def test_big(fit_intercept, is_sparse):
with dask.config.set(scheduler='synchronous'):
X, y = make_classification(is_sparse=is_sparse)
lr = LogisticRegression(fit_intercept=fit_intercept)
lr.fit(X, y)
lr.predict(X)
lr.predict_proba(X)
if fit_intercept:
assert lr.intercept_ is not None
def test_fit(fit_intercept, is_sparse):
X, y = make_classification(n_samples=100,
n_features=5,
chunksize=10,
is_sparse=is_sparse)
lr = LogisticRegression(fit_intercept=fit_intercept)
lr.fit(X, y)
lr.predict(X)
lr.predict_proba(X)
def test_big(fit_intercept):
import dask
dask.set_options(get=dask.get)
X, y = make_classification()
lr = LogisticRegression(fit_intercept=fit_intercept)
lr.fit(X, y)
lr.predict(X)
lr.predict_proba(X)
if fit_intercept:
assert lr.intercept_ is not None
def test_gridsearch():
from sklearn.pipeline import make_pipeline
dcv = pytest.importorskip('dask_searchcv')
X, y = make_classification(n_samples=100, n_features=5, chunksize=10)
grid = {'logisticregression__lamduh': [.001, .01, .1, .5]}
pipe = make_pipeline(DoNothingTransformer(), LogisticRegression())
search = dcv.GridSearchCV(pipe, grid, cv=3)
search.fit(X, y)
def test_big(fit_intercept, is_sparse, is_cupy):
with dask.config.set(scheduler='synchronous'):
X, y = make_classification(is_sparse=is_sparse)
if is_cupy and not is_sparse:
cupy = pytest.importorskip('cupy')
X, y = to_dask_cupy_array_xy(X, y, cupy)
lr = LogisticRegression(fit_intercept=fit_intercept)
lr.fit(X, y)
lr.predict(X)
lr.predict_proba(X)
if fit_intercept:
assert lr.intercept_ is not None
def test_fit(fit_intercept, is_sparse, is_cupy):
X, y = make_classification(n_samples=100,
n_features=5,
chunksize=10,
is_sparse=is_sparse)
if is_cupy and not is_sparse:
cupy = pytest.importorskip('cupy')
X, y = to_dask_cupy_array_xy(X, y, cupy)
lr = LogisticRegression(fit_intercept=fit_intercept)
lr.fit(X, y)
lr.predict(X)
lr.predict_proba(X)
def __init__(self,
*,
client=None,
fit_intercept=True,
solver="admm",
penalty="l2",
C=1.0,
max_iter=100,
verbose=False,
**kwargs):
super(LogisticRegression, self).__init__(client=client,
verbose=verbose,
**kwargs)
if not has_daskglm("0.2.1.dev"):
raise ImportError(
"dask-glm >= 0.2.1.dev was not found, please install it"
" to use multi-GPU logistic regression.")
from dask_glm.estimators import LogisticRegression \
as LogisticRegressionGLM
self.fit_intercept = fit_intercept
self.solver = solver
self.penalty = penalty
self.C = C
self.max_iter = max_iter
if self.penalty not in ("l2", "l1", "elastic_net"):
raise TypeError("Only l2, l1, and elastic_net penalties are"
" currently supported.")
self.solver_model = LogisticRegressionGLM(
solver=self.solver,
fit_intercept=self.fit_intercept,
regularizer=self.penalty,
max_iter=self.max_iter,
lamduh=1 / self.C,
)
def test_in_pipeline():
from sklearn.pipeline import make_pipeline
X, y = make_classification(n_samples=100, n_features=5, chunksize=10)
pipe = make_pipeline(DoNothingTransformer(), LogisticRegression())
pipe.fit(X, y)
def test_lr_init(solver):
LogisticRegression(solver=solver)
class LogisticRegression(BaseEstimator):
"""
Distributed Logistic Regression for Binary classification.
Parameters
----------
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.
solver : 'admm'
Solver to use. Only admm is supported currently.
penalty : {'l1', 'l2', 'elastic_net'} (default = 'l2')
Regularization technique for the solver.
C: float (default = 1.0)
Inverse of regularization strength; must be a positive float.
max_iter: int (default = 100)
Maximum number of iterations taken for the solvers to converge.
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.
