def evaluate(X_tr, y_tr, X_te, y_te, interactions=None):
grid = sg.RegularGridConfiguration()
grid.dim_ = 64
grid.level_ = 2
grid.type_ = sg.GridType_ModLinear
adapt = sg.AdaptivityConfiguration()
adapt.numRefinements_ = 0
adapt.noPoints_ = 0
solv = sg.SLESolverConfiguration()
solv.maxIterations_ = 50
solv.eps_ = 10e-6
solv.threshold_ = 10e-6
solv.type_ = sg.SLESolverType_CG
final_solv = solv
final_solv.maxIterations = 200
regular = sg.RegularizationConfiguration()
regular.type_ = sg.RegularizationType_Identity
regular.exponentBase_ = 1.0
regular.lambda_ = 0.1
X_tr = sg.DataMatrix(X_tr)
y_tr = sg.DataVector(y_tr)
X_te = sg.DataMatrix(X_te)
y_te = sg.DataVector(y_te)
if interactions is None:
estimator = sg.ClassificationLearner(grid, adapt, solv, final_solv,regular)
else:
estimator = sg.ClassificationLearner(grid, adapt, solv, final_solv,regular, interactions)
estimator.train(X_tr,y_tr)
return estimator.getAccuracy(X_te,y_te)
def test_laplace(grid, lmax):
resolution = 100000
grid.getGenerator().regular(lmax)
gridStorage = grid.getStorage()
size = gridStorage.getSize()
b = getBasis(grid)
op = pysgpp.createOperationLaplace(grid)
alpha = pysgpp.DataVector(size)
result = pysgpp.DataVector(size)
for point_i in range(size):
for point_j in range(size):
gp_i = gridStorage.getPoint(point_i)
gp_j = gridStorage.getPoint(point_j)
print("--------")
for i in range(0, size):
alpha[i] = 0
alpha[point_i] = 1
op.mult(alpha, result)
xs = np.linspace(0, 1, resolution)
approx = sum([
b.evalDx(gp_i.getLevel(0), gp_i.getIndex(0), x) *
b.evalDx(gp_j.getLevel(0), gp_j.getIndex(0), x) for x in xs
]) / resolution
print("i,j: {},{} result: {} approx:{}".format(
point_i, point_j, result[point_j], approx))
if (abs(result.get(point_j) - approx) > 1e-1):
print("--------")
print("points: {},{} ".format(point_i, point_j))
print("approx:{}".format(approx))
print("result:{}".format(result.get(point_j)))
# print result
print("--------")
def getNextTodoPoints(maxPoints, precalculatedValues, dim, gridType, degree,
numTimeSteps, gridResolution, normalization, residual,
wave_type, distribution, minimum_allowed_height):
okushiriFunc = okushiri(dim, numTimeSteps, gridResolution, normalization,
residual, wave_type, distribution,
minimum_allowed_height)
lb, ub = okushiriFunc.getDomain()
objFunc = vectorObjFuncSGpp(okushiriFunc)
pdfs = objFunc.getDistributions()
reSurf = pysgpp.SplineResponseSurfaceVector(
objFunc, pysgpp.DataVector(lb), pysgpp.DataVector(ub),
pysgpp.Grid.stringToGridType(gridType), degree)
reSurf.regular(initialLevel)
todoPointsDetermined = False
counter = 0
while not todoPointsDetermined:
previousSize = reSurf.getSize()
if previousSize > maxPoints:
print(
f"nothing to calculate for a maximum of {maxPoints} grid points"
)
return todoPointsDetermined, [], reSurf.getSize()
reSurf.nextSurplusAdaptiveGrid(numRefine, verbose)
#reSurf.nextDistributionAdaptiveGrid(numRefine, pdfs, verbose)
todoPointsDetermined, todoPoints = checkPrecalc(
reSurf, precalculatedValues)
if not todoPointsDetermined:
counter = counter + 1
print(f"refining ({counter}), grid size: {reSurf.getSize()}")
reSurf.refineSurplusAdaptive(numRefine, verbose)
#reSurf.refineDistributionAdaptive(numRefine, pdfs, verbose)
return todoPointsDetermined, todoPoints, reSurf.getSize()
def evaluate_one(estimator, X, y, train, test):
