[ONNX] Added support for negative axes. (#3557)

* [ONNX] Added function for converting negative axes.

* [ONNX] Bugfix

* [ONNX] Code formatting

* [ONNX] Changed name of function.

* Added validation for axis/axes.

* Changed abs to llabs for int64 argument.

* Changed validation function.

* [ONNX] Added function for converting negative axes.

* [ONNX] Bugfix

* [ONNX] Code formatting

* [ONNX] Changed name of function.

* Added validation for axis/axes.

* Changed abs to llabs for int64 argument.

* Changed validation function.

* Changed validation function.

* Changed validation function

* Changed rank size in onehot

* Apply suggestions from code review
Co-Authored-By: Tomasz Socha <tomasz.socha@intel.com>

* Code refactoring.

* Update flatten.cpp

src/ngraph/frontend/onnx_import/op/concat.cpp

@@ -33,12 +33,8 @@ namespace ngraph
                 {
                     NodeVector inputs{node.get_ng_inputs()};
                     std::int64_t axis = node.get_attribute_value<std::int64_t>("axis");
-
                     size_t valid_axis =
-                        common::convert_negative_axis(axis, inputs.at(0)->get_shape().size());
-
-                    ASSERT_VALID_ARGUMENT(node, valid_axis >= 0)
-                        << "Incorrect value of axis attribute: " << axis;
+                        common::validate_axis(node, axis, inputs.at(0)->get_shape().size());
 
                     return {std::make_shared<ngraph::op::Concat>(inputs, valid_axis)};
                 }

src/ngraph/frontend/onnx_import/op/flatten.cpp

@@ -19,6 +19,7 @@
 #include "exceptions.hpp"
 #include "flatten.hpp"
 #include "ngraph/builder/reshape.hpp"
+#include "utils/common.hpp"
 
 namespace ngraph
 {
@@ -33,13 +34,12 @@ namespace ngraph
                     NodeVector inputs{node.get_ng_inputs()};
                     auto data = inputs.at(0);
                     auto axis = node.get_attribute_value<std::int64_t>("axis", 1);
+                    auto data_rank = data->get_shape().size();
+                    // Accepted range is [-r, r] where r = rank(input).
+                    auto valid_axis =
+                        common::validate_axis(node, axis, data_rank, -data_rank, data_rank);
 
-                    ASSERT_VALID_ARGUMENT(
-                        node,
-                        (axis >= 0) && (axis <= static_cast<int64_t>(data->get_shape().size())))
-                        << "provided 'axis' attribute is not valid.";
-
-                    return {ngraph::builder::flatten(data, axis)};
+                    return {ngraph::builder::flatten(data, valid_axis)};
                 }
 
             } // namespace set_1

src/ngraph/frontend/onnx_import/op/gather.hpp

@@ -19,6 +19,7 @@
 #include "core/node.hpp"
 #include "ngraph/node.hpp"
 #include "ngraph/op/gather.hpp"
+#include "utils/common.hpp"
 
 namespace ngraph
 {
@@ -34,12 +35,9 @@ namespace ngraph
                     auto data = ng_inputs.at(0);
                     auto indices = ng_inputs.at(1);
                     auto axis = node.get_attribute_value<int64_t>("axis", 0);
-                    if (axis < 0)
-                    {
-                        axis += data->get_shape().size();
-                    }
+                    auto valid_axis = common::validate_axis(node, axis, data->get_shape().size());
 
-                    return {std::make_shared<ngraph::op::Gather>(data, indices, axis)};
+                    return {std::make_shared<ngraph::op::Gather>(data, indices, valid_axis)};
                 }
 
             } // namespace set_1

src/ngraph/frontend/onnx_import/op/hardmax.cpp

@@ -35,13 +35,10 @@ namespace ngraph
                     const auto& input_shape = input->get_shape();
                     auto axis = node.get_attribute_value<std::int64_t>("axis", 1);
 
-                    ASSERT_VALID_ARGUMENT(
-                        node, axis >= 0 && axis < static_cast<int64_t>(input_shape.size()))
-                        << "The provided axis value " << axis
-                        << " does not match the input tensor dimensions";
+                    auto valid_axis = common::validate_axis(node, axis, input_shape.size());
 
