def testTooManyNeighbors(self):
"""
Test that GaussianProcess checks if too many neighbours are requested
"""
nData = 100 # number of data points
dimen = 10 # dimension of each point
data = np.zeros((nData,dimen))
fn = np.zeros(nData)
gg = gp.GaussianProcessD(data, fn, gp.SquaredExpCovariogramD())
test = np.zeros(dimen)
sigma = np.empty(1)
mu_arr = np.empty(1)
self.assertRaises(pex.Exception,gg.interpolate,sigma,test,2*nData)
self.assertRaises(pex.Exception,gg.interpolate,sigma,test,-5)
self.assertRaises(pex.Exception,gg.selfInterpolate,sigma,0,2*nData)
self.assertRaises(pex.Exception,gg.selfInterpolate,sigma,0,-5)
self.assertRaises(pex.Exception,gg.selfInterpolate,sigma,-1,nData-1)
# the following segfaults, for unknown reasons, so run directly instead
#self.assertRaises(pex.Exception,gg.selfInterpolate,sigma,nData,nData-1)
try:
gg.interpolate(mu_arr,sigma,2*nData)
self.fail("gg.interpolate(mu_arr,sigma,2*nData) did not fail")
except pex.Exception:
pass
self.assertRaises(pex.Exception,gg.interpolate,mu_arr,sigma,2*nData)
self.assertRaises(pex.Exception,gg.interpolate,mu_arr,sigma,-5)
def testSubtraction(self):
"""
This will test interpolate after subtracting points
"""
tol=1.0e-3
data=np.zeros((2000,10), dtype = float)
fn=np.zeros((2000,4), dtype = float)
mu=np.zeros((4), dtype=float)
sig=np.zeros((4), dtype = float)
mushld=np.zeros((4), dtype = float)
vv=np.zeros((10), dtype = float)
kk=30
f=open("tests/data/gp_subtraction_data.sav","r")
ff=f.readlines()
f.close()
for i in range(len(ff)):
s=ff[i].split()
for j in range(10):
data[i][j]=float(s[j])
for j in range(4):
fn[i][j]=float(s[j+10])
xx=gp.SquaredExpCovariogramD()
xx.setEllSquared(2.3)
try:
gg=gp.GaussianProcessD(data,fn,xx);
except pex.Exception as e:
print e.what()
gg.setLambda(0.002)
j=1
for i in range(1000):
try:
gg.removePoint(j)
except pex.Exception as e:
print e.what()
j=j+1
worstMuErr=-1.0
worstSigErr=-1.0
f=open("tests/data/gp_subtraction_solutions.sav","r")
ff=f.readlines()
f.close()
for i in range(len(ff)):
s=ff[i].split()
for j in range(10):
vv[j]=float(s[j])
for j in range(4):
mushld[j]=float(s[j+10])
sigshld=float(s[14])
gg.interpolate(mu,sig,vv,kk)
for j in range(4):
muErr= (mu[j]-mushld[j])/mushld[j]
sigErr = (sig[j]-sigshld)/sigshld
if muErr < 0.0:
muErr = -1.0 * muErr
if sigErr < 0.0:
sigErr = -1.0 * sigErr
if (muErr > worstMuErr):
worstMuErr=muErr
if (sigErr > worstSigErr):
worstSigErr=sigErr
print "\nThe errors for subtraction interpolation\n"
print "worst mu error ",worstMuErr
print "worst sig2 error ",worstSigErr
self.assertTrue(worstMuErr < tol)
self.assertTrue(worstSigErr < tol)
worstMuErr=-1.0
worstSigErr=-1.0
f=open("tests/data/gp_subtraction_selfinterpolate_solutions.sav","r")
ff=f.readlines()
f.close()
for i in range(len(ff)):
s=ff[i].split()
try:
gg.selfInterpolate(mu,sig,i,kk);
except pex.Exception as e:
print e.what()
for j in range(4):
mushld[j]=float(s[j])
sigshld=float(s[4])
for j in range(4):
muErr=(mu[j]-mushld[j])/mushld[j]
if muErr < 0.0:
muErr=-1.0 * muErr
if muErr>worstMuErr:
worstMuErr=muErr
sigErr=(sig[j]-sigshld)/sigshld
if sigErr<0.0:
sigErr = -1.0*sigErr
if sigErr>worstSigErr:
worstSigErr=sigErr
print "\nThe errors for subtraction self interpolation\n"
print "worst mu error ",worstMuErr
print "worst sig2 error ",worstSigErr
self.assertTrue(worstMuErr < tol)
self.assertTrue(worstSigErr < tol)
def testInterpolate(self):
"""
This will test GaussianProcess.interpolate using both the squared
exponential covariogram and the neural network covariogram on data
that was generated with known answers.
The test will check that the code returns the correct values of both
mu (interpolated function value) and sig2 (the variance)
This test uses the GaussianProcess constructor that does not normalize
coordinate values with minima and maxima.
