https://github.com/GPflow/GPflow
Revision bb08f22e337d1487b8d9ab9944d8b9f7fff853ff authored by Vincent Dutordoir on 18 June 2018, 17:04:06 UTC, committed by Artem Artemev on 18 June 2018, 17:04:06 UTC
* Introduction of MultiOutputFeatures (Mof) and MultiOutputKernels (Mok).
These are used to specify a particular setup of multi-output correlation.

* Multiple-dispatch for conditional. This allows GPflow to select the most efficient conditional code depending on your choice of Mof and Mok.

* Multiple-dispatch for Kuu and Kuf. Previously Kuu(.) and Kuf(.) were member functions of the feature class. This became cumbersome as the calculation of Kuu and Kuf also depends on the kernel used. In line with conditional we now also use multiple-dispatch to calculate Kuu and Kuf for a particular combination of Mok and Mof.

* The actual maths to efficiently calculate the output-correlated conditional (credits to @markvdw )

* sample_conditional function that makes sure that the most efficient code is used to get a sample from the conditional distribution.

* Minor: we updated a couple of models to use the new multi-output conditional.
1 parent 6baeb43
Raw File
Tip revision: bb08f22e337d1487b8d9ab9944d8b9f7fff853ff authored by Vincent Dutordoir on 18 June 2018, 17:04:06 UTC
Multi-output conditionals (#724)
Tip revision: bb08f22
test_kldiv.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.

# -*- coding: utf-8 -*-

import numpy as np
import tensorflow as tf

import gpflow
from gpflow.kullback_leiblers import gauss_kl
from numpy.testing import assert_almost_equal
import pytest
from gpflow import settings
from gpflow.test_util import session_tf

def squareT(A):
    """
    Returns (A Aᵀ)
    """
    return A.dot(A.T)

def make_sqrt_data(rng, N, M):
    return np.array([np.tril(rng.randn(M, M)) for _ in range(N)]) # N x M x M

def make_K_batch_data(rng, N, M):
    K_np = rng.randn(N, M, M)
    beye = np.array([np.eye(M) for _ in range(N)])
    return .1 * (K_np + np.transpose(K_np, (0, 2, 1))) + beye

class Datum:
    M, N = 5, 4
    rng = np.random.RandomState(0)
    mu_data = rng.randn(M, N)  # M x N
    K_data = squareT(rng.randn(M, M)) + 1e-6 * np.eye(M)  # M x M
    I = np.eye(M) # M x M
    sqrt_data = make_sqrt_data(rng, N, M) # N x M x M
    sqrt_diag_data = rng.randn(M, N) # M x N
    K_batch_data = make_K_batch_data(rng, N, M)

@pytest.fixture
def mu(session_tf):
    return tf.convert_to_tensor(Datum.mu_data)

@pytest.fixture
def sqrt_diag(session_tf):
    return tf.convert_to_tensor(Datum.sqrt_diag_data)

@pytest.fixture
def K(session_tf):
    return tf.convert_to_tensor(Datum.K_data)

@pytest.fixture
def K_batch(session_tf):
    return tf.convert_to_tensor(Datum.K_batch_data)

@pytest.fixture
def sqrt(session_tf):
    return tf.convert_to_tensor(Datum.sqrt_data)

@pytest.fixture()
def I(session_tf):
    return tf.convert_to_tensor(Datum.I)

@pytest.mark.parametrize('white', [True, False])
def test_diags(session_tf, white, mu, sqrt_diag, K):
    """
    The covariance of q(x) can be Cholesky matrices or diagonal matrices.
    Here we make sure the behaviours overlap.
    """
    # the chols are diagonal matrices, with the same entries as the diag representation.
    chol_from_diag = tf.stack([tf.diag(sqrt_diag[:, i]) for i in range(Datum.N)]) # N x M x M
    # run
    kl_diag = gauss_kl(mu, sqrt_diag, K if white else None)
    kl_dense = gauss_kl(mu, chol_from_diag, K if white else None)

    np.testing.assert_allclose(kl_diag.eval(), kl_dense.eval())

