# create shogun feature representation features=RealFeatures(data) # compute median data distance in order to use for Gaussian kernel width # 0.5*median_distance normally (factor two in Gaussian kernel) # However, shoguns kernel width is different to usual parametrization # Therefore 0.5*2*median_distance^2 # Use a subset of data for that, only 200 elements. Median is stable # Using all distances here would blow up memory subset=Math.randperm_vec(features.get_num_vectors()) subset=subset[0:200] features.add_subset(subset) dist=EuclideanDistance(features, features) distances=dist.get_distance_matrix() features.remove_subset() median_distance=Statistics.matrix_median(distances, True) sigma=median_distance**2 print "median distance for Gaussian kernel:", sigma kernel=GaussianKernel(10,sigma) # use biased statistic mmd=LinearTimeMMD(kernel,features, m) # sample alternative distribution alt_samples=zeros(num_null_samples) for i in range(len(alt_samples)): data=DataGenerator.generate_mean_data(m,dim,difference) features.set_feature_matrix(data) alt_samples[i]=mmd.compute_statistic() # sample from null distribution
def statistics_quadratic_time_mmd():
from shogun.Features import RealFeatures
from shogun.Features import MeanShiftDataGenerator
from shogun.Kernel import GaussianKernel
from shogun.Statistics import QuadraticTimeMMD
from shogun.Statistics import BOOTSTRAP, MMD2_SPECTRUM, MMD2_GAMMA, BIASED, UNBIASED
from shogun.Distance import EuclideanDistance
from shogun.Mathematics import Statistics, IntVector
# note that the quadratic time mmd has to store kernel matrices
# which upper bounds the sample size
n = 100
dim = 2
difference = 0.5
# streaming data generator for mean shift distributions
gen_p = MeanShiftDataGenerator(0, dim)
gen_q = MeanShiftDataGenerator(difference, dim)
# Stream examples and merge them in order to compute median on joint sample
# alternative is to call a different constructor of QuadraticTimeMMD
features = gen_p.get_streamed_features(n)
features = features.create_merged_copy(gen_q.get_streamed_features(n))
# use data generator class to produce example data
data = features.get_feature_matrix()
print "dimension means of X", mean(data.T[0:n].T)
print "dimension means of Y", mean(data.T[n : 2 * n + 1].T)
# compute median data distance in order to use for Gaussian kernel width
# 0.5*median_distance normally (factor two in Gaussian kernel)
# However, shoguns kernel width is different to usual parametrization
# Therefore 0.5*2*median_distance^2
# Use a subset of data for that, only 200 elements. Median is stable
# Use a permutation set to temporarily merge features in merged examples
subset = IntVector.randperm_vec(features.get_num_vectors())
subset = subset[0:200]
features.add_subset(subset)
dist = EuclideanDistance(features, features)
distances = dist.get_distance_matrix()
features.remove_subset()
median_distance = Statistics.matrix_median(distances, True)
sigma = median_distance ** 2
print "median distance for Gaussian kernel:", sigma
kernel = GaussianKernel(10, sigma)
mmd = QuadraticTimeMMD(kernel, features, n)
# perform test: compute p-value and test if null-hypothesis is rejected for
# a test level of 0.05 using different methods to approximate
# null-distribution
statistic = mmd.compute_statistic()
alpha = 0.05
print "computing p-value using bootstrapping"
mmd.set_null_approximation_method(BOOTSTRAP)
# normally, at least 250 iterations should be done, but that takes long
mmd.set_bootstrap_iterations(10)
# bootstrapping allows usage of unbiased or biased statistic
mmd.set_statistic_type(UNBIASED)
p_value = mmd.compute_p_value(statistic)
print "p_value:", p_value
print "p_value <", alpha, ", i.e. test sais p!=q:", p_value < alpha
# only can do this if SHOGUN was compiled with LAPACK so check
if "sample_null_spectrum" in dir(QuadraticTimeMMD):
print "computing p-value using spectrum method"
mmd.set_null_approximation_method(MMD2_SPECTRUM)
# normally, at least 250 iterations should be done, but that takes long
mmd.set_num_samples_sepctrum(50)
mmd.set_num_eigenvalues_spectrum(n - 10)
# spectrum method computes p-value for biased statistics only
mmd.set_statistic_type(BIASED)
p_value = mmd.compute_p_value(statistic)
print "p_value:", p_value
print "p_value <", alpha, ", i.e. test sais p!=q:", p_value < alpha
print "computing p-value using gamma method"
mmd.set_null_approximation_method(MMD2_GAMMA)
# gamma method computes p-value for biased statistics only
mmd.set_statistic_type(BIASED)
p_value = mmd.compute_p_value(statistic)
print "p_value:", p_value
print "p_value <", alpha, ", i.e. test sais p!=q:", p_value < alpha
# sample from null distribution (these may be plotted or whatsoever)
# mean should be close to zero, variance stronly depends on data/kernel
# bootstrapping, biased statistic
print "sampling null distribution using bootstrapping"
mmd.set_null_approximation_method(BOOTSTRAP)
mmd.set_statistic_type(BIASED)
mmd.set_bootstrap_iterations(10)
null_samples = mmd.bootstrap_null()
print "null mean:", mean(null_samples)
print "null variance:", var(null_samples)
# sample from null distribution (these may be plotted or whatsoever)
# mean should be close to zero, variance stronly depends on data/kernel
# spectrum, biased statistic
print "sampling null distribution using spectrum method"
mmd.set_null_approximation_method(MMD2_SPECTRUM)
mmd.set_statistic_type(BIASED)
# 200 samples using 100 eigenvalues
null_samples = mmd.sample_null_spectrum(50, 10)
print "null mean:", mean(null_samples)
print "null variance:", var(null_samples)
