Merge branch 'master' into krovatkin/ml_reuse_debug

CODEOWNERS

@@ -6,11 +6,11 @@
 
 /cmake/                             @rkimballn1 @silee2
 
-/.ci/                               @aslepko @crlishka
-/.ci/travis/                        @postrational
-/.ci/onnx/                          @postrational
-/contrib/docker/                    @aslepko @crlishka
-/.travis.yml                        @postrational
+/.ci/                               @aslepko
+/.ci/travis/                        @aslepko @postrational
+/.ci/onnx/                          @aslepko @postrational
+/contrib/docker/                    @aslepko
+/.travis.yml                        @aslepko @postrational
 
 /.clang-format                      @rkimballn1
 /.gitattributes                     @rkimballn1

python/CMakeLists.txt

@@ -42,3 +42,7 @@ add_custom_command(
 )
 
 add_custom_target(python_wheel DEPENDS ngraph ${CMAKE_BINARY_DIR}/python/dist/)
+if (NGRAPH_CPU_ENABLE)
+    add_dependencies(python_wheel ext_mkldnn)
+endif()
+

python/ngraph/__init__.py

@@ -15,8 +15,12 @@
 # ******************************************************************************
 """ngraph module namespace, exposing factory functions for all ops and other classes."""
 
-from pkg_resources import get_distribution
-__version__ = get_distribution('ngraph-core').version
+from pkg_resources import get_distribution, DistributionNotFound
+
+try:
+    __version__ = get_distribution('ngraph-core').version
+except DistributionNotFound:
+    __version__ = '0.0.0-dev'
 
 from ngraph.ops import absolute
 from ngraph.ops import absolute as abs

python/setup.py

@@ -374,6 +374,10 @@ class BuildExt(build_ext):
         build_ext.build_extensions(self)
 
 
+if sys.platform == 'darwin':
+    # This turns out to be needed when building using Anaconda python on macOS.
+    os.environ['MACOSX_DEPLOYMENT_TARGET'] = '10.9'
+
 with open(os.path.join(PYNGRAPH_ROOT_DIR, 'requirements.txt')) as req:
     requirements = req.read().splitlines()
 

src/ngraph/frontend/onnx_import/onnx.cpp

@@ -99,6 +99,14 @@ namespace ngraph
             return op_list;
         }
 
+        bool is_operator_supported(const std::string& op_name,
+                                   std::int64_t version,
+                                   const std::string& domain)
+        {
+            return OperatorsBridge::is_operator_registered(
+                op_name, version, domain == "ai.onnx" ? "" : domain);
+        }
+
     } // namespace onnx_import
 
 } // namespace ngraph

src/ngraph/frontend/onnx_import/onnx.hpp

@@ -52,6 +52,18 @@ namespace ngraph
         std::set<std::string> get_supported_operators(std::int64_t version,
                                                       const std::string& domain);
 
+        /// \brief      Determines whether ONNX operator is supported.
+        ///
+        /// \param[in]  op_name  The ONNX operator name.
+        /// \param[in]  version  The ONNX operator set version.
+        /// \param[in]  domain   The domain the ONNX operator is registered to.
+        ///
+        /// \return     True if operator is supported, False otherwise.
+        ///
+        bool is_operator_supported(const std::string& op_name,
+                                   std::int64_t version,
+                                   const std::string& domain = "ai.onnx");
+
         /// \brief Convert an ONNX model to nGraph function
         /// The function translated serialized ONNX model to nGraph function. The serialized
         /// ONNX model is read from input stream.

src/ngraph/frontend/onnx_import/ops_bridge.cpp

@@ -21,6 +21,7 @@
 #include <unordered_map>
 
 #include "core/attribute.hpp"
+#include "ngraph/log.hpp"
 #include "op/abs.hpp"
 #include "op/acos.hpp"
 #include "op/add.hpp"
@@ -102,20 +103,19 @@ namespace ngraph
     {
         namespace detail
         {
-            const Operator& find(const std::string& name,
-                                 std::int64_t version,
-                                 const std::string& domain,
-                                 const std::map<std::int64_t, Operator>& map)
+            const std::map<std::int64_t, Operator>::const_iterator
+                find(std::int64_t version, const std::map<std::int64_t, Operator>& map)
             {
+                std::map<std::int64_t, Operator>::const_iterator it{};
                 while (version > 0)
                 {
-                    const auto it = map.find(version--);
+                    it = map.find(version--);
                     if (it != std::end(map))
                     {
-                        return it->second;
+                        return it;
                     }
                 }
-                throw error::UnsupportedVersion{name, version, domain};
+                return it;
             }
         }
 
