apunkt/GPy
1
0
Fork 0
mirror of https://github.com/SheffieldML/GPy.git synced 2026-05-30 14:35:15 +02:00
Code Issues Projects Releases Packages Wiki Activity Actions

merge the changes

Browse source
This commit is contained in:
Fariba 2015-07-27 17:42:01 +01:00
commit fc0e8f3e7e
66 changed files with 55761 additions and 702 deletions
Download patch file Download diff file Expand all files Collapse all files

1
GPy/core/parameterization/index_operations.py
View file

@ -5,6 +5,7 @@ import numpy
from numpy.lib.function_base import vectorize
from .lists_and_dicts import IntArrayDict
from functools import reduce
from .transformations import Transformation
def extract_properties_to_index(index, props):
prop_index = dict()

5
GPy/core/parameterization/param.py
View file

@ -38,6 +38,11 @@ class Param(Parameterizable, ObsAr):
Fixing parameters will fix them to the value they are right now. If you change
the fixed value, it will be fixed to the new value!
Important Note:
Multilevel indexing (e.g. self[:2][1:]) is not supported and might lead to unexpected behaviour.
Try to index in one go, using boolean indexing or the numpy builtin
np.index function.
See :py:class:`GPy.core.parameterized.Parameterized` for more details on constraining etc.
"""

43
GPy/core/parameterization/parameter_core.py
View file

@ -430,23 +430,38 @@ class Indexable(Nameable, Updateable):
def log_prior(self):
"""evaluate the prior"""
if self.priors.size > 0:
x = self.param_array
#py3 fix
#return reduce(lambda a, b: a + b, (p.lnpdf(x[ind]).sum() for p, ind in self.priors.iteritems()), 0)
return reduce(lambda a, b: a + b, (p.lnpdf(x[ind]).sum() for p, ind in self.priors.items()), 0)
return 0.
if self.priors.size == 0:
return 0.
x = self.param_array
#evaluate the prior log densities
log_p = reduce(lambda a, b: a + b, (p.lnpdf(x[ind]).sum() for p, ind in self.priors.items()), 0)
#account for the transformation by evaluating the log Jacobian (where things are transformed)
log_j = 0.
priored_indexes = np.hstack([i for p, i in self.priors.items()])
for c,j in self.constraints.items():
if not isinstance(c, Transformation):continue
for jj in j:
if jj in priored_indexes:
log_j += c.log_jacobian(x[jj])
return log_p + log_j
def _log_prior_gradients(self):
"""evaluate the gradients of the priors"""
if self.priors.size > 0:
x = self.param_array
ret = np.zeros(x.size)
#py3 fix
#[np.put(ret, ind, p.lnpdf_grad(x[ind])) for p, ind in self.priors.iteritems()]
[np.put(ret, ind, p.lnpdf_grad(x[ind])) for p, ind in self.priors.items()]
return ret
return 0.
if self.priors.size == 0:
return 0.
x = self.param_array
ret = np.zeros(x.size)
#compute derivate of prior density
[np.put(ret, ind, p.lnpdf_grad(x[ind])) for p, ind in self.priors.items()]
#add in jacobian derivatives if transformed
priored_indexes = np.hstack([i for p, i in self.priors.items()])
for c,j in self.constraints.items():
if not isinstance(c, Transformation):continue
for jj in j:
if jj in priored_indexes:
ret[jj] += c.log_jacobian_grad(x[jj])
return ret
#===========================================================================
# Tie parameters together

2
GPy/core/parameterization/parameterized.py
View file

@ -9,7 +9,7 @@ from .param import ParamConcatenation
from .parameter_core import HierarchyError, Parameterizable, adjust_name_for_printing
import logging
from GPy.core.parameterization.index_operations import ParameterIndexOperationsView
from .index_operations import ParameterIndexOperationsView
logger = logging.getLogger("parameters changed meta")
class ParametersChangedMeta(type):

