Addes backprop to BatchDot op, allows fusion in training. (#2297)

* batch dot bprop WIP.

* WIP.

* testing.

* clean up debug code.

* comments and var name change.

* clean up.

* format style, batch dot differentiable pass.

* removed debug output.

* added unit test to autodiff, refactored make_function -> make_function_from_file.

* fixed build warning.

* fixed gpu build error.

* clang format fix.

* all test_tools.cpp to find SERIALIZED_ZOO

* remove cmake redef.

* fix unused macro.

* making test cpu only.

* testing build var

* macro test

* verbose makefile test

* style fix

* verbose make

* test/util needs test/models.

* removed debug output.

* refactor fusion type.

* refactor fusion type.

CMakeLists.txt

@@ -360,6 +360,7 @@ if (NGRAPH_UNIT_TEST_ENABLE)
     add_subdirectory(test)
     message(STATUS "unit tests enabled")
 else()
+    add_subdirectory(test/models)
     add_subdirectory(test/util)
     message(STATUS "unit tests disabled")
 endif()

src/ngraph/pass/pass.hpp

@@ -34,6 +34,14 @@ namespace ngraph
         class NodePass;
         class CallGraphPass;
         class Manager;
+        enum FusionType
+        {
+            //`DIFFERENTIABLE_FUSIONS` produce ops that support autodiff
+            // i.e. implement `generate_adjoints`
+            DIFFERENTIABLE_FUSIONS = 0x1,
+            REGULAR_FUSIONS = 0x2,
+            ALL_FUSIONS = 0xFFFFFFFF
+        };
     }
 }
 

src/ngraph/runtime/cpu/op/batch_dot.cpp

@@ -16,6 +16,7 @@
 
 #include "batch_dot.hpp"
 #include "ngraph/log.hpp"
+#include "ngraph/op/reshape.hpp"
 #include "ngraph/util.hpp"
 
 using namespace std;
@@ -69,3 +70,41 @@ op::BatchDot::BatchDot(shared_ptr<Node> a, shared_ptr<Node> b, bool transpose_a,
 
     set_output_type(0, a->get_element_type(), dot_shape);
 }
+
+void op::BatchDot::generate_adjoints(autodiff::Adjoints& adjoints, const NodeVector& deltas)
+{
+    auto delta = deltas.at(0); // NxIxK
+
+    auto a = get_inputs().at(0).get_output().get_node(); // NxIxJ (maybe transposed)
+    auto b = get_inputs().at(1).get_output().get_node(); // NxJxK (maybe transposed)
+
+    auto batch_transpose = [](const shared_ptr<Node>& node) {
+        const auto& batch_shape = node->get_shape();
+        // index 0 is the batch, only transposing the others.
+        AxisVector input_order{0, 2, 1};
+        Shape output_shape{batch_shape[0], batch_shape[2], batch_shape[1]};
+        return make_shared<op::Reshape>(node, input_order, output_shape);
+    };
+
+    // if b is already transposed, it does not need to be transposed again
+    auto delta_dot_b = make_shared<op::BatchDot>(delta, b, false, !m_transpose_b); // IK.KJ->IJ
+    // if a is transposed, the result need to be transposed to match original a shape.
+    if (m_transpose_a)
+    {
+        adjoints.add_delta(a, batch_transpose(delta_dot_b));
+    }
+    else
+    {
+        adjoints.add_delta(a, delta_dot_b);
+    }
+
+    auto a_dot_delta = make_shared<BatchDot>(a, delta, !m_transpose_a, false); // JI.IK->JK
+    if (m_transpose_b)
+    {
+        adjoints.add_delta(b, batch_transpose(a_dot_delta));
+    }
+    else
+    {
+        adjoints.add_delta(b, a_dot_delta);
+    }
+}

src/ngraph/runtime/cpu/op/batch_dot.hpp

@@ -37,6 +37,9 @@ namespace ngraph
             virtual std::shared_ptr<Node>
                 copy_with_new_args(const NodeVector& new_args) const override;
 
