Пример #1
0
def main():
    plot.close('all')

    #       0 Read data, selecting digits 3 and 8, dimension reduction
    print
    print "Read data, selecting digits 3 and 8, dimension reduction"
    print

    test_path = "test.h5"
    training_path = "train.h5"

    images_train = vigra.readHDF5(test_path, "images")
    labels_train = vigra.readHDF5(test_path, "labels")

    images_test = vigra.readHDF5(training_path, "images")
    labels_test = vigra.readHDF5(training_path, "labels")

    print 'Size of the training set: {}'.format(np.shape(images_train))
    print np.shape(labels_train)
    print 'Size of the test set: {}'.format(np.shape(images_test))
    print np.shape(labels_test)

    # Reshape data

    n = images_train.shape[0]
    d = images_train.shape[1]
    images_train = images_train.reshape(n, d * d)

    n = images_test.shape[0]
    assert d != images_test.shape[
        0], 'Test and training sets have different dim'
    images_test = images_test.reshape(n, d * d)

    # Select 3s and 8s

    ind3 = (labels_train == 3)
    ind8 = (labels_train == 8)

    images_train_38 = images_train[ind3 + ind8]
    labels_train_38 = labels_train[ind3 + ind8]

    ind3 = (labels_test == 3)
    ind8 = (labels_test == 8)

    images_test_38 = images_test[ind3 + ind8]
    labels_test_38 = labels_test[ind3 + ind8]

    # Dimension reduction

    rimages_train_38 = dr(images_train_38, labels_train_38, [3, 8])
    rimages_test_38 = dr(images_test_38, labels_test_38, [3, 8])

    print 'Size of the training set of 3s and 8s: {}'.format(
        np.shape(rimages_train_38))
    print 'Size of the test set of 3s and 8s: {}'.format(
        np.shape(rimages_test_38))

    print
    print "1 Naive Bayes"
    print
    print "1.1 Classification"
    print

    # Training: priors and likelihood for each d=1,2
    # for each feature and class individual histograms <=> 4 histogramms

    n = rimages_train_38.shape[0]
    d = rimages_train_38.shape[1]

    # Choose bin width
    L, dx = chooseBinSize(rimages_train_38)

    # train classifier for each class separatly
    p3, pdf3 = naiveBayes_train_single_class(rimages_train_38, \
                                                  labels_train_38, 3, L, dx)
    p8, pdf8 = naiveBayes_train_single_class(rimages_train_38, \
                                                  labels_train_38, 8, L, dx)

    rimages_test_38_predict = naiveBayesClassifier(rimages_test_38, \
                                                p3, p8, pdf3, pdf8, L, dx)

    ccr_naiveBayes = correctClassRate(rimages_test_38_predict,\
                                      labels_test_38, [3,8], \
                                      print_confMatrix = True)

    print 'Correct Classification rate on the test set:{}'.format(
        ccr_naiveBayes)
    print 'Error rate on the test set:{}'.format(1 - ccr_naiveBayes)

    plot_histogram(pdf3, dx, "Histograms of the class 3", "histograms3.png")
    plot_histogram(pdf8, dx, "Histograms of the class 8", "histograms8.png")

    plot_likelihood(pdf3, dx, "Likelihoods of the class 3", "likelihoods3.png")
    plot_likelihood(pdf8, dx, "Likelihoods of the class 8", "likelihoods8.png")

    print
    print "1.2 Generate Threes"
    print

    # use function  naiveBayes_train_single_class to compute the likelihood
    # for all feature dimension

    n = images_train_38.shape[0]
    d = images_train_38.shape[1]

    # Choose bin width
    L, dx = chooseBinSize(images_train_38)

    # train classifier for each class separatly
    p3, pdf3 = naiveBayes_train_single_class(images_train_38, \
                                                  labels_train_38, 3, L, dx)
    # generate 3 new threes
    new3th = np.zeros((3, d), dtype=np.int32)
    for i in range(0, 3):
        new3th[i, :] = generate3naiveBayes(pdf3, dx)

        img = new3th[i, :].reshape(np.sqrt(d), np.sqrt(d))
        plot.figure()
        plot.gray()
        plot.imshow(img)
        plot.show()
    # end for i

    print
    print "2 Density Tree"
    print
    print "Naive splitting"
    print

    tstart = time.time()

    # class 3
    prior3, DT3 = DT_learning(rimages_train_38, labels_train_38, 3, 'naive')
    DT_visualize2D(DT3, rimages_train_38, labels_train_38, 3, "naiveDT3.png")
    # class 8
    prior8, DT8 = DT_learning(rimages_train_38, labels_train_38, 8, 'naive')
    DT_visualize2D(DT8, rimages_train_38, labels_train_38, 8, "naiveDT8.png")

    tstop = time.time()
    print "DT learning time (naive splitting) {}".format(tstop - tstart)

    tstart = time.time()
    rimages_test_38_predict = DT_Classifier_2classes(rimages_test_38, prior3,
                                                     prior8, DT3, DT8, [3, 8])
    tstop = time.time()
    print "DT classification time (naive splitting) {}".format(tstop - tstart)

    ccr_DT = correctClassRate(rimages_test_38_predict,\
                                      labels_test_38, [3,8], \
                                      print_confMatrix = True)

    print 'Correct Classification rate on the test set:{}'.format(ccr_DT)
    print 'Error rate on the test set:{}'.format(1 - ccr_DT)

    print
    print 'Generate new threes: '

    d = images_train_38.shape[1]

