demo.pl

一种新的SVM算法

#!/usr/bin/perl

##########################################################################
# This script demonstrates the syntax and usage of 
# ginisvm toolkit
##########################################################################

print "###################################################################\n";
print "# GiniSVM toolkit data format is in a column format where each     \n";
print "# entity is seperated by a newline character.                      \n";
print "# Each file (training/test) has a header file which contains the   \n";
print "# information of the data dimension, total number of classes and   \n";
print "# total number of data points present in the file.                 \n";
print "#                                                                  \n";
print "# A sample header for a 2 feature dimension, 9 classes and 1000    \n";
print "# data points look like this                                       \n";
print "# 2                                                                \n";
print "# 9                                                                \n";
print "# 1000                                                             \n";
print "#                                                                  \n";
print "# The training data is then arranged as                            \n";
print "# <label>                                                          \n";
print "# <C weight>                                                       \n";
print "# <features>                                                       \n";
print "#                                                                  \n";
print "# GiniSVM toolkit supports labels in two formats                   \n";
print "# Case 1)  If your aim is just classification: You know the        \n";
print "#          which class this feature belongs to. In this case the   \n";
print "#          label is simply a single scalar number representing the \n";
print "#          membership class                                        \n";
print "#                                                                  \n";
print "# Case 2) If you have the prior information of the probabilities   \n";
print "#         of different classes for this data and want to do a      \n";
print "#         probability regression then the label is a vector of     \n";
print "#         prior probabilities. For eg for a 2 class problem a      \n";
print "#         possible labels could be   0      0.9    0.2             \n";
print "#                                    1      0.1    0.8             \n";
print "#                                                                  \n";
print "# <C weight> is a optional parameter which specifies the weight    \n";
print "#            of the current data point. More noisy the point lesser\n";
print "#            value it gets. This information is used when ginisvm  \n";
print "#            is used in an EM training loop. If -cflag is present  \n";
print "#            then ginitrain assumes that <C weight> is not present \n";
print "#            This weight is multiplied by the global C specified   \n";
print "#            at the input argument to obtain effective C for each  \n";
print "#            data point                                            \n";
print "#                                                                  \n";
print "# <features> can be represented in a sparse or non-sparse format   \n";
print "# Case 1)    Non-sparse format, the features are in a column format\n";
print "#            with entries equal to the dimension value in the      \n";
print "#            header.                                               \n";
print "#                                                                  \n";
print "# Case 2)    Sparse format, the first entry is the total number of \n";
print "#            non-zero entries in this vector, and then index of    \n";
print "#            non-zero entries with its value. For a feature with   \n";
print "#            2 non zero entries for eg   2                         \n";
print "#                                        10                        \n";
print "#                                        0.32412                   \n";
print "#                                        21                        \n";
print "#                                        -0.1232                   \n";
print "###################################################################\n";
print "# TEST File Format                                                 \n";
print "###################################################################\n";
print "# Test file header is similar to training file but it may / maynot \n";
print "# contain <C weight> or <labels> whose presence is indicated by    \n";
print "# -cflag and -label training option. However the label format for  \n";
print "# test file is fixed and is a scale value                          \n";
print "###################################################################\n";
print "# Demo run with sample data (Press any key to continue )           \n";
print "###################################################################\n";

$inpkey =<STDIN>;
$TRAIN_MOD = "./ginitrain";
$TEST_MOD = "./ginitest";

print "###################################################################\n";
print "# Training                                                         \n";
print "###################################################################\n";
print "# 9 class, 2 dimension, 1000 data points, non-sparse format, gaussian \n";
print "# kernel with parameter 1.0, simple labels, no data C weight present, globa C = 1 \n";
print "###################################################################\n";
print "$TRAIN_MOD -k 1 -p1 1 -C 1 ../data/train_9c2d1000_nsp.dat ../data/train.cfg\n";
print "# (Press any key to continue )                                     \n";
print "###################################################################\n";

$inpkey =<STDIN>;
system("$TRAIN_MOD -k 1 -p1 1 -C 1 ../data/train_9c2d1000_nsp.dat ../data/train.cfg");

print "###################################################################\n";
print "# Testing with labels present                                      \n";
print "###################################################################\n";
print "$TEST_MOD -k 1 -label ../data/cv_9c2d1000_nsp.dat ../data/train.cfg ../data/train.out\n";
print "# (Press any key to continue )                                     \n";
print "###################################################################\n";

$inpkey =<STDIN>;
system("$TEST_MOD -k 1 -label ../data/cv_9c2d1000_nsp.dat ../data/train.cfg ../data/train.out");


print "###################################################################\n";
print "# Training                                                         \n";
print "###################################################################\n";
print "# 9 class, 2 dimension, 1000 data points, SPARSE format, gaussian \n";
print "# kernel with parameter 1.0, simple labels, no data C weight present, globa C = 1 \n";
print "###################################################################\n";
print "$TRAIN_MOD -k 1 -p1 1 -C 1 -sp ../data/train_9c2d1000_sp.dat ../data/train.cfg\n";
print "# (Press any key to continue )                                     \n";
print "###################################################################\n";

$inpkey =<STDIN>;
system("$TRAIN_MOD -k 1 -p1 1 -C 1 -sp ../data/train_9c2d1000_sp.dat ../data/train.cfg");

print "###################################################################\n";
print "# Testing                                                         \n";
print "###################################################################\n";
print "$TEST_MOD -k 1 -sp -label ../data/cv_9c2d1000_sp.dat ../data/train.cfg ../data/train.out\n";
print "# (Press any key to continue )                                     \n";
print "###################################################################\n";

$inpkey =<STDIN>;
system("$TEST_MOD -k 1 -sp -label ../data/cv_9c2d1000_sp.dat ../data/train.cfg ../data/train.out");

print "###################################################################\n";
print "# Training                                                         \n";
print "###################################################################\n";
print "# 9 class, 2 dimension, 1000 data points, non-sparse format, gaussian \n";
print "# kernel with parameter 1.0, VECTOR labels, no data C weight present, globa C = 1 \n";
print "###################################################################\n";
print "$TRAIN_MOD -k 1 -p1 1 -C 1 -pflag ../data/train_9c2d1000_nsp_pflag.dat ../data/train.cfg\n";
print "# (Press any key to continue )                                     \n";
print "###################################################################\n";

$inpkey =<STDIN>;
system("$TRAIN_MOD -k 1 -p1 1 -C 1 -pflag ../data/train_9c2d1000_nsp_pflag.dat ../data/train.cfg");

print "###################################################################\n";
print "# Testing                                                         \n";
print "###################################################################\n";
print "$TEST_MOD -k 1 -label ../data/cv_9c2d1000_nsp.dat ../data/train.cfg ../data/train.out\n";
print "# (Press any key to continue )                                     \n";
print "###################################################################\n";

$inpkey =<STDIN>;
system("$TEST_MOD -k 1 -label ../data/cv_9c2d1000_nsp.dat ../data/train.cfg ../data/train.out");

print "###################################################################\n";
print "# END DEMO !!                                                      \n";
print "###################################################################\n";