[LPT] Concat precision selection fix (#6069)

inference-engine/src/low_precision_transformations/src/concat.cpp

@@ -43,19 +43,21 @@ bool ConcatTransformation::transform(TransformationContext& context, ngraph::pat
         return false;
     }
 
-    // precisions can be different
+    // Concat operations precision is defined:
+    // 1. consumers after Concat
+    // 2. FakeQuantize precisions without zero point
     ngraph::Node& quantizationLayer = *subgraph.quantizationLayers[0];
     std::shared_ptr<ngraph::opset1::FakeQuantize> fq = ngraph::as_type_ptr<ngraph::opset1::FakeQuantize>(quantizationLayer.shared_from_this());
     if (!NetworkHelper::isQuantizeSupported(fq)) {
         return false;
     }
-
+    DataPrecision dataPrecision = getDataPrecision(fq, QuantizationDetails::getDetails(fq), false);
-    std::vector<element::Type> concatParentsChildrensPrecisions = precisionsOnActivations;
+    if (dataPrecision.precision == ngraph::element::undefined) {
-    fillAvailablePrecisions(subgraph.quantizationLayers[0], concatParentsChildrensPrecisions);
-    if (concatParentsChildrensPrecisions.empty()) {
         return false;
     }
 
+    std::vector<element::Type> concatChildrenPrecisions = precisionsOnActivations;
+
     for (size_t i = 0; i < subgraph.quantizationLayers.size(); ++i) {
         fq = ngraph::as_type_ptr<ngraph::opset1::FakeQuantize>(subgraph.quantizationLayers[i]);
         if (fq == nullptr) {
@@ -72,20 +74,28 @@ bool ConcatTransformation::transform(TransformationContext& context, ngraph::pat
         if (quantizationDetails.inputHighValues.size() != 1ul) {
             return false;
         }
-        std::vector<element::Type> fqChildrensPrecisions = precisionsOnActivations;
-        fillAvailablePrecisions(subgraph.quantizationLayers[i], fqChildrensPrecisions);
-        concatParentsChildrensPrecisions = NetworkHelper::precisionIntersection(concatParentsChildrensPrecisions, fqChildrensPrecisions);
 
-        if (concatParentsChildrensPrecisions.empty()) {
+        // define concatenation operation consumers precisions
+        std::vector<element::Type> fqChildrenPrecisions = precisionsOnActivations;
+        fillAvailablePrecisions(subgraph.quantizationLayers[i], fqChildrenPrecisions);
+        concatChildrenPrecisions = NetworkHelper::precisionIntersection(concatChildrenPrecisions, fqChildrenPrecisions);
+        if (concatChildrenPrecisions.empty()) {
             return false;
         }
+
+        // define FakeQuantize precisions without zero point
+        const DataPrecision dataPrecision2 = getDataPrecision(subgraph.quantizationLayers[i]->shared_from_this(), quantizationDetails, false);
+        if (dataPrecision2.precision == ngraph::element::undefined) {
+            return false;
+        }
+
+        if (dataPrecision.precision != dataPrecision2.precision) {
+            dataPrecision = dataPrecision.precision.is_signed() ? dataPrecision : dataPrecision2;
+        }
     }
 
-    DataPrecision dataPrecision;
+    if (std::find(concatChildrenPrecisions.begin(), concatChildrenPrecisions.end(), dataPrecision.precision) == concatChildrenPrecisions.end()) {
-    if (std::find(concatParentsChildrensPrecisions.begin(), concatParentsChildrensPrecisions.end(), element::i8) != concatParentsChildrensPrecisions.end()) {
+        dataPrecision = DataPrecision(concatChildrenPrecisions[0]);
-        dataPrecision = DataPrecision(element::i8);
-    } else {
-        dataPrecision = DataPrecision(concatParentsChildrensPrecisions[0]);
     }
 
     std::vector<QuantizationDetails> quantizationLayersDetails;

inference-engine/src/low_precision_transformations/src/concat_multi_channels.cpp

@@ -64,14 +64,23 @@ bool ConcatMultiChannelsTransformation::transform(TransformationContext& context
 
     DataPrecision dataPrecision;
     {
+        std::vector<element::Type> concatChildrenPrecisions = precisionsOnActivations;
         for (auto quantizationLayer : subgraph.quantizationLayers) {
             std::shared_ptr<ngraph::opset1::FakeQuantize> fq = ngraph::as_type_ptr<ngraph::opset1::FakeQuantize>(quantizationLayer->shared_from_this());
             if (!NetworkHelper::isQuantizeSupported(fq)) {
                 return false;
             }
 