@@ -136,13 +136,51 @@ namespace ngraph
             {
                 throw error::UnknownDomain{domain};
             }
+            if (version > OperatorsBridge::LATEST_SUPPORTED_OPSET_VERSION)
+            {
+                NGRAPH_WARN << "Currently operator set version: " << version << " is unsupported."
+                            << " Falling back to: "
+                            << OperatorsBridge::LATEST_SUPPORTED_OPSET_VERSION;
+            }
             for (const auto& op : dm->second)
             {
-                result.emplace(op.first, detail::find(op.first, version, domain, op.second));
+                const auto& it = detail::find(version, op.second);
+                if (it == std::end(op.second))
+                {
+                    throw error::UnsupportedVersion{op.first, version, domain};
+                }
+                result.emplace(op.first, it->second);
             }
             return result;
         }
 
+        bool OperatorsBridge::_is_operator_registered(const std::string& name,
+                                                      std::int64_t version,
+                                                      const std::string& domain)
+        {
+            // search for domain
+            auto dm_map = m_map.find(domain);
+            if (dm_map == std::end(m_map))
+            {
+                return false;
+            }
+            // search for name
+            auto op_map = dm_map->second.find(name);
+            if (op_map == std::end(dm_map->second))
+            {
+                return false;
+            }
+
+            if (detail::find(version, op_map->second) != std::end(op_map->second))
+            {
+                return true;
+            }
+            else
+            {
+                return false;
+            }
+        }
+
 #define REGISTER_OPERATOR(name_, ver_, fn_)                                                        \
     m_map[""][name_].emplace(ver_, std::bind(op::set_##ver_::fn_, std::placeholders::_1))
 

src/ngraph/frontend/onnx_import/ops_bridge.hpp

@@ -62,6 +62,8 @@ namespace ngraph
         class OperatorsBridge
         {
         public:
+            static constexpr const int LATEST_SUPPORTED_OPSET_VERSION = ONNX_OPSET_VERSION;
+
             OperatorsBridge(const OperatorsBridge&) = delete;
             OperatorsBridge& operator=(const OperatorsBridge&) = delete;
             OperatorsBridge(OperatorsBridge&&) = delete;
@@ -80,6 +82,13 @@ namespace ngraph
                 instance()._register_operator(name, version, domain, std::move(fn));
             }
 
+            static bool is_operator_registered(const std::string& name,
+                                               std::int64_t version,
+                                               const std::string& domain)
+            {
+                return instance()._is_operator_registered(name, version, domain);
+            }
+
         private:
             std::unordered_map<std::string,
                                std::unordered_map<std::string, std::map<std::int64_t, Operator>>>
@@ -98,6 +107,9 @@ namespace ngraph
                                     const std::string& domain,
                                     Operator fn);
             OperatorSet _get_operator_set(std::int64_t version, const std::string& domain);
+            bool _is_operator_registered(const std::string& name,
+                                         std::int64_t version,
+                                         const std::string& domain);
         };
 
     } // namespace onnx_import

src/ngraph/pass/reshape_sinking.cpp

@@ -29,22 +29,30 @@
 #include "ngraph/op/convolution.hpp"
 #include "ngraph/op/dequantize.hpp"
 #include "ngraph/op/get_output_element.hpp"
+#include "ngraph/op/pad.hpp"
 #include "ngraph/op/quantize.hpp"
 #include "ngraph/op/reshape.hpp"
+#include "ngraph/op/slice.hpp"
 #include "ngraph/op/util/binary_elementwise_arithmetic.hpp"
 #include "ngraph/op/util/unary_elementwise_arithmetic.hpp"
+#include "ngraph/pattern/op/label.hpp"
 #include "ngraph/util.hpp"
 
 using namespace ngraph;
 
-extern template ngraph::AxisVector
-    ngraph::apply_permutation<ngraph::AxisVector>(ngraph::AxisVector input,
-                                                  ngraph::AxisVector order);
+using ReshapeMap = std::unordered_map<std::shared_ptr<Node>, std::shared_ptr<op::Reshape>>;
 
