apunkt/GPy
1
0
Fork 0
mirror of https://github.com/SheffieldML/GPy.git synced 2026-05-09 03:52:39 +02:00
Code Issues Projects Releases Packages Wiki Activity Actions

new gradient handling way nicer

Browse source
This commit is contained in:
Max Zwiessele 2014-01-24 15:07:28 +00:00
parent f0ac290eb3
commit e128059377
6 changed files with 15 additions and 26 deletions
Download patch file Download diff file Expand all files Collapse all files

4
GPy/core/gp.py
View file

@ -54,8 +54,8 @@ class GP(Model):
print "defaulting to ", inference_method, "for latent function inference"
self.inference_method = inference_method
self.add_parameter(self.kern, gradient=self.dL_dtheta_K)
self.add_parameter(self.likelihood, gradient=lambda:self.posterior.dL_dtheta_lik)
self.add_parameter(self.kern)
self.add_parameter(self.likelihood)
self.parameters_changed()

6
GPy/core/model.py
View file

@ -34,9 +34,9 @@ class Model(Parameterized):
g = np.zeros(self.size)
try:
#[g.__setitem__(s, self.gradient_mapping[p]().flat) for p, s in itertools.izip(self._parameters_, self._param_slices_) if not p.is_fixed]
[g.__setitem__(s, p.gradient.flat) for p, s in itertools.izip(self._parameters_, self._param_slices_) if not p.is_fixed]
except KeyError:
raise KeyError, 'Gradient for {} not defined, please specify gradients for parameters to optimize'.format(p.name)
[p._collect_gradient(g[s]) for p, s in itertools.izip(self._parameters_, self._param_slices_) if not p.is_fixed]
except ValueError:
raise ValueError, 'Gradient for {} not defined, please specify gradients for parameters to optimize'.format(p.name)
return g
raise NotImplementedError, "this needs to be implemented to use the model class"

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

@ -26,9 +26,6 @@ class Param(ObservableArray, Constrainable):
:param name: name of the parameter to be printed
:param input_array: array which this parameter handles
:param gradient: callable with one argument, which is the model of this parameter
:param args: additional arguments to gradient
:param kwargs: additional keyword arguments to gradient
You can add/remove constraints by calling constrain on the parameter itself, e.g:
@ -156,6 +153,8 @@ class Param(ObservableArray, Constrainable):
@property
def _parameters_(self):
return []
def _collect_gradient(self, target):
target[:] = self.gradient
#===========================================================================
# Fixing Parameters:
#===========================================================================

13
GPy/core/parameterization/parameterized.py
View file

@ -75,7 +75,6 @@ class Parameterized(Constrainable, Pickleable, Observable):
if not self._has_fixes():
self._fixes_ = None
self._connect_parameters()
self.gradient_mapping = {}
self._added_names_ = set()
del self._in_init_
@ -118,12 +117,10 @@ class Parameterized(Constrainable, Pickleable, Observable):
def _has_fixes(self):
return hasattr(self, "_fixes_") and self._fixes_ is not None
def add_parameter(self, param, gradient=None, index=None):
def add_parameter(self, param, index=None):
"""
:param parameters: the parameters to add
:type parameters: list of or one :py:class:`GPy.core.param.Param`
:param [gradients]: gradients for each param,
one gradient per param
:param [index]: index of where to put parameters
@ -167,8 +164,6 @@ class Parameterized(Constrainable, Pickleable, Observable):
self._fixes_ = np.ones(self.size+param.size, dtype=bool)
self._fixes_[ins:ins+param.size] = fixes_param
self.size += param.size
if gradient:
self.gradient_mapping[param] = gradient
self._connect_parameters()
# make sure the constraints are pulled over:
if hasattr(param, "_constraints_") and param._constraints_ is not None:
@ -206,7 +201,7 @@ class Parameterized(Constrainable, Pickleable, Observable):
add self as a listener to the param, such that
updates get passed through. See :py:function:``GPy.core.param.Observable.add_observer``
"""
# will be called as soon as paramters have changed
# will be called as soon as parameters have changed
pass
def _connect_parameters(self):
@ -282,13 +277,11 @@ class Parameterized(Constrainable, Pickleable, Observable):
self._constraints_,
self._parameters_,
self._name,
#self.gradient_mapping,
self._added_names_,
]
def setstate(self, state):
self._added_names_ = state.pop()
#self.gradient_mapping = state.pop()
self._name = state.pop()
self._parameters_ = state.pop()
self._connect_parameters()
@ -547,6 +540,8 @@ class Parameterized(Constrainable, Pickleable, Observable):
return [adjust_name_for_printing(self.name) + "." + xi for x in self._parameters_ for xi in x._parameter_names(add_name=True)]
return [xi for x in self._parameters_ for xi in x._parameter_names(add_name=True)]
parameter_names = property(_parameter_names, doc="Names for all parameters handled by this parameterization object -- will add hirarchy name entries for printing")
def _collect_gradient(self, target):
[p._collect_gradient(target[s]) for p, s in itertools.izip(self._parameters_, self._param_slices_)]
@property
def flattened_parameters(self):
return [xi for x in self._parameters_ for xi in x.flattened_parameters]

1
GPy/kern/kern.py
View file

@ -346,6 +346,7 @@ class kern(Parameterized):
return target
def update_gradients_full(self, dL_dK, X):
[p.update_gradients_full(dL_dK, X) for p in self._parameters_]
pass
def update_gradients_sparse(self, dL_dKmm, dL_dKnm, dL_dKdiag, X, Z):
pass

12
GPy/kern/parts/rbf.py
View file

@ -80,9 +80,6 @@ class RBF(Kernpart):
self._X, self._X2 = np.empty(shape=(2, 1))
self._Z, self._mu, self._S = np.empty(shape=(3, 1)) # cached versions of Z,mu,S
def K(self, X, X2, target):
self._K_computations(X, X2)
target += self.variance * self._K_dvar
@ -101,11 +98,8 @@ class RBF(Kernpart):
self._psi_computations(Z, mu, S)
target += self._psi2
def update_gradients_full(self, dL_dK, X):
self._K_computations(X, X2)
self._K_computations(X, None)
self.variance.gradient = np.sum(self._K_dvar * dL_dK)
if self.ARD:
self.lengthscale.gradient = self._dL_dlengthscales_via_K(dL_dK, X, None)
@ -123,7 +117,7 @@ class RBF(Kernpart):
self.lengthscales.gradient = self._dL_dlengthscales_via_K(dL_dKnm, X, Z)
else:
self.lengthscale.gradient = (self.variance / self.lengthscale) * np.sum(self._K_dvar * self._K_dist2 * dL_dK)
self.lengthscale.gradient = (self.variance / self.lengthscale) * np.sum(self._K_dvar * self._K_dist2 * dL_dKmm)
#from Kmm
self._K_computations(Z, None)
@ -131,7 +125,7 @@ class RBF(Kernpart):
if self.ARD:
self.lengthscales.gradient += self._dL_dlengthscales_via_K(dL_dKmm, Z, None)
else:
self.lengthscale.gradient += (self.variance / self.lengthscale) * np.sum(self._K_dvar * self._K_dist2 * dL_dK)
self.lengthscale.gradient += (self.variance / self.lengthscale) * np.sum(self._K_dvar * self._K_dist2 * dL_dKmm)
def update_gradients_variational(self, dL_dKmm, dL_dpsi0, dL_dpsi1, dL_dpsi2, mu, S, Z):
self._psi_computations(Z, mu, S)