-            const DataPrecision tmp = getDataPrecision(fq, QuantizationDetails::getDetails(fq), false);
+            // define concatenation operation consumers precisions
+            std::vector<element::Type> fqChildrenPrecisions = precisionsOnActivations;
+            fillAvailablePrecisions(quantizationLayer, fqChildrenPrecisions);
+            concatChildrenPrecisions = NetworkHelper::precisionIntersection(concatChildrenPrecisions, fqChildrenPrecisions);
+            if (concatChildrenPrecisions.empty()) {
+                return false;
+            }
 
+            // define FakeQuantize precisions without zero point
+            const DataPrecision tmp = getDataPrecision(fq, QuantizationDetails::getDetails(fq), false);
             if (dataPrecision.precision == ngraph::element::undefined) {
                 dataPrecision = tmp;
                 continue;
@@ -81,6 +90,10 @@ bool ConcatMultiChannelsTransformation::transform(TransformationContext& context
                 dataPrecision = tmp;
             }
         }
+
+        if (std::find(concatChildrenPrecisions.begin(), concatChildrenPrecisions.end(), dataPrecision.precision) == concatChildrenPrecisions.end()) {
+            dataPrecision = DataPrecision(concatChildrenPrecisions[0]);
+        }
     }
 
     for (size_t i = 0; i < subgraph.quantizationLayers.size(); ++i) {

inference-engine/tests/functional/inference_engine/lp_transformations/concat_with_intermediate_precision_selection_transformation.cpp

