def hessian_params(self, points: np.ndarray, params: Union[List[float], List[Interval]]) -> Union[List[List[np.ndarray]], List[List[Interval]]]:
a, b, theta = params
x = points[:, 0]
y = points[:, 1]
z = points[:, 2]
zero = iv.zeros_like(z)
return [
[
iv.stack([
z * (iv.cos(theta) * x - iv.sin(theta) * y),
z * (iv.sin(theta) * x + iv.cos(theta) * y),
zero
], axis=1, convert=isinstance(a, Interval)),
iv.stack([
zero,
zero,
zero
], axis=1, convert=isinstance(a, Interval)),
iv.stack([
-a * z * (iv.sin(theta) * x + iv.cos(theta) * y),
a * z * (iv.cos(theta) * x - iv.sin(theta) * y),
zero
], axis=1, convert=isinstance(a, Interval))
],
[
iv.stack([
zero,
zero,
zero
], axis=1, convert=isinstance(a, Interval)),
iv.stack([
zero,
zero,
zero
], axis=1, convert=isinstance(a, Interval)),
iv.stack([
-z * (iv.sin(theta) * x + iv.cos(theta) * y),
z * (iv.cos(theta) * x - iv.sin(theta) * y),
zero
], axis=1, convert=isinstance(a, Interval))
],
[
iv.stack([
-a * z * (iv.sin(theta) * x + iv.cos(theta) * y),
a * z * (iv.cos(theta) * x - iv.sin(theta) * y),
zero
], axis=1, convert=isinstance(a, Interval)),
iv.stack([
-z * (iv.sin(theta) * x + iv.cos(theta) * y),
z * (iv.cos(theta) * x - iv.sin(theta) * y),
zero
], axis=1, convert=isinstance(a, Interval)),
iv.stack([
(0.5 * iv.square(a) * z + b * z + 1) * (-iv.cos(theta) * x + iv.sin(theta) * y),
(0.5 * iv.square(a) * z + b * z + 1) * (-iv.sin(theta) * x - iv.cos(theta) * y),
zero
], axis=1, convert=isinstance(a, Interval))
]
]
def transform(
self, points: np.ndarray,
params: Union[List[float],
List[Interval]]) -> Union[np.ndarray, Interval]:
a, b = params
x = points[:, 0]
y = points[:, 1]
z = points[:, 2]
x_transformed = (0.5 * iv.square(a) * z + b * z + 1) * x
y_transformed = (0.5 * iv.square(a) * z + b * z + 1) * y
z_transformed = z
return iv.stack([x_transformed, y_transformed, z_transformed],
axis=1,
convert=isinstance(a, Interval))
def hessian_points(self, points: np.ndarray, params: Union[List[float], List[Interval]]) -> Union[List[List[np.ndarray]], List[List[Interval]]]:
a, b, theta = params
z = points[:, 2]
zero = iv.zeros_like(z)
one = iv.ones_like(z)
return [
[
iv.stack([
zero,
zero,
zero
], axis=1, convert=isinstance(a, Interval)),
iv.stack([
zero,
zero,
zero
], axis=1, convert=isinstance(a, Interval)),
iv.stack([
(0.5 * iv.square(a) + b) * iv.cos(theta) * one,
(0.5 * iv.square(a) + b) * iv.sin(theta) * one,
zero
], axis=1, convert=isinstance(a, Interval)),
],
[
iv.stack([
zero,
zero,
zero
], axis=1, convert=isinstance(a, Interval)),
iv.stack([
zero,
zero,
zero
], axis=1, convert=isinstance(a, Interval)),
iv.stack([
-(0.5 * iv.square(a) + b) * iv.sin(theta) * one,
(0.5 * iv.square(a) + b) * iv.cos(theta) * one,
zero
], axis=1, convert=isinstance(a, Interval))
],
[
iv.stack([
(0.5 * iv.square(a) + b) * iv.cos(theta) * one,
(0.5 * iv.square(a) + b) * iv.sin(theta) * one,
zero
], axis=1, convert=isinstance(a, Interval)),
iv.stack([
