def testBijector(self):
with self.test_session():
for fwd in [
bijectors.Identity(),
bijectors.Exp(event_ndims=1),
bijectors.Affine(shift=[0., 1.],
scale_diag=[2., 3.],
event_ndims=1),
bijectors.Softplus(event_ndims=1),
bijectors.SoftmaxCentered(event_ndims=1),
bijectors.SigmoidCentered(),
]:
rev = bijectors.Invert(fwd)
self.assertEqual("_".join(["invert", fwd.name]), rev.name)
x = [[[1., 2.], [2., 3.]]]
self.assertAllClose(
fwd.inverse(x).eval(),
rev.forward(x).eval())
self.assertAllClose(
fwd.forward(x).eval(),
rev.inverse(x).eval())
self.assertAllClose(
fwd.forward_log_det_jacobian(x).eval(),
rev.inverse_log_det_jacobian(x).eval())
self.assertAllClose(
fwd.inverse_log_det_jacobian(x).eval(),
rev.forward_log_det_jacobian(x).eval())
def testDistWithBatchShapeOneThenTransformedThroughSoftplus(self):
# This complex combination of events resulted in a loss of static shape
# information when tensor_util.constant_value(self._needs_rotation) was
# being used incorrectly (resulting in always rotating).
# Batch shape = [1], event shape = [3]
mu = array_ops.zeros((1, 3))
diag = array_ops.ones((1, 3))
with self.cached_session():
base_dist = ds.MultivariateNormalDiag(mu, diag, validate_args=True)
dist = ds.TransformedDistribution(base_dist,
validate_args=True,
bijector=bijectors.Softplus())
samps = dist.sample(5) # Shape [5, 1, 3].
self.assertAllEqual([5, 1], dist.log_prob(samps).get_shape())
def testInvertible(self):
# Generate random inputs from an unconstrained space, with
# event size 6 to specify 3x3 triangular matrices.
batch_shape = [2, 1]
x = np.float32(self._rng.randn(*(batch_shape + [6])))
b = bijectors.ScaleTriL(diag_bijector=bijectors.Softplus(),
diag_shift=3.14159)
y = self.evaluate(b.forward(x))
self.assertAllEqual(y.shape, batch_shape + [3, 3])
x_ = self.evaluate(b.inverse(y))
self.assertAllClose(x, x_, rtol=1e-4)
fldj = self.evaluate(b.forward_log_det_jacobian(x, event_ndims=1))
ildj = self.evaluate(b.inverse_log_det_jacobian(y, event_ndims=2))
self.assertAllClose(fldj, -ildj, rtol=1e-4)