Attributes
----------
coef_: device array (n_features, 1)
The estimated coefficients for the logistic regression model.
intercept_: device array (1,)
The independent term. If `fit_intercept` is False, will be 0.
solver: string
Algorithm to use in the optimization process. Currently only `admm` is
supported.
Notes
------
This estimator is a wrapper class around Dask-GLM's
Logistic Regression estimator. Several methods in this wrapper class
duplicate code from Dask-GLM to create a user-friendly namespace.
"""
def __init__(self,
*,
client=None,
fit_intercept=True,
solver="admm",
penalty="l2",
C=1.0,
max_iter=100,
verbose=False,
**kwargs):
super(LogisticRegression, self).__init__(client=client,
verbose=verbose,
**kwargs)
if not has_daskglm("0.2.1.dev"):
raise ImportError(
"dask-glm >= 0.2.1.dev was not found, please install it"
" to use multi-GPU logistic regression.")
from dask_glm.estimators import LogisticRegression \
as LogisticRegressionGLM
self.fit_intercept = fit_intercept
self.solver = solver
self.penalty = penalty
self.C = C
self.max_iter = max_iter
if self.penalty not in ("l2", "l1", "elastic_net"):
raise TypeError("Only l2, l1, and elastic_net penalties are"
" currently supported.")
self.solver_model = LogisticRegressionGLM(
solver=self.solver,
fit_intercept=self.fit_intercept,
regularizer=self.penalty,
max_iter=self.max_iter,
lamduh=1 / self.C,
)
@with_cupy_rmm
def fit(self, X, y):
"""
Fit the model with X and y.
Parameters
----------
X : Dask cuDF dataframe or CuPy backed Dask Array (n_rows, n_features)
Features for regression
y : Dask cuDF Series or CuPy backed Dask Array (n_rows,)
Label (outcome values)
"""
X = self._input_to_dask_cupy_array(X)
y = self._input_to_dask_cupy_array(y)
self.solver_model.fit(X, y)
self._finalize_coefs()
return self
@with_cupy_rmm
def predict(self, X):
"""
Predicts the ŷ for X.
Parameters
----------
X : Dask cuDF dataframe or CuPy backed Dask Array (n_rows, n_features)
Distributed dense matrix (floats or doubles) of shape
(n_samples, n_features).
Returns
-------
y : Dask cuDF Series or CuPy backed Dask Array (n_rows,)
"""
return self.predict_proba(X) > 0.5
@with_cupy_rmm
def predict_proba(self, X):
from dask_glm.utils import sigmoid
X = self._input_to_dask_cupy_array(X)
return sigmoid(self.decision_function(X))
@with_cupy_rmm
def decision_function(self, X):
X = self._input_to_dask_cupy_array(X)
X_ = self._maybe_add_intercept(X)
return np.dot(X_, self._coef)
@with_cupy_rmm
def score(self, X, y):
from dask_glm.utils import accuracy_score
X = self._input_to_dask_cupy_array(X)
y = self._input_to_dask_cupy_array(y)
return accuracy_score(y, self.predict(X))
@with_cupy_rmm
def _finalize_coefs(self):
# _coef contains coefficients and (potentially) intercept
self._coef = cp.asarray(self.solver_model._coef)
if self.fit_intercept:
self.coef_ = self._coef[:-1]
self.intercept_ = self.solver_model._coef[-1]
else:
self.coef_ = self._coef
@with_cupy_rmm
def _maybe_add_intercept(self, X):
from dask_glm.utils import add_intercept
if self.fit_intercept:
return add_intercept(X)
else:
return X
@with_cupy_rmm
def _input_to_dask_cupy_array(self, X):
if (is_dataframe_like(X) or is_series_like(X)) and hasattr(X, "dask"):
if not isinstance(X._meta, (cudf.Series, cudf.DataFrame)):
raise TypeError("Please convert your Dask DataFrame"
" to a Dask-cuDF DataFrame using dask_cudf.")
X = X.values
X._meta = cp.asarray(X._meta)
elif is_arraylike(X) and hasattr(X, "dask"):
if not isinstance(X._meta, cp.ndarray):
raise TypeError("Please convert your CPU Dask Array"
" to a GPU Dask Array using"
" arr.map_blocks(cp.asarray).")
else:
raise TypeError("Please pass a GPU backed Dask DataFrame"
" or Dask Array.")
X.compute_chunk_sizes()
return X
def get_param_names(self):
return list(self.kwargs.keys())