def test_kl_divergence_mv_gaussian_v2_chol(self):
_, mu1, covar1 = self._random_normal_params(cov_rep.CovarianceFull)
chol_covariance1 = np.linalg.cholesky(covar1)
_, mu2, covar2 = self._random_normal_params(cov_rep.CovarianceFull)
chol_covariance2 = np.linalg.cholesky(covar2)
tf_mvnd1 = dist.MultivariateNormalFullCovariance(
loc=mu1, covariance_matrix=covar1)
tf_mvnd2 = dist.MultivariateNormalFullCovariance(
loc=mu2, covariance_matrix=covar2)
tf_kldiv = dist.kl_divergence(tf_mvnd1, tf_mvnd2)
mu1_tf, mu2_tf = self._convert_to_tensor(mu1, mu2)
covar1 = cov_rep.CovarianceCholesky(
chol_covariance=tf.convert_to_tensor(chol_covariance1))
covar2 = cov_rep.CovarianceCholesky(
chol_covariance=tf.convert_to_tensor(chol_covariance2))
covar_kldiv = kl_divergence_mv_gaussian_v2(sigma1=covar1,
sigma2=covar2,
mu1=mu1_tf,
mu2=mu2_tf,
mean_batch=False)
self._asset_allclose_tf_feed(tf_kldiv, covar_kldiv)
tf_kldiv = tf.reduce_mean(tf_kldiv)
covar_kldiv = kl_divergence_mv_gaussian_v2(sigma1=covar1,
sigma2=covar2,
mu1=mu1_tf,
mu2=mu2_tf)
self._asset_allclose_tf_feed(tf_kldiv, covar_kldiv)
def _kl_mvnd_tfmvnd(a, b, name=None):
"""Batched KL divergence `KL(a || b)` for multivariate Normals, when "a" is a
tf.contrib.distributions.MultivariateNormal* distribution"""
a_cov_obj = cov_rep.CovarianceCholesky(chol_covariance=a.scale.to_dense())
return kl_divergence_mv_gaussian_v2(mu1=a.loc,
mu2=b.loc,
sigma1=a_cov_obj,
sigma2=b.cov_obj,
mean_batch=False,
name=name)
def _create_covariance_instance(self):
self.cov_object = cov_rep.CovarianceCholesky(
chol_covariance=self.tf_input,
inversion_method=self.inversion_method)
self.np_covariance = self.create_random_covariance(
self.batch_size, self.features_size, self.dtype, False)
self._create_np_precision_cholesky()
self.np_input = self.np_chol_covariance
def __init__(self,
loc,
chol_covariance=None,
chol_precision=None,
log_diag_chol_covariance=None,
log_diag_chol_precision=None,
validate_args=False,
allow_nan_stats=True,
name="MultivariateNormalChol"):
parameters = locals()
cov_obj = None
graph_parents = None
if chol_covariance is not None:
assert log_diag_chol_covariance is not None, 'Must provide log_diag of Cholesky matrix'
chol_covariance = tf.convert_to_tensor(chol_covariance)
log_diag_chol_covariance = tf.convert_to_tensor(
log_diag_chol_covariance)
cov_obj = cov_rep.CovarianceCholesky(
chol_covariance=chol_covariance)
cov_obj.log_diag_chol_covariance = log_diag_chol_covariance
graph_parents = [chol_covariance, log_diag_chol_covariance]
assert chol_precision is None
if chol_precision is not None:
assert log_diag_chol_precision is not None, 'Must provide log_diag of Cholesky matrix'
chol_precision = tf.convert_to_tensor(chol_precision)
log_diag_chol_precision = tf.convert_to_tensor(
log_diag_chol_precision)
cov_obj = cov_rep.PrecisionCholesky(chol_precision=chol_precision)
cov_obj.log_diag_chol_precision = log_diag_chol_precision
graph_parents = [chol_precision, log_diag_chol_precision]
assert chol_covariance is None
if cov_obj is None:
raise RuntimeError(
'Must provide chol_covariance or chol_precision')
super().__init__(loc=loc,
cov_obj=cov_obj,
validate_args=validate_args,
allow_nan_stats=allow_nan_stats,
name=name)
self._parameters = parameters
def test_neg_log_likelihood_mv_gaussian_chol(self):
x, mu, covariance = self._random_normal_params(
cov_rep.CovarianceCholesky)
chol_covariance = np.linalg.cholesky(covariance)
tf_mvnd = tfd.MultivariateNormalFullCovariance(
loc=mu, covariance_matrix=covariance)
tf_nll = -tf_mvnd.log_prob(x)
covar = cov_rep.CovarianceCholesky(
chol_covariance=tf.convert_to_tensor(chol_covariance))
r_tf = tf.convert_to_tensor(x - mu)
nll = neg_log_likelihood_mv_gaussian(
r_tf,
x_precision_x=covar.x_precision_x(r_tf),
log_det_cov=covar.log_det_covariance(),
mean_batch=False)
self._asset_allclose_tf_feed(nll, tf_nll)