def __init__(self,config,ds,run,nsample,thin,burnin,levels=-1,randomize=True,randomSamplerOrder=True):
self.config = Configuration(config,randomizePriors=randomSamplerOrder)
if randomize:
self.config.randomize()
self.run = run
self.nsample = nsample
self.thin = thin
self.burnin = burnin
self.levels = levels
self.randomSamplerOrder = randomSamplerOrder
# default to using all levels
if self.levels == -1:
self.levels = self.config.levels
# how many functions to use
self.f = sum(self.config.nf[:self.levels+1])
self.x, self.p, self.dm, self.kernels, self.priors = self.config.get()
self.yKernel = self.kernels[0]
for i in range(self.levels+1):
k = 'k%d'%(i+1)
self.__dict__[k] = self.kernels[i+1]
self.kernels = self.kernels[1:self.levels+2]
if self.config.hierarchy and self.levels < self.config.levels:
# self.yKernel = self.yKernel.kernels[-1]
args = self.yKernel.kernels[:self.levels]+[self.yKernel.kernels[-1]]
self.yKernel = Hierarchical(*args)
self.y = pd.read_csv(os.path.join(config,ds,'data.csv')).values
self.dm = self.dm[:self.f,:]
# self.model = self.config.model
self.model = Model(self.x,self.y,self.dm)
if randomize:
self.model.beta = self.config.model.beta[:,:self.f]
self._buildSamplers()
kwargs = {}
kwargs['yKernel'] = self.yKernel
for i,k in enumerate(self.kernels):
kwargs['k%d'%(i+1)] = k
self.freeze = Freezer(model=self.model,**kwargs)
self.startSample = self.freeze.freeze()
self.samples = []
def __init__(self,config,ds,run,nsample,thin,burnin,levels=-1,*args,**kwargs):
Sample.__init__(self,config,ds,run,nsample,thin,burnin,levels,*args,**kwargs)
self.levels=self.config.levels-1
kern = self.__dict__['k%d'%(self.levels+2)]
self.yKernelOld = self.yKernel
self.yKernel = Addition(self.yKernel,kern)
del self.kernels[-1]
kwargs = {}
kwargs['yKernel'] = self.yKernel
for i,k in enumerate(self.kernels):
kwargs['k%d'%(i+1)] = k
self.freeze = Freezer(model=self.model,**kwargs)
self.startSample = self.freeze.freeze()
class Sample(object):
def __init__(self,config,ds,run,nsample,thin,burnin,levels=-1,randomize=True,randomSamplerOrder=True):
self.config = Configuration(config,randomizePriors=randomSamplerOrder)
if randomize:
self.config.randomize()
self.run = run
self.nsample = nsample
self.thin = thin
self.burnin = burnin
self.levels = levels
self.randomSamplerOrder = randomSamplerOrder
# default to using all levels
if self.levels == -1:
self.levels = self.config.levels
# how many functions to use
self.f = sum(self.config.nf[:self.levels+1])
self.x, self.p, self.dm, self.kernels, self.priors = self.config.get()
self.yKernel = self.kernels[0]
for i in range(self.levels+1):
k = 'k%d'%(i+1)
self.__dict__[k] = self.kernels[i+1]
self.kernels = self.kernels[1:self.levels+2]
if self.config.hierarchy and self.levels < self.config.levels:
# self.yKernel = self.yKernel.kernels[-1]
args = self.yKernel.kernels[:self.levels]+[self.yKernel.kernels[-1]]
self.yKernel = Hierarchical(*args)
self.y = pd.read_csv(os.path.join(config,ds,'data.csv')).values
self.dm = self.dm[:self.f,:]
# self.model = self.config.model
self.model = Model(self.x,self.y,self.dm)
