def test_sigma_mad(self):
"""Test sigma_mad."""
npt.assert_almost_equal(
stats.sigma_mad(self.data1),
2.9651999999999998,
err_msg='Incorrect sigma from MAD',
)
def reweight(self, x_new):
""" Updat the weights.
Parameters
----------
x_new: ndarray
the current primal solution.
Returns
-------
sigma_est: ndarray
the variance estimate on each scale.
"""
self.linear_op.op(x_new)
weights = np.empty((0, ), dtype=self.weights.dtype)
sigma_est = []
for scale in range(self.linear_op.transform.nb_scale):
bands_array, _ = flatten(self.linear_op.transform[scale])
if scale == (self.linear_op.transform.nb_scale - 1):
std_at_scale_i = 0.
else:
std_at_scale_i = sigma_mad(bands_array)
sigma_est.append(std_at_scale_i)
thr = np.ones(bands_array.shape, dtype=weights.dtype)
thr *= self.thresh_factor * std_at_scale_i
weights = np.concatenate((weights, thr))
self.weights = weights
return sigma_est
def get_weights(data, psf, filters, wave_thresh_factor=np.array([3, 3, 4])):
"""Get Sparsity Weights
Parameters
----------
data : np.ndarray
Input data, 2D image
psf : np.ndarray
Input PSF, 2D image
filters : np.ndarray
Wavelet filters
wave_thresh_factor : np.ndarray, optional
Threshold factors for each wavelet scale (default is
np.array([3, 3, 4]))
Returns
-------
np.ndarray weights
"""
noise_est = sigma_mad(data)
filter_conv = filter_convolve(np.rot90(psf, 2), filters)
filter_norm = np.array([
np.linalg.norm(a) * b * np.ones(data.shape)
for a, b in zip(filter_conv, wave_thresh_factor)
])
return noise_est * filter_norm
def set_noise(data, **kwargs):
"""Set the noise level
This method calculates the noise standard deviation using the median
absolute deviation (MAD) of the input data and adds it to the keyword
arguments.
Parameters
----------
data : np.ndarray
Input noisy data (3D array)
Returns
-------
dict Updated keyword arguments
"""
# It the noise is not already provided calculate it using the MAD
if isinstance(kwargs['noise_est'], type(None)):
kwargs['noise_est'] = sigma_mad(data)
print(' - Noise Estimate:', kwargs['noise_est'])
if 'log' in kwargs:
kwargs['log'].info(' - Noise Estimate: ' + str(kwargs['noise_est']))
return kwargs
def set_noise(data, **kwargs):
"""Set the noise level
This method calculates the noise standard deviation using the median
absolute deviation (MAD) of the input data and adds it to the keyword
arguments.
Parameters
----------
data : np.ndarray
Input noisy data (3D array)
Returns
-------
dict Updated keyword arguments
"""
# It the noise is not already provided calculate it using the MAD
if isinstance(kwargs['noise_est'], type(None)):
kwargs['noise_est'] = sigma_mad(data)
print(' - Noise Estimate:', kwargs['noise_est'])
if 'log' in kwargs:
kwargs['log'].info(' - Noise Estimate: ' + str(kwargs['noise_est']))
return kwargs
def sigma_mad_sparse(grad_op, linear_op):
""" Estimate the std from the mad routine on each approximation scale.
Parameters
----------
grad_op: instance
gradient operator.
linear_op: instance
linear operator.
Returns
-------
sigma: list of float
a list of std estimate for each scale.
"""
linear_op.op(grad_op.grad)
return [
sigma_mad(flatten(linear_op.transform[scale])[0])
for scale in range(linear_op.transform.nb_scale)
]
def condatvu(gradient_op,
linear_op,
dual_regularizer,
cost_op,
max_nb_of_iter=150,
tau=None,
sigma=None,
relaxation_factor=1.0,
x_init=None,
std_est=None,
std_est_method=None,
std_thr=2.,
nb_of_reweights=1,
metric_call_period=5,
metrics={},
verbose=0):
""" The Condat-Vu sparse reconstruction with reweightings.