+            virtual void generate_adjoints(autodiff::Adjoints& adjoints,
+                                           const NodeVector& deltas) override;
+
         private:
             bool m_transpose_a;
             bool m_transpose_b;

src/ngraph/runtime/cpu/op/conv_bias.hpp

@@ -53,7 +53,8 @@ namespace ngraph
             virtual std::shared_ptr<Node>
                 copy_with_new_args(const NodeVector& new_args) const override;
 
-            void generate_adjoints(autodiff::Adjoints& adjoints, const NodeVector& deltas) override;
+            virtual void generate_adjoints(autodiff::Adjoints& adjoints,
+                                           const NodeVector& deltas) override;
 
         protected:
             Strides m_window_movement_strides;

src/ngraph/runtime/cpu/op/group_conv.hpp

@@ -49,7 +49,8 @@ namespace ngraph
             virtual std::shared_ptr<Node>
                 copy_with_new_args(const NodeVector& new_args) const override;
 
-            void generate_adjoints(autodiff::Adjoints& adjoints, const NodeVector& deltas) override;
+            virtual void generate_adjoints(autodiff::Adjoints& adjoints,
+                                           const NodeVector& deltas) override;
 
         protected:
             Strides m_window_movement_strides;

src/ngraph/runtime/cpu/pass/cpu_fusion.hpp

@@ -35,27 +35,16 @@ namespace ngraph
 class ngraph::runtime::cpu::pass::CPUFusion : public ngraph::pass::GraphRewrite
 {
 public:
-    // 30 different fusion groups that we can nest/mix&match/etc
-    // should be good enough for quite a while
-    enum fusions
-    {
-        //`DIFFERENTIABLE_FUSIONS` produce ops that support autodiff
-        // i.e. implement `generate_adjoints`
-        DIFFERENTIABLE_FUSIONS = 0x1,
-        REGULAR_FUSIONS = 0x2,
-        ALL = 0xFFFFFFFF
-    };
-
-    CPUFusion(int fusions = ALL)
+    CPUFusion(ngraph::pass::FusionType fusions = ngraph::pass::ALL_FUSIONS)
         : GraphRewrite()
     {
-        if (fusions & DIFFERENTIABLE_FUSIONS)
+        if (fusions & ngraph::pass::DIFFERENTIABLE_FUSIONS)
         {
             construct_conv_bias();
             construct_sigmoid_multiply();
         }
 
-        if (fusions & REGULAR_FUSIONS)
+        if (fusions & ngraph::pass::REGULAR_FUSIONS)
         {
             construct_matmul();
             construct_matmulbias();

src/ngraph/runtime/cpu/pass/cpu_mat_fusion.cpp

@@ -578,16 +578,21 @@ bool runtime::cpu::pass::CPUBatchFusion::run_on_function(std::shared_ptr<Functio
         const Node& node = *n;
         if (TI(node) == TI(op::Concat))
         {
-            auto fused_node = fuse_batch_dot(n);
-            if (fused_node)
+            if (m_fusion_type & ngraph::pass::DIFFERENTIABLE_FUSIONS)
             {
-                func->replace_node(n, fused_node);
-                modified = true;
+                if (auto fused_node = fuse_batch_dot(n))
+                {
+                    func->replace_node(n, fused_node);
+                    modified = true;
+                }
             }
-            else if (auto fused_conv = fuse_group_convolution(n))
+            if (m_fusion_type & ngraph::pass::REGULAR_FUSIONS)
             {
-                func->replace_node(n, fused_conv);
-                modified = true;
+                if (auto fused_conv = fuse_group_convolution(n))
+                {
+                    func->replace_node(n, fused_conv);
+                    modified = true;
+                }
             }
         }
     }