49
GPy/core/parameterization/priors.py
View file

@ -522,16 +522,9 @@ class DGPLVM(Prior):
"""
domain = _REAL
# _instances = []
# def __new__(cls, mu, sigma): # Singleton:
# if cls._instances:
# cls._instances[:] = [instance for instance in cls._instances if instance()]
# for instance in cls._instances:
# if instance().mu == mu and instance().sigma == sigma:
# return instance()
# o = super(Prior, cls).__new__(cls, mu, sigma)
# cls._instances.append(weakref.ref(o))
# return cls._instances[-1]()
def __new__(cls, sigma2, lbl, x_shape):
return super(Prior, cls).__new__(cls, sigma2, lbl, x_shape)
def __init__(self, sigma2, lbl, x_shape):
self.sigma2 = sigma2
@ -730,7 +723,7 @@ class DGPLVM(Prior):
# ******************************************
from parameterized import Parameterized
from .. import Parameterized
from .. import Param
class DGPLVM_Lamda(Prior, Parameterized):
"""
@ -758,12 +751,12 @@ class DGPLVM_Lamda(Prior, Parameterized):
self.sigma2 = sigma2
# self.x = x
self.lbl = lbl
self.lamda = lamda
self.lamda = lamda
self.classnum = lbl.shape[1]
self.datanum = lbl.shape[0]
self.x_shape = x_shape
self.dim = x_shape[1]
self.lamda = Param('lamda', np.diag(lamda))
self.lamda = Param('lamda', np.diag(lamda))
self.link_parameter(self.lamda)
def get_class_label(self, y):
@ -789,7 +782,7 @@ class DGPLVM_Lamda(Prior, Parameterized):
M_i = np.zeros((self.classnum, self.dim))
for i in cls:
# Mean of each class
class_i = cls[i]
class_i = cls[i]
M_i[i] = np.mean(class_i, axis=0)
return M_i
@ -899,8 +892,8 @@ class DGPLVM_Lamda(Prior, Parameterized):
#!!!!!!!!!!!!!!!!!!!!!!!!!!!
#self.lamda.values[:] = self.lamda.values/self.lamda.values.sum()
xprime = x.dot(np.diagflat(self.lamda))
x = xprime
xprime = x.dot(np.diagflat(self.lamda))
x = xprime
# print x
cls = self.compute_cls(x)
M_0 = np.mean(x, axis=0)
@ -916,8 +909,8 @@ class DGPLVM_Lamda(Prior, Parameterized):
# This function calculates derivative of the log of prior function
def lnpdf_grad(self, x):
x = x.reshape(self.x_shape)
xprime = x.dot(np.diagflat(self.lamda))
x = xprime
xprime = x.dot(np.diagflat(self.lamda))
x = xprime
# print x
cls = self.compute_cls(x)
M_0 = np.mean(x, axis=0)
@ -951,14 +944,14 @@ class DGPLVM_Lamda(Prior, Parameterized):
# Because of the GPy we need to transpose our matrix so that it gets the same shape as out matrix (denominator layout!!!)
DPxprim_Dx = DPxprim_Dx.T
DPxprim_Dlamda = DPx_Dx.dot(x)
DPxprim_Dlamda = DPx_Dx.dot(x)
# Because of the GPy we need to transpose our matrix so that it gets the same shape as out matrix (denominator layout!!!)
DPxprim_Dlamda = DPxprim_Dlamda.T
DPxprim_Dlamda = DPxprim_Dlamda.T
self.lamda.gradient = np.diag(DPxprim_Dlamda)
self.lamda.gradient = np.diag(DPxprim_Dlamda)
# print DPxprim_Dx
return DPxprim_Dx
return DPxprim_Dx
# def frb(self, x):
@ -1139,8 +1132,8 @@ class DGPLVM_T(Prior):
# This function calculates log of our prior
def lnpdf(self, x):
x = x.reshape(self.x_shape)
xprim = x.dot(self.vec)
x = xprim
xprim = x.dot(self.vec)
x = xprim
# print x
cls = self.compute_cls(x)
M_0 = np.mean(x, axis=0)
@ -1156,11 +1149,11 @@ class DGPLVM_T(Prior):
# This function calculates derivative of the log of prior function
def lnpdf_grad(self, x):
x = x.reshape(self.x_shape)
xprim = x.dot(self.vec)
x = xprim
x = x.reshape(self.x_shape)
xprim = x.dot(self.vec)
x = xprim
# print x
cls = self.compute_cls(x)
cls = self.compute_cls(x)
M_0 = np.mean(x, axis=0)
M_i = self.compute_Mi(cls)
Sb = self.compute_Sb(cls, M_i, M_0)