    # new threes
    new3th = np.zeros((3, d), dtype=np.int32)
    for i in range(0, 3):
        new3th[i, :] = generate3DT(images_train_38, labels_train_38, 3)
        img = new3th[i, :].reshape(np.sqrt(d), np.sqrt(d))

        plot.figure()
        plot.gray()
        plot.imshow(img)
        plot.show()
    # end for i

    print
    print "Clever splitting"
    print

    #    # class 3
    #    prior3, DT3 = DT_learning(rimages_train_38, labels_train_38, 3, 'clever')
    #    DT_visualize2D(DT3, rimages_train_38, labels_train_38, 3, "naiveDT3.png")
    #    # class 8
    #    prior8, DT8 = DT_learning(rimages_train_38, labels_train_38, 8, 'clever')
    #    DT_visualize2D(DT8, rimages_train_38, labels_train_38, 8, "naiveDT8.png")
    #
    #
    #    rimages_test_38_predict = DT_Classifier_2classes(rimages_test_38,
    #                                                prior3, prior8, DT3, DT8, [3,8])
    #
    #    ccr_DT = correctClassRate(rimages_test_38_predict,\
    #                                      labels_test_38, [3,8], \
    #                                      print_confMatrix = True)
    #
    #    print 'Correct Classification rate on the test set:{}'.format(ccr_DT)
    #    print 'Error rate on the test set:{}'.format(1-ccr_DT)
    #
    #    print
    #    print "Generate Threes"
    #    print

    print
    print "3 Combine DT and Naive Bayes"
    print

    n = images_train_38.shape[0]
    d = images_train_38.shape[1]

    print
    print "Learning phase"
    print

    # train 1D-histogramms for each feature and class

    tstart = time.time()
    L, dx = chooseBinSize(images_train_38)  # number of bins, bins size

    # pdf dxL matrices
    prior3, pdf3 = naiveBayes_train_single_class(images_train_38, \
                                                  labels_train_38, 3, L, dx)
    prior8, pdf8 = naiveBayes_train_single_class(images_train_38, \
                                                  labels_train_38, 8, L, dx)
    # compute the cdf of each histogramm
    cdf3 = []
    cdf8 = []
    pdf3sorted = []
    pdf8sorted = []
    for j in range(0, d):
        tmp = np.sort(pdf3[j])
        pdf3sorted.append(tmp)
        cdf3.append(np.cumsum(tmp))

        tmp = np.sort(pdf8[j])
        pdf8sorted.append(tmp)
        cdf8.append(np.cumsum(tmp))
    # end for

    tstop = time.time()
    print "Learning 1D histograms and computing cdf's took {} sec".\
                                                        format(tstop-tstart)

    # map data to copula using rank order transformation
    u = np.zeros(images_train_38.shape, dtype=np.float32)
    for j in range(0, d):
        ind = np.argsort(images_train_38[:, j])
        u[:, j] = ind[:] / float(n + 1)
    # end for j

    print
    print "Generate threes"

    new3th = np.zeros((3, d), dtype=np.int32)
    for i in range(0, 3):

        u1 = generate3(u, labels_train_38, 3)

        for j in range(0, d):
            dist = abs(cdf3[j] - u1[j])
            binx = np.argsort(dist)
            new3th[i, j] = np.floor(dx[j] * binx[0] / 2.)
        # end for j
        img = new3th[i, :].reshape(np.sqrt(d), np.sqrt(d))

        plot.figure()
        plot.gray()
        plot.imshow(img)
        plot.show()
Пример #2
0
def main():
    plot.close('all')

    #       0 Read data, selecting digits 3 and 8, dimension reduction
    print
    print "Read data, selecting digits 3 and 8, dimension reduction"
    print

    test_path = "test.h5"
    training_path = "train.h5"

    images_train = vigra.readHDF5(test_path, "images")
    labels_train = vigra.readHDF5(test_path, "labels")

    images_test = vigra.readHDF5(training_path, "images")
    labels_test = vigra.readHDF5(training_path, "labels")

    print 'Size of the training set: {}'.format(np.shape(images_train))
    print np.shape(labels_train)
    print 'Size of the test set: {}'.format(np.shape(images_test))
    print np.shape(labels_test)

    # Reshape data

    n = images_train.shape[0]
    d = images_train.shape[1]
    images_train = images_train.reshape(n, d * d)

    n = images_test.shape[0]
    assert d != images_test.shape[
        0], 'Test and training sets have different dim'
    images_test = images_test.reshape(n, d * d)