@@ -0,0 +1,317 @@
+// Copyright (C) 2018-2021 Intel Corporation
+// SPDX-License-Identifier: Apache-2.0
+//
+
+#include "layer_transformation.hpp"
+
+#include <string>
+#include <sstream>
+#include <memory>
+
+#include <gtest/gtest.h>
+
+#include <transformations/utils/utils.hpp>
+#include <transformations/init_node_info.hpp>
+#include <low_precision/transformer.hpp>
+#include <low_precision/avg_pool.hpp>
+#include <low_precision/concat.hpp>
+#include <low_precision/concat_multi_channels.hpp>
+#include <low_precision/max_pool.hpp>
+
+#include "common_test_utils/ngraph_test_utils.hpp"
+#include "lpt_ngraph_functions/concat_function.hpp"
+#include "lpt_ngraph_functions/common/fake_quantize_on_data.hpp"
+#include "simple_low_precision_transformer.hpp"
+
+using namespace testing;
+using namespace ngraph;
+using namespace ngraph::pass;
+
+namespace {
+
+class ConcatTransformationActualValues {
+public:
+    ngraph::builder::subgraph::FakeQuantizeOnData fakeQuantize1;
+    ngraph::builder::subgraph::FakeQuantizeOnData fakeQuantize2;
+};
+
+inline std::ostream& operator<<(std::ostream& out, const ConcatTransformationActualValues& values) {
+    return out << "_" << values.fakeQuantize1 << "_" << values.fakeQuantize2;
+}
+
+class ConcatTransformationResultValues {
+public:
+    ngraph::builder::subgraph::FakeQuantizeOnData fakeQuantize1;
+    ngraph::builder::subgraph::FakeQuantizeOnData fakeQuantize2;
+    ngraph::element::Type precisionBeforeOp;
+    ngraph::builder::subgraph::DequantizationOperations dequantizationBefore1;
+    ngraph::builder::subgraph::DequantizationOperations dequantizationBefore2;
+    ngraph::element::Type precisionAfterOperation;
+    ngraph::builder::subgraph::DequantizationOperations dequantizationAfter1;
+    ngraph::builder::subgraph::DequantizationOperations dequantizationAfter2;
+};
+
+inline std::ostream& operator<<(std::ostream& out, const ConcatTransformationResultValues& values) {
+    return out << "_" <<
+        values.fakeQuantize1 << "_" <<
+        values.fakeQuantize2 << "_" <<
+        values.dequantizationAfter1 << "_" <<
+        values.dequantizationAfter2;
+}
+
+class ConcatTransformationTestValues {
+public:
+    ngraph::pass::low_precision::LayerTransformation::Params params;
+    bool multiChannels;
+    ConcatTransformationActualValues actual;
+    ConcatTransformationResultValues result;
+};
+
+inline std::ostream& operator<<(std::ostream& out, const ConcatTransformationTestValues& values) {
+    return out << "_" << values.multiChannels << "_" << values.actual << "_" << values.result;
+}
+
+typedef std::tuple <
+    ngraph::element::Type,
+    ngraph::Shape,
+    ConcatTransformationTestValues
+> ConcatTransformationParams;
+
+class ConcatWithIntermediatePrecisionSelectionTransformation : public LayerTransformation, public testing::WithParamInterface<ConcatTransformationParams> {
+public:
+    void SetUp() override {
+        const ngraph::element::Type precision = std::get<0>(GetParam());
+        const ngraph::Shape shape = std::get<1>(GetParam());
+        ConcatTransformationTestValues testValues = std::get<2>(GetParam());
+
+        actualFunction = ngraph::builder::subgraph::ConcatFunction::getOriginalWithIntermediateAvgPool(
+            precision,
+            shape,
+            testValues.actual.fakeQuantize1,
+            testValues.actual.fakeQuantize2);
+
+        SimpleLowPrecisionTransformer transform;
+        if (testValues.multiChannels) {
+            transform.addBranchSpecific<ngraph::pass::low_precision::ConcatMultiChannelsTransformation, ngraph::opset1::Concat>(testValues.params);
+        } else {
+            transform.addBranchSpecific<ngraph::pass::low_precision::ConcatTransformation, ngraph::opset1::Concat>(testValues.params);
+        }
+        transform.add<ngraph::pass::low_precision::MaxPoolTransformation, ngraph::opset1::MaxPool>(testValues.params);
+        transform.add<ngraph::pass::low_precision::AvgPoolTransformation, ngraph::opset1::AvgPool>(testValues.params);
+        transform.transform(actualFunction);
+
+        referenceFunction = ngraph::builder::subgraph::ConcatFunction::getReferenceWithIntermediateAvgPool(
+            precision,
+            shape,
+            testValues.result.fakeQuantize1,
+            testValues.result.fakeQuantize2,
+            testValues.result.precisionBeforeOp,
+            testValues.result.dequantizationBefore1,
+            testValues.result.dequantizationBefore2,
+            testValues.result.precisionAfterOperation,
+            testValues.result.dequantizationAfter1,
+            testValues.result.dequantizationAfter2);
+    }
+
+    static std::string getTestCaseName(testing::TestParamInfo<ConcatTransformationParams> obj) {
+        const ngraph::element::Type precision = std::get<0>(obj.param);
+        const ngraph::Shape shape = std::get<1>(obj.param);
+        const ConcatTransformationTestValues testValues = std::get<2>(obj.param);
+
+        std::ostringstream result;
+        result <<
+            LayerTransformation::getTestCaseNameByParams(precision, shape, testValues.params) << "_" <<
+            (testValues.multiChannels ? "multiChannels_" : "notMultiChannels_") <<
+            testValues.actual << "_" <<
+            testValues.result << "_";
+        return result.str();
+    }
+};
+
+TEST_P(ConcatWithIntermediatePrecisionSelectionTransformation, CompareFunctions) {
+    actualFunction->validate_nodes_and_infer_types();
+    auto res = compare_functions(referenceFunction, actualFunction, true, false, true);
+    ASSERT_TRUE(res.first) << res.second;
+}
+
+const std::vector<ngraph::element::Type> precisions = {
+    ngraph::element::f32,
+    // ngraph::element::f16
+};
+
+const std::vector<ConcatTransformationTestValues> testValues = {
+    // Concat: FakeQuantize operations with signed intervals but consumer requires U8
+    {
+        LayerTransformation::createParamsU8I8(),
+        false,
+        {
+            { 256ul, ngraph::Shape({}), {-1.28f}, {1.27f}, {-1.28f}, {1.27f} },
+            { 256ul, ngraph::Shape({}), {-1.28f / 2.f}, {1.27f / 2.f}, {-1.28f / 2.f}, {1.27f / 2.f} }
+        },
+        {
+            { 256ul, ngraph::Shape({}), {-1.28f}, {1.27f}, {0.f}, {255.f} },
+            { 256ul, ngraph::Shape({}), {-1.28f / 2.f}, {1.27f / 2.f}, {64.f}, {192.f} },
+            ngraph::element::u8,
+            {{}, {}, {}},
+            {{}, {}, {}},
+            ngraph::element::u8,
+            { ngraph::element::f32, { 128.f }, { 0.01f } },
+            { {}, { 128.f }, { 0.01f } }
+        }
+    },
+
+    // Concat: FakeQuantize operations with unsigned intervals but consumer requires I8
+    {
+        LayerTransformation::createParamsI8I8(),
+        false,
+        {
+            { 256ul, ngraph::Shape({}), {0.f}, {2.55f}, {0.f}, {2.55f} },
+            { 256ul, ngraph::Shape({}), {0.f}, {2.55f / 2.f}, {0.f}, {2.55f / 2.f} }
+        },
+        {
+            { 256ul, ngraph::Shape({}), {0.f}, {2.55f}, {-128.f}, {127.f} },
+            { 256ul, ngraph::Shape({}), {0.f}, {2.55f / 2.f}, {-128.f}, { -0.f} },
+            ngraph::element::i8,
+            {{}, {}, {}},
+            {{}, {}, {}},
+            ngraph::element::i8,
+            { ngraph::element::f32, { -128.f }, { 0.01f } },
+            { {}, { -128.f }, { 0.01f } }
+        }
+    },
+
+    // ConcatMultichannel: FakeQuantize operations with signed intervals but consumer requires U8
+    {
+        LayerTransformation::createParamsU8I8(),
+        true,
+        {
+            { 256ul, ngraph::Shape({}), {-1.28f}, {1.27f}, {-1.28f}, {1.27f} },
+            { 256ul, ngraph::Shape({}), {-1.28f / 2.f}, {1.27f / 2.f}, {-1.28f / 2.f}, {1.27f / 2.f} }
+        },
+        {
+            { 256ul, ngraph::Shape({}), {-1.28f}, {1.27f}, {0.f}, {255.f} },
+            { 256ul, ngraph::Shape({}), {-1.28f / 2.f}, {1.27f / 2.f}, {0.f}, { 255.f} },
+            ngraph::element::u8,
+            {},
+            {},
+            ngraph::element::u8,
+            { ngraph::element::f32, { 128.f }, {{ 0.01f, 0.01f, 0.01f, 0.005f, 0.005f, 0.005f }} },
+            { {}, { 128.f }, { 0.005f } }
+        }
+    },
+
+    // ConcatMultichannel: FakeQuantize operations with unsigned intervals but consumer requires I8
+    {
+        LayerTransformation::createParamsI8I8(),
+        true,
+        {
+            { 256ul, ngraph::Shape({}), {0.f}, {2.55f}, {0.f}, {2.55f} },
+            { 256ul, ngraph::Shape({}), {0.f}, {2.55f / 2.f}, {0.f}, {2.55f / 2.f} }
