svmbr_20090226
所属分类:其他
开发工具:C++
文件大小:870KB
下载次数:3
上传日期:2013-02-27 00:00:37
上 传 者:
mba123
说明: SVn training program
文件列表:
svmbr (0, 2009-02-27)
svmbr\uml (0, 2009-02-27)
svmbr\test (0, 2009-02-27)
svmbr\src (0, 2009-02-27)
svmbr\Release (0, 2009-02-27)
svmbr\docs (0, 2009-02-27)
svmbr\.svn (0, 2009-02-27)
svmbr\uml\.svn (0, 2009-02-27)
svmbr\test\.svn (0, 2009-02-27)
svmbr\src\.svn (0, 2009-02-27)
svmbr\Release\.svn (0, 2009-02-27)
svmbr\docs\.svn (0, 2009-02-27)
svmbr\.svn\tmp (0, 2009-02-27)
svmbr\.svn\props (0, 2009-02-27)
svmbr\.svn\prop-base (0, 2009-02-27)
svmbr\.svn\text-base (0, 2009-02-27)
svmbr\uml\.svn\tmp (0, 2009-02-27)
svmbr\uml\.svn\props (0, 2009-02-27)
svmbr\uml\.svn\prop-base (0, 2009-02-27)
svmbr\uml\.svn\text-base (0, 2009-02-27)
svmbr\test\.svn\tmp (0, 2009-02-27)
svmbr\test\.svn\props (0, 2009-02-27)
svmbr\test\.svn\prop-base (0, 2009-02-27)
svmbr\test\.svn\text-base (0, 2009-02-27)
svmbr\src\.svn\tmp (0, 2009-02-27)
svmbr\src\.svn\props (0, 2009-02-27)
svmbr\src\.svn\prop-base (0, 2009-02-27)
svmbr\src\.svn\text-base (0, 2009-02-27)
svmbr\Release\.svn\tmp (0, 2009-02-27)
svmbr\Release\.svn\props (0, 2009-02-27)
svmbr\Release\.svn\prop-base (0, 2009-02-27)
svmbr\Release\.svn\text-base (0, 2009-02-27)
svmbr\docs\.svn\tmp (0, 2009-02-27)
svmbr\docs\.svn\props (0, 2009-02-27)
svmbr\docs\.svn\prop-base (0, 2009-02-27)
svmbr\docs\.svn\text-base (0, 2009-02-27)
svmbr\.svn\tmp\props (0, 2009-02-27)
svmbr\.svn\tmp\prop-base (0, 2009-02-27)
svmbr\.svn\tmp\text-base (0, 2009-02-27)
svmbr\uml\.svn\tmp\props (0, 2009-02-27)
... ...
+----------------------------------------------+
| SVMBR - A program for training SVMs |
+----------------------------------------------+
| * Creator, modeling, coding: |
| Marcelo Barros de Almeida |
| barros@smar.com.br |
| http://litc.cpdee.ufmg.br/~barros/ |
| * Coding, improvements, bug fixes: |
| Bernardo Penna |
| bprc@brfree.com.br |
+----------------------------------------------+
Copyright(c) 2002 by Marcelo Barros de Almeida
All rights reserved
------------------------------------------------
Install notes
------------------------------------------------
Download the SMOBR and unpack it in a new directory
using gunzip/tar or winzip.
Windows users
SMOBR can be compiled using DJGPP (free GCC based compiler)
or Visual C++. For DJGPP there exist an interface called
RHIDE (like the old Borland C for DOS). I renamed libstdcxx
to libstdc++ to compile in DJGPP. Use Makefile provided with
sources.
For Visual C++ there is a project ready to use.
Checks if there is a definition for _MSC_VER
and WIN32 (or define it yourself with #define _MSC_VER).
Unix/Linux users
make all
;-)
------------------------------------------------
Using SVMBR
------------------------------------------------
Usage:(1 of 2)
svmbr -option1 value1 -option2 value2 ...
Options (type in lower case !):
KERNELS:
=======
-k [poly|rbf|perceptron|linear] | Chooses the SVM kernel.
| Kernel expressions:
| rbf : exp(-|x1-x2|^2/(2*p3*p1^2))
| poly : ((x1*x2+p2)/p3)^p1
| perceptron: tanh(p1*x1*x2/p3+p2)
| linear : (x1*x2+p2)/p3
|
-p1 value | Kernel parameters.
-p2 value | For RBF, p1 is the variance and p2
-p3 value | is not used. For polynomial p1 is the
| degree, p2 is bias. Perceptron has p1 as
| slope and p2 as offset. Linear uses p2 as
| bias. Scale parameter is provided for
| all kernels by p3 parameter.
| Default: p1=1, p2=0, p3=1.
|
FILES:
=====
-tp filename | Training patterns.
| First two numbers specifies the amount of
| training vectors and the input vector,
| dimension, respectively. After these two
| values, other points are considered as
| data. Put one data vector per row,
| followed by its target. Numbers must be
| separated by spaces. For instance, dataset
| for XOR problem is stated as (please,
| remove comments):
| ----------------------------------------
| 4 // number of training vector
| 2 // input dimension
| 0 0 -1 // vector 1 + target
| 0 1 1 // vector 2 + target
| 1 0 1 // vector 3 + target
| 1 1 -1 // vector 4 + target
| ----------------------------------------
|
-gp filename | Testing patterns (generalization).
| First two numbers specifies the amount of
| testing vectors and the input vector,
| dimension, respectively. After these two
| values, other points are considered as
| data. Put one data vector per row,
| followed by its target. Numbers must be
| separated by spaces. For instance, dataset
| for XOR problem is stated as (please,
| remove comments):
| ----------------------------------------
| 4 // number of training vector
| 2 // input dimension
| 0 0 -1 // vector 1 + target
| 0 1 1 // vector 2 + target
| 1 0 1 // vector 3 + target
| 1 1 -1 // vector 4 + target
| ----------------------------------------
|
-go filename | Testing output file.