    # Select 3s and 8s

    ind3 = (labels_train == 3)
    ind8 = (labels_train == 8)

    images_train_38 = images_train[ind3 + ind8]
    labels_train_38 = labels_train[ind3 + ind8]

    ind3 = (labels_test == 3)
    ind8 = (labels_test == 8)

    images_test_38 = images_test[ind3 + ind8]
    labels_test_38 = labels_test[ind3 + ind8]

    # Dimension reduction

    rimages_train_38 = dr(images_train_38, labels_train_38, [3, 8])
    rimages_test_38 = dr(images_test_38, labels_test_38, [3, 8])

    print 'Size of the training set of 3s and 8s: {}'.format(
        np.shape(rimages_train_38))
    print 'Size of the test set of 3s and 8s: {}'.format(
        np.shape(rimages_test_38))

    print
    print "1 Naive Bayes"
    print
    print "1.1 Classification"
    print

    # Training: priors and likelihood for each d=1,2
    # for each feature and class individual histograms <=> 4 histogramms

    n = rimages_train_38.shape[0]
    d = rimages_train_38.shape[1]

    # Choose bin width
    L, dx = chooseBinSize(rimages_train_38)

    # train classifier for each class separatly
    p3, pdf3 = naiveBayes_train_single_class(rimages_train_38, \
                                                  labels_train_38, 3, dx, L)
    p8, pdf8 = naiveBayes_train_single_class(rimages_train_38, \
                                                  labels_train_38, 8, dx, L)

    rimages_test_38_predict = naiveBayesClassifier(rimages_test_38, \
                                                p3, p8, pdf3, pdf8, dx)

    ccr_naiveBayes = correctClassRate(rimages_test_38_predict,\
                                      labels_test_38, [3,8], \
                                      print_confMatrix = True)

    print 'Correct Classification rate on the test set:{}'.format(
        ccr_naiveBayes)
    print 'Error rate on the test set:{}'.format(1 - ccr_naiveBayes)

    plot_histogram(pdf3, dx, "Histograms of the class 3", "histograms3.png")
    plot_histogram(pdf8, dx, "Histograms of the class 8", "histograms8.png")

    plot_likelihood(pdf3, dx, "Likelihoods of the class 3", "likelihoods3.png")
    plot_likelihood(pdf8, dx, "Likelihoods of the class 8", "likelihoods8.png")

    print
    print "1.2 Generate Threes"
    print

    # use function  naiveBayes_train_single_class to compute the likelihood
    # for all feature dimension

    n = images_train_38.shape[0]
    d = images_train_38.shape[1]

    # Choose bin width
    # Choose bin width
    L, dx = chooseBinSize(images_train_38)

    # train classifier for each class separatly
    p3, pdf3 = naiveBayes_train_single_class(images_train_38, \
                                                  labels_train_38, 3, dx, L)

    # generate 5 new threes
    new3th = np.zeros((5, d), dtype=np.int32)
    for i in range(0, 3):
        new3th[i, :] = generate3naiveBayes(pdf3, dx)

        img = new3th[i, :].reshape(np.sqrt(d), np.sqrt(d))
        plot.figure()
        plot.gray()
        plot.imshow(img)
        plot.show()
    # end for i

    print
    print "2 Density Tree"
    print
    print "Naive splitting"
    print

    # class 3

    tstart = time.time()

    prior3, DT3 = DT_learning(rimages_train_38, labels_train_38, 3, 'naive')
    DT_visualize2D(DT3, rimages_train_38, labels_train_38, 3, "naiveDT3.png")
    # class 8
    prior8, DT8 = DT_learning(rimages_train_38, labels_train_38, 8, 'naive')
    DT_visualize2D(DT8, rimages_train_38, labels_train_38, 8, "naiveDT8.png")

    tstop = time.time()
    print "DT learning time (naive splitting) {}".format(tstop - tstart)

    tstart = time.time()
    rimages_test_38_predict = DT_Classifier_2classes(rimages_test_38, prior3,
                                                     prior8, DT3, DT8, [3, 8])
    tstop = time.time()
    print "DT classification time (naive splitting) {}".format(tstop - tstart)

    ccr_DT = correctClassRate(rimages_test_38_predict,\
                                      labels_test_38, [3,8], \
                                      print_confMatrix = True)

    print 'Correct Classification rate on the test set:{}'.format(ccr_DT)
    print 'Error rate on the test set:{}'.format(1 - ccr_DT)

    print
    print 'Generate new threes: '
    print
    d = images_train_38.shape[1]
    prior3, DT3 = DT_learning(images_train_38, labels_train_38, 3, 'naive')
    # new threes
    new3th = np.zeros((5, d), dtype=np.int32)
    for i in range(0, 3):
        new3th[i, :] = generate3DT(DT3, images_train_38, labels_train_38, 3)
        img = new3th[i, :].reshape(np.sqrt(d), np.sqrt(d))