src/ngraph/runtime/cpu/pass/cpu_mat_fusion.hpp

@@ -29,12 +29,22 @@ namespace ngraph
                 class CPURnnMatFusion : public ngraph::pass::FunctionPass
                 {
                 public:
-                    bool run_on_function(std::shared_ptr<ngraph::Function> function) override;
+                    virtual bool
+                        run_on_function(std::shared_ptr<ngraph::Function> function) override;
                 };
                 class CPUBatchFusion : public ngraph::pass::FunctionPass
                 {
                 public:
-                    bool run_on_function(std::shared_ptr<ngraph::Function> function) override;
+                    CPUBatchFusion(ngraph::pass::FusionType type = ngraph::pass::ALL_FUSIONS)
+                        : FunctionPass()
+                        , m_fusion_type(type)
+                    {
+                    }
+                    virtual bool
+                        run_on_function(std::shared_ptr<ngraph::Function> function) override;
+
+                private:
+                    ngraph::pass::FusionType m_fusion_type;
                 };
             }
         }

test/autodiff.in.cpp

@@ -21,13 +21,18 @@
 
 #include "gtest/gtest.h"
 
+// clang-format off
+#define AUTODIFF_BACKEND_${BACKEND_NAME}
+// clang-format on
+
 #include "ngraph/ngraph.hpp"
+#include "ngraph/pass/manager.hpp"
+#include "ngraph/runtime/cpu/pass/cpu_mat_fusion.hpp"
 #include "ngraph/runtime/reference/avg_pool.hpp"
 #include "util/autodiff/backprop_function.hpp"
 #include "util/autodiff/numeric_compare.hpp"
 #include "util/random.hpp"
 #include "util/test_control.hpp"
-
 using namespace std;
 using namespace ngraph;
 
@@ -847,6 +852,30 @@ NGRAPH_TEST(${BACKEND_NAME}, backwards_dot_tensor3_tensor3)
     EXPECT_TRUE(autodiff_numeric_compare<float>(backend.get(), make_graph, {x0, x1}, .01f, .01f));
 }
 
+#ifdef AUTODIFF_BACKEND_CPU
+NGRAPH_TEST(${BACKEND_NAME}, backwards_batchdot_tensor2_tensor2)
+{
+    auto backend = runtime::Backend::create("${BACKEND_NAME}");
+    std::string backend_name = "${BACKEND_NAME}";
+
+    const std::string file_name("mxnet/batch_dot_3.json");
+    auto f = make_function_from_file(file_name);
+
+    test::Uniform<float> rng(-1.0f, 1.0f);
+    std::vector<std::shared_ptr<ngraph::runtime::Tensor>> args;
+    for (shared_ptr<op::Parameter> param : f->get_parameters())
+    {
+        args.push_back(rng.initialize(backend->create_tensor<float>(param->get_shape())));
+    }
+
+    auto g = make_function_from_file(file_name);
+    pass::Manager pass_manager;
+    pass_manager.register_pass<runtime::cpu::pass::CPUBatchFusion>();
+    pass_manager.run_passes(g);
+    EXPECT_TRUE(autodiff_numeric_compare<float>(backend.get(), f, g, args, .01f, .01f));
+}
+#endif
+
 NGRAPH_TEST(${BACKEND_NAME}, backwards_exp)
 {
     auto backend = runtime::Backend::create("${BACKEND_NAME}");
@@ -1760,3 +1789,9 @@ NGRAPH_TEST(${BACKEND_NAME}, backwards_reverse_sequence_n4d2c3h2w2)
     backend->call_with_validate(handle, {da, db}, {a, b, c});
     ASSERT_EQ(read_vector<int>(da), expected);
 }
+
+// clang-format off
+#ifdef AUTODIFF_BACKEND_${BACKEND_NAME}
+#undef AUTODIFF_BACKEND_${BACKEND_NAME}
+#endif
+// clang-format on