@pytest.mark.parametrize('diag', [True, False])
def test_whitened(session_tf, diag, mu, sqrt_diag, I):
    """
    Check that K=Identity and K=None give same answer
    """
    chol_from_diag = tf.stack([tf.diag(sqrt_diag[:, i]) for i in range(Datum.N)]) # N x M x M
    s = sqrt_diag if diag else chol_from_diag

    kl_white = gauss_kl(mu, s)
    kl_nonwhite = gauss_kl(mu, s, I)

    np.testing.assert_allclose(kl_white.eval(), kl_nonwhite.eval())

@pytest.mark.parametrize('shared_k', [True, False])
@pytest.mark.parametrize('diag', [True, False])
def test_sumkl_equals_batchkl(session_tf, shared_k, diag, mu,
                              sqrt, sqrt_diag, K_batch, K):
    """
    gauss_kl implicitely performs a sum of KL divergences
    This test checks that doing the sum outside of the function is equivalent
    For q(X)=prod q(x_l) and p(X)=prod p(x_l), check that sum KL(q(x_l)||p(x_l)) = KL(q(X)||p(X))
    Here, q(X) has covariance L x M x M
    p(X) has covariance L x M x M ( or M x M )
    Here, q(x_i) has covariance 1 x M x M
    p(x_i) has covariance M x M
    """
    s = sqrt_diag if diag else sqrt
    kl_batch = gauss_kl(mu,s,K if shared_k else K_batch)
    kl_sum = []
    for n in range(Datum.N):
        kl_sum.append(gauss_kl(mu[:, n][:,None], # M x 1
            sqrt_diag[:, n][:, None] if diag else sqrt[n, :, :][None, :, :], # 1 x M x M or M x 1
            K if shared_k else K_batch[n, :, :][None,:,:])) # 1 x M x M or M x M
    kl_sum =tf.reduce_sum(kl_sum)
    assert_almost_equal(kl_sum.eval(), kl_batch.eval())

def tf_kl_1d(q_mu, q_sigma, p_var=1.0):
    p_var = tf.ones_like(q_sigma) if p_var is None else p_var
    q_var = tf.square(q_sigma)
    kl = 0.5 * (q_var / p_var + tf.square(q_mu) / p_var - 1 + tf.log(p_var / q_var))
    return tf.reduce_sum(kl)

@pytest.mark.parametrize('white', [True, False])
def test_oned(session_tf, white, mu, sqrt, K_batch):
    """
    Check that the KL divergence matches a 1D by-hand calculation.
    """
    m = 0
    mu1d = mu[m,:][None,:] # 1 x N
    s1d = sqrt[:,m,m][:,None,None] # N x 1 x 1
    K1d = K_batch[:,m,m][:,None,None] # N x 1 x 1

    kl = gauss_kl(mu1d,s1d,K1d if not white else None)
    kl_tf = tf_kl_1d(tf.reshape(mu1d,(-1,)), # N
                   tf.reshape(s1d,(-1,)), # N
                   None if white else tf.reshape(K1d,(-1,))) # N
    np.testing.assert_allclose(kl.eval(), kl_tf.eval())


def test_unknown_size_inputs(session_tf):
    """
    Test for #725 and #734. When the shape of the Gaussian's mean had at least
    one unknown parameter, `gauss_kl` would blow up. This happened because
    `tf.size` can only output types `tf.int32` or `tf.int64`.
    """
    mu_ph = tf.placeholder(settings.float_type, [None, None])
    sqrt_ph = tf.placeholder(settings.float_type, [None, None, None])
    mu = np.ones([1, 4], dtype=settings.float_type)
    sqrt = np.ones([4, 1, 1], dtype=settings.float_type)
    
    feed_dict = {mu_ph: mu, sqrt_ph: sqrt}
    known_shape_tf = gauss_kl(*map(tf.constant, [mu, sqrt]))
    unknown_shape_tf = gauss_kl(mu_ph, sqrt_ph)
    
    known_shape = session_tf.run(known_shape_tf)
    unknown_shape = session_tf.run(unknown_shape_tf, feed_dict=feed_dict)
    
    np.testing.assert_allclose(known_shape, unknown_shape)


if __name__ == "__main__":
    tf.test.main()
back to top