Filter nonfinite calibration references (#2262)

* Filter nonfinite calibration references

* Ignore unusable bias calibration references

* Test nonfinite calibration references

Author: FlorianPfaff

src/pyrecest/calibration/bias.py

@@ -167,7 +167,8 @@ def make_bias_training_examples(
         if features.shape[0] != measurements.shape[0]:
             raise ValueError("feature_values rows must match measurement_values rows")
 
-    if reference_times.size == 0:
+    finite_reference = np.isfinite(reference_times) & np.isfinite(references).all(axis=1)
+    if not finite_reference.any():
         return BiasTrainingExamples(
             measured=np.empty((0, measurements.shape[1])),
             reference=np.empty((0, measurements.shape[1])),
@@ -181,6 +182,8 @@ def make_bias_training_examples(
             time_delta_s=np.empty(0),
         )
 
+    reference_times = reference_times[finite_reference]
+    references = references[finite_reference]
     order = np.argsort(reference_times)
     reference_times = reference_times[order]
     references = references[order]
@@ -189,7 +192,6 @@ def make_bias_training_examples(
     valid = (
         np.isfinite(measurement_times)
         & np.isfinite(measurements).all(axis=1)
-        & np.isfinite(references[nearest]).all(axis=1)
         & (delta_s <= float(max_time_delta_s))
     )
     valid &= np.isfinite(features).all(axis=1)

src/pyrecest/calibration/time_offset.py

@@ -94,15 +94,34 @@ def _validate_max_time_delta(max_time_delta_s: float | None) -> None:
         raise ValueError("max_time_delta_s must be nonnegative")
 
 
+def _finite_reference_rows(
+    reference_times_s: np.ndarray,
+    reference_values: np.ndarray | None = None,
+) -> np.ndarray:
+    """Return a mask selecting reference rows that are usable for matching."""
+
+    reference_times = np.asarray(reference_times_s, dtype=float).reshape(-1)
+    finite = np.isfinite(reference_times)
+    if reference_values is not None:
+        values = np.asarray(reference_values, dtype=float)
+        if values.ndim == 1:
+            values = values.reshape(-1, 1)
+        finite &= np.isfinite(values).all(axis=1)
+    return finite
+
+
 def nearest_time_indices(
     reference_times_s: np.ndarray, query_times_s: np.ndarray
 ) -> np.ndarray:
-    """Return original indices of nearest reference times for each query time."""
+    """Return original indices of nearest finite reference times for each query time."""
 
     reference = np.asarray(reference_times_s, dtype=float).reshape(-1)
     query = np.asarray(query_times_s, dtype=float).reshape(-1)
-    if reference.size == 0:
-        raise ValueError("reference_times_s must not be empty")
+    finite_reference = _finite_reference_rows(reference)
+    if not finite_reference.any():
+        raise ValueError("reference_times_s must contain at least one finite value")
+    original_indices = np.flatnonzero(finite_reference)
+    reference = reference[finite_reference]
     order = np.argsort(reference)
     sorted_reference = reference[order]
     insertion = np.searchsorted(sorted_reference, query)
@@ -111,7 +130,7 @@ def nearest_time_indices(
     use_right = np.abs(sorted_reference[right] - query) < np.abs(
         sorted_reference[left] - query
     )
-    return order[np.where(use_right, right, left)]
+    return original_indices[order[np.where(use_right, right, left)]]
 
 
 def interpolate_reference_values(
@@ -133,8 +152,11 @@ def interpolate_reference_values(
         reference_values = reference_values.reshape(-1, 1)
     if reference_times.size != reference_values.shape[0]:
         raise ValueError("reference_times_s length must match reference_values rows")
-    if reference_times.size < 2:
-        raise ValueError("at least two reference times are required for interpolation")
+    finite_reference = _finite_reference_rows(reference_times, reference_values)
+    if np.count_nonzero(finite_reference) < 2:
+        raise ValueError("at least two finite reference rows are required for interpolation")
+    reference_times = reference_times[finite_reference]
+    reference_values = reference_values[finite_reference]
     order = np.argsort(reference_times)
     reference_times = reference_times[order]
     reference_values = reference_values[order]
@@ -145,7 +167,11 @@ def interpolate_reference_values(
             for dim in range(reference_values.shape[1])
         ]
     )
-    valid = (query_times >= reference_times[0]) & (query_times <= reference_times[-1])
+    valid = (
+        np.isfinite(query_times)
+        & (query_times >= reference_times[0])
+        & (query_times <= reference_times[-1])
+    )
     if max_time_delta_s is not None:
         nearest = nearest_time_indices(reference_times, query_times)
         valid &= np.abs(reference_times[nearest] - query_times) <= float(