train_X = sg.DataMatrix(X[train])
train_y = sg.DataVector(y[train])
test_X = sg.DataMatrix(X[test])
test_y = sg.DataVector(y[test])
estimator.train(train_X, train_y)
error = estimator.getMSE(test_X, test_y)
return error
def getDomain(self):
lb = pysgpp.DataVector(self.dim)
ub = pysgpp.DataVector(self.dim)
for d in range(self.dim):
bounds = self.pdfs.get(d).getBounds()
lb[d] = bounds[0]
ub[d] = bounds[1]
return lb, ub
def generate_friedman1(seed):
(X, y) = data.make_friedman1(n_samples=2000, random_state=seed, noise=1.0)
# transform values to DataMatrix/DataVector types
X = sg.DataMatrix(X)
y = sg.DataVector(y)
return (X, y)
def __init__(self, d, f):
self.f = f
self.d = d
self.grid = pysgpp.Grid.createBsplineClenshawCurtisGrid(d, 3)
self.gridStorage = self.grid.getStorage()
try :
self.hierarch = pysgpp.createOperationHierarchisation(self.grid)
except :
self.hierarch = pysgpp.createOperationMultipleHierarchisation(self.grid)
self.opeval = pysgpp.createOperationEvalNaive(self.grid)
self.alpha = pysgpp.DataVector(self.gridStorage.getSize())
def create_interpolation(self, grid_lvl):
self.gridStorage.clear()
self.grid.getGenerator().regular(grid_lvl)
self.alpha = pysgpp.DataVector(self.gridStorage.getSize())
self.min_f = float('inf')
for i in range(self.gridStorage.getSize()):
gp = self.gridStorage.getPoint(i)
x = [self.gridStorage.getCoordinate(gp, j) for j in range(self.d)]
self.alpha[i] = self.f(x)
if self.alpha[i] < self.min_f:
self.min_f = self.alpha[i]
self.min_x = x
self.hierarch.doHierarchisation(self.alpha)
def eval(self, x):
# map x from [0,1]^D to real parameter space reallb, realub
transX = pysgpp.DataVector(self.dim)
for d in range(self.dim):
transX[d] = self.reallb[d] + (self.realub[d] -
self.reallb[d]) * x[d]
# print(f"{x.toString()}")
# print(f"{transX.toString()}\n")
y = self.okushiriStorage.eval(transX)
maxRunUp = np.max(y)
time = float(np.argmax(y))
return maxRunUp
def test_LTwoDotImplicit(grid, l):
grid.getGenerator().regular(l)
gridStorage = grid.getStorage()
size = gridStorage.getSize()
# print(size)
m = pysgpp.DataMatrix(size, size)
opExplicit = pysgpp.createOperationLTwoDotExplicit(m, grid)
op = pysgpp.createOperationLTwoDotProduct(grid)
alpha = pysgpp.DataVector(size)
resultExplicit = pysgpp.DataVector(size)
result = pysgpp.DataVector(size)
for i in range(size):
alpha[i] = 1
opExplicit.mult(alpha, resultExplicit)
op.mult(alpha, result)
for i in range(size):
if result[i] != resultExplicit[i]:
print("Error result entry {} differs".format(i))
if abs(result[i] - resultExplicit[i]) > 1e-16:
# print result[i] - resultExplicit[i]
print("result:{}".format(result[i]))
print("resultExplicit:{}".format(resultExplicit[i]))
def ct_to_pce():
start_time = time.time()
# initialize model function
func = pysgpp.multiFunc(expModel)
numDims = 2
# regular sparse grid level q
q = 6
# create polynomial basis
config = pysgpp.OrthogonalPolynomialBasis1DConfiguration()
config.polyParameters.type_ = pysgpp.OrthogonalPolynomialBasisType_LEGENDRE
basisFunction = pysgpp.OrthogonalPolynomialBasis1D(config)
# create sprarse grid interpolation operation
op = pysgpp.CombigridOperation.createExpClenshawCurtisPolynomialInterpolation(
numDims, func)
# start with regular level q and add some levels adaptively
op.getLevelManager().addRegularLevels(q)
op.getLevelManager().addLevelsAdaptiveByNumLevels(5)
