}
INSTANTIATE_TEST_CASE_P(OCL_ML, KNN, Combine(Values(6, 5), Values(Size(200, 400), Size(300, 600)),
Values(4, 3), Values(false, true)));
-#endif // HAVE_OPENCL
\ No newline at end of file
+
+////////////////////////////////SVM/////////////////////////////////////////////////
+PARAM_TEST_CASE(SVM_OCL, int, int, int)
+{
+ cv::Size size;
+ int kernel_type;
+ int svm_type;
+ Mat src, labels, samples, labels_predict;
+ int K;
+ cv::RNG rng ;
+
+ virtual void SetUp()
+ {
+
+ kernel_type = GET_PARAM(0);
+ svm_type = GET_PARAM(1);
+ K = GET_PARAM(2);
+ rng = TS::ptr()->get_rng();
+ cv::Size size = cv::Size(MWIDTH, MHEIGHT);
+ src.create(size, CV_32FC1);
+ labels.create(1, size.height, CV_32SC1);
+ int row_idx = 0;
+ const int max_number = size.height / K - 1;
+ CV_Assert(K <= size.height);
+ for(int i = 0; i < K; i++ )
+ {
+ Mat center_row_header = src.row(row_idx);
+ center_row_header.setTo(0);
+ int nchannel = center_row_header.channels();
+ for(int j = 0; j < nchannel; j++)
+ {
+ center_row_header.at<float>(0, i * nchannel + j) = 500.0;
+ }
+ labels.at<int>(0, row_idx) = i;
+ for(int j = 0; (j < max_number) ||
+ (i == K - 1 && j < max_number + size.height % K); j ++)
+ {
+ Mat cur_row_header = src.row(row_idx + 1 + j);
+ center_row_header.copyTo(cur_row_header);
+ Mat tmpmat = randomMat(rng, cur_row_header.size(), cur_row_header.type(), 1, 100, false);
+ cur_row_header += tmpmat;
+ labels.at<int>(0, row_idx + 1 + j) = i;
+ }
+ row_idx += 1 + max_number;
+ }
+ labels.convertTo(labels, CV_32FC1);
+ cv::Size test_size = cv::Size(MWIDTH, 100);
+ samples.create(test_size, CV_32FC1);
+ labels_predict.create(1, test_size.height, CV_32SC1);
+ const int max_number_test = test_size.height / K - 1;
+ row_idx = 0;
+ for(int i = 0; i < K; i++ )
+ {
+ Mat center_row_header = samples.row(row_idx);
+ center_row_header.setTo(0);
+ int nchannel = center_row_header.channels();
+ for(int j = 0; j < nchannel; j++)
+ {
+ center_row_header.at<float>(0, i * nchannel + j) = 500.0;
+ }
+ labels_predict.at<int>(0, row_idx) = i;
+ for(int j = 0; (j < max_number_test) ||
+ (i == K - 1 && j < max_number_test + test_size.height % K); j ++)
+ {
+ Mat cur_row_header = samples.row(row_idx + 1 + j);
+ center_row_header.copyTo(cur_row_header);
+ Mat tmpmat = randomMat(rng, cur_row_header.size(), cur_row_header.type(), 1, 100, false);
+ cur_row_header += tmpmat;
+ labels_predict.at<int>(0, row_idx + 1 + j) = i;
+ }
+ row_idx += 1 + max_number_test;
+ }
+ labels_predict.convertTo(labels_predict, CV_32FC1);
+ }
+};
+TEST_P(SVM_OCL, Accuracy)
+{
+ CvSVMParams params;
+ params.degree = 0.4;
+ params.gamma = 1;
+ params.coef0 = 1;
+ params.C = 1;
+ params.nu = 0.5;
+ params.p = 1;
+ params.svm_type = svm_type;
+ params.kernel_type = kernel_type;
+
+ params.term_crit = cvTermCriteria(CV_TERMCRIT_ITER, 1000, 0.001);
+
+ CvSVM SVM;
+ SVM.train(src, labels, Mat(), Mat(), params);
+
+ cv::ocl::CvSVM_OCL SVM_OCL;
+ SVM_OCL.train(src, labels, Mat(), Mat(), params);
+
+ int c = SVM.get_support_vector_count();
+ int c1 = SVM_OCL.get_support_vector_count();
+
+ Mat sv(c, MHEIGHT, CV_32FC1);
+ Mat sv_ocl(c1, MHEIGHT, CV_32FC1);
+ for(int i = 0; i < c; i++)
+ {
+ const float* v = SVM.get_support_vector(i);
+
+ for(int j = 0; j < MHEIGHT; j++)
+ {
+ sv.at<float>(i, j) = v[j];
+ }
+ }
+ for(int i = 0; i < c1; i++)
+ {
+ const float* v_ocl = SVM_OCL.get_support_vector(i);
+
+ for(int j = 0; j < MHEIGHT; j++)
+ {
+ sv_ocl.at<float>(i, j) = v_ocl[j];
+ }
+ }
+ cv::BFMatcher matcher(cv::NORM_L2);
+ std::vector<cv::DMatch> matches;
+ matcher.match(sv, sv_ocl, matches);
+ int count = 0;
+
+ for(std::vector<cv::DMatch>::iterator itr = matches.begin(); itr != matches.end(); itr++)
+ {
+ if((*itr).distance < 0.1)
+ {
+ count ++;
+ }
+ }
+ if(c != 0)
+ {
+ float matchedRatio = (float)count / c;
+ EXPECT_GT(matchedRatio, 0.95);
+ }
+ if(c != 0)
+ {
+ CvMat *result = cvCreateMat(1, samples.rows, CV_32FC1);
+ CvMat test_samples = samples;
+
+ CvMat *result_ocl = cvCreateMat(1, samples.rows, CV_32FC1);
+
+ SVM.predict(&test_samples, result);
+
+ SVM_OCL.predict(&test_samples, result_ocl);
+
+ int true_resp = 0, true_resp_ocl = 0;
+ for (int i = 0; i < samples.rows; i++)
+ {
+ if (result->data.fl[i] == labels_predict.at<float>(0, i))
+ {
+ true_resp++;
+ }
+ }
+ float matchedRatio = (float)true_resp / samples.rows;
+
+ for (int i = 0; i < samples.rows; i++)
+ {
+ if (result_ocl->data.fl[i] == labels_predict.at<float>(0, i))
+ {
+ true_resp_ocl++;
+ }
+ }
+ float matchedRatio_ocl = (float)true_resp_ocl / samples.rows;
+
+ if(matchedRatio != 0 && true_resp_ocl < true_resp)
+ {
+ EXPECT_NEAR(matchedRatio_ocl, matchedRatio, 0.03);
+ }
+ }
+}
+INSTANTIATE_TEST_CASE_P(OCL_ML, SVM_OCL, testing::Combine(
+ Values(CvSVM::LINEAR, CvSVM::POLY, CvSVM::RBF, CvSVM::SIGMOID),
+ Values(CvSVM::C_SVC, CvSVM::NU_SVC, CvSVM::ONE_CLASS, CvSVM::EPS_SVR, CvSVM::NU_SVR),
+ Values(2, 3, 4)
+ ));
+#endif // HAVE_OPENCL