50
GPy/core/parameterization/transformations.py
View file

@ -31,6 +31,16 @@ class Transformation(object):
raise NotImplementedError
def finv(self, model_param):
raise NotImplementedError
def log_jacobian(self, model_param):
"""
compute the log of the jacobian of f, evaluated at f(x)= model_param
"""
raise NotImplementedError
def log_jacobian_grad(self, model_param):
"""
compute the drivative of the log of the jacobian of f, evaluated at f(x)= model_param
"""
raise NotImplementedError
def gradfactor(self, model_param, dL_dmodel_param):
""" df(opt_param)_dopt_param evaluated at self.f(opt_param)=model_param, times the gradient dL_dmodel_param,
@ -74,9 +84,33 @@ class Logexp(Transformation):
if np.any(f < 0.):
print("Warning: changing parameters to satisfy constraints")
return np.abs(f)
def log_jacobian(self, model_param):
return np.where(model_param>_lim_val, model_param, np.log(np.exp(model_param+1e-20) - 1.)) - model_param
def log_jacobian_grad(self, model_param):
return 1./(np.exp(model_param)-1.)
def __str__(self):
return '+ve'
class Exponent(Transformation):
domain = _POSITIVE
def f(self, x):
return np.where(x<_lim_val, np.where(x>-_lim_val, np.exp(x), np.exp(-_lim_val)), np.exp(_lim_val))
def finv(self, x):
return np.log(x)
def gradfactor(self, f, df):
return np.einsum('i,i->i', df, f)
def initialize(self, f):
if np.any(f < 0.):
print("Warning: changing parameters to satisfy constraints")
return np.abs(f)
def log_jacobian(self, model_param):
return np.log(model_param)
def log_jacobian_grad(self, model_param):
return 1./model_param
def __str__(self):
return '+ve'
class NormalTheta(Transformation):
"Do not use, not officially supported!"
@ -417,22 +451,6 @@ class LogexpClipped(Logexp):
def __str__(self):
return '+ve_c'
class Exponent(Transformation):
# TODO: can't allow this to go to zero, need to set a lower bound. Similar with negative Exponent below. See old MATLAB code.
domain = _POSITIVE
def f(self, x):
return np.where(x<_lim_val, np.where(x>-_lim_val, np.exp(x), np.exp(-_lim_val)), np.exp(_lim_val))
def finv(self, x):
return np.log(x)
def gradfactor(self, f, df):
return np.einsum('i,i->i', df, f)
def initialize(self, f):
if np.any(f < 0.):
print("Warning: changing parameters to satisfy constraints")
return np.abs(f)
def __str__(self):
return '+ve'
class NegativeExponent(Exponent):
domain = _NEGATIVE
def f(self, x):