-(0.5 * iv.square(a) + b) * iv.sin(theta) * one,
(0.5 * iv.square(a) + b) * iv.cos(theta) * one,
zero
], axis=1, convert=isinstance(a, Interval)),
iv.stack([
zero,
zero,
zero
], axis=1, convert=isinstance(a, Interval))
],
]
def hessian_params(
self, points: np.ndarray, params: Union[List[float], List[Interval]]
) -> Union[List[List[np.ndarray]], List[List[Interval]]]:
assert len(params) == 1
alpha = params[0]
x = points[:, 0]
y = points[:, 1]
z = points[:, 2]
d_alpha_alpha_x = -iv.square(z) * (x * iv.cos(alpha * z) -
y * iv.sin(alpha * z))
d_alpha_alpha_y = -iv.square(z) * (x * iv.sin(alpha * z) +
y * iv.cos(alpha * z))
d_alpha_alpha_z = iv.zeros_like(z)
return [[
iv.stack([d_alpha_alpha_x, d_alpha_alpha_y, d_alpha_alpha_z],
axis=1,
convert=isinstance(alpha, Interval))
]]
def gradient_points(
self, points: np.ndarray, params: Union[List[float], List[Interval]]
) -> Union[List[np.ndarray], List[Interval]]:
a, b = params
x = points[:, 0]
y = points[:, 1]
z = points[:, 2]
zero = iv.zeros_like(z)
return [
iv.stack([(0.5 * iv.square(a) * z + b * z + 1), zero, zero],
axis=1,
convert=isinstance(a, Interval)),
iv.stack([zero, (0.5 * iv.square(a) * z + b * z + 1), zero],
axis=1,
convert=isinstance(a, Interval)),
iv.stack([(0.5 * iv.square(a) + b) * x,
(0.5 * iv.square(a) + b) * y,
iv.ones_like(z)],
axis=1,
convert=isinstance(a, Interval)),
]
def hessian_points(
self, points: np.ndarray, params: Union[List[float], List[Interval]]
) -> Union[List[List[np.ndarray]], List[List[Interval]]]:
a, b = params
x = points[:, 0]
zero = iv.zeros_like(x)
one = iv.ones_like(x)
return [[
iv.stack([zero, zero, zero],
axis=1,
convert=isinstance(a, Interval)),
iv.stack([zero, zero, zero],
axis=1,
convert=isinstance(a, Interval)),
iv.stack([(0.5 * iv.square(a) + b) * one, zero, zero],
axis=1,
convert=isinstance(a, Interval)),
],
[
iv.stack([zero, zero, zero],
axis=1,
convert=isinstance(a, Interval)),
iv.stack([zero, zero, zero],
axis=1,
convert=isinstance(a, Interval)),
iv.stack([zero, (0.5 * iv.square(a) + b) * one, zero],
axis=1,
convert=isinstance(a, Interval)),
],
[
iv.stack([(0.5 * iv.square(a) + b) * one, zero, zero],
axis=1,
convert=isinstance(a, Interval)),
iv.stack([zero, (0.5 * iv.square(a) + b) * one, zero],
axis=1,
convert=isinstance(a, Interval)),
iv.stack([zero, zero, zero],
axis=1,
convert=isinstance(a, Interval)),
]]
def encode(transformation: Transformation, points: np.ndarray, params: List[Interval]) -> LinearBounds:
center = [p.center() for p in params]
zero_order_term = transformation.transform(points, center)
gradients = transformation.gradient_params(points, center)
hessian = transformation.hessian_params(points, params)
offset = zero_order_term
for i in range(len(params)):
offset = offset - (gradients[i] * center[i])
second_order_bounds = 0.0
for i, j in itertools.product(range(len(params)), range(len(params))):
factor = iv.square(params[i] - center[i]) if i == j else (params[i] - center[i]) * (params[j] - center[j])