+        },
+        {
+            { 256ul, ngraph::Shape({}), {0.f}, {2.55f}, {-128.f}, {127.f} },
+            { 256ul, ngraph::Shape({}), {0.f}, {2.55f / 2.f}, {-128.f}, { 127.f} },
+            ngraph::element::i8,
+            {{}, {}, {}},
+            {{}, {}, {}},
+            ngraph::element::i8,
+            { ngraph::element::f32, { -128.f }, {{ 0.01f, 0.01f, 0.01f, 0.005f, 0.005f, 0.005f }} },
+            { {}, { -128.f }, { 0.005f } }
+        }
+    },
+
+    // Concat: FakeQuantize operations with unsigned intervals, no consumer limitations: FQ were decomposed to U8 precision
+    {
+        LayerTransformation::createParamsU8I8AndI8(),
+        false,
+        {
+            { 256ul, ngraph::Shape({}), {0.f}, {2.55f}, {0.f}, {2.55f} },
+            { 256ul, ngraph::Shape({}), {0.f}, {2.55f / 2.f}, {0.f}, {2.55f / 2.f} }
+        },
+        {
+            { 256ul, ngraph::Shape({}), {0.f}, {2.55f}, {0.f}, {255.f} },
+            { 256ul, ngraph::Shape({}), {0.f}, {2.55f / 2.f}, {0.f}, { 128.f} },
+            ngraph::element::u8,
+            {{}, {}, {}},
+            {{}, {}, {}},
+            ngraph::element::u8,
+            { ngraph::element::f32, {}, { 0.01f } },
+            { {}, {}, { 0.01f } }
+        }
+    },
+
+    // Concat: FakeQuantize operations with signed intervals, no consumer limitations: FQ were decomposed to I8 precision
+    {
+        LayerTransformation::createParamsU8I8AndI8(),
+        false,
+        {
+            { 256ul, ngraph::Shape({}), {-1.28f}, {1.27f}, {-1.28f}, {1.27f} },
+            { 256ul, ngraph::Shape({}), {-1.28f / 2.f}, {1.27f / 2.f}, {-1.28f / 2.f}, {1.27f / 2.f} }
+        },
+        {
+            { 256ul, ngraph::Shape({}), {-1.28f}, {1.27f}, {-128.f}, {127.f} },
+            { 256ul, ngraph::Shape({}), {-1.28f / 2.f}, {1.27f / 2.f}, {-64.f}, {64.f} },
+            ngraph::element::i8,
+            {{}, {}, {}},
+            {{}, {}, {}},
+            ngraph::element::i8,
+            { ngraph::element::f32, {}, { 0.01f } },
+            { {}, {}, { 0.01f } }
+        }
+    },
+
+    // ConcatMultichannel: FakeQuantize operations with unsigned intervals, no consumer limitations: FQ were decomposed to U8 precision
+    {
+        LayerTransformation::createParamsU8I8AndI8(),
+        true,
+        {
+            { 256ul, ngraph::Shape({}), {0.f}, {2.55f}, {0.f}, {2.55f} },
+            { 256ul, ngraph::Shape({}), {0.f}, {2.55f / 2.f}, {0.f}, {2.55f / 2.f} }
+        },
+        {
+            { 256ul, ngraph::Shape({}), {0.f}, {2.55f}, {0.f}, {255.f} },
+            { 256ul, ngraph::Shape({}), {0.f}, {2.55f / 2.f}, {0.f}, {255.f} },
+            ngraph::element::u8,
+            {{}, {}, {}},
+            {{}, {}, {}},
+            ngraph::element::u8,
+            { ngraph::element::f32, {}, {{ 0.01f, 0.01f, 0.01f, 0.005f, 0.005f, 0.005f }} },
+            { {}, {}, { 0.005f } }
+        }
+    },
+
+    // ConcatMultichannel: FakeQuantize operations with signed intervals, no consumer limitations: FQ were decomposed to I8 precision
+    {
+        LayerTransformation::createParamsU8I8AndI8(),
+        true,
+        {
+            { 256ul, ngraph::Shape({}), {-1.28f}, {1.27f}, {-1.28f}, {1.27f} },
+            { 256ul, ngraph::Shape({}), {-1.28f / 2.f}, {1.27f / 2.f}, {-1.28f / 2.f}, {1.27f / 2.f} }
+        },
+        {
+            { 256ul, ngraph::Shape({}), {-1.28f}, {1.27f}, {-128.f}, {127.f} },
+            { 256ul, ngraph::Shape({}), {-1.28f / 2.f}, {1.27f / 2.f}, {-128.f}, {127.f} },
+            ngraph::element::i8,
+            {{}, {}, {}},
+            {{}, {}, {}},
+            ngraph::element::i8,
+            { ngraph::element::f32, {}, {{ 0.01f, 0.01f, 0.01f, 0.005f, 0.005f, 0.005f }} },
+            { {}, {}, { 0.005f } }
+        }
+    }
+};
+
+const std::vector<ngraph::Shape> shapes = {
+    { 1, 3, 9, 9 },
+    { 4, 3, 9, 9 }
+};
+
+INSTANTIATE_TEST_CASE_P(
+    smoke_LPT,
+    ConcatWithIntermediatePrecisionSelectionTransformation,
+    ::testing::Combine(
+        ::testing::ValuesIn(precisions),
+        ::testing::ValuesIn(shapes),
+        ::testing::ValuesIn(testValues)),
+    ConcatWithIntermediatePrecisionSelectionTransformation::getTestCaseName);
+}  // namespace