## and construct a PCE representation to easily calculate statistical features of our model.
# create polynomial chaos surrogate from sparse grid
surrogateConfig = pysgpp.CombigridSurrogateModelConfiguration()
surrogateConfig.type = pysgpp.CombigridSurrogateModelsType_POLYNOMIAL_CHAOS_EXPANSION
surrogateConfig.loadFromCombigridOperation(op)
surrogateConfig.basisFunction = basisFunction
pce = pysgpp.createCombigridSurrogateModel(surrogateConfig)
# compute sobol indices
sobol_indices = pysgpp.DataVector(1)
total_indices = pysgpp.DataVector(1)
pce.getComponentSobolIndices(sobol_indices)
pce.getTotalSobolIndices(total_indices)
# print results
print("Mean: {} Variance: {}".format(pce.mean(), pce.variance()))
print("Sobol indices {}".format(sobol_indices.toString()))
print("Total Sobol indices {}".format(total_indices.toString()))
print("Sum {}\n".format(sobol_indices.sum()))
print("Elapsed time: {} s".format(time.time() - start_time))
def example3():
## Use Leja points unlike example 2 and use CombigridMultiOperation for evaluation at multiple
## points.
operation = pysgpp.CombigridMultiOperation.createLinearLejaPolynomialInterpolation(
d, func)
## We slightly deviate from the C++ example here and pass the interpolation points via a DataMatrix.
## We will use 2 interpolation points.
## IMPORTANT: For python, the parameters matrix needs to be transposed
firstParam = [0.2, 0.6, 0.7]
secondParam = [0.3, 0.9, 1.0]
params = np.array([firstParam, secondParam])
parameters = pysgpp.DataMatrix(params.transpose())
print(parameters)
## Let's use the simple interface for this example and stop the time:
stopwatch = pysgpp.Stopwatch()
result = operation.evaluate(3, parameters)
stopwatch.log()
print("First result: " + str(result[0]) + ", function value: " +
str(func(pysgpp.DataVector(firstParam))))
print("Second result: " + str(result[1]) + ", function value: " +
str(func(pysgpp.DataVector(secondParam))))
def example5():
## First, we want to configure which grid points to use in which dimension.
## We use Chebyshev points in the 0th dimension. To make them nested, we have to use at least \f$n
## = 3^l\f$ points at level \f$l\f$. This is why this method contains the prefix exp.
## CombiHierarchies provides some matching configurations for grid points. If you nevertheless
## need your own configuration or you want to know which growth strategy and ordering fit to which
## point distribution, look up the implementation details in CombiHierarchies, it is not
## difficult to implement your own configuration.
grids = pysgpp.AbstractPointHierarchyVector()
grids.push_back(pysgpp.CombiHierarchies.expChebyshev())
## Our next set of grid points are Leja points with linear growth (\f$n = 1 + 3l\f$).
## For the last dimension, we use equidistant points with boundary. These are suited for linear
## interpolation. To make them nested, again the slowest possible exponential growth is selected
## by the CombiHierarchies class.
grids.push_back(pysgpp.CombiHierarchies.linearLeja(3))
grids.push_back(pysgpp.CombiHierarchies.expUniformBoundary())
## The next thing we have to configure is the linear operation that is performed in those
## directions. We will use polynomial interpolation in the 0th dimension, quadrature in the 1st
## dimension and linear interpolation in the 2nd dimension.
## Roughly spoken, this means that a quadrature is performed on the 1D function that is the
## interpolated function with two fixed parameters. But since those operators "commute", the
## result is invariant under the order that the operations are applied in.
## The CombiEvaluators class also provides analogous methods and typedefs for the multi-evaluation
## case.
evaluators = pysgpp.FloatScalarAbstractLinearEvaluatorVector()
evaluators.push_back(pysgpp.CombiEvaluators.polynomialInterpolation())
evaluators.push_back(pysgpp.CombiEvaluators.quadrature())
evaluators.push_back(pysgpp.CombiEvaluators.linearInterpolation())
## To create a CombigridOperation object with our own configuration, we have to provide a
## LevelManager as well:
levelManager = pysgpp.WeightedRatioLevelManager()
operation = pysgpp.CombigridOperation(grids, evaluators, levelManager,
func)
