skglm.solvers.AndersonCD

class skglm.solvers.AndersonCD(max_iter=50, max_epochs=50000, p0=10, tol=0.0001, ws_strategy='subdiff', fit_intercept=True, warm_start=False, verbose=0)

Coordinate descent solver with working sets and Anderson acceleration.

fit_interceptbool

Whether or not to fit an intercept.

max_iterint, optional

The maximum number of iterations (definition of working set and resolution of problem restricted to features in working set).

max_epochsint, optional

Maximum number of (block) CD epochs on each subproblem.

p0int, optional

First working set size.

tolfloat, optional

The tolerance for the optimization.

ws_strategy(‘subdiff’|’fixpoint’), optional

The score used to build the working set.

verbosebool or int, optional

Amount of verbosity. 0/False is silent.

References

[1]

Bertrand, Q. and Klopfenstein, Q. and Bannier, P.-A. and Gidel, G. and Massias, M. “Beyond L1: Faster and Better Sparse Models with skglm”, 2022 https://arxiv.org/abs/2204.07826

[2]

Bertrand, Q. and Massias, M. “Anderson acceleration of coordinate descent”, AISTATS, 2021 https://proceedings.mlr.press/v130/bertrand21a.html code: https://github.com/mathurinm/andersoncd

__init__(max_iter=50, max_epochs=50000, p0=10, tol=0.0001, ws_strategy='subdiff', fit_intercept=True, warm_start=False, verbose=0)

Methods

__init__([max_iter, max_epochs, p0, tol, ...])
custom_checks(X, y, datafit, penalty)	Ensure the solver is suited for the datafit + penalty problem.
path(X, y, datafit, penalty[, alphas, ...])
solve(X, y, datafit, penalty[, w_init, ...])	Solve the optimization problem after validating its compatibility.