"""
pp = 2000 #number of data points
dd = 10 #number of dimensions
kk = 15 #number of nearest neighbors being used
tol = 1.0e-3 #the largest relative error that will be tolerated
data = np.zeros((pp,dd),dtype = float) #input data points
fn = np.zeros((pp),dtype = float) #input function values
test = np.zeros((dd),dtype = float) #query points
sigma = np.zeros((1),dtype = float) #variance
xx=gp.SquaredExpCovariogramD()
xx.setEllSquared(100.0)
#read in the input data
f = open("tests/data/gp_exp_covar_data.sav")
ff = f.readlines()
f.close()
for i in range(len(ff)):
s = ff[i].split()
fn[i] = float(s[10])
for j in range(10):
data[i][j] = float(s[j])
#first try the squared exponential covariogram (the default)
try:
gg = gp.GaussianProcessD(data,fn,xx)
except pex.Exception as e:
print e.what()
gg.setLambda(0.001)
#now, read in the test points and their corresponding known solutions
f = open("tests/data/gp_exp_covar_solutions.sav")
ff = f.readlines()
f.close()
worstMuErr = -1.0 #keep track of the worst fractional error in mu
worstSigErr = -1.0 #keep track of the worst fractional error in the variance
for z in range(len(ff)):
s = ff[z].split() #s will store the zth line of the solution file
for i in range(dd):
test[i] = float(s[i]) #read in the test point coordinates
mushld = float(s[dd + kk]) #read in what mu should be
sigshld = float(s[dd + kk + 1]) #read in what the variance should be
mu = gg.interpolate(sigma,test,kk)
err = (mu - mushld)
if mushld != 0.0:
err = err/mushld
if err < 0.0:
err = -1.0 * err
if z == 0 or err > worstMuErr:
worstMuErr = err
err = (sigma[0] - sigshld)
if sigshld != 0.0:
err = err/sigshld
if err < 0.0:
err = -1.0 * err
if z == 0 or err > worstSigErr:
worstSigErr = err
print "\nThe errors for squared exponent covariogram\n"
print "worst mu error ",worstMuErr
print "worst sig2 error ",worstSigErr
self.assertTrue(worstMuErr < tol)
self.assertTrue(worstSigErr < tol)
#now try with the Neural Network covariogram
kk = 50
nn=gp.NeuralNetCovariogramD()
nn.setSigma0(1.23)
nn.setSigma1(0.452)
gg.setCovariogram(nn)
gg.setLambda(0.0045)
f = open("tests/data/gp_nn_solutions.sav")
ff = f.readlines()
f.close()
worstMuErr = -1.0
worstSigErr = -1.0
for z in range(len(ff)):
s = ff[z].split()
for i in range(dd):
test[i] = float(s[i])
mushld = float(s[dd + kk])
sigshld = float(s[dd + kk + 1])
mu = gg.interpolate(sigma,test,kk)
err = (mu - mushld)
if mushld != 0.0:
err = err/mushld
if err < 0.0:
err = -1.0 * err
if z == 0 or err > worstMuErr:
worstMuErr = err
err = (sigma[0] - sigshld)
if sigshld != 0.0:
err = err/sigshld
if err < 0.0:
err = -1.0 * err
if z == 0 or err > worstSigErr:
worstSigErr = err
print "\nThe errors for neural net covariogram\n"
print "worst mu error ",worstMuErr
print "worst sig2 error ",worstSigErr
self.assertTrue(worstMuErr < tol)
self.assertTrue(worstSigErr < tol)
def testSelf(self):
"""
This test will test GaussianProcess.selfInterpolation
"""
pp = 2000
dd = 10
tol = 1.0e-3
kk = 20
data = np.zeros((pp,dd),dtype = float)
fn = np.zeros((pp),dtype = float)
f = open("tests/data/gp_exp_covar_data.sav","r");
ff = f.readlines()
f.close()
for i in range(pp):
s = ff[i].split()
for j in range(dd):
data[i][j] = float(s[j])
fn[i] = float(s[dd])
xx=gp.SquaredExpCovariogramD()
xx.setEllSquared(20.0)
try:
gg = gp.GaussianProcessD(data,fn,xx)
except pex.Exception as e:
print e.what()
gg.setKrigingParameter(30.0)
gg.setLambda(0.00002)
f = open("tests/data/gp_self_solutions.sav","r")
ff = f.readlines()
f.close()
variance = np.zeros((1),dtype = float)
worstMuErr = -1.0
worstSigErr = -1.0
for i in range(pp):
s = ff[i].split()
mushld = float(s[0])
sig2shld = float(s[1])
try:
mu = gg.selfInterpolate(variance,i,kk)
except pex.Exception as e:
print e.what()
err = mu - mushld
if mushld != 0.0:
err = err/mushld
if err < 0.0:
err = err * (-1.0)
if i == 0 or err > worstMuErr:
worstMuErr = err
err = variance[0] - sig2shld
if sig2shld != 0.0:
err = err/sig2shld
if err < 0.0:
