def test_pointwise_norm_gradient_real(exponent):
# The operator is not differentiable for exponent 'inf'
if exponent == float('inf'):
fspace = odl.uniform_discr([0, 0], [1, 1], (2, 2))
vfspace = ProductSpace(fspace, 1)
pwnorm = PointwiseNorm(vfspace, exponent)
point = vfspace.one()
with pytest.raises(NotImplementedError):
pwnorm.derivative(point)
return
# 1d
fspace = odl.uniform_discr([0, 0], [1, 1], (2, 2))
vfspace = ProductSpace(fspace, 1)
pwnorm = PointwiseNorm(vfspace, exponent)
point = noise_element(vfspace)
direction = noise_element(vfspace)
# Computing expected result
tmp = pwnorm(point).ufuncs.power(1 - exponent)
v_field = vfspace.element()
for i in range(len(v_field)):
v_field[i] = tmp * point[i] * np.abs(point[i])**(exponent - 2)
pwinner = odl.PointwiseInner(vfspace, v_field)
expected_result = pwinner(direction)
func_pwnorm = pwnorm.derivative(point)
assert all_almost_equal(func_pwnorm(direction), expected_result)
# 3d
fspace = odl.uniform_discr([0, 0], [1, 1], (2, 2))
vfspace = ProductSpace(fspace, 3)
pwnorm = PointwiseNorm(vfspace, exponent)
point = noise_element(vfspace)
direction = noise_element(vfspace)
# Computing expected result
tmp = pwnorm(point).ufuncs.power(1 - exponent)
v_field = vfspace.element()
for i in range(len(v_field)):
v_field[i] = tmp * point[i] * np.abs(point[i])**(exponent - 2)
pwinner = odl.PointwiseInner(vfspace, v_field)
expected_result = pwinner(direction)
func_pwnorm = pwnorm.derivative(point)
assert all_almost_equal(func_pwnorm(direction), expected_result)
def inf_action(self, lie_alg_element):
assert lie_alg_element in self.lie_group.associated_algebra
pts = self.domain.points()
deformed_pts = lie_alg_element.arr.dot(pts.T) - pts.T
deformed_pts = self.domain.tangent_bundle.element(deformed_pts)
pointwise_inner = odl.PointwiseInner(self.gradient.range, deformed_pts)
return pointwise_inner * self.gradient
def _call(self, x, out):
xi = vfield
Id = IdentityOperator(domain)
xiT = odl.PointwiseInner(domain, xi)
xixiT = odl.BroadcastOperator(*[x * xiT for x in xi])
gamma = 1
P = (Id - gamma * xixiT)
out.assign(P(x))
def gradient(space,
sinfo=None,
mode=None,
gamma=1,
eta=1e-2,
show_sinfo=False,
prefix=None):
grad = odl.Gradient(space, method='forward', pad_mode='symmetric')
if sinfo is not None:
if mode == 'direction':
norm = odl.PointwiseNorm(grad.range)
grad_sinfo = grad(sinfo)
ngrad_sinfo = norm(grad_sinfo)
for i in range(len(grad_sinfo)):
grad_sinfo[i] /= ngrad_sinfo.ufuncs.max()
ngrad_sinfo = norm(grad_sinfo)
ngrad_sinfo_eta = np.sqrt(ngrad_sinfo**2 + eta**2)
xi = grad.range.element([g / ngrad_sinfo_eta
for g in grad_sinfo]) # UGLY
Id = odl.operator.IdentityOperator(grad.range)
xiT = odl.PointwiseInner(grad.range, xi)
xixiT = odl.BroadcastOperator(*[x * xiT for x in xi])
grad = (Id - gamma * xixiT) * grad
if show_sinfo:
misc.save_image(ngrad_sinfo, prefix + '_sinfo_norm')
misc.save_vfield(xi.asarray(), filename=prefix + '_sinfo_xi')
misc.save_vfield_cmap(filename=prefix + '_sinfo_xi_cmap')
elif mode == 'location':
norm = odl.PointwiseNorm(grad.range)
ngrad_sinfo = norm(grad(sinfo))
