def statistics_linear_time_mmd_kernel_choice():
from shogun.Features import RealFeatures, CombinedFeatures
from shogun.Features import MeanShiftRealDataGenerator
from shogun.Kernel import GaussianKernel, CombinedKernel
from shogun.Statistics import LinearTimeMMD
from shogun.Statistics import BOOTSTRAP, MMD1_GAUSSIAN
# note that the linear time statistic is designed for much larger datasets
n = 50000
dim = 5
difference = 2
# 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 kernels/features to choose from
# here: just a bunch of Gaussian Kernels with different widths
# real sigmas are 2^-5, ..., 2^10
sigmas = array([pow(2, x) for x in range(-5, 10)])
# shogun has a different parametrization of the Gaussian kernel
shogun_sigmas = array([x * x * 2 for x in sigmas])
# We will use multiple kernels
kernel = CombinedKernel()
# two separate feature objects here, could also be one with appended data
features = CombinedFeatures()
# all kernels work on same features
for i in range(len(sigmas)):
kernel.append_kernel(GaussianKernel(10, shogun_sigmas[i]))
features.append_feature_obj(RealFeatures(data))
mmd = LinearTimeMMD(kernel, features, n)
print "start learning kernel weights"
mmd.set_opt_regularization_eps(10E-5)
mmd.set_opt_low_cut(10E-5)
mmd.set_opt_max_iterations(1000)
mmd.set_opt_epsilon(10E-7)
mmd.optimize_kernel_weights()
weights = kernel.get_subkernel_weights()
print "learned weights:", weights
#pyplot.plot(array(range(len(sigmas))), weights)
#pyplot.show()
print "index of max weight", weights.argmax()
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)
# 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
# bootstrapping, biased statistic
mmd.set_null_approximation_method(BOOTSTRAP)
mmd.set_bootstrap_iterations(num_null_samples)
null_samples_boot = mmd.bootstrap_null()
# fit normal distribution to null and sample a normal distribution
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_mmd_kernel_selection_combined(m, distance, stretch, num_blobs,
angle, selection_method):
from shogun.Features import RealFeatures
from shogun.Features import GaussianBlobsDataGenerator
from shogun.Kernel import GaussianKernel, CombinedKernel
from shogun.Statistics import LinearTimeMMD
from shogun.Statistics import MMDKernelSelectionCombMaxL2
from shogun.Statistics import MMDKernelSelectionCombOpt
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
# results for this low number will be bad (unstable, type I error wrong)
# streaming data generator
gen_p = GaussianBlobsDataGenerator(num_blobs, distance, 1, 0)
gen_q = GaussianBlobsDataGenerator(num_blobs, distance, stretch, angle)
# stream some data and plot
num_plot = 1000
features = gen_p.get_streamed_features(num_plot)
features = features.create_merged_copy(
gen_q.get_streamed_features(num_plot))
data = features.get_feature_matrix()
#figure()
#subplot(2,2,1)
#grid(True)
#plot(data[0][0:num_plot], data[1][0:num_plot], 'r.', label='$x$')
#title('$X\sim p$')
#subplot(2,2,2)
#grid(True)
#plot(data[0][num_plot+1:2*num_plot], data[1][num_plot+1:2*num_plot], 'b.', label='$x$', alpha=0.5)
#title('$Y\sim q$')
# create combined kernel with Gaussian kernels inside (shoguns Gaussian kernel is
# different to the standard form, see documentation)
sigmas = [2**x for x in range(-3, 10)]
widths = [x * x * 2 for x in sigmas]
combined = CombinedKernel()
for i in range(len(sigmas)):
combined.append_kernel(GaussianKernel(10, widths[i]))
# mmd instance using streaming features, blocksize of 10000
block_size = 10000
mmd = LinearTimeMMD(combined, gen_p, gen_q, m, block_size)
# kernel selection instance (this can easily replaced by the other methods for selecting
# combined kernels
if selection_method == "opt":
selection = MMDKernelSelectionCombOpt(mmd)
elif selection_method == "l2":
selection = MMDKernelSelectionCombMaxL2(mmd)
# perform kernel selection (kernel is automatically set)
kernel = selection.select_kernel()
kernel = CombinedKernel.obtain_from_generic(kernel)
#print "selected kernel weights:", kernel.get_subkernel_weights()
#subplot(2,2,3)
#plot(kernel.get_subkernel_weights())
#title("Kernel weights")
# compute tpye I and II error (use many more trials). Type I error is only
# estimated to check MMD1_GAUSSIAN method for estimating the null
# distribution. Note that testing has to happen on difference data than
# kernel selecting, but the linear time mmd does this implicitly
mmd.set_null_approximation_method(MMD1_GAUSSIAN)
# number of trials should be larger to compute tight confidence bounds
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 kernel, typeIerrors, typeIIerrors
