GPy/GPy/kern/_src/symbolic.py
2014-04-03 09:10:50 +02:00

432 lines
20 KiB
Python

# Check Matthew Rocklin's blog post.
try:
import sympy as sp
sympy_available=True
from sympy.utilities.lambdify import lambdify
except ImportError:
sympy_available=False
import numpy as np
from kern import Kern
from ...core.parameterization import Param
from ...core.parameterization.transformations import Logexp
class Symbolic(Kern):
"""
A kernel object, where all the hard work is done by sympy.
:param k: the covariance function
:type k: a positive definite sympy function of x_0, z_0, x_1, z_1, x_2, z_2...
To construct a new sympy kernel, you'll need to define:
- a kernel function using a sympy object. Ensure that the kernel is of the form k(x,z).
- that's it! we'll extract the variables from the function k.
Note:
- to handle multiple inputs, call them x_1, z_1, etc
- to handle multpile correlated outputs, you'll need to add parameters with an index, such as lengthscale_i and lengthscale_j.
"""
def __init__(self, input_dim, k=None, output_dim=1, name='symbolic', param=None, active_dims=None, operators=None):
if k is None:
raise ValueError, "You must provide an argument for the covariance function."
super(Sympykern, self).__init__(input_dim, active_dims, name)
self._sp_k = k
# pull the variable names out of the symbolic covariance function.
sp_vars = [e for e in k.atoms() if e.is_Symbol]
self._sp_x= sorted([e for e in sp_vars if e.name[0:2]=='x_'],key=lambda x:int(x.name[2:]))
self._sp_z= sorted([e for e in sp_vars if e.name[0:2]=='z_'],key=lambda z:int(z.name[2:]))
# Check that variable names make sense.
assert all([x.name=='x_%i'%i for i,x in enumerate(self._sp_x)])
assert all([z.name=='z_%i'%i for i,z in enumerate(self._sp_z)])
assert len(self._sp_x)==len(self._sp_z)
x_dim=len(self._sp_x)
self._sp_kdiag = k
for x, z in zip(self._sp_x, self._sp_z):
self._sp_kdiag = self._sp_kdiag.subs(z, x)
# If it is a multi-output covariance, add an input for indexing the outputs.
self._real_input_dim = x_dim
# Check input dim is number of xs + 1 if output_dim is >1
assert self.input_dim == x_dim + int(output_dim > 1)
self.output_dim = output_dim
# extract parameter names from the covariance
thetas = sorted([e for e in sp_vars if not (e.name[0:2]=='x_' or e.name[0:2]=='z_')],key=lambda e:e.name)
# Look for parameters with index (subscripts), they are associated with different outputs.
if self.output_dim>1:
self._sp_theta_i = sorted([e for e in thetas if (e.name[-2:]=='_i')], key=lambda e:e.name)
self._sp_theta_j = sorted([e for e in thetas if (e.name[-2:]=='_j')], key=lambda e:e.name)
# Make sure parameter appears with both indices!
assert len(self._sp_theta_i)==len(self._sp_theta_j)
assert all([theta_i.name[:-2]==theta_j.name[:-2] for theta_i, theta_j in zip(self._sp_theta_i, self._sp_theta_j)])
# Extract names of shared parameters (those without a subscript)
self._sp_theta = [theta for theta in thetas if theta not in self._sp_theta_i and theta not in self._sp_theta_j]
self.num_split_params = len(self._sp_theta_i)
self._split_theta_names = ["%s"%theta.name[:-2] for theta in self._sp_theta_i]
# Add split parameters to the model.
for theta in self._split_theta_names:
# TODO: what if user has passed a parameter vector, how should that be stored and interpreted?
setattr(self, theta, Param(theta, np.ones(self.output_dim), None))
self.add_parameter(getattr(self, theta))
self.num_shared_params = len(self._sp_theta)
for theta_i, theta_j in zip(self._sp_theta_i, self._sp_theta_j):
self._sp_kdiag = self._sp_kdiag.subs(theta_j, theta_i)
else:
self.num_split_params = 0
self._split_theta_names = []
self._sp_theta = thetas
self.num_shared_params = len(self._sp_theta)
# Add parameters to the model.
for theta in self._sp_theta:
val = 1.0
# TODO: what if user has passed a parameter vector, how should that be stored and interpreted? This is the old way before params class.