ngrad_sinfo /= ngrad_sinfo.ufuncs.max()
w = eta / np.sqrt(ngrad_sinfo**2 + eta**2)
grad = odl.DiagonalOperator(odl.MultiplyOperator(w), 2) * grad
if show_sinfo:
misc.save_image(ngrad_sinfo, prefix + '_sinfo_norm')
misc.save_image(w, prefix + '_w')
else:
grad = None
return grad
sideinfo_grad = gradient(sideinfo)
gradient_space = gradient.range
norm = odl.PointwiseNorm(gradient_space, 2)
norm_sideinfo_grad = norm(sideinfo_grad)
max_norm = np.max(norm_sideinfo_grad)
eta_scaled = eta * max(max_norm, 1e-4)
norm_eta_sideinfo_grad = np.sqrt(norm_sideinfo_grad**2 + eta_scaled**2)
xi = gradient_space.element(
[g / norm_eta_sideinfo_grad for g in sideinfo_grad]) # UGLY
I = odl.operator.IdentityOperator(gradient_space)
xiT = odl.PointwiseInner(gradient_space, xi)
xixiT = odl.BroadcastOperator(*[x * xiT for x in xi])
D = (I - xixiT) * gradient
norm_D = misc.norm(D, '{}/norm_D.npy'.format(folder_param))
c = float(norm_K) / float(norm_D)
D *= c
norm_D *= c
L1 = (alpha / c) * odl.solvers.GroupL1Norm(D.range)
L164 = (alpha / c) * odl.solvers.GroupL1Norm(D.range.astype('float64'))
g = odl.solvers.IndicatorBox(X, lower=0)
obj_fun = KL * K + L1 * D + g # objective functional
fldr = '{}/pics'.format(folder_param)
def __init__(self,
domain,
alpha=1.,
sinfo=None,
NonNeg=False,
prox_options={},
gamma=1,
eta=1e-2):
if isinstance(domain, odl.ProductSpace):
grad_basic = odl.Gradient(domain[0],
method='forward',
pad_mode='symmetric')
pd = [
odl.discr.diff_ops.PartialDerivative(domain[0],
i,
method='forward',
pad_mode='symmetric')
for i in range(2)
]
cp = [
odl.operator.ComponentProjection(domain, i) for i in range(2)
]
if sinfo is None:
self.grad = odl.BroadcastOperator(
*[pd[i] * cp[j] for i in range(2) for j in range(2)])
else:
vfield = gamma * generate_vfield_from_sinfo(
sinfo, grad_basic, eta)
inner = odl.PointwiseInner(domain, vfield) * grad_basic
self.grad = odl.BroadcastOperator(*[
pd[i] * cp[j] - vfield[i] * inner * cp[j] for i in range(2)
for j in range(2)
])
self.grad.norm = self.grad.norm(estimate=True)
else:
grad_basic = odl.Gradient(domain,
method='forward',
pad_mode='symmetric')
if sinfo is None:
self.grad = grad_basic
else:
vfield = gamma * generate_vfield_from_sinfo(
sinfo, grad_basic, eta)
P = project_on_fixed_vfield(grad_basic.range, vfield)
self.grad = P * grad_basic
grad_norm = 2 * np.sqrt(sum(1 / grad_basic.domain.cell_sides**2))
self.grad.norm = grad_norm
self.tv = total_variation(domain, grad=self.grad)
if NonNeg is True:
self.nn = IndicatorBox(domain, 0, np.inf)
else:
self.nn = ZeroFunctional(domain)
if 'name' not in prox_options:
prox_options['name'] = 'FGP'
if 'warmstart' not in prox_options:
prox_options['warmstart'] = True
if 'niter' not in prox_options:
prox_options['niter'] = 5
if 'p' not in prox_options:
prox_options['p'] = None
if 'tol' not in prox_options:
prox_options['tol'] = None
prox_options['grad'] = self.grad
prox_options['proj_P'] = self.tv.left.convex_conj.proximal(0)
prox_options['proj_C'] = self.nn.proximal(1)
self.prox_options = prox_options
self.alpha = alpha
super().__init__(space=domain, linear=False, grad_lipschitz=0)
