def testGetAndMakeFromParameters(self):
one = tf.constant(1.0)
d = tfp.distributions.Normal(loc=one, scale=3.0, validate_args=True)
d = tfp.bijectors.Tanh()(d)
d = tfp.bijectors.Tanh()(d)
p = utils.get_parameters(d)
expected_p = utils.Params(
tfp.distributions.TransformedDistribution,
params={
'bijector':
utils.Params(tfp.bijectors.Chain,
params={
'bijectors': [
utils.Params(tfp.bijectors.Tanh,
params={}),
utils.Params(tfp.bijectors.Tanh,
params={}),
]
}),
'distribution':
utils.Params(tfp.distributions.Normal,
params={
'validate_args': True,
'scale': 3.0,
'loc': one
})
})
self.compare_params(p, expected_p)
d_recreated = utils.make_from_parameters(p)
points = [0.01, 0.25, 0.5, 0.75, 0.99]
self.assertAllClose(d.log_prob(points), d_recreated.log_prob(points))
def testGetAndMakeNontrivialBijectorFromParameters(self):
scale_matrix = tf.Variable([[1.0, 2.0], [-1.0, 0.0]])
d = tfp.distributions.MultivariateNormalDiag(loc=[1.0, 1.0],
scale_diag=[2.0, 3.0],
validate_args=True)
b = tfp.bijectors.ScaleMatvecLinearOperator(
scale=tf.linalg.LinearOperatorFullMatrix(matrix=scale_matrix),
adjoint=True)
b_d = b(d)
p = utils.get_parameters(b_d)
expected_p = utils.Params(
tfp.distributions.TransformedDistribution,
params={
'bijector':
utils.Params(tfp.bijectors.ScaleMatvecLinearOperator,
params={
'adjoint':
True,
'scale':
utils.Params(
tf.linalg.LinearOperatorFullMatrix,
params={'matrix': scale_matrix})
}),
'distribution':
utils.Params(tfp.distributions.MultivariateNormalDiag,
params={
'validate_args': True,
'scale_diag': [2.0, 3.0],
'loc': [1.0, 1.0]
})
})
self.compare_params(p, expected_p)
b_d_recreated = utils.make_from_parameters(p)
points = [[-1.0, -2.0], [0.0, 0.0], [3.0, -5.0], [5.0, 5.0],
[1.0, np.inf], [-np.inf, 0.0]]
self.assertAllClose(b_d.log_prob(points),
b_d_recreated.log_prob(points))
def merge_and_convert(spec, params):
return distribution_utils.make_from_parameters(
distribution_utils.merge_to_parameters_from_dict(
spec.parameters, params))