do not assume that by default you are dealing only with tensors

theano/gradient.py

@@ -250,10 +250,12 @@ def Rop(f, wrt, eval_points):
     for pack in enumerate(zip(wrt, eval_points)):
         i = pack[0]
         wrt_elem, eval_point = pack[1]
-
+        if not isinstance(wrt_elem, gof.Variable):
             wrt_elem = as_tensor_variable(wrt_elem)
+        if not isinstance(eval_point, gof.Variable):
             eval_point = as_tensor_variable(eval_point)
 
+        try:
             wrt_dim = len(wrt_elem.type.broadcastable)
             eval_dim = len(eval_point.type.broadcastable)
 
@@ -265,6 +267,10 @@ def Rop(f, wrt, eval_points):
                                  str(wrt_dim) +
                                  ' versus ' +
                                  str(eval_dim))
+        except:
+            # wrt_elem and eval_point can be non-tensor variable which do
+            # not have broadcastable flags
+            pass
 
     seen_nodes = {}