ngraph/core/src/op/grn.cpp

   1 //*****************************************************************************
   2 // Copyright 2017-2020 Intel Corporation
   3 //
   4 // Licensed under the Apache License, Version 2.0 (the "License");
   5 // you may not use this file except in compliance with the License.
   6 // You may obtain a copy of the License at
   7 //
   8 //     http://www.apache.org/licenses/LICENSE-2.0
   9 //
  10 // Unless required by applicable law or agreed to in writing, software
  11 // distributed under the License is distributed on an "AS IS" BASIS,
  12 // WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
  13 // See the License for the specific language governing permissions and
  14 // limitations under the License.
  15 //*****************************************************************************
  16 #include <algorithm>
  17 #include <iterator>
  18
  19 #include "grn.hpp"
  20 #include "ngraph/attribute_visitor.hpp"
  21 #include "ngraph/axis_set.hpp"
  22 #include "ngraph/builder/norm.hpp"
  23 #include "ngraph/builder/reshape.hpp"
  24 #include "ngraph/op/broadcast.hpp"
  25 #include "ngraph/op/constant.hpp"
  26 #include "ngraph/op/divide.hpp"
  27 #include "ngraph/shape.hpp"
  28
  29 using namespace std;
  30 using namespace ngraph;
  31
  32 NGRAPH_SUPPRESS_DEPRECATED_START
  33
  34 constexpr NodeTypeInfo op::GRN::type_info;
  35
  36 op::GRN::GRN(const Output<Node>& data, float bias)
  37     : FusedOp({data})
  38     , m_bias(bias)
  39 {
  40     constructor_validate_and_infer_types();
  41 }
  42
  43 bool ngraph::op::v0::GRN::visit_attributes(AttributeVisitor& visitor)
  44 {
  45     visitor.on_attribute("bias", m_bias);
  46     return true;
  47 }
  48
  49 void op::GRN::pre_validate_and_infer_types()
  50 {
  51     const auto& data_pshape = get_input_partial_shape(0);
  52
  53     if (data_pshape.is_static())
  54     {
  55         const Shape& data_shape{data_pshape.to_shape()};
  56
  57         // Input data must be 2, 3 or 4D tensor.
  58         NODE_VALIDATION_CHECK(this,
  59                               (data_shape.size() >= 2 && data_shape.size() <= 4),
  60                               "Input tensor rank must be 2, 3 or 4 dimensional (actual input "
  61                               "shape: ",
  62                               data_shape,
  63                               ").");
  64     }
  65 }
  66
  67 OutputVector op::GRN::decompose_op() const
  68 {
  69     Output<Node> data{input_value(0)};
  70     const Shape& input_shape{data.get_shape()};
  71
  72     // Reshape to 4D tensor.
  73     if (input_shape.size() != 4)
  74     {
  75         Shape data_shape(4 - input_shape.size(), 1);
  76         copy(begin(input_shape), end(input_shape), back_inserter(data_shape));
  77         data = builder::opset1::reshape(data, data_shape);
  78     }
  79
  80     const auto axis_set_const = op::Constant::create(element::i64, {}, {1});
  81     // Calculate l2 norm across channels.
  82     shared_ptr<Node> norm = builder::opset1::l2_norm(data, axis_set_const, m_bias);
  83     // Get back reduced axis.
  84     norm = std::make_shared<Broadcast>(norm, data.get_shape(), AxisSet{1});
  85     data = data / norm;
  86
  87     // get back original input tensor rank
  88     if (input_shape.size() != 4)
  89     {
  90         data = builder::opset1::reshape(data, input_shape);
  91     }
  92
  93     return OutputVector{data};
  94 }
  95
  96 shared_ptr<Node> op::GRN::clone_with_new_inputs(const OutputVector& new_args) const
  97 {
  98     if (new_args.size() != 1)
  99     {
 100         throw ngraph_error("Incorrect number of new arguments");
 101     }
 102     return make_shared<GRN>(new_args.at(0), m_bias);
 103 }