LinearPhysicsMultiScaler#

class deepinv.physics.LinearPhysicsMultiScaler(physics, img_shape, filter='sinc', factors=(2, 4, 8), device='cpu', **kwargs)[source]#

Bases: PhysicsMultiScaler, LinearPhysics

Multi-scale wrapper for linear physics operators.

See PhysicsMultiScaler for details.

Examples:

A multiscale BlurFFT operator can be created as follows:

>>> import torch
>>> import deepinv as dinv
>>> physics = dinv.physics.BlurFFT(img_size=(1, 32, 32), filter=dinv.physics.blur.gaussian_blur(.2))
>>> x = torch.rand((1, 1, 8, 8))  # define an image 4 times smaller than the physics input size (scale = 2)
>>> new_physics = dinv.physics.LinearPhysicsMultiScaler(physics, (1, 32, 32), factors=[2, 4, 8])  # define a multiscale physics with base img size (1, 32, 32)
>>> y = new_physics(x, scale=2)  # applying physics at scale 2
>>> print(y.shape)
torch.Size([1, 1, 32, 32])
Parameters:

  • physics (deepinv.physics.Physics) – base physics operator.

  • img_shape (tuple) – shape of the input image (C, H, W).

  • filter (str) – type of filter to use for upsampling, e.g., ‘sinc’, ‘nearest’, ‘bilinear’.

  • factors (list[int]) – list of factors to use for upsampling.

  • device (str, torch.device, str) – device to use for the upsampling operator, e.g., ‘cpu’, ‘cuda’.

A_dagger(y, scale=None, **kwargs)[source]#

Computes the pseudo-inverse of the linear operator \(A\).

If the scale is set to 0, it uses the base physics pseudo-inverse, which might have a more efficient implementation.

Parameters:

y (torch.Tensor) – measurements tensor

Returns:

(torch.Tensor) estimated signal tensor

prox_l2(z, y, gamma, solver='CG', max_iter=None, tol=None, verbose=False, scale=None, **kwargs)[source]#

Computes proximal operator of \(f(x) = \frac{1}{2}\|Ax-y\|^2\), i.e.,

\[\underset{x}{\arg\min} \; \frac{\gamma}{2}\|Ax-y\|^2 + \frac{1}{2}\|x-z\|^2\]

If the scale is set to 0, it uses the base physics proximal operator, which might have a more efficient implementation.

Parameters:

  • y (torch.Tensor) – measurements tensor

  • z (torch.Tensor) – signal tensor

  • gamma (float) – hyperparameter of the proximal operator

  • solver (str) – solver to use for the proximal operator, see deepinv.optim.utils.least_squares() for details

  • max_iter (int) – maximum number of iterations for iterative solvers

  • tol (float) – tolerance for iterative solvers

  • verbose (bool) – whether to print information during the solver execution

  • scale (int) – scale at which to apply the physics operator

Returns:

(torch.Tensor) estimated signal tensor