def testHalfCauchyKLDivergence(self):
shape = (3, )
regularizer = ed.regularizers.get('half_cauchy_kl_divergence')
variational_posterior = ed.Independent(ed.LogNormal(
loc=tf.zeros(shape), scale=1.).distribution,
reinterpreted_batch_ndims=1)
kl = regularizer(variational_posterior)
kl_value = self.evaluate(kl)
self.assertGreaterEqual(kl_value, 0.)
def testHalfCauchyKLDivergence(self):
shape = (3,)
regularizer = ed.regularizers.get('half_cauchy_kl_divergence')
variational_posterior = ed.Independent(
ed.LogNormal(loc=tf.zeros(shape), scale=1.).distribution,
reinterpreted_batch_ndims=1)
kl = regularizer(variational_posterior)
# KL uses a single-sample estimate, which is not necessarily >0. We only
# check shape.
self.assertEqual(kl.shape, ())
def testLogNormalKLDivergence(self):
shape = (3,)
regularizer = ed.regularizers.get('log_normal_kl_divergence')
variational_posterior = ed.Independent(
ed.LogNormal(loc=tf.zeros(shape), scale=1.).distribution,
reinterpreted_batch_ndims=1)
kl = regularizer(variational_posterior)
self.assertGreaterEqual(kl, 0.)
dataset_size = 100
scale_factor = 1. / dataset_size
regularizer = ed.regularizers.LogNormalKLDivergence(
scale_factor=scale_factor)
scaled_kl = regularizer(variational_posterior)
self.assertEqual(scale_factor * kl, scaled_kl)