def test_taylor_approximation(self, x, params):
manual_transformation = RotationZX()
auto_transformation = Composition(RotationZ(), RotationX())
manual = taylor.encode(manual_transformation, x, params)
auto = taylor.encode(auto_transformation, x, params)
self.assertSound(manual, params,
partial(manual_transformation.transform, x))
self.assertSound(auto, params, partial(auto_transformation.transform,
x))
def test_taylor_approximation(self, x, params):
manual_transformation = TaperingRotation()
auto_transformation = Composition(TaperingZ(), RotationZ())
expected = taylor.encode(manual_transformation, x, params)
actual = taylor.encode(auto_transformation, x, params)
self.assertSound(expected, params,
partial(manual_transformation.transform, x))
self.assertSound(actual, params,
partial(auto_transformation.transform, x))
self.assertAlmostEqualBounds(expected, actual)
def concretize(self, x, A, sign=-1, aux=None):
n_batch = A.shape[0]
n_outputs = A.shape[1]
n_values = A.shape[2]
n_points = x.shape[1]
n_coords = x.shape[2]
n_params = len(self.params)
assert n_values == n_points * n_coords
# Computing linear constraints based on taylor relaxations of transformation
x_np = x.cpu().numpy()
x_np = np.reshape(x_np, (n_batch * n_points, n_coords))
bounds = taylor.encode(self.transformation, x_np, self.params)
lower_offset = torch.tensor(bounds.lower_offset.reshape(
(n_batch, n_values, 1)),
device=x.device)
upper_offset = torch.tensor(bounds.upper_offset.reshape(
(n_batch, n_values, 1)),
device=x.device)
lower_slope = torch.tensor(bounds.lower_slope.reshape(
(n_batch, n_values, n_params)),
device=x.device)
upper_slope = torch.tensor(bounds.upper_slope.reshape(
(n_batch, n_values, n_params)),
device=x.device)
# Backwards propagate coefficients through linear relaxation of transformation
if sign == -1: # computing lower bound
new_A = torch.matmul(A.clamp(min=0.0), lower_slope) + torch.matmul(
A.clamp(max=0.0), upper_slope)
offset = torch.matmul(A.clamp(min=0.0),
lower_offset) + torch.matmul(
A.clamp(max=0.0), upper_offset)
elif sign == 1: # computing upper bound
new_A = torch.matmul(A.clamp(min=0.0), upper_slope) + torch.matmul(
A.clamp(max=0.0), lower_slope)
offset = torch.matmul(A.clamp(min=0.0),
upper_offset) + torch.matmul(
A.clamp(max=0.0), lower_offset)
else:
raise RuntimeError(f"Invalid sign value: {sign}")
# Instantiate bounds based on valid parameter ranges. Same implementation as for L-inf perturbation in PerturbationLpNorm
lb = torch.tensor([[p.lower_bound] for p in self.params],
dtype=x.dtype,
device=x.device).reshape((1, n_params, 1))
ub = torch.tensor([[p.upper_bound] for p in self.params],
dtype=x.dtype,
device=x.device).reshape((1, n_params, 1))
center = (ub + lb) / 2.0
diff = (ub - lb) / 2.0
bound = new_A.matmul(center) + sign * new_A.abs().matmul(diff)
result = bound + offset
assert result.shape == (n_batch, n_outputs, 1)
return result.squeeze(-1)
if settings.relaxation == 'interval':
bounds = transformation.transform(np_points, params)
if settings.implicit_intervals > 1:
bounds = split_implicitly(
bounds=bounds, intervals=settings.implicit_intervals,
encode_rotation=partial(transformation.transform, np_points),
params=params
)
(dominant_classes, nlb, nub) = eran.analyze_segmentation_box(bounds, label, valid_classes, num_total_classes)
elif settings.relaxation == 'taylor':
bounds = transformation.transform(np_points, params)
constraints = taylor.encode(transformation, np_points, params)
if settings.implicit_intervals > 1:
bounds = split_implicitly(
bounds=bounds, intervals=settings.implicit_intervals,
encode_rotation=partial(transformation.transform, np_points),
params=params
)
(dominant_classes, nlb, nub) = eran.analyze_segmentation_linear(bounds, constraints, params, label, valid_classes, num_total_classes)
dominant_classes = np.array(dominant_classes)
assert np.all(np.logical_or(label == dominant_classes, dominant_classes == -1)), \
f"Wrong dominant class! label {label}, dominant_class: {dominant_classes}"
else:
def test_transformation_taylor(self, x, params):
transformation = TwistingZ()
actual = taylor.encode(transformation, x, params)
self.assertSound(actual, params, partial(transformation.transform, x))
def test_taylor_twisting_rotation(self, x, params):
transformation = Composition(TwistingZ(), RotationZ())
actual = taylor.encode(transformation, x, params)
self.assertSound(actual, params, partial(transformation.transform, x))
def test_taylor_rot_zyx(self, x, params):
transformation = Composition(RotationZ(),
Composition(RotationY(), RotationX()))
actual = taylor.encode(transformation, x, params)
self.assertSound(actual, params, partial(transformation.transform, x))