        plot.figure()
        plot.gray()
        plot.imshow(img)
        plot.show()
    # end for i

    print
    print "Clever splitting"
    print

    #    # class 3
    #    prior3, DT3 = DT_learning(rimages_train_38, labels_train_38, 3, 'clever')
    ##    DT_visualize2D(DT3, rimages_train_38, labels_train_38, 3, "naiveDT3.png")
    #    # class 8
    #    prior8, DT8 = DT_learning(rimages_train_38, labels_train_38, 8, 'clever')
    ##    DT_visu    alize2D(DT8, rimages_train_38, labels_train_38, 8, "naiveDT8.png")
    #
    #
    #    rimages_test_38_predict = DT_Classifier_2classes(rimages_test_38,
    #                                                prior3, prior8, DT3, DT8, [3,8])
    #
    #    ccr_DT = correctClassRate(rimages_test_38_predict,\
    #                                      labels_test_38, [3,8], \
    #                                      print_confMatrix = True)
    #
    #    print 'Correct Classification rate on the test set:{}'.format(ccr_DT)
    #    print 'Error rate on the test set:{}'.format(1-ccr_DT)

    print
    print "3 Combine DT and Naive Bayes"
    print

    n = images_train_38.shape[0]
    d = images_train_38.shape[1]

    print
    print "Learning phase"
    print

    # train 1D-histogramms for each feature and class

    tstart = time.time()
    L, dx = chooseBinSize(images_train_38)  # number of bins, bins size

    # pdf dxL matrices
    prior3, pdf3 = naiveBayes_train_single_class(images_train_38, \
                                                  labels_train_38, 3, dx, L)
    prior8, pdf8 = naiveBayes_train_single_class(images_train_38, \
                                                  labels_train_38, 8, dx, L)
    # compute the cdf of each histogramm
    cdf3 = np.zeros(pdf3.shape, dtype=np.float32)
    cdf8 = np.zeros(pdf8.shape, dtype=np.float32)
    for j in range(0, d):
        cdf3[j, :] = np.cumsum(pdf3[j, :])
        cdf8[j, :] = np.cumsum(pdf8[j, :])
    # end for

    tstop = time.time()
    print "Learning 1D histograms and computing cdf's took {} sec".\
                                                        format(tstop-tstart)

    # map data to copula using rank order transformation
    u = np.zeros(images_train_38.shape, dtype=np.float32)
    for j in range(0, d):
        ind = np.sort(images_train_38[:, j])
        u[:, j] = ind[:] / float(n + 1)
    # end for j

    # train a DT on u
    tstart = time.time()

    prior3, DT3 = DT_learning(u, labels_train_38, 3, 'naive')
    prior8, DT8 = DT_learning(u, labels_train_38, 8, 'naive')

    tstop = time.time()
    print "Learning DTs took {} sec".format(tstop - tstart)

    print
    print "Classification"
    print

    ntest = images_test.shape[0]
    prediction = np.zeros(ntest, dtype=np.int8)

    for i in range(0, n):
        x = images_test[i, :]
        u3 = np.zeros(d, dtype=np.float32)
        u8 = np.zeros(d, dtype=np.float32)

        naiveBayesDensity3 = 1.
        naiveBayesDensity8 = 1.
        for j in range(0, d):
            l = np.floor(x[j] / dx[j]) + 1  # bin number
            if l > L - 1:
                l = L - 1
            naiveBayesDensity3 *= pdf3[j, l]
            naiveBayesDensity8 *= pdf8[j, l]

            u3[j] = cdf3[j, l]
            u8[j] = cdf8[j, l]
        # end for j

        copulaDensity3 = 0
        for node in DT3:
            if point_in_region(u3, node.region):
                copulaDensity3 = node.p
                break
            # end if
        # end for node

        copulaDensity8 = 0
        for node in DT8:
            if point_in_region(u8, node.region):
                copulaDensity8 = node.p
                break
            # end if
        # end for node

        p_y3_x = naiveBayesDensity3 * copulaDensity3 * prior3
        p_y8_x = naiveBayesDensity8 * copulaDensity8 * prior8

        # argmax (p_y3_x, p_y8_x)
        #print p_y3_x
        if p_y3_x > p_y8_x:
            prediction[i] = 3
        else:
            prediction[i] = 8
        # end if
    # end for i

    ccr = correctClassRate(prediction, labels_test_38, [3,8], \
                                      print_confMatrix = True)

    print 'Correct Classification rate on the test set:{}'.format(ccr)
    print 'Error rate on the test set:{}'.format(1 - ccr)
Пример #3
0
def main():
    plot.close('all')  
    