-extern template ngraph::Shape ngraph::apply_permutation<ngraph::Shape>(ngraph::Shape input,
-                                                                       ngraph::AxisVector order);
+static std::string describe_reshape(std::shared_ptr<Node> node)
+{
+    std::stringstream ss;
+    auto reshape = std::dynamic_pointer_cast<op::Reshape>(node);
+    ss << reshape->get_name()
+       << " ( axis order = " << ngraph::vector_to_string(reshape->get_input_order())
+       << " , shape = " << vector_to_string(reshape->get_shape()) << " ) "
+       << " , child = " << reshape->get_argument(0)->get_name();
 
-using ReshapeMap = std::unordered_map<std::shared_ptr<Node>, std::shared_ptr<op::Reshape>>;
+    return ss.str();
+}
 
 static std::shared_ptr<op::Reshape> combine_reshapes(std::shared_ptr<op::Reshape> r1,
                                                      std::shared_ptr<op::Reshape> r2)
@@ -64,18 +72,6 @@ static void
     target->get_inputs().at(input_index).replace_output(new_reshape->get_outputs().at(0));
 }
 
-std::string describe_reshape(std::shared_ptr<Node> node)
-{
-    std::stringstream ss;
-    auto reshape = std::dynamic_pointer_cast<op::Reshape>(node);
-    ss << reshape->get_name()
-       << " ( axis order = " << ngraph::vector_to_string(reshape->get_input_order())
-       << " , shape = " << vector_to_string(reshape->get_shape()) << " ) "
-       << " , child = " << reshape->get_argument(0)->get_name();
-
-    return ss.str();
-}
-
 static void delete_reshape(std::shared_ptr<Node> reshape)
 {
     NGRAPH_DEBUG << "Removing reshape " << reshape->get_name();
@@ -256,6 +252,7 @@ static void sink_reshape(std::shared_ptr<op::Reshape> reshape,
         mark_reshape_for_deletion(orig_reshape, reshapes_to_delete);
         //replace reshape with combined one
         ngraph::replace_node(reshape, new_reshape);
+        mark_reshape_for_deletion(new_reshape, reshapes_to_delete);
         reorders[new_reshape] = new_reshape;
         NGRAPH_DEBUG << "Combining " << describe_reshape(orig_reshape) << " and"
                      << describe_reshape(reshape) << " into  " << describe_reshape(new_reshape);
@@ -309,6 +306,61 @@ static void sink_binary(std::shared_ptr<op::util::BinaryElementwiseArithmetic> b
     }
 }
 
