minimal example of GridSearchCV showing compilation issue

Author: floriankozikowski

skglm/estimators.py

@@ -188,24 +188,105 @@ class GeneralizedLinearEstimator(LinearModel):
         Number of subproblems solved to reach the specified tolerance.
     """
 
+    _parameter_constraints: dict = {
+        "datafit": [None, "object"],
+        "penalty": [None, "object"],
+        "solver": [None, "object"],
+    }
+
     def __init__(self, datafit=None, penalty=None, solver=None):
         super(GeneralizedLinearEstimator, self).__init__()
         self.penalty = penalty
         self.datafit = datafit
         self.solver = solver
 
     def __repr__(self):
-        """Get string representation of the estimator.
+        """String representation."""
+        penalty_name = self.penalty.__class__.__name__ if self.penalty else "None"
+        datafit_name = self.datafit.__class__.__name__ if self.datafit else "None"
+
+        if self.penalty and hasattr(self.penalty, 'alpha'):
+            penalty_alpha = self.penalty.alpha
+        else:
+            penalty_alpha = None
 
-        Returns
-        -------
-        repr : str
-            String representation.
-        """
         return (
             'GeneralizedLinearEstimator(datafit=%s, penalty=%s, alpha=%s)'
-            % (self.datafit.__class__.__name__, self.penalty.__class__.__name__,
-               self.penalty.alpha))
+            % (datafit_name, penalty_name, penalty_alpha))
+
+    def get_params(self, deep=True):
+        """Get parameters, including nested datafit, penalty, solver hyper-parameters."""
+        # First get the top-level params (datafit, penalty, solver)
+        params = super().get_params(deep=False)
+
+        if not deep:
+            return params
+
+        # For each sub-estimator, ask its own get_params if available
+        for comp_name in ('datafit', 'penalty', 'solver'):
+            comp = getattr(self, comp_name)
+            if comp is None:
+                continue
+
+            # If it implements sklearn API, use get_params
+            if hasattr(comp, 'get_params'):
+                sub_params = comp.get_params(deep=True)
+            else:
+                # Otherwise fall back to its public attributes
+                sub_params = {k: v for k, v in vars(
+                    comp).items() if not k.startswith('_')}
+
+            for sub_key, sub_val in sub_params.items():
+                params[f'{comp_name}__{sub_key}'] = sub_val
+
+        return params
+
+    def set_params(self, **params):
+        """Set parameters, including nested ones for datafit, penalty, solver."""
+        if not params:
+            return self
+
+        # Top-level valid args are exactly those in __init__
+        valid_top = set(self.get_params(deep=False))
+
+        # Collect sub-params for each component
+        nested = {'datafit': {}, 'penalty': {}, 'solver': {}}
+
+        # First pass: split top-level vs nested
+        for key, val in params.items():
+            if '__' in key:
+                comp, sub_key = key.split('__', 1)
+                if comp not in nested:
+                    raise ValueError(f"Invalid parameter: {key}")
+                nested[comp][sub_key] = val
+            else:
+                if key not in valid_top:
+                    raise ValueError(f"Invalid parameter: {key}")
+                setattr(self, key, val)
+
+        # Second pass: apply nested updates
+        for comp, comp_params in nested.items():
+            if not comp_params:
+                continue
+            current = getattr(self, comp)
+            if current is not None and hasattr(current, 'set_params'):
+                current.set_params(**comp_params)
+            elif current is not None:
+                # fallback to simple setattr
+                for sub_key, sub_val in comp_params.items():
+                    if not hasattr(current, sub_key):
+                        raise ValueError(f"{comp} has no parameter {sub_key}")
+                    setattr(current, sub_key, sub_val)
+            else:
+                # instantiate missing component
+                if comp == 'datafit':
+                    self.datafit = Quadratic(**comp_params)
+                elif comp == 'penalty':
+                    self.penalty = L1(**comp_params)
+                elif comp == 'solver':
+                    self.solver = AndersonCD(**comp_params)
+
+        return self
 
     def fit(self, X, y):
         """Fit estimator.
@@ -269,31 +350,6 @@ def predict(self, X):
         else:
             return self._decision_function(X)
 