## The two interpolations need a parameter \f$(x, z)\f$. If \f$\tilde{f}\f$ is the interpolated
## function, the operation approximates the result of \f$\int_0^1 \tilde{f}(x, y, z) \,dy\f$.
parameters = pysgpp.DataVector([0.777, 0.14159])
result = operation.evaluate(2, parameters)
print("Result: " + str(result))
def plotFunction(opEval, surpluses, X):
if not doPlot: return
# generate a meshgrid for plotting
xx0 = np.linspace(0, 1, 65)
xx1 = np.linspace(0, 1, 65)
XX0, XX1 = np.meshgrid(xx0, xx1)
XX = pysgpp.DataMatrix(np.column_stack([XX0.flatten(), XX1.flatten()]))
# evaluate interpolant at meshgrid
YY = pysgpp.DataVector(0)
opEval.multiEval(surpluses, XX, YY)
# convert resulting sgpp::base::DataVector to NumPy array
YY = np.reshape(np.array([YY[k] for k in range(YY.getSize())]), XX0.shape)
# actual plotting
fig = plt.figure(figsize=(6, 6))
ax = fig.gca(projection="3d")
ax.plot_surface(XX0, XX1, YY)
ax.plot(X[:, 0], X[:, 1], "k.", zs=f(X[:, 0], X[:, 1]), ms=10)
def getPercentiles(reSurf, percentages, numSamples, distribution='normal'):
# create a large set of samples
dim = reSurf.getNumDim()
numTimeSteps = reSurf.getNumRes()
if distribution == 'normal':
mean = 1.0
sd = 0.125
lower = 0.5
upper = 1.5
rng = truncnorm((lower - mean) / sd, (upper - mean) / sd, loc=mean, scale=sd)
sampleSet = rng.rvs((numSamples, dim))
elif distribution == 'uniform':
lb = 0.5
ub = 1.5
unitpoints = np.random.rand(numSamples, dim)
sampleSet = np.zeros((numSamples, dim))
for i, point in enumerate(unitpoints):
for d in range(dim):
sampleSet[i, d] = lb + (ub-lb)*point[d]
point = pysgpp.DataVector(dim)
results = np.zeros((numSamples, numTimeSteps))
for i in range(numSamples):
for d in range(dim):
point[d] = sampleSet[i, d]
res = reSurf.eval(point)
for n in range(numTimeSteps):
results[i, n] = res[n]
print(f'max of all {numSamples} percentile samples: {np.max(results)*400:.4f}m')
# calculate the percentiles from the set of samples
# (Also Monte Carlo based mean to see if Stochastic Collocation
# gives a significant increase in accuracy over simple MC.)
percentiles = np.zeros((len(percentages), numTimeSteps))
mcMeans = np.zeros(numTimeSteps)
for n in range(numTimeSteps):
mcMeans[n] = np.mean(results[:, n])
for p in range(len(percentages)):
percentiles[p, n] = np.percentile(results[:, n], percentages[p])
return percentiles, mcMeans
def test_firstMoment(grid, lmax):
grid.getGenerator().regular(lmax)
resolution = 100000
gridStorage = grid.getStorage()
b = grid.getBasis()
op = pysgpp.createOperationFirstMoment(grid)
alpha = pysgpp.DataVector(grid.getSize(), 1.0)
bounds = pysgpp.DataMatrix(1, 2, 0.0)
bounds.set(0, 1, 1.0)
res = 0.0
for i in range(grid.getSize()):
lev = gridStorage.getPoint(i).getLevel(0)
ind = gridStorage.getPoint(i).getIndex(0)
temp_res = 0.0
for c in range(resolution):
x = float(c) / resolution
temp_res += x * b.eval(lev, ind, x)
res += alpha.get(i) * temp_res / resolution
print("--FirstMoment--")
print(res)
print(op.doQuadrature(alpha, bounds))
print(res - op.doQuadrature(alpha, bounds))
def example6():
## To create a CombigridOperation, we currently have to use the longer way as in example 5.
grids = pysgpp.AbstractPointHierarchyVector(
d, pysgpp.CombiHierarchies.expUniformBoundary())
evaluators = pysgpp.FloatScalarAbstractLinearEvaluatorVector(
d, pysgpp.CombiEvaluators.cubicSplineInterpolation())
levelManager = pysgpp.WeightedRatioLevelManager()
