TinySVM: Support Vector Machines

Last update: $Date: 2001/01/16 22:25:52 $

TinySVM

TinySVM is an implementation of Support Vector Machines (SVMs) [Vapnik 95], [Vapnik 98] for the problem of pattern recognition. Support Vector Machines is a new generation learning algorithms based on recent advances in statistical learning theory, and applied to large number of real-world applications, such as text categorization, hand-written character recognition.

List of Contents

• What's new
• Features
• libtinysvm Reference Manual
• Using included tools
• Download
  • Sources
  • RPM package for RedHat Linux
  • Binary package for MS-Windows
  • CVS
• Install
• Language Bindings
  • Perl
  • Ruby
  • Python
• Questions and Bug Reports
• TODO
• References
• Related Link

What's new

• 2001-01-17: TinySVM 0.0.2 Released
  
  • Supports Support Vector Regression
  • Ruby module is available

Features

• Fast optimization algorithms stemmed from SVM_light,[Joachims/99a].
  • Working set selection based on steepest feasible descent.
  • "Shrinking" (an effective heuristics to reduce working sets dynamically)[Joachims/99a]
  • Uses sparse vector representation
  • Can handle several ten-thousands of training examples, and hundred-thousands of feature dimension.
  • Optimization for handling binary features, twice faster than SVM_light.
• Written in C++ with OO style. Provides useful class libraries.
• Provides Perl/Ruby module

Using the included Tools

Format of training data

TinySVM accepts the same representation of training data as SVM_light uses. This format has potential to handle large sparse feature vectors.

(BNF-like representation)
<class> .=. +1 | -1
<feature> .=. integer (>=1)
<value> .=. real
<line> .=. <class> <feature>:<value><feature>:<value> ... <feature>:<value>


Example (SVM)
+1 201:1.2 3148:1.8 3983:1 4882:1
-1 874:0.3 3652:1.1 3963:1 6179:1
+1 1168:1.2 3318:1.2 3938:1.8 4481:1
+1 350:1 3082:1.5 3965:1 6122:0.2
-1 99:1 3057:1 3957:1 5838:0.3

See tests/train.svmdata and tests/test.svmdata in the package.

The case of SVR, you can give real value as class label.

Example (SVR)
0.23 201:1.2 3148:1.8 3983:1 4882:1
0.33 874:0.3 3652:1.1 3963:1 6179:1
-0.12 1168:1.2 3318:1.2 3938:1.8 4481:1

See tests/train.svrdata in the package

svm_learn (simple learner)

"svm_learn" accepts two arguments --- file name of training data stored in the above formats and model file in which the SVs and their weights (alpha) are stored after training.
Try --help option for finding out other options.

% svm_learn -t 1 -d 2 -c 1 train.svmdata model
TinySVM - tiny SVM package
Copyright (C) 2000 Taku Kudoh All rights reserved.

  1000 examples, cache size: 1524
....................   1000   1000 0.0165 1286/ 64.3%   1286/ 64.3%
............
Checking optimality of inactive variables  re-activated: 0

Done! 1627 iterations

Number of SVs (BSVs)            719 (4)
Empirical Risk:                 0.002 (2/1000)
L1 Loss:                        4.22975
CPU Time:                       0:00:01

% ls -al model
  -rw-r--r--    1 taku-ku  is-stude    26455 Dec  7 13:40

TinySVM prints the following messages during the learning iterations,

....................   1000  15865  2412 1.6001  33.2%  33.2%
....................   2000  15864  2412 1.3847  39.5%  36.4%

• 1st column: One "." means 50 iterations.
• 2nd column: Represents the total iterations processed
• 3rd column: Represents the size of current working set. It will become smaller during the shrinking process.
• 4th column: Represents the current cache size.
• 5th column: Max KKT violation value. If this value reaches termination-criterion (default 0.001), it means that the 1 st stage of the learning process has completed. After the 1st stage, re-activates all shrinked examples for final check. You can omit final check using -f option.
• 7th column: Cache hit rate during last 1000 iterations.
• 7th column: Cache hit rate during all iterations.

svm_classify (simple classifier)

"svm_classify" accepts two arguments --- file name of test data stored in above formats and model file produced by svm_learn program. "svm_classify" simply outputs the accuracy for given test data. You can also employ interactive classification by giving "-" as file name of test example.
Try --help option for finding out other options.