test/cpu_fusion.cpp

@@ -278,8 +278,7 @@ TEST(cpu_fusion, cpu_fusion_pass_basic)
     auto add = dot + broadcast;
     auto graph = make_shared<op::Abs>(add);
     pass::Manager pass_manager;
-    pass_manager.register_pass<runtime::cpu::pass::CPUFusion>(
-        runtime::cpu::pass::CPUFusion::REGULAR_FUSIONS);
+    pass_manager.register_pass<runtime::cpu::pass::CPUFusion>(pass::REGULAR_FUSIONS);
     auto func = make_shared<Function>(graph, ParameterVector{A, B, C});
     pass_manager.run_passes(func);
     ASSERT_NE(std::dynamic_pointer_cast<op::MatmulBias>(graph->get_argument(0)), nullptr);
@@ -300,8 +299,7 @@ TEST(cpu_fusion, commutative_matmul_bias)
     auto add = broadcast + dot;
     auto graph = make_shared<op::Abs>(add);
     pass::Manager pass_manager;
-    pass_manager.register_pass<runtime::cpu::pass::CPUFusion>(
-        runtime::cpu::pass::CPUFusion::REGULAR_FUSIONS);
+    pass_manager.register_pass<runtime::cpu::pass::CPUFusion>(pass::REGULAR_FUSIONS);
     auto func = make_shared<Function>(graph, ParameterVector{A, B, C});
     pass_manager.run_passes(func);
     ASSERT_NE(std::dynamic_pointer_cast<op::MatmulBias>(graph->get_argument(0)), nullptr);
@@ -323,8 +321,7 @@ TEST(cpu_fusion, cpu_fusion_pass_matmul_bias)
 
     auto graph = make_shared<op::Abs>(add);
     pass::Manager pass_manager;
-    pass_manager.register_pass<runtime::cpu::pass::CPUFusion>(
-        runtime::cpu::pass::CPUFusion::REGULAR_FUSIONS);
+    pass_manager.register_pass<runtime::cpu::pass::CPUFusion>(pass::REGULAR_FUSIONS);
     auto func = make_shared<Function>(graph, ParameterVector{W, x, b});
     pass_manager.run_passes(func);
     auto gmm = graph->get_argument(0);
@@ -345,8 +342,7 @@ TEST(cpu_fusion, cpu_fusion_pass_matmul_no_bias)
     auto graph = make_shared<op::Abs>(re_dot);
 
     pass::Manager pass_manager;
-    pass_manager.register_pass<runtime::cpu::pass::CPUFusion>(
-        runtime::cpu::pass::CPUFusion::REGULAR_FUSIONS);
+    pass_manager.register_pass<runtime::cpu::pass::CPUFusion>(pass::REGULAR_FUSIONS);
     auto func = make_shared<Function>(graph, ParameterVector{W, x});
     pass_manager.run_passes(func);
     size_t mmb = count_ops_of_type<op::MatmulBias>(func);
@@ -360,8 +356,7 @@ TEST(cpu_fusion, gemm_mlp)
     stringstream ss(json_string);
     shared_ptr<Function> func = ngraph::deserialize(ss);
     pass::Manager pass_manager;
-    pass_manager.register_pass<runtime::cpu::pass::CPUFusion>(
-        runtime::cpu::pass::CPUFusion::REGULAR_FUSIONS);
+    pass_manager.register_pass<runtime::cpu::pass::CPUFusion>(pass::REGULAR_FUSIONS);
     pass_manager.run_passes(func);
     auto mmbs = count_ops_of_type<op::MatmulBias>(func);
     ASSERT_EQ(mmbs, 3);
@@ -372,8 +367,7 @@ TEST(cpu_fusion, fuse_fprop_bn)
     pass::Manager pass_manager;
     pass_manager.register_pass<pass::VisualizeTree>("bn_fprop_before_fusion.png");
     pass_manager.register_pass<ngraph::pass::ReshapeElimination>();
-    pass_manager.register_pass<runtime::cpu::pass::CPUFusion>(
-        runtime::cpu::pass::CPUFusion::REGULAR_FUSIONS);
+    pass_manager.register_pass<runtime::cpu::pass::CPUFusion>(pass::REGULAR_FUSIONS);
     pass_manager.register_pass<pass::VisualizeTree>("bn_fprop_after_fusion.png");
     const string json_path = file_util::path_join(SERIALIZED_ZOO, "mxnet/bn_fprop_b2c3h2w2.json");
     const string json_string = file_util::read_file_to_string(json_path);
@@ -503,8 +497,7 @@ TEST(cpu_fusion, fuse_conv_bias)
 {
     pass::Manager pass_manager;
     pass_manager.register_pass<ngraph::pass::ReshapeElimination>();
-    pass_manager.register_pass<runtime::cpu::pass::CPUFusion>(
-        runtime::cpu::pass::CPUFusion::DIFFERENTIABLE_FUSIONS);
+    pass_manager.register_pass<runtime::cpu::pass::CPUFusion>(pass::DIFFERENTIABLE_FUSIONS);
     const string json_path = file_util::path_join(SERIALIZED_ZOO, "conv_bias.json");
     const string json_string = file_util::read_file_to_string(json_path);
     stringstream ss(json_string);
@@ -851,8 +844,7 @@ TEST(cpu_fusion, fuse_conv_relu)
     auto func = make_shared<Function>(abs_node, ParameterVector{A, weights});
 