Parameters
----------
gradient_op: instance of class GradBase
the gradient operator.
linear_op: instance of LinearBase
the linear operator: seek the sparsity, ie. a wavelet transform.
dual_regularizer: instance of ProximityParent
the dual regularization operator
cost_op: instance of costObj
the cost function used to check for convergence during the
optimization.
max_nb_of_iter: int, default 150
the maximum number of iterations in the Condat-Vu proximal-dual
splitting algorithm.
tau, sigma: float, default None
parameters of the Condat-Vu proximal-dual splitting algorithm.
If None estimates these parameters.
relaxation_factor: float, default 0.5
parameter of the Condat-Vu proximal-dual splitting algorithm.
If 1, no relaxation.
x_init: np.ndarray (optional, default None)
the initial guess of image
std_est: float, default None
the noise std estimate.
If None use the MAD as a consistent estimator for the std.
std_est_method: str, default None
if the standard deviation is not set, estimate this parameter using
the mad routine in the image ('primal') or in the sparse wavelet
decomposition ('dual') domain.
std_thr: float, default 2.
use this treshold expressed as a number of sigma in the residual
proximity operator during the thresholding.
relaxation_factor: float, default 0.5
parameter of the Condat-Vu proximal-dual splitting algorithm.
If 1, no relaxation.
nb_of_reweights: int, default 1
the number of reweightings.
metric_call_period: int (default 5)
the period on which the metrics are compute.
metrics: dict (optional, default None)
the list of desired convergence metrics: {'metric_name':
[@metric, metric_parameter]}. See modopt for the metrics API.
verbose: int, default 0
the verbosity level.
Returns
-------
x_final: ndarray
the estimated CONDAT-VU solution.
costs: list of float
the cost function values.
metrics: dict
the requested metrics values during the optimization.
y_final: ndarrat
the estimated dual CONDAT-VU solution
"""
# Check inputs
start = time.clock()
if std_est_method not in (None, "primal", "dual"):
raise ValueError(
"Unrecognize std estimation method '{0}'.".format(std_est_method))
# Define the initial primal and dual solutions
if x_init is None:
x_init = np.squeeze(
np.zeros((linear_op.n_coils, *gradient_op.fourier_op.shape),
dtype=np.complex))
primal = x_init
dual = linear_op.op(primal)
weights = dual
# Define the weights used during the thresholding in the dual domain,
# the reweighting strategy, and the prox dual operator
# Case1: estimate the noise std in the image domain
if std_est_method == "primal":
if std_est is None:
std_est = sigma_mad(gradient_op.MtX(gradient_op.obs_data))
weights[...] = std_thr * std_est
reweight_op = cwbReweight(weights)
dual_regularizer.weights = reweight_op.weights
# Case2: estimate the noise std in the sparse wavelet domain
elif std_est_method == "dual":
if std_est is None:
std_est = 0.0
weights[...] = std_thr * std_est
reweight_op = mReweight(weights, linear_op, thresh_factor=std_thr)
dual_regularizer.weights = reweight_op.weights
# Case3: manual regularization mode, no reweighting
else:
reweight_op = None
nb_of_reweights = 0
# Define the Condat Vu optimizer: define the tau and sigma in the
# Condat-Vu proximal-dual splitting algorithm if not already provided.
# Check also that the combination of values will lead to convergence.
norm = linear_op.l2norm(x_init.shape)
lipschitz_cst = gradient_op.spec_rad
if sigma is None:
sigma = 0.5
if tau is None:
# to avoid numerics troubles with the convergence bound
eps = 1.0e-8
# due to the convergence bound
tau = 1.0 / (lipschitz_cst / 2 + sigma * norm**2 + eps)
convergence_test = (1.0 / tau - sigma * norm**2 >= lipschitz_cst / 2.0)
# Welcome message
if verbose > 0:
print(" - mu: ", dual_regularizer.weights)
print(" - lipschitz constant: ", gradient_op.spec_rad)
print(" - tau: ", tau)
print(" - sigma: ", sigma)
print(" - rho: ", relaxation_factor)
print(" - std: ", std_est)
print(" - 1/tau - sigma||L||^2 >= beta/2: ", convergence_test)
print(" - data: ", gradient_op.fourier_op.shape)
if hasattr(linear_op, "nb_scale"):
print(" - wavelet: ", linear_op, "-", linear_op.nb_scale)
print(" - max iterations: ", max_nb_of_iter)
print(" - number of reweights: ", nb_of_reweights)
print(" - primal variable shape: ", primal.shape)
print(" - dual variable shape: ", dual.shape)
print("-" * 40)
prox_op = Identity()
# Define the optimizer
opt = Condat(x=primal,
y=dual,
grad=gradient_op,
prox=prox_op,
prox_dual=dual_regularizer,
linear=linear_op,
cost=cost_op,
rho=relaxation_factor,
sigma=sigma,
tau=tau,
rho_update=None,
sigma_update=None,
tau_update=None,
auto_iterate=False,
metric_call_period=metric_call_period,
metrics=metrics)
cost_op = opt._cost_func
# Perform the first reconstruction
if verbose > 0:
print("Starting optimization...")