# 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, m, 10000) # sample alternative distribution, stream ensures different samples each run alt_samples=zeros(num_null_samples) for i in range(len(alt_samples)): alt_samples[i]=mmd.compute_statistic() # sample from null distribution # bootstrapping, biased statistic mmd.set_null_approximation_method(BOOTSTRAP) mmd.set_bootstrap_iterations(num_null_samples) null_samples_boot=mmd.bootstrap_null() # fit normal distribution to null and sample a normal distribution mmd.set_null_approximation_method(MMD1_GAUSSIAN) variance=mmd.compute_variance_estimate()
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 linear_time_mmd_graphical():
# parameters, change to get different results
m = 1000 # set to 10000 for a good test result
dim = 2
# setting the difference of the first dimension smaller makes a harder test
difference = 1
# number of samples taken from null and alternative distribution
num_null_samples = 150
# streaming data generator for mean shift distributions
gen_p = MeanShiftDataGenerator(0, dim)
gen_q = MeanShiftDataGenerator(difference, dim)
# use the median kernel selection
# create combined kernel with Gaussian kernels inside (shoguns Gaussian kernel is
# 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
sigmas = [2**x for x in range(-3, 10)]
widths = [x * x * 2 for x in sigmas]
print "kernel widths:", widths
combined = CombinedKernel()
for i in range(len(sigmas)):
combined.append_kernel(GaussianKernel(10, widths[i]))
# mmd instance using streaming features, blocksize of 10000
block_size = 1000
mmd = LinearTimeMMD(combined, gen_p, gen_q, m, block_size)
# kernel selection instance (this can easily replaced by the other methods for selecting
# single kernels
selection = MMDKernelSelectionOpt(mmd)
# perform kernel selection
kernel = selection.select_kernel()
kernel = GaussianKernel.obtain_from_generic(kernel)
mmd.set_kernel(kernel)
print "selected kernel width:", kernel.get_width()
# sample alternative distribution, stream ensures different samples each run
alt_samples = zeros(num_null_samples)
for i in range(len(alt_samples)):
alt_samples[i] = mmd.compute_statistic()
# sample from null distribution
# bootstrapping, biased statistic
mmd.set_null_approximation_method(BOOTSTRAP)
mmd.set_bootstrap_iterations(num_null_samples)
null_samples_boot = mmd.bootstrap_null()
# fit normal distribution to null and sample a normal distribution
mmd.set_null_approximation_method(MMD1_GAUSSIAN)
variance = mmd.compute_variance_estimate()
null_samples_gaussian = normal(0, sqrt(variance), num_null_samples)
# to plot data, sample a few examples from stream first
features = gen_p.get_streamed_features(m)
features = features.create_merged_copy(gen_q.get_streamed_features(m))
data = features.get_feature_matrix()
# plot
figure()
# plot data of p and q
subplot(2, 3, 1)
grid(True)
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
plot(data[0][0:m], data[1][0:m], 'ro', label='$x$')
plot(data[0][m + 1:2 * m],
data[1][m + 1:2 * m],
'bo',
label='$x$',
alpha=0.5)
title('Data, shift in $x_1$=' + str(difference) + '\nm=' + str(m))
xlabel('$x_1, y_1$')
ylabel('$x_2, y_2$')
# histogram of first data dimension and pdf
subplot(2, 3, 2)
grid(True)
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
hist(data[0], bins=50, alpha=0.5, facecolor='r', normed=True)
hist(data[1], bins=50, alpha=0.5, facecolor='b', normed=True)
xs = linspace(min(data[0]) - 1, max(data[0]) + 1, 50)
plot(xs, normpdf(xs, 0, 1), 'r', linewidth=3)
plot(xs, normpdf(xs, difference, 1), 'b', linewidth=3)
xlabel('$x_1, y_1$')
ylabel('$p(x_1), p(y_1)$')
title('Data PDF in $x_1, y_1$')
# compute threshold for test level
alpha = 0.05
null_samples_boot.sort()
null_samples_gaussian.sort()
thresh_boot = null_samples_boot[floor(
len(null_samples_boot) * (1 - alpha))]
thresh_gaussian = null_samples_gaussian[floor(
len(null_samples_gaussian) * (1 - alpha))]
type_one_error_boot = sum(
null_samples_boot < thresh_boot) / float(num_null_samples)
type_one_error_gaussian = sum(
null_samples_gaussian < thresh_boot) / float(num_null_samples)
# plot alternative distribution with threshold
subplot(2, 3, 4)
grid(True)
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
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_gaussian)]),
max([max(null_samples_boot),
max(null_samples_gaussian)])
]
# plot null distribution with threshold
subplot(2, 3, 3)
grid(True)
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
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 gaussian
subplot(2, 3, 5)
grid(True)
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
hist(null_samples_gaussian, 20, range=hist_range, normed=True)
axvline(thresh_gaussian, 0, 1, linewidth=2, color='red')
title('Null Dist. Gaussian\nType I error is ' +
str(type_one_error_gaussian))
# pull plots a bit apart
subplots_adjust(hspace=0.5)
subplots_adjust(wspace=0.5)
