https://github.com/GPflow/GPflow
Tip revision: b819db324fb3c64cab4db52c8f618ab8ff0f5778 authored by st-- on 14 September 2020, 17:03:08 UTC
Merge pull request #1565 from GPflow/develop
Merge pull request #1565 from GPflow/develop
Tip revision: b819db3
kullback_leiblers.py
# Copyright 2016-2020 The GPflow Contributors. All Rights Reserved.
#
# 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.
# -*- coding: utf-8 -*-
import tensorflow as tf
from .config import default_float, default_jitter
from .covariances.kuus import Kuu
from .inducing_variables import InducingVariables
from .kernels import Kernel
from .utilities import Dispatcher, to_default_float
prior_kl = Dispatcher("prior_kl")
@prior_kl.register(InducingVariables, Kernel, object, object)
def _(inducing_variable, kernel, q_mu, q_sqrt, whiten=False):
if whiten:
return gauss_kl(q_mu, q_sqrt, None)
else:
K = Kuu(inducing_variable, kernel, jitter=default_jitter()) # [P, M, M] or [M, M]
return gauss_kl(q_mu, q_sqrt, K)
def gauss_kl(q_mu, q_sqrt, K=None, *, K_cholesky=None):
"""
Compute the KL divergence KL[q || p] between
q(x) = N(q_mu, q_sqrt^2)
and
p(x) = N(0, K) if K is not None
p(x) = N(0, I) if K is None
We assume L multiple independent distributions, given by the columns of
q_mu and the first or last dimension of q_sqrt. Returns the *sum* of the
divergences.
q_mu is a matrix ([M, L]), each column contains a mean.
q_sqrt can be a 3D tensor ([L, M, M]), each matrix within is a lower
triangular square-root matrix of the covariance of q.
q_sqrt can be a matrix ([M, L]), each column represents the diagonal of a
square-root matrix of the covariance of q.
K is the covariance of p (positive-definite matrix). The K matrix can be
passed either directly as `K`, or as its Cholesky factor, `K_cholesky`. In
either case, it can be a single matrix [M, M], in which case the sum of the
L KL divergences is computed by broadcasting, or L different covariances
[L, M, M].
Note: if no K matrix is given (both `K` and `K_cholesky` are None),
`gauss_kl` computes the KL divergence from p(x) = N(0, I) instead.
"""
if (K is not None) and (K_cholesky is not None):
raise ValueError(
"Ambiguous arguments: gauss_kl() must only be passed one of `K` or `K_cholesky`."
)
is_white = (K is None) and (K_cholesky is None)
is_diag = len(q_sqrt.shape) == 2
shape_constraints = [
(q_mu, ["M", "L"]),
(q_sqrt, (["M", "L"] if is_diag else ["L", "M", "M"])),
]
if not is_white:
if K is not None:
shape_constraints.append((K, (["L", "M", "M"] if len(K.shape) == 3 else ["M", "M"])))
else:
shape_constraints.append(
(K_cholesky, (["L", "M", "M"] if len(K_cholesky.shape) == 3 else ["M", "M"]))
)
tf.debugging.assert_shapes(shape_constraints, message="gauss_kl() arguments")
M, L = tf.shape(q_mu)[0], tf.shape(q_mu)[1]
if is_white:
alpha = q_mu # [M, L]
else:
if K is not None:
Lp = tf.linalg.cholesky(K) # [L, M, M] or [M, M]
elif K_cholesky is not None:
Lp = K_cholesky # [L, M, M] or [M, M]
is_batched = len(Lp.shape) == 3
q_mu = tf.transpose(q_mu)[:, :, None] if is_batched else q_mu # [L, M, 1] or [M, L]
alpha = tf.linalg.triangular_solve(Lp, q_mu, lower=True) # [L, M, 1] or [M, L]
if is_diag:
Lq = Lq_diag = q_sqrt
Lq_full = tf.linalg.diag(tf.transpose(q_sqrt)) # [L, M, M]
else:
Lq = Lq_full = tf.linalg.band_part(q_sqrt, -1, 0) # force lower triangle # [L, M, M]
Lq_diag = tf.linalg.diag_part(Lq) # [M, L]
# Mahalanobis term: μqᵀ Σp⁻¹ μq
mahalanobis = tf.reduce_sum(tf.square(alpha))
# Constant term: - L * M
constant = -to_default_float(tf.size(q_mu, out_type=tf.int64))
# Log-determinant of the covariance of q(x):
logdet_qcov = tf.reduce_sum(tf.math.log(tf.square(Lq_diag)))
# Trace term: tr(Σp⁻¹ Σq)
if is_white:
trace = tf.reduce_sum(tf.square(Lq))
else:
if is_diag and not is_batched:
# K is [M, M] and q_sqrt is [M, L]: fast specialisation
LpT = tf.transpose(Lp) # [M, M]
Lp_inv = tf.linalg.triangular_solve(
Lp, tf.eye(M, dtype=default_float()), lower=True
) # [M, M]
K_inv = tf.linalg.diag_part(tf.linalg.triangular_solve(LpT, Lp_inv, lower=False))[
:, None
] # [M, M] -> [M, 1]
trace = tf.reduce_sum(K_inv * tf.square(q_sqrt))
else:
# TODO: broadcast instead of tile when tf allows -- tf2.1 segfaults
# (https://github.com/tensorflow/tensorflow/issues/37584).
# See # https://github.com/GPflow/GPflow/issues/1321
Lp_full = Lp if is_batched else tf.tile(tf.expand_dims(Lp, 0), [L, 1, 1])
LpiLq = tf.linalg.triangular_solve(Lp_full, Lq_full, lower=True)
trace = tf.reduce_sum(tf.square(LpiLq))
twoKL = mahalanobis + constant - logdet_qcov + trace
# Log-determinant of the covariance of p(x):
if not is_white:
log_sqdiag_Lp = tf.math.log(tf.square(tf.linalg.diag_part(Lp)))
sum_log_sqdiag_Lp = tf.reduce_sum(log_sqdiag_Lp)
# If K is [L, M, M], num_latent_gps is no longer implicit, no need to multiply the single kernel logdet
scale = 1.0 if is_batched else to_default_float(L)
twoKL += scale * sum_log_sqdiag_Lp
tf.debugging.assert_shapes([(twoKL, ())], message="gauss_kl() return value") # returns scalar
return 0.5 * twoKL