关于SVM参数c&g选取的总结帖[matlab-libsvm]

写了个程序
来选取SVM中参数
c和g的最佳值.

[写这个的目的是方便大家用这个小程序直接来寻找


c和g的最佳值,不用再另外编写东西了.

]


其实原本libsvm
C语言版本中有相应的子程序可以找到最佳的c和g,需装载python语言然后用py 那个画图
就可以找到最佳的c和g,我写了个matlab版本的.算是弥补了libsvm在matlab版本下的空缺.

测试数据
还是我视频
里的wine data.

寻找最佳c和g的思想仍然是让c和g在一定的范围里跑(比如
c = 2^(-5),2^(-4),...,2^(5),g = 2^(-5),2^(-4),...,2^(5)),然后用cross
validation的想法找到是的准确率最高的c和g,在这里我做了一点修改(纯粹是个人的一点小经验和想法),我改进的是:
因为会有不同的c和g都对应最高的的准确率,我把具有最小c的那组c和g认为是最佳的c和g,因为惩罚参数不能设置
太高,很高的惩罚参数能使得validation数据的准确率提高,但过高的惩罚参数c会造成过学习状态,反正从我用SVM到现在,往往都是惩罚参数c过高会导致最终测试集合的准确率并不是很理想..


在使用这个程序时也有小技巧,可以先大范围粗糙的找 比较理想的c和g,然后再细范围找更加理想的c和g.

比如首先让 c = 2^(-5),2^(-4),...,2^(5),g = 2^(-5),2^(-4),...,2^(5)在这个范围找比较理想的c和g,如图:

======

======

此时bestc = 0.5,bestg=1,bestacc = 98.8764[cross validation 的准确率]

最终测试集合的准确率 Accuracy = 96.6292% (86/89) (classification)

======

此时看到可以把c和g的范围缩小.还有步进的大小也可以缩小(程序里都有参数可以自己调节,也有默认值可不调节).

让 c = 2^(-2),2^(-1.5),...,2^(4),g = 2^(-4),2^(-3.5),...,2^(4)在这个范围找比较理想的c和g,如图:

=============

===============

此时bestc = 0.3536,bestg=0.7017,bestacc = 98.8764[cross validation 的准确率]

最终测试集合的准确率 Accuracy = 96.6292% (86/89) (classification)

===================

上面第二个的测试的代码

:

复制内容到剪贴板
代码:

load wine_SVM;

train_wine = [wine(1:30,:);wine(60:95,:);wine(131:153,:)];

train_wine_labels = [wine_labels(1:30);wine_labels(60:95);wine_labels(131:153)];

test_wine = [wine(31:59,:);wine(96:130,:);wine(154:178,:)];

test_wine_labels = [wine_labels(31:59);wine_labels(96:130);wine_labels(154:178)];

[train_wine,pstrain] = mapminmax(train_wine');

pstrain.ymin = 0;

pstrain.ymax = 1;

[train_wine,pstrain] = mapminmax(train_wine,pstrain);

[test_wine,pstest] = mapminmax(test_wine');

pstest.ymin = 0;

pstest.ymax = 1;

[test_wine,pstest] = mapminmax(test_wine,pstest);

train_wine = train_wine';

test_wine = test_wine';

[bestacc,bestc,bestg] = SVMcg(train_wine_labels,train_wine,-2,4,-4,4,3,0.5,0.5,0.9);

cmd = ['-c ',num2str(bestc),' -g ',num2str(bestg)];

model = svmtrain(train_wine_labels,train_wine,cmd);

[pre,acc] = svmpredict(test_wine_labels,test_wine,model);

============我写的那个选取SVM中参数c和g的最佳值.的程序的代码 SVMcg.m====================

复制内容到剪贴板
代码:

function [bestacc,bestc,bestg] = SVMcg(train_label,train,cmin,cmax,gmin,gmax,v,cstep,gstep,accstep)

%SVMcg cross validation by faruto

%Email:farutoliyang@gmail.com QQ:516667408 http://blog.sina.com.cn/faruto BNU

%last modified 2009.8.23

%Super Moderator @ www.ilovematlab.cn

%% about the parameters of SVMcg

if nargin < 10

accstep = 1.5;

end

if nargin < 8

accstep = 1.5;

cstep = 1;

gstep = 1;

end

if nargin < 7

accstep = 1.5;

v = 3;

cstep = 1;

gstep = 1;

end

if nargin < 6

accstep = 1.5;

v = 3;

cstep = 1;

gstep = 1;

gmax = 5;

end

if nargin < 5

accstep = 1.5;

v = 3;

cstep = 1;

gstep = 1;

gmax = 5;

gmin = -5;

end

if nargin < 4

accstep = 1.5;

v = 3;

cstep = 1;

gstep = 1;

gmax = 5;

gmin = -5;

cmax = 5;

end

if nargin < 3

accstep = 1.5;

v = 3;

cstep = 1;

gstep = 1;

gmax = 5;

gmin = -5;

cmax = 5;

cmin = -5;

end

%% X:c Y:g cg:acc

[X,Y] = meshgrid(cmin:cstep:cmax,gmin:gstep:gmax);

[m,n] = size(X);

cg = zeros(m,n);

%% record acc with different c & g,and find the bestacc with the smallest c

bestc = 0;

bestg = 0;

bestacc = 0;

basenum = 2;

for i = 1:m

for j = 1:n

cmd = ['-v ',num2str(v),' -c ',num2str( basenum^X(i,j) ),' -g ',num2str( basenum^Y(i,j) )];

cg(i,j) = svmtrain(train_label, train, cmd);

if cg(i,j) > bestacc

bestacc = cg(i,j);

bestc = basenum^X(i,j);

bestg = basenum^Y(i,j);

end

if ( cg(i,j) == bestacc && bestc > basenum^X(i,j) )

bestacc = cg(i,j);

bestc = basenum^X(i,j);

bestg = basenum^Y(i,j);

end

end

end

%% to draw the acc with different c & g

[C,h] = contour(X,Y,cg,60:accstep:100);

clabel(C,h,'FontSize',10,'Color','r');

xlabel('log2c','FontSize',10);

ylabel('log2g','FontSize',10);

grid on;

=====================================

这样那个libsvm-matlab工具箱
我就有了自己的一个升级版本的了.大家可以把这个SVMcg.m加进去 一起用了...

里面有SVMcg.m使用说明.如下:

[bestacc,bestc,bestg] = SVMcg(train_label,train,cmin,cmax,gmin,gmax,v,cstep,gstep,accstep)

train_label:训练
集标签.要求与libsvm工具箱中要求一致.

train:训练集.要求与libsvm工具箱中要求一致.

cmin:惩罚参数c的变化范围的最小值(取以2为底的对数后),即 c_min = 2^(cmin).默认为 -5

cmax:惩罚参数c的变化范围的最大值(取以2为底的对数后),即 c_max = 2^(cmax).默认为 5

gmin:参数g的变化范围的最小值(取以2为底的对数后),即 g_min = 2^(gmin).默认为 -5

gmax:参数g的变化范围的最小值(取以2为底的对数后),即 g_min = 2^(gmax).默认为 5

v:cross validation的参数,即给测试集分为几部分进行cross validation.默认为 3

cstep:参数c步进的大小.默认为 1

gstep:参数g步进的大小.默认为 1

accstep:最后显示准确率图时的步进大小. 默认为 1.5

[上面这些参数大家可以更改以期达到最佳效果,也可不改用默认值]

====================