if param is not None:
if param.has_key(theta.name):
val = param[theta.name]
setattr(self, theta.name, Param(theta.name, val, None))
self.add_parameters(getattr(self, theta.name))
# Differentiate with respect to parameters.
derivative_arguments = self._sp_x + self._sp_theta
if self.output_dim > 1:
derivative_arguments += self._sp_theta_i
self.derivatives = {theta.name : sp.diff(self._sp_k,theta).simplify() for theta in derivative_arguments}
self.diag_derivatives = {theta.name : sp.diff(self._sp_kdiag,theta).simplify() for theta in derivative_arguments}
# This gives the parameters for the arg list.
self.arg_list = self._sp_x + self._sp_z + self._sp_theta
self.diag_arg_list = self._sp_x + self._sp_theta
if self.output_dim > 1:
self.arg_list += self._sp_theta_i + self._sp_theta_j
self.diag_arg_list += self._sp_theta_i
# Check if there are additional linear operators on the covariance.
self._sp_operators = operators
# TODO: Deal with linear operators
#if self._sp_operators:
# for operator in self._sp_operators:
# psi_stats aren't yet implemented.
if False:
self.compute_psi_stats()
# generate the code for the covariance functions
self._gen_code()
def __add__(self,other):
return spkern(self._sp_k+other._sp_k)
def _gen_code(self):
#fn_theano = theano_function([self.arg_lists], [self._sp_k + self.derivatives], dims={x: 1}, dtypes={x_0: 'float64', z_0: 'float64'})
self._K_function = lambdify(self.arg_list, self._sp_k, 'numpy')
for key in self.derivatives.keys():
setattr(self, '_K_diff_' + key, lambdify(self.arg_list, self.derivatives[key], 'numpy'))
self._Kdiag_function = lambdify(self.diag_arg_list, self._sp_kdiag, 'numpy')
for key in self.derivatives.keys():
setattr(self, '_Kdiag_diff_' + key, lambdify(self.diag_arg_list, self.diag_derivatives[key], 'numpy'))
def K(self,X,X2=None):
self._K_computations(X, X2)
return self._K_function(**self._arguments)
def Kdiag(self,X):
self._K_computations(X)
return self._Kdiag_function(**self._diag_arguments)
def _param_grad_helper(self,partial,X,Z,target):
pass
def gradients_X(self, dL_dK, X, X2=None):
#if self._X is None or X.base is not self._X.base or X2 is not None:
self._K_computations(X, X2)
gradients_X = np.zeros((X.shape[0], X.shape[1]))
for i, x in enumerate(self._sp_x):
gf = getattr(self, '_K_diff_' + x.name)
gradients_X[:, i] = (gf(**self._arguments)*dL_dK).sum(1)
if X2 is None:
gradients_X *= 2
return gradients_X
def gradients_X_diag(self, dL_dK, X):
self._K_computations(X)
dX = np.zeros_like(X)
for i, x in enumerate(self._sp_x):
gf = getattr(self, '_Kdiag_diff_' + x.name)
dX[:, i] = gf(**self._diag_arguments)*dL_dK
return dX
def update_gradients_full(self, dL_dK, X, X2=None):
# Need to extract parameters to local variables first
self._K_computations(X, X2)
for theta in self._sp_theta:
parameter = getattr(self, theta.name)
gf = getattr(self, '_K_diff_' + theta.name)
gradient = (gf(**self._arguments)*dL_dK).sum()
if X2 is not None:
gradient += (gf(**self._reverse_arguments)*dL_dK).sum()
setattr(parameter, 'gradient', gradient)
if self.output_dim>1:
for theta in self._sp_theta_i:
parameter = getattr(self, theta.name[:-2])
gf = getattr(self, '_K_diff_' + theta.name)
A = gf(**self._arguments)*dL_dK
gradient = np.asarray([A[np.where(self._output_ind==i)].T.sum()
for i in np.arange(self.output_dim)])
if X2 is None:
gradient *= 2
else:
A = gf(**self._reverse_arguments)*dL_dK.T
gradient += np.asarray([A[np.where(self._output_ind2==i)].T.sum()
for i in np.arange(self.output_dim)])
setattr(parameter, 'gradient', gradient)
def update_gradients_diag(self, dL_dKdiag, X):
self._K_computations(X)
for theta in self._sp_theta:
parameter = getattr(self, theta.name)
gf = getattr(self, '_Kdiag_diff_' + theta.name)
setattr(parameter, 'gradient', (gf(**self._diag_arguments)*dL_dKdiag).sum())
if self.output_dim>1:
for theta in self._sp_theta_i:
parameter = getattr(self, theta.name[:-2])
gf = getattr(self, '_Kdiag_diff_' + theta.name)
a = gf(**self._diag_arguments)*dL_dKdiag
setattr(parameter, 'gradient',
np.asarray([a[np.where(self._output_ind==i)].sum()
for i in np.arange(self.output_dim)]))
def _K_computations(self, X, X2=None):
"""Set up argument lists for the derivatives."""
