Fixed bug in local_add_specialize that would create a zero ndarray w wrong number of dims.

theano/tensor/opt.py

@@ -2313,15 +2313,21 @@ def local_add_specialize(node):
                 y = get_constant_value(input)
             except TypeError:
                 y = input
-            if N.all(y == 0.0):
+            if numpy.all(y == 0.0):
                 continue
             new_inputs.append(input)
 
         if len(new_inputs) < len(node.inputs):
             if len(new_inputs) == 0:
                 #we got rid of the entire expression!
-                return fill_chain(T.TensorConstant(T.TensorType(dtype=node.outputs[0].type.dtype,
-                    broadcastable = [True] * node.outputs[0].ndim), N.asarray(0)))
+                ndim = node.outputs[0].type.ndim
+                dtype = node.outputs[0].type.dtype
+                return fill_chain(
+                        T.TensorConstant(
+                            T.TensorType(
+                                dtype=dtype,
+                                broadcastable = [True] * ndim),
+                            numpy.zeros((1,)*ndim, dtype=dtype)))
 
             if len(new_inputs) == 1:
                 return fill_chain(new_inputs[0])

theano/tensor/tests/test_opt.py

@@ -2174,3 +2174,15 @@ def test_local_mul_to_neg():
     aval = numpy.random.randint(0,10,(2,2)).astype('int32')
     assert f1(aval).dtype == a.dtype
     assert f2(aval).dtype == 'float64'
+
+def test_local_add_specialize():
+
+    # test of non-zero dimension
+    a = TT.vector()
+    s = TT.add(TT.zeros_like(a))
+    assert local_add_specialize.transform(s.owner)
+
+    # test of 0-d
+    a = TT.scalar()
+    s = TT.add(TT.zeros_like(a))
+    assert local_add_specialize.transform(s.owner)