diff --git a/modules/ml/include/opencv2/ml.hpp b/modules/ml/include/opencv2/ml.hpp index e596ead0e..bcd22c46b 100644 --- a/modules/ml/include/opencv2/ml.hpp +++ b/modules/ml/include/opencv2/ml.hpp @@ -1626,10 +1626,10 @@ public: * stepDecreasingPower = 1; * termCrit.maxCount = 100000; * termCrit.epsilon = 0.00001; - * @param svmsgdType is the type of SVMSGD classifier. Legal values are SvmsgdType::SGD and SvmsgdType::ASGD. - * Recommended value is SvmsgdType::ASGD (by default). - * @param marginType is the type of margin constraint. Legal values are MarginType::SOFT_MARGIN and MarginType::HARD_MARGIN. - * Default value is MarginType::SOFT_MARGIN. + * @param svmsgdType is the type of SVMSGD classifier. Legal values are SVMSGD::SvmsgdType::SGD and SVMSGD::SvmsgdType::ASGD. + * Recommended value is SVMSGD::SvmsgdType::ASGD (by default). + * @param marginType is the type of margin constraint. Legal values are SVMSGD::MarginType::SOFT_MARGIN and SVMSGD::MarginType::HARD_MARGIN. + * Default value is SVMSGD::MarginType::SOFT_MARGIN. */ CV_WRAP virtual void setOptimalParameters(int svmsgdType = SVMSGD::ASGD, int marginType = SVMSGD::SOFT_MARGIN) = 0; diff --git a/modules/ml/src/svmsgd.cpp b/modules/ml/src/svmsgd.cpp index 0a4048168..9a03ed9ac 100644 --- a/modules/ml/src/svmsgd.cpp +++ b/modules/ml/src/svmsgd.cpp @@ -142,6 +142,7 @@ void SVMSGDImpl::normalizeSamples(Mat &samples, Mat &average, float &multiplier) int samplesCount = samples.rows; average = Mat(1, featuresCount, samples.type()); + CV_Assert(average.type() == CV_32FC1); for (int featureIndex = 0; featureIndex < featuresCount; featureIndex++) { average.at(featureIndex) = static_cast(mean(samples.col(featureIndex))[0]); @@ -170,11 +171,11 @@ void SVMSGDImpl::makeExtendedTrainSamples(const Mat &trainSamples, Mat &extended cv::hconcat(normalizedTrainSamples, onesCol, extendedTrainSamples); } -void SVMSGDImpl::updateWeights(InputArray _sample, bool firstClass, float stepSize, Mat& weights) +void SVMSGDImpl::updateWeights(InputArray _sample, bool positive, float stepSize, Mat& weights) { Mat sample = _sample.getMat(); - int response = firstClass ? 1 : -1; // ensure that trainResponses are -1 or 1 + int response = positive ? 1 : -1; // ensure that trainResponses are -1 or 1 if ( sample.dot(weights) * response > 1) { @@ -197,6 +198,7 @@ float SVMSGDImpl::calcShift(InputArray _samples, InputArray _responses) const Mat trainResponses = _responses.getMat(); + CV_Assert(trainResponses.type() == CV_32FC1); for (int samplesIndex = 0; samplesIndex < trainSamplesCount; samplesIndex++) { Mat currentSample = trainSamples.row(samplesIndex); @@ -261,7 +263,7 @@ bool SVMSGDImpl::train(const Ptr& data, int) RNG rng(0); - CV_Assert (params.termCrit.type & TermCriteria::COUNT || params.termCrit.type & TermCriteria::EPS); + CV_Assert ((params.termCrit.type & TermCriteria::COUNT || params.termCrit.type & TermCriteria::EPS) && (trainResponses.type() == CV_32FC1)); int maxCount = (params.termCrit.type & TermCriteria::COUNT) ? params.termCrit.maxCount : INT_MAX; double epsilon = (params.termCrit.type & TermCriteria::EPS) ? params.termCrit.epsilon : 0; @@ -300,7 +302,7 @@ bool SVMSGDImpl::train(const Ptr& data, int) weights_ = extendedWeights(roi); weights_ *= multiplier; - CV_Assert(params.marginType == SOFT_MARGIN || params.marginType == HARD_MARGIN); + CV_Assert((params.marginType == SOFT_MARGIN || params.marginType == HARD_MARGIN) && (extendedWeights.type() == CV_32FC1)); if (params.marginType == SOFT_MARGIN) { @@ -332,7 +334,7 @@ float SVMSGDImpl::predict( InputArray _samples, OutputArray _results, int ) cons else { CV_Assert( nSamples == 1 ); - results = Mat(1, 1, CV_32F, &result); + results = Mat(1, 1, CV_32FC1, &result); } for (int sampleIndex = 0; sampleIndex < nSamples; sampleIndex++) diff --git a/modules/ml/test/test_svmsgd.cpp b/modules/ml/test/test_svmsgd.cpp index 9a206e47e..b6aed3c7e 