Support init arguments in MNMG LogisticRegression (#5519)

The init arguments are for LBFGS (the only algorithm in the current MNMG LogisticRegression).  

The key code changes should be a few lines after [PR 5516 for predict](#5516) gets merged. Key code changes can be reviewed from [here](https://github.com/rapidsai/cuml/pull/5519/files/d058d884c992661984224d0190c3bbcc0a23caf4..fbbaa5c6aef47ddc7100f5bea2a751851ca6d1b4)

Authors:
  - Jinfeng Li (https://github.com/lijinf2)

Approvers:
  - Corey J. Nolet (https://github.com/cjnolet)

URL: #5519

python/cuml/dask/linear_model/logistic_regression.py

@@ -33,8 +33,112 @@
 
 
 class LogisticRegression(LinearRegression):
-    def __init__(self, *, client=None, verbose=False, **kwargs):
-        super().__init__(client=client, verbose=verbose, **kwargs)
+    """
+    LogisticRegression is a linear model that is used to model probability of
+    occurrence of certain events, for example probability of success or fail of
+    an event.
+
+    cuML's dask Logistic Regression (multi-node multi-gpu) expects dask cuDF
+    DataFrame and provides an algorithms, L-BFGS, to fit the logistic model. It
+    currently supports single class, l2 regularization, and sigmoid loss.
+
+    Note that, just like in Scikit-learn, the bias will not be regularized.
+
+    Examples
+    --------
+
+    .. code-block:: python
+
+        >>> from dask_cuda import LocalCUDACluster
+        >>> from dask.distributed import Client
+        >>> import dask_cudf
+        >>> import cudf
+        >>> import numpy as np
+        >>> from cuml.dask.linear_model import LogisticRegression
+
+        >>> cluster = LocalCUDACluster(CUDA_VISIBLE_DEVICES="0,1")
+        >>> client = Client(cluster)
+
+        >>> X = cudf.DataFrame()
+        >>> X['col1'] = np.array([1,1,2,2], dtype = np.float32)
+        >>> X['col2'] = np.array([1,2,2,3], dtype = np.float32)
+        >>> y = cudf.Series(np.array([0.0, 0.0, 1.0, 1.0], dtype=np.float32))
+
+        >>> X_ddf = dask_cudf.from_cudf(X, npartitions=2)
+        >>> y_ddf = dask_cudf.from_cudf(y, npartitions=2)
+
+        >>> reg = LogisticRegression()
+        >>> reg.fit(X_ddf, y_ddf)
+        LogisticRegression()
+
+        >>> print(reg.coef_)
+                 0         1
+        0  0.69861  0.570058
+        >>> print(reg.intercept_)
+        0   -2.188068
+        dtype: float32
+
+        >>> X_new = cudf.DataFrame()
+        >>> X_new['col1'] = np.array([1,5], dtype = np.float32)
+        >>> X_new['col2'] = np.array([2,5], dtype = np.float32)
+        >>> X_new_ddf = dask_cudf.from_cudf(X_new, npartitions=2)
+        >>> preds = reg.predict(X_new_ddf)
+
+        >>> print(preds.compute())
+        0    0.0
+        1    1.0
+        dtype: float32
+
+    Parameters
+    ----------
+    tol : float (default = 1e-4)
+        Tolerance for stopping criteria.
+        The exact stopping conditions depend on the L-BFGS solver.
+        Check the solver's documentation for more details:
+
+          * :class:`Quasi-Newton (L-BFGS)<cuml.QN>`
+
+    C : float (default = 1.0)
+        Inverse of regularization strength; must be a positive float.
+    fit_intercept : boolean (default = True)
+        If True, the model tries to correct for the global mean of y.
+        If False, the model expects that you have centered the data.
+    max_iter : int (default = 1000)
+        Maximum number of iterations taken for the solvers to converge.
+    linesearch_max_iter : int (default = 50)
+        Max number of linesearch iterations per outer iteration used in the
+        solver.
+    verbose : int or boolean, default=False
+        Sets logging level. It must be one of `cuml.common.logger.level_*`.
+        See :ref:`verbosity-levels` for more info.
+    output_type : {'input', 'array', 'dataframe', 'series', 'df_obj', \
+        'numba', 'cupy', 'numpy', 'cudf', 'pandas'}, default=None
+        Return results and set estimator attributes to the indicated output
+        type. If None, the output type set at the module level
+        (`cuml.global_settings.output_type`) will be used. See
+        :ref:`output-data-type-configuration` for more info.
+
+    Attributes
+    ----------
+    coef_: dev array, dim (n_classes, n_features) or (n_classes, n_features+1)
+        The estimated coefficients for the linear regression model.
+    intercept_: device array (n_classes, 1)
+        The independent term. If `fit_intercept` is False, will be 0.
+
+    Notes
+    -----
+    cuML's LogisticRegression uses a different solver that the equivalent
+    Scikit-learn, except when there is no penalty and `solver=lbfgs` is
+    used in Scikit-learn. This can cause (smaller) differences in the
+    coefficients and predictions of the model, similar to
+    using different solvers in Scikit-learn.
+
+    For additional information, see `Scikit-learn's LogisticRegression
+    <https://scikit-learn.org/stable/modules/generated/sklearn.linear_model.LogisticRegression.html>`_.
+    """
+
+    def __init__(self, **kwargs):
+        super().__init__(**kwargs)
 
     def fit(self, X, y):
         """
@@ -64,7 +168,7 @@ def _create_model(sessionId, datatype, **kwargs):
         )
 
         handle = get_raft_comm_state(sessionId, get_worker())["handle"]
-        return LogisticRegressionMG(handle=handle)
+        return LogisticRegressionMG(handle=handle, **kwargs)
 