     pass::Manager pass_manager;
-    pass_manager.register_pass<runtime::cpu::pass::CPUFusion>(
-        runtime::cpu::pass::CPUFusion::REGULAR_FUSIONS);
+    pass_manager.register_pass<runtime::cpu::pass::CPUFusion>(pass::REGULAR_FUSIONS);
     pass_manager.run_passes(func);
     size_t cb = count_ops_of_type<op::ConvolutionRelu>(func);
     ASSERT_GT(cb, 0);
@@ -1280,8 +1272,7 @@ std::vector<shared_ptr<runtime::Tensor>> rnn_matrix_fusion_eval(const size_t tim
     {
         pass::Manager pass_manager;
         pass_manager.register_pass<runtime::cpu::pass::CPURnnMatFusion>();
-        pass_manager.register_pass<runtime::cpu::pass::CPUFusion>(
-            runtime::cpu::pass::CPUFusion::REGULAR_FUSIONS);
+        pass_manager.register_pass<runtime::cpu::pass::CPUFusion>(pass::REGULAR_FUSIONS);
         pass_manager.run_passes(func);
         // check all of our dot/add are converted to a single MatmulBias op.
         size_t count = count_ops_of_type<op::MatmulBias>(func);
@@ -1340,8 +1331,7 @@ TEST(cpu_fusion, rnn_fusion_from_json_model)
 {
     pass::Manager pass_manager;
     pass_manager.register_pass<runtime::cpu::pass::CPURnnMatFusion>();
-    pass_manager.register_pass<runtime::cpu::pass::CPUFusion>(
-        runtime::cpu::pass::CPUFusion::REGULAR_FUSIONS);
+    pass_manager.register_pass<runtime::cpu::pass::CPUFusion>(pass::REGULAR_FUSIONS);
     const string json_path =
         file_util::path_join(SERIALIZED_ZOO, "mxnet/rnn-10-step-fusion-test.json");
     const string json_string = file_util::read_file_to_string(json_path);
@@ -2301,20 +2291,11 @@ TEST(cpu_fusion, fuse_1_layer_rnn)
     }
 }
 
-static std::shared_ptr<Function> make_function(const std::string& file_name)
-{
-    const string json_path = file_util::path_join(SERIALIZED_ZOO, file_name);
-    const string json_string = file_util::read_file_to_string(json_path);
-    stringstream ss(json_string);
-    shared_ptr<Function> func = ngraph::deserialize(ss);
-    return func;
-}
-
 TEST(cpu_fusion, rnn_fusion_1lstm_cell)
 {
     const std::string file_name("mxnet/1_lstm_cell_forward.json");
-    auto cpu_f = make_function(file_name);
-    auto int_f = make_function(file_name);
+    auto cpu_f = make_function_from_file(file_name);
+    auto int_f = make_function_from_file(file_name);
     test::Uniform<float> rng(-1.0f, 1.0f);
     vector<vector<float>> args;
 
