/**
* wrap up various variables to save memeory and do some housekeeping after optimization has
* finished.
*
* @throws Exception if something goes wrong
*/
protected void wrapUp() throws Exception {
m_target = null;
m_nEvals = m_kernel.numEvals();
m_nCacheHits = m_kernel.numCacheHits();
if ((m_SVM.getKernel() instanceof PolyKernel)
&& ((PolyKernel) m_SVM.getKernel()).getExponent() == 1.0) {
// convert alpha's to weights
double[] weights = new double[m_data.numAttributes()];
for (int k = m_supportVectors.getNext(-1); k != -1; k = m_supportVectors.getNext(k)) {
for (int j = 0; j < weights.length; j++) {
if (j != m_classIndex) {
weights[j] += (m_alpha[k] - m_alphaStar[k]) * m_data.instance(k).value(j);
}
}
}
m_weights = weights;
// release memory
m_alpha = null;
m_alphaStar = null;
m_kernel = null;
}
m_bModelBuilt = true;
}
/**
* initializes the algorithm
*
* @param data the data to work with
* @throws Exception if m_SVM is null
*/
protected void init(Instances data) throws Exception {
if (m_SVM == null) {
throw new Exception("SVM not initialized in optimizer. Use RegOptimizer.setSVMReg()");
}
m_C = m_SVM.getC();
m_data = data;
m_classIndex = data.classIndex();
m_nInstances = data.numInstances();
// Initialize kernel
m_kernel = Kernel.makeCopy(m_SVM.getKernel());
m_kernel.buildKernel(data);
// init m_target
m_target = new double[m_nInstances];
for (int i = 0; i < m_nInstances; i++) {
m_target[i] = data.instance(i).classValue();
}
m_random = new Random(m_nSeed);
// initialize alpha and alpha* array to all zero
m_alpha = new double[m_target.length];
m_alphaStar = new double[m_target.length];
m_supportVectors = new SMOset(m_nInstances);
m_b = 0.0;
m_nEvals = 0;
m_nCacheHits = -1;
}
/**
* SVMOutput of an instance in the training set, m_data This uses the cache, unlike
* SVMOutput(Instance)
*
* @param index index of the training instance in m_data
* @return the SVM output
* @throws Exception if something goes wrong
*/
protected double SVMOutput(int index) throws Exception {
double result = -m_b;
for (int i = m_supportVectors.getNext(-1); i != -1; i = m_supportVectors.getNext(i)) {
result += (m_alpha[i] - m_alphaStar[i]) * m_kernel.eval(index, i, m_data.instance(index));
}
return result;
}
/**
* Compute the value of the objective function.
*
* @return the score
* @throws Exception if something goes wrong
*/
protected double getScore() throws Exception {
double res = 0;
double t = 0, t2 = 0;
double sumAlpha = 0.0;
for (int i = 0; i < m_nInstances; i++) {
sumAlpha += (m_alpha[i] - m_alphaStar[i]);
for (int j = 0; j < m_nInstances; j++) {
t +=
(m_alpha[i] - m_alphaStar[i])
* (m_alpha[j] - m_alphaStar[j])
* m_kernel.eval(i, j, m_data.instance(i));
}
// switch(m_nLossType) {
// case L1:
// t2 += m_data.instance(i).classValue() * (m_alpha[i] - m_alpha_[i]);
// break;
// case L2:
// t2 += m_data.instance(i).classValue() * (m_alpha[i] - m_alpha_[i]) - (0.5/m_SVM.getC())
// * (m_alpha[i]*m_alpha[i] + m_alpha_[i]*m_alpha_[i]);
// break;
// case HUBER:
// t2 += m_data.instance(i).classValue() * (m_alpha[i] - m_alpha_[i]) -
// (0.5*m_SVM.getEpsilon()/m_SVM.getC()) * (m_alpha[i]*m_alpha[i] + m_alpha_[i]*m_alpha_[i]);
// break;
// case EPSILON:
// t2 += m_data.instance(i).classValue() * (m_alpha[i] - m_alphaStar[i]) - m_epsilon *
// (m_alpha[i] + m_alphaStar[i]);
t2 += m_target[i] * (m_alpha[i] - m_alphaStar[i]) - m_epsilon * (m_alpha[i] + m_alphaStar[i]);
// break;
// }
}
res += -0.5 * t + t2;
return res;
}
/**
* @param inst
* @return
* @throws Exception
*/
public double SVMOutput(Instance inst) throws Exception {
double result = -m_b;
// Is the machine linear?
if (m_weights != null) {
// Is weight vector stored in sparse format?
for (int i = 0; i < inst.numValues(); i++) {
if (inst.index(i) != m_classIndex) {
result += m_weights[inst.index(i)] * inst.valueSparse(i);
}
}
} else {
for (int i = m_supportVectors.getNext(-1); i != -1; i = m_supportVectors.getNext(i)) {
result += (m_alpha[i] - m_alphaStar[i]) * m_kernel.eval(-1, i, inst);
}
}
return result;
}