tests/calibration/test_time_offset_bias.py

@@ -38,6 +38,13 @@ def test_nearest_time_indices_accepts_unsorted_reference_times(self):
 
         npt.assert_array_equal(indices, np.array([1, 2, 0]))
 
+    def test_nearest_time_indices_ignores_nonfinite_reference_times(self):
+        indices = nearest_time_indices(
+            np.array([np.nan, 10.0, 0.0]), np.array([0.2, 9.0])
+        )
+
+        npt.assert_array_equal(indices, np.array([2, 1]))
+
     def test_aggregate_time_offset_sweeps_preserves_rmse_and_max_semantics(self):
         aggregated = aggregate_time_offset_sweeps(
             [
@@ -133,6 +140,27 @@ def test_interpolation_rejects_negative_max_time_delta(self):
                 max_time_delta_s=-1.0,
             )
 
+    def test_interpolation_skips_nonfinite_reference_rows(self):
+        interpolated, valid = interpolate_reference_values(
+            np.array([0.0, 1.0, 2.0, np.nan, 3.0]),
+            np.array([[0.0], [np.nan], [2.0], [99.0], [3.0]]),
+            np.array([0.5, 1.5, 2.5]),
+        )
+
+        npt.assert_allclose(interpolated, np.array([[0.5], [1.5], [2.5]]))
+        npt.assert_array_equal(valid, np.array([True, True, True]))
+
+    def test_interpolation_rejects_without_two_finite_reference_rows(self):
+        with self.assertRaisesRegex(
+            ValueError,
+            "at least two finite reference rows are required for interpolation",
+        ):
+            interpolate_reference_values(
+                np.array([0.0, 1.0]),
+                np.array([[0.0], [np.nan]]),
+                np.array([0.5]),
+            )
+
     def test_interpolation_rejects_scalar_reference_values(self):
         with self.assertRaisesRegex(
             ValueError,
@@ -193,6 +221,34 @@ def test_make_bias_training_examples_uses_nearest_reference(self):
 
         npt.assert_allclose(examples.residual, np.array([[1.0], [1.0]]))
 
+    def test_make_bias_training_examples_skips_nonfinite_reference_rows(self):
+        examples = make_bias_training_examples(
+            np.array([0.0, 1.0, 2.0]),
+            np.array([[1.0], [3.0], [5.0]]),
+            np.array([0.0, 1.0, 2.0, np.nan]),
+            np.array([[0.0], [np.nan], [4.0], [99.0]]),
+            max_time_delta_s=0.25,
+        )
+
+        npt.assert_allclose(examples.measured, np.array([[1.0], [5.0]]))
+        npt.assert_allclose(examples.reference, np.array([[0.0], [4.0]]))
+        npt.assert_allclose(examples.residual, np.array([[1.0], [1.0]]))
+        npt.assert_allclose(examples.time_delta_s, np.array([0.0, 0.0]))
+
+    def test_make_bias_training_examples_returns_empty_without_finite_reference_rows(
+        self,
+    ):
+        examples = make_bias_training_examples(
+            np.array([0.0, 1.0]),
+            np.array([[1.0], [2.0]]),
+            np.array([np.nan, 1.0]),
+            np.array([[0.0], [np.nan]]),
+            feature_values=np.array([[1.0], [2.0]]),
+        )
+
+        self.assertEqual(examples.measured.shape, (0, 1))
+        self.assertEqual(examples.features.shape, (0, 1))
+
     def test_make_bias_training_examples_validates_feature_rows_without_references(
         self,
     ):