def cv(self,dataTest=None):
"""
yxy"
start with different initial estimates and cross-validating,
Cross validation value used for comparing different function optimization outputs
can be computed with a separate dataset test or with training dataset
<math>
nll(testing dataset)/nll(zero function)
dataTest (Optional):
If declared, will be used instead of self.data for computing the nll
"""
if dataTest is None:
data = self.data
else:
data = data_validation.data_structure(dataTest)
nll_f = self.nll(dataTest=data)
nll_z = -np.sum(stats.norm.logpdf(np.zeros_like(data['Y']),loc=data['Y'],
scale=data['Var']**0.5))
return nll_f/nll_z
def __init__(self,data,**kwargs):
self.data = data_validation.data_structure(data)
self.debug = kwargs.get('debug',0)
self.name = 'Log Function with levmar optimization'
#self.params = kwargs.get('params_0',None)
self.params = np.ones((self.data['Y'].shape[1],4))
def __init__(self, data, **kwargs):
super(GPcore, self).__init__()
#Main Data
self.name = 'GPcore'
self.data = data_validation.data_structure(data)
self.all_mac_dict = kwargs.pop('all_mac_dict', None)
self.ndp, self.nap = self.data['Y'].shape #data points, #access points
#check data
assert self.data['X'].shape[0] == self.ndp
assert self.data['X'].shape[1] == 2
#Likelihood params
self.likelihood = kwargs.pop('likelihood', lh.Gaussian(**kwargs))
#raise error if likelihood is not a derived class of Likelihood
assert isinstance(self.likelihood, lh.Likelihood)
self.add_noise_var = kwargs.pop('add_noise_var', .0005)
#Xtest
self.Xtest = kwargs.pop('Xtest', None)
self.Xtest_num = kwargs.pop(
'Xtest_num', 75) #num of points for each Xtest dimension
self.Xtest_factor = kwargs.pop(
'Xtest_factor', .25) #Xtest = min/max x +- x_factor*span
if self.Xtest is None:
x0 = np.min(self.data['X'][:, 0])
x1 = np.max(self.data['X'][:, 0])
xspn = x1 - x0
self.xmin = x0 - self.Xtest_factor * xspn
self.xmax = x1 + self.Xtest_factor * xspn
y0 = np.min(self.data['X'][:, 1])
y1 = np.max(self.data['X'][:, 1])
yspn = y1 - y0
self.ymin = y0 - self.Xtest_factor * yspn
self.ymax = y1 + self.Xtest_factor * yspn
temp_x = np.linspace(self.xmin, self.xmax, self.Xtest_num)
temp_y = np.linspace(self.ymin, self.ymax, self.Xtest_num)
self.Xtest = mesh(temp_x, temp_y)
#Sampling params
self.sampling = kwargs.pop('sampling',
sampling.accept_reject_by_regions_map)
self.nsamples = kwargs.get('nsamples', 800)
#Debug params
self.debug = kwargs.get('debug', False)
self.verbose = kwargs.get('verbose', False)
if self.debug:
print 'class GP init works'
def __init__(self,data,**kwargs):
super(PathLoss,self).__init__()
self.data = data_validation.data_structure(data)
self.debug = kwargs.get('debug',0)
self.params = kwargs.get('params_0',None)
if self.params is None:
self.params = self.params_init()
else:
if self.params.ndim == 1: #casting into [1,np]
self.params = np.reshape(self.params,(1,self.params.shape[0]))
#Verify params_0 has adequate format
assert params.shape[0] == data['Y'].shape[1]
if self.debug:
print 'class PathLoss init'
def nll(self,dataTest=None):
"""
Negative log likelihood
dataTest (Optional):
If declared, will be used instead of self.data for computing the nll
"""
if dataTest is None:
data = self.data
else:
data = data_validation.data_structure(dataTest)
Ys = self.mean_val(data['X']) #bounded estimation
log_prob = stats.norm.logpdf(Ys,loc=data['Y'],scale=data['Var']**0.5)
return -np.sum(log_prob)