#       0 Read data, selecting digits 3 and 8, dimension reduction
    print
    print "Read data, selecting digits 3 and 8, dimension reduction"
    print
    
    test_path     = "test.h5"
    training_path = "train.h5"
    
    images_train = vigra.readHDF5(test_path, "images")
    labels_train = vigra.readHDF5(test_path, "labels")
    
    images_test = vigra.readHDF5(training_path, "images")
    labels_test = vigra.readHDF5(training_path, "labels")
    
    print 'Size of the training set: {}'. format(np.shape(images_train))
    print np.shape(labels_train)
    print 'Size of the test set: {}'. format(np.shape(images_test))
    print np.shape(labels_test)
    
    # Reshape data

    n = images_train.shape[0]
    d = images_train.shape[1]
    images_train  =  images_train.reshape(n,d*d)
    
    n = images_test.shape[0]
    assert d!=images_test.shape[0], 'Test and training sets have different dim'
    images_test  =  images_test.reshape(n,d*d)

    # Select 3s and 8s 

    ind3 = (labels_train==3)
    ind8 = (labels_train==8)

    images_train_38 = images_train[ind3+ind8]
    labels_train_38 = labels_train[ind3+ind8]
    
    ind3 = (labels_test==3)
    ind8 = (labels_test==8)

    images_test_38 = images_test[ind3+ind8]
    labels_test_38 = labels_test[ind3+ind8]
  
    # Dimension reduction

    rimages_train_38 = dr(images_train_38, labels_train_38, [3,8])
    rimages_test_38 = dr(images_test_38, labels_test_38, [3,8])
    
    print 'Size of the training set of 3s and 8s: {}'. format(np.shape(rimages_train_38))
    print 'Size of the test set of 3s and 8s: {}'. format(np.shape(rimages_test_38))
          
    
    print
    print "1 Naive Bayes"
    print
    print "1.1 Classification"
    print
    
    # Training: priors and likelihood for each d=1,2
    # for each feature and class individual histograms <=> 4 histogramms   
    
    n = rimages_train_38.shape[0]
    d = rimages_train_38.shape[1]
    
    # Choose bin width 
    L, dx = chooseBinSize(rimages_train_38)


    # train classifier for each class separatly                        
    p3, pdf3 = naiveBayes_train_single_class(rimages_train_38, \
                                                  labels_train_38, 3, L, dx)
    p8, pdf8 = naiveBayes_train_single_class(rimages_train_38, \
                                                  labels_train_38, 8, L, dx)
                                                    
    rimages_test_38_predict = naiveBayesClassifier(rimages_test_38, \
                                                p3, p8, pdf3, pdf8, L, dx)
                                                
    ccr_naiveBayes = correctClassRate(rimages_test_38_predict,\
                                      labels_test_38, [3,8], \
                                      print_confMatrix = True)            
                                            
    print 'Correct Classification rate on the test set:{}'.format(ccr_naiveBayes)   
    print 'Error rate on the test set:{}'.format(1-ccr_naiveBayes)   

    
    plot_histogram(pdf3, dx, "Histograms of the class 3", "histograms3.png")
    plot_histogram(pdf8, dx, "Histograms of the class 8", "histograms8.png")
    
    plot_likelihood(pdf3, dx, "Likelihoods of the class 3", "likelihoods3.png")
    plot_likelihood(pdf8, dx, "Likelihoods of the class 8", "likelihoods8.png")

    print
    print "1.2 Generate Threes"
    print
    
    # use function  naiveBayes_train_single_class to compute the likelihood 
    # for all feature dimension

    n = images_train_38.shape[0]
    d = images_train_38.shape[1]
    
    # Choose bin width 
    L, dx = chooseBinSize(images_train_38)

    # train classifier for each class separatly                        
    p3, pdf3 = naiveBayes_train_single_class(images_train_38, \
                                                  labels_train_38, 3, L, dx)    
    # generate 3 new threes
    new3th = np.zeros((3,d), dtype = np.int32)
    for i in range(0,3) :    
        new3th[i,:] = generate3naiveBayes(pdf3, dx)
        
        img = new3th[i,:].reshape(np.sqrt(d),np.sqrt(d))
        plot.figure()
        plot.gray()
        plot.imshow(img);
        plot.show()
    # end for i
        
  
    print
    print "2 Density Tree"
    print
    print "Naive splitting"
    print  

    
    tstart = time.time()
    
    # class 3    
    prior3, DT3 = DT_learning(rimages_train_38, labels_train_38, 3, 'naive')            
    DT_visualize2D(DT3, rimages_train_38, labels_train_38, 3, "naiveDT3.png")        
    # class 8
    prior8, DT8 = DT_learning(rimages_train_38, labels_train_38, 8, 'naive')        
    DT_visualize2D(DT8, rimages_train_38, labels_train_38, 8, "naiveDT8.png")        
    
    tstop = time.time()
    print "DT learning time (naive splitting) {}". format(tstop-tstart)
        
    tstart = time.time()    
    rimages_test_38_predict = DT_Classifier_2classes(rimages_test_38, 
                                                prior3, prior8, DT3, DT8, [3,8])
    tstop = time.time()
    print "DT classification time (naive splitting) {}". format(tstop-tstart)
                                                
    ccr_DT = correctClassRate(rimages_test_38_predict,\
                                      labels_test_38, [3,8], \
                                      print_confMatrix = True)     

    print 'Correct Classification rate on the test set:{}'.format(ccr_DT)   
    print 'Error rate on the test set:{}'.format(1-ccr_DT)                 
      
    print 
    print 'Generate new threes: '
    
    d = images_train_38.shape[1]

