def sampleGP(trainingSet,covar,mean=None,nsamples=100,
step=None,fix=None,metropolis=False,markovpy=False):
"""
NAME:
sampleGP
PURPOSE:
sample a GP
INPUT:
trainingSet - a trainingSet instance
covar - an instance of your covariance function of
choice, with initialized parameters
mean= an instance of your mean function of choice, with initialized
parameters
nsamples - number of samples desired
step= step-size for slice creation or metropolis sampling
(number or list)
fix= None or list of parameters to hold fixed
metropolis= if True, use Metropolis sampling
markovpy= if True, use markovpy sampling
OUTPUT:
list of outcovarFunc
HISTORY:
2010-08-08 - Written - Bovy (NYU)
"""
#Put in dummy mean if mean is None
if mean is None:
from flexgp.zeroMean import meanFunc
mean= meanFunc()
noMean= True
out= [covar]
else:
noMean= False
out= [(covar,mean)]
#Pack the covariance parameters
(params,packing)= pack_params(covar,mean,fix)
if step is None:
step= [0.1 for ii in range(len(params))]
#Grab the covariance class
covarFuncName= inspect.getmodule(covar).__name__
thisCovarClass= __import__(covarFuncName)
meanFuncName= inspect.getmodule(mean).__name__
thisMeanClass= __import__(meanFuncName)
#Set up isDomainFinite, domain, and create_method, even when metropolis
isDomainFinite, domain, create_method= [], [], []
covarIsDomainFinite= covar.isDomainFinite()
covarDomain= covar.paramsDomain()
covarCreate= covar.create_method()
meanIsDomainFinite= mean.isDomainFinite()
meanDomain= mean.paramsDomain()
meanCreate= mean.create_method()
for ii in range(len(packing.GPhyperPackingList)):
p= packing.GPhyperPackingList[ii]
try:
for jj in range(packing.GPhyperPackingDim[ii]):
isDomainFinite.append(covarIsDomainFinite[p])
domain.append(covarDomain[p])
create_method.append(covarCreate[p])
except KeyError:
for jj in range(len(packing.GPhyperPackingDim[ii])):
isDomainFinite.append(meanIsDomainFinite[p])
domain.append(meanDomain[p])
create_method.append(meanCreate[p])
if len(packing.GPhyperPackingList) == 1: #one-d
isDomainFinite= isDomainFinite[0]
domain= domain[0]
create_method= create_method[0]
if isinstance(step,(list,numpy.ndarray)): step= step[0]
if not metropolis and not markovpy:
#slice sample the marginal likelihood
samples= bovy_mcmc.slice(params,step,
_lnpdf,(trainingSet,packing,thisCovarClass,
thisMeanClass),
isDomainFinite=isDomainFinite,
domain=domain,
nsamples=nsamples,
create_method=create_method)
elif metropolis:
samples, faccept= bovy_mcmc.metropolis(params,step,
_lnpdf,(trainingSet,packing,
thisCovarClass,
thisMeanClass),
symmetric=True,
nsamples=nsamples)
if numpy.any((faccept < 0.15)) or numpy.any((faccept > 0.6)):
print "WARNING: Metropolis acceptance ratio was < 0.15 or > 0.6 for a direction"
print "Full acceptance ratio list:"
print faccept
elif markovpy:
samples= bovy_mcmc.markovpy(params,step,
_lnpdf,(trainingSet,packing,thisCovarClass,
thisMeanClass),
isDomainFinite=isDomainFinite,
domain=domain,
nsamples=nsamples)
print nsamples, len(samples)
if noMean:
for ii in range(nsamples):
hyperParamsDict= unpack_params(samples[ii],packing)
out.append(packing.covarFunc(**hyperParamsDict))
return out
else:
for ii in range(nsamples):
hyperParamsDict= unpack_params(samples[ii],packing)
out.append((packing.covarFunc(**hyperParamsDict),
packing.meanFunc(**hyperParamsDict)))
return out
def trainGP(trainingSet,covar,mean=None,filename=None,ext='h5',
useDerivs=True,fix=None):
"""
NAME:
trainGP
PURPOSE:
train a GP
INPUT:
trainingSet - a trainingSet instance
covar - an instance of your covariance function of
choice, with initialized parameters
mean= an instance of your mean function of choice, with initialized
parameters
filename - filename for output
ext - format of the output file (checks extension if none given,
if this is not 'fit' or 'fits' assumes HDF5
useDerivs - use the derivative of the objective function to optimize
fix= fix these parameters of the mean and covariance functions
OUTPUT:
trained GP= outcovarFunc
HISTORY:
2010-07-25 - Written - Bovy (NYU)
"""
#Put in dummy mean if mean is None
if mean is None:
from flexgp.zeroMean import meanFunc
mean= meanFunc()
noMean= True
else: noMean= False
#Pack the covariance and mean parameters
(params,packing)= pack_params(covar,mean,fix)
#Grab the covariance and mean class
covarFuncName= inspect.getmodule(covar).__name__
thisCovarClass= __import__(covarFuncName)
meanFuncName= inspect.getmodule(mean).__name__
thisMeanClass= __import__(meanFuncName)
#Optimize the marginal likelihood
#Sort of good ftol, assuming data has been pre-processed
try:
if useDerivs:
raise NotImplementedError()
params= optimize.fmin_cg(marginalLikelihood,params,
fprime=derivMarginalLikelihood,
args=(trainingSet,packing,
thisCovarClass,useDerivs))
else:
params= optimize.fmin_powell(marginalLikelihood,params,
args=(trainingSet,packing,
thisCovarClass,
thisMeanClass,
useDerivs))
except numpy.linalg.linalg.LinAlgError:
raise
if not filename == None:
write_train_out(params,packing,filename=filename,ext=ext)
if packing.nhyperParams == 1: params= numpy.array([params])
hyperParamsDict= unpack_params(params,packing)
outcovarFunc= packing.covarFunc(**hyperParamsDict)
if noMean:
return outcovarFunc
outmeanFunc= packing.meanFunc(**hyperParamsDict)
return (outcovarFunc,outmeanFunc)
