GPy/GPy/likelihoods/symbolic.py

279 lines
12 KiB
Python

# Copyright (c) 2014 GPy Authors
# Licensed under the BSD 3-clause license (see LICENSE.txt)
import sympy as sym
import numpy as np
from likelihood import Likelihood
from ..core.symbolic import Symbolic_core
class Symbolic(Likelihood, Symbolic_core):
"""
Symbolic likelihood.
Likelihood where the form of the likelihood is provided by a sympy expression.
"""
def __init__(self, log_pdf=None, logZ=None, missing_log_pdf=None, gp_link=None, name='symbolic', log_concave=False, parameters=None, func_modules=[]):
if gp_link is None:
gp_link = link_functions.Identity()
if log_pdf is None:
raise ValueError, "You must provide an argument for the log pdf."
Likelihood.__init__(self, gp_link, name=name)
functions = {'log_pdf':log_pdf}
self.cacheable = ['F', 'Y']
self.missing_data = False
if missing_log_pdf:
self.missing_data = True
functions['missing_log_pdf']=missing_log_pdf
self.ep_analytic = False
if logZ:
self.ep_analytic = True
functions['logZ'] = logZ
self.cacheable += ['M', 'V']
Symbolic_core.__init__(self, functions, cacheable=self.cacheable, derivatives = ['F', 'theta'], parameters=parameters, func_modules=func_modules)
# TODO: Is there an easy way to check whether the pdf is log
self.log_concave = log_concave
def _set_derivatives(self, derivatives):
# these are arguments for computing derivatives.
print "Whoop"
Symbolic_core._set_derivatives(self, derivatives)
# add second and third derivatives for Laplace approximation.
derivative_arguments = []
if derivatives is not None:
for derivative in derivatives:
derivative_arguments += self.variables[derivative]
exprs = ['log_pdf']
if self.missing_data:
exprs.append('missing_log_pdf')
for expr in exprs:
self.expressions[expr]['second_derivative'] = {theta.name : self.stabilize(sym.diff(self.expressions[expr]['derivative']['f_0'], theta)) for theta in derivative_arguments}
self.expressions[expr]['third_derivative'] = {theta.name : self.stabilize(sym.diff(self.expressions[expr]['second_derivative']['f_0'], theta)) for theta in derivative_arguments}
if self.ep_analytic:
derivative_arguments = [M]
# add second derivative for EP
exprs = ['logZ']
if self.missing_data:
exprs.append('missing_logZ')
for expr in exprs:
self.expressions[expr]['second_derivative'] = {theta.name : self.stabilize(sym.diff(self.expressions[expr]['derivative'], theta)) for theta in derivative_arguments}
def eval_update_cache(self, Y, **kwargs):
# TODO: place checks for inf/nan in here
# for all provided keywords
Symbolic_core.eval_update_cache(self, Y=Y, **kwargs)
# Y = np.atleast_2d(Y)
# for variable, code in sorted(self.code['parameters_changed'].iteritems()):
# self._set_attribute(variable, eval(code, self.namespace))
# for i, theta in enumerate(self.variables['Y']):
# missing = np.isnan(Y[:, i])
# self._set_attribute('missing_' + str(i), missing)
# self._set_attribute(theta.name, value[missing, i][:, None])
# for variable, value in kwargs.items():
# # update their cached values
# if value is not None:
# if variable == 'F' or variable == 'M' or variable == 'V' or variable == 'Y_metadata':
# for i, theta in enumerate(self.variables[variable]):
# self._set_attribute(theta.name, value[:, i][:, None])
# else:
# self._set_attribute(theta.name, value[:, i])
# for variable, code in sorted(self.code['update_cache'].iteritems()):
# self._set_attribute(variable, eval(code, self.namespace))
def parameters_changed(self):
pass
def update_gradients(self, grads):
"""
Pull out the gradients, be careful as the order must match the order
in which the parameters are added
"""
# The way the Laplace approximation is run requires the
# covariance function to compute the true gradient (because it
# is dependent on the mode). This means we actually compute
# the gradient outside this object. This function would
# normally ask the object to update its gradients internally,
# but here it provides them externally, because they are
# computed in the inference code. TODO: Thought: How does this
# effect EP? Shouldn't this be done by a separate
# Laplace-approximation specific call?
for theta, grad in zip(self.variables['theta'], grads):
parameter = getattr(self, theta.name)
setattr(parameter, 'gradient', grad)
def pdf_link(self, f, y, Y_metadata=None):
"""
Likelihood function given inverse link of f.
:param f: inverse link of latent variables.
:type f: Nx1 array
:param y: data
:type y: Nx1 array
:param Y_metadata: Y_metadata which is not used in student t distribution
:returns: likelihood evaluated for this point
:rtype: float
"""
return np.exp(self.logpdf_link(f, y, Y_metadata=None))
def logpdf_link(self, f, y, Y_metadata=None):
"""
Log Likelihood Function given inverse link of latent variables.
:param f: latent variables (inverse link of f)
:type f: Nx1 array
:param y: data
:type y: Nx1 array
:param Y_metadata: Y_metadata
:returns: likelihood evaluated for this point
:rtype: float
"""
assert np.atleast_1d(f).shape == np.atleast_1d(y).shape
if self.missing_data:
missing_flag = np.isnan(y)
not_missing_flag = np.logical_not(missing_flag)
ll = self.eval_function('missing_log_pdf', F=f[missing_flag]).sum()
ll += self.eval_function('log_pdf', F=f[not_missing_flag], Y=y[not_missing_flag], Y_metadata=Y_metadata[not_missing_flag]).sum()
else:
ll = self.eval_function('log_pdf', F=f, Y=y, Y_metadata=Y_metadata).sum()
return ll
def dlogpdf_dlink(self, f, y, Y_metadata=None):
"""
Gradient of log likelihood with respect to the inverse link function.