                     // reshape to 2D - "batch size" x "input feature dimensions" (NxD)
-                    const auto coerced_tensor = ngraph::builder::flatten(input, axis);
+                    const auto coerced_tensor = ngraph::builder::flatten(input, valid_axis);
                     const auto& coerced_shape = coerced_tensor->get_shape();
 
                     const std::shared_ptr<ngraph::Node> argmax_2d =

src/ngraph/frontend/onnx_import/op/lp_norm.cpp

@@ -23,6 +23,7 @@
 #include "ngraph/builder/norm.hpp"
 #include "ngraph/op/broadcast.hpp"
 #include "ngraph/op/divide.hpp"
+#include "utils/common.hpp"
 
 namespace ngraph
 {
@@ -37,17 +38,14 @@ namespace ngraph
                     const std::shared_ptr<ngraph::Node> data{node.get_ng_inputs().at(0)};
                     std::int64_t axis{node.get_attribute_value<std::int64_t>("axis", -1)};
                     const std::int64_t p_norm{node.get_attribute_value<std::int64_t>("p", 2)};
-
-                    if (axis < 0)
-                    {
-                        axis += data->get_shape().size();
-                    }
+                    std::size_t valid_axis =
+                        common::validate_axis(node, axis, data->get_shape().size());
 
                     ASSERT_VALID_ARGUMENT(node, p_norm == 1 || p_norm == 2)
                         << "Invalid `p` attribute value: " << p_norm
                         << "Only normalization of 1st or 2nd order is supported.";
 
-                    const AxisSet reduction_axes{static_cast<std::size_t>(axis)};
+                    const AxisSet reduction_axes{valid_axis};
                     std::shared_ptr<ngraph::Node> norm = ngraph::builder::lp_norm(
                         data, reduction_axes, static_cast<std::size_t>(p_norm));
                     norm = std::make_shared<ngraph::op::Broadcast>(

src/ngraph/frontend/onnx_import/op/mean_variance_normalization.cpp

@@ -19,6 +19,7 @@
 #include "mean_variance_normalization.hpp"
 #include "ngraph/axis_set.hpp"
 #include "ngraph/op/fused/mvn.hpp"
+#include "utils/common.hpp"
 
 namespace ngraph
 {
@@ -47,9 +48,11 @@ namespace ngraph
                 NodeVector mean_variance_normalization(const Node& node)
                 {
                     auto data = node.get_ng_inputs().at(0);
-                    auto axes = node.get_attribute_value<std::vector<size_t>>("axes", {0, 2, 3});
+                    auto axes = node.get_attribute_value<std::vector<int64_t>>("axes", {0, 2, 3});
+                    std::vector<std::size_t> valid_axes =
+                        common::validate_axes(node, axes, data->get_shape().size());
 
-                    return {std::make_shared<ngraph::op::MVN>(data, AxisSet(axes))};
+                    return {std::make_shared<ngraph::op::MVN>(data, AxisSet(valid_axes))};
                 }
 
             } // namespace set_9

src/ngraph/frontend/onnx_import/op/onehot.cpp

@@ -28,6 +28,7 @@
 #include "ngraph/op/subtract.hpp"
 #include "ngraph/op/util/broadcasting.hpp"
 #include "onehot.hpp"
+#include "utils/common.hpp"
 
 namespace ngraph
 {
@@ -51,15 +52,13 @@ namespace ngraph
                         std::make_shared<ngraph::op::Slice>(values, Coordinate{1}, Coordinate{2});
                     auto axis = node.get_attribute_value<std::int64_t>("axis", -1);
 
-                    if (axis < 0)
-                    {
-                        axis += indices_shape.size() + 1;
-                    }
-
-                    ASSERT_VALID_ARGUMENT(
-                        node, (axis >= 0) && (axis <= static_cast<int64_t>(indices_shape.size())))
-                        << "invalid 'axis' attribute: "
-                        << node.get_attribute_value<std::int64_t>("axis", -1);
+                    // Accepted range for axis is [-r-1, r] where r = rank(indices). Validate
+                    // against rank+1.
+                    std::size_t valid_axis = common::validate_axis(node,
+                                                                   axis,
+                                                                   indices_shape.size() + 1,
+                                                                   -indices_shape.size() - 1,
+                                                                   indices_shape.size());
 
                     auto constant_depth = ngraph::as_type_ptr<ngraph::op::Constant>(depth);
 