# Could check if this needs doing or not, there could
# definitely be some computational savings by checking for
# parameter updates here.
self._arguments = {}
self._diag_arguments = {}
for i, x in enumerate(self._sp_x):
self._arguments[x.name] = X[:, i][:, None]
self._diag_arguments[x.name] = X[:, i][:, None]
if self.output_dim > 1:
self._output_ind = np.asarray(X[:, -1], dtype='int')
for i, theta in enumerate(self._sp_theta_i):
self._arguments[theta.name] = np.asarray(getattr(self, theta.name[:-2])[self._output_ind])[:, None]
self._diag_arguments[theta.name] = self._arguments[theta.name]
for theta in self._sp_theta:
self._arguments[theta.name] = np.asarray(getattr(self, theta.name))
self._diag_arguments[theta.name] = self._arguments[theta.name]
if X2 is not None:
for i, z in enumerate(self._sp_z):
self._arguments[z.name] = X2[:, i][None, :]
if self.output_dim > 1:
self._output_ind2 = np.asarray(X2[:, -1], dtype='int')
for i, theta in enumerate(self._sp_theta_j):
self._arguments[theta.name] = np.asarray(getattr(self, theta.name[:-2])[self._output_ind2])[None, :]
else:
for z in self._sp_z:
self._arguments[z.name] = self._arguments['x_'+z.name[2:]].T
if self.output_dim > 1:
self._output_ind2 = self._output_ind
for theta in self._sp_theta_j:
self._arguments[theta.name] = self._arguments[theta.name[:-2] + '_i'].T
if X2 is not None:
# These arguments are needed in gradients when X2 is not equal to X.
self._reverse_arguments = self._arguments
for x, z in zip(self._sp_x, self._sp_z):
self._reverse_arguments[x.name] = self._arguments[z.name].T
self._reverse_arguments[z.name] = self._arguments[x.name].T
if self.output_dim > 1:
for theta_i, theta_j in zip(self._sp_theta_i, self._sp_theta_j):
self._reverse_arguments[theta_i.name] = self._arguments[theta_j.name].T
self._reverse_arguments[theta_j.name] = self._arguments[theta_i.name].T
if False:
class Symcombine(CombinationKernel):
"""
Combine list of given sympy covariances together with the provided operations.
"""
def __init__(self, subkerns, operations, name='sympy_combine'):
super(Symcombine, self).__init__(subkerns, name)
for subkern, operation in zip(subkerns, operations):
self._sp_k += self._k_double_operate(subkern._sp_k, operation)
#def _double_operate(self, k, operation):
@Cache_this(limit=2, force_kwargs=['which_parts'])
def K(self, X, X2=None, which_parts=None):
"""
Combine covariances with a linear operator.
"""
assert X.shape[1] == self.input_dim
if which_parts is None:
which_parts = self.parts
elif not isinstance(which_parts, (list, tuple)):
# if only one part is given
which_parts = [which_parts]
return reduce(np.add, (p.K(X, X2) for p in which_parts))
@Cache_this(limit=2, force_kwargs=['which_parts'])
def Kdiag(self, X, which_parts=None):
assert X.shape[1] == self.input_dim
if which_parts is None:
which_parts = self.parts
elif not isinstance(which_parts, (list, tuple)):
# if only one part is given
which_parts = [which_parts]
return reduce(np.add, (p.Kdiag(X) for p in which_parts))
def update_gradients_full(self, dL_dK, X, X2=None):
[p.update_gradients_full(dL_dK, X, X2) for p in self.parts]
def update_gradients_diag(self, dL_dK, X):
[p.update_gradients_diag(dL_dK, X) for p in self.parts]
def gradients_X(self, dL_dK, X, X2=None):
"""Compute the gradient of the objective function with respect to X.
:param dL_dK: An array of gradients of the objective function with respect to the covariance function.