    # new threes
    new3th = np.zeros((3,d), dtype = np.int32)    
    for i in range(0,3) :    
        new3th[i,:] = generate3DT(images_train_38, labels_train_38, 3 )
        img = new3th[i,:].reshape(np.sqrt(d),np.sqrt(d))
        
        plot.figure()
        plot.gray()
        plot.imshow(img);
        plot.show()
    # end for i
      
    print
    print "Clever splitting"
    print

#    # class 3    
#    prior3, DT3 = DT_learning(rimages_train_38, labels_train_38, 3, 'clever')        
#    DT_visualize2D(DT3, rimages_train_38, labels_train_38, 3, "naiveDT3.png")        
#    # class 8
#    prior8, DT8 = DT_learning(rimages_train_38, labels_train_38, 8, 'clever')        
#    DT_visualize2D(DT8, rimages_train_38, labels_train_38, 8, "naiveDT8.png")        
#    
#        
#    rimages_test_38_predict = DT_Classifier_2classes(rimages_test_38, 
#                                                prior3, prior8, DT3, DT8, [3,8])
#                                                
#    ccr_DT = correctClassRate(rimages_test_38_predict,\
#                                      labels_test_38, [3,8], \
#                                      print_confMatrix = True)     
#
#    print 'Correct Classification rate on the test set:{}'.format(ccr_DT)   
#    print 'Error rate on the test set:{}'.format(1-ccr_DT)           
#   
#    print   
#    print "Generate Threes"
#    print
    
    print    
    print "3 Combine DT and Naive Bayes"
    print
    
    
    n = images_train_38.shape[0]
    d = images_train_38.shape[1]


    print    
    print "Learning phase"
    print
    
    # train 1D-histogramms for each feature and class      
    
    tstart = time.time() 
    L, dx = chooseBinSize(images_train_38) # number of bins, bins size
    
    # pdf dxL matrices                  
    prior3, pdf3 = naiveBayes_train_single_class(images_train_38, \
                                                  labels_train_38, 3, L, dx)    
    prior8, pdf8 = naiveBayes_train_single_class(images_train_38, \
                                                  labels_train_38, 8, L, dx)    
    # compute the cdf of each histogramm
    cdf3 = [] 
    cdf8 = []
    pdf3sorted = []
    pdf8sorted = []
    for j in range(0, d):
        tmp = np.sort(pdf3[j])
        pdf3sorted.append(tmp)
        cdf3.append(np.cumsum(tmp) )

        tmp = np.sort(pdf8[j])
        pdf8sorted.append(tmp)        
        cdf8.append(np.cumsum(tmp) )
    # end for                                               

    tstop = time.time()
    print "Learning 1D histograms and computing cdf's took {} sec".\
                                                        format(tstop-tstart)
    
    # map data to copula using rank order transformation
    u = np.zeros(images_train_38.shape, dtype = np.float32)
    for j in range(0,d):
        ind = np.argsort(images_train_38[:,j])
        u[:,j] = ind[:]/float(n+1)
    # end for j    
    
    print
    print "Generate threes"
    
    new3th = np.zeros((3,d), dtype = np.int32)  
    for i in range(0,3) :

        u1 = generate3(u, labels_train_38 , 3 )

        for j in range(0,d):
            dist =abs(cdf3[j] - u1[j])
            binx = np.argsort(dist)
            new3th[i,j] = np.floor(dx[j]*binx[0]/2.)
        # end for j
        img = new3th[i,:].reshape(np.sqrt(d),np.sqrt(d))
        
        plot.figure()
        plot.gray()
        plot.imshow(img);
        plot.show()
Пример #4
0
def main():
    plot.close('all')  
    
#       0 Read data, selecting digits 3 and 8, dimension reduction
    print
    print "Read data, selecting digits 3 and 8, dimension reduction"
    print
    
    test_path     = "test.h5"
    training_path = "train.h5"
    
    images_train = vigra.readHDF5(test_path, "images")
    labels_train = vigra.readHDF5(test_path, "labels")
    
    images_test = vigra.readHDF5(training_path, "images")
    labels_test = vigra.readHDF5(training_path, "labels")
    
    print 'Size of the training set: {}'. format(np.shape(images_train))
    print np.shape(labels_train)
    print 'Size of the test set: {}'. format(np.shape(images_test))
    print np.shape(labels_test)
    
    # Reshape data

    n = images_train.shape[0]
    d = images_train.shape[1]
    images_train  =  images_train.reshape(n,d*d)
    
    n = images_test.shape[0]
    assert d!=images_test.shape[0], 'Test and training sets have different dim'
    images_test  =  images_test.reshape(n,d*d)

