cpan.org/src/KWILLIAMS/AI-Categorizer-0.07/lib/AI/ >> >> Categorizer/Storable.pm >> >> If you throw something like this into the bottom of >> one of your perl >> files, you should be able just to call >> $gen->store_state('filename') and then >> $gen->restore_state >> ('filename') (where $gen is an instance of >> AI::Genetic) > > [snip code] > > Thanks. That's an excellent suggestion. I'll add that > to AI::Genetic and upload a new version soon. > > Thanks, > --Ala > > > > > ______________________________________________________________________ > ______________ > Do you Yahoo!? > Everyone is raving about the all-new Yahoo! Mail beta. > http://new.mail.yahoo.com http://www.nntp.perl.org/group/perl.ai/2006/12/msg552.html Sat, 09 Dec 2006 03:16:17 +0000 Re: AI::Genetic by Ala Qumsieh --- Benjamin Tucker <ben@greenriver.org> wrote: > I don't actually have any experience with > AI::Genetic, but > Storable.pm is probably your best bet. Take a look > at how > AI::Categorizer interfaces with it: > http://search.cpan.org/src/KWILLIAMS/AI-Categorizer-0.07/lib/AI/ > > Categorizer/Storable.pm > > If you throw something like this into the bottom of > one of your perl > files, you should be able just to call > $gen->store_state('filename') and then > $gen->restore_state > ('filename') (where $gen is an instance of > AI::Genetic) [snip code] Thanks. That's an excellent suggestion. I'll add that to AI::Genetic and upload a new version soon. Thanks, --Ala ____________________________________________________________________________________ Do you Yahoo!? Everyone is raving about the all-new Yahoo! Mail beta. http://new.mail.yahoo.com http://www.nntp.perl.org/group/perl.ai/2006/12/msg551.html Fri, 08 Dec 2006 07:56:29 +0000 Re: AI::Genetic by Benjamin Tucker I don't actually have any experience with AI::Genetic, but Storable.pm is probably your best bet. Take a look at how AI::Categorizer interfaces with it: http://search.cpan.org/src/KWILLIAMS/AI-Categorizer-0.07/lib/AI/ Categorizer/Storable.pm If you throw something like this into the bottom of one of your perl files, you should be able just to call $gen->store_state('filename') and then $gen->restore_state ('filename') (where $gen is an instance of AI::Genetic) package AI::Genetic; use strict; use Storable; use File::Spec (); use File::Path (); sub save_state { my ($self, $path) = @_; if (-e $path) { File::Path::rmtree($path) or die "Couldn't overwrite $path: $!"; } mkdir($path, 0777) or die "Can't create $path: $!"; Storable::nstore($self, File::Spec->catfile($path, 'self')); } sub restore_state { my ($package, $path) = @_; return Storable::retrieve(File::Spec->catfile($path, 'self')); } 1; Ben On Dec 8, 2006, at 9:37 AM, Brad Larsen wrote: > One (possibly stupid) suggestion is to look at Data::Dumper. It > should work, but may be very slow if the object in question is > large. Let us know if you find anything better. > > Cheers, > Brad Larsen > > greggallen@gmail.com wrote: >> I know this is going to turn out to be a stupid question, but >> could someone tell me the easiest way to store and retrieve the >> state of the entire AI::Genetic colony, and parameters, to a disk >> file so it can be read in and out at will? >> I'm doing some constrained optimization experiments that can take >> several days, even a week, to run in the background, but I have a >> computer (Mac OS X 10.4.8) that is shared, and I need to install >> software and restart it almost daily. >> I would like to save the entire thing about every hour, but I can >> handle the timing part myself. >> Sincerely, >> Gregg Allen >> Cerebra, Inc. http://www.nntp.perl.org/group/perl.ai/2006/12/msg550.html Fri, 08 Dec 2006 07:13:17 +0000 Re: AI::Genetic by Brad Larsen One (possibly stupid) suggestion is to look at Data::Dumper. It should work, but may be very slow if the object in question is large. Let us know if you find anything better. Cheers, Brad Larsen greggallen@gmail.com wrote: > > I know this is going to turn out to be a stupid question, but could > someone tell me the easiest way to store and retrieve the state of the > entire AI::Genetic