:type dL_dK: np.ndarray (num_samples x num_inducing)
:param X: Observed data inputs
:type X: np.ndarray (num_samples x input_dim)
:param X2: Observed data inputs (optional, defaults to X)
:type X2: np.ndarray (num_inducing x input_dim)"""
target = np.zeros(X.shape)
[target.__iadd__(p.gradients_X(dL_dK, X, X2)) for p in self.parts]
return target
def gradients_X_diag(self, dL_dKdiag, X):
target = np.zeros(X.shape)
[target.__iadd__(p.gradients_X_diag(dL_dKdiag, X)) for p in self.parts]
return target
def psi0(self, Z, variational_posterior):
return reduce(np.add, (p.psi0(Z, variational_posterior) for p in self.parts))
def psi1(self, Z, variational_posterior):
return reduce(np.add, (p.psi1(Z, variational_posterior) for p in self.parts))
def psi2(self, Z, variational_posterior):
psi2 = reduce(np.add, (p.psi2(Z, variational_posterior) for p in self.parts))
#return psi2
# compute the "cross" terms
from static import White, Bias
from rbf import RBF
#from rbf_inv import RBFInv
from linear import Linear
#ffrom fixed import Fixed
for p1, p2 in itertools.combinations(self.parts, 2):
# i1, i2 = p1.active_dims, p2.active_dims
# white doesn;t combine with anything
if isinstance(p1, White) or isinstance(p2, White):
pass
# rbf X bias
#elif isinstance(p1, (Bias, Fixed)) and isinstance(p2, (RBF, RBFInv)):
elif isinstance(p1, Bias) and isinstance(p2, (RBF, Linear)):
tmp = p2.psi1(Z, variational_posterior)
psi2 += p1.variance * (tmp[:, :, None] + tmp[:, None, :])
#elif isinstance(p2, (Bias, Fixed)) and isinstance(p1, (RBF, RBFInv)):
elif isinstance(p2, Bias) and isinstance(p1, (RBF, Linear)):
tmp = p1.psi1(Z, variational_posterior)
psi2 += p2.variance * (tmp[:, :, None] + tmp[:, None, :])
elif isinstance(p2, (RBF, Linear)) and isinstance(p1, (RBF, Linear)):
assert np.intersect1d(p1.active_dims, p2.active_dims).size == 0, "only non overlapping kernel dimensions allowed so far"
tmp1 = p1.psi1(Z, variational_posterior)
tmp2 = p2.psi1(Z, variational_posterior)
psi2 += (tmp1[:, :, None] * tmp2[:, None, :]) + (tmp2[:, :, None] * tmp1[:, None, :])
else:
raise NotImplementedError, "psi2 cannot be computed for this kernel"
return psi2
def update_gradients_expectations(self, dL_dpsi0, dL_dpsi1, dL_dpsi2, Z, variational_posterior):
from static import White, Bias
for p1 in self.parts:
#compute the effective dL_dpsi1. Extra terms appear becaue of the cross terms in psi2!
eff_dL_dpsi1 = dL_dpsi1.copy()
for p2 in self.parts:
if p2 is p1:
continue
if isinstance(p2, White):
continue
elif isinstance(p2, Bias):
eff_dL_dpsi1 += dL_dpsi2.sum(1) * p2.variance * 2.
else:# np.setdiff1d(p1.active_dims, ar2, assume_unique): # TODO: Careful, not correct for overlapping active_dims
eff_dL_dpsi1 += dL_dpsi2.sum(1) * p2.psi1(Z, variational_posterior) * 2.
p1.update_gradients_expectations(dL_dpsi0, eff_dL_dpsi1, dL_dpsi2, Z, variational_posterior)
def gradients_Z_expectations(self, dL_dpsi1, dL_dpsi2, Z, variational_posterior):
from static import White, Bias
target = np.zeros(Z.shape)
for p1 in self.parts:
#compute the effective dL_dpsi1. extra terms appear becaue of the cross terms in psi2!
eff_dL_dpsi1 = dL_dpsi1.copy()
for p2 in self.parts:
if p2 is p1:
continue
if isinstance(p2, White):
continue
elif isinstance(p2, Bias):
eff_dL_dpsi1 += dL_dpsi2.sum(1) * p2.variance * 2.
else:
eff_dL_dpsi1 += dL_dpsi2.sum(1) * p2.psi1(Z, variational_posterior) * 2.
target += p1.gradients_Z_expectations(eff_dL_dpsi1, dL_dpsi2, Z, variational_posterior)
return target
def gradients_qX_expectations(self, dL_dpsi0, dL_dpsi1, dL_dpsi2, Z, variational_posterior):
from static import White, Bias
target_mu = np.zeros(variational_posterior.shape)
target_S = np.zeros(variational_posterior.shape)
for p1 in self._parameters_:
#compute the effective dL_dpsi1. extra terms appear becaue of the cross terms in psi2!
eff_dL_dpsi1 = dL_dpsi1.copy()
for p2 in self._parameters_:
if p2 is p1:
continue
if isinstance(p2, White):
continue
elif isinstance(p2, Bias):
eff_dL_dpsi1 += dL_dpsi2.sum(1) * p2.variance * 2.
else:
eff_dL_dpsi1 += dL_dpsi2.sum(1) * p2.psi1(Z, variational_posterior) * 2.
a, b = p1.gradients_qX_expectations(dL_dpsi0, eff_dL_dpsi1, dL_dpsi2, Z, variational_posterior)
target_mu += a
target_S += b
return target_mu, target_S
def _getstate(self):
"""
Get the current state of the class,
here just all the indices, rest can get recomputed
"""
return super(Add, self)._getstate()
def _setstate(self, state):
super(Add, self)._setstate(state)
def add(self, other, name='sum'):
if isinstance(other, Add):
other_params = other._parameters_.copy()
for p in other_params:
other.remove_parameter(p)
self.add_parameters(*other_params)
else: self.add_parameter(other)
return self