def printDebug(self):
assert self._debug is True
from tabulate import tabulate
iters = [self._debugUACalls[i].innerIters for i in range(len(self._debugUACalls))]
sig02 = [self._debugUACalls[i].sig02 for i in range(len(self._debugUACalls))]
sig12 = [self._debugUACalls[i].sig12 for i in range(len(self._debugUACalls))]
sign2 = [self._debugUACalls[i].sign2 for i in range(len(self._debugUACalls))]
gradMeans = [NP.mean(abs(self._debugUACalls[i].lastGrad)) for i in range(len(self._debugUACalls))]
Pr.prin("*** Update approximation ***")
Pr.prin("calls: %d" % (len(self._debugUACalls),))
table = [["", "min", "max", "mean"],
["iters", min(iters), max(iters), NP.mean(iters)],
["|grad|_{mean}", min(gradMeans), max(gradMeans), NP.mean(gradMeans)],
["sig01", min(sig02), max(sig02), NP.mean(sig02)],
["sig11", min(sig12), max(sig12), NP.mean(sig12)],
["sign1", min(sign2), max(sign2), NP.mean(sign2)]]
Pr.prin(tabulate(table))
def printDebug(self):
assert self._debug is True
from tabulate import tabulate
iters = [
self._debugUACalls[i].innerIters
for i in range(len(self._debugUACalls))
]
sig02 = [
self._debugUACalls[i].sig02 for i in range(len(self._debugUACalls))
]
sig12 = [
self._debugUACalls[i].sig12 for i in range(len(self._debugUACalls))
]
sign2 = [
self._debugUACalls[i].sign2 for i in range(len(self._debugUACalls))
]
gradMeans = [
NP.mean(abs(self._debugUACalls[i].lastGrad))
for i in range(len(self._debugUACalls))
]
Pr.prin("*** Update approximation ***")
Pr.prin("calls: %d" % (len(self._debugUACalls), ))
table = [["", "min", "max", "mean"],
["iters", min(iters),
max(iters), NP.mean(iters)],
[
"|grad|_{mean}",
min(gradMeans),
max(gradMeans),
NP.mean(gradMeans)
], ["sig01", min(sig02),
max(sig02),
NP.mean(sig02)],
["sig11", min(sig12),
max(sig12), NP.mean(sig12)],
["sign1", min(sign2),
max(sign2), NP.mean(sign2)]]
Pr.prin(tabulate(table))
def _updateApproximation(self):
'''
Calculates the Laplace approximation for the posterior.
It can be defined by two variables: f mode and W at f mode.
'''
if self._updateApproximationCount == 0:
return
if self._is_kernel_zero():
self._updateApproximationCount = 0
return
self._updateApproximationBegin()
gradEpsStop = 1e-10
objEpsStop = 1e-8
gradEpsErr = 1e-3
self._mean = self._calculateMean()
m = self._mean
if self._lasta is None or self._lasta.shape[0] != self._N:
aprev = NP.zeros(self._N)
else:
aprev = self._lasta
fprev = self._rdotK(aprev) + m
objprev = self._likelihood.log(fprev, self._y) - (fprev-m).dot(aprev)/2.0
ii = 0
line_search = False
maxIter = 1000
failed = False
failedMsg = ''
while ii < maxIter:
grad = self._calculateUAGrad(fprev, aprev)
if NP.mean(abs(grad)) < gradEpsStop:
a = aprev
f = fprev
break
# The following is just a Newton step (eq. (3.18) [1]) to maximize
# log(p(F|X,y)) over F
g = self._likelihood.gradient_log(fprev, self._y)
W = self._calculateW(fprev)
b = W*(fprev-m) + g
a = self._calculateUAa(b, W)
if line_search:
(f, a, obj) = self._lineSearch(a, aprev, m)
else:
f = self._rdotK(a) + m
obj = self._likelihood.log(f, self._y) - (f-m).dot(a)/2.0
if abs(objprev-obj) < objEpsStop :
grad = self._calculateUAGrad(f, a)
break
if obj > objprev:
fprev = f
objprev = obj
aprev = a
else:
if line_search:
grad = self._calculateUAGrad(fprev, aprev)
a = aprev
f = fprev
break
line_search = True
ii+=1
self._lasta = a
err = NP.mean(abs(grad))
if err > gradEpsErr:
failed = True
failedMsg = 'Gradient not too small in the Laplace update approximation.\n'
failedMsg = failedMsg+"Problem in the f mode estimation. |grad|_{mean} = %.6f." % (err,)
if ii>=maxIter:
failed = True
failedMsg = 'Laplace update approximation did not converge in less than maxIter.'
if self._debug:
self._debugUACalls.append(self.DebugUACall(
ii, grad, not failed, self.beta, self._sig02, self._sig12, self._sign2))
if failed:
Pr.prin('Laplace update approximation failed. The failure message is the following.')
Pr.prin(failedMsg)
sys.exit('Stopping program.')
self._updateApproximationEnd(f, a)
self._updateApproximationCount = 0
def _updateApproximation(self):
'''
Calculates the Laplace approximation for the posterior.
