@Override
public AnalysisCollector analyze(CollectionReader cr) throws AnalyzerFailureException {
svm_problem svmProblem = loadData(cr);
svm_problem sampled = null;
if (findBestParameters) {
if (sample < 1d) {
logger.debug("Sampling.");
sampled = do_sample(svmProblem);
}
logger.debug("Performing grid search.");
do_find_best_parameters(sampled != null ? sampled : svmProblem);
}
svm_parameter svmParam = getDefaultSvmParameters();
svmParam.probability = 1;
svmParam.C = c;
svmParam.gamma = gamma;
setWeights(svmParam);
logger.debug("Training with C=" + c + " gamma=" + gamma);
svm_model model = svm.svm_train(svmProblem, svmParam);
logger.debug("Done!");
return new SingletonAnalysisCollector(
new LibSvmTrainerAnalysis(model, scaler, labelList, c, gamma));
}
private void setWeights(svm_parameter svmParam) {
if (weights != null) {
svmParam.nr_weight = weights.size();
svmParam.weight_label = new int[weights.size()];
svmParam.weight = new double[weights.size()];
for (int i = 0; i < weights.size(); ++i) {
svmParam.weight_label[i] = i;
svmParam.weight[i] = weights.get(i);
}
logger.debug("Class weights: " + weights);
}
}
public svm_parameter getDefaultSvmParameters() {
svm_parameter param = new svm_parameter();
// default values
param.svm_type = svm_parameter.C_SVC;
param.kernel_type = svm_parameter.RBF;
param.degree = 3;
param.gamma = 0; // 1/num_features
param.coef0 = 0;
param.nu = 0.5;
param.cache_size = 100;
param.C = 1;
param.eps = 1e-3;
param.p = 0.1;
param.shrinking = 1;
param.probability = 0;
param.nr_weight = 0;
param.weight_label = new int[0];
param.weight = new double[0];
return param;
}
private void do_find_best_parameters(svm_problem svmProblem) {
svm_parameter svmParam = getDefaultSvmParameters();
setWeights(svmParam);
int maxIter =
((int) Math.ceil(Math.abs((log2cEnd - log2cBegin) / log2cStep)) + 1)
* ((int) Math.ceil(Math.abs((log2gEnd - log2gBegin) / log2gStep)) + 1);
// Run the grid search in separate CV threads
ExecutorService executorService = Executors.newFixedThreadPool(numberOfThreads);
List<CvParams> cvParamsList = new ArrayList<CvParams>();
for (double log2c = log2cBegin;
(log2cBegin < log2cEnd && log2c <= log2cEnd)
|| (log2cBegin >= log2cEnd && log2c >= log2cEnd);
log2c += log2cStep) {
double c1 = Math.pow(2, log2c);
for (double log2g = log2gBegin;
(log2gBegin < log2gEnd && log2g <= log2gEnd)
|| (log2gBegin >= log2gEnd && log2g >= log2gEnd);
log2g += log2gStep) {
double gamma1 = Math.pow(2, log2g);
svm_parameter svmParam1 = (svm_parameter) svmParam.clone();
svmParam1.C = c1;
svmParam1.gamma = gamma1;
executorService.execute(
new RunnableSvmCrossValidator(svmProblem, svmParam1, nrFold, cvParamsList));
}
}
// now wait for all threads to complete by calling shutdown
// note that this will NOT terminate the currently running threads, it just signals the thread
// pool to closeWriter
// once all work is completed
executorService.shutdown();
while (!executorService.isTerminated()) {
try {
Thread.sleep(1000);
} catch (InterruptedException e) {
// don't care if we get interrupted
}
// every second, report statistics
logger.debug(
String.format("%% complete: %5.2f", cvParamsList.size() / (double) maxIter * 100));
CvParams best = getBestCvParams(cvParamsList);
CvParams worst = getWorstcvParams(cvParamsList);
if (best != null) {
logger.debug("Best accuracy: " + best.accuracy);
logger.debug("Best C: " + best.c);
logger.debug("Best Gamma: " + best.gamma);
}
if (worst != null) {
logger.debug("Worst accuracy: " + worst.accuracy);
}
}
CvParams best = getBestCvParams(cvParamsList);
CvParams worst = getWorstcvParams(cvParamsList);
if (best != null) {
logger.debug("Best accuracy: " + best.accuracy);
logger.debug("Best C: " + best.c);
logger.debug("Best Gamma: " + best.gamma);
c = best.c;
gamma = best.gamma;
} else {
logger.error("Best CV parameters is null.");
}
if (worst != null) {
logger.debug("Worst accuracy: " + worst.accuracy);
}
}