## We have to specify if the function always produces the same value for the same grid points.
## This can make the storage smaller if the grid points are nested. In this implementation, this
## is true. However, it would be false in the PDE case, so we set it to false here.
exploitNesting = False
## Now create an operation as usual and evaluate the interpolation with a test parameter.
operation = pysgpp.CombigridOperation(grids, evaluators, levelManager,
pysgpp.gridFunc(gf), exploitNesting)
parameter = pysgpp.DataVector([0.1, 0.2, 0.3])
result = operation.evaluate(4, parameter)
print("Target function value: " + str(func(parameter)))
print("Numerical result: " + str(result))
def example8(dist_type="uniform"):
operation = pysgpp.CombigridOperation.createExpClenshawCurtisPolynomialInterpolation(
d, func)
config = pysgpp.OrthogonalPolynomialBasis1DConfiguration()
if dist_type == "beta":
config.polyParameters.type_ = pysgpp.OrthogonalPolynomialBasisType_JACOBI
config.polyParameters.alpha_ = 5
config.polyParameters.alpha_ = 4
U = J(
[Beta(config.polyParameters.alpha_, config.polyParameters.beta_)] *
d)
else:
config.polyParameters.type_ = pysgpp.OrthogonalPolynomialBasisType_LEGENDRE
U = J([Uniform(0, 1)] * d)
basisFunction = pysgpp.OrthogonalPolynomialBasis1D(config)
basisFunctions = pysgpp.OrthogonalPolynomialBasis1DVector(d, basisFunction)
q = 3
operation.getLevelManager().addRegularLevels(q)
print("Total function evaluations: %i" % operation.numGridPoints())
## compute variance of the interpolant
surrogateConfig = pysgpp.CombigridSurrogateModelConfiguration()
surrogateConfig.type = pysgpp.CombigridSurrogateModelsType_POLYNOMIAL_CHAOS_EXPANSION
surrogateConfig.loadFromCombigridOperation(operation)
surrogateConfig.basisFunction = basisFunction
pce = pysgpp.createCombigridSurrogateModel(surrogateConfig)
n = 10000
values = [g(pysgpp.DataVector(xi)) for xi in U.rvs(n)]
print("E(u) = %g ~ %g" % (np.mean(values), pce.mean()))
print("Var(u) = %g ~ %g" % (np.var(values), pce.variance()))
class objFuncSGpp(pysgpp.ScalarFunction):
def __init__(self, dim):
super(objFuncSGpp, self).__init__(dim)
def eval(self, x):
# input x: pysgpp.DataVector
return x[0] * x[1] + x[1]
# create an instance of the objective
dim = 2
objFunc = objFuncSGpp(dim)
# set the objectives domain
lb = pysgpp.DataVector([0, 0]) # domain has lower bounds 0
ub = pysgpp.DataVector([1, 1]) # domain has upper bounds 1
# set up response surface
degree = 3
gridType = 'nakBsplineBoundary'
# create a response surface object
reSurf = pysgpp.SplineResponseSurface(objFunc, lb, ub,
pysgpp.Grid.stringToGridType(gridType),
degree)
# create surrogate with regular sparse grid
# reSurf.regular(2)
# create surrogate with spatially adaptive sparse grid
# accuracy of the extension principle
numberOfAlphaSegments = 100
## We use regular sparse grids for the sparse grid surrogates.
print("Constructing the sparse grids...")
gridBSpline = pysgpp.Grid.createBsplineBoundaryGrid(d, p, b)
gridBSpline.getGenerator().regular(n)
gridLinear = pysgpp.Grid.createBsplineBoundaryGrid(d, 1, b)
gridLinear.getGenerator().regular(n)
N = gridBSpline.getSize()
gridStorage = gridBSpline.getStorage()
functionValues = pysgpp.DataVector(N)
x = pysgpp.DataVector(d)
for k in range(N):
gridStorage.getPoint(k).getStandardCoordinates(x)
functionValues[k] = f.eval(x)
## For the hierarchization for the B-spline surrogate, we solve the corresponding
## system of linear equations and create the interpolant and its gradient.
print("Hierarchizing (B-spline coefficients)...")
