Stabilize DVS polarity sign inference

src/pyrecest/experimental/dvs/normal_flow.py

@@ -2,6 +2,8 @@
 
 from __future__ import annotations
 
+import math
+
 import numpy as np
 
 from .active_contour import (
@@ -185,14 +187,23 @@ def infer_polarity_contrast_sign(
     if flows.shape != polarities.shape:
         raise ValueError("event_polarities must have one value per signed normal flow")
 
-    score = 0.0
+    contributions: list[float] = []
     for signed_flow, event_polarity in zip(flows, polarities, strict=True):
         if signed_scalar_sign(signed_flow, zero_tolerance=tolerance) == 0.0:
             continue
-        score += event_polarity_sign(event_polarity) * float(signed_flow)
-    if abs(score) <= tolerance:
+        contributions.append(
+            event_polarity_sign(event_polarity) * float(signed_flow)
+        )
+    if not contributions:
+        return 1.0
+
+    scale = max(abs(contribution) for contribution in contributions)
+    scaled_score = math.fsum(
+        contribution / scale for contribution in contributions
+    )
+    if abs(scaled_score) <= tolerance / scale:
         return 1.0
-    return 1.0 if score > 0.0 else -1.0
+    return 1.0 if scaled_score > 0.0 else -1.0
 
 
 def polarity_consistency_for_signed_flow(

tests/experimental/test_dvs_normal_flow.py

@@ -73,6 +73,17 @@ def test_infer_polarity_contrast_sign_from_batch():
     assert infer_polarity_contrast_sign(signed_flows, polarities, -1.0) == -1.0
 
 
+def test_infer_polarity_contrast_sign_avoids_overflow():
+    largest = np.finfo(float).max
+    signed_flows = largest * np.array([0.75, 0.75, -0.75, -0.9])
+    polarities = np.ones(signed_flows.shape)
+
+    with np.errstate(over="raise", invalid="raise"):
+        inferred = infer_polarity_contrast_sign(signed_flows, polarities, "infer")
+
+    assert inferred == -1.0
+
+
 def test_infer_polarity_contrast_sign_rejects_invalid_batches():
     with pytest.raises(ValueError, match="signed_normal_flows"):
         infer_polarity_contrast_sign([1.0, np.nan], [1.0, 0.0], "infer")