+static void sink_slice(std::shared_ptr<op::Slice> n,
+                       ReshapeMap& reorders,
+                       std::set<std::shared_ptr<Node>>& reshapes_to_delete)
+{
+    auto arg_reshape = reorders.at(n->get_argument(0));
+    auto order = arg_reshape->get_input_order();
+
+    // we need the correct input shape to produce the right output shape
+    // we are going to create a label of the right input shape,
+    // so a new slice will have the right shape
+    auto def_order = ngraph::get_permutation_to_default_order(order);
+    auto input_shape = ngraph::apply_permutation(arg_reshape->get_shape(), def_order);
+    auto dummy_correct_shape =
+        std::make_shared<pattern::op::Label>(arg_reshape->get_element_type(), input_shape);
+
+    auto new_lower = ngraph::apply_permutation(n->get_lower_bounds(), def_order);
+    auto new_upper = ngraph::apply_permutation(n->get_upper_bounds(), def_order);
+    auto new_strides = ngraph::apply_permutation(n->get_strides(), def_order);
+    auto new_slice =
+        std::make_shared<op::Slice>(dummy_correct_shape, new_lower, new_upper, new_strides);
+    ngraph::replace_node(dummy_correct_shape, n->get_argument(0));
+    NGRAPH_DEBUG << "Replacing " << n->get_name() << " with " << new_slice->get_name();
+    ngraph::replace_node(n, new_slice);
+
+    auto new_reshape = std::make_shared<op::Reshape>(new_slice, order, n->get_shape());
+    NGRAPH_DEBUG << "Propagating " << describe_reshape(new_reshape) << " for " << n->get_name();
+    reorders[new_slice] = new_reshape;
+}
+
+static void sink_pad(std::shared_ptr<op::Pad> n,
+                     ReshapeMap& reorders,
+                     std::set<std::shared_ptr<Node>>& reshapes_to_delete)
+{
+    auto arg_reshape = reorders.at(n->get_argument(0));
+    auto order = arg_reshape->get_input_order();
+    // we need the correct input shape to produce the right output shape
+    // we are going to create a label of the right input shape,
+    // so a new pad will have the right shape
+    auto def_order = ngraph::get_permutation_to_default_order(order);
+    auto input_shape = ngraph::apply_permutation(arg_reshape->get_shape(), def_order);
+    auto dummy_correct_shape =
+        std::make_shared<pattern::op::Label>(arg_reshape->get_element_type(), input_shape);
+
+    auto new_lower = ngraph::apply_permutation(n->get_padding_below(), def_order);
+    auto new_upper = ngraph::apply_permutation(n->get_padding_above(), def_order);
+    auto new_interior = ngraph::apply_permutation(n->get_padding_interior(), def_order);
+    auto new_pad = std::make_shared<op::Pad>(
+        dummy_correct_shape, n->get_argument(1), new_lower, new_upper, new_interior);
+    ngraph::replace_node(dummy_correct_shape, n->get_argument(0));
+    NGRAPH_DEBUG << "Replacing " << n->get_name() << " with " << new_pad->get_name();
+    ngraph::replace_node(n, new_pad);
+    auto new_reshape = std::make_shared<op::Reshape>(new_pad, order, n->get_shape());
+    NGRAPH_DEBUG << "Propagating " << describe_reshape(new_reshape) << " for " << n->get_name();
+    reorders[new_pad] = new_reshape;
+}
 static void sink_quantize(std::shared_ptr<op::Quantize> quantize,
                           ReshapeMap& reorders,
                           std::set<std::shared_ptr<Node>>& reshapes_to_delete)
@@ -419,6 +471,14 @@ bool ngraph::pass::ReshapeSinking::run_on_function(std::shared_ptr<ngraph::Funct
         {
             sink_dequantize(dequantize, reorders, reshapes_to_delete);
         }
+        else if (auto slice = std::dynamic_pointer_cast<op::Slice>(n))
+        {
+            sink_slice(slice, reorders, reshapes_to_delete);
+        }
+        else if (auto pad = std::dynamic_pointer_cast<op::Pad>(n))
+        {
+            sink_pad(pad, reorders, reshapes_to_delete);
+        }
         else
         {
             materialize_shapes(n, reorders, reshapes_to_delete);

src/ngraph/pass/reshape_sinking.hpp

@@ -17,6 +17,7 @@
 #pragma once
 
 #include "ngraph/pass/pass.hpp"
+#include "ngraph/util.hpp"
 
 namespace ngraph
 {
@@ -29,3 +30,17 @@ namespace ngraph
         };
     }
 }
+
+extern template ngraph::AxisVector
+    ngraph::apply_permutation<ngraph::AxisVector>(ngraph::AxisVector input,
+                                                  ngraph::AxisVector order);
+
+extern template ngraph::Coordinate
+    ngraph::apply_permutation<ngraph::Coordinate>(ngraph::Coordinate input,
+                                                  ngraph::AxisVector order);
+
+extern template ngraph::Strides
+    ngraph::apply_permutation<ngraph::Strides>(ngraph::Strides input, ngraph::AxisVector order);
+
+extern template ngraph::Shape ngraph::apply_permutation<ngraph::Shape>(ngraph::Shape input,
+                                                                       ngraph::AxisVector order);

src/ngraph/runtime/cpu/cpu_backend.cpp

@@ -145,3 +145,8 @@ vector<runtime::PerformanceCounter>
     }
     return rc;
 }
+
+bool runtime::cpu::CPU_Backend::is_supported(const Node& op) const
+{
+    return true;
+}

src/ngraph/runtime/cpu/cpu_backend.hpp

@@ -58,6 +58,8 @@ namespace ngraph
                 std::vector<PerformanceCounter>
                     get_performance_data(std::shared_ptr<Function> func) const override;
 
