ridgesketch should not be listed as an option for solvers

facebookresearch / RidgeSketch

A Fast sketching based solver for large scale ridge regression

Other

17 stars 5 forks source link

ridgesketch should not be listed as an option for solvers #1

Open Sciroccogti opened 9 months ago

Sciroccogti commented 9 months ago

In the comment of class RidgeSketch, "ridgesketch" is listed as an option, but in fact it is not implemented.

solver (str): defaults to 'auto'.
    If dimensions small, uses direct solver.
    Otherwise, randomized solver.
    Options are "direct" or "ridgesketch" or "conjugate gradients".

Currently available solver seems to be:

    SKLEARN_SOLVERS = {
        "svd",
        "cholesky",
        "lsqr",
        "sag",
        "saga",
    }

    SKETCH_SOLVERS = {
        "direct",
        "cg",
        "subsample",
        "coordinate descent",
        "gaussian",
        "count",
        "subcount",
        "hadamard",
    }

ngazagna commented 9 months ago

Sorry for late reply @Sciroccogti , you are right the available options are

direct for using a standard direct solver defined here https://github.com/facebookresearch/RidgeSketch/blob/1073b4da0423b28093ab9dd5ac41939b187b8660/ridge_sketch.py#L341
cg for the conjugate gradient method defined here https://github.com/facebookresearch/RidgeSketch/blob/1073b4da0423b28093ab9dd5ac41939b187b8660/ridge_sketch.py#L575
and the other options "subsample", "coordinate descent", "gaussian", "count", "subcount", "hadamard" which are referring to the name of the sketch method used in the ridgesketch algorithm

I will clarify the documentation.