100644 --- a/modules/ml/test/test_svmsgd.cpp +++ b/modules/ml/test/test_svmsgd.cpp @@ -62,8 +62,7 @@ public: private: virtual void run( int start_from ); static float decisionFunction(const Mat &sample, const Mat &weights, float shift); - void makeTrainData(Mat weights, float shift); - void makeTestData(Mat weights, float shift); + void makeData(int samplesCount, Mat weights, float shift, RNG rng, Mat &samples, Mat & responses); void generateSameBorders(int featureCount); void generateDifferentBorders(int featureCount); @@ -108,46 +107,28 @@ void CV_SVMSGDTrainTest::generateDifferentBorders(int featureCount) } } -void CV_SVMSGDTrainTest::makeTrainData(Mat weights, float shift) +float CV_SVMSGDTrainTest::decisionFunction(const Mat &sample, const Mat &weights, float shift) +{ + return static_cast(sample.dot(weights)) + shift; +} + +void CV_SVMSGDTrainTest::makeData(int samplesCount, Mat weights, float shift, RNG rng, Mat &samples, Mat & responses) { - int datasize = 100000; int featureCount = weights.cols; - RNG rng(0); - - cv::Mat samples = cv::Mat::zeros(datasize, featureCount, CV_32FC1); + samples.create(samplesCount, featureCount, CV_32FC1); for (int featureIndex = 0; featureIndex < featureCount; featureIndex++) { rng.fill(samples.col(featureIndex), RNG::UNIFORM, borders[featureIndex].first, borders[featureIndex].second); } - cv::Mat responses = cv::Mat::zeros(datasize, 1, CV_32FC1); - for (int sampleIndex = 0; sampleIndex < datasize; sampleIndex++) + responses.create(samplesCount, 1, CV_32FC1); + + for (int i = 0 ; i < samplesCount; i++) { - responses.at(sampleIndex) = decisionFunction(samples.row(sampleIndex), weights, shift) > 0 ? 1.f : -1.f; + responses.at(i) = decisionFunction(samples.row(i), weights, shift) > 0 ? 1.f : -1.f; } - data = TrainData::create(samples, cv::ml::ROW_SAMPLE, responses); -} - -void CV_SVMSGDTrainTest::makeTestData(Mat weights, float shift) -{ - int testSamplesCount = 100000; - int featureCount = weights.cols; - cv::RNG rng(42); - - testSamples.create(testSamplesCount, featureCount, CV_32FC1); - for (int featureIndex = 0; featureIndex < featureCount; featureIndex++) - { - rng.fill(testSamples.col(featureIndex), RNG::UNIFORM, borders[featureIndex].first, borders[featureIndex].second); - } - - testResponses.create(testSamplesCount, 1, CV_32FC1); - - for (int i = 0 ; i < testSamplesCount; i++) - { - testResponses.at(i) = decisionFunction(testSamples.row(i), weights, shift) > 0 ? 1.f : -1.f; - } } CV_SVMSGDTrainTest::CV_SVMSGDTrainTest(const Mat &weights, float shift, TrainDataType _type, double _precision) @@ -169,13 +150,16 @@ CV_SVMSGDTrainTest::CV_SVMSGDTrainTest(const Mat &weights, float shift, TrainDat CV_Error(CV_StsBadArg, "Unknown train data type"); } - makeTrainData(weights, shift); - makeTestData(weights, shift); -} + RNG rng(0); -float CV_SVMSGDTrainTest::decisionFunction(const Mat &sample, const Mat &weights, float shift) -{ - return static_cast(sample.dot(weights)) + shift; + Mat trainSamples; + Mat trainResponses; + int trainSamplesCount = 10000; + makeData(trainSamplesCount, weights, shift, rng, trainSamples, trainResponses); + data = TrainData::create(trainSamples, cv::ml::ROW_SAMPLE, trainResponses); + + int testSamplesCount = 100000; + makeData(testSamplesCount, weights, shift, rng, testSamples, testResponses); } void CV_SVMSGDTrainTest::run( int /*start_from*/ ) @@ -205,7 +189,6 @@ void CV_SVMSGDTrainTest::run( int /*start_from*/ ) } } - void makeWeightsAndShift(int featureCount, Mat &weights, float &shift) { weights.create(1, featureCount, CV_32FC1); @@ -253,7 +236,7 @@ TEST(ML_SVMSGD, trainSameScale100) float shift = 0; makeWeightsAndShift(featureCount, weights, shift); - CV_SVMSGDTrainTest test(weights, shift, CV_SVMSGDTrainTest::UNIFORM_SAME_SCALE); + CV_SVMSGDTrainTest test(weights, shift, CV_SVMSGDTrainTest::UNIFORM_SAME_SCALE, 0.02); test.safe_run(); }