inference-engine/tests/functional/inference_engine/lp_transformations/simple_low_precision_transformer.cpp

@@ -49,19 +49,41 @@ bool SimpleLowPrecisionTransformer::isPrecisionPreserved(const std::shared_ptr<n
 }
 
 void SimpleLowPrecisionTransformer::transform(std::shared_ptr<ngraph::Function>& function) {
+    // initialization
+    for (auto it : branchSpecificTransformations) {
+        ngraph::pass::low_precision::LayerTransformationPtr transformation = it.second;
+        transformation->setParamsManager(this);
+        transformation->setLayerTransformationsManager(this);
+    }
+
+    for (auto it : transformations) {
+        ngraph::pass::low_precision::LayerTransformationPtr transformation = it.second;
+        transformation->setParamsManager(this);
+        transformation->setLayerTransformationsManager(this);
+    }
+
+    // transformation
     {
         ngraph::pass::low_precision::TypeRelaxedReplacer pass;
         pass.run_on_function(function);
     }
 
     ngraph::pass::low_precision::TransformationContext context(function);
-    GraphRewrite pass;
+    {
-    for (auto it : transformations) {
+        GraphRewrite pass;
-        ngraph::pass::low_precision::LayerTransformationPtr transformation = it.second;
+        for (auto it : branchSpecificTransformations) {
-
+            ngraph::pass::low_precision::LayerTransformationPtr transformation = it.second;
-        transformation->setParamsManager(this);
+            transformation->registerMatcherIn(pass, context);
-        transformation->setLayerTransformationsManager(this);
+        }
-        transformation->registerMatcherIn(pass, context);
+        pass.run_on_function(function);
+    }
+
+    {
+        GraphRewrite pass;
+        for (auto it : transformations) {
+            ngraph::pass::low_precision::LayerTransformationPtr transformation = it.second;
+            transformation->registerMatcherIn(pass, context);
+        }
+        pass.run_on_function(function);
     }
-    pass.run_on_function(function);
 }

inference-engine/tests/functional/inference_engine/lp_transformations/simple_low_precision_transformer.hpp

@@ -28,9 +28,22 @@ public:
     bool isQuantized(const std::shared_ptr<ngraph::Node>& layer) const noexcept override;
     bool isPrecisionPreserved(const std::shared_ptr<ngraph::Node>& layer) const noexcept override;
 
+    template <class T, class Operation>
+    ngraph::pass::low_precision::LayerTransformationPtr addBranchSpecific(const ngraph::pass::low_precision::LayerTransformation::Params& params) {
+        const std::string typeName = ngraph::pass::low_precision::LowPrecisionTransformations::getType<Operation>();
+
+        const auto it = branchSpecificTransformations.find(typeName);
+        if (it != branchSpecificTransformations.end()) {
+            branchSpecificTransformations.erase(it);
+        }
+
+        auto transformation = std::make_shared<T>(params);
+        branchSpecificTransformations.emplace(typeName, transformation);
+        return transformation;
+    }
+
     template <class T, class Operation>
     ngraph::pass::low_precision::LayerTransformationPtr add(const ngraph::pass::low_precision::LayerTransformation::Params& params) {
-        // const std::string typeName = typeid(ngraph::op::TypeRelaxed<Operation>).name();
         const std::string typeName = ngraph::pass::low_precision::LowPrecisionTransformations::getType<Operation>();
 
         const auto it = transformations.find(typeName);
@@ -46,5 +59,6 @@ public:
     void transform(std::shared_ptr<ngraph::Function>& function);
 
 private:
+    std::map<std::string, ngraph::pass::low_precision::LayerTransformationPtr> branchSpecificTransformations;
     std::map<std::string, ngraph::pass::low_precision::LayerTransformationPtr> transformations;
 };

inference-engine/tests/ngraph_helpers/lpt_ngraph_functions/include/lpt_ngraph_functions/concat_function.hpp

@@ -51,6 +51,12 @@ public:
         const FakeQuantizeOnData& fqOnData1,
         const FakeQuantizeOnData& fqOnData2);
 
+    static std::shared_ptr<ngraph::Function> getOriginalWithIntermediateAvgPool(
+        const ngraph::element::Type precision,
+        const ngraph::Shape& inputShape,
+        const FakeQuantizeOnData& fqOnData1,
+        const FakeQuantizeOnData& fqOnData2);
+
     static std::shared_ptr<ngraph::Function> getOriginalWithSplitedIntermediate(
         const ngraph::element::Type precision,
         const ngraph::Shape& inputShape,
@@ -134,6 +140,7 @@ public:
         const std::string& neighborType,
         const std::string& additionalLayer);
 
+    // TODO: refactor: dequantizationBefore2 <=> dequantizationOperations2
     static std::shared_ptr<ngraph::Function> getReferenceWithIntermediate(
         const ngraph::element::Type precision,
         const ngraph::Shape& inputShape,
@@ -142,6 +149,18 @@ public:
         const FakeQuantizeOnData& fqOnData2,
         const ngraph::element::Type precisionBeforeOp,
         const DequantizationOperations& dequantizationBefore1,
+        const DequantizationOperations& dequantizationOperations2,
+        const ngraph::element::Type precisionAfterOperation,
+        const DequantizationOperations& dequantizationOperations1,
+        const DequantizationOperations& dequantizationBefore2);
+
+    static std::shared_ptr<ngraph::Function> getReferenceWithIntermediateAvgPool(
+        const ngraph::element::Type precision,
+        const ngraph::Shape& inputShape,
+        const FakeQuantizeOnData& fqOnData1,
+        const FakeQuantizeOnData& fqOnData2,
+        const ngraph::element::Type precisionBeforeOp,
+        const DequantizationOperations& dequantizationBefore1,
         const DequantizationOperations& dequantizationBefore2,
         const ngraph::element::Type precisionAfterOperation,
         const DequantizationOperations& dequantizationOperations1,

inference-engine/tests/ngraph_helpers/lpt_ngraph_functions/src/concat_function.cpp

@@ -272,6 +272,58 @@ std::shared_ptr<ngraph::Function> ConcatFunction::getOriginalWithIntermediate(
     return function;
 }
 
+std::shared_ptr<ngraph::Function> ConcatFunction::getOriginalWithIntermediateAvgPool(
+    const ngraph::element::Type precision,
+    const ngraph::Shape& inputShape,
+    const FakeQuantizeOnData& fqOnData1,
+    const FakeQuantizeOnData& fqOnData2) {
+    const std::vector<size_t> inputShape1 = { inputShape[0], inputShape[1], inputShape[2] - 2, inputShape[3] - 2 };
+
+    const auto input1 = std::make_shared<ngraph::opset1::Parameter>(precision, ngraph::Shape(inputShape1));
+    input1->set_friendly_name("input1");
+    const auto fakeQuantize1 = makeFakeQuantize(input1, precision, fqOnData1);
+    fakeQuantize1->set_friendly_name("fakeQuantize1");
+
+    const std::vector<size_t> inputShape2 = { inputShape[0], inputShape[1], inputShape[2], inputShape[3] };
+    const auto input2 = std::make_shared<ngraph::opset1::Parameter>(precision, ngraph::Shape(inputShape2));
+    input2->set_friendly_name("input2");
+
+    const auto fakeQuantize2 = makeFakeQuantize(input2, precision, fqOnData2);
+    fakeQuantize2->set_friendly_name("fakeQuantize2");
+
+    std::shared_ptr<Node> intermediateOp = makeMaxPool(fakeQuantize2->output(0), { 3, 3 });
+    intermediateOp->set_friendly_name("intermediate");
+
+    const std::shared_ptr<ngraph::opset1::Concat> concat = std::make_shared<ngraph::opset1::Concat>(
+        ngraph::OutputVector{ fakeQuantize1->output(0), intermediateOp->output(0) }, 1);
+    concat->set_friendly_name("concat");
+
+    auto& rtInfo = concat->get_rt_info();
+    rtInfo["Variant::std::string"] = std::make_shared<VariantWrapper<std::string>>("concat");
+
+    std::shared_ptr<Node> parent2 = std::make_shared<ngraph::opset1::AvgPool>(
+        intermediateOp,
+        Strides{ 1, 1 },
+        Shape{ 1, 1 },
+        Shape{ 0, 0 },
+        Shape{ 2, 2 },
+        true,
+        op::RoundingType::FLOOR);
+    parent2->set_friendly_name("avgPool");
+
+    ngraph::ResultVector results {
+        std::make_shared<ngraph::opset1::Result>(concat),
+        std::make_shared<ngraph::opset1::Result>(parent2)
+    };
+
+    std::shared_ptr<ngraph::Function> function = std::make_shared<ngraph::Function>(
+        results,
+        ngraph::ParameterVector{ input1, input2 },
+        "ConcatWithIntermediateTransformation");
+
+    return function;
+}
+
 std::shared_ptr<ngraph::Function> ConcatFunction::getOriginalWithSplitedIntermediate(
     const ngraph::element::Type precision,
     const ngraph::Shape& inputShape,
@@ -1056,6 +1108,77 @@ std::shared_ptr<ngraph::Function> ConcatFunction::getReferenceWithIntermediate(
     return function;
 }
 
