# Copyright 2016 James Hensman, alexggmatthews # # 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 . import settings from .decors import name_scope @name_scope() 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) We assume N multiple independent distributions, given by the columns of q_mu and the 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, [M, M] or [L, M, M] K_cholesky is the cholesky of the covariance of p, [M, M] or [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. 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]. """ 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`.") white = (K is None) and (K_cholesky is None) diag = q_sqrt.get_shape().ndims == 2 M, B = tf.shape(q_mu)[0], tf.shape(q_mu)[1] if white: alpha = q_mu # [M, B] else: if K is not None: Lp = tf.cholesky(K) # [B, M, M] or [M, M] elif K_cholesky is not None: Lp = K_cholesky # [B, M, M] or [M, M] batched = Lp.get_shape().ndims == 3 q_mu = tf.transpose(q_mu)[:, :, None] if batched else q_mu # [B, M, 1] or [M, B] alpha = tf.matrix_triangular_solve(Lp, q_mu, lower=True) # [B, M, 1] or [M, B] if diag: Lq = Lq_diag = q_sqrt Lq_full = tf.matrix_diag(tf.transpose(q_sqrt)) # [B, M, M] else: Lq = Lq_full = tf.matrix_band_part(q_sqrt, -1, 0) # force lower triangle # [B, M, M] Lq_diag = tf.matrix_diag_part(Lq) # [M, B] # Mahalanobis term: μqᵀ Σp⁻¹ μq mahalanobis = tf.reduce_sum(tf.square(alpha)) # Constant term: - B * M constant = - tf.cast(tf.size(q_mu, out_type=tf.int64), dtype=settings.float_type) # Log-determinant of the covariance of q(x): logdet_qcov = tf.reduce_sum(tf.log(tf.square(Lq_diag))) # Trace term: tr(Σp⁻¹ Σq) if white: trace = tf.reduce_sum(tf.square(Lq)) else: if diag and not batched: # K is [M, M] and q_sqrt is [M, B]: fast specialisation LpT = tf.transpose(Lp) # [M, M] Lp_inv = tf.matrix_triangular_solve(Lp, tf.eye(M, dtype=settings.float_type),lower=True) # [M, M] K_inv = tf.matrix_diag_part(tf.matrix_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 (not implemented in tf <= 1.12) Lp_full = Lp if batched else tf.tile(tf.expand_dims(Lp, 0), [B, 1, 1]) LpiLq = tf.matrix_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 white: log_sqdiag_Lp = tf.log(tf.square(tf.matrix_diag_part(Lp))) sum_log_sqdiag_Lp = tf.reduce_sum(log_sqdiag_Lp) # If K is B x M x M, num_latent is no longer implicit, no need to multiply the single kernel logdet scale = 1.0 if batched else tf.cast(B, settings.float_type) twoKL += scale * sum_log_sqdiag_Lp return 0.5 * twoKL