fix(models): compare array-valued LSS family options by value

deeptab/models/lss_base.py

@@ -28,6 +28,29 @@
 from deeptab.training import TaskModel
 
 
+def _as_numpy(value):
+    """Return *value* as a NumPy array; tensors are moved to the CPU first."""
+    return value.detach().cpu().numpy() if isinstance(value, torch.Tensor) else np.asarray(value)
+
+
+def _same_family_options(options: dict, stored: dict) -> bool:
+    """Compare family options, treating arrays and tensors (e.g. quantile levels) by value.
+
+    A plain ``options != stored`` raises for array values, because NumPy and
+    torch cannot turn an element-wise comparison into a single truth value.
+    """
+    if options.keys() != stored.keys():
+        return False
+    for key, value in options.items():
+        other = stored[key]
+        if isinstance(value, np.ndarray | torch.Tensor) or isinstance(other, np.ndarray | torch.Tensor):
+            if not np.array_equal(_as_numpy(value), _as_numpy(other)):
+                return False
+        elif value != other:
+            return False
+    return True
+
+
 class SklearnBaseLSS(SklearnBase):
     """Distributional regression base class (LSS variant of SklearnBase).
 
@@ -59,7 +82,9 @@ def _configure_distribution(self, y, family=None, distributional_kwargs=None, *,
                 raise ValueError("Dirichlet num_classes must match the number of target components")
             options["num_classes"] = count
         if not allow_change:
-            if family != getattr(self, "family_name", None) or options != getattr(self, "distributional_kwargs_", {}):
+            if family != getattr(self, "family_name", None) or not _same_family_options(
+                options, getattr(self, "distributional_kwargs_", {})
+            ):
                 raise ValueError("Changing the distribution family or options requires rebuild=True")
             return
         self.family = get_distribution(family, **options)

tests/integration/test_lss_base.py

@@ -532,6 +532,33 @@ def test_reused_build_rejects_changed_options_without_mutating_family():
     assert model.family_name == "normal"
 
 
+@pytest.mark.parametrize(
+    "quantiles", [np.array([0.1, 0.5, 0.9]), torch.tensor([0.1, 0.5, 0.9], dtype=torch.float64)], ids=["numpy", "torch"]
+)
+def test_reused_build_accepts_array_options(quantiles):
+    model = _small_lss()
+    model.build_model(_X[:24], _Y[:24], family="quantile", distributional_kwargs={"quantiles": quantiles})
+    model.fit(_X[:24], _Y[:24], family="quantile", rebuild=False)
+    model.fit(_X[:24], _Y[:24], family="quantile", distributional_kwargs={"quantiles": [0.1, 0.5, 0.9]}, rebuild=False)
+    assert model.is_fitted_
+    assert model.predict(_X[:4]).shape == (4, 3)
+
+
+def test_reused_build_rejects_different_numpy_array_options():
+    model = _small_lss()
+    model.build_model(
+        _X[:24], _Y[:24], family="quantile", distributional_kwargs={"quantiles": np.array([0.1, 0.5, 0.9])}
+    )
+    with pytest.raises(ValueError, match="requires rebuild=True"):
+        model.fit(
+            _X[:24],
+            _Y[:24],
+            family="quantile",
+            distributional_kwargs={"quantiles": np.array([0.2, 0.5, 0.8])},
+            rebuild=False,
+        )
+
+
 def test_legacy_lss_bundle_loads_without_distributional_kwargs(tmp_path, fitted_mlplss):
     path = str(tmp_path / "legacy.deeptab")
     fitted_mlplss.save(path)