TVL1Prior#

class deepinv.optim.TVL1Prior(def_crit=1e-8, n_it_max=1000, *args, **kwargs)[source]#

Bases: TVPrior

Total Variation (TV) prior with an L1 norm.

This prior computes the isotropic total variation regularization term.

The prior is defined as:

\[\mathrm{TV}(x) = \sum_i \|\nabla x_i\|_1\]

where \(\nabla x\) denotes the discrete gradient of x.

Parameters:

  • def_crit (float) – default convergence criterion for the inner solver of the TV denoiser; default value: 1e-8.

  • n_it_max (int) – maximal number of iterations for the inner solver of the TV denoiser; default value: 1000.

fn(x, *args, **kwargs)[source]#

Computes the regularizer

\[\reg{x} = \|Dx\|_{1}\]

where D is the finite differences linear operator.

Parameters:

x (torch.Tensor) – Variable \(x\) at which the prior is computed.

Returns:

(torch.Tensor) prior \(g(x)\).

Return type:

Tensor