Categorical: preserve ints when NaN are present

`Categorical([1, np.nan])` would end up with a single `1.` float level.
This commit ensures that if `values` is a list of ints and contains np.nan,
the float conversation does not take place.

Author: jankatins

pandas/core/categorical.py

@@ -220,9 +220,17 @@ def __init__(self, values, levels=None, ordered=None, name=None, fastpath=False,
             inferred = com._possibly_infer_to_datetimelike(values)
             if not isinstance(inferred, np.ndarray):
                 from pandas.core.series import _sanitize_array
-                values = _sanitize_array(values, None)
+                safe_dtype = None
+                if isinstance(values, list) and np.nan in values:
+                    # On list with NaNs, int values will be converted to float. Use "object" dtype
+                    # to prvent this. In the end objects will be casted to int/... in the level
+                    # assignment step.
+                    safe_dtype = "object"
+                values = _sanitize_array(values, None, dtype=safe_dtype)
 
         if levels is None:
+            # object is needed to preserve ints in case we have np.nan in values
+            values = np.asarray(values, dtype="object")
             try:
                 codes, levels = factorize(values, sort=True)
                 # If the underlying data structure was sortable, and the user doesn't want to

pandas/tests/test_categorical.py

@@ -111,6 +111,18 @@ def test_constructor(self):
         cat = pd.Categorical([1,2,3,np.nan], levels=[1,2,3])
         self.assertTrue(com.is_integer_dtype(cat.levels))
 
+        # https://github.com/pydata/pandas/issues/3678
+        cat = pd.Categorical([np.nan,1, 2, 3])
+        self.assertTrue(com.is_integer_dtype(cat.levels))
+
+        # this should result in floats
+        cat = pd.Categorical([np.nan, 1, 2., 3 ])
+        self.assertTrue(com.is_float_dtype(cat.levels))
+
+        cat = pd.Categorical([np.nan, 1., 2., 3. ])
+        self.assertTrue(com.is_float_dtype(cat.levels))
+
+
     def test_from_codes(self):
 
         # too few levels