surplusesBSpline = pysgpp.DataVector(N)
hierSLEBSpline = pysgpp.HierarchisationSLE(gridBSpline)
sleSolverBSpline = pysgpp.AutoSLESolver()
if not sleSolverBSpline.solve(hierSLEBSpline, functionValues,
def eval_op_transpose(x, op, size):
result_vec = sg.DataVector(size)
x = sg.DataVector(np.array(x).flatten())
op.multTranspose(x, result_vec)
return result_vec.array().copy()
print("--------------------------------------------------------------------------------------")
dim = 2
radius = 0.1
degree = 2
grid = pysgpp.Grid.createWEBsplineGrid(dim, degree)
gridStorage = grid.getStorage()
print("dimensionality: {}".format(gridStorage.getDimension()))
level = 3
grid.getGenerator().regular(level)
print("number of grid points: {}".format(gridStorage.getSize()))
alpha = pysgpp.DataVector(gridStorage.getSize(),0.0)
beta = pysgpp.DataVector(gridStorage.getSize(),0.0)
print("length of alpha vector: {}".format(len(alpha)))
print("length of beta vector: {}".format(len(beta)))
printLine()
for i in range(gridStorage.getSize()):
gp = gridStorage.getPoint(i)
alpha[i] = gp.getStandardCoordinate(0)
beta[i] = gp.getStandardCoordinate(1)
#print("alpha: {}".format(alpha))
#print("beta: {}".format(beta))
x = np.zeros((len(alpha),dim))
eval_circle= np.zeros(len(alpha))
print("dimensionality: {}".format(dim))
# create regular grid, level 3
level = 3
gridGen = grid.getGenerator()
gridGen.regular(level)
print("number of grid points: {}".format(gridStorage.getSize()))
## Calculate the surplus vector alpha for the interpolant of \f$
## f(x)\f$. Since the function can be evaluated at any
## point. Hence. we simply evaluate it at the coordinates of the
## grid points to obtain the nodal values. Then we use
## hierarchization to obtain the surplus value.
# create coefficient vector
alpha = pysgpp.DataVector(gridStorage.getSize())
for i in range(gridStorage.getSize()):
gp = gridStorage.getPoint(i)
p = tuple([gp.getStandardCoordinate(j) for j in range(dim)])
alpha[i] = f(p)
pysgpp.createOperationHierarchisation(grid).doHierarchisation(alpha)
## Now we compute and compare the quadrature using four different methods available in SG++.
# direct quadrature
opQ = pysgpp.createOperationQuadrature(grid)
res = opQ.doQuadrature(alpha)
print("exact integral value: {}".format(res))
# Monte Carlo quadrature using 100000 paths
def main():
# Generate data
print("generate dataset... ", end=' ')
data_tr,_ = generate_friedman1(123456)
print("Done")
print("generated a friedman1 dataset (10D) with 2000 samples")
# Config grid
print("create grid config... ", end=' ')
grid = sg.RegularGridConfiguration()
grid.dim_ = 10
grid.level_ = 3
grid.type_ = sg.GridType_Linear
print("Done")
# Config adaptivity
print("create adaptive refinement config... ", end=' ')
adapt = sg.AdaptivityConfiguration()
adapt.numRefinements_ = 0
adapt.noPoints_ = 10
print("Done")
# Config solver
print("create solver config... ", end=' ')
solv = sg.SLESolverConfiguration()
solv.maxIterations_ = 1000
solv.eps_ = 1e-14
solv.threshold_ = 1e-14
solv.type_ = sg.SLESolverType_CG
print("Done")
# Config regularization
print("create regularization config... ", end=' ')
regular = sg.RegularizationConfiguration()
regular.regType_ = sg.RegularizationType_Laplace
print("Done")
# Config cross validation for learner
print("create learner config... ", end=' ')
#crossValid = sg.CrossvalidationConfiguration()
crossValid = sg.CrossvalidationConfiguration()
crossValid.enable_ = False
crossValid.kfold_ = 3
crossValid.lambda_ = 3.16228e-06
crossValid.lambdaStart_ = 1e-1
crossValid.lambdaEnd_ = 1e-10
crossValid.lambdaSteps_ = 3
crossValid.logScale_ = True
crossValid.shuffle_ = True
crossValid.seed_ = 1234567
crossValid.silent_ = False
print("Done")
#
# Create the learner with the given configuration
#
print("create the learner... ")
learner = sg.LearnerSGDE(grid, adapt, solv, regular, crossValid)
learner.initialize(data_tr)
# Train the learner
print("start training... ")
learner.train()
print("done training")
#
# Estimate the probability density function (pdf) via a Gaussian kernel
# density estimation (KDE) and print the corresponding values
#
kde = sg.KernelDensityEstimator(data_tr)
x = sg.DataVector(learner.getDim())
x.setAll(0.5)
print("-----------------------------------------------")
print(learner.getSurpluses().getSize(), " -> ", learner.getSurpluses().sum())
print("pdf_SGDE(x) = ", learner.pdf(x), " ~ ", kde.pdf(x), " = pdf_KDE(x)")
print("mean_SGDE = ", learner.mean(), " ~ ", kde.mean(), " = mean_KDE")
print("var_SGDE = ", learner.variance(), " ~ ", kde.variance(), " = var_KDE")
# Print the covariances
C = sg.DataMatrix(grid.dim_, grid.dim_)
print("----------------------- Cov_SGDE -----------------------")
learner.cov(C)
print(C)
print("----------------------- Cov_KDE -----------------------")
kde.cov(C)
print(C)
