public SVM() { // CvSVM // SVMを利用して2次元ベクトルの3クラス分類問題を解く const int S = 1000; const int SIZE = 400; CvRNG rng = new CvRNG((ulong)DateTime.Now.Ticks); // (1)画像領域の確保と初期化 using (IplImage img = new IplImage(SIZE, SIZE, BitDepth.U8, 3)) { img.Zero(); // (2)学習データの生成 CvPoint[] pts = new CvPoint[S]; int[] res = new int[S]; for (int i = 0; i < S; i++) { pts[i].X = (int)(rng.RandInt() % SIZE); pts[i].Y = (int)(rng.RandInt() % SIZE); if (pts[i].Y > 50 * Math.Cos(pts[i].X * Cv.PI / 100) + 200) { img.Line(new CvPoint(pts[i].X - 2, pts[i].Y - 2), new CvPoint(pts[i].X + 2, pts[i].Y + 2), new CvColor(255, 0, 0)); img.Line(new CvPoint(pts[i].X + 2, pts[i].Y - 2), new CvPoint(pts[i].X - 2, pts[i].Y + 2), new CvColor(255, 0, 0)); res[i] = 1; } else { if (pts[i].X > 200) { img.Line(new CvPoint(pts[i].X - 2, pts[i].Y - 2), new CvPoint(pts[i].X + 2, pts[i].Y + 2), new CvColor(0, 255, 0)); img.Line(new CvPoint(pts[i].X + 2, pts[i].Y - 2), new CvPoint(pts[i].X - 2, pts[i].Y + 2), new CvColor(0, 255, 0)); res[i] = 2; } else { img.Line(new CvPoint(pts[i].X - 2, pts[i].Y - 2), new CvPoint(pts[i].X + 2, pts[i].Y + 2), new CvColor(0, 0, 255)); img.Line(new CvPoint(pts[i].X + 2, pts[i].Y - 2), new CvPoint(pts[i].X - 2, pts[i].Y + 2), new CvColor(0, 0, 255)); res[i] = 3; } } } // (3)学習データの表示 Cv.NamedWindow("SVM", WindowMode.AutoSize); Cv.ShowImage("SVM", img); Cv.WaitKey(0); // (4)学習パラメータの生成 float[] data = new float[S * 2]; for (int i = 0; i < S; i++) { data[i * 2] = ((float)pts[i].X) / SIZE; data[i * 2 + 1] = ((float)pts[i].Y) / SIZE; } // (5)SVMの学習 using (CvSVM svm = new CvSVM()) { CvMat data_mat = new CvMat(S, 2, MatrixType.F32C1, data); CvMat res_mat = new CvMat(S, 1, MatrixType.S32C1, res); CvTermCriteria criteria = new CvTermCriteria(1000, float.Epsilon); CvSVMParams param = new CvSVMParams(SVMType.CSvc, SVMKernelType.Rbf, 10.0, 8.0, 1.0, 10.0, 0.5, 0.1, null, criteria); svm.Train(data_mat, res_mat, null, null, param); // (6)学習結果の描画 for (int i = 0; i < SIZE; i++) { for (int j = 0; j < SIZE; j++) { float[] a = { (float)j / SIZE, (float)i / SIZE }; CvMat m = new CvMat(1, 2, MatrixType.F32C1, a); float ret = svm.Predict(m); CvColor color = new CvColor(); switch ((int)ret) { case 1: color = new CvColor(100, 0, 0); break; case 2: color = new CvColor(0, 100, 0); break; case 3: color = new CvColor(0, 0, 100); break; } img[i, j] = color; } } // (7)トレーニングデータの再描画 for (int i = 0; i < S; i++) { CvColor color = new CvColor(); switch (res[i]) { case 1: color = new CvColor(255, 0, 0); break; case 2: color = new CvColor(0, 255, 0); break; case 3: color = new CvColor(0, 0, 255); break; } img.Line(new CvPoint(pts[i].X - 2, pts[i].Y - 2), new CvPoint(pts[i].X + 2, pts[i].Y + 2), color); img.Line(new CvPoint(pts[i].X + 2, pts[i].Y - 2), new CvPoint(pts[i].X - 2, pts[i].Y + 2), color); } // (8)サポートベクターの描画 int sv_num = svm.GetSupportVectorCount(); for (int i = 0; i < sv_num; i++) { var support = svm.GetSupportVector(i); img.Circle(new CvPoint((int)(support[0] * SIZE), (int)(support[1] * SIZE)), 5, new CvColor(200, 200, 200)); } // (9)画像の表示 Cv.NamedWindow("SVM", WindowMode.AutoSize); Cv.ShowImage("SVM", img); Cv.WaitKey(0); Cv.DestroyWindow("SVM"); } } }