-    def get_params(self, deep=False):
-        """Get parameters of the estimators including the datafit's and penalty's.
-
-        Parameters
-        ----------
-        deep : bool
-            Whether or not return the parameters for contained subobjects estimators.
-
-        Returns
-        -------
-        params : dict
-            The parameters of the estimator.
-        """
-        params = super().get_params(deep)
-        filtered_types = (float, int, str, np.ndarray)
-        penalty_params = [('penalty__', p, getattr(self.penalty, p)) for p in
-                          dir(self.penalty) if p[0] != "_" and
-                          type(getattr(self.penalty, p)) in filtered_types]
-        datafit_params = [('datafit__', p, getattr(self.datafit, p)) for p in
-                          dir(self.datafit) if p[0] != "_" and
-                          type(getattr(self.datafit, p)) in filtered_types]
-        for p_prefix, p_key, p_val in penalty_params + datafit_params:
-            params[p_prefix + p_key] = p_val
-        return params
-
 
 class Lasso(RegressorMixin, LinearModel):
     r"""Lasso estimator based on Celer solver and primal extrapolation.

skglm/tests/test_gridsearch.py

@@ -0,0 +1,48 @@
+from sklearn.model_selection import GridSearchCV
+from sklearn.datasets import make_regression
+from skglm import GeneralizedLinearEstimator
+from skglm.datafits import Quadratic
+from skglm.penalties import L1
+from skglm.solvers import AndersonCD
+
+
+def test_gridsearch_compatibility():
+    # Generate synthetic data
+    X, y = make_regression(n_samples=100, n_features=10, random_state=42)
+
+    # Create base estimator
+    base_estimator = GeneralizedLinearEstimator(
+        datafit=Quadratic(),
+        penalty=L1(1.0),
+        solver=AndersonCD()
+    )
+
+    # Define parameter grid
+    param_grid = {
+        'penalty__alpha': [0.1, 1.0, 10.0],
+        'solver__max_iter': [10, 20],
+        'solver__tol': [1e-3, 1e-4]
+    }
+
+    # Create GridSearchCV
+    grid_search = GridSearchCV(
+        base_estimator,
+        param_grid,
+        cv=3,
+        scoring='neg_mean_squared_error'
+    )
+
+    # Fit GridSearchCV
+    grid_search.fit(X, y)
+
+    # Verify that GridSearchCV worked
+    assert hasattr(grid_search, 'best_params_')
+    assert hasattr(grid_search, 'best_score_')
+    assert hasattr(grid_search, 'best_estimator_')
+
+    # Verify that best_estimator_ has the correct parameters
+    best_estimator = grid_search.best_estimator_
+    assert isinstance(best_estimator, GeneralizedLinearEstimator)
+    assert best_estimator.penalty.alpha in [0.1, 1.0, 10.0]
+    assert best_estimator.solver.max_iter in [10, 20]
+    assert best_estimator.solver.tol in [1e-3, 1e-4]

skglm/utils/jit_compilation.py

@@ -59,7 +59,17 @@ def jit_cached_compile(klass, spec, to_float32=False):
     if to_float32:
         spec = spec_to_float32(spec)
 
-    return jitclass(spec)(klass)
+    # Create a new class without slots
+    class CompiledClass:
+        pass
+
+    # Copy over all methods and attributes from the original class
+    for name, value in klass.__dict__.items():
+        # Skip __slots__ and __slotnames__ but keep other special methods
+        if name not in ['__slots__', '__slotnames__']:
+            setattr(CompiledClass, name, value)
+
+    return jitclass(spec)(CompiledClass)
 
 
 def compiled_clone(instance, to_float32=False):