@@ -75,10 +74,11 @@ namespace ngraph
                     // axis = 1
                     // depth = 10
                     // output_shape = (2, 10, 2)
-                    output_shape.insert(std::next(std::begin(output_shape), axis), depth_value);
+                    output_shape.insert(std::next(std::begin(output_shape), valid_axis),
+                                        depth_value);
 
                     std::shared_ptr<ngraph::Node> one_hot = std::make_shared<ngraph::op::Convert>(
-                        std::make_shared<ngraph::op::OneHot>(indices, output_shape, axis),
+                        std::make_shared<ngraph::op::OneHot>(indices, output_shape, valid_axis),
                         values->get_element_type());
                     auto broadcasted_values =
                         ngraph::op::numpy_style_broadcast({one_hot, on_value, off_value});

src/ngraph/frontend/onnx_import/op/reverse_sequence.cpp

@@ -21,6 +21,7 @@
 #include "ngraph/op/convert.hpp"
 #include "ngraph/op/reverse_sequence.hpp"
 #include "ngraph/type/element_type.hpp"
+#include "utils/common.hpp"
 
 namespace ngraph
 {
@@ -40,22 +41,26 @@ namespace ngraph
                         node.get_ng_inputs().at(1), element::i32);
 
                     const auto batch_axis = node.get_attribute_value<int64_t>("batch_axis", 1);
+                    std::size_t valid_batch_axis =
+                        common::validate_axis(node, batch_axis, data->get_shape().size());
                     const auto time_axis = node.get_attribute_value<int64_t>("time_axis", 0);
+                    std::size_t valid_time_axis =
+                        common::validate_axis(node, time_axis, data->get_shape().size());
 
-                    NGRAPH_CHECK(batch_axis == 0 || batch_axis == 1,
+                    NGRAPH_CHECK(valid_batch_axis == 0 || valid_batch_axis == 1,
                                  "Allowed values of the 'batch_axis' attribute for ReverseSequence "
                                  "operator are 0 and 1");
 
-                    NGRAPH_CHECK(time_axis == 0 || time_axis == 1,
+                    NGRAPH_CHECK(valid_time_axis == 0 || valid_time_axis == 1,
                                  "Allowed values of the 'time_axis' attribute for ReverseSequence "
                                  "operator are 0 and 1");
 
-                    NGRAPH_CHECK(batch_axis != time_axis,
+                    NGRAPH_CHECK(valid_batch_axis != valid_time_axis,
                                  "'batch_axis' and 'time_axis' attributes of the ReverseSequence "
                                  "operator can't point to the same dimension");
 
                     return {std::make_shared<ngraph::op::ReverseSequence>(
-                        data, sequence_lengths_i32, batch_axis, time_axis)};
+                        data, sequence_lengths_i32, valid_batch_axis, valid_time_axis)};
                 }
 
             } // namespace set_1

src/ngraph/frontend/onnx_import/op/softmax.cpp

@@ -19,6 +19,7 @@
 #include "exceptions.hpp"
 #include "ngraph/op/softmax.hpp"
 #include "softmax.hpp"
+#include "utils/common.hpp"
 
 namespace ngraph
 {
@@ -35,22 +36,13 @@ namespace ngraph
                     auto data_shape = data->get_shape();
 
                     int axis = node.get_attribute_value<int64_t>("axis", 1);
-
-                    if (axis < 0)
-                    {
-                        axis = data_shape.size() + axis;
-                    }
-
-                    ASSERT_VALID_ARGUMENT(node, axis < static_cast<int64_t>(data_shape.size()))
-                        << "provided 'axis' value:" << axis
-                        << " is out of input tensor dimensions range.";
+                    std::size_t valid_axis = common::validate_axis(node, axis, data_shape.size());
 