def statistics_quadratic_time_mmd (): from shogun.Features import RealFeatures from shogun.Features import MeanShiftRealDataGenerator from shogun.Kernel import GaussianKernel from shogun.Statistics import QuadraticTimeMMD from shogun.Statistics import BOOTSTRAP, MMD2_SPECTRUM, MMD2_GAMMA, BIASED, UNBIASED from shogun.Distance import EuclideanDistance from shogun.Mathematics import Statistics, IntVector # note that the quadratic time mmd has to store kernel matrices # which upper bounds the sample size n=500 dim=2 difference=0.5 # streaming data generator for mean shift distributions gen_p=MeanShiftRealDataGenerator(0, dim) gen_q=MeanShiftRealDataGenerator(difference, dim) # Stream examples and merge them in order to compute median on joint sample # alternative is to call a different constructor of QuadraticTimeMMD features=gen_p.get_streamed_features(n) features=features.create_merged_copy(gen_q.get_streamed_features(n)) # use data generator class to produce example data data=features.get_feature_matrix() print "dimension means of X", mean(data.T[0:n].T) print "dimension means of Y", mean(data.T[n:2*n+1].T) # compute median data distance in order to use for Gaussian kernel width # 0.5*median_distance normally (factor two in Gaussian kernel) # However, shoguns kernel width is different to usual parametrization # Therefore 0.5*2*median_distance^2 # Use a subset of data for that, only 200 elements. Median is stable # Use a permutation set to temporarily merge features in merged examples subset=IntVector.randperm_vec(features.get_num_vectors()) subset=subset[0:200] features.add_subset(subset) dist=EuclideanDistance(features, features) distances=dist.get_distance_matrix() features.remove_subset() median_distance=Statistics.matrix_median(distances, True) sigma=median_distance**2 print "median distance for Gaussian kernel:", sigma kernel=GaussianKernel(10,sigma) mmd=QuadraticTimeMMD(kernel,features, n) # perform test: compute p-value and test if null-hypothesis is rejected for # a test level of 0.05 using different methods to approximate # null-distribution statistic=mmd.compute_statistic() alpha=0.05 print "computing p-value using bootstrapping" mmd.set_null_approximation_method(BOOTSTRAP) # normally, at least 250 iterations should be done, but that takes long mmd.set_bootstrap_iterations(10) # bootstrapping allows usage of unbiased or biased statistic mmd.set_statistic_type(UNBIASED) p_value=mmd.compute_p_value(statistic) print "p_value:", p_value print "p_value <", alpha, ", i.e. test sais p!=q:", p_value<alpha # only can do this if SHOGUN was compiled with LAPACK so check if "sample_null_spectrum" in dir(QuadraticTimeMMD): print "computing p-value using spectrum method" mmd.set_null_approximation_method(MMD2_SPECTRUM) # normally, at least 250 iterations should be done, but that takes long mmd.set_num_samples_sepctrum(50) mmd.set_num_eigenvalues_spectrum(n-10) # spectrum method computes p-value for biased statistics only mmd.set_statistic_type(BIASED) p_value=mmd.compute_p_value(statistic) print "p_value:", p_value print "p_value <", alpha, ", i.e. test sais p!=q:", p_value<alpha print "computing p-value using gamma method" mmd.set_null_approximation_method(MMD2_GAMMA) # gamma method computes p-value for biased statistics only mmd.set_statistic_type(BIASED) p_value=mmd.compute_p_value(statistic) print "p_value:", p_value print "p_value <", alpha, ", i.e. test sais p!=q:", p_value<alpha # sample from null distribution (these may be plotted or whatsoever) # mean should be close to zero, variance stronly depends on data/kernel # bootstrapping, biased statistic print "sampling null distribution using bootstrapping" mmd.set_null_approximation_method(BOOTSTRAP) mmd.set_statistic_type(BIASED) mmd.set_bootstrap_iterations(10) null_samples=mmd.bootstrap_null() print "null mean:", mean(null_samples) print "null variance:", var(null_samples) # sample from null distribution (these may be plotted or whatsoever) # mean should be close to zero, variance stronly depends on data/kernel # spectrum, biased statistic print "sampling null distribution using spectrum method" mmd.set_null_approximation_method(MMD2_SPECTRUM) mmd.set_statistic_type(BIASED) # 200 samples using 100 eigenvalues null_samples=mmd.sample_null_spectrum(50,10) print "null mean:", mean(null_samples) print "null variance:", var(null_samples)
def statistics_hsic (): from shogun.Features import RealFeatures from shogun.Features import DataGenerator from shogun.Kernel import GaussianKernel from shogun.Statistics import HSIC from shogun.Statistics import BOOTSTRAP, HSIC_GAMMA from shogun.Distance import EuclideanDistance from shogun.Mathematics import Statistics, IntVector # note that the HSIC has to store kernel matrices # which upper bounds the sample size n=250 difference=3 angle=pi/3 # use data generator class to produce example data data=DataGenerator.generate_sym_mix_gauss(n,difference,angle) #plot(data[0], data[1], 'x');show() # create shogun feature representation features_x=RealFeatures(array([data[0]])) features_y=RealFeatures(array([data[1]])) # compute median data distance in order to use for Gaussian kernel width # 0.5*median_distance normally (factor two in Gaussian kernel) # However, shoguns kernel width is different to usual parametrization # Therefore 0.5*2*median_distance^2 # Use a subset of data for that, only 200 elements. Median is stable subset=IntVector.randperm_vec(features_x.get_num_vectors()) subset=subset[0:200] features_x.add_subset(subset) dist=EuclideanDistance(features_x, features_x) distances=dist.get_distance_matrix() features_x.remove_subset() median_distance=Statistics.matrix_median(distances, True) sigma_x=median_distance**2 features_y.add_subset(subset) dist=EuclideanDistance(features_y, features_y) distances=dist.get_distance_matrix() features_y.remove_subset() median_distance=Statistics.matrix_median(distances, True) sigma_y=median_distance**2 print "median distance for Gaussian kernel on x:", sigma_x print "median distance for Gaussian kernel on y:", sigma_y kernel_x=GaussianKernel(10,sigma_x) kernel_y=GaussianKernel(10,sigma_y) hsic=HSIC(kernel_x,kernel_y,features_x,features_y) # perform test: compute p-value and test if null-hypothesis is rejected for # a test level of 0.05 using different methods to approximate # null-distribution statistic=hsic.compute_statistic() print "HSIC:", statistic alpha=0.05 print "computing p-value using bootstrapping" hsic.set_null_approximation_method(BOOTSTRAP) # normally, at