FixedPoint
- class deepinv.optim.FixedPoint(iterator=None, update_params_fn=None, update_data_fidelity_fn=None, update_prior_fn=None, init_iterate_fn=None, init_metrics_fn=None, update_metrics_fn=None, check_iteration_fn=None, check_conv_fn=None, max_iter=50, early_stop=True, anderson_acceleration=False, history_size=5, beta_anderson_acc=1.0, eps_anderson_acc=0.0001, verbose=False, show_progress_bar=False)[source]
Bases:
Module
Fixed-point iterations module.
This module implements the fixed-point iteration algorithm given a specific fixed-point iterator (e.g. proximal gradient iteration, the ADMM iteration, see
deepinv.optim.optim_iterators()
), that is for \(k=1,2,...\)\[\qquad (x_{k+1}, u_{k+1}) = \operatorname{FixedPoint}(x_k, u_k, f, g, A, y, ...) \hspace{2cm} (1)\]where \(f\) is the data-fidelity term, \(g\) is the prior, \(A\) is the physics model, \(y\) is the data.
- Examples:
This example shows how to use the
FixedPoint
class to solve the problem \(\min_x 0.5*||Ax-y||_2^2 + \lambda*||x||_1\) with the PGD algorithm, where A is the identity operator, \(\lambda = 1\) and \(y = [2, 2]\).>>> import deepinv as dinv >>> # Create the measurement operator A >>> A = torch.tensor([[1, 0], [0, 1]], dtype=torch.float64) >>> A_forward = lambda v: A @ v >>> A_adjoint = lambda v: A.transpose(0, 1) @ v >>> # Define the physics model associated to this operator >>> physics = dinv.physics.LinearPhysics(A=A_forward, A_adjoint=A_adjoint) >>> # Define the measurement y >>> y = torch.tensor([2, 2], dtype=torch.float64) >>> # Define the data fidelity term >>> data_fidelity = dinv.optim.data_fidelity.L2() >>> # Define the prior term >>> prior = dinv.optim.prior.L1Prior() >>> # Define the parameters of the algorithm >>> params_algo = {"g_param": 1.0, "stepsize": 1.0, "lambda": 1.0, "beta": 1.0} >>> # Choose the iterator associated to the PGD algorithm >>> iterator = dinv.optim.optim_iterators.PGDIteration() >>> # Iterate the iterator >>> x_init = torch.tensor([2, 2], dtype=torch.float64) # Define initialisation of the algorithm >>> X = {"est": (x_init ,), "cost": []} # Iterates are stored in a dictionary of the form {'est': (x,z), 'cost': F} >>> max_iter = 50 >>> for it in range(max_iter): ... X = iterator(X, data_fidelity, prior, params_algo, y, physics) >>> # Return the solution >>> X["est"][0] tensor([1., 1.], dtype=torch.float64)
- Parameters:
iterator (deepinv.optim.optim_iterators.optim_iterator) – function that takes as input the current iterate, as well as parameters of the optimization problem (prior, measurements, etc.)
update_params_fn (function) – function that returns the parameters to be used at each iteration. Default:
None
.update_prior_fn (function) – function that returns the prior to be used at each iteration. Default:
None
.init_iterate_fn (function) – function that returns the initial iterate. Default:
None
.init_metrics_fn (function) – function that returns the initial metrics. Default:
None
.check_iteration_fn (function) – function that performs a check on the last iteration and returns a bool indicating if we can proceed to next iteration. Default:
None
.check_conv_fn (function) – function that checks the convergence after each iteration, returns a bool indicating if convergence has been reached. Default:
None
.max_iter (int) – maximum number of iterations. Default:
50
.early_stop (bool) – if True, the algorithm stops when the convergence criterion is reached. Default:
True
.anderson_acceleration (bool) – if True, the Anderson acceleration is used. Default:
False
.history_size (int) – size of the history used for the Anderson acceleration. Default:
5
.beta_anderson_acc (float) – momentum of the Anderson acceleration step. Default:
1.0
.eps_anderson_acc (float) – regularization parameter of the Anderson acceleration step. Default:
1e-4
.
- anderson_acceleration_step(it, X_prev, TX_prev, x_hist, T_hist, H, q, cur_data_fidelity, cur_prior, cur_params, *args)[source]
Anderson acceleration step.
- Parameters:
it (int) – current iteration.
X_prev (dict) – previous iterate.
TX_prev (dict) – output of the fixed-point operator evaluated at X_prev
x_hist (torch.Tensor) – history of last
history-size
iterates.T_hist (torch.Tensor) – history of T evlauaton at the last
history-size
, where T is the fixed-point operator.H (torch.Tensor) – H in the Anderson acceleration linear system Hp = q .
q (torch.Tensor) – q in the Anderson acceleration linear system Hp = q .
cur_data_fidelity (deepinv.optim.DataFidelity) – Instance of the DataFidelity class defining the current data_fidelity.
cur_prior (deepinv.optim.prior) – Instance of the Prior class defining the current prior.
cur_params (dict) – Dictionary containing the current parameters of the algorithm.
args – arguments for the iterator.
- forward(*args, compute_metrics=False, x_gt=None, **kwargs)[source]
Loops over the fixed-point iterator as (1) and returns the fixed point.
The iterates are stored in a dictionary of the form
X = {'est': (x_k, u_k), 'cost': F_k}
where:est
is a tuple containing the current primal and auxiliary iterates,cost
is the value of the cost function at the current iterate.
Since the prior and parameters (stepsize, regularisation parameter, etc.) can change at each iteration, the prior and parameters are updated before each call to the iterator.
- Parameters:
compute_metrics (bool) – if
True
, the metrics are computed along the iterations. Default:False
.x_gt (torch.Tensor) – ground truth solution. Default:
None
.args – optional arguments for the iterator. Commonly (y,physics) where
y
(torch.Tensor y) is the measurement andphysics
(deepinv.physics) is the physics model.kwargs – optional keyword arguments for the iterator.
- Return tuple:
(x,metrics)
withx
the fixed-point solution (dict) andmetrics
the computed along the iterations ifcompute_metrics
isTrue
orNone
otherwise.