51
GPy/core/parameterization/variational.py
View file

@ -36,8 +36,9 @@ class NormalPrior(VariationalPrior):
variational_posterior.variance.gradient -= (1. - (1. / (variational_posterior.variance))) * 0.5
class SpikeAndSlabPrior(VariationalPrior):
def __init__(self, pi=None, learnPi=False, variance = 1.0, name='SpikeAndSlabPrior', **kw):
super(SpikeAndSlabPrior, self).__init__(name=name, **kw)
def __init__(self, pi=None, learnPi=False, variance = 1.0, group_spike=False, name='SpikeAndSlabPrior', **kw):
super(SpikeAndSlabPrior, self).__init__(name=name, **kw)
self.group_spike = group_spike
self.variance = Param('variance',variance)
self.learnPi = learnPi
if learnPi:
@ -50,7 +51,10 @@ class SpikeAndSlabPrior(VariationalPrior):
def KL_divergence(self, variational_posterior):
mu = variational_posterior.mean
S = variational_posterior.variance
gamma = variational_posterior.gamma.values
if self.group_spike:
gamma = variational_posterior.gamma.values[0]
else:
gamma = variational_posterior.gamma.values
if len(self.pi.shape)==2:
idx = np.unique(variational_posterior.gamma._raveled_index()/gamma.shape[-1])
pi = self.pi[idx]
@ -65,14 +69,21 @@ class SpikeAndSlabPrior(VariationalPrior):
def update_gradients_KL(self, variational_posterior):
mu = variational_posterior.mean
S = variational_posterior.variance
gamma = variational_posterior.gamma.values
if self.group_spike:
gamma = variational_posterior.gamma.values[0]
else:
gamma = variational_posterior.gamma.values
if len(self.pi.shape)==2:
idx = np.unique(variational_posterior.gamma._raveled_index()/gamma.shape[-1])
pi = self.pi[idx]
else:
pi = self.pi
variational_posterior.binary_prob.gradient -= np.log((1-pi)/pi*gamma/(1.-gamma))+((np.square(mu)+S)/self.variance-np.log(S)+np.log(self.variance)-1.)/2.
if self.group_spike:
dgamma = np.log((1-pi)/pi*gamma/(1.-gamma))/variational_posterior.num_data
else:
dgamma = np.log((1-pi)/pi*gamma/(1.-gamma))
variational_posterior.binary_prob.gradient -= dgamma+((np.square(mu)+S)/self.variance-np.log(S)+np.log(self.variance)-1.)/2.
mu.gradient -= gamma*mu/self.variance
S.gradient -= (1./self.variance - 1./S) * gamma /2.
if self.learnPi:
@ -154,13 +165,31 @@ class SpikeAndSlabPosterior(VariationalPosterior):
'''
The SpikeAndSlab distribution for variational approximations.
'''
def __init__(self, means, variances, binary_prob, name='latent space'):
def __init__(self, means, variances, binary_prob, group_spike=False, sharedX=False, name='latent space'):
"""
binary_prob : the probability of the distribution on the slab part.
"""
super(SpikeAndSlabPosterior, self).__init__(means, variances, name)
self.gamma = Param("binary_prob",binary_prob,Logistic(0.,1.))
self.link_parameter(self.gamma)
self.group_spike = group_spike
self.sharedX = sharedX
if sharedX:
self.mean.fix(warning=False)
self.variance.fix(warning=False)
if group_spike:
self.gamma_group = Param("binary_prob_group",binary_prob.mean(axis=0),Logistic(1e-10,1.-1e-10))
self.gamma = Param("binary_prob",binary_prob, __fixed__)
self.link_parameters(self.gamma_group,self.gamma)
else:
self.gamma = Param("binary_prob",binary_prob,Logistic(1e-10,1.-1e-10))
self.link_parameter(self.gamma)
def propogate_val(self):
if self.group_spike:
self.gamma.values[:] = self.gamma_group.values
def collate_gradient(self):
if self.group_spike:
self.gamma_group.gradient = self.gamma.gradient.reshape(self.gamma.shape).sum(axis=0)
def set_gradients(self, grad):
self.mean.gradient, self.variance.gradient, self.gamma.gradient = grad
@ -179,15 +208,15 @@ class SpikeAndSlabPosterior(VariationalPosterior):
n.parameters[dc['variance']._parent_index_] = dc['variance']
n.parameters[dc['binary_prob']._parent_index_] = dc['binary_prob']
n._gradient_array_ = None
oversize = self.size - self.mean.size - self.variance.size
n.size = n.mean.size + n.variance.size + oversize
oversize = self.size - self.mean.size - self.variance.size - self.gamma.size
n.size = n.mean.size + n.variance.size + n.gamma.size + oversize
n.ndim = n.mean.ndim
n.shape = n.mean.shape
n.num_data = n.mean.shape[0]
n.input_dim = n.mean.shape[1] if n.ndim != 1 else 1
return n
else:
return super(VariationalPrior, self).__getitem__(s)
return super(SpikeAndSlabPosterior, self).__getitem__(s)
def plot(self, *args, **kwargs):
"""