#
# Apply the inverse Rosenblatt transformatio to a matrix of random points. To
# do this, first generate the random points uniformly, then initialize an
# inverse Rosenblatt transformation operation and apply it to the points.
# Finally print the calculated values
#
print("-----------------------------------------------")
opInvRos = sg.createOperationInverseRosenblattTransformation(learner.getGrid())
points = sg.DataMatrix(randu_mat(12, grid.dim_))
print(points)
pointsCdf = sg.DataMatrix(points.getNrows(), points.getNcols())
opInvRos.doTransformation(learner.getSurpluses(), points, pointsCdf)
#
# To check whether the results are correct perform a Rosenform transformation on
# the data that has been created by the inverse Rosenblatt transformation above
# and print the calculated values
#
points.setAll(0.0)
opRos = sg.createOperationRosenblattTransformation(learner.getGrid())
opRos.doTransformation(learner.getSurpluses(), pointsCdf, points)
print("-----------------------------------------------")
print(pointsCdf)
print("-----------------------------------------------")
print(points)
def unitSample(self):
dv = pysgpp.DataVector(self.k)
self.gen.getSample(dv)
return [dv[i] for i in range(self.k)]
## With the iterative grid generator, we generate adaptively a sparse grid.
printLine()
print("Generating grid...\n")
if not gridGen.generate():
print("Grid generation failed, exiting.")
sys.exit(1)
## Then, we hierarchize the function values to get hierarchical B-spline
## coefficients of the B-spline sparse grid interpolant
## \f$\tilde{f}\colon [0, 1]^d \to \mathbb{R}\f$.
printLine()
print("Hierarchizing...\n")
functionValues = gridGen.getFunctionValues()
coeffs = pysgpp.DataVector(len(functionValues))
hierSLE = pysgpp.OptHierarchisationSLE(grid)
sleSolver = pysgpp.OptAutoSLESolver()
# solve linear system
if not sleSolver.solve(hierSLE, gridGen.getFunctionValues(), coeffs):
print("Solving failed, exiting.")
sys.exit(1)
## We define the interpolant \f$\tilde{f}\f$ and its gradient
## \f$\nabla\tilde{f}\f$ for use with the gradient method (steepest descent).
## Of course, one can also use other optimization algorithms from
## sgpp::optimization::optimizer.
printLine()
print("Optimizing smooth interpolant...\n")
ft = pysgpp.OptInterpolantScalarFunction(grid, coeffs)
def example2():
## This time, we use Clenshaw-Curtis points with exponentially growing number of points per level.
## This is helpful for CC points to make them nested. Nested means that the set of grid points at
## one level is a subset of the set of grid points at the next level. Nesting can drastically
## reduce the number of needed function evaluations. Using these grid points, we will do
## polynomial interpolation at a single point.
#operation = pysgpp.CombigridOperation.createExpClenshawCurtisPolynomialInterpolation(d, func)
operation = pysgpp.CombigridOperation.createExpL2LejaPolynomialInterpolation(
d, func)
## Now create a point where to evaluate the interpolated function:
evaluationPoint = pysgpp.DataVector([0.1572, 0.6627, 0.2378])
## We can now evaluate the interpolation at this point (using 3 as a bound for the 1-norm of the
## level multi-index):
result = operation.evaluate(3, evaluationPoint)
## Now compare the result to the actual function value:
print("Interpolation result: " + str(result) + ", function value: " +
str(func(evaluationPoint)))
## Again, print the number of function evaluations:
print("Function evaluations: " + str(operation.numGridPoints()))
## Now, let's do another (more sophisticated) evaluation at a different point, so change the point
## and re-set the parameter. This method will automatically clear all intermediate values that
## have been computed internally up to now.
evaluationPoint[0] = 0.4444
print("Target function value: " + str(func(evaluationPoint)))
operation.setParameters(evaluationPoint)
## The level manager provides more options for combigrid evaluation, so let's get it:
levelManager = operation.getLevelManager()
## We can add regular levels like before:
levelManager.addRegularLevels(3)
## The result can be fetched from the CombigridOperation:
print("Regular result 1: " + str(operation.getResult()))
print("Total function evaluations: " + str(operation.numGridPoints()))