                     // create vector of capacity data_dimensions - axis_divider position
-                    std::vector<size_t> axes(data_shape.size() - axis);
-                    std::iota(std::begin(axes), std::end(axes), axis);
+                    std::vector<size_t> axes(data_shape.size() - valid_axis);
+                    std::iota(std::begin(axes), std::end(axes), valid_axis);
                     return {std::make_shared<ngraph::op::Softmax>(data, axes)};
                 }
-
             } // namespace set_1
 
         } // namespace op

src/ngraph/frontend/onnx_import/op/split.cpp

@@ -19,6 +19,7 @@
 
 #include "ngraph/op/fused/split.hpp"
 #include "op/split.hpp"
+#include "utils/common.hpp"
 
 namespace ngraph
 {
@@ -33,13 +34,15 @@ namespace ngraph
                     const auto input = node.get_ng_inputs().at(0);
                     const auto outputs_number = node.get_output_names().size();
                     const auto axis = node.get_attribute_value<int64_t>("axis", 0);
+                    std::size_t valid_axis =
+                        common::validate_axis(node, axis, input->get_shape().size());
 
                     try
                     {
                         const auto length_parts =
                             node.get_attribute_value<std::vector<std::size_t>>("split");
                         const auto fused_split =
-                            std::make_shared<ngraph::op::Split>(input, axis, length_parts);
+                            std::make_shared<ngraph::op::Split>(input, valid_axis, length_parts);
 
                         return fused_split->decompose_op();
                     }
@@ -49,7 +52,7 @@ namespace ngraph
                         // the 'split' attribute - this means we should split the input tensor
                         // into same-length parts equal to the number of node outputs
                         const auto fused_split =
-                            std::make_shared<ngraph::op::Split>(input, axis, outputs_number);
+                            std::make_shared<ngraph::op::Split>(input, valid_axis, outputs_number);
 
                         return fused_split->decompose_op();
                     }

src/ngraph/frontend/onnx_import/op/squeeze.cpp

@@ -20,6 +20,7 @@
 #include "ngraph/op/constant.hpp"
 #include "ngraph/op/fused/squeeze.hpp"
 #include "squeeze.hpp"
+#include "utils/common.hpp"
 
 namespace ngraph
 {
@@ -32,17 +33,12 @@ namespace ngraph
                 NodeVector squeeze(const Node& node)
                 {
                     auto data = node.get_ng_inputs().at(0);
-                    auto axes = node.get_attribute_value<std::vector<std::int64_t>>("axes", {});
-
-                    for (auto axis : axes)
-                    {
-                        ASSERT_VALID_ARGUMENT(node, axis >= 0)
-                            << "provided axes attribute is invalid. Only non-negative "
-                            << "integers are allowed, got " << axis << ".";
-                    }
-
+                    std::vector<std::int64_t> axes =
+                        node.get_attribute_value<std::vector<std::int64_t>>("axes", {});
+                    std::vector<std::size_t> valid_axes =
+                        common::validate_axes(node, axes, data->get_shape().size());
                     auto axes_node = std::make_shared<ngraph::op::Constant>(
-                        element::u64, Shape{axes.size()}, axes);
+                        element::u64, Shape{valid_axes.size()}, valid_axes);
                     return {std::make_shared<ngraph::op::Squeeze>(data, axes_node)};
                 }
 

src/ngraph/frontend/onnx_import/op/topk.cpp

@@ -23,6 +23,7 @@
 #include "ngraph/op/topk.hpp"
 #include "ngraph/type/element_type.hpp"
 #include "topk.hpp"
+#include "utils/common.hpp"
 
 namespace ngraph
 {
@@ -35,21 +36,14 @@ namespace ngraph
                 NodeVector topk(const Node& node)
                 {
                     auto data = node.get_ng_inputs().at(0);
-                    std::int64_t axis{node.get_attribute_value<std::int64_t>("axis", -1)};
                     std::int64_t k{node.get_attribute_value<std::int64_t>("k")};
-
                     auto num_dimensions = data->get_shape().size();
 
