add test for the change to MakeVector.

theano/tensor/opt.py

@@ -230,8 +230,14 @@ class MakeVector(T.Op):
             dtype=theano.scalar.upcast(self.dtype,*[i.dtype for i in inputs])
             #upcast the input to the determined dtype, but don't upcast downcast anything
             assert dtype==self.dtype, "Upcast the input of MakeVector to dtype gived in init without precissino loss only."
+            if not all(self.dtype == T.cast(i,dtype=dtype).dtype for a in inputs):
+                raise TypeError("MakeVector.make_node expected inputs upcastable to %s. got %s"%(
+                        self.dtype,
+                        str([i.dtype for i in inputs])
+                        ))
             inputs=[T.cast(i,dtype=dtype) for i in inputs]
-        assert all(a.type == inputs[0].type for a in inputs)
+        assert all(self.dtype == a.dtype for a in inputs)
+        assert all(a.ndim==0 for a in inputs)
 
         if inputs:
             dtype = inputs[0].type.dtype

theano/tensor/tests/test_opt.py

@@ -1538,7 +1538,7 @@ class T_local_sum(unittest.TestCase):
     def test_local_sum_all_to_none(self):
         a = T.tensor3()
         input=numpy.arange(3*3*3).reshape(3,3,3)
-        f = theano.function([a],a.sum()),mode=self.mode)
+        f = theano.function([a],a.sum(),mode=self.mode)
         assert len(f.maker.env.nodes)==1
         assert numpy.allclose(f(input),input.sum())
 
@@ -1632,6 +1632,33 @@ class T_local_sum_dimshuffle(unittest.TestCase):
     # test_local_sum_prod_dimshuffle (a * b * c)
     # test_local_sum_divprod_dimshuffle ((a * b) / (c * d))
 
+def test_make_vector_upcast():
+    b = T.bscalar()
+    i = T.iscalar()
+    d = T.dscalar()
+    
+    opt.MakeVector(dtype="int8")(b,b)
+    opt.MakeVector(dtype="int32")(i,b)
+    opt.MakeVector(dtype="int32")(b,i)
+    opt.MakeVector(dtype="float64")(b,i)
+    opt.MakeVector(dtype="float64")(b,d)
+    opt.MakeVector(dtype="float64")(d,i)
+
+    #should fail
+    for (dtype,inputs) in [("int8",(b,i)),
+                           ("int8",(i,b)),
+                           ("int8",(b,d)),
+                           ("int8",(i,i)),
+                           ("int32",(d,i)),
+                           ("int32",(i,d)),
+                           ("float32",(i,d)),
+                           ]:
+        try:
+            opt.MakeVector(dtype=dtype)(*inputs)
+            raise Exception("Theano should have raised an error")
+        except AssertionError:
+            pass
+
 if __name__ == '__main__':
 #    unittest.main()
     test_fusion().tes_memory_leak()