Graph should only have one (input) kParam node (onnx#1088)

that has all graph level outputs as its outputs

Author: bddppq

onnx/optimizer/passes/extract_constant_to_initializer.h

@@ -28,18 +28,18 @@ struct ExtractConstantToInitializer final : public OptimizePass {
       if (n->kind() == kConstant) {
         const auto name = n->output()->uniqueName();
         Tensor t = n->t(kvalue);
+
+        // add a new graph input
+        Value* input = graph.addInput();
+        input->setUniqueName(name);
+        input->setSizes({t.sizes().begin(), t.sizes().end()});
+        input->setElemType(t.elem_type());
+        n->output()->replaceAllUsesWith(input);
+
+        // copy the tensor to initializer
         t.setName(name);
-        std::vector<Dimension> tsizes;
-        for (auto v : t.sizes()) {
-          tsizes.push_back(v);
-        }
-        Node* param = graph.create(kParam, 1);
-        param->output()->setUniqueName(name);
-        param->output()->setSizes(tsizes);
-        param->output()->setElemType(t.elem_type());
         graph.addInitializer(std::move(t), name);
-        graph.addInput()->copyMetadata(param->output());
-        n->replaceAllUsesWith(param);
+
         it.destroyCurrent();
       }
     }

onnx/test/optimizer_test.py

@@ -276,14 +276,19 @@ def test_extract_constant_to_initializer(self):  # type: () -> None
             [helper.make_tensor_value_info("X", TensorProto.FLOAT, (1, 5, 3, 3)),
              helper.make_tensor_value_info("Y", TensorProto.FLOAT, (16, 5, 3, 3))],
             [helper.make_tensor_value_info("B", TensorProto.FLOAT, (1, 16, 3, 3))],
-            value_info=[
-                helper.make_tensor_value_info("A", TensorProto.FLOAT, (16, 1, 1)),
-            ]
         )
         optimized_model = self._optimized(graph, ["extract_constant_to_initializer"])
-        assert len(list(optimized_model.graph.initializer)) == 1
-        assert len(list(optimized_model.graph.node)) == 2
-        assert "A" in [i.name for i in optimized_model.graph.initializer]
+        self.assertEqual(
+            set(vi.name for vi in optimized_model.graph.input),
+            {'X', 'Y', 'A'})
+
+        self.assertEqual(len(optimized_model.graph.initializer), 1)
+        init = optimized_model.graph.initializer[0]
+        self.assertEqual(init.name, 'A')
+        self.assertEqual(init.dims, [16])
+        self.assertEqual(init.data_type, TensorProto.FLOAT)
+
+        self.assertEqual([n.op_type for n in optimized_model.graph.node], ['Conv', 'Add'])
 
     def test_fuse_transpose(self):  # type: () -> None
         nodes = [helper.make_node("Transpose", ["X"], ["Y"], perm=[1, 0, 2]),