@@ -2335,8 +2316,8 @@ TEST(cpu_fusion, rnn_fusion_1lstm_cell)
 TEST(cpu_fusion, rnn_fusion_1rnn_layer_3lstm_cell)
 {
     const std::string file_name("mxnet/1rnn_layer_3lstm_cell.json");
-    auto cpu_f = make_function(file_name);
-    auto int_f = make_function(file_name);
+    auto cpu_f = make_function_from_file(file_name);
+    auto int_f = make_function_from_file(file_name);
     test::Uniform<float> rng(-1.0f, 1.0f);
     vector<vector<float>> args;
 
@@ -2357,8 +2338,8 @@ TEST(cpu_fusion, rnn_fusion_1rnn_layer_3lstm_cell)
 TEST(cpu_fusion, rnn_fusion_2rnn_layer_3lstm_cell)
 {
     const std::string file_name("mxnet/2rnn_layer_3lstm_cell.json");
-    auto cpu_f = make_function(file_name);
-    auto int_f = make_function(file_name);
+    auto cpu_f = make_function_from_file(file_name);
+    auto int_f = make_function_from_file(file_name);
     test::Uniform<float> rng(-1.0f, 1.0f);
     vector<vector<float>> args;
 
@@ -3016,8 +2997,8 @@ TEST(cpu_fusion, fuse_batch_dot_forward)
     pass_manager.register_pass<runtime::cpu::pass::CPUBatchFusion>();
 
     const std::string file_name("mxnet/batch_dot_3.json");
-    auto cpu_f = make_function(file_name);
-    auto int_f = make_function(file_name);
+    auto cpu_f = make_function_from_file(file_name);
+    auto int_f = make_function_from_file(file_name);
     pass_manager.run_passes(cpu_f);
     test::Uniform<float> rng(0.0f, 1.0f);
     vector<vector<float>> args;
@@ -3036,11 +3017,42 @@ TEST(cpu_fusion, fuse_batch_dot_forward)
     }
 }
 
+TEST(cpu_fusion, fuse_batch_dot_backward)
+{
+    const std::string file_name("mxnet/batch_dot_3.json");
+    auto cpu_f = make_function_from_file(file_name);
+    auto int_f = make_function_from_file(file_name);
+
+    pass::Manager pass_manager;
+    pass_manager.register_pass<runtime::cpu::pass::CPUBatchFusion>();
+    pass_manager.run_passes(cpu_f);
+
+    auto int_df = autodiff::backprop_function(int_f);
+    auto cpu_df = autodiff::backprop_function(cpu_f);
+
+    test::Uniform<float> rng(-1.0f, 1.0f);
+    vector<vector<float>> args;
+    for (shared_ptr<op::Parameter> param : cpu_df->get_parameters())
+    {
+        vector<float> tensor_val(shape_size(param->get_shape()));
+        rng.initialize(tensor_val);
+        args.push_back(tensor_val);
+    }
+
+    auto int_results = execute(int_df, args, "INTERPRETER");
+    auto cpu_results = execute(cpu_df, args, "CPU");
+
+    for (size_t i = 0; i < cpu_results.size(); i++)
+    {
+        EXPECT_TRUE(test::all_close(cpu_results.at(i), int_results.at(i), 1.0e-4f, 1.0e-4f));
+    }
+}
+
 TEST(cpu_fusion, fuse_rnn_across_layer_2layer_3timestep)
 {
     const std::string file_name("mxnet/2layer_3timestep_ic100oc100.json");
-    auto cpu_f = make_function(file_name);
-    auto int_f = make_function(file_name);
+    auto cpu_f = make_function_from_file(file_name);
+    auto int_f = make_function_from_file(file_name);
     test::Uniform<float> rng(-1.0f, 1.0f);
     vector<vector<float>> args;
 

