Safer assert_all_finite.

Fabian Pedregosa · larsmans · commit 71953c493486 · 2011-09-20T15:32:48.000+02:00
Check for Inf/NaN when X.sum() is not finite. This avoids the risk of false positives because of sum overflow while remaining performant for the usual case where all values are finite. This idea is taken from: tecki/numpy@57167f7
diff --git a/sklearn/utils/__init__.py b/sklearn/utils/__init__.py
@@ -2,18 +2,16 @@
 import scipy.sparse as sp
 import warnings
 
-_FLOAT_CODES = np.typecodes['AllFloat']
-
-
 def assert_all_finite(X):
     """Throw a ValueError if X contains NaN or infinity.
     Input MUST be an np.ndarray instance or a scipy.sparse matrix."""
 
-    # O(n) time, O(1) solution. XXX: will fail if the sum over X is
-    # *extremely* large. A proper solution would be a C-level loop to check
-    # each element.
-    if X.dtype.char in _FLOAT_CODES and not np.isfinite(X.sum()):
-        raise ValueError("array contains NaN or infinity")
+    # First try an O(n) time, O(1) space solution for the common case that
+    # there everything is finite; fall back to O(n) space np.isfinite to
+    # prevent false positives from overflow in sum method.
+    if X.dtype.char in np.typecodes['AllFloat'] and not np.isfinite(X.sum()) \
+      and not np.isfinite(X).all():
+            raise ValueError("array contains NaN or infinity")
 
 
 def safe_asanyarray(X, dtype=None, order=None):