/// <summary> /// SVM /// </summary> /// <param name="dataFilename"></param> /// <param name="filenameToSave"></param> /// <param name="filenameToLoad"></param> private void BuildSvmClassifier(string dataFilename, string filenameToSave, string filenameToLoad) { //C_SVCのパラメータ const float SvmC = 1000; //RBFカーネルのパラメータ const float SvmGamma = 0.1f; CvMat data = null; CvMat responses = null; CvMat sampleIdx = null; int nsamplesAll = 0, ntrainSamples = 0; double trainHr = 0, testHr = 0; CvSVM svm = new CvSVM(); CvTermCriteria criteria = new CvTermCriteria(100, 0.001); try { ReadNumClassData(dataFilename, 16, out data, out responses); } catch { Console.WriteLine("Could not read the database {0}", dataFilename); return; } Console.WriteLine("The database {0} is loaded.", dataFilename); nsamplesAll = data.Rows; ntrainSamples = (int)(nsamplesAll * 0.2); // Create or load Random Trees classifier if (filenameToLoad != null) { // load classifier from the specified file svm.Load(filenameToLoad); ntrainSamples = 0; if (svm.GetSupportVectorCount() == 0) { Console.WriteLine("Could not read the classifier {0}", filenameToLoad); return; } Console.WriteLine("The classifier {0} is loaded.", filenameToLoad); } else { // create classifier by using <data> and <responses> Console.Write("Training the classifier ..."); // 2. create sample_idx sampleIdx = new CvMat(1, nsamplesAll, MatrixType.U8C1); { CvMat mat; Cv.GetCols(sampleIdx, out mat, 0, ntrainSamples); mat.Set(CvScalar.RealScalar(1)); Cv.GetCols(sampleIdx, out mat, ntrainSamples, nsamplesAll); mat.SetZero(); } // 3. train classifier // 方法、カーネルにより使わないパラメータは0で良く、 // 重みについてもNULLで良い svm.Train(data, responses, null, sampleIdx, new CvSVMParams(CvSVM.C_SVC, CvSVM.RBF, 0, SvmGamma, 0, SvmC, 0, 0, null, criteria)); Console.WriteLine(); } // compute prediction error on train and test data for (int i = 0; i < nsamplesAll; i++) { double r; CvMat sample; Cv.GetRow(data, out sample, i); r = svm.Predict(sample); // compare results Console.WriteLine( "predict: {0}, responses: {1}, {2}", (char)r, (char)responses.DataArraySingle[i], Math.Abs((double)r - responses.DataArraySingle[i]) <= float.Epsilon ? "Good!" : "Bad!" ); r = Math.Abs((double)r - responses.DataArraySingle[i]) <= float.Epsilon ? 1 : 0; if (i < ntrainSamples) trainHr += r; else testHr += r; } testHr /= (double)(nsamplesAll - ntrainSamples); trainHr /= (double)ntrainSamples; Console.WriteLine("Gamma={0:F5}, C={1:F5}", SvmGamma, SvmC); if (filenameToLoad != null) { Console.WriteLine("Recognition rate: test = {0:F1}%", testHr * 100.0); } else { Console.WriteLine("Recognition rate: train = {0:F1}%, test = {1:F1}%", trainHr * 100.0, testHr * 100.0); } Console.WriteLine("Number of Support Vector: {0}", svm.GetSupportVectorCount()); // Save SVM classifier to file if needed if (filenameToSave != null) { svm.Save(filenameToSave); } Console.Read(); if (sampleIdx != null) sampleIdx.Dispose(); data.Dispose(); responses.Dispose(); svm.Dispose(); }