Remove Rebroadcast in the input of Join.

theano/tensor/basic.py

@@ -3506,13 +3506,7 @@ class Join(Op):
                 except IndexError:
                     raise ValueError('Join argument "axis" is out of range'
                                      ' (given input dimensions)')
-                as_tensor_variable_args = [unbroadcast(x, axis)
-                                           for x in as_tensor_variable_args]
             else:
-                # These unbroadcasts are for the gradient... not sure exactly
-                # why...
-                as_tensor_variable_args = [unbroadcast(x, *range(x.type.ndim))
-                                           for x in as_tensor_variable_args]
                 # When the axis may vary, no dimension can be guaranteed to be
                 # broadcastable.
                 bcastable = [False] * len(
@@ -3590,7 +3584,11 @@ class Join(Op):
             # If there is only one split, it might not be in a list.
             if not isinstance(split_gz, list):
                 split_gz = [split_gz]
-
+            # Split.make_node isn't always able to infer the right
+            # broadcast. As the grad need to keep the information,
+            # readd it if needed.
+            split_gz = [patternbroadcast(g, t.broadcastable)
+                        for t, g in zip(tensors, split_gz)]
             rval = rval + split_gz
         else:
             # the output has integer type, so the gradient through it