colony, and parameters, to a disk file so it can be > read in and out at will? > > I'm doing some constrained optimization experiments that can take > several days, even a week, to run in the background, but I have a > computer (Mac OS X 10.4.8) that is shared, and I need to install > software and restart it almost daily. > > I would like to save the entire thing about every hour, but I can > handle the timing part myself. > > > Sincerely, > > Gregg Allen > Cerebra, Inc. > > > http://www.nntp.perl.org/group/perl.ai/2006/12/msg549.html Fri, 08 Dec 2006 06:38:17 +0000 AI::Genetic by greggallen I know this is going to turn out to be a stupid question, but could someone tell me the easiest way to store and retrieve the state of the entire AI::Genetic colony, and parameters, to a disk file so it can be read in and out at will? I'm doing some constrained optimization experiments that can take several days, even a week, to run in the background, but I have a computer (Mac OS X 10.4.8) that is shared, and I need to install software and restart it almost daily. I would like to save the entire thing about every hour, but I can handle the timing part myself. Sincerely, Gregg Allen Cerebra, Inc. http://www.nntp.perl.org/group/perl.ai/2006/12/msg548.html Fri, 08 Dec 2006 00:10:45 +0000 Re: [ANNOUNCE] AI::FANN by Salvador "Fandiño" --- Ovid <publiustemp-perlai@yahoo.com> wrote: > I think these are a typos: > > FANN::AI->new_standard(@layer_sizes) > FANN::AI->new_sparse($connection_rate, @layer_sizes) > FANN::AI->new_shortcut(@layer_sizes) > FANN::AI->new_from_file($filename) oh, yes, thank you for pointing them. Cheers, - Salva __________________________________________________ Do You Yahoo!? Tired of spam? Yahoo! Mail has the best spam protection around http://mail.yahoo.com http://www.nntp.perl.org/group/perl.ai/2006/04/msg547.html Fri, 14 Apr 2006 10:51:21 +0000 Re: [ANNOUNCE] AI::FANN by Ovid --- Salvador Fandiï¿½o <sfandino@yahoo.com> wrote: > Hi, > > I have uploaded the new AI::FANN module to CPAN: > > http://search.cpan.org/~salva/AI-FANN/ Looks great (from the docs, haven't tested it). I think these are a typos: FANN::AI->new_standard(@layer_sizes) FANN::AI->new_sparse($connection_rate, @layer_sizes) FANN::AI->new_shortcut(@layer_sizes) FANN::AI->new_from_file($filename) Cheers, Ovid -- If this message is a response to a question on a mailing list, please send follow up questions to the list. Web Programming with Perl -- http://users.easystreet.com/ovid/cgi_course/ http://www.nntp.perl.org/group/perl.ai/2006/04/msg546.html Fri, 14 Apr 2006 09:26:47 +0000 [ANNOUNCE] AI::FANN by Salvador Fandiño Hi, I have uploaded the new AI::FANN module to CPAN: http://search.cpan.org/~salva/AI-FANN/ It is a wrapper for the Fast Artificial Neural Network library (http://fann.sf.net): Fast Artificial Neural Network Library is a free open source neural network library, which implements multilayer artificial neural networks in C with support for both fully connected and sparsely connected networks. Cross-platform execution in both fixed and floating point are supported. It includes a framework for easy handling of training data sets. It is easy to use, versatile, well documented, and fast. PHP, C++, .NET, Python, Delphi, Octave, Ruby, Pure Data and Mathematica bindings are available. A reference manual accompanies the library with examples and recommendations on how to use the library. A graphical user interface is also available for the library. This is an early release that may contain critical bugs, though most things seem to be working properly. The documentation focus on the differences with the C library, and both versions should be consulted in order to use the module. Training an ANN to emulate a XOR gate with AI::FANN looks like that: use AI::FANN qw(:all); # create an ANN with 2 inputs, a hidden layer with 3 neurons # and an output layer with 1 neuron: my $ann = AI::FANN->new_standard(2, 3, 1); $ann->hidden_activation_function(FANN_SIGMOID_SYMMETRIC); $ann->output_activation_function(FANN_SIGMOID_SYMMETRIC); # create the training data for a XOR operator: my $xor_train = AI::FANN::TrainData->new( [-1, -1], [-1], [-1, 1], [1], [1, -1], [1], [1, 1], [-1] ); $ann->train_on_data($xor_train, 500000, 1000, 0.001); $ann->save("xor.ann"); And using the trained ANN: use AI::FANN; my $ann = AI::FANN->new_from_file("xor.ann"); for my $a (-1, 1) { for my $b (-1, 1) { my $out = $ann->run([$a, $b]); printf "xor(%f, %f) = %f\n", $a, $b, $out->[0]; } } Comments and feedback are very welcome! Cheers, - Salva. http://www.nntp.perl.org/group/perl.ai/2006/04/msg545.html Fri, 14 Apr 2006 03:51:59 +0000 Author of Language::Prolog::Yaswi ? by Steffen Schwigon Hi perl-ai people! I'm trying to contact the author of Language::Prolog::Yaswi, Salvador "Fandiño" García, but didn't hear anything from him since september 2005. Does anyone know something about him? (Greeti+Tha)nX Steffen -- Steffen Schwigon <schwigon@webit.de> Dresden Perl Mongers <http://dresden-pm.org/> http://www.nntp.perl.org/group/perl.ai/2006/02/msg544.html Wed, 01 Feb 2006 07:30:12 +0000 New AI::NeuralNet::Simple - feedback welcome by Ovid Hi all, For those interested in neural nets, there's a new version of AI::NeuralNet::Simple out (http://search.cpan.org/dist/AI-NeuralNet-Simple/). This version incorporates a good-sized patch from Raphael Manfredi, the author of Storable. New features: Added tanh activation function. train_set() now accepts a maximum error rate target. Multiple network support. Persistence via storable. AI::NeuralNet::Simple is an easy to use "feed forward, back propogation neural network." Since the core of the module is written in C, it's very fast. The only significant (known) limitation of the module at this point is that the number of layers (3) is fixed. However, this is a very common number of layers for this type of network, so it's not too bad. The docs make loud warnings about the code being alpha and it claims to be a "simple learning module" but I think it's solid enough at this point that it might actually be useful. Cheers, Ovid -- If this message is a response to a question on a mailing list, please send follow up questions to the list. Web Programming with Perl -- http://users.easystreet.com/ovid/cgi_course/ http://www.nntp.perl.org/group/perl.ai/2006/01/msg543.html Mon, 09 Jan 2006 09:01:56 +0000 Yaswi with modules by Steffen Schwigon #! /usr/bin/perl use strict; use warnings; use Language::Prolog::Types::overload; use Language::Prolog::Yaswi qw(:query :load :context); use Language::Prolog::Sugar functors => { give_me_sth => 'give_me_sth' }, vars => [qw( Answer )] ; swi_use_modules ( "./a.swipl", "./b.swipl" ); sub yaswi_give_me_sth { local $swi_module = 'a'; ### or 'b' # Variante 1 swi_set_query( give_me_sth(2, Answer) ); my $answer = swi_var(Answer) if swi_next; swi_cut if swi_next; print "Answer: $answer \n"; } yaswi_give_me_sth(); http://www.nntp.perl.org/group/perl.ai/2006/01/msg542.html Mon, 09 Jan 2006 07:30:14 +0000 [Job Posting] Research Scientist, Eagan MN by Ken Williams Hi all, My company, Thomson Legal and Regulatory (the parent company for West Publishing, FindLaw, and other legal information services) is looking for a good Natural Language Processing person. Our R&D group is about 20 or so people, about 10 of whom are Research Scientists (including me). Since so much of our business is in text data, our entire group specializes in NLP. This position is certainly not perl-specific, but as researchers we can generally choose the tools that we want to use. Personally I tend to choose perl a lot, but as you'll note below, we also use Java, C, or whatever is appropriate to our tasks - sometimes prolog or python or smaller niche languages. Thus I thought it would be appropriate to post in this forum. This position has also been posted on various public job boards, including Monster.com and our company's career web site: http://jobsearch.monster.com/getjob.asp?JobID=37099311 http://www.thomsoncareercenter.com/search/view_job_xml.asp? src=rs&jobID=154545&loc=Ext Note that I am not the hiring manager or an HR person, I'm a fellow Research Scientist. Eagan is a suburb of Minneapolis/St. Paul. Thanks, -Ken *********************************************************************** Research Scientist – Opportunities at Thomson Legal and Regulatory The Research & Development department for Thomson Legal and Regulatory would like to invite qualified applicants to apply for an open Research Scientist position in their Eagan, MN offices. The Research & Development department performs applied research in natural language processing, document retrieval, information extraction, text classification, summarization, and named entity recognition. The ideal candidate would have significant expertise in one or more of these research areas. Preference will be given to candidates with research and work related experience in the area of named entity recognition and resolution. Principal duties: • Conducting applied research in information retrieval, information extraction, text categorization, text mining, or related areas in the context of large online delivery environments, such as Westlaw. • Execution of such projects, including • implementation of prototypes and the design of experiments to evaluate them • performing of experiments to validate key algorithms and architectures associated with such prototypes, followed by written reports • liaison with other departments concerning transition of prototypes into production • collaboration with software engineers engaged in the construction of prototypes. • Custom development for other TLR departments on key projects with a research component. • Monitoring of research literature through reading. Prerequisites: • Applicants should have a graduate degree in Computer Science or a related discipline. • Relevant experience in one or more of the following areas: natural language processing, machine learning, information retrieval, information extraction, document classification, summarization, and named entity extraction and resolution (Applicants with Masters degree must have some additional work related experience). • Substantial experience with UNIX or Windows environment. • Proficiency in a programming language such as Java or C++. • Good oral and written communication skills (as demonstrated through prior technical publications). • Relevant publications in Journals and (refereed) conferences a plus. Reports to: Director of Research. Contact: Khalid Al-Kofahi <Khalid.Al-Kofahi@thomson.com> http://www.nntp.perl.org/group/perl.ai/2005/12/msg541.html Thu, 15 Dec 2005 10:29:22 +0000 Re: Yaswi question by Steffen Schwigon Hi! Thanks for all your answers. And of course you are right, when you say that I just have to benchmark it. After struggling with some minor problems with Yaswi from mod_perl, I can now say that the performance impact is indeed quite small when using Language::Prolog::Yaswi. I compared - Language::Prolog::Yaswi from within a mod_perl application, - the same mod_perl framework without calls ti L::P::Yaswi, - the pure prolog httpd that is delivered with SWI Prolog. I used the threaded stress tester that comes with the swi prolog htttpd and measured time needed for 10000 requests from 8 threads). (Apache 1.3, mod_perl 1, Athlon 1800+) - mod_perl+Yaswi takes about 29sec. - plain mod_perl takes about 26sec. - httpd from swi takes about 20sec. As you can see, using Yaswi doesn't make my application much slower than it already is. GreetinX Steffen -- Steffen Schwigon <schwigon@webit.de> Dresden Perl Mongers <http://dresden-pm.org/> http://www.nntp.perl.org/group/perl.ai/2005/10/msg540.html Tue, 04 Oct 2005 08:41:44 +0000 Re: Yaswi question by Salvador "Fandiño" --- Steffen Schwigon <schwigon@webit.de> wrote: > Hi! > > I'm about to intermix a Perl web application with SWI-Prolog. > Currently Language::Prolog::Yaswi seems to be useful. > > The prolog part ought to solve only one particular problem, > everything else is a mod_perl driven web app. > > Now I'm not sure about its performance. I expect about 1 to 3 > requests > per second at peak times and I don't know yet how long my prolog > programm will take. > > Does Language::Prolog::Yaswi start a new "pl" process for every > query? No. It uses SWI-Prolog as a library, and the queries run