def noise_accumulator_test_template(test_case, model, num_steps):
"""Tests `model`'s transition_power_noise_accumulator."""
transition_matrix = ops.convert_to_tensor(model.get_state_transition(),
dtype=model.dtype)
noise_transform = ops.convert_to_tensor(model.get_noise_transform(),
dtype=model.dtype)
state_dimension = tensor_shape.dimension_value(transition_matrix.shape[0])
state_noise_dimension = tensor_shape.dimension_value(
noise_transform.shape[1])
gen_noise_addition = math_utils.sign_magnitude_positive_definite(
raw=random_ops.random_normal(
shape=[state_noise_dimension, state_noise_dimension],
dtype=model.dtype))
gen_starting_noise = math_utils.sign_magnitude_positive_definite(
random_ops.random_normal(shape=[state_dimension, state_dimension],
dtype=model.dtype))
starting_noise = array_ops.placeholder(
shape=[state_dimension, state_dimension], dtype=model.dtype)
step_number = array_ops.placeholder(shape=[], dtype=dtypes.int64)
starting_transitioned = math_ops.matmul(math_ops.matmul(
transition_matrix, starting_noise),
transition_matrix,
adjoint_b=True)
with test_case.test_session():
evaled_starting_noise = gen_starting_noise.eval()
current_starting_noise_transitioned = evaled_starting_noise
current_noise = evaled_starting_noise
evaled_noise_addition = gen_noise_addition.eval()
evaled_noise_addition_transformed = math_ops.matmul(
math_ops.matmul(noise_transform, evaled_noise_addition),
noise_transform,
adjoint_b=True).eval()
model.state_transition_noise_covariance = evaled_noise_addition
model._window_initializer( # pylint: disable=protected-access
times=math_ops.range(num_steps + 1)[..., None],
state=(None, None, 0))
model_update = model.transition_power_noise_accumulator(
num_steps=step_number)
for iteration_number in range(num_steps):
model_new_noise = model_update.eval(
feed_dict={step_number: iteration_number})
test_case.assertAllClose(
current_noise,
model_new_noise + current_starting_noise_transitioned,
rtol=1e-8 if current_noise.dtype == numpy.float64 else 1e-3)
current_starting_noise_transitioned = starting_transitioned.eval(
feed_dict={
starting_noise: current_starting_noise_transitioned
})
current_noise = (starting_transitioned.eval(
feed_dict={starting_noise: current_noise}) +
evaled_noise_addition_transformed)
def noise_accumulator_test_template(test_case, model, num_steps):
"""Tests `model`'s transition_power_noise_accumulator."""
transition_matrix = ops.convert_to_tensor(
model.get_state_transition(), dtype=model.dtype)
noise_transform = ops.convert_to_tensor(
model.get_noise_transform(), dtype=model.dtype)
state_dimension = transition_matrix.get_shape()[0].value
state_noise_dimension = noise_transform.get_shape()[1].value
gen_noise_addition = math_utils.sign_magnitude_positive_definite(
raw=random_ops.random_normal(
shape=[state_noise_dimension, state_noise_dimension],
dtype=model.dtype))
gen_starting_noise = math_utils.sign_magnitude_positive_definite(
random_ops.random_normal(
shape=[state_dimension, state_dimension], dtype=model.dtype))
starting_noise = array_ops.placeholder(
shape=[state_dimension, state_dimension], dtype=model.dtype)
step_number = array_ops.placeholder(shape=[], dtype=dtypes.int64)
starting_transitioned = math_ops.matmul(
math_ops.matmul(transition_matrix, starting_noise),
transition_matrix,
adjoint_b=True)
with test_case.test_session():
evaled_starting_noise = gen_starting_noise.eval()
current_starting_noise_transitioned = evaled_starting_noise
current_noise = evaled_starting_noise
evaled_noise_addition = gen_noise_addition.eval()
evaled_noise_addition_transformed = math_ops.matmul(
math_ops.matmul(noise_transform, evaled_noise_addition),
noise_transform,
adjoint_b=True).eval()
model.state_transition_noise_covariance = evaled_noise_addition
model._window_initializer( # pylint: disable=protected-access
times=math_ops.range(num_steps + 1)[..., None], state=(None, None, 0))
model_update = model.transition_power_noise_accumulator(
num_steps=step_number)
for iteration_number in range(num_steps):
model_new_noise = model_update.eval(
feed_dict={step_number: iteration_number})
test_case.assertAllClose(
current_noise,
model_new_noise + current_starting_noise_transitioned,
rtol=1e-8 if current_noise.dtype == numpy.float64 else 1e-3)
current_starting_noise_transitioned = starting_transitioned.eval(
feed_dict={starting_noise: current_starting_noise_transitioned})
current_noise = (
starting_transitioned.eval(
feed_dict={starting_noise: current_noise})
+ evaled_noise_addition_transformed)
def test_sign_magnitude_positive_definite(self):
for dtype in [dtypes.float32, dtypes.float64]:
with self.cached_session():
matrix_tensor = math_utils.sign_magnitude_positive_definite(
raw=constant_op.constant([[-1., -2.], [3., 4.]], dtype=dtype),
off_diagonal_scale=constant_op.constant(-1., dtype=dtype),
overall_scale=constant_op.constant(1., dtype=dtype))
matrix_evaled = matrix_tensor.eval()
self.assertAllClose(matrix_evaled, matrix_evaled.T)
self.assertTrue(numpy.all(numpy.linalg.eigvals(matrix_evaled) > 0))
def test_zero_size_matrix(self):
raw = numpy.zeros([0, 0])
with self.cached_session():
constructed = math_utils.sign_magnitude_positive_definite(raw=raw).eval()
self.assertEqual((0, 0), constructed.shape)