Prevent overflow in exponential link functions by clipping linear predictor

gen_imgs.py

@@ -323,7 +323,6 @@ def chicago_tensor():
 
 
 def expectiles():
-
     """Generate expectiles visualization."""
     X, y = mcycle(return_X_y=True)
 

pygam/links.py

@@ -118,7 +118,7 @@ def mu(self, lp, dist):
         -------
         mu : np.array of length n
         """
-        elp = np.exp(lp)
+        elp = np.exp(np.clip(lp, -500, 500))
         return dist.levels * elp / (elp + 1)
 
     def gradient(self, mu, dist):
@@ -178,7 +178,7 @@ def mu(self, lp, dist):
         -------
         mu : np.array of length n
         """
-        return np.exp(lp)
+        return np.exp(np.clip(lp, -500, 500))
 
     def gradient(self, mu, dist):
         """

pygam/tests/test_links.py

@@ -0,0 +1,21 @@
+import numpy as np
+
+from pygam.distributions import BinomialDist
+from pygam.links import LogitLink, LogLink
+
+
+def test_log_link_no_overflow():
+    """LogLink.mu should not overflow for large linear predictor values."""
+    link = LogLink()
+    lp = np.array([-1000.0, 0.0, 1000.0])
+    result = link.mu(lp, dist=None)
+    assert np.all(np.isfinite(result)), "LogLink.mu produced inf or nan"
+
+
+def test_logit_link_no_overflow():
+    """LogitLink.mu should not overflow for large linear predictor values."""
+    link = LogitLink()
+    dist = BinomialDist()
+    lp = np.array([-1000.0, 0.0, 1000.0])
+    result = link.mu(lp, dist)
+    assert np.all(np.isfinite(result)), "LogitLink.mu produced inf or nan"