IVGCVSW-3650 Add TfLite Parser support for Transpose Convolution layer

* Added ParseTransposeConv to TfLite Parser
* New TransposeConv test file
* Updated documentation for supported Transpose Convolution

Change-Id: Id7356d8525556805c164af693ae2b16f6a8492fa
Signed-off-by: Matthew Jackson <matthew.jackson@arm.com>

5 files changed, 270 insertions, 3 deletions

diff --git a/CMakeLists.txt b/CMakeLists.txt
index 05dd0ec16d..9c186e3e6e 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -603,6 +603,7 @@ if(BUILD_UNIT_TESTS)
              src/armnnTfLiteParser/test/Squeeze.cpp
              src/armnnTfLiteParser/test/StridedSlice.cpp
              src/armnnTfLiteParser/test/Sub.cpp
+             src/armnnTfLiteParser/test/TransposeConv.cpp
              src/armnnTfLiteParser/test/Unpack.cpp
              src/armnnTfLiteParser/test/LoadModel.cpp
              src/armnnTfLiteParser/test/GetBuffer.cpp
diff --git a/src/armnnTfLiteParser/TensorFlowLiteSupport.md b/src/armnnTfLiteParser/TensorFlowLiteSupport.md
index 7acbf28397..a4c48ca6a0 100644
--- a/src/armnnTfLiteParser/TensorFlowLiteSupport.md
+++ b/src/armnnTfLiteParser/TensorFlowLiteSupport.md
@@ -60,6 +60,8 @@ The Arm NN SDK TensorFlow Lite parser currently supports the following operators
 
 * TANH
 
+* TRANSPOSE_CONV
+
 * UNPACK
 
 ## Custom Operator
diff --git a/src/armnnTfLiteParser/TfLiteParser.cpp b/src/armnnTfLiteParser/TfLiteParser.cpp
index f345d4a6e1..0e11a5c3e1 100644
--- a/src/armnnTfLiteParser/TfLiteParser.cpp
+++ b/src/armnnTfLiteParser/TfLiteParser.cpp
@@ -460,6 +460,7 @@ TfLiteParser::TfLiteParser()
     m_ParserFunctions[tflite::BuiltinOperator_PAD]               =  &TfLiteParser::ParsePad;
     m_ParserFunctions[tflite::BuiltinOperator_SPLIT]             =  &TfLiteParser::ParseSplit;
     m_ParserFunctions[tflite::BuiltinOperator_TANH]              =  &TfLiteParser::ParseTanH;
+    m_ParserFunctions[tflite::BuiltinOperator_TRANSPOSE_CONV]    =  &TfLiteParser::ParseTransposeConv;
     m_ParserFunctions[tflite::BuiltinOperator_UNPACK]            =  &TfLiteParser::ParseUnpack;
 }
 
@@ -737,7 +738,7 @@ void TfLiteParser::ParseConv2D(size_t subgraphIndex, size_t operatorIndex)
     auto filterTensorAndData = CreateConstTensor(inputs[1],
                                                  filterTensorInfo,
                                                  armnn::Optional<armnn::PermutationVector&>());
-    armnn::IConnectableLayer* layer;
+    armnn::IConnectableLayer* layer = nullptr;
 
     auto layerName = boost::str(boost::format("Conv2D:%1%:%2%") % subgraphIndex % operatorIndex);
 
@@ -826,7 +827,7 @@ void TfLiteParser::ParseDepthwiseConv2D(size_t subgraphIndex, size_t operatorInd
                 desc.m_DilationX, desc.m_PadLeft, desc.m_PadRight, options->padding);
 
     auto filterTensorAndData = CreateConstTensor(inputs[1], filterTensorInfo, permutationVector);
-    armnn::IConnectableLayer* layer;
+    armnn::IConnectableLayer* layer = nullptr;
     auto layerName = boost::str(boost::format("DepthwiseConv2D:%1%:%2%") % subgraphIndex % operatorIndex);
 
     if (inputs.size() == 3)
@@ -864,6 +865,91 @@ void TfLiteParser::ParseDepthwiseConv2D(size_t subgraphIndex, size_t operatorInd
     RegisterOutputSlots(subgraphIndex, operatorIndex, layer, {outputTensorIndexes[0]});
 }
 
+void TfLiteParser::ParseTransposeConv(size_t subgraphIndex, size_t operatorIndex)
+{
+    CHECK_MODEL(m_Model, subgraphIndex, operatorIndex);
+
+    const auto & operatorPtr = m_Model->subgraphs[subgraphIndex]->operators[operatorIndex];
+    const auto * options = operatorPtr->builtin_options.AsTransposeConvOptions();
+
+    TransposeConvolution2dDescriptor desc;
+    desc.m_BiasEnabled = false;
+    desc.m_StrideX = CHECKED_NON_NEGATIVE(options->stride_w);
+    desc.m_StrideY = CHECKED_NON_NEGATIVE(options->stride_h);
+    desc.m_DataLayout = armnn::DataLayout::NHWC;
+
+    auto inputs = GetInputs(m_Model, subgraphIndex, operatorIndex);
+    CHECK_VALID_SIZE(inputs.size(), 2, 3);
+
+    auto outputs = GetOutputs(m_Model, subgraphIndex, operatorIndex);
+    CHECK_VALID_