if randomize:
self.model.beta = self.config.model.beta[:,:self.f]
self._buildSamplers()
kwargs = {}
kwargs['yKernel'] = self.yKernel
for i,k in enumerate(self.kernels):
kwargs['k%d'%(i+1)] = k
self.freeze = Freezer(model=self.model,**kwargs)
self.startSample = self.freeze.freeze()
self.samples = []
def load(self,f):
samples = Analysis.loadSamples(f)
self.startSample = samples[0]
self.samples = samples
self.freeze.push(**self.samples[-1])
def _buildSamplers(self):
self.samplers = []
if self.config.hierarchy:
self.samplers.append((self.yKernel.__dict__['k%d'%(self.levels+1)],'sigma',
Slice('ySigma',
lambda x: self.model.dataLikelihood(self.yKernel,**{'k%d_sigma'%(self.config.levels+1):x}),
lambda x: self.priors['yKernel']['sigma'].logpdf(x),
self.config.config.getfloat('yKernel','slice-w'),self.config.config.getfloat('yKernel','slice-m'),logspace=True)
))
for i in range(self.levels):
self.samplers.append((self.yKernel.__dict__['k%d'%(i+1)],'sigma',
Slice('sigma',
lambda x,i=i: self.model.dataLikelihood(self.yKernel,**{'k%d_sigma'%i:x}),
lambda x: self.priors['k%d'%(i+1)]['sigma'].logpdf(x),
self.config.config.getfloat('k%d'%(i+1),'slice-w'),self.config.config.getfloat('k%d'%(i+1),'slice-m'),logspace=True)
))
self.samplers.append((self.yKernel.__dict__['k%d'%(i+1)],'lengthscale',
Slice('lengthscale',
lambda x,i=i: self.model.dataLikelihood(self.yKernel,**{'k%d_lengthscale'%i:x}),
lambda x: self.priors['k%d'%(i+1)]['lengthscale'].logpdf(x),
self.config.config.getfloat('k%d'%(i+1),'slice-w'),self.config.config.getfloat('k%d'%(i+1),'slice-m'),logspace=True)
))
self.samplers.append((self.k1,'sigma',
PriorSlice('k1-sigma', self.model,
self.priors['functions'][0], self.priors['k1']['sigma'],
'sigma',
self.config.config.getfloat('k1','slice-w'),
self.config.config.getfloat('k1','slice-m'),
logspace=True)
))
self.samplers.append((self.k1,'lengthscale',
PriorSlice('k1-lengthscale', self.model,
self.priors['functions'][0], self.priors['k1']['lengthscale'],
'lengthscale',
self.config.config.getfloat('k1','slice-w'),
self.config.config.getfloat('k1','slice-m'),
logspace=True)
))
else:
self.samplers.append((self.yKernel,'sigma',
Slice('ySigma',
lambda x: self.model.dataLikelihood(self.yKernel,sigma=x),
lambda x: self.priors['yKernel']['sigma'].logpdf(x),
self.config.config.getfloat('yKernel','slice-w'),self.config.config.getfloat('yKernel','slice-m'),logspace=True)
))
for i in range(self.levels+1):
k = 'k%d'%(i+1)
self.samplers.append((self.__dict__[k],'sigma',
PriorSlice('%s-sigma'%k, self.model,
self.priors['functions'][i], self.priors[k]['sigma'],
'sigma',
self.config.config.getfloat(k,'slice-w'),
self.config.config.getfloat(k,'slice-m'),
logspace=True)
))
self.samplers.append((self.__dict__[k],'lengthscale',
PriorSlice('%s-lengthscale'%k, self.model,
self.priors['functions'][i], self.priors[k]['lengthscale'],
'lengthscale',
self.config.config.getfloat(k,'slice-w'),
self.config.config.getfloat(k,'slice-m'),
logspace=True)
))
def sample(self):
for i in range(len(self.samples),self.nsample):
self._sampleIteration()