opt.iterate(max_iter=max_nb_of_iter)
# Loop through the number of reweightings
for reweight_index in range(nb_of_reweights):
# Generate the new weights following reweighting prescription
if std_est_method == "primal":
reweight_op.reweight(linear_op.op(opt._x_new))
else:
std_est = reweight_op.reweight(opt._x_new)
# Welcome message
if verbose > 0:
print(" - reweight: ", reweight_index + 1)
print(" - std: ", std_est)
# Update the weights in the dual proximity operator
dual_regularizer.weights = reweight_op.weights
# Perform optimisation with new weights
opt.iterate(max_iter=max_nb_of_iter)
# Goodbye message
end = time.clock()
if verbose > 0:
if hasattr(cost_op, "cost"):
print(" - final iteration number: ", cost_op._iteration)
print(" - final cost value: ", cost_op.cost)
print(" - converged: ", opt.converge)
print("Done.")
print("Execution time: ", end - start, " seconds")
print("-" * 40)
# Get the final solution
x_final = opt.x_final
y_final = opt.y_final
if hasattr(cost_op, "cost"):
costs = cost_op._cost_list
else:
costs = None
return x_final, costs, opt.metrics, y_final
def sparse_rec_condatvu(gradient_op,
linear_op,
std_est=None,
std_est_method=None,
std_thr=2.,
mu=1e-6,
tau=None,
sigma=None,
relaxation_factor=1.0,
nb_of_reweights=1,
max_nb_of_iter=150,
add_positivity=False,
atol=1e-4,
verbose=0):
""" The Condat-Vu sparse reconstruction with reweightings.
.. note:: At the moment, supports only 2D data.
Parameters
----------
data: ndarray
the data to reconstruct: observation are expected in Fourier space.
wavelet_name: str
the wavelet name to be used during the decomposition.
samples: np.ndarray
the mask samples in the Fourier domain.
nb_scales: int, default 4
the number of scales in the wavelet decomposition.
std_est: float, default None
the noise std estimate.
If None use the MAD as a consistent estimator for the std.
std_est_method: str, default None
if the standard deviation is not set, estimate this parameter using
the mad routine in the image ('primal') or in the sparse wavelet
decomposition ('dual') domain.
std_thr: float, default 2.
use this treshold expressed as a number of sigma in the residual
proximity operator during the thresholding.
mu: float, default 1e-6
regularization hyperparameter.
tau, sigma: float, default None
parameters of the Condat-Vu proximal-dual splitting algorithm.
If None estimates these parameters.
relaxation_factor: float, default 0.5
parameter of the Condat-Vu proximal-dual splitting algorithm.
If 1, no relaxation.
nb_of_reweights: int, default 1
the number of reweightings.
max_nb_of_iter: int, default 150
the maximum number of iterations in the Condat-Vu proximal-dual
splitting algorithm.
add_positivity: bool, default False
by setting this option, set the proximity operator to identity or
positive.
atol: float, default 1e-4
tolerance threshold for convergence.
verbose: int, default 0
the verbosity level.
Returns
-------
x_final: ndarray
the estimated CONDAT-VU solution.
transform: a WaveletTransformBase derived instance
the wavelet transformation instance.