test/gpu_fusion.cpp

@@ -183,15 +183,6 @@ TEST(DISABLED_gpu_fusion, fuse_1_layer_rnn)
     }
 }
 
-static std::shared_ptr<Function> make_function(const std::string& file_name)
-{
-    const string json_path = file_util::path_join(SERIALIZED_ZOO, file_name);
-    const string json_string = file_util::read_file_to_string(json_path);
-    stringstream ss(json_string);
-    shared_ptr<Function> func = ngraph::deserialize(ss);
-    return func;
-}
-
 TEST(gpu_fusion, lstm_analytic)
 {
     auto input_xt = std::make_shared<op::Parameter>(element::f32, Shape{1, 1});
@@ -432,8 +423,8 @@ TEST(gpu_fusion, fuse_2_layer_rnn_1lstm_analytic)
 TEST(gpu_fusion, rnn_fusion_inter_vs_gpu_1lstm_cell)
 {
     const std::string file_name("mxnet/1_lstm_cell_forward.json");
-    auto gpu_f = make_function(file_name);
-    auto int_f = make_function(file_name);
+    auto gpu_f = make_function_from_file(file_name);
+    auto int_f = make_function_from_file(file_name);
     test::Uniform<float> rng(-10.0f, 10.0f);
     vector<vector<float>> args;
 
@@ -454,8 +445,8 @@ TEST(gpu_fusion, rnn_fusion_inter_vs_gpu_1lstm_cell)
 TEST(DISABLED_gpu_fusion, rnn_fusion_inter_vs_gpu_1rnn_layer_3lstm_cell)
 {
     const std::string file_name("mxnet/1rnn_layer_3lstm_cell.json");
-    auto gpu_f = make_function(file_name);
-    auto int_f = make_function(file_name);
+    auto gpu_f = make_function_from_file(file_name);
+    auto int_f = make_function_from_file(file_name);
     test::Uniform<float> rng(-10.0f, 10.0f);
     vector<vector<float>> args;
 
@@ -476,8 +467,8 @@ TEST(DISABLED_gpu_fusion, rnn_fusion_inter_vs_gpu_1rnn_layer_3lstm_cell)
 TEST(gpu_fusion, rnn_fusion_inter_vs_gpu_2rnn_layer_3lstm_cell)
 {
     const std::string file_name("mxnet/2rnn_layer_3lstm_cell.json");
-    auto gpu_f = make_function(file_name);
-    auto int_f = make_function(file_name);
+    auto gpu_f = make_function_from_file(file_name);
+    auto int_f = make_function_from_file(file_name);
     test::Uniform<float> rng(-10.0f, 10.0f);
     vector<vector<float>> args;
 
@@ -516,8 +507,8 @@ TEST(gpu_fusion, fuse_rnn_across_layer)
 TEST(gpu_fusion, fuse_rnn_across_2layer_1timestep)
 {
     const std::string file_name("mxnet/2rnn_layer_1timestep.json");
-    auto gpu_f = make_function(file_name);
-    auto int_f = make_function(file_name);
+    auto gpu_f = make_function_from_file(file_name);
+    auto int_f = make_function_from_file(file_name);
     test::Uniform<float> rng(-10.0f, 10.0f);
     vector<vector<float>> args;
 

test/util/test_tools.cpp

@@ -301,3 +301,12 @@ string
 
     return ss.str();
 }
+
+std::shared_ptr<Function> make_function_from_file(const std::string& file_name)
+{
+    const string json_path = file_util::path_join(SERIALIZED_ZOO, file_name);
+    const string json_string = file_util::read_file_to_string(json_path);
+    stringstream ss(json_string);
+    shared_ptr<Function> func = ngraph::deserialize(ss);
+    return func;
+}

test/util/test_tools.hpp

@@ -38,6 +38,7 @@ namespace ngraph
 
 bool validate_list(const std::list<std::shared_ptr<ngraph::Node>>& nodes);
 std::shared_ptr<ngraph::Function> make_test_graph();
+std::shared_ptr<ngraph::Function> make_function_from_file(const std::string& file_name);
 
 template <typename T>
 void copy_data(std::shared_ptr<ngraph::runtime::Tensor> tv, const std::vector<T>& data)