least 250 iterations should be done, but that takes long hsic.set_bootstrap_iterations(100) # bootstrapping allows usage of unbiased or biased statistic p_value=hsic.compute_p_value(statistic) thresh=hsic.compute_threshold(alpha) print "p_value:", p_value print "threshold for 0.05 alpha:", thresh print "p_value <", alpha, ", i.e. test sais p and q are dependend:", p_value<alpha print "computing p-value using gamma method" hsic.set_null_approximation_method(HSIC_GAMMA) p_value=hsic.compute_p_value(statistic) thresh=hsic.compute_threshold(alpha) print "p_value:", p_value print "threshold for 0.05 alpha:", thresh print "p_value <", alpha, ", i.e. test sais p and q are dependend::", p_value<alpha # sample from null distribution (these may be plotted or whatsoever) # mean should be close to zero, variance stronly depends on data/kernel # bootstrapping, biased statistic print "sampling null distribution using bootstrapping" hsic.set_null_approximation_method(BOOTSTRAP) hsic.set_bootstrap_iterations(100) null_samples=hsic.bootstrap_null() print "null mean:", mean(null_samples) print "null variance:", var(null_samples)
def statistics_linear_time_mmd (n,dim,difference): from shogun.Features import RealFeatures from shogun.Features import MeanShiftDataGenerator from shogun.Kernel import GaussianKernel from shogun.Statistics import LinearTimeMMD from shogun.Statistics import BOOTSTRAP, MMD1_GAUSSIAN from shogun.Distance import EuclideanDistance from shogun.Mathematics import Statistics, Math # init seed for reproducability Math.init_random(1) # note that the linear time statistic is designed for much larger datasets # so increase to get reasonable results # streaming data generator for mean shift distributions gen_p=MeanShiftDataGenerator(0, dim) gen_q=MeanShiftDataGenerator(difference, dim) # compute median data distance in order to use for Gaussian kernel width # 0.5*median_distance normally (factor two in Gaussian kernel) # However, shoguns kernel width is different to usual parametrization # Therefore 0.5*2*median_distance^2 # Use a subset of data for that, only 200 elements. Median is stable # Stream examples and merge them in order to compute median on joint sample features=gen_p.get_streamed_features(100) features=features.create_merged_copy(gen_q.get_streamed_features(100)) # compute all pairwise distances dist=EuclideanDistance(features, features) distances=dist.get_distance_matrix() # compute median and determine kernel width (using shogun) median_distance=Statistics.matrix_median(distances, True) sigma=median_distance**2 #print "median distance for Gaussian kernel:", sigma kernel=GaussianKernel(10,sigma) # mmd instance using streaming features, blocksize of 10000 mmd=LinearTimeMMD(kernel, gen_p, gen_q, n, 10000) # perform test: compute p-value and test if null-hypothesis is rejected for # a test level of 0.05 statistic=mmd.compute_statistic() #print "test statistic:", statistic # do the same thing using two different way to approximate null-dstribution # bootstrapping and gaussian approximation (ony for really large samples) alpha=0.05 #print "computing p-value using bootstrapping" mmd.set_null_approximation_method(BOOTSTRAP) mmd.set_bootstrap_iterations(50) # normally, far more iterations are needed p_value_boot=mmd.compute_p_value(statistic) #print "p_value_boot:", p_value_boot #print "p_value_boot <", alpha, ", i.e. test sais p!=q:", p_value_boot<alpha #print "computing p-value using gaussian approximation" mmd.set_null_approximation_method(MMD1_GAUSSIAN) p_value_gaussian=mmd.compute_p_value(statistic) #print "p_value_gaussian:", p_value_gaussian #print "p_value_gaussian <", alpha, ", i.e. test sais p!=q:", p_value_gaussian<alpha # sample from null distribution (these may be plotted or whatsoever) # mean should be close to zero, variance stronly depends on data/kernel mmd.set_null_approximation_method(BOOTSTRAP) mmd.set_bootstrap_iterations(10) # normally, far more iterations are needed null_samples=mmd.bootstrap_null() #print "null mean:", mean(null_samples) #print "null variance:", var(null_samples) # compute type I and type II errors for Gaussian approximation # number of trials should be larger to compute tight confidence bounds mmd.set_null_approximation_method(MMD1_GAUSSIAN) num_trials=5; alpha=0.05 # test power typeIerrors=[0 for x in range(num_trials)] typeIIerrors=[0 for x in range(num_trials)] for i in range(num_trials): # this effectively means that p=q - rejecting is tpye I error mmd.set_simulate_h0(True) typeIerrors[i]=mmd.perform_test()>alpha mmd.set_simulate_h0(False) typeIIerrors[i]=mmd.perform_test()>alpha #print "type I error:", mean(typeIerrors), ", type II error:", mean(typeIIerrors) return statistic, p_value_boot, p_value_gaussian, null_samples, typeIerrors, typeIIerrors
def statistics_linear_time_mmd (): from shogun.Features import RealFeatures from shogun.Features import DataGenerator from shogun.Kernel import GaussianKernel from shogun.Statistics import LinearTimeMMD from shogun.Statistics import BOOTSTRAP, MMD1_GAUSSIAN from shogun.Distance import EuclideanDistance from shogun.Mathematics import Statistics, Math # note that the linear time statistic is designed for much larger datasets n=10000 dim=2 difference=0.5 # use data generator class to produce example data # in pratice, this generate data function could be replaced by a method # that obtains data from a stream data=DataGenerator.generate_mean_data(n,dim,difference) print "dimension means of X", mean(data.T[0:n].T) print "dimension means of Y", mean(data.T[n:2*n+1].T) # create shogun feature representation features=RealFeatures(data) # compute median data distance in order to use for Gaussian kernel width # 0.5*median_distance normally (factor two in Gaussian kernel) # However, shoguns kernel width is different to usual parametrization # Therefore 0.5*2*median_distance^2 # Use a subset of data for that, only 200 elements. Median is stable # Using all distances here would blow up memory