Merge pull request #6512 from noahnovsak/dask-logisticregression

[ENH] Dask: Logistic Regression

Orange/base.py

@@ -141,7 +141,7 @@ def __call__(self, data, progress_callback=None):
 
         progress_callback(0.1, "Fitting...")
         model = self._fit_model(data)
-        model.used_vals = [np.unique(y).astype(int) for y in data.Y[:, None].T]
+        model.used_vals = [np.asarray(np.unique(y), dtype=int) for y in data.Y[:, None].T]
         if not hasattr(model, "domain") or model.domain is None:
             # some models set domain themself and it should be respected
             # e.g. calibration learners set the base_learner's domain which

Orange/classification/logistic_regression.py

@@ -1,11 +1,20 @@
+import warnings
+
 import numpy as np
+import dask.array as da
 import sklearn.linear_model as skl_linear_model
 
+try:
+    import dask_ml.linear_model as dask_linear_model
+except ImportError:
+    dask_linear_model = skl_linear_model
+
 from Orange.classification import SklLearner, SklModel
 from Orange.preprocess import Normalize
 from Orange.preprocess.score import LearnerScorer
 from Orange.data import Variable, DiscreteVariable
 
+
 __all__ = ["LogisticRegressionLearner"]
 
 
@@ -22,11 +31,11 @@ def score(self, data):
 class LogisticRegressionClassifier(SklModel):
     @property
     def intercept(self):
-        return self.skl_model.intercept_
+        return np.atleast_1d(self.skl_model.intercept_)
 
     @property
     def coefficients(self):
-        return self.skl_model.coef_
+        return np.atleast_2d(self.skl_model.coef_)
 
 
 class LogisticRegressionLearner(SklLearner, _FeatureScorerMixin):
@@ -43,9 +52,23 @@ def __init__(self, penalty="l2", dual=False, tol=0.0001, C=1.0,
 
     def _initialize_wrapped(self, X=None, Y=None):
         params = self.params.copy()
+        solver = params.pop("solver")
+        penalty = params.get("penalty") or "none"
+
+        if isinstance(X, da.Array) or isinstance(Y, da.Array):
+            if dask_linear_model is skl_linear_model:
+                warnings.warn("dask_ml is not installed, using sklearn instead.")
+            else:
+                if solver == "auto":
+                    if penalty in "none":
+                        solver = "gradient_descent"
+                    else:
+                        solver = "admm"
+                params["solver"], params["penalty"] = solver, penalty
+                return dask_linear_model.LogisticRegression(**params)
+
         # The default scikit-learn solver `lbfgs` (v0.22) does not support the
         # l1 penalty.
-        solver, penalty = params.pop("solver"), params.get("penalty")
         if solver == "auto":
             if penalty == "l1":
                 solver = "liblinear"

Orange/tests/test_logistic_regression.py

@@ -9,6 +9,7 @@
 from Orange.data import Table, ContinuousVariable, Domain
 from Orange.classification import LogisticRegressionLearner, Model
 from Orange.evaluation import CrossValidation, CA
+from Orange.tests.test_dasktable import temp_dasktable
 
 
 class TestLogisticRegressionLearner(unittest.TestCase):
@@ -114,9 +115,9 @@ def test_coefficients(self):
 
     def test_predict_on_instance(self):
         lr = LogisticRegressionLearner()
-        m = lr(self.zoo)
-        probs = m(self.zoo[50], m.Probs)
-        probs2 = m(self.zoo[50, :], m.Probs)
+        m = lr(self.heart_disease)
+        probs = m(self.heart_disease[50], m.Probs)
+        probs2 = m(self.heart_disease[50, :], m.Probs)
         np.testing.assert_almost_equal(probs, probs2[0])
 