% svm_classify test.svmdata model
Accuracy:   77.80000% (389/500)
Precision:  66.82927% (137/205)
Recall:     76.11111% (137/180)
System/Answer p/p p/n n/p n/n: 137 68 43 252

% svm_classify -V test.svmdata model
-1 -1.04404
-1 -1.26626
-1 -0.545195
.. snip
Accuracy:   77.80000% (389/500)

svm_model (model browser)

"svm_model" shows the estimated margin, VC dimension and number of SVs of given some model file.
Try --help option for finding out other options.

% svm_model model
File Name:                      model
Margin:                         0.181666
Number of SVs:                  719
Number of BSVs:                 4
Size of training data:          1000
L1 Loss (Empirical Risk):       4.16917
Estimated VC dimension:         728.219
Estimated xi-alpha(2):          573

Download

TinySVM is free software distributed under the GNU Lesser General Public License.

Source

• TinySVM-0.02.tar.gz: HTTP

Binary/Source package for RedHat Linux

• RedHat 6.x i386: HTTP
• RedHat 6.x SRPMS: HTTP
• RedHat 7.x i386: HTTP
• RedHat 7.x SRPMS: HTTP

Binary package for MS-Windows

• HTTP

CVS

Development of TinySVM uses CVS. So latest developing version is available at CVS.
We are willing to welcome you to join CVS based development.

% cvs -d :pserver:anonymous@chasen.aist-nara.ac.jp:/cvsroot login
CVS password:   # Just hit return/enter.
% cvs -d :pserver:anonymous@chasen.aist-nara.ac.jp:/cvsroot co TinySVM

Install

 % ./configure 
 % make
 % make check
 % su
 # make install

You can change default install path by using --prefix option of configure script.
Try --help option for finding out other options.

Language Bindings

• Perl: see perl/README file.
• Ruby: see ruby/README file
• Python: Under construction.
• Java: Under construction.

Questions and Bug Reports

If you find bugs or you have any questions please contact me via email taku-ku@is.aist-nara.ac.jp.
(Japanese mail is also (more) acceptable.)

TODO

• MultiClass SVM (one-for-all-others, pairwise)
• nu-SVM and nu-SVR [Schölkopf1998]
• Transductive SVM [Vapnik 98],[Joachims 99]
• Span SVM [Vapnik 2000]
• Ordinal SVR [Herbrich 2000]
• Provide some Wrapper class to handle STRING features.
• Provide DLL (Dynamic Link Libraries) for Windows environment.
• Provide an API for user customizable kernel function.
  

References

• [Joachims 99a] T. Joachims, 11 in: Making large-Scale SVM Learning Practical. Advances in Kernel Methods pp.169-
• [Vapnik 95] Vladimir N. Vapnik, The Nature of Statistical Learning Theory. Springer, 1998.
• [Vapnik 98] Vladimir N. Vapnik, The Statisitcal Learning Theory. Springer, 1998.
• [Joachims 99c] T. Joachims, Transductive Inference for Text Classification using Support Vector Machines. International Conference on Machine Learning (ICML), 1999.
• [Vapnik 2000] Vladimir N. Vapnik, 14 in: Bounds on Error Expectation for SVM. in Advances in Large Margin Classifiers. pp. 261-
• [Herbrich 2000] Ralf Herbrich et al, 7 in: Large Margin Rank Boundaries for Ordinal Regression. in Advances in Large Margin Classifiers. pp. 116-
• [Schölkopf 1998] B. Schölkopf, A. Smola, R. Williamson, and P. L. Bartlett. New support vector algorithms. NeuroCOLT Technical Report NC-TR-98-031, Royal Holloway College, University of London, UK, 1998. To appear in Neural Computation.

Related Links

• SVM_light Yet another SVM package.
• libsvmYet another Fast and useful SVM package
• Kernel Machine Website
• SVM Application List
• Support Vectors and Kernel Methods
• An introduction to Suppor Vector Machines

$Id: index.html,v 1.10 2001/01/16 22:25:52 taku-ku Exp $;