Merge pull request #439 from lamblin/finish_careduce_dtype

theano/tensor/basic.py

@@ -2703,27 +2703,27 @@ def mean(input, axis=None, dtype=None, op=False):
                   necessarily be the dtype of the output (in particular
                   if it is a discrete (int/uint) dtype, the output will
                   be in a float type).
-                  If None, then we use float64 for a discrete input, and the
-                  same rules as `sum()` for a continuous input.
+                  If None, then we use the same rules as `sum()`.
     :type dtype: None or string
 
     :note: for gpu, if you specify dtype=float32, everything will be done
            on the gpu.
     """
     if op:
+        if dtype not in (None, 'float64'):
+            raise NotImplementedError(
+                    'The Mean op does not support the dtype argument, '
+                    'and will always use float64. If you want to specify '
+                    'the dtype, call tensor.mean(..., op=False).',
+                    dtype)
         return Mean(axis)(input)
 
     if dtype is not None:
         # The summation will be done with the specified dtype.
         # sum() will complain if it is not suitable.
         sum_dtype = dtype
-    elif input.dtype in discrete_dtypes:
-        # we need to cast eventually anyway, and this helps
-        # to prevents overflow. Numpy uses 'float64'.
-        # TODO: use floatX? let casting_policy decide?
-        sum_dtype = 'float64'
     else:
-        # Let sum() infer the appropriate dtype
+        # Let sum() infer the appropriate dtype.
         sum_dtype = None
 
     s = sum(input, axis=axis, dtype=sum_dtype)
@@ -2741,8 +2741,9 @@ def mean(input, axis=None, dtype=None, op=False):
         axis = range(input.ndim)
     elif isinstance(axis, int):
         axis = [axis]
+
     for i in axis:
-        s = s / shp[i]
+        s = true_div(s, shp[i])
 
     return s