def testJacobianWithTensors(self):
bijector = tfb.CholeskyToInvCholesky()
x = np.array([[[3., 0.], [1., 4.]], [[2., 0.], [7., 1.]]],
dtype=np.float32)
fldj = bijector.forward_log_det_jacobian(x, event_ndims=2)
fldj0 = bijector.forward_log_det_jacobian(x[0], event_ndims=2)
fldj1 = bijector.forward_log_det_jacobian(x[1], event_ndims=2)
fldj_, fldj0_, fldj1_ = self.evaluate([fldj, fldj0, fldj1])
self.assertAllClose(fldj_[0], fldj0_)
self.assertAllClose(fldj_[1], fldj1_)
def testBijector(self):
bijector = tfb.CholeskyToInvCholesky()
self.assertEqual("cholesky_to_inv_cholesky", bijector.name)
x = np.array([[3., 0.], [2., 7.]], dtype=np.float32)
m = x.dot(x.T)
m_inv = np.linalg.inv(m)
y = np.linalg.cholesky(m_inv)
x_fwd = bijector.forward(x)
y_inv = bijector.inverse(x_fwd)
x_fwd_, y_inv_ = self.evaluate([x_fwd, y_inv])
self.assertAllClose(y, x_fwd_, atol=1.e-5, rtol=1.e-5)
self.assertAllClose(x, y_inv_, atol=1.e-5, rtol=1.e-5)
def testBijectorWithTensors(self):
bijector = tfb.CholeskyToInvCholesky()
x = np.array([[[3., 0.], [1., 4.]], [[2., 0.], [7., 1.]]],
dtype=np.float32)
y = bijector.forward(x)
y0 = bijector.forward(x[0, :])
y1 = bijector.forward(x[1, :])
y_inv = bijector.inverse(y)
y_inv0 = bijector.inverse(y[0, :])
y_inv1 = bijector.inverse(y[1, :])
y_, y0_, y1_, y_inv_, y_inv0_, y_inv1_ = self.evaluate(
[y, y0, y1, y_inv, y_inv0, y_inv1])
self.assertAllClose(y_[0, :], y0_, atol=1.e-5, rtol=1.e-5)
self.assertAllClose(y_[1, :], y1_, atol=1.e-5, rtol=1.e-5)
self.assertAllClose(y_inv_[0, :], y_inv0_, atol=1.e-5, rtol=1.e-5)
self.assertAllClose(y_inv_[1, :], y_inv1_, atol=1.e-5, rtol=1.e-5)
self.assertAllClose(y_inv_, x, atol=1.e-5, rtol=1.e-5)
def testJacobian(self):
cholesky_to_vector = tfb.Invert(
tfb.FillScaleTriL(diag_bijector=tfb.Exp(), diag_shift=None))
bijector = tfb.CholeskyToInvCholesky()
for x in [
np.array([[2.]], dtype=np.float64),
np.array([[2., 0.], [3., 4.]], dtype=np.float64),
np.array([[2., 0., 0.], [3., 4., 0.], [5., 6., 7.]],
dtype=np.float64)
]:
fldj = bijector.forward_log_det_jacobian(x, event_ndims=2)
fldj_numerical = self._get_fldj_numerical(
bijector,
x,
event_ndims=2,
input_to_vector=cholesky_to_vector,
output_to_vector=cholesky_to_vector)
fldj_, fldj_numerical_ = self.evaluate([fldj, fldj_numerical])
self.assertAllClose(fldj_, fldj_numerical_, rtol=1e-2)
def testJacobian(self):
cholesky_to_vector = tfb.Chain([
tfb.Invert(tfb.FillTriangular()),
tfb.TransformDiagonal(tfb.Invert(tfb.Exp()))
])
bijector = tfb.CholeskyToInvCholesky()
for x in [np.array([[2.]],
dtype=np.float64),
np.array([[2., 0.], [3., 4.]],
dtype=np.float64),
np.array([[2., 0., 0.], [3., 4., 0.], [5., 6., 7.]],
dtype=np.float64)]:
fldj = bijector.forward_log_det_jacobian(x, event_ndims=2)
fldj_numerical = self._get_fldj_numerical(
bijector, x, event_ndims=2, eps=1.e-6,
input_to_vector=cholesky_to_vector,
output_to_vector=cholesky_to_vector)
fldj_, fldj_numerical_ = self.evaluate([fldj, fldj_numerical])
self.assertAllClose(fldj_, fldj_numerical_)