if i % self.thin == 0 and i > self.burnin:
self.samples.append(self.freeze.freeze())
def _sampleIteration(self):
order = range(self.levels+1+len(self.samplers))
if self.randomSamplerOrder:
order = np.random.choice(order,len(order),replace=False)
for o in order:
if o < self.config.f:
prior = self.priors['functions'][o]
prior.sample(self.model,self.yKernel)
else:
obj, param, sampler = self.samplers[o-self.levels-1]
obj.__dict__[param] = sampler.sample(obj.__dict__[param])
# print
# for f,prior in self.priors['functions'].iteritems():
# if f > self.levels:
# continue
# prior.sample(self.model,self.yKernel)
#
# for obj, param, sampler in self.samplers:
# obj.__dict__[param] = sampler.sample(obj.__dict__[param])
def save(self,dir):
self.freeze.save(self.samples,os.path.join(dir,'samples.json'))
self.freeze.save([self.startSample],os.path.join(dir,'startSample.json'))
ind = min(np.where(f < config.cumnf)[0]) + 1
# print f,ind
kernel = kernels[ind]
betaTrue[:, f] = scipy.stats.multivariate_normal.rvs(
np.zeros(x.shape[0]), kernel.K(x))
# y = np.dot(betaTrue,dm) + scipy.stats.multivariate_normal.rvs(np.zeros(x.shape[0]),kernels[0].K(x),size=p).reshape((x.shape[0],p))
y = np.zeros((x.shape[0], dm.shape[1]))
for i in range(dm.shape[1]):
y[:, i] = np.dot(betaTrue,
dm[:, i]) + scipy.stats.multivariate_normal.rvs(
np.zeros(x.shape[0]), kernels[0].K(x))
model = Model(x, y, dm)
model.beta = betaTrue
kwargs = {
'k%d' % (i + 1): kernels[i + 1]
for i in range(config.levels + 1)
}
freeze = Freezer(yKernel=kernels[0], model=model, **kwargs)
trueSample = freeze.freeze()
freeze.save([trueSample],
os.path.join(args.configuration, args.label,
'parameters-true.json'))
pd.DataFrame(y).to_csv(os.path.join(args.configuration, args.label,
'data.csv'),
index=False)
class SampleIntegrate(Sample):
"""Integrate out the lowest level of hierarchical variation."""
def __init__(self,config,ds,run,nsample,thin,burnin,levels=-1,*args,**kwargs):
Sample.__init__(self,config,ds,run,nsample,thin,burnin,levels,*args,**kwargs)
self.levels=self.config.levels-1
kern = self.__dict__['k%d'%(self.levels+2)]
self.yKernelOld = self.yKernel
self.yKernel = Addition(self.yKernel,kern)
del self.kernels[-1]
kwargs = {}
kwargs['yKernel'] = self.yKernel
for i,k in enumerate(self.kernels):
kwargs['k%d'%(i+1)] = k
self.freeze = Freezer(model=self.model,**kwargs)
self.startSample = self.freeze.freeze()
def _buildSamplers(self):
self.samplers = []
self.samplers.append((self.yKernel,'k1_sigma',
Slice('ySigma',
lambda x: self.model.dataLikelihood(self.yKernel,k1_sigma=x),
lambda x: self.priors['yKernel']['sigma'].logpdf(x),
self.config.config.getfloat('yKernel','slice-w'),self.config.config.getfloat('yKernel','slice-m'),logspace=True)
))
for i in range(self.levels+1):
k = 'k%d'%(i+1)
self.samplers.append((self.__dict__[k],'sigma',
PriorSlice('%s-sigma'%k, self.model,
self.priors['functions'][i], self.priors[k]['sigma'],
'sigma',
self.config.config.getfloat(k,'slice-w'),
self.config.config.getfloat(k,'slice-m'),
logspace=True)
))
self.samplers.append((self.__dict__[k],'lengthscale',
PriorSlice('%s-lengthscale'%k, self.model,