:param f: latent variables (inverse link of f)
:type f: Nx1 array
:param y: data
:type y: Nx1 array
:param Y_metadata: Y_metadata
:returns: gradient of likelihood with respect to each point.
:rtype: Nx1 array
"""
assert np.atleast_1d(f).shape == np.atleast_1d(y).shape
self.eval_update_cache(F=f, Y=y, Y_metadata=Y_metadata)
if self.missing_data:
return np.where(np.isnan(y),
eval(self.code['missing_log_pdf']['derivative']['f_0'], self.namespace),
eval(self.code['log_pdf']['derivative']['f_0'], self.namespace))
else:
return np.where(np.isnan(y),
0.,
eval(self.code['log_pdf']['derivative']['f_0'], self.namespace))
def d2logpdf_dlink2(self, f, y, Y_metadata=None):
"""
Hessian of log likelihood given inverse link of latent variables with respect to that inverse link.
i.e. second derivative logpdf at y given inv_link(f_i) and inv_link(f_j) w.r.t inv_link(f_i) and inv_link(f_j).
:param f: inverse link of the latent variables.
:type f: Nx1 array
:param y: data
:type y: Nx1 array
:param Y_metadata: Y_metadata which is not used in student t distribution
:returns: Diagonal of Hessian matrix (second derivative of likelihood evaluated at points f)
:rtype: Nx1 array
.. Note::
Returns diagonal of Hessian, since every where else it is
0, as the likelihood factorizes over cases (the
distribution for y_i depends only on link(f_i) not on
link(f_(j!=i))
"""
assert np.atleast_1d(f).shape == np.atleast_1d(y).shape
self.eval_update_cache(F=f, Y=y, Y_metadata=Y_metadata)
if self.missing_data:
return np.where(np.isnan(y),
eval(self.code['missing_log_pdf']['second_derivative']['f_0'], self.namespace),
eval(self.code['log_pdf']['second_derivative']['f_0'], self.namespace))
else:
return np.where(np.isnan(y),
0.,
eval(self.code['log_pdf']['second_derivative']['f_0'], self.namespace))
def d3logpdf_dlink3(self, f, y, Y_metadata=None):
assert np.atleast_1d(f).shape == np.atleast_1d(y).shape
self.eval_update_cache(F=f, Y=y, Y_metadata=Y_metadata)
if self.missing_data:
return np.where(np.isnan(y),
eval(self.code['missing_log_pdf']['third_derivative']['f_0'], self.namespace),
eval(self.code['log_pdf']['third_derivative']['f_0'], self.namespace))
else:
return np.where(np.isnan(y),
0.,
eval(self.code['log_pdf']['third_derivative']['f_0'], self.namespace))
def dlogpdf_link_dtheta(self, f, y, Y_metadata=None):
assert np.atleast_1d(f).shape == np.atleast_1d(y).shape
self.eval_update_cache(F=f, Y=y, Y_metadata=Y_metadata)
g = np.zeros((np.atleast_1d(y).shape[0], len(self.variables['theta'])))
for i, theta in enumerate(self.variables['theta']):
if self.missing_data:
g[:, i:i+1] = np.where(np.isnan(y),
eval(self.code['missing_log_pdf']['derivative'][theta.name], self.namespace),
eval(self.code['log_pdf']['derivative'][theta.name], self.namespace))
else:
g[:, i:i+1] = np.where(np.isnan(y),
0.,
eval(self.code['log_pdf']['derivative'][theta.name], self.namespace))
return g.sum(0)
def dlogpdf_dlink_dtheta(self, f, y, Y_metadata=None):
assert np.atleast_1d(f).shape == np.atleast_1d(y).shape
self.eval_update_cache(F=f, Y=y, Y_metadata=Y_metadata)
g = np.zeros((np.atleast_1d(y).shape[0], len(self.variables['theta'])))
for i, theta in enumerate(self.variables['theta']):
if self.missing_data:
g[:, i:i+1] = np.where(np.isnan(y),
eval(self.code['missing_log_pdf']['second_derivative'][theta.name], self.namespace),
eval(self.code['log_pdf']['second_derivative'][theta.name], self.namespace))
else:
g[:, i:i+1] = np.where(np.isnan(y),
0.,
eval(self.code['log_pdf']['second_derivative'][theta.name], self.namespace))
return g
def d2logpdf_dlink2_dtheta(self, f, y, Y_metadata=None):
assert np.atleast_1d(f).shape == np.atleast_1d(y).shape
self.eval_update_cache(F=f, Y=y, Y_metadata=Y_metadata)
g = np.zeros((np.atleast_1d(y).shape[0], len(self.variables['theta'])))
for i, theta in enumerate(self.variables['theta']):
if self.missing_data:
g[:, i:i+1] = np.where(np.isnan(y),
eval(self.code['missing_log_pdf']['third_derivative'][theta.name], self.namespace),
eval(self.code['log_pdf']['third_derivative'][theta.name], self.namespace))
else:
g[:, i:i+1] = np.where(np.isnan(y),
0.,
eval(self.code['log_pdf']['third_derivative'][theta.name], self.namespace))
return g
def predictive_mean(self, mu, sigma, Y_metadata=None):
raise NotImplementedError
def predictive_variance(self, mu,variance, predictive_mean=None, Y_metadata=None):
raise NotImplementedError
def conditional_mean(self, gp):
raise NotImplementedError
def conditional_variance(self, gp):
raise NotImplementedError
def samples(self, gp, Y_metadata=None):
raise NotImplementedError