-                    if (axis < 0)
-                    {
-                        axis += num_dimensions;
-                    }
-
-                    ASSERT_VALID_ARGUMENT(node, axis < static_cast<int64_t>(num_dimensions))
-                        << "`axis` parameter is out of range: " << axis;
+                    std::int64_t axis{node.get_attribute_value<std::int64_t>("axis", -1)};
+                    std::int64_t valid_axis = common::validate_axis(node, axis, num_dimensions);
 
                     std::shared_ptr<ngraph::Node> top_k =
-                        std::make_shared<ngraph::op::TopK>(data, axis, element::i64, k);
+                        std::make_shared<ngraph::op::TopK>(data, valid_axis, element::i64, k);
 
                     std::shared_ptr<ngraph::Node> indices =
                         std::make_shared<ngraph::op::GetOutputElement>(top_k, 0);

src/ngraph/frontend/onnx_import/utils/common.cpp

@@ -46,6 +46,53 @@ namespace ngraph
                                        static_cast<onnx::TensorProto_DataType>(onnx_type)));
             }
 
+            std::size_t validate_axis(const ngraph::onnx_import::Node& node,
+                                      std::int64_t axis,
+                                      std::int64_t tensor_rank)
+            {
+                // Accepted range of value for axis is [-tensor_rank, tensor_rank-1].
+                return validate_axis(node, axis, tensor_rank, -tensor_rank, tensor_rank - 1);
+            }
+
+            std::size_t validate_axis(const ngraph::onnx_import::Node& node,
+                                      std::int64_t axis,
+                                      std::int64_t tensor_rank,
+                                      std::int64_t axis_range_min,
+                                      std::int64_t axis_range_max)
+            {
+                // Accepted range of value for axis is [axis_range_min, axis_range_max].
+                NGRAPH_CHECK(((axis >= axis_range_min) && (axis <= axis_range_max)),
+                             node.get_description(),
+                             "Parameter axis ",
+                             axis,
+                             " out of the tensor rank [-",
+                             axis_range_min,
+                             ", ",
+                             axis_range_max,
+                             "].");
+
+                if (axis < 0)
+                {
+                    axis = axis + tensor_rank;
+                }
+
+                return static_cast<size_t>(axis);
+            }
+
+            std::vector<std::size_t> validate_axes(const ngraph::onnx_import::Node& node,
+                                                   std::vector<std::int64_t> axes,
+                                                   std::int64_t tensor_rank)
+            {
+                std::vector<std::size_t> new_axes;
+
+                for (auto a : axes)
+                {
+                    new_axes.push_back(validate_axis(node, a, tensor_rank));
+                }
+
+                return new_axes;
+            }
+
         } // namespace  common
     }     // namespace onnx_import
 } // namespace ngraph

src/ngraph/frontend/onnx_import/utils/common.hpp

@@ -25,6 +25,7 @@
 #include <type_traits> // std::enable_if
 #include <vector>
 
+#include "core/node.hpp"
 #include "ngraph/op/constant.hpp"
 #include "ngraph/op/util/broadcasting.hpp"
 #include "ngraph/shape.hpp"
@@ -67,28 +68,52 @@ namespace ngraph
                 return range;
             }
 