"""
# Check inputs
# analysis = True
# if hasattr(gradient_op, 'linear_op'):
# analysis = False
start = time.clock()
if std_est_method not in (None, "primal", "dual"):
raise ValueError(
"Unrecognize std estimation method '{0}'.".format(std_est_method))
# Define the initial primal and dual solutions
x_init = np.zeros(gradient_op.fourier_op.shape, dtype=np.complex)
weights = linear_op.op(x_init)
# Define the weights used during the thresholding in the dual domain,
# the reweighting strategy, and the prox dual operator
# Case1: estimate the noise std in the image domain
if std_est_method == "primal":
if std_est is None:
std_est = sigma_mad(gradient_op.MtX(gradient_op.y))
weights[...] = std_thr * std_est
reweight_op = cwbReweight(weights)
prox_dual_op = Threshold(reweight_op.weights)
# Case2: estimate the noise std in the sparse wavelet domain
elif std_est_method == "dual":
if std_est is None:
std_est = 0.0
weights[...] = std_thr * std_est
reweight_op = mReweight(weights, linear_op, thresh_factor=std_thr)
prox_dual_op = Threshold(reweight_op.weights)
# Case3: manual regularization mode, no reweighting
else:
weights[...] = mu
reweight_op = None
prox_dual_op = Threshold(weights)
nb_of_reweights = 0
# Define the Condat Vu optimizer: define the tau and sigma in the
# Condat-Vu proximal-dual splitting algorithm if not already provided.
# Check also that the combination of values will lead to convergence.
norm = linear_op.l2norm(gradient_op.fourier_op.shape)
lipschitz_cst = gradient_op.spec_rad
if sigma is None:
sigma = 0.5
if tau is None:
# to avoid numerics troubles with the convergence bound
eps = 1.0e-8
# due to the convergence bound
tau = 1.0 / (lipschitz_cst / 2 + sigma * norm**2 + eps)
convergence_test = (1.0 / tau - sigma * norm**2 >= lipschitz_cst / 2.0)
# Define initial primal and dual solutions
primal = np.zeros(gradient_op.fourier_op.shape, dtype=np.complex)
dual = linear_op.op(primal)
dual[...] = 0.0
# Welcome message
if verbose > 0:
print(condatvu_logo())
print(" - mu: ", mu)
print(" - lipschitz constant: ", gradient_op.spec_rad)
print(" - tau: ", tau)
print(" - sigma: ", sigma)
print(" - rho: ", relaxation_factor)
print(" - std: ", std_est)
print(" - 1/tau - sigma||L||^2 >= beta/2: ", convergence_test)
print(" - data: ", gradient_op.obs_data.shape)
print(" - max iterations: ", max_nb_of_iter)
print(" - number of reweights: ", nb_of_reweights)
print(" - primal variable shape: ", primal.shape)
print(" - dual variable shape: ", dual.shape)
print("-" * 40)
# Define the proximity operator
if add_positivity:
prox_op = Positivity()
else:
prox_op = Identity()
# Define the cost function
cost_op = DualGapCost(linear_op=linear_op,
initial_cost=1e6,
tolerance=1e-4,
cost_interval=1,
test_range=4,
verbose=0,
plot_output=None)
# Define the optimizer
opt = Condat(x=primal,
y=dual,
grad=gradient_op,
prox=prox_op,
prox_dual=prox_dual_op,
linear=linear_op,
cost=cost_op,
rho=relaxation_factor,
sigma=sigma,
tau=tau,
rho_update=None,
sigma_update=None,
tau_update=None,
auto_iterate=False)
# Perform the first reconstruction
if verbose > 0:
print("Starting optimization...")
for i in range(max_nb_of_iter):
opt._update()
opt.x_final = opt._x_new
opt.y_final = opt._y_new
# Loop through the number of reweightings
for reweight_index in range(nb_of_reweights):
# Generate the new weights following reweighting prescription
if std_est_method == "primal":
reweight_op.reweight(linear_op.op(opt._x_new))
else:
std_est = reweight_op.reweight(opt._x_new)
# Welcome message
if verbose > 0:
print(" - reweight: ", reweight_index + 1)
print(" - std: ", std_est)
# Update the weights in the dual proximity operator
prox_dual_op.weights = reweight_op.weights
# Perform optimisation with new weights
opt.iterate(max_iter=max_nb_of_iter)
# Goodbye message
end = time.clock()
if verbose > 0:
print(" - final iteration number: ", cost_op._iteration)
print(" - final cost value: ", cost_op.cost)
print(" - converged: ", opt.converge)
print("Done.")
print("Execution time: ", end - start, " seconds")
print("-" * 40)
# Get the final solution
x_final = opt.x_final
linear_op.transform.analysis_data = unflatten(opt.y_final,
linear_op.coeffs_shape)
return x_final, linear_op.transform