subset=Math.randperm_vec(features.get_num_vectors()) subset=subset[0:200] features.add_subset(subset) dist=EuclideanDistance(features, features) distances=dist.get_distance_matrix() features.remove_subset() median_distance=Statistics.matrix_median(distances, True) sigma=median_distance**2 print "median distance for Gaussian kernel:", sigma kernel=GaussianKernel(10,sigma) mmd=LinearTimeMMD(kernel,features, n) # perform test: compute p-value and test if null-hypothesis is rejected for # a test level of 0.05 statistic=mmd.compute_statistic() print "test statistic:", statistic # do the same thing using two different way to approximate null-dstribution # bootstrapping and gaussian approximation (ony for really large samples) alpha=0.05 print "computing p-value using bootstrapping" mmd.set_null_approximation_method(BOOTSTRAP) mmd.set_bootstrap_iterations(50) # normally, far more iterations are needed p_value=mmd.compute_p_value(statistic) print "p_value:", p_value print "p_value <", alpha, ", i.e. test sais p!=q:", p_value<alpha print "computing p-value using gaussian approximation" mmd.set_null_approximation_method(MMD1_GAUSSIAN) p_value=mmd.compute_p_value(statistic) print "p_value:", p_value print "p_value <", alpha, ", i.e. test sais p!=q:", p_value<alpha # sample from null distribution (these may be plotted or whatsoever) # mean should be close to zero, variance stronly depends on data/kernel mmd.set_null_approximation_method(BOOTSTRAP) mmd.set_bootstrap_iterations(10) # normally, far more iterations are needed null_samples=mmd.bootstrap_null() print "null mean:", mean(null_samples) print "null variance:", var(null_samples)
def statistics_linear_time_mmd (): from shogun.Features import RealFeatures from shogun.Features import MeanShiftRealDataGenerator from shogun.Kernel import GaussianKernel from shogun.Statistics import LinearTimeMMD from shogun.Statistics import BOOTSTRAP, MMD1_GAUSSIAN from shogun.Distance import EuclideanDistance from shogun.Mathematics import Statistics, Math # note that the linear time statistic is designed for much larger datasets n=10000 dim=2 difference=0.5 # streaming data generator for mean shift distributions gen_p=MeanShiftRealDataGenerator(0, dim) gen_q=MeanShiftRealDataGenerator(difference, dim) # compute median data distance in order to use for Gaussian kernel width # 0.5*median_distance normally (factor two in Gaussian kernel) # However, shoguns kernel width is different to usual parametrization # Therefore 0.5*2*median_distance^2 # Use a subset of data for that, only 200 elements. Median is stable # Stream examples and merge them in order to compute median on joint sample features=gen_p.get_streamed_features(100) features=features.create_merged_copy(gen_q.get_streamed_features(100)) # compute all pairwise distances dist=EuclideanDistance(features, features) distances=dist.get_distance_matrix() # compute median and determine kernel width (using shogun) median_distance=Statistics.matrix_median(distances, True) sigma=median_distance**2 print "median distance for Gaussian kernel:", sigma kernel=GaussianKernel(10,sigma) # mmd instance using streaming features, blocksize of 10000 mmd=LinearTimeMMD(kernel, gen_p, gen_q, n, 10000) # perform test: compute p-value and test if null-hypothesis is rejected for # a test level of 0.05 statistic=mmd.compute_statistic() print "test statistic:", statistic # do the same thing using two different way to approximate null-dstribution # bootstrapping and gaussian approximation (ony for really large samples) alpha=0.05 print "computing p-value using bootstrapping" mmd.set_null_approximation_method(BOOTSTRAP) mmd.set_bootstrap_iterations(50) # normally, far more iterations are needed p_value=mmd.compute_p_value(statistic) print "p_value:", p_value print "p_value <", alpha, ", i.e. test sais p!=q:", p_value<alpha print "computing p-value using gaussian approximation" mmd.set_null_approximation_method(MMD1_GAUSSIAN) p_value=mmd.compute_p_value(statistic) print "p_value:", p_value print "p_value <", alpha, ", i.e. test sais p!=q:", p_value<alpha # sample from null distribution (these may be plotted or whatsoever) # mean should be close to zero, variance stronly depends on data/kernel mmd.set_null_approximation_method(BOOTSTRAP) mmd.set_bootstrap_iterations(10) # normally, far more iterations are needed null_samples=mmd.bootstrap_null() print "null mean:", mean(null_samples) print "null variance:", var(null_samples)
def hsic_graphical():
# parameters, change to get different results
m=250
difference=3
# setting the angle lower makes a harder test
angle=pi/30
# number of samples taken from null and alternative distribution
num_null_samples=500
# use data generator class to produce example data
data=DataGenerator.generate_sym_mix_gauss(m,difference,angle)
# create shogun feature representation
features_x=RealFeatures(array([data[0]]))
features_y=RealFeatures(array([data[1]]))
# compute median data distance in order to use for Gaussian kernel width
# 0.5*median_distance normally (factor two in Gaussian kernel)
# However, shoguns kernel width is different to usual parametrization
# Therefore 0.5*2*median_distance^2
# Use a subset of data for that, only 200 elements. Median is stable
subset=int32(array([x for x in range(features_x.get_num_vectors())])) # numpy
subset=random.permutation(subset) # numpy permutation
subset=subset[0:200]
features_x.add_subset(subset)
dist=EuclideanDistance(features_x, features_x)
distances=dist.get_distance_matrix()
features_x.remove_subset()
median_distance=Statistics.matrix_median(distances, True)
sigma_x=median_distance**2
features_y.add_subset(subset)
dist=EuclideanDistance(features_y, features_y)
distances=dist.get_distance_matrix()
features_y.remove_subset()
median_distance=Statistics.matrix_median(distances, True)
sigma_y=median_distance**2
print "median distance for Gaussian kernel on x:", sigma_x
print "median distance for Gaussian kernel on y:", sigma_y
kernel_x=GaussianKernel(10,sigma_x)
kernel_y=GaussianKernel(10,sigma_y)