-            /// \brief      Handle negative axis value.
+            /// \brief      Handle out of range axis.
             ///
-            /// \param[in]  axis        The requested axis value.
-            /// \param[in]  tensor_dim  The corresponding tensor dimensionality.
+            /// \param[in]  node         The node with requested axis.
+            /// \param[in]  axis         The requested axis value.
+            /// \param[in]  tensor_rank  The corresponding tensor rank.
             ///
-            /// \tparam     T           Provided axis value type.
+            /// \return    Checking if axis is in range [-tensor_rank, tensor_rank-1], otherwise
+            /// returns error.
+            ///            If negative axis, it counts from the last to the first axis, by adding
+            ///            tensor_rank to axis.
             ///
-            /// \return     If negative axis, then return sum of tensor dimension and axis.
+            std::size_t validate_axis(const ngraph::onnx_import::Node& node,
+                                      std::int64_t axis,
+                                      std::int64_t tensor_rank);
+
+            /// \brief      Handle out of range axis.
             ///
-            template <typename T,
-                      typename std::enable_if<std::is_integral<T>::value, int>::type = 0>
-            std::int64_t convert_negative_axis(T axis, std::size_t tensor_dim)
-            {
-                if (axis >= 0)
-                {
-                    return std::min(axis, static_cast<T>(tensor_dim));
-                }
-                else
-                {
-                    return static_cast<std::int64_t>(tensor_dim) + axis;
-                }
-            }
+            /// \param[in]  node            The node with requested axis.
+            /// \param[in]  axis            The requested axis value.
+            /// \param[in]  tensor_rank     The corresponding tensor rank.
+            /// \param[in]  axis_range_min  The min value of accepted range for axis.
+            /// \param[in]  axis_range_max  The max value of accepted range for axis.
+            ///
+            /// \return     Checking if axis is in range [axis_range_min, axis_range_max], otherwise
+            /// returns error.
+            ////            If negative axis, it counts from the last to the first axis, by adding
+            /// tensor_rank to axis.
+            ///
+            std::size_t validate_axis(const ngraph::onnx_import::Node& node,
+                                      std::int64_t axis,
+                                      std::int64_t tensor_rank,
+                                      std::int64_t axis_range_min,
+                                      std::int64_t axis_range_max);
+
+            /// \brief      Handle out of range axes in vector.
+            ///
+            /// \param[in]  node         The node with requested axes.
+            /// \param[in]  axes         The requested vector of axes.
+            /// \param[in]  tensor_rank  The corresponding tensor rank.
+            ///
+            /// \return     If any negative axis in vector, it counts from the last to the first
+            /// axis, by adding tensor_rank to axis.
+            ///
+            std::vector<std::size_t> validate_axes(const ngraph::onnx_import::Node& node,
+                                                   std::vector<std::int64_t> axes,
+                                                   std::int64_t tensor_rank);
 
             /// \brief Creates a shifted square identity matrix.
             /// \note Shifting in the context of this operator means that

src/ngraph/frontend/onnx_import/utils/reduction.cpp

@@ -32,13 +32,17 @@ namespace ngraph
                 AxisSet get_reduction_axes(const Node& node)
                 {
                     auto reduction_axes =
-                        node.get_attribute_value<std::vector<std::size_t>>("axes", {});
+                        node.get_attribute_value<std::vector<std::int64_t>>("axes", {});
+                    std::vector<std::size_t> valid_reduction_axes = common::validate_axes(
+                        node, reduction_axes, node.get_ng_inputs().at(0)->get_shape().size());
+
                     if (reduction_axes.empty())
                     {
-                        reduction_axes = onnx_import::common::get_monotonic_range<std::size_t>(
-                            node.get_ng_inputs().at(0)->get_shape().size());
+                        valid_reduction_axes =
+                            onnx_import::common::get_monotonic_range<std::size_t>(
+                                node.get_ng_inputs().at(0)->get_shape().size());
                     }
-                    return AxisSet{reduction_axes};
+                    return AxisSet{valid_reduction_axes};
                 }
             } // namespace  detail
 

src/ngraph/frontend/onnx_import/utils/reduction.hpp

@@ -26,6 +26,7 @@
 #include "ngraph/op/reshape.hpp"
 #include "ngraph/shape.hpp"
 #include "ngraph/util.hpp"
+#include "utils/common.hpp"
 #include "utils/reshape.hpp"
 
 namespace ngraph
@@ -64,8 +65,10 @@ namespace ngraph
                 auto axis = node.get_attribute_value<std::int64_t>("axis", 0);
                 auto keepdims = node.get_attribute_value<std::int64_t>("keepdims", 1);
                 auto input_node = node.get_ng_inputs().at(0);
+                auto valid_axis = common::validate_axis(node, axis, input_node->get_shape().size());
 
-                auto op_node = std::make_shared<IndexReduction>(input_node, axis, element::i64);
+                auto op_node =
+                    std::make_shared<IndexReduction>(input_node, valid_axis, element::i64);
 
                 if (keepdims == 0)
                 {
@@ -76,7 +79,7 @@ namespace ngraph
                 auto convert_node = std::make_shared<ngraph::op::Convert>(op_node, element::f32);
 
                 auto output_shape = input_node->get_shape();
-                output_shape.at(axis) = 1;
+                output_shape.at(valid_axis) = 1;
                 auto reshape_node = std::make_shared<ngraph::op::Reshape>(
                     convert_node,
                     ngraph::get_default_order(op_node->get_shape().size()),