BSREMIteration

class deepinv.optim.optim_iterators.BSREMIteration(eps=1e-6, sensitivity_threshold=1e-2, cost_fn=None, **kwargs)

Bases: OptimIterator

Performs a single BSREM epoch, updating the estimate once per measurement subset. See deepinv.optim.BSREM for algorithm details.

Parameters:

• eps (float) – Lower bound for division denominators and the reconstructed image. Default: 1e-6.

• sensitivity_threshold (float) – Sensitivity threshold defining the reconstruction support. Default: 1e-2.

• cost_fn (Callable) – Custom cost function evaluated after each epoch. Default: None.

forward(X, cur_data_fidelity, cur_prior, cur_params, y, physics, sensitivities, *args, **kwargs)

Perform one Block Sequential Regularized EM epoch.

Parameters:

• X (dict) – Dictionary containing the current iterate and estimated cost.

• cur_data_fidelity (deepinv.optim.StackedPhysicsDataFidelity) – Data-fidelity terms corresponding to the physics subsets.

• cur_prior (deepinv.optim.Prior) – Differentiable prior used for each subset update.

• cur_params (dict) – Algorithm parameters "stepsize", "lambda", and "g_param".

• y (deepinv.utils.TensorList) – Measurement subsets.

• physics (deepinv.physics.StackedLinearPhysics) – Physics operators corresponding to the measurement subsets.

• sensitivities (list[torch.Tensor]) – Precomputed sensitivity maps \(A_l^T\mathbf{1}\) for each subset.

Returns:

Dictionary {"est": (x, None), "cost": F, "it": k + 1} containing the updated iterate and estimated cost.

Return type:

dict[str, tuple[torch.Tensor, None] | torch.Tensor | int | None]