+                bool is_supported(const Node& node) const override;
+
             private:
                 class FunctionInstance
                 {

src/ngraph/runtime/cpu/cpu_external_function.cpp

@@ -1982,10 +1982,10 @@ void runtime::cpu::CPU_ExternalFunction::build()
                         file_util::path_join(s_debug_dir, m_function_name + "_debug.txt");
                     std::stringstream ss;
 
-                    ss << "EXECUTION PLAN:\n";
+                    ss << "\nEXECUTION PLAN:\n";
                     for (size_t i = 0; i < functors.size(); i++)
                     {
-                        ss << op_names.at(i) << "will be executed with the following inputs:\n";
+                        ss << op_names.at(i) << " will be executed with the following inputs:\n";
                         for (auto is : this->m_op_attrs.at(i).Inputs)
                         {
                             ss << "\t" << is << " = " << this->get_tensor_data(is) << std::endl;

src/ngraph/runtime/hybrid/hybrid_backend.cpp

@@ -18,13 +18,10 @@
 #include "ngraph/graph_util.hpp"
 #include "ngraph/pass/manager.hpp"
 #include "ngraph/pass/visualize_tree.hpp"
-#include "ngraph/runtime/gpu/gpu_backend.hpp"
-#include "ngraph/runtime/gpu/gpu_tensor.hpp"
 #include "ngraph/runtime/host_tensor.hpp"
 #include "ngraph/runtime/hybrid/hybrid_util.hpp"
 #include "ngraph/runtime/hybrid/pass/assign_placement.hpp"
 #include "ngraph/runtime/hybrid/pass/fix_get_output_element.hpp"
-#include "ngraph/runtime/interpreter/int_backend.hpp"
 #include "ngraph/runtime/tensor.hpp"
 
 using namespace ngraph;
@@ -205,32 +202,6 @@ bool runtime::hybrid::HybridBackend::is_supported(const Node& node) const
     return true;
 }
 
-string runtime::hybrid::HybridBackend::get_placement_name(const runtime::Tensor* t)
-{
-    string rc;
-    if (dynamic_cast<const runtime::HostTensor*>(t) != nullptr)
-    {
-        rc = "HostTensor";
-    }
-    else if (dynamic_cast<const runtime::gpu::GPUTensor*>(t) != nullptr)
-    {
-        rc = "GPUTensor";
-    }
-    return rc;
-}
-string runtime::hybrid::HybridBackend::get_placement_name(const runtime::Backend* t)
-{
-    string rc;
-    if (dynamic_cast<const runtime::interpreter::INTBackend*>(t) != nullptr)
-    {
-        rc = "INTBackend";
-    }
-    else if (dynamic_cast<const runtime::gpu::GPU_Backend*>(t) != nullptr)
-    {
-        rc = "GPU_Backend";
-    }
-    return rc;
-}
 size_t runtime::hybrid::HybridBackend::get_placement(const runtime::Tensor* t)
 {
     size_t index = 0;

src/ngraph/runtime/hybrid/hybrid_backend.hpp

@@ -70,7 +70,5 @@ private:
     std::map<std::shared_ptr<ngraph::Function>, FunctionInstance> m_function_map;
     std::vector<std::shared_ptr<runtime::Backend>> m_backend_list;
 
-    std::string get_placement_name(const runtime::Tensor* t);
-    std::string get_placement_name(const runtime::Backend* t);
     size_t get_placement(const runtime::Tensor* t);
 };

src/ngraph/util.cpp

@@ -478,6 +478,10 @@ T ngraph::apply_permutation(T input, AxisVector order)
 
 template AxisVector ngraph::apply_permutation<AxisVector>(AxisVector input, AxisVector order);
 template Shape ngraph::apply_permutation<Shape>(Shape input, AxisVector order);
+template ngraph::Coordinate ngraph::apply_permutation<ngraph::Coordinate>(ngraph::Coordinate input,
+                                                                          ngraph::AxisVector order);
+template ngraph::Strides ngraph::apply_permutation<ngraph::Strides>(ngraph::Strides input,
+                                                                    ngraph::AxisVector order);
 
 AxisVector ngraph::get_default_order(const Shape& shape)
 {

test/CMakeLists.txt

@@ -65,13 +65,6 @@ set(SRC
 set_source_files_properties(includes.cpp PROPERTIES COMPILE_DEFINITIONS
     NGRAPH_INCLUDES="${PROJECT_SOURCE_DIR}/src/ngraph")
 
