Remove predict option

docs/examples/example2_student_admission.py

@@ -189,7 +189,7 @@
 # model.
 #
 pred_constr = add_predictor_constr(
-    m, pipe, students_opt_data, predict_function="predict_proba"
+    m, pipe, students_opt_data, predict_function="decision_function"
 )
 pred_constr.print_stats()
 m.update()
@@ -204,7 +204,7 @@
 # model we are only interested in the probability of class 1, i.e.
 # the column of index 1. We store this as a pandas series
 
-y = pd.Series(data=pred_constr.output[:, 1].tolist(), index=x.index)
+y = pd.Series(data=pred_constr.output[:, 0].tolist(), index=x.index)
 
 y
 

src/gurobi_ml/modeling/softmax.py

@@ -100,14 +100,20 @@ def _addGenConstrIndicatorMvarV10(binvar, binval, lhs, sense, rhs, name):
 def logistic(predictor_model: AbstractPredictorConstr, linear_predictor: gp.MVar):
     log_result = predictor_model.output[:, 1]
 
+    print(linear_predictor.shape)
+    print(log_result.shape)
     if _HAS_NL_EXPR:
         predictor_model.gp_model.addConstr(
             log_result == nlfunc.logistic(linear_predictor[:, 0])
         )
     else:
+        linear_predictor_vars = predictor_model.gp_model.addMVar(
+            log_result.shape[0], lb=-gp.GRB.INFINITY, name="linear_predictor"
+        )
+        predictor_model.gp_model.addConstr(linear_predictor_vars == linear_predictor)
         for index in np.ndindex(log_result.shape):
             predictor_model.gp_model.addGenConstrLogistic(
-                linear_predictor[index],
+                linear_predictor_vars[index],
                 log_result[index],
                 name=predictor_model._indexed_name(index, "logistic"),
             )

src/gurobi_ml/sklearn/logistic_regression.py

@@ -22,7 +22,8 @@
 
 import gurobipy as gp
 
-from ..modeling.softmax import hardmax, logistic, max2, softmax
+from ..modeling.base_predictor_constr import AbstractPredictorConstr
+from ..modeling.softmax import logistic, softmax
 
 try:
     pass
@@ -37,7 +38,6 @@
     _HAS_NL_EXPR = False
 
 from ..exceptions import ParameterError
-from .base_regressions import BaseSKlearnRegressionConstr
 from .skgetter import SKClassifier
 
 
@@ -46,7 +46,7 @@ def add_logistic_regression_constr(
     logistic_regression,
     input_vars,
     output_vars=None,
-    predict_function="predict",
+    predict_function="predict_proba",
     epsilon=0.0,
     pwl_attributes=None,
     **kwargs,
@@ -146,7 +146,7 @@ def add_logistic_regression_constr(
     )
 
 
-class LogisticRegressionConstr(SKClassifier, BaseSKlearnRegressionConstr):
+class LogisticRegressionConstr(SKClassifier, AbstractPredictorConstr):
     """Class to formulate a trained
     :external+sklearn:py:class:`sklearn.linear_model.LogisticRegression` in a gurobipy model.
 
@@ -159,21 +159,21 @@ def __init__(
         predictor,
         input_vars,
         output_vars=None,
-        predict_function="predict",
+        predict_function="predict_proba",
         epsilon=0.0,
         pwl_attributes=None,
         **kwargs,
     ):
-        if predict_function not in ("predict", "predict_proba", "decision_function"):
+        if predict_function not in ("predict_proba", "decision_function"):
             raise ParameterError(
                 "predict_function should be either 'predict' or 'predict_proba'"
             )
-        if predict_function == "predict" and pwl_attributes is not None:
+        if predict_function != "predict_proba" and pwl_attributes is not None:
             message = """
 pwl_attributes are not required for classification.  The problem is
 formulated without requiring the non-linear logistic function."""
             warnings.warn(message)
-        elif predict_function != "predict":
+        elif predict_function == "predict_proba":
             self.attributes = (
                 self.default_pwl_attributes()
                 if pwl_attributes is None
@@ -184,10 +184,11 @@ def __init__(
         self._default_name = "log_reg"
         self.linear_predictor = None
         SKClassifier.__init__(self, predictor, input_vars, predict_function)
-        BaseSKlearnRegressionConstr.__init__(
+        if self._output_shape == 2 and predict_function == "decision_function":
+            self._output_shape = 1
+        AbstractPredictorConstr.__init__(
             self,
             gp_model,
-            predictor,
             input_vars,
             output_vars,
             **kwargs,
@@ -223,41 +224,29 @@ def default_pwl_attributes() -> dict:
 
     def _two_classes_model(self, **kwargs):
         """Add the prediction constraints to Gurobi."""
-        m, _ = self.output.shape
-
-        linear_predictor = self.gp_model.addMVar(
-            (m, 1), lb=-gp.GRB.INFINITY, name="linear_predictor"
-        )
-        self._add_regression_constr(output=linear_predictor)
+        coefs = self.predictor.coef_
+        intercept = self.predictor.intercept_
 
-        self.gp_model.addConstr(self.output.sum(axis=1) == 1)
+        linreg = self.input @ coefs.T + intercept
 
-        self.linear_predictor = linear_predictor
+        self.linear_predictor = linreg
 
-        if self.predict_function == "predict":
-            max2(self, linear_predictor, self.epsilon)
+        if self.predict_function == "predict_proba":
+            self.gp_model.addConstr(self.output.sum(axis=1) == 1)
+            logistic(self, linreg)
         else:
-            logistic(self, linear_predictor)
+            self.gp_model.addConstr(self.output[:, 0] == linreg[:, 0])
 
         self.gp_model.update()
 
-    @property
-    def linear_predictor_variables(self) -> gp.MVar:
-        """Variables that store the result of the linear_predictor
-        (i.e. before applying the logistic function).
-        """
-        return self.linear_predictor
-
     def _multi_class_model(self, **kwargs):
         """Add the prediction constraints to Gurobi."""
         coefs = self.predictor.coef_
         intercept = self.predictor.intercept_
 
         linreg = self.input @ coefs.T + intercept
 
-        if self.predict_function == "predict":
-            hardmax(self, linreg, **kwargs)
-        elif self.predict_function == "predict_proba":
+        if self.predict_function == "predict_proba":
             softmax(self, linreg, **kwargs)
         else:
             self.gp_model.addConstr(self.output == linreg)

src/gurobi_ml/sklearn/pipeline.py

@@ -95,8 +95,6 @@ def _build_submodel(self, gp_model, *args, **kwargs):
         self._mip_model(**kwargs)
         assert self.output is not None
         assert self.input is not None
-        # We can call validate only after the model is created
-        self._validate()
         return self
 
     def _mip_model(self, **kwargs):

tests/test_sklearn/test_sklearn_formulations.py

@@ -115,8 +115,10 @@ def test_iris_clf(self):
 
         for regressor in cases:
             onecase = cases.get_case(regressor)
-            self.do_one_case(onecase, X, 5, "all", predict_function="predict")
-            self.do_one_case(onecase, X, 6, "pairs", predict_function="predict")
+            self.do_one_case(onecase, X, 5, "all", predict_function="decision_function")
+            self.do_one_case(
+                onecase, X, 6, "pairs", predict_function="decision_function"
+            )
 
     def test_iris_pwl_args(self):
         data = datasets.load_iris()