    # Select 3s and 8s 

    ind3 = (labels_train==3)
    ind8 = (labels_train==8)

    images_train_38 = images_train[ind3+ind8]
    labels_train_38 = labels_train[ind3+ind8]
    
    ind3 = (labels_test==3)
    ind8 = (labels_test==8)

    images_test_38 = images_test[ind3+ind8]
    labels_test_38 = labels_test[ind3+ind8]
  
    # Dimension reduction

    rimages_train_38 = dr(images_train_38, labels_train_38, [3,8])
    rimages_test_38 = dr(images_test_38, labels_test_38, [3,8])
    
    print 'Size of the training set of 3s and 8s: {}'. format(np.shape(rimages_train_38))
    print 'Size of the test set of 3s and 8s: {}'. format(np.shape(rimages_test_38))
          
    
    print
    print "1 Naive Bayes"
    print
    print "1.1 Classification"
    print
    
    # Training: priors and likelihood for each d=1,2
    # for each feature and class individual histograms <=> 4 histogramms   
    
    n = rimages_train_38.shape[0]
    d = rimages_train_38.shape[1]
    
    # Choose bin width 
    L, dx = chooseBinSize(rimages_train_38)

    # train classifier for each class separatly                        
    p3, pdf3 = naiveBayes_train_single_class(rimages_train_38, \
                                                  labels_train_38, 3, dx, L)
    p8, pdf8 = naiveBayes_train_single_class(rimages_train_38, \
                                                  labels_train_38, 8, dx, L)
                                                    
    rimages_test_38_predict = naiveBayesClassifier(rimages_test_38, \
                                                p3, p8, pdf3, pdf8, dx)
                                                
    ccr_naiveBayes = correctClassRate(rimages_test_38_predict,\
                                      labels_test_38, [3,8], \
                                      print_confMatrix = True)            
                                            
    print 'Correct Classification rate on the test set:{}'.format(ccr_naiveBayes)   
    print 'Error rate on the test set:{}'.format(1-ccr_naiveBayes)   


    plot_histogram(pdf3, dx, "Histograms of the class 3", "histograms3.png")
    plot_histogram(pdf8, dx, "Histograms of the class 8", "histograms8.png")
    
    plot_likelihood(pdf3, dx, "Likelihoods of the class 3", "likelihoods3.png")
    plot_likelihood(pdf8, dx, "Likelihoods of the class 8", "likelihoods8.png")


    print
    print "1.2 Generate Threes"
    print
    
    # use function  naiveBayes_train_single_class to compute the likelihood 
    # for all feature dimension

    n = images_train_38.shape[0]
    d = images_train_38.shape[1]
    
    # Choose bin width 
    # Choose bin width 
    L, dx = chooseBinSize(images_train_38)

    # train classifier for each class separatly                        
    p3, pdf3 = naiveBayes_train_single_class(images_train_38, \
                                                  labels_train_38, 3, dx, L)    

    
    # generate 5 new threes
    new3th = np.zeros((5,d), dtype = np.int32)
    for i in range(0,3) :    
        new3th[i,:] = generate3naiveBayes(pdf3, dx)
        
        img = new3th[i,:].reshape(np.sqrt(d),np.sqrt(d))
        plot.figure()
        plot.gray()
        plot.imshow(img);
        plot.show()
    # end for i
        
  
    print
    print "2 Density Tree"
    print
    print "Naive splitting"
    print  

    # class 3    
    
    tstart = time.time()
    
    prior3, DT3 = DT_learning(rimages_train_38, labels_train_38, 3, 'naive')        
    DT_visualize2D(DT3, rimages_train_38, labels_train_38, 3, "naiveDT3.png")        
    # class 8
    prior8, DT8 = DT_learning(rimages_train_38, labels_train_38, 8, 'naive')        
    DT_visualize2D(DT8, rimages_train_38, labels_train_38, 8, "naiveDT8.png")        
    
    tstop = time.time()
    print "DT learning time (naive splitting) {}". format(tstop-tstart)
        
    tstart = time.time()    
    rimages_test_38_predict = DT_Classifier_2classes(rimages_test_38, 
                                                prior3, prior8, DT3, DT8, [3,8])
    tstop = time.time()
    print "DT classification time (naive splitting) {}". format(tstop-tstart)
                                                
    ccr_DT = correctClassRate(rimages_test_38_predict,\
                                      labels_test_38, [3,8], \
                                      print_confMatrix = True)     

    print 'Correct Classification rate on the test set:{}'.format(ccr_DT)   
    print 'Error rate on the test set:{}'.format(1-ccr_DT)           
      
