Implemented Pascals comments. Improved test for local_elemwise_sub_zeros.

theano/tensor/opt.py

@@ -4299,9 +4299,6 @@ def local_useless_reshape(node):
     # This could hide errors if the user provides inconsistent shapes.
     if (input.ndim == 1 and output.ndim == 1 and
             input.broadcastable == output.broadcastable):
-        # Copy over stack trace
-        copy_stack_trace(node.outputs[0], input)
-
         return [input]
 
     # Second case: all the shapes match the input shape
@@ -4309,9 +4306,6 @@ def local_useless_reshape(node):
     if output_shape.owner and isinstance(output_shape.owner.op, Shape):
         shape_input = output_shape.owner.inputs[0]
         if shape_input == input:
-            # Copy over stack trace
-            copy_stack_trace(node.outputs[0], input)
-
             return [input]
 
     # Match Reshape(x, [x.shape[0], ..., x.shape[-1]]), accounting for
@@ -4362,9 +4356,6 @@ def local_useless_reshape(node):
                     continue
 
         if all(shape_match):
-            # Copy over stack trace
-            copy_stack_trace(node.outputs[0], input)
-
             return [input]
 
         # TODO later: if all the shapes except one match, we may want to

theano/tensor/tests/test_opt.py

@@ -3502,6 +3502,8 @@ def test_local_elemwise_sub_zeros():
     f = function([scalar], scalar - scalar, mode=mode)
     # Check optimized graph is correct
     assert isinstance(f.maker.fgraph.toposort()[0].op, T.Elemwise)
+    assert f.maker.fgraph.toposort()[0].op.name\
+        == 'Elemwise{second,no_inplace}'
     assert isinstance(f.maker.fgraph.toposort()[0].inputs[1],
                       T.TensorConstant) or\
         isinstance(f.maker.fgraph.toposort()[0].inputs[1],
@@ -3514,6 +3516,8 @@ def test_local_elemwise_sub_zeros():
     f = function([vect], vect - vect, mode=mode)
     # Check optimized graph is correct
     assert isinstance(f.maker.fgraph.toposort()[0].op, T.Elemwise)
+    assert f.maker.fgraph.toposort()[0].op.name\
+        == 'Elemwise{second,no_inplace}'
     assert isinstance(f.maker.fgraph.toposort()[0].inputs[1],
                       T.TensorConstant) or\
         isinstance(f.maker.fgraph.toposort()[0].inputs[1],
@@ -3526,6 +3530,8 @@ def test_local_elemwise_sub_zeros():
     f = function([mat], mat - mat, mode=mode)
     # Check optimized graph is correct
     assert isinstance(f.maker.fgraph.toposort()[0].op, T.Elemwise)
+    assert f.maker.fgraph.toposort()[0].op.name\
+        == 'Elemwise{second,no_inplace}'
     assert isinstance(f.maker.fgraph.toposort()[0].inputs[1],
                       T.TensorConstant) or\
         isinstance(f.maker.fgraph.toposort()[0].inputs[1],
@@ -5416,8 +5422,6 @@ class T_local_sum_prod(unittest.TestCase):
                            (s1_val * s2_val * v_val * m_val).prod(), 2)
 
     def test_local_sum_prod_all_to_none(self):
-        # Julian: It appears that the opt local_sum_prod_mul_by_scalar
-        #         is never used in any of these tests...
         a = T.tensor3()
         input = numpy.arange(3 * 4 * 5, dtype=config.floatX).reshape(3, 4, 5)
         # test sum
@@ -5652,10 +5656,8 @@ class T_local_sum_prod(unittest.TestCase):
         finally:
             config.on_opt_error = backup
 
-    def test_stack_trace(self):
-        """
-        Test that stack trace is copied over correctly.
-        """
+    def test_local_sum_prod_mul_by_scalar_stack_trace(self):
+        # Test that stack trace is copied over correctly for local_sum_prod_mul_by_scalar.
         m0 = theano.compile.get_default_mode()\
             .excluding('inplace_elemwise_opt')\
             .including('canonicalize', 'specialize')