Support initializer not in inputs (#396)

* support onnx file with initializer data not an input

* fix format

* fix review comments and format issue

* format

* reorder alphabetically

src/onnx/onnx.cpp

@@ -1422,21 +1422,14 @@ struct onnx_parser
     void parse_graph(const onnx::GraphProto& graph)
     {
         nodes = get_nodes(graph);
-        std::unordered_map<std::string, onnx::TensorProto> initializer_data;
         for(auto&& f : graph.initializer())
-        {
-            initializer_data[f.name()] = f;
-        }
+            instructions[f.name()] = prog.add_literal(parse_tensor(f));
+
         for(auto&& input : graph.input())
         {
             const std::string& name = input.name();
-            // Does the input have an initializer?
-            if(contains(initializer_data, name))
-            {
-                auto t             = initializer_data[name];
-                instructions[name] = prog.add_literal(parse_tensor(t));
-            }
-            else
+            // input not in initializer_data, so it is a real input
+            if(!contains(instructions, name))
             {
                 // TODO: Get shape of input parameter
                 shape s            = parse_type(input.type());

test/onnx/gen_onnx.py

@@ -838,6 +838,26 @@ def implicit_sub_bcast_test():
     return ([node], [arg0, arg1], [arg_out])
 
 
+@onnx_test
+def initializer_not_an_input():
+    values = np.array([[1, 2, 3, 4], [5, 6, 7, 8]])
+    w = helper.make_tensor(name='w',
+                           data_type=TensorProto.FLOAT,
+                           dims=values.shape,
+                           vals=values.flatten().astype(np.float))
+
+    x = helper.make_tensor_value_info('x', TensorProto.FLOAT, [5, 2])
+    y = helper.make_tensor_value_info('y', TensorProto.FLOAT, [5, 4])
+
+    node = onnx.helper.make_node(
+        'Gemm',
+        inputs=['x', 'w'],
+        outputs=['y'],
+    )
+
+    return ([node], [x], [y], [w])
+
+
 @onnx_test
 def leaky_relu_test():
     x = helper.make_tensor_value_info('0', TensorProto.FLOAT, [3])

test/onnx/onnx_test.cpp

@@ -48,8 +48,8 @@ TEST_CASE(add_fp16_test)
 TEST_CASE(add_scalar_test)
 {
     migraphx::program p;
-    auto l0 = p.add_parameter("0", migraphx::shape{migraphx::shape::float_type, {2, 3, 4, 5}});
     auto l1 = p.add_literal(migraphx::literal{migraphx::shape{migraphx::shape::float_type}, {1}});
+    auto l0 = p.add_parameter("0", migraphx::shape{migraphx::shape::float_type, {2, 3, 4, 5}});
     auto m0 = p.add_instruction(migraphx::op::multibroadcast{{2, 3, 4, 5}}, l0);
     auto m1 = p.add_instruction(migraphx::op::multibroadcast{{2, 3, 4, 5}}, l1);
     p.add_instruction(migraphx::op::add{}, m0, m1);
@@ -585,6 +585,19 @@ TEST_CASE(implicit_sub_bcast_test)
     EXPECT(p == prog);
 }
 
+TEST_CASE(initializer_not_an_input)
+{
+    migraphx::program p;
+    std::vector<float> w = {1, 2, 3, 4, 5, 6, 7, 8};
+    auto l1 = p.add_literal(migraphx::literal({migraphx::shape::float_type, {2, 4}}, w));
+    auto l0 = p.add_parameter("x", migraphx::shape{migraphx::shape::float_type, {5, 2}});
+    p.add_instruction(migraphx::op::dot{}, l0, l1);
+
+    auto prog = migraphx::parse_onnx("initializer_not_an_input.onnx");
+
+    EXPECT(p == prog);
+}
+
 TEST_CASE(leaky_relu_test)
 {
     migraphx::program p;
@@ -858,9 +871,9 @@ TEST_CASE(reshape_test)
     migraphx::program p;
     migraphx::op::reshape op;
     std::vector<int64_t> reshape_dims{3, 8};
-    auto l0 = p.add_parameter("0", migraphx::shape{migraphx::shape::float_type, {4, 2, 3}});
     p.add_literal(
         migraphx::literal{migraphx::shape{migraphx::shape::int64_type, {2}}, reshape_dims});
+    auto l0 = p.add_parameter("0", migraphx::shape{migraphx::shape::float_type, {4, 2, 3}});
     op.dims = reshape_dims;
     p.add_instruction(op, l0);
     p.add_instruction(op, l0);