     @staticmethod
     def _func_fit(f, data, n_rows, n_cols, partsToSizes, rank):

python/cuml/linear_model/logistic_regression.pyx

@@ -170,7 +170,7 @@ class LogisticRegression(UniversalBase,
     Attributes
     ----------
     coef_: dev array, dim (n_classes, n_features) or (n_classes, n_features+1)
-        The estimated coefficients for the linear regression model.
+        The estimated coefficients for the logistic regression model.
     intercept_: device array (n_classes, 1)
         The independent term. If `fit_intercept` is False, will be 0.
 

python/cuml/linear_model/logistic_regression_mg.pyx

@@ -82,8 +82,8 @@ cdef extern from "cuml/linear_model/qn_mg.hpp" namespace "ML::GLM::opg" nogil:
 
 class LogisticRegressionMG(MGFitMixin, LogisticRegression):
 
-    def __init__(self, *, handle=None):
-        super().__init__(handle=handle)
+    def __init__(self, **kwargs):
+        super(LogisticRegressionMG, self).__init__(**kwargs)
 
     @property
     @cuml.internals.api_base_return_array_skipall

python/cuml/tests/dask/test_dask_logistic_regression.py

@@ -177,13 +177,90 @@ def imp():
     assert_array_equal(preds, y, strict=True)
 
 
+def test_lbfgs_init(client):
+    def imp():
+        import cuml.comm.serialize  # NOQA
+
+    client.run(imp)
+
+    X = np.array([(1, 2), (1, 3), (2, 1), (3, 1)], dtype=np.float32)
+    y = np.array([1.0, 1.0, 0.0, 0.0], dtype=np.float32)
+
+    X_df, y_df = _prep_training_data(
+        c=client, X_train=X, y_train=y, partitions_per_worker=2
+    )
+
+    from cuml.dask.linear_model.logistic_regression import (
+        LogisticRegression as cumlLBFGS_dask,
+    )
+
+    def assert_params(
+        tol,
+        C,
+        fit_intercept,
+        max_iter,
+        linesearch_max_iter,
+        verbose,
+        output_type,
+    ):
+
+        lr = cumlLBFGS_dask(
+            tol=tol,
+            C=C,
+            fit_intercept=fit_intercept,
+            max_iter=max_iter,
+            linesearch_max_iter=linesearch_max_iter,
+            verbose=verbose,
+            output_type=output_type,
+        )
+
+        lr.fit(X_df, y_df)
+        qnpams = lr.qnparams.params
+        assert qnpams["grad_tol"] == tol
+        assert qnpams["loss"] == 0  # "sigmoid" loss
+        assert qnpams["penalty_l1"] == 0.0
+        assert qnpams["penalty_l2"] == 1.0 / C
+        assert qnpams["fit_intercept"] == fit_intercept
+        assert qnpams["max_iter"] == max_iter
+        assert qnpams["linesearch_max_iter"] == linesearch_max_iter
+        assert (
+            qnpams["verbose"] == 5 if verbose is True else 4
+        )  # cuml Verbosity Levels
+        assert (
+            lr.output_type == "input" if output_type is None else output_type
+        )  # cuml.global_settings.output_type
+
+    assert_params(
+        tol=1e-4,
+        C=1.0,
+        fit_intercept=True,
+        max_iter=1000,
+        linesearch_max_iter=50,
+        verbose=False,
+        output_type=None,
+    )
+
+    assert_params(
+        tol=1e-6,
+        C=1.5,
+        fit_intercept=False,
+        max_iter=200,
+        linesearch_max_iter=100,
+        verbose=True,
+        output_type="cudf",
+    )
+
+
 @pytest.mark.mg
 @pytest.mark.parametrize("nrows", [1e5])
 @pytest.mark.parametrize("ncols", [20])
 @pytest.mark.parametrize("n_parts", [2, 23])
+@pytest.mark.parametrize("fit_intercept", [False, True])
 @pytest.mark.parametrize("datatype", [np.float32])
 @pytest.mark.parametrize("delayed", [True, False])
-def test_lbfgs(nrows, ncols, n_parts, datatype, delayed, client):
+def test_lbfgs(
+    nrows, ncols, n_parts, fit_intercept, datatype, delayed, client
+):
     tolerance = 0.005
 
     def imp():
@@ -203,12 +280,12 @@ def imp():
 
     X_df, y_df = _prep_training_data(client, X, y, n_parts)
 
-    lr = cumlLBFGS_dask()
+    lr = cumlLBFGS_dask(fit_intercept=fit_intercept)
     lr.fit(X_df, y_df)
     lr_coef = lr.coef_.to_numpy()
     lr_intercept = lr.intercept_.to_numpy()
 
-    sk_model = skLR()
+    sk_model = skLR(fit_intercept=fit_intercept)
     sk_model.fit(X, y)
     sk_coef = sk_model.coef_
     sk_intercept = sk_model.intercept_