| Generalization results are saved on this
| file, like SVM's output and MSE.
|
-sv filename | SVM configuration.
| This file describes the SVM structure. It
| is the output file from SVMBR.
|
-d [normal|sparse|binary] | Data representation.
| This parameter affects how the vectors are
| treated by the program. Set sparse when
| using training vector built mainly by
| zeros and binary when input vectors have
| only 0 and 1 values.
| Default: d=normal
|
OPTIMIZATION:
============
-solver [edr|smo|boosting] | Method used to find the solution.
| Default: solver = smo
|
-c value | Upper limit for Lagrange multipliers.
| Default: c=1
|
-t value | Tolerance for KKT condition.
| Default: t=0.001
|
-e value | Epsilon for bound precision.
| Default: e=0.001
|
-edr value | Error dependent repetition.
| Number of scans performed on the training
| set.
| Default: edr = -1 (off)
|
-power value | Power for comparison function in EDR.
| (err_i)^power > j*err_max/edr.
| Default: power = 1
|
-maxiter value | Maximum number of iterations of SMO when
| using together with EDR.
| Default: maxiter = -1
|
-chunk value | Chunking size.
| Number of vectors in chunk.
| Default: chunk = -1 (not using chunk).
|
OTHERS:
======
-l [0|1] | Hard (1) or soft limit (0) when
| evaluating output.
| Default: l=0
|
-version | Print the current version.
|
-h | Print this help message.
|
3) Usage:(2 of 2)
svmbr file_with_all_parameters
Options (type in upper case !):
[GLOBAL]
NUMCLASSES=value | Number of classes.
| Default: 2
|
DATAREPRESENTATION= | Data representation.
[normal|sparse|binary] | This parameter affects how the vectors are
| treated by the program. Set sparse when
| using training vector built mainly by
| zeros and binary when input vectors have
| only 0 and 1 values.
| Default: d=normal
|
SOLUTIONMETHOD=[edr|smo|boosting]| Method used to find the solution.
| Default: smo
|
CHUNKSIZE=value | Chunking size.
| Number of vectors in chunk.
| Default: chunk = -1 (not using chunk).
|
TRAININGPATTERNFILE=filename | Training patterns.
| First two numbers specifies the amount of
| training vectors and the input vector,
| dimension, respectively. After these two
| values, other points are considered as
| data. Put one data vector per row,
| followed by its target. Numbers must be
| separated by spaces. For instance, dataset
| for XOR problem is stated as (please,
| remove comments):
| ----------------------------------------
| 4 // number of training vector
| 2 // input dimension
| 0 0 -1 // vector 1 + target
| 0 1 1 // vector 2 + target
| 1 0 1 // vector 3 + target
| 1 1 -1 // vector 4 + target
| ----------------------------------------
| For multi class problems, the targets are
| represented by 0,1,2,3... For instance:
| 6 // number of training vector
| 2 // input dimension
| 0 7 0 // vector 1 + target
| 2 1 0 // vector 2 + target
| 4 4 1 // vector 3 + target
| 3 1 2 // vector 4 + target
| 2 4 0 // vector 5 + target
| 7 6 1 // vector 6 + target
|
TESTINGPATTERNFILE=filename | Testing patterns (generalization).
| First two numbers specifies the amount of
| testing vectors and the input vector,
| dimension, respectively. After these two
| values, other points are considered as
| data. Put one data vector per row,
| followed by its target. Numbers must be
| separated by spaces. For instance, dataset
| for XOR problem is stated as (please,
| remove comments):
| ----------------------------------------
| 4 // number of training vector
| 2 // input dimension
| 0 0 -1 // vector 1 + target
| 0 1 1 // vector 2 + target
| 1 0 1 // vector 3 + target
| 1 1 -1 // vector 4 + target
| ----------------------------------------
| For multi class problems, the targets are
| represented by 0,1,2,3... For instance:
| 6 // number of training vector
| 2 // input dimension
| 0 7 0 // vector 1 + target
| 2 1 0 // vector 2 + target
| 4 4 1 // vector 3 + target
| 3 1 2 // vector 4 + target
| 2 4 0 // vector 5 + target
| 7 6 1 // vector 6 + target
|
TESTINGOUTFILE=filename | Testing output file.
| Generalization results are saved on this
| file, like SVM's output and MSE.
|
[GLOBAL] or [CLASS:x]
(x is the number of the class: 0,1,2,3...)
SVMFILE=filename | SVM configuration.
| This file describes the SVM structure. It
| is the output file from SVMBR
| Only for 2 classes it is global, if not
| it is not global, it has to exist in each
| class parameters.
[CLASS:x]
(x is the number of the class: 0,1,2,3...)
KERNELTYPE= | Chooses the SVM kernel.
[poly|rbf|perceptron|linear] | Kernel expressions:
| rbf : exp(-|x1-x2|^2/(2*p3*p1^2))
| poly : ((x1*x2+p2)/p3)^p1
| perceptron: tanh(p1*x1*x2/p3+p2)
| linear : (x1*x2+p2)/p3
|
P1=value | Kernel parameters.
P2=value | For RBF, p1 is the variance and p2
P3=value | is not used. For polynomial p1 is the
| degree, p2 is bias. Perceptron has p1 as
| slope and p2 as offset. Linear uses p2 as
| bias. Scale parameter is provided for
| all kernels by p3 parameter.
| Default: P1=1, P2=0, P3=1.
|
|
C=value | Upper limit for Lagrange multipliers.
| Default: 1
... ...
近期下载者:
相关文件:
收藏者: