The best parameter for prob outputs may be different from the setting

[pango99]

The best parameter for prob outputs may be different from the setting
without..
See also the following faq

Q: Why using svm-predict -b 0 and -b 1 gives different accuracy values?

Let's just consider two-class classification here. After probability
information is obtained in training, we do not have

prob > = 0.5 if and only if decision value >= 0.

So predictions may be different with -b 0 and 1.

On 2019-09-27 01:22, Zheng Li wrote:

Hi, I use libSVM to develop my face recognition system,I use the
libSVM to find the closet face feature and identify target person,but
I found when I call svm_predict_probability() to do the prediction,the
accuracy is very low,and when I call svm_predict(), the accuracy is
high,because I need the probability param to check similarity , so I
want to know why svm_predict_probability()'s accuracy is so low?

below is my code to create the svm_model:

inline void fillLibSvmNodes(std::vector &fea, std::vector<svm_node>
&line_x_space)
{
size_t featuresNb = fea.size();
assert((fea.size() + 1) == line_x_space.size());
for (int i = 0; i < featuresNb; i++)
{
line_x_space[i].index = i + 1;
line_x_space[i].value = (double)fea[i];
}
line_x_space[featuresNb].index = -1;
}

void testLibSVM()
{
// dbFaceFeatures is a global variable to store the load face
features, it is vector<pair<string,vector>> type,the face feature is
512D;
assert(dbFaceFeatures.size() > 0);
int featuresNb = (int)dbFaceFeatures[0].second.size(); //
featuresNb=512

memset(&param, 0, sizeof(svm_parameter));
param.svm_type = C_SVC;
param.kernel_type = LINEAR;
param.cache_size = 512;
param.eps = cv::TermCriteria(cv::TermCriteria::MAX_ITER, 20000,
1e-06).epsilon;
param.degree = 0;
param.gamma = 1.0;
param.coef0 = 0;
param.nu = 0;
param.C = 1.0;
param.p = 0;
param.probability = 1; // do probability estimate
param.shrinking = 0;
param.nr_weight = 0;
param.weight = nullptr;
param.weight_label = nullptr;

memset(&prob, 0, sizeof(svm_problem));
prob.l = (int)dbFaceFeatures.size();

y_space.resize(dbFaceFeatures.size()); // y_space is vector
type
prob.y = y_space.data();
for (size_t i = 0; i < y_space.size(); i++)
y_space.at(i) = (double)(i + 1);

prob.x = (svm_node **)malloc(sizeof(svm_node
*)*dbFaceFeatures.size());
x_space.resize(dbFaceFeatures.size()); // x_space is
vector<vector<svm_node>> type
for(int n=0;n<(int)dbFaceFeatures.size();n++)
{
std::vector<svm_node> &line_x_space = x_space.at(n);
line_x_space.resize(featuresNb + 1);
prob.x[n] = line_x_space.data();
fillLibSvmNodes(dbFaceFeatures.at(n).second, line_x_space);
}

const char * err_msg = svm_check_parameter(&prob, &param);
if (err_msg)
{
std::cerr << "svm_check_parameter() fail " << err_msg <<
std::endl;
assert(false);
return;
}
lib_svm = svm_train(&prob, &param);
assert(lib_svm);
int chkRet = svm_check_probability_model(lib_svm);
if(chkRet == 0)
{
std::cerr << "svm_check_probability_model() fail" << std::endl;
assert(false);
return;
}

}

--
You are receiving this because you are subscribed to this thread.
Reply to this email directly, view it on GitHub [1], or mute the
thread [2]. [ { "@context": "http://schema.org", "@type":
"EmailMessage", "potentialAction": { "@type": "ViewAction", "target":
"https://github.com//issues/152?email_source=notifications\u0026email_token=ABI3BHWM3YKIZAE6BTHHJCTQLW7DTA5CNFSM4I3DKNV2YY3PNVWWK3TUL52HS4DFUVEXG43VMWVGG33NNVSW45C7NFSM4HOC2G3A",
"url":
"https://github.com//issues/152?email_source=notifications\u0026email_token=ABI3BHWM3YKIZAE6BTHHJCTQLW7DTA5CNFSM4I3DKNV2YY3PNVWWK3TUL52HS4DFUVEXG43VMWVGG33NNVSW45C7NFSM4HOC2G3A",
"name": "View Issue" }, "description": "View this Issue on GitHub",
"publisher": { "@type": "Organization", "name": "GitHub", "url":
"https://github.com" } } ]

Links:

[1]
https://github.com/cjlin1/libsvm/issues/152?email_source=notifications&amp;email_token=ABI3BHWM3YKIZAE6BTHHJCTQLW7DTA5CNFSM4I3DKNV2YY3PNVWWK3TUL52HS4DFUVEXG43VMWVGG33NNVSW45C7NFSM4HOC2G3A
[2]
https://github.com/notifications/unsubscribe-auth/ABI3BHUFUBG3VCMXEIEHU2DQLW7DTANCNFSM4I3DKNVQ

Originally posted by @cjlin1 in #152 (comment)