# create hsic instance. Note that this is a convienience constructor which copies
# feature data. features_x and features_y are not these used in hsic.
# This is only for user-friendlyness. Usually, its ok to do this.
# Below, the alternative distribution is sampled, which means
# that new feature objects have to be created in each iteration (slow)
# However, normally, the alternative distribution is not sampled
hsic=HSIC(kernel_x,kernel_y,features_x,features_y)
# sample alternative distribution
alt_samples=zeros(num_null_samples)
for i in range(len(alt_samples)):
data=DataGenerator.generate_sym_mix_gauss(m,difference,angle)
features_x.set_feature_matrix(array([data[0]]))
features_y.set_feature_matrix(array([data[1]]))
# re-create hsic instance everytime since feature objects are copied due to
# useage of convienience constructor
hsic=HSIC(kernel_x,kernel_y,features_x,features_y)
alt_samples[i]=hsic.compute_statistic()
# sample from null distribution
# bootstrapping, biased statistic
hsic.set_null_approximation_method(BOOTSTRAP)
hsic.set_bootstrap_iterations(num_null_samples)
null_samples_boot=hsic.bootstrap_null()
# fit gamma distribution, biased statistic
hsic.set_null_approximation_method(HSIC_GAMMA)
gamma_params=hsic.fit_null_gamma()
# sample gamma with parameters
null_samples_gamma=array([gamma(gamma_params[0], gamma_params[1]) for _ in range(num_null_samples)])
# plot
figure()
# plot data x and y
subplot(2,2,1)
gca().xaxis.set_major_locator( MaxNLocator(nbins = 4) ) # reduce number of x-ticks
gca().yaxis.set_major_locator( MaxNLocator(nbins = 4) ) # reduce number of x-ticks
grid(True)
plot(data[0], data[1], 'o')
title('Data, rotation=$\pi$/'+str(1/angle*pi)+'\nm='+str(m))
xlabel('$x$')
ylabel('$y$')
# compute threshold for test level
alpha=0.05
null_samples_boot.sort()
null_samples_gamma.sort()
thresh_boot=null_samples_boot[floor(len(null_samples_boot)*(1-alpha))];
thresh_gamma=null_samples_gamma[floor(len(null_samples_gamma)*(1-alpha))];
type_one_error_boot=sum(null_samples_boot<thresh_boot)/float(num_null_samples)
type_one_error_gamma=sum(null_samples_gamma<thresh_boot)/float(num_null_samples)
# plot alternative distribution with threshold
subplot(2,2,2)
gca().xaxis.set_major_locator( MaxNLocator(nbins = 3) ) # reduce number of x-ticks
gca().yaxis.set_major_locator( MaxNLocator(nbins = 3) ) # reduce number of x-ticks
grid(True)
hist(alt_samples, 20, normed=True);
axvline(thresh_boot, 0, 1, linewidth=2, color='red')
type_two_error=sum(alt_samples<thresh_boot)/float(num_null_samples)
title('Alternative Dist.\n' + 'Type II error is ' + str(type_two_error))
# compute range for all null distribution histograms
hist_range=[min([min(null_samples_boot), min(null_samples_gamma)]), max([max(null_samples_boot), max(null_samples_gamma)])]
# plot null distribution with threshold
subplot(2,2,3)
gca().xaxis.set_major_locator( MaxNLocator(nbins = 3) ) # reduce number of x-ticks
gca().yaxis.set_major_locator( MaxNLocator(nbins = 3) ) # reduce number of x-ticks
grid(True)
hist(null_samples_boot, 20, range=hist_range, normed=True);
axvline(thresh_boot, 0, 1, linewidth=2, color='red')
title('Bootstrapped Null Dist.\n' + 'Type I error is ' + str(type_one_error_boot))
# plot null distribution gamma
subplot(2,2,4)
gca().xaxis.set_major_locator( MaxNLocator(nbins = 3) ) # reduce number of x-ticks
gca().yaxis.set_major_locator( MaxNLocator(nbins = 3) ) # reduce number of x-ticks
grid(True)
hist(null_samples_gamma, 20, range=hist_range, normed=True);
axvline(thresh_gamma, 0, 1, linewidth=2, color='red')
title('Null Dist. Gamma\nType I error is ' + str(type_one_error_gamma))
grid(True)
# pull plots a bit apart
subplots_adjust(hspace=0.5)
subplots_adjust(wspace=0.5)
# create shogun feature representation features_x=RealFeatures(array([data[0]])) features_y=RealFeatures(array([data[1]])) # compute median data distance in order to use for Gaussian kernel width # 0.5*median_distance normally (factor two in Gaussian kernel) # However, shoguns kernel width is different to usual parametrization # Therefore 0.5*2*median_distance^2 # Use a subset of data for that, only 200 elements. Median is stable subset=Math.randperm_vec(features_x.get_num_vectors()) subset=subset[0:200] features_x.add_subset(subset) dist=EuclideanDistance(features_x, features_x) distances=dist.get_distance_matrix() features_x.remove_subset() median_distance=Statistics.matrix_median(distances, True) sigma_x=median_distance**2 features_y.add_subset(subset) dist=EuclideanDistance(features_y, features_y) distances=dist.get_distance_matrix() features_y.remove_subset() median_distance=Statistics.matrix_median(distances, True) sigma_y=median_distance**2 print "median distance for Gaussian kernel on x:", sigma_x print "median distance for Gaussian kernel on y:", sigma_y kernel_x=GaussianKernel(10,sigma_x) kernel_y=GaussianKernel(10,sigma_y) # create hsic instance. Note that this is a convienience constructor which copies # feature data. features_x and features_y are not these used in hsic. # This is only for user-friendlyness. Usually, its ok to do this.