It can be defined by two variables: f mode and W at f mode.
'''
if self._updateApproximationCount == 0:
return
if self._is_kernel_zero():
self._updateApproximationCount = 0
return
self._updateApproximationBegin()
gradEpsStop = 1e-10
objEpsStop = 1e-8
gradEpsErr = 1e-3
self._mean = self._calculateMean()
m = self._mean
if self._lasta is None or self._lasta.shape[0] != self._N:
aprev = NP.zeros(self._N)
else:
aprev = self._lasta
fprev = self._rdotK(aprev) + m
objprev = self._likelihood.log(fprev,
self._y) - (fprev - m).dot(aprev) / 2.0
ii = 0
line_search = False
maxIter = 1000
failed = False
failedMsg = ''
while ii < maxIter:
grad = self._calculateUAGrad(fprev, aprev)
if NP.mean(abs(grad)) < gradEpsStop:
a = aprev
f = fprev
break
# The following is just a Newton step (eq. (3.18) [1]) to maximize
# log(p(F|X,y)) over F
g = self._likelihood.gradient_log(fprev, self._y)
W = self._calculateW(fprev)
b = W * (fprev - m) + g
a = self._calculateUAa(b, W)
if line_search:
(f, a, obj) = self._lineSearch(a, aprev, m)
else:
f = self._rdotK(a) + m
obj = self._likelihood.log(f, self._y) - (f - m).dot(a) / 2.0
if abs(objprev - obj) < objEpsStop:
grad = self._calculateUAGrad(f, a)
break
if obj > objprev:
fprev = f
objprev = obj
aprev = a
else:
if line_search:
grad = self._calculateUAGrad(fprev, aprev)
a = aprev
f = fprev
break
line_search = True
ii += 1
self._lasta = a
err = NP.mean(abs(grad))
if err > gradEpsErr:
failed = True
failedMsg = 'Gradient not too small in the Laplace update approximation.\n'
failedMsg = failedMsg + "Problem in the f mode estimation. |grad|_{mean} = %.6f." % (
err, )
if ii >= maxIter:
failed = True
failedMsg = 'Laplace update approximation did not converge in less than maxIter.'
if self._debug:
self._debugUACalls.append(
self.DebugUACall(ii, grad, not failed, self.beta, self._sig02,
self._sig12, self._sign2))
if failed:
Pr.prin(
'Laplace update approximation failed. The failure message is the following.'
)
Pr.prin(failedMsg)
sys.exit('Stopping program.')
self._updateApproximationEnd(f, a)
self._updateApproximationCount = 0
def _updateApproximation(self):
'''
Calculates the Laplace approximation for the posterior.
It can be defined by two variables: f mode and W at f mode.
'''
if self._updateApproximationCount == 0:
return
if self._is_kernel_zero():
self._updateApproximationCount = 0
return
self._updateApproximationBegin()
self._mean = self._calculateMean()
m = self._mean
ttau = NP.zeros(self._N)
tnu = NP.zeros(self._N)
sig2_ = NP.zeros(self._N)
mu_ = NP.zeros(self._N)
prevsig2 = self._dKn()
prevmu = m.copy()
converged = False
outeriter = 1
iterMax = 1000
while outeriter <= iterMax and not converged:
tau_ = 1.0 / prevsig2 - ttau
nu_ = prevmu / prevsig2 - tnu
tt = tau_**2 + tau_
stt = NP.sqrt(tt)
c = (self._y * nu_) / stt
# dan: using _cdf and _pdf instead of cdf and pdf I avoid
# a lot of overhead due to error checking and other things
nc_hz = NP.exp(ST.norm._logpdf(c) - ST.norm._logcdf(c))
hmu = nu_ / tau_ + nc_hz * self._y / stt
hsig2 = 1.0 / tau_ - (nc_hz / tt) * (c + nc_hz)
tempttau = 1.0 / hsig2 - tau_
ok = NP.bitwise_and(~NP.isnan(tempttau), abs(tempttau) > 1e-8)
ttau[ok] = tempttau[ok]
tnu[ok] = hmu[ok] / hsig2[ok] - nu_[ok]
(sig2, mu) = self._calculateSig2Mu(ttau, tnu, m)
lastMuDiff = abs(mu - prevmu).max()
lastSig2Diff = abs(sig2 - prevsig2).max()
if lastMuDiff < 1e-6 and lastSig2Diff < 1e-6:
converged = True
prevmu = mu
prevsig2 = sig2
outeriter += 1
if outeriter > iterMax:
Pr.prin('EP did not converge. Printing debug information...')
Pr.prin('beta ' + str(self.beta))
Pr.prin('sig02 ' + str(self.sig02) + ' sig12 ' + str(self.sig12) +
' sign2 ' + str(self.sign2))
Pr.prin('abs(mu-prevmu).max() ' + str(lastMuDiff))
Pr.prin('abs(sig2-prevsig2).max() ' + str(lastSig2Diff))
self._updateApproximationEnd(ttau, tnu, tau_, nu_)
self._updateApproximationCount = 0