-if (NGRAPH_ONNX_IMPORT_ENABLE)
-    list(APPEND SRC onnx_import.cpp)
-    if (NGRAPH_ONNXIFI_ENABLE)
-        list(APPEND SRC onnxifi.cpp onnxifi_span.cpp)
-    endif()
-endif()
-
 if (NGRAPH_INTERPRETER_ENABLE)
     list(APPEND SRC
         backend_debug_api.cpp
@@ -140,13 +133,22 @@ set(MULTI_TEST_SRC
     backend_unary_elementwise.in.cpp
     convolution_test.in.cpp
 )
+
 if(NGRAPH_DISTRIBUTED_ENABLE)
     list(APPEND MULTI_TEST_SRC distributed.in.cpp)
 endif()
+
 if (NGRAPH_CPU_ENABLE)
     list(APPEND MULTI_TEST_SRC backend_graph_comparison.in.cpp)
 endif()
 
+if (NGRAPH_ONNX_IMPORT_ENABLE)
+    list(APPEND MULTI_TEST_SRC onnx_import.in.cpp)
+    if (NGRAPH_ONNXIFI_ENABLE)
+        list(APPEND SRC onnxifi.cpp onnxifi_span.cpp)
+    endif()
+endif()
+
 foreach(BACKEND_NAME ${ACTIVE_BACKEND_LIST})
     # Some---but not all---autodiff tests go through multiple iterations with
     # different random seeds. On the CPU backend this is currently very slow

test/reshape_sinking.cpp

@@ -163,3 +163,43 @@ TEST(reshape_sinking, nasnet_pooladd)
     size_t before_after = count_ops_of_type<op::Reshape>(func);
     ASSERT_LE(before_after, before_count);
 }
+
+TEST(reshape_sinking, slice_pad)
+{
+    Shape shape_a{100, 8, 8, 1};
+
+    AxisVector to_nhwc{0, 2, 3, 1};
+    AxisVector to_nchw{0, 3, 1, 2};
+
+    auto A = make_shared<op::Parameter>(element::f32, shape_a);
+    auto pad_value = op::Constant::create<float>(element::f32, Shape{}, std::vector<float>{0.0f});
+
+    Shape padding_below{0, 0, 0, 0};
+    Shape padding_above{0, 1, 1, 0};
+    Shape padding_interior{0, 0, 0, 0};
+
+    auto reshape1 = make_shared<op::Reshape>(A, to_nchw, Shape{100, 1, 8, 8});
+    auto maxpool =
+        make_shared<op::MaxPool>(reshape1, Shape{1, 1}, Strides{2, 2}, Shape{0, 0}, Shape{0, 0});
+    auto reshape2 = make_shared<op::Reshape>(maxpool, to_nhwc, Shape{100, 4, 4, 1});
+    auto pad =
+        make_shared<op::Pad>(reshape2, pad_value, padding_below, padding_above, padding_interior);
+    auto slice = make_shared<op::Slice>(
+        pad, Coordinate{0, 1, 1, 0}, Coordinate{100, 5, 5, 1}, Strides{1, 1, 1, 1});
+
+    auto reshape3 = make_shared<op::Reshape>(slice, to_nchw, Shape{100, 1, 4, 4});
+    auto avgpool = make_shared<op::AvgPool>(reshape3, Shape{1, 1}, Strides{2, 2});
+    auto reshape4 = make_shared<op::Reshape>(avgpool, to_nhwc, Shape{100, 1, 2, 2});
+    auto f = make_shared<Function>(reshape4, ParameterVector{A});
+
+    pass::Manager pass_manager;
+    size_t before_count = count_ops_of_type<op::Reshape>(f);
+    pass_manager.register_pass<pass::VisualizeTree>("before.pdf");
+    pass_manager.register_pass<pass::ReshapeSinking>();
+    pass_manager.register_pass<pass::ReshapeElimination>();
+    pass_manager.register_pass<pass::CommonSubexpressionElimination>();
+    pass_manager.register_pass<pass::VisualizeTree>("after.pdf");
+    pass_manager.run_passes(f);
+    size_t before_after = count_ops_of_type<op::Reshape>(f);
+    ASSERT_LE(before_after, before_count);
+}