     def test_single_class(self):
@@ -151,3 +152,54 @@ def test_auto_solver(self):
         skl_clf = lr._initialize_wrapped()
         self.assertEqual(skl_clf.solver, "liblinear")
         self.assertEqual(skl_clf.penalty, "l1")
+
+
+class TestLogisticRegressionOnDask(TestLogisticRegressionLearner):
+    @classmethod
+    def setUpClass(cls):
+        cls.iris = temp_dasktable(Table('iris'))
+        cls.heart_disease = temp_dasktable(Table('heart_disease.tab'))
+
+    def test_learner_scorer(self):
+        # for some reason dask_ml and sklearn yield different results
+        learner = LogisticRegressionLearner()
+        scores = learner.score_data(self.heart_disease)
+        self.assertEqual('major vessels colored',
+                         self.heart_disease.domain.attributes[np.argmax(scores)].name)
+        self.assertEqual(scores.shape, (1, len(self.heart_disease.domain.attributes)))
+
+    @unittest.skip("Discretizer not yet implemented")
+    def test_learner_scorer_previous_transformation(self):
+        super().test_learner_scorer_previous_transformation()
+
+    @unittest.skip("Dask-ML does not support multiclass regression")
+    def test_learner_scorer_multiclass(self):
+        super().test_learner_scorer_multiclass()
+
+    @unittest.skip("Dask-ML does not support multiclass regression")
+    def test_learner_scorer_multiclass_feature(self):
+        super().test_learner_scorer_multiclass_feature()
+
+    @unittest.skip("Dask-ML accepts single class")
+    def test_single_class(self):
+        super().test_single_class()
+
+    @unittest.skip("Dask-ML accepts single class")
+    def test_sklearn_single_class(self):
+        super().test_sklearn_single_class()
+
+    def test_auto_solver(self):
+        lr = LogisticRegressionLearner(penalty="l2", solver="auto")
+        skl_clf = lr._initialize_wrapped(self.iris.X)
+        self.assertEqual(skl_clf.solver, "admm")
+        self.assertEqual(skl_clf.penalty, "l2")
+
+        lr = LogisticRegressionLearner(penalty="l1", solver="auto")
+        skl_clf = lr._initialize_wrapped(self.iris.X)
+        self.assertEqual(skl_clf.solver, "admm")
+        self.assertEqual(skl_clf.penalty, "l1")
+
+        lr = LogisticRegressionLearner(penalty=None, solver="auto")
+        skl_clf = lr._initialize_wrapped(self.iris.X)
+        self.assertEqual(skl_clf.solver, "gradient_descent")
+        self.assertEqual(skl_clf.penalty, "none")

Orange/widgets/evaluate/owtestandscore.py

@@ -950,7 +950,7 @@ def __update(self):
                     do_stratify = False
                 elif self.data.domain.class_var.is_discrete:
                     least = min(filter(None,
-                                       np.bincount(self.data.Y.astype(int))))
+                                       np.bincount(np.asarray(self.data.Y, dtype=int))))
                     if least < k:
                         self.Warning.cant_stratify(k, least)
                         do_stratify = False

Orange/widgets/model/owlogisticregression.py

@@ -5,6 +5,7 @@
 from orangewidget.report import bool_str
 
 from Orange.data import Table, Domain, ContinuousVariable, StringVariable
+from Orange.data.dask import DaskTable
 from Orange.classification.logistic_regression import LogisticRegressionLearner
 from Orange.widgets import settings, gui
 from Orange.widgets.utils.owlearnerwidget import OWBaseLearner
@@ -139,6 +140,15 @@ def get_learner_parameters(self):
             self.penalty_types[self.penalty_type], self.C_s[self.C_index],
             bool_str(self.class_weight))),)
 
+    def check_data(self):
+        valid = super().check_data()
+        if valid and isinstance(self.data, DaskTable) \
+                and len(self.data.domain.class_var.values) > 2 \
+                and len(np.unique(self.data).compute()) > 2:
+            self.Error.data_error("Data contains too many target values.")
+            valid = False
+        return valid
+
 
 def create_coef_table(classifier):
     i = classifier.intercept