err = err * (-1.0)
if i == 0 or err > worstSigErr:
worstSigErr = err
print "\nThe errors for self interpolation\n"
print "worst mu error ",worstMuErr
print "worst sig2 error ",worstSigErr
self.assertTrue(worstMuErr < tol)
self.assertTrue(worstSigErr < tol)
def testBatch(self):
"""
This test will test GaussianProcess.batchInterpolate both with
and without variance calculation
"""
pp = 100
dd = 10
tol = 1.0e-3
data = np.zeros((pp,dd),dtype = float)
fn = np.zeros((pp),dtype = float)
f = open("tests/data/gp_exp_covar_data.sav","r");
ff = f.readlines()
f.close()
for i in range(100):
s = ff[i].split()
for j in range(dd):
data[i][j] = float(s[j])
fn[i] = float(s[dd])
xx=gp.SquaredExpCovariogramD();
xx.setEllSquared(2.0)
try:
gg = gp.GaussianProcessD(data,fn,xx)
except pex.Exception as e:
print e.what()
gg.setLambda(0.0032)
f = open("tests/data/gp_batch_solutions.sav","r")
ff = f.readlines()
f.close()
ntest = len(ff)
mushld = np.zeros((ntest),dtype = float)
varshld = np.zeros((ntest),dtype = float)
mu = np.zeros((ntest),dtype = float)
var = np.zeros((ntest),dtype = float)
queries = np.zeros((ntest,dd),dtype = float)
for i in range(ntest):
s = ff[i].split()
for j in range(dd):
queries[i][j] = float(s[j])
mushld[i] = float(s[dd])
varshld[i] = float(s[dd + 1])
#test with variance calculation
gg.batchInterpolate(mu,var,queries)
worstMuErr = -1.0
worstVarErr = -1.0
for i in range(ntest):
err = mu[i]-mushld[i]
if mushld[i] != 0.0:
err = err/mushld[i]
if err < 0.0:
err = -1.0 * err
if err > worstMuErr:
worstMuErr = err
err = var[i]-varshld[i]
if varshld[i] != 0.0:
err = err/varshld[i]
if err < 0.0:
err = -1.0 * err
if err > worstVarErr:
worstVarErr = err
#test without variance interpolation
#continue keeping track of worstMuErr
gg.batchInterpolate(mu,queries)
for i in range(ntest):
err = mu[i]-mushld[i]
if mushld[i] != 0.0:
err = err/mushld[i]
if err < 0.0:
err = -1.0 * err
if err > worstMuErr:
worstMuErr = err
self.assertTrue(worstMuErr < tol)
self.assertTrue(worstVarErr < tol)
print "\nThe errors for batch interpolation\n"
print "worst mu error ",worstMuErr
print "worst sig2 error ",worstVarErr
def testAddition(self):
"""
This will test the performance of interpolation after adding new points
to GaussianProcess' data set
"""
pp = 1000
dd = 10
kk = 15
tol = 1.0e-4
data = np.zeros((pp,dd),dtype = float)
fn = np.zeros((pp),dtype = float)
test = np.zeros((dd),dtype = float)
sigma = np.zeros((1),dtype = float)
xx=gp.SquaredExpCovariogramD()
xx.setEllSquared(5.0)
f = open("tests/data/gp_additive_test_root.sav")
ff = f.readlines()
f.close()
for i in range(len(ff)):
s=ff[i].split()
fn[i] = float(s[10])
for j in range(10):
data[i][j] = float(s[j])
#establish the Gaussian Process
try:
gg = gp.GaussianProcessD(data,fn,xx)
except pex.Exception as e:
print e.what()
gg.setLambda(0.002)
#now add new points to it and see if GaussianProcess.interpolate performs
#correctly
f = open("tests/data/gp_additive_test_data.sav")
ff = f.readlines()
f.close()
for z in range(len(ff)):
s = ff[z].split()
for i in range(dd):
test[i] = float(s[i])
mushld = float(s[dd])
try:
gg.addPoint(test,mushld)
except pex.Exception as e:
print e.what()
f = open("tests/data/gp_additive_test_solutions.sav")
ff = f.readlines()
f.close()
worstMuErr = -1.0
worstSigErr = -1.0
for z in range(len(ff)):
s = ff[z].split()
for i in range(dd):
test[i] = float(s[i])
mushld = float(s[dd + kk])
sigshld = float(s[dd + kk + 1])
mu = gg.interpolate(sigma,test,kk)
err = (mu - mushld)
if mushld != 0:
err = err/mushld
if err < 0.0:
err = -1.0 * err
if z == 0 or err > worstMuErr:
worstMuErr = err
err = (sigma[0] - sigshld)
if sigshld != 0:
err = err/sigshld
if err < 0.0:
err = -1.0 * err
if z == 0 or err > worstSigErr:
worstSigErr = err
print "\nThe errors for the test of adding points to the Gaussian process\n"
print "worst mu error ",worstMuErr
print "worst sig2 error ",worstSigErr
self.assertTrue(worstMuErr < tol)
self.assertTrue(worstSigErr < tol)