+std::shared_ptr<ngraph::Function> ConcatFunction::getReferenceWithIntermediateAvgPool(
+    const ngraph::element::Type precision,
+    const ngraph::Shape& inputShape,
+    const FakeQuantizeOnData& fqOnData1,
+    const FakeQuantizeOnData& fqOnData2,
+    const ngraph::element::Type precisionBeforeOp,
+    const DequantizationOperations& dequantizationBefore1,
+    const DequantizationOperations& dequantizationBefore2,
+    const ngraph::element::Type precisionAfterOperation,
+    const DequantizationOperations& dequantizationAfter1,
+    const DequantizationOperations& dequantizationAfter2) {
+    const std::vector<size_t> inputShape1 = { inputShape[0], inputShape[1], inputShape[2] - 2, inputShape[3] - 2};
+    const auto input1 = std::make_shared<ngraph::opset1::Parameter>(precision, ngraph::Shape(inputShape1));
+    input1->set_friendly_name("input1");
+
+    const auto fakeQuantize1 = makeFakeQuantizeTypeRelaxed(input1, precision, fqOnData1);
+    low_precision::NetworkHelper::setOutDataPrecisionForTypeRelaxed(fakeQuantize1, precisionBeforeOp);
+    fakeQuantize1->set_friendly_name("fakeQuantize1");
+    const auto deqBefore1 = makeDequantization(fakeQuantize1, dequantizationBefore1);
+
+    const std::vector<size_t> inputShape2 = { inputShape[0], inputShape[1], inputShape[2], inputShape[3] };
+    const auto input2 = std::make_shared<ngraph::opset1::Parameter>(precision, ngraph::Shape(inputShape2));
+    input2->set_friendly_name("input2");
+
+    const auto fakeQuantize2 = makeFakeQuantizeTypeRelaxed(input2, precision, fqOnData2);
+    low_precision::NetworkHelper::setOutDataPrecisionForTypeRelaxed(fakeQuantize2, precisionBeforeOp);
+    fakeQuantize2->set_friendly_name("fakeQuantize2");
+    const auto deqBefore2 = makeDequantization(fakeQuantize2, dequantizationBefore2);
+
+    std::shared_ptr<Node> intermediateOp  = makeMaxPool(deqBefore2, { 3, 3 });
+    intermediateOp->set_friendly_name("intermediate");
+
+    const std::shared_ptr<ngraph::opset1::Concat> concat = std::make_shared<ngraph::opset1::Concat>(
+        ngraph::OutputVector { deqBefore1, intermediateOp },
+        1);
+    concat->set_friendly_name("concat");
+    low_precision::NetworkHelper::setOutDataPrecision(concat, precisionAfterOperation);
+
+    auto& rtInfo = concat->get_rt_info();
+    rtInfo["Variant::std::string"] = std::make_shared<VariantWrapper<std::string>>("concat");
+
+    const std::shared_ptr<ngraph::Node> parent1 = makeDequantization(concat, dequantizationAfter1);
+    parent1->set_friendly_name("concat");
+
+    std::shared_ptr<Node> parent2 = std::make_shared<ngraph::op::TypeRelaxed<ngraph::opset1::AvgPool>>(
+        std::vector<ngraph::element::Type>{ element::f32, element::f32 },
+        std::vector<ngraph::element::Type>{ element::f32 },
+        ngraph::op::TemporaryReplaceOutputType(intermediateOp, element::f32).get(),
+        Strides{ 1, 1 },
+        Shape{ 1, 1 },
+        Shape{ 0, 0 },
+        Shape{ 2, 2 },
+        true,
+        op::RoundingType::FLOOR);
+    parent2->set_friendly_name("avgPool");
+
+    parent2 = makeDequantization(parent2, dequantizationAfter2);
+
+    ngraph::ResultVector results {
+        std::make_shared<ngraph::opset1::Result>(parent1),
+        std::make_shared<ngraph::opset1::Result>(parent2)
+    };
+
+    std::shared_ptr<ngraph::Function> function = std::make_shared<ngraph::Function>(
+        results,
+        ngraph::ParameterVector{ input1, input2 },
+        "ConcatWithIntermediateTransformation");
+
+    return function;
+}
+
 std::shared_ptr<ngraph::Function> ConcatFunction::getReferenceWithSplitedIntermediate(
     const ngraph::element::Type precision,
     const ngraph::Shape& inputShape,