[BUG] roll back state when set_params fails validation

skbase/base/_base.py

@@ -384,6 +384,14 @@ def set_params(self, **params):
             return self
         valid_params = self.get_params(deep=True)
 
+        # snapshot instance state, to restore it if the reset call below raises,
+        # e.g., through a parameter validation failure in __init__.
+        # Must be taken before the setattr writes below, since those writes are
+        # what leave self in a state that __init__ could not have produced.
+        # This is a shallow copy of the instance __dict__, i.e., it stores
+        # references to attribute values, it does not copy the values themselves.
+        prev_state = self.__dict__.copy()
+
         unmatched_keys = []
 
         nested_params = defaultdict(dict)  # grouped by prefix
@@ -405,7 +413,19 @@ def set_params(self, **params):
 
         # all matched params have now been set
         # reset object to clean post-init state with those params
-        self.reset()
+        try:
+            self.reset()
+        except Exception as e:
+            # restore the pre-call state, then report what happened,
+            # the original exception is chained via "from e"
+            self.__dict__.clear()
+            self.__dict__.update(prev_state)
+            raise RuntimeError(
+                f"Error in {type(self).__name__}.set_params, the parameter values "
+                f"passed were rejected when re-running __init__, which raised "
+                f"{type(e).__name__}: {e}. The object has been restored to its "
+                f"state before the set_params call."
+            ) from e
 
         # recurse in components
         for key, sub_params in nested_params.items():

skbase/tests/test_base.py

@@ -903,6 +903,54 @@ def test_set_params_with_no_param_to_set_returns_object(
     )
 
 
+class ValidatingObject(BaseObject):
+    """BaseObject whose __init__ writes then validates a parameter.
+
+    Regression fixture for https://github.com/sktime/sktime/issues/10695 :
+    __init__ assigns ``self.x`` before raising, mirroring the common pattern
+    of estimators that write a hyper-parameter to ``self`` ahead of a
+    validation check.
+    """
+
+    def __init__(self, x=1):
+        self.x = x
+        if x < 0:
+            raise ValueError("x must be non-negative")
+        super().__init__()
+
+
+def test_set_params_rolls_back_state_on_invalid_value():
+    """Test a failed set_params leaves the object in its pre-call state.
+
+    A failed set_params previously left self holding the rejected value,
+    since set_params writes parameters to self before reset() re-runs
+    __init__ for validation. A state __init__ could never have produced
+    would then survive the raise, and break get_params, clone and repeated
+    set_params calls from that point on.
+    """
+    obj = ValidatingObject(x=5)
+
+    with pytest.raises(RuntimeError, match="restored to its state") as exc_info:
+        obj.set_params(x=-1)
+
+    # the exception names the failing object and reports the restore
+    assert "ValidatingObject.set_params" in str(exc_info.value)
+    # the original __init__ exception is chained, not discarded
+    assert isinstance(exc_info.value.__cause__, ValueError)
+    assert "x must be non-negative" in str(exc_info.value.__cause__)
+
+    assert obj.x == 5
+    assert obj.get_params() == {"x": 5}
+
+    # clone should still work off the pre-call state
+    cloned = obj.clone()
+    assert cloned.get_params() == {"x": 5}
+
+    # a subsequent valid set_params call should still work normally
+    obj.set_params(x=9)
+    assert obj.get_params() == {"x": 9}
+
+
 # This section tests the clone functionality
 # These have been adapted from sklearn's tests of clone to use the clone
 # method that is included as part of the BaseObject interface