inside the perl process. L::P::Y just converts data structures between perl and prolog. The overhead is relatively small, though it depends on how big the data structures passing through the interface are. Cheers, - Salvador __________________________________ Yahoo! Mail - PC Magazine Editors' Choice 2005 http://mail.yahoo.com http://www.nntp.perl.org/group/perl.ai/2005/09/msg539.html Thu, 29 Sep 2005 08:26:32 +0000 Re: Yaswi question by Ovid --- Steffen Schwigon <schwigon@webit.de> wrote: > Now I'm not sure about its performance. I expect about 1 to 3 > requests > per second at peak times and I don't know yet how long my prolog > programm will take. > > Does Language::Prolog::Yaswi start a new "pl" process for every > query? > The README talks about threads, so maybe it already does something > clever about this. The reason threads are mentioned is because Perl and SWI-Prolog must both have threads enabled or both NOT have threads enabled. Threads are not a requirement. I'm not aware that SWI-Prolog does start another process but I certainly wouldn't place a bet on that. I would strongly recommend Devel::Profiler or something similar on your code to find out where your true bottlenecks are. If you do find out your Prolog calls are slow, then you can spend the time tuning those queries (of course, if it's just too many Prolog queries in too short of a time, tuning might not help). On the other hand, if profiling can get you significant gains in the Perl portion, then you may get a win just by tuning Perl. If you can't speed up the Prolog, trying to replicate it's functionality in Perl may be difficult. Of course, there are tricks you can use to push much of the logic into a database and making multiple calls from Perl. That's probably slower but a good RDBMS has optimizations that Prolog often doesn't so it's worth exploring if you're stuck. If you really have problems with Prolog, I would suggest issuing a bunch of short Prolog queries and timing them and then issue some long-running queries. Time them both from Perl and directly in SWI-Prolog as if you were running it there. That can give you some idea as to the performance impact from Perl. Cheers, Ovid -- If this message is a response to a question on a mailing list, please send follow up questions to the list. Web Programming with Perl -- http://users.easystreet.com/ovid/cgi_course/ http://www.nntp.perl.org/group/perl.ai/2005/09/msg538.html Thu, 29 Sep 2005 08:15:47 +0000 RE: Yaswi question by BUDNEY, DANIEL L If you are worried about the Language::Prolog::Yaswi module being a bottleneck, the proper action is to make up a group of test requests and write a script that calls the L:P:Y module repeatedly for these cases. You can then time the number of "typical" requests that can be handled each minute (and you can even compare the results of a "short" request vs. a "long" request). Run each case 100 or 1000 times in a loop. For the most meaningful results, you should run the tests on the actual server you are using for the website. -----Original Message----- From: Steffen Schwigon [mailto:schwigon@webit.de] Sent: Thursday, September 29, 2005 4:02 AM To: perl-ai@perl.org Subject: Yaswi question Hi! I'm about to intermix a Perl web application with SWI-Prolog. Currently Language::Prolog::Yaswi seems to be useful. The prolog part ought to solve only one particular problem, everything else is a mod_perl driven web app. Now I'm not sure about its performance. I expect about 1 to 3 requests per second at peak times and I don't know yet how long my prolog programm will take. Does Language::Prolog::Yaswi start a new "pl" process for every query? The README talks about threads, so maybe it already does something clever about this. Or is there another recommended way to set up kind of an "swi prolog application server" (a process that always runs and answers queries, eg. via .*-RPC), that's accessible from Perl? (Greeti+Tha)nX Steffen -- Steffen Schwigon <schwigon@webit.de> Dresden Perl Mongers <http://dresden-pm.org/> http://www.nntp.perl.org/group/perl.ai/2005/09/msg537.html Thu, 29 Sep 2005 07:08:23 +0000