shrinkage.HybridCoordinateDescent

Shrinkage for Python · python · 0.2.0

HybridCoordinateDescent(
    *,
    stopping: StoppingCriterion | None = None,
    max_iterations: int = 10000,
    max_outer_iterations: int = 100,
    max_inner_iterations: int = 1000,
    max_outer_backtracks: int = 100,
    cluster_sweeps: int = 5,
    check_every: int = 10,
    screening: bool = True,
    working_set_policy: shrinkage.WorkingSetPolicy = <factory>,
    step_policy: shrinkage.HybridStep = <factory>
)

Coordinate descent with periodic proximal-gradient steps and binomial IRLS.

Start with a PG transition, then alternate cluster_sweeps block sweeps with PG. max_iterations counts committed PG transitions plus sweeps; logistic fits share it across quadratic subproblems. Outer iterations count accepted IRLS steps. Inner iterations limit each quadratic model. Gaussian fitting ignores the outer and inner budgets.

Checks run initially, every check_every iterations, and at the budget. Only the full original-loss mapping certifies convergence. stopping=None uses absolute mapping tolerance 1e-6. Gap and KKT criteria are rejected. Screening and working sets are enabled independently by default.

stopping: StoppingCriterion | None = field(default=None, kw_only=True)

max_iterations: int = field(default=10_000, kw_only=True)

max_outer_iterations: int = field(default=100, kw_only=True)

max_inner_iterations: int = field(default=1_000, kw_only=True)

max_outer_backtracks: int = field(default=100, kw_only=True)

cluster_sweeps: int = field(default=5, kw_only=True)

check_every: int = field(default=10, kw_only=True)

screening: bool = field(default=True, kw_only=True)

working_set_policy: shrinkage.WorkingSetPolicy = field(
        default_factory=WorkingSetPolicy.auto, kw_only=True
    )

step_policy: shrinkage.HybridStep = field(
        default_factory=HybridStep.backtracking, kw_only=True
    )