self.priors['functions'][i], self.priors[k]['lengthscale'],
'lengthscale',
self.config.config.getfloat(k,'slice-w'),
self.config.config.getfloat(k,'slice-m'),
logspace=True)
))
k = 'k%d'%(self.levels+1)
self.samplers.append((self.yKernel,'k2_sigma',
Slice('ySigma2',
lambda x: self.model.dataLikelihood(self.yKernel,k2_sigma=x),
lambda x: self.priors[k]['sigma'].logpdf(x),
self.config.config.getfloat(k,'slice-w'),self.config.config.getfloat(k,'slice-m'),logspace=True)
))
self.samplers.append((self.yKernel,'k2_lengthscale',
Slice('yLengthscale2',
lambda x: self.model.dataLikelihood(self.yKernel,k2_lengthscale=x),
lambda x: self.priors[k]['lengthscale'].logpdf(x),
self.config.config.getfloat(k,'slice-w'),self.config.config.getfloat(k,'slice-m'),logspace=True)
))
def _sampleIteration(self):
order = range(self.levels+1+len(self.samplers))
if self.randomSamplerOrder:
order = np.random.choice(order,len(order),replace=False)
for o in order:
if o <= self.levels:
prior = self.priors['functions'][o]
prior.sample(self.model,self.yKernel)
else:
obj, param, sampler = self.samplers[o-self.levels-1]
# print param, obj
if obj == self.yKernelOld:
obj = self.yKernel
if param == 'k1_sigma':
obj.__dict__['k1'].sigma = sampler.sample(obj.__dict__['k1'].sigma)
elif param == 'k2_sigma':
obj.__dict__['k2'].sigma = sampler.sample(obj.__dict__['k2'].sigma)
elif param == 'k2_lengthscale':
obj.__dict__['k2'].lengthscale = sampler.sample(obj.__dict__['k2'].lengthscale)
else:
raise ValueError("unknown parameter %s!"%param)
else:
obj.__dict__[param] = sampler.sample(obj.__dict__[param])
lambda x: prior2.loglikelihood(model.beta, sigma=x),
lambda x: scipy.stats.uniform(1e-2, 1e1).logpdf(x),
.1,
10,
logspace=True)
k2LengthscaleSlice = Slice(
'k2Lengthscale',
lambda x: prior2.loglikelihood(model.beta, lengthscale=x),
lambda x: scipy.stats.uniform(1e-2, 1e1).logpdf(x),
.1,
10,
logspace=True)
samples = []
freeze = Freezer(yKernel=yKernel, k1=k1, k2=k2, model=model)
thin = 10
burnin = 0
nsample = 4000
ll = []
for i in range(nsample):
prior.sample(model, yKernel)
prior2.sample(model, yKernel)
yKernel.sigma = ySigmaSlice.sample(yKernel.sigma)
k1.sigma = kSigmaSlice.sample(k1.sigma)
k1.lengthscale = kLengthscaleSlice.sample(k1.lengthscale)
k2.sigma = k2SigmaSlice.sample(k2.sigma)
lambda x: model.dataLikelihood(yKernel,sigma=x),
lambda x: scipy.stats.uniform(1e-6,1e1).logpdf(x),
.1,2,logspace=True)
kSigmaSlice = Slice('kSigma',
lambda x: prior.loglikelihood(model.beta,sigma=x),
lambda x: scipy.stats.uniform(1e-2,1e1).logpdf(x),
.2,5,logspace=True)
kLengthscaleSlice = Slice('kLengthscale',
lambda x: prior.loglikelihood(model.beta,lengthscale=x),
lambda x: scipy.stats.uniform(1e-2,1e1).logpdf(x),
.2,5,logspace=True)
samples = []
freeze = Freezer(yKernel=yKernel,k1=k1,model=model)
trueParams = freeze.freeze()
thin = 10
burnin = 0
nsample = 1000
ll = []
for i in range(nsample):
prior.sample(model,yKernel)
yKernel.sigma = ySigmaSlice.sample(yKernel.sigma)
k1.sigma = kSigmaSlice.sample(k1.sigma)
k1.lengthscale = kLengthscaleSlice.sample(k1.lengthscale)