Check graph outputs are defined (#6083)

The current checker does not check that all graph outputs are defined.
This PR adds such a check. (See issue
#6040 ).

Note: We probably also want a separate check to ensure that a
graph-input is not used as a graph-output. But leaving that for a
separate PR, since it is a distinctly different check (and perhaps might
require a different discussion).

Redo of PR #6070

Signed-off-by: Ganesan Ramalingam <grama@microsoft.com>

onnx/checker.cc

@@ -709,6 +709,12 @@ void check_graph(const GraphProto& graph, const CheckerContext& ctx, const Lexic
       lex_ctx.add(output);
     }
   }
+  for (const auto& value_info : graph.output()) {
+    if (!lex_ctx.this_graph_has(value_info.name())) {
+      fail_check("Graph output '", value_info.name(), "' is not an output of any node in graph.");
+    }
+  }
+
   print_warning_if_has_experimental(used_experimental_ops);
 }
 

onnx/test/checker_test.py

@@ -797,7 +797,7 @@ def test_loop_with_different_initializer_input_below_ir4(self) -> None:
                             ],
                             outputs=[
                                 helper.make_tensor_value_info(
-                                    "cond___while_Identity_graph_outputs_Identity__3_0",
+                                    "cond___while_Less__13_0",
                                     TensorProto.BOOL,
                                     shape=[],
                                 ),
@@ -825,7 +825,7 @@ def test_loop_with_different_initializer_input_below_ir4(self) -> None:
                                     outputs=["cond___while_Less__13_0"],
                                     name="cond___while_Less__13",
                                     domain="",
-                                    to=TensorProto.FLOAT,
+                                    to=TensorProto.BOOL,
                                 ),
                             ],
                         ),
@@ -1076,6 +1076,39 @@ def test_check_model_supports_unicode_path(self):
             onnx.save(model, unicode_model_path)
             checker.check_model(unicode_model_path, full_check=True)
 
+    def test_graph_output_is_defined(self):
+        model = onnx.parser.parse_model(
+            """
+            <ir_version: 7, opset_import: [ "" : 17]>
+            agraph (float[N] x) => (float[N] y, float[N] z)
+            {
+                y = Add(x, x)
+            }
+            # Error: z is not defined
+        """
+        )
+        self.assertRaises(checker.ValidationError, checker.check_model, model)
+
+    def test_graph_output_is_defined_within_sub_graph(self):
+        model = onnx.parser.parse_model(
+            """
+            <ir_version: 7, opset_import: [ "" : 17]>
+            agraph (float[N] x, bool cond) => (float[N] y)
+            {
+                sum = Add (x, x)
+                prod = Mul (x, x)
+                y = If (cond) <
+                    then_branch = then_graph () => (sum) {},
+                    else_branch = else_graph () => (prod) {}
+                >
+            }
+            # Error: sum/prod are accessible inside if-then-else branches, but cannot
+            # be used as outputs of the then/else branch implicitly.
+            # An explicit "Identity(sum)" must be used to return sum as output.
+        """
+        )
+        self.assertRaises(checker.ValidationError, checker.check_model, model)
+
 
 if __name__ == "__main__":
     unittest.main()

onnx/test/compose_test.py

@@ -87,7 +87,7 @@ def _make_sparse_tensor(name: str) -> SparseTensorProto:
     {
         C0 = Add(B01, B11)
         C1 = Sub(B11, B21)
-        M1 = Mul(C0, C1)
+        D0 = Mul(C0, C1)
     }
     """