register opt local_sum_broadcastable only at specialize step

theano/tensor/opt.py

@@ -3144,7 +3144,12 @@ def local_cut_useless_reduce(node):
             return [summed]
 
 
-@register_canonicalize
+#Enabling this optimization at canonicalization step break this test:
+#theano/tensor/tests/test_opt.py:T_local_sum.test_local_sum_broadcast_some_0
+# see gh-790 issue.
+#
+#@register_canonicalize
+@register_specialize
 @gof.local_optimizer([])
 def local_sum_broadcastable(node):
     """Remove reduction over broadcastable dimensions"""
@@ -3177,6 +3182,7 @@ def local_sum_broadcastable(node):
                     # -- in this case we can remove the reduction completely
                     return [new_reduced.astype(odtype)]
 
+
 @register_specialize
 @gof.local_optimizer([])
 def local_sum_alloc(node):

theano/tensor/tests/test_opt.py

@@ -3188,7 +3188,8 @@ class test_local_remove_switch_const_cond(unittest.TestCase):
 
 class T_local_sum(unittest.TestCase):
     def setUp(self):
-        self.mode = theano.compile.get_default_mode().including('canonicalize')
+        self.mode = theano.compile.get_default_mode().including('canonicalize',
+                                                                'specialize')
 
     def test_local_sum_all_to_none(self):
         a = T.tensor3()