second_order_bounds = second_order_bounds + (0.5 * hessian[i][j] * factor)
return LinearBounds(
upper_slope=np.stack(gradients, axis=2),
lower_slope=np.stack(gradients, axis=2),
upper_offset=offset + second_order_bounds.upper_bound,
lower_offset=offset + second_order_bounds.lower_bound,
)
def gradient_points(self, points: np.ndarray, params: Union[List[float], List[Interval]]) -> Union[List[np.ndarray], List[Interval]]:
a, b, theta = params
x = points[:, 0]
y = points[:, 1]
z = points[:, 2]
zero = iv.zeros_like(z)
one = iv.ones_like(z)
return [
iv.stack([
(0.5 * iv.square(a) * z + b * z + 1) * iv.cos(theta),
(0.5 * iv.square(a) * z + b * z + 1) * iv.sin(theta),
zero
], axis=1, convert=isinstance(a, Interval)),
iv.stack([
-(0.5 * iv.square(a) * z + b * z + 1) * iv.sin(theta),
(0.5 * iv.square(a) * z + b * z + 1) * iv.cos(theta),
zero
], axis=1, convert=isinstance(a, Interval)),
iv.stack([
(0.5 * iv.square(a) + b) * (iv.cos(theta) * x - iv.sin(theta) * y),
(0.5 * iv.square(a) + b) * (iv.sin(theta) * x + iv.cos(theta) * y),
one
], axis=1, convert=isinstance(a, Interval))
]
def hessian_points(
self, points: np.ndarray, params: Union[List[float], List[Interval]]
) -> Union[List[List[np.ndarray]], List[List[Interval]]]:
(alpha, ) = params
x = points[:, 0]
y = points[:, 1]
z = points[:, 2]
zero = iv.zeros_like(x)
return [
[
iv.stack([
zero,
zero,
zero,
],
axis=1,
convert=isinstance(alpha, Interval)),
iv.stack([
zero,
zero,
zero,
],
axis=1,
convert=isinstance(alpha, Interval)),
iv.stack([
-alpha * iv.sin(alpha * z),
alpha * iv.cos(alpha * z),
zero,
],
axis=1,
convert=isinstance(alpha, Interval)),
],
[
iv.stack([
zero,
zero,
zero,
],
axis=1,
convert=isinstance(alpha, Interval)),
iv.stack([
zero,
zero,
zero,
],
axis=1,
convert=isinstance(alpha, Interval)),
iv.stack([
-alpha * iv.cos(alpha * z),
-alpha * iv.sin(alpha * z),
zero,
],
axis=1,
convert=isinstance(alpha, Interval)),
],
[
iv.stack([
-alpha * iv.sin(alpha * z),
alpha * iv.cos(alpha * z),
zero,
],
axis=1,
convert=isinstance(alpha, Interval)),
iv.stack([
-alpha * iv.cos(alpha * z),
-alpha * iv.sin(alpha * z),
zero,
],
axis=1,
convert=isinstance(alpha, Interval)),
iv.stack([
iv.square(alpha) *
(-iv.cos(alpha * z) * x + iv.sin(alpha * z) * y),
iv.square(alpha) *
(-iv.sin(alpha * z) * x - iv.cos(alpha * z) * y),
zero,
],
axis=1,
convert=isinstance(alpha, Interval)),
],
]
def test_single_scalar(self, a: float, expected: float):
self.assertAlmostEqual(expected, iv.square(a))
def test_single_interval(self, a: Interval, expected: Interval):
self.assertAlmostEqual(expected.lower_bound, iv.square(a).lower_bound)
self.assertAlmostEqual(expected.upper_bound, iv.square(a).upper_bound)
self.assertSound(iv.square(a), [a], np.square)
def test_nd_array(self, a):
self.assertAlmostEqualNumpy(np.square(a), iv.square(a))
def test_nd_interval(self, a):
self.assertSound(iv.square(a), [a], np.square)
