merge

theano/tensor/elemwise.py

@@ -686,12 +686,15 @@ class CAReduce(Op):
             self.axis = axis
         self.ufunc = numpy.frompyfunc(scalar_op.impl, 2, 1)
 
+    def _output_dtype(self, input_dtype):
+        return input_dtype
+
     def make_node(self, input):
         input = as_tensor_variable(input)
         axis = self.axis
         if axis is None:
             axis = range(len(input.type.broadcastable))
-        output = TensorType(dtype = input.type.dtype,
+        output = TensorType(dtype = self._output_dtype(input.type.dtype),
                              broadcastable = [x for i, x in enumerate(input.type.broadcastable) if i not in axis])()
         return Apply(self, [input], [output])
 
@@ -817,6 +820,13 @@ class Sum(CAReduce):
     def __init__(self, axis = None):
         CAReduce.__init__(self, scalar.add, axis)
 
+    def _output_dtype(self, idtype):
+        if idtype.startswith('int'):
+            return 'int64' #we want to protect against overflow
+        else:
+            return idtype
+
+
     def grad(self, (x, ), (gz, )):
         gz = as_tensor_variable(gz)
         axis = self.axis

theano/tensor/tests/test_basic.py

@@ -1887,6 +1887,12 @@ def test_var():
     f = function([a], var(a, axis=2))
     assert numpy.allclose(numpy.var(a_val, axis=2), f(a_val))
 
+def test_sum_overflow():
+    """Ensure that overflow errors are a little bit harder to get"""
+    a = Tensor(dtype='int8', broadcastable=[False])()
+    f = function([a], sum(a))
+    assert f([1]*300) == 300
+
 if __name__ == '__main__':
     if len(sys.argv) >= 2 and sys.argv[1] == 'OPT':
         default_mode = compile.Mode(linker = 'c&py',