Revision 48270681afc13081094f7f398a1e194c6b07ba9b authored by vdutor on 03 January 2018, 17:44:53 UTC, committed by Mark van der Wilk on 03 January 2018, 17:44:53 UTC
* Outline of new expectations code. * Quadrature code now uses TensorFlow shape inference. * General expectations work. * Expectations RBF kern, not tested * Add Identity mean function * General unittests for Expectations * Add multipledispatch package to travis * Update tests_expectations * Expectations of mean functions * Mean function uncertain conditional * Uncertain conditional with mean_function. Tested. * Support for Add and Prod kernels and quadrature fallback decorator * Refactor expectations unittests * Psi stats Linear kernel * Split expectations in different files * Expectation Linear kernel and Linear mean function * Remove None's from expectations api * Removed old ekernels framework * Add multipledispatch to setup file * Work on PR feedback, not finished * Addressed PR feedback * Support for pairwise xKxz * Enable expectations unittests * Renamed `TimeseriesGaussian` to `MarkovGaussian` and added tests. * Rename some variable, plus note for later test of <x Kxz>_q. * Update conditionals.py Add comment * Change order of inputs to (feat, kern) * Stef/expectations (#601) * adding gaussmarkov quad * don't override the markvogaussian in the quadrature * can't test * adding external test * quadrature code done and works for MarkovGauss * MarkovGaussian with quad implemented. All tests pass * Shape comments. * Removed superfluous autoflow functions for kernel expectations * Update kernels.py * Update quadrature.py
1 parent 2182bf0
probability_distributions.py
# Copyright 2017 the GPflow authors.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.from __future__ import print_function
# Eventually, it would be nice to not have to have our own classes for
# proability distributions. The TensorFlow "distributions" framework would
# be a good replacement.
class ProbabilityDistribution:
"""
This is the base class for a probability distributions,
over which we take the expectations in the expectations framework.
"""
pass
class Gaussian(ProbabilityDistribution):
def __init__(self, mu, cov):
self.mu = mu # N x D
self.cov = cov # N x D x D
class DiagonalGaussian(ProbabilityDistribution):
def __init__(self, mu, var):
self.mu = mu # N x D
self.var = var # N x D
class MarkovGaussian(ProbabilityDistribution):
"""
Gaussian distribution with Markov structure.
Only covariances and covariances between t and t+1 need to be
parameterised. We use the solution proposed by Carl Rasmussen, i.e. to
represent
Var[x_t] = cov[x_t, :, :] * cov[x_t, :, :].T
Cov[x_t, x_{t+1}] = cov[t, :, :] * cov[t+1, :, :]
"""
def __init__(self, mu, cov):
self.mu = mu # N+1 x D
self.cov = cov # 2 x (N+1) x D x D
Computing file changes ...