bartorch.optim.ADMM

bartorch.optim.ADMM#

class bartorch.optim.ADMM(regularizers=None, *, maxiter=30, rho=0.5, cg_maxiter=10, hogwild=False, cclambda=0.0, biases=None, dynamic_rho=False, dynamic_tau=False, relative_norm=False, fast=False, alpha=1.6, mu=3.0, tau_max=20.0, abstol=0.0, reltol=0.0, cg_maxiter_first=None, precond=None)#

Alternating direction method of multipliers (pics --admm).

Parameters:

  • regularizers (Regularizer or iterable of Regularizer, optional) – Terms from bartorch.prox.

  • maxiter (int)

  • rho (float) – Penalty parameter (pics -u); 0.5 is BART’s default.

  • cg_maxiter (int) – Conjugate-gradient iterations per step (pics -C); 10 is BART’s default.

  • hogwild (bool) – BART’s hogwild setting (pics -H), which doubles rho after ten steps, then twenty, then forty. Not combinable with dynamic_rho, which BART asserts against.

  • cclambda (float) – Weight of an identity added to the normal operator (pics -q).

  • biases (sequence of tensor, optional) – The b_j of f_j(G_j x - b_j), one per term, each of its term’s transformed shape.

  • dynamic_rho (bool) – Move rho with the residuals (pics --admm_dynamic_rho): up by tau when the primal residual leads, down when the dual does. The dual variables are rescaled to match, so the split stays where it was.

  • dynamic_tau (bool) – Choose tau from the residuals too (pics --admm_dynamic_tau), as sqrt(r / s) clipped to [1 / tau_max, tau_max]. Together with dynamic_rho and relative_norm this is the residual balancing of Wohlberg (2017).

  • relative_norm (bool) – Compare the residuals to their scalings rather than to each other (pics --admm_relative_norm).

  • fast (bool) – Skip the residuals entirely, and with them the stopping test.

  • alpha (float) – Over-relaxation; BART’s default of 1.6 is what pics runs. Out of reach of in_library(), which italgo_config gives no way to set.

  • mu (float) – How far the residuals must part before dynamic_rho moves rho. Out of reach of in_library().

  • tau_max (float) – The clip on tau. Out of reach of in_library().

  • abstol (float) – Boyd’s absolute and relative tolerances, which stop the iteration when both residuals are inside them. italgo_config sets both to zero, whatever iter_admm_defaults says, so the budget is what stops pics; these are out of reach of in_library().

  • reltol (float) – Boyd’s absolute and relative tolerances, which stop the iteration when both residuals are inside them. italgo_config sets both to zero, whatever iter_admm_defaults says, so the budget is what stops pics; these are out of reach of in_library().

  • cg_maxiter_first (int, optional) – A separate budget for the first step’s inner solve. Not BART’s – it is riesling’s iters0, the one thing that implementation has that BART’s does not – so it is out of reach of in_library() as well.

__init__(regularizers=None, *, maxiter=30, rho=0.5, cg_maxiter=10, hogwild=False, cclambda=0.0, biases=None, dynamic_rho=False, dynamic_tau=False, relative_norm=False, fast=False, alpha=1.6, mu=3.0, tau_max=20.0, abstol=0.0, reltol=0.0, cg_maxiter_first=None, precond=None)#

Methods

__init__([regularizers, maxiter, rho, ...])

fixed_point(image_shape, *[, trainable])

Refused: the fixed point of an alternating-direction step is not in x.

in_library(y, A[, x0])

BART's own loop, which cannot be given everything this solver takes.

unrolled(image_shape, *[, trainable])

This solver as a network of maxiter steps, trained end to end.