## We can also add more points in a regular structure, using at most 50 new function evaluations.
## Most level-adding variants of levelManager also have a parallelized version. This version
## executes the calls to func in parallel with a specified number of threads, which is okay here
## since func supports parallel evaluations. Since func takes very little time to evaluate and the
## parallelization only concerns function evaluations and not the computations on the resulting
## function values, parallel evaluation is not actually useful in this case.
## We will use 4 threads for the function evaluations.
levelManager.addRegularLevelsByNumPointsParallel(50, 4)
print("Regular result 2: " + str(operation.getResult()))
print("Total function evaluations: " + str(operation.numGridPoints()))
## We can also use adaptive level generation. The adaption strategy depends on the subclass of
## LevelManager that is used. If you do not want to use the default LevelManager, you can specify
## your own LevelManager:
operation.setLevelManager(pysgpp.AveragingLevelManager())
levelManager = operation.getLevelManager()
## It was necessary to use setLevelManager(), because this links the LevelManager to the
## computation. Now, let's add at most 60 more function evaluations adaptively.
## Note that the adaption here is only based on the result at our single evaluation point, which
## might give inaccurate results. The same holds for quadrature.
## In practice, you should probably do an interpolation at a lot of Monte-Carlo points via
## CombigridMultiOperation (cf. Example 3) and then transfer the generated level structure to
## another CombigridOperation or CombigridMultiOperation for your actual evaluation (cf. Example
## 4).
levelManager.addLevelsAdaptive(60)
print("Adaptive result: " + str(operation.getResult()))
print("Total function evaluations: " + str(operation.numGridPoints()))
## We can also fetch the used grid points and plot the grid:
grid = levelManager.getGridPointMatrix()
fig = plt.figure()
ax = fig.add_subplot(111, projection='3d')
gridList = [[grid.get(r, c) for c in range(grid.getNcols())]
for r in range(grid.getNrows())]
ax.scatter(gridList[0], gridList[1], gridList[2], c='r', marker='o')
ax.set_xlabel('x')
ax.set_ylabel('y')
ax.set_zlabel('z')
plt.show()
def __call__(self, x):
if (self.d == 1 and not isinstance(x, list)):
x = [x]
return self.opeval.eval(self.alpha, pysgpp.DataVector(x))
def generate_friedman1(seed):
(X, y) = data.make_friedman1(n_samples=10000, random_state=seed, noise=1.0)
y = sg.DataVector(y)
X = sg.DataMatrix(X)
return X, y
means = [m / meanNormalizingFactor for m in means]
### Multiply everything by 400 so that we get real scale, not model scale ###
for p in range(len(percentiles)):
for t in range(numTimeSteps):
percentiles[p, t] *= 400
maxTimeline = [maxTimeline[t]*400 for t in range(numTimeSteps)]
means = [means[t]*400 for t in range(numTimeSteps)]
mcMeans *= 400
print(
f'average mean error: {np.average(np.abs(means-reference_means))} worst diff {np.max(np.abs(means-reference_means))}')
# Calculate VARIANCES
if calcVar:
dummyMeanSquares = pysgpp.DataVector(1)
dummyMeans = pysgpp.DataVector(1)
start = time.time()
variances = reSurf.getVariances(pdfs, quadOrder, dummyMeans, dummyMeanSquares)
variances_py = [variances[t] for t in range(numTimeSteps)]
variances_py = [v*400 for v in variances_py]
print(f'Calculating the variances took {time.time()-start}s')
np.savetxt('/home/rehmemk/git/anugasgpp/Okushiri/plots/variances.txt', variances_py)
else:
if plotVar:
print('LOADING VARIANCES FROM FILE. NO GUARANTEES FOR THEM MATCHING THE CURRENT CALCULATIONS!')
variances_py = np.loadtxt('/home/rehmemk/git/anugasgpp/Okushiri/plots/variances.txt')
### Plotting ###
time = [t*22.5/450 for t in range(451)]
