inference-engine/tests/ngraph_functions/src/low_precision_transformations/convert_mul_or_add_finally_with_dequantization_function.cpp

   1 // Copyright (C) 2020 Intel Corporation
   2 // SPDX-License-Identifier: Apache-2.0
   3 //
   4
   5 #include "ngraph_functions/low_precision_transformations/convert_mul_or_add_finally_with_dequantization_function.hpp"
   6
   7 #include <memory>
   8 #include <vector>
   9 #include <ngraph/ngraph.hpp>
  10
  11
  12 #include <ngraph/opsets/opset1.hpp>
  13 #include "ngraph_functions/subgraph_builders.hpp"
  14 #include "ngraph_functions/low_precision_transformations/common/builders.hpp"
  15 #include "low_precision/network_helper.hpp"
  16 #include <legacy/ngraph_ops/scaleshift.hpp>
  17 #include "low_precision/common/dequantization_op.hpp"
  18
  19 namespace ngraph {
  20 namespace builder {
  21 namespace subgraph {
  22
  23 using namespace ngraph::pass;
  24
  25 std::shared_ptr<ngraph::Function> ConvertMulOrAddWithDequantizationFunction::getOriginal(
  26     const ngraph::Shape& inputShape,
  27     const ngraph::element::Type inputPrecision,
  28     const std::vector<float>& multiplyConst) {
  29     const auto input = std::make_shared<ngraph::opset1::Parameter>(inputPrecision, inputShape);
  30     const auto reluOriginal = ngraph::opset1::Relu(
  31         ngraph::op::TemporaryReplaceOutputType(input, element::f32).get());
  32
  33     std::shared_ptr<ngraph::opset1::Relu> relu = std::make_shared<ngraph::op::TypeRelaxed<ngraph::opset1::Relu>>(
  34         reluOriginal,
  35         std::vector<element::Type>{ element::f32, element::f32 },
  36         std::vector<element::Type>{});
  37
  38
  39     const auto multiply = std::make_shared<ngraph::pass::low_precision::DequantizationMultiply>(relu,
  40                                                             std::make_shared<opset1::Constant>(element::f32, inputShape, multiplyConst));
  41
  42     multiply->set_friendly_name("output");
  43
  44     ngraph::ResultVector results{ std::make_shared<ngraph::opset1::Result>(multiply) };
  45     return std::make_shared<ngraph::Function>(results, ngraph::ParameterVector{ input },
  46                                               "ConvertMulOrAddTransformationWithDequantization");
  47 }
  48
  49 std::shared_ptr<ngraph::Function> ConvertMulOrAddWithDequantizationFunction::getReference(
  50     const ngraph::Shape& inputShape,
  51     const ngraph::element::Type inputPrecision,
  52     const std::vector<float>& multiplyConst) {
  53     const auto input = std::make_shared<ngraph::opset1::Parameter>(inputPrecision, inputShape);
  54     const auto reluOriginal = ngraph::opset1::Relu(
  55         ngraph::op::TemporaryReplaceOutputType(input, element::f32).get());
  56
  57     std::shared_ptr<ngraph::opset1::Relu> relu = std::make_shared<ngraph::op::TypeRelaxed<ngraph::opset1::Relu>>(
  58         reluOriginal,
  59         std::vector<element::Type>{ element::f32, element::f32 },
  60         std::vector<element::Type>{});
  61
  62     const auto weights = std::make_shared<opset1::Constant>(element::f32, inputShape, multiplyConst);
  63     const auto bias = std::make_shared<opset1::Constant>(element::f32, inputShape, 0.0);
  64     std::shared_ptr<Node> scaleShift = std::make_shared<ngraph::op::ScaleShiftIE>(relu, weights, bias);
  65
  66     scaleShift = low_precision::NetworkHelper::markAsDequantizationOp(scaleShift);
  67
  68     scaleShift->set_friendly_name("output");
  69
  70     ngraph::ResultVector results{ std::make_shared<ngraph::opset1::Result>(scaleShift) };
  71     return std::make_shared<ngraph::Function>(results, ngraph::ParameterVector{ input }, "ConvertMulOrAddTransformationWithDequantization");
  72 }
  73
  74 }  // namespace subgraph
  75 }  // namespace builder
  76 }  // namespace ngraph