scikit-learn · rth · Feb 3, 2021 · Feb 2, 2021 · Feb 2, 2021 · Feb 2, 2021
diff --git a/sklearn/utils/arrayfuncs.pyx b/sklearn/utils/arrayfuncs.pyx
@@ -19,35 +19,26 @@ np.import_array()
 
 
 def min_pos(np.ndarray X):
-   """
-   Find the minimum value of an array over positive values
-
-   Returns a huge value if none of the values are positive
-   """
-   if X.dtype.name == 'float32':
-      return _float_min_pos(<float *> X.data, X.size)
-   elif X.dtype.name == 'float64':
-      return _double_min_pos(<double *> X.data, X.size)
-   else:
-      raise ValueError('Unsupported dtype for array X')
-
-
-cdef float _float_min_pos(float *X, Py_ssize_t size):
-   cdef Py_ssize_t i
-   cdef float min_val = DBL_MAX
-   for i in range(size):
-      if 0. < X[i] < min_val:
-         min_val = X[i]
-   return min_val
-
-
-cdef double _double_min_pos(double *X, Py_ssize_t size):
-   cdef Py_ssize_t i
-   cdef np.float64_t min_val = FLT_MAX
-   for i in range(size):
-      if 0. < X[i] < min_val:
-         min_val = X[i]
-   return min_val
+    """Find the minimum value of an array over positive values
+
+    Returns the maximum representable value of the input dtype if none of the
+    values are positive.
+    """
+    if X.dtype == np.float32:
+        return _min_pos[float](<float *> X.data, X.size)
+    elif X.dtype == np.float64:
+        return _min_pos[double](<double *> X.data, X.size)
+    else:
+        raise ValueError('Unsupported dtype for array X')
+
+
+cdef floating _min_pos(floating* X, Py_ssize_t size):
+    cdef Py_ssize_t i
+    cdef floating min_val = FLT_MAX if floating is float else DBL_MAX
+    for i in range(size):
+        if 0. < X[i] < min_val:
+            min_val = X[i]
+    return min_val
 
 
 # General Cholesky Delete.

diff --git a/sklearn/utils/tests/test_arrayfuncs.py b/sklearn/utils/tests/test_arrayfuncs.py
@@ -0,0 +1,26 @@
+import pytest
+import numpy as np
+
+from sklearn.utils._testing import assert_allclose
+from sklearn.utils.arrayfuncs import min_pos
+
+
+def test_min_pos():
+    # Check that min_pos returns a positive value and that it's consistent
+    # between float and double
+    X = np.random.RandomState(0).randn(100)
+
+    min_double = min_pos(X)
+    min_float = min_pos(X.astype(np.float32))
+
+    assert_allclose(min_double, min_float)
+    assert min_double >= 0
+
+
+@pytest.mark.parametrize("dtype", [np.float32, np.float64])
+def test_min_pos_no_positive(dtype):
+    # Check that the return value of min_pos is the maximum representable
+    # value of the input dtype when all input elements are <= 0 (#19328)
+    X = np.full(100, -1.).astype(dtype, copy=False)
+
+    assert min_pos(X) == np.finfo(dtype).max