      
      
      
    print 
    print 'Generate new threes: '
    print 
    d = images_train_38.shape[1]
    prior3, DT3 = DT_learning(images_train_38, labels_train_38, 3, 'naive')     
    # new threes
    new3th = np.zeros((5,d), dtype = np.int32)    
    for i in range(0,3) :    
        new3th[i,:] = generate3DT(DT3, images_train_38, labels_train_38, 3 )
        img = new3th[i,:].reshape(np.sqrt(d),np.sqrt(d))
        
        plot.figure()
        plot.gray()
        plot.imshow(img);
        plot.show()
    # end for i
      
    print
    print "Clever splitting"
    print

#    # class 3    
#    prior3, DT3 = DT_learning(rimages_train_38, labels_train_38, 3, 'clever')        
##    DT_visualize2D(DT3, rimages_train_38, labels_train_38, 3, "naiveDT3.png")        
#    # class 8
#    prior8, DT8 = DT_learning(rimages_train_38, labels_train_38, 8, 'clever')        
##    DT_visu    alize2D(DT8, rimages_train_38, labels_train_38, 8, "naiveDT8.png")        
#    
#        
#    rimages_test_38_predict = DT_Classifier_2classes(rimages_test_38, 
#                                                prior3, prior8, DT3, DT8, [3,8])
#                                                
#    ccr_DT = correctClassRate(rimages_test_38_predict,\
#                                      labels_test_38, [3,8], \
#                                      print_confMatrix = True)     
#
#    print 'Correct Classification rate on the test set:{}'.format(ccr_DT)   
#    print 'Error rate on the test set:{}'.format(1-ccr_DT)           
   
    
    print    
    print "3 Combine DT and Naive Bayes"
    print
    
    
    n = images_train_38.shape[0]
    d = images_train_38.shape[1]


    print    
    print "Learning phase"
    print
    
    # train 1D-histogramms for each feature and class      
    
    tstart = time.time() 
    L, dx = chooseBinSize(images_train_38) # number of bins, bins size
    
    # pdf dxL matrices                  
    prior3, pdf3 = naiveBayes_train_single_class(images_train_38, \
                                                  labels_train_38, 3, dx, L)    
    prior8, pdf8 = naiveBayes_train_single_class(images_train_38, \
                                                  labels_train_38, 8, dx, L)    
    # compute the cdf of each histogramm
    cdf3 = np.zeros(pdf3.shape, dtype = np.float32)   
    cdf8 = np.zeros(pdf8.shape, dtype = np.float32)   
    for j in range(0, d):
        cdf3[j,:] = np.cumsum(pdf3[j,:])
        cdf8[j,:] = np.cumsum(pdf8[j,:])
    # end for                                               

    tstop = time.time()
    print "Learning 1D histograms and computing cdf's took {} sec".\
                                                        format(tstop-tstart)
    
    # map data to copula using rank order transformation
    u = np.zeros(images_train_38.shape, dtype = np.float32)
    for j in range(0,d):
        ind = np.sort(images_train_38[:,j])
        u[:,j] = ind[:]/float(n+1)
    # end for j    
    
    # train a DT on u
    tstart = time.time() 
    
    prior3, DT3 = DT_learning(u, labels_train_38, 3, 'naive') 
    prior8, DT8 = DT_learning(u, labels_train_38, 8, 'naive') 
    
    tstop = time.time()
    print "Learning DTs took {} sec". format(tstop-tstart)    
    
    print    
    print "Classification"
    print    
    
    ntest = images_test.shape[0]
    prediction = np.zeros(ntest, dtype = np.int8)
        
    for i in range(0,n):
        x = images_test[i,:]
        u3 = np.zeros(d, dtype = np.float32)        
        u8 = np.zeros(d, dtype = np.float32)                
        
        naiveBayesDensity3 = 1.
        naiveBayesDensity8 = 1.        
        for j in range(0, d):
            l= np.floor(x[j]/dx[j])+1 # bin number
            if l>L-1:
                l=L-1
            naiveBayesDensity3 *= pdf3[j,l]            
            naiveBayesDensity8 *= pdf8[j,l]                      
            
            u3[j]= cdf3[j,l]
            u8[j]= cdf8[j,l]
        # end for j
        
        copulaDensity3 = 0
        for node in DT3:
            if point_in_region(u3, node.region):
                copulaDensity3 = node.p
                break 
            # end if
        # end for node
                
        copulaDensity8 = 0
        for node in DT8:
            if point_in_region(u8, node.region):
                copulaDensity8 = node.p
                break 
            # end if
        # end for node
                
        p_y3_x = naiveBayesDensity3*copulaDensity3*prior3
        p_y8_x = naiveBayesDensity8*copulaDensity8*prior8
        
        # argmax (p_y3_x, p_y8_x) 
        #print p_y3_x
        if p_y3_x>p_y8_x :
            prediction[i] = 3
        else:
            prediction[i] = 8
        # end if        
    # end for i    
 
    ccr = correctClassRate(prediction, labels_test_38, [3,8], \
                                      print_confMatrix = True)     

    print 'Correct Classification rate on the test set:{}'.format(ccr)   
    print 'Error rate on the test set:{}'.format(1-ccr)