def hsic_graphical():
# parameters, change to get different results
m = 250
difference = 3
# setting the angle lower makes a harder test
angle = pi / 30
# number of samples taken from null and alternative distribution
num_null_samples = 500
# use data generator class to produce example data
data = DataGenerator.generate_sym_mix_gauss(m, difference, angle)
# create shogun feature representation
features_x = RealFeatures(array([data[0]]))
features_y = RealFeatures(array([data[1]]))
# compute median data distance in order to use for Gaussian kernel width
# 0.5*median_distance normally (factor two in Gaussian kernel)
# However, shoguns kernel width is different to usual parametrization
# Therefore 0.5*2*median_distance^2
# Use a subset of data for that, only 200 elements. Median is stable
subset = int32(array([x for x in range(features_x.get_num_vectors())
])) # numpy
subset = random.permutation(subset) # numpy permutation
subset = subset[0:200]
features_x.add_subset(subset)
dist = EuclideanDistance(features_x, features_x)
distances = dist.get_distance_matrix()
features_x.remove_subset()
median_distance = Statistics.matrix_median(distances, True)
sigma_x = median_distance**2
features_y.add_subset(subset)
dist = EuclideanDistance(features_y, features_y)
distances = dist.get_distance_matrix()
features_y.remove_subset()
median_distance = Statistics.matrix_median(distances, True)
sigma_y = median_distance**2
print "median distance for Gaussian kernel on x:", sigma_x
print "median distance for Gaussian kernel on y:", sigma_y
kernel_x = GaussianKernel(10, sigma_x)
kernel_y = GaussianKernel(10, sigma_y)
# create hsic instance. Note that this is a convienience constructor which copies
# feature data. features_x and features_y are not these used in hsic.
# This is only for user-friendlyness. Usually, its ok to do this.
# Below, the alternative distribution is sampled, which means
# that new feature objects have to be created in each iteration (slow)
# However, normally, the alternative distribution is not sampled
hsic = HSIC(kernel_x, kernel_y, features_x, features_y)
# sample alternative distribution
alt_samples = zeros(num_null_samples)
for i in range(len(alt_samples)):
data = DataGenerator.generate_sym_mix_gauss(m, difference, angle)
features_x.set_feature_matrix(array([data[0]]))
features_y.set_feature_matrix(array([data[1]]))
# re-create hsic instance everytime since feature objects are copied due to
# useage of convienience constructor
hsic = HSIC(kernel_x, kernel_y, features_x, features_y)
alt_samples[i] = hsic.compute_statistic()
# sample from null distribution
# bootstrapping, biased statistic
hsic.set_null_approximation_method(BOOTSTRAP)
hsic.set_bootstrap_iterations(num_null_samples)
null_samples_boot = hsic.bootstrap_null()
# fit gamma distribution, biased statistic
hsic.set_null_approximation_method(HSIC_GAMMA)
gamma_params = hsic.fit_null_gamma()
# sample gamma with parameters
null_samples_gamma = array([
gamma(gamma_params[0], gamma_params[1])
for _ in range(num_null_samples)
])
# plot
figure()
# plot data x and y
subplot(2, 2, 1)
gca().xaxis.set_major_locator(
MaxNLocator(nbins=4)) # reduce number of x-ticks
gca().yaxis.set_major_locator(
MaxNLocator(nbins=4)) # reduce number of x-ticks
grid(True)
plot(data[0], data[1], 'o')
title('Data, rotation=$\pi$/' + str(1 / angle * pi) + '\nm=' + str(m))
xlabel('$x$')
ylabel('$y$')
# compute threshold for test level
alpha = 0.05
null_samples_boot.sort()
null_samples_gamma.sort()
thresh_boot = null_samples_boot[floor(
len(null_samples_boot) * (1 - alpha))]
thresh_gamma = null_samples_gamma[floor(
len(null_samples_gamma) * (1 - alpha))]
type_one_error_boot = sum(
null_samples_boot < thresh_boot) / float(num_null_samples)
type_one_error_gamma = sum(
null_samples_gamma < thresh_boot) / float(num_null_samples)
# plot alternative distribution with threshold
subplot(2, 2, 2)
gca().xaxis.set_major_locator(
MaxNLocator(nbins=3)) # reduce number of x-ticks
gca().yaxis.set_major_locator(
MaxNLocator(nbins=3)) # reduce number of x-ticks
grid(True)
hist(alt_samples, 20, normed=True)
axvline(thresh_boot, 0, 1, linewidth=2, color='red')
type_two_error = sum(alt_samples < thresh_boot) / float(num_null_samples)
title('Alternative Dist.\n' + 'Type II error is ' + str(type_two_error))
# compute range for all null distribution histograms
hist_range = [
min([min(null_samples_boot),
min(null_samples_gamma)]),
max([max(null_samples_boot),
max(null_samples_gamma)])
]
# plot null distribution with threshold
subplot(2, 2, 3)
gca().xaxis.set_major_locator(
MaxNLocator(nbins=3)) # reduce number of x-ticks
gca().yaxis.set_major_locator(
MaxNLocator(nbins=3)) # reduce number of x-ticks
grid(True)
hist(null_samples_boot, 20, range=hist_range, normed=True)
axvline(thresh_boot, 0, 1, linewidth=2, color='red')
title('Bootstrapped Null Dist.\n' + 'Type I error is ' +
str(type_one_error_boot))
# plot null distribution gamma
subplot(2, 2, 4)
gca().xaxis.set_major_locator(
MaxNLocator(nbins=3)) # reduce number of x-ticks
gca().yaxis.set_major_locator(
MaxNLocator(nbins=3)) # reduce number of x-ticks
grid(True)
hist(null_samples_gamma, 20, range=hist_range, normed=True)
axvline(thresh_gamma, 0, 1, linewidth=2, color='red')
title('Null Dist. Gamma\nType I error is ' + str(type_one_error_gamma))
grid(True)
# pull plots a bit apart
subplots_adjust(hspace=0.5)
subplots_adjust(wspace=0.5)
def statistics_linear_time_mmd():
from shogun.Features import RealFeatures
from shogun.Features import MeanShiftRealDataGenerator
from shogun.Kernel import GaussianKernel
from shogun.Statistics import LinearTimeMMD
from shogun.Statistics import BOOTSTRAP, MMD1_GAUSSIAN
from shogun.Distance import EuclideanDistance
from shogun.Mathematics import Statistics, Math
# note that the linear time statistic is designed for much larger datasets
n = 10000
dim = 2
difference = 0.5
# streaming data generator for mean shift distributions
gen_p = MeanShiftRealDataGenerator(0, dim)
gen_q = MeanShiftRealDataGenerator(difference, dim)
# compute median data distance in order to use for Gaussian kernel width
# 0.5*median_distance normally (factor two in Gaussian kernel)
# However, shoguns kernel width is different to usual parametrization
# Therefore 0.5*2*median_distance^2
# Use a subset of data for that, only 200 elements. Median is stable
# Stream examples and merge them in order to compute median on joint sample
features = gen_p.get_streamed_features(100)
features = features.create_merged_copy(gen_q.get_streamed_features(100))
# compute all pairwise distances
dist = EuclideanDistance(features, features)
distances = dist.get_distance_matrix()
# compute median and determine kernel width (using shogun)
median_distance = Statistics.matrix_median(distances, True)
sigma = median_distance**2
print "median distance for Gaussian kernel:", sigma
kernel = GaussianKernel(10, sigma)
# mmd instance using streaming features, blocksize of 10000
mmd = LinearTimeMMD(kernel, gen_p, gen_q, n, 10000)
# perform test: compute p-value and test if null-hypothesis is rejected for
# a test level of 0.05
statistic = mmd.compute_statistic()
print "test statistic:", statistic
# do the same thing using two different way to approximate null-dstribution
# bootstrapping and gaussian approximation (ony for really large samples)
alpha = 0.05
print "computing p-value using bootstrapping"
mmd.set_null_approximation_method(BOOTSTRAP)
mmd.set_bootstrap_iterations(
50) # normally, far more iterations are needed
p_value = mmd.compute_p_value(statistic)
print "p_value:", p_value
print "p_value <", alpha, ", i.e. test sais p!=q:", p_value < alpha
print "computing p-value using gaussian approximation"
mmd.set_null_approximation_method(MMD1_GAUSSIAN)
p_value = mmd.compute_p_value(statistic)
print "p_value:", p_value
print "p_value <", alpha, ", i.e. test sais p!=q:", p_value < alpha
# sample from null distribution (these may be plotted or whatsoever)
# mean should be close to zero, variance stronly depends on data/kernel
mmd.set_null_approximation_method(BOOTSTRAP)
mmd.set_bootstrap_iterations(
10) # normally, far more iterations are needed
null_samples = mmd.bootstrap_null()
print "null mean:", mean(null_samples)
print "null variance:", var(null_samples)
def statistics_hsic (n, difference, angle): from shogun.Features import RealFeatures from shogun.Features import DataGenerator from shogun.Kernel import GaussianKernel from shogun.Statistics import HSIC from shogun.Statistics import BOOTSTRAP, HSIC_GAMMA from shogun.Distance import EuclideanDistance from shogun.Mathematics import Math, Statistics, IntVector # init seed for reproducability Math.init_random(1) # note that the HSIC has to store kernel matrices # which upper bounds the sample size # use data generator class to produce example data data=DataGenerator.generate_sym_mix_gauss(n,difference,angle) #plot(data[0], data[1], 'x');show() # create shogun feature representation features_x=RealFeatures(array([data[0]])) features_y=RealFeatures(array([data[1]])) # compute median data distance in order to use for Gaussian kernel width # 0.5*median_distance normally (factor two in Gaussian kernel) # However, shoguns kernel width is different to usual parametrization # Therefore 0.5*2*median_distance^2 # Use a subset of data for that, only 200 elements. Median is stable subset=IntVector.randperm_vec(features_x.get_num_vectors()) subset=subset[0:200] features_x.add_subset(subset) dist=EuclideanDistance(features_x, features_x) distances=dist.get_distance_matrix() features_x.remove_subset() median_distance=Statistics.matrix_median(distances, True) sigma_x=median_distance**2 features_y.add_subset(subset) dist=EuclideanDistance(features_y, features_y) distances=dist.get_distance_matrix() features_y.remove_subset() median_distance=Statistics.matrix_median(distances, True) sigma_y=median_distance**2 #print "median distance for Gaussian kernel on x:", sigma_x #print "median distance for Gaussian kernel on y:", sigma_y kernel_x=GaussianKernel(10,sigma_x) kernel_y=GaussianKernel(10,sigma_y) hsic=HSIC(kernel_x,kernel_y,features_x,features_y) # perform test: compute p-value and test if null-hypothesis is rejected for # a test level of 0.05 using different methods to approximate # null-distribution statistic=hsic.compute_statistic() #print "HSIC:", statistic alpha=0.05 #print "computing p-value using bootstrapping" hsic.set_null_approximation_method(BOOTSTRAP) # normally, at least 250 iterations should be done, but that takes long hsic.set_bootstrap_iterations(100) # bootstrapping allows usage of unbiased or biased statistic p_value_boot=hsic.compute_p_value(statistic) thresh_boot=hsic.compute_threshold(alpha) #print "p_value:", p_value_boot #print "threshold for 0.05 alpha:", thresh_boot #print "p_value <", alpha, ", i.e. test sais p and q are dependend:", p_value_boot<alpha #print "computing p-value using gamma method" hsic.set_null_approximation_method(HSIC_GAMMA) p_value_gamma=hsic.compute_p_value(statistic) thresh_gamma=hsic.compute_threshold(alpha) #print "p_value:", p_value_gamma #print "threshold for 0.05 alpha:", thresh_gamma #print "p_value <", alpha, ", i.e. test sais p and q are dependend::", p_value_gamma<alpha # sample from null distribution (these may be plotted or whatsoever) # mean should be close to zero, variance stronly depends on data/kernel # bootstrapping, biased statistic #print "sampling null distribution using bootstrapping" hsic.set_null_approximation_method(BOOTSTRAP) hsic.set_bootstrap_iterations(100) null_samples=hsic.bootstrap_null() #print "null mean:", mean(null_samples) #print "null variance:", var(null_samples) #hist(null_samples, 100); show() return p_value_boot, thresh_boot, p_value_gamma, thresh_gamma, statistic, null_samples
def statistics_linear_time_mmd(n, dim, difference):
from shogun.Features import RealFeatures
from shogun.Features import MeanShiftDataGenerator
from shogun.Kernel import GaussianKernel
from shogun.Statistics import LinearTimeMMD
from shogun.Statistics import BOOTSTRAP, MMD1_GAUSSIAN
from shogun.Distance import EuclideanDistance
from shogun.Mathematics import Statistics, Math
# init seed for reproducability
Math.init_random(1)
# note that the linear time statistic is designed for much larger datasets
# so increase to get reasonable results
# streaming data generator for mean shift distributions
gen_p = MeanShiftDataGenerator(0, dim)
gen_q = MeanShiftDataGenerator(difference, dim)
# compute median data distance in order to use for Gaussian kernel width
# 0.5*median_distance normally (factor two in Gaussian kernel)
# However, shoguns kernel width is different to usual parametrization
# Therefore 0.5*2*median_distance^2
# Use a subset of data for that, only 200 elements. Median is stable
# Stream examples and merge them in order to compute median on joint sample
features = gen_p.get_streamed_features(100)
features = features.create_merged_copy(gen_q.get_streamed_features(100))
# compute all pairwise distances
dist = EuclideanDistance(features, features)
distances = dist.get_distance_matrix()
# compute median and determine kernel width (using shogun)
median_distance = Statistics.matrix_median(distances, True)
sigma = median_distance**2
#print "median distance for Gaussian kernel:", sigma
kernel = GaussianKernel(10, sigma)
# mmd instance using streaming features, blocksize of 10000
mmd = LinearTimeMMD(kernel, gen_p, gen_q, n, 10000)
# perform test: compute p-value and test if null-hypothesis is rejected for
# a test level of 0.05
statistic = mmd.compute_statistic()
#print "test statistic:", statistic
# do the same thing using two different way to approximate null-dstribution
# bootstrapping and gaussian approximation (ony for really large samples)
alpha = 0.05
#print "computing p-value using bootstrapping"
mmd.set_null_approximation_method(BOOTSTRAP)
mmd.set_bootstrap_iterations(
50) # normally, far more iterations are needed
p_value_boot = mmd.compute_p_value(statistic)
#print "p_value_boot:", p_value_boot
#print "p_value_boot <", alpha, ", i.e. test sais p!=q:", p_value_boot<alpha
#print "computing p-value using gaussian approximation"
mmd.set_null_approximation_method(MMD1_GAUSSIAN)
p_value_gaussian = mmd.compute_p_value(statistic)
#print "p_value_gaussian:", p_value_gaussian
#print "p_value_gaussian <", alpha, ", i.e. test sais p!=q:", p_value_gaussian<alpha
# sample from null distribution (these may be plotted or whatsoever)
# mean should be close to zero, variance stronly depends on data/kernel
mmd.set_null_approximation_method(BOOTSTRAP)
mmd.set_bootstrap_iterations(
10) # normally, far more iterations are needed
null_samples = mmd.bootstrap_null()
#print "null mean:", mean(null_samples)
#print "null variance:", var(null_samples)
# compute type I and type II errors for Gaussian approximation
# number of trials should be larger to compute tight confidence bounds
mmd.set_null_approximation_method(MMD1_GAUSSIAN)
num_trials = 5
alpha = 0.05 # test power
typeIerrors = [0 for x in range(num_trials)]
typeIIerrors = [0 for x in range(num_trials)]
for i in range(num_trials):
# this effectively means that p=q - rejecting is tpye I error
mmd.set_simulate_h0(True)
typeIerrors[i] = mmd.perform_test() > alpha
mmd.set_simulate_h0(False)
typeIIerrors[i] = mmd.perform_test() > alpha
#print "type I error:", mean(typeIerrors), ", type II error:", mean(typeIIerrors)
return statistic, p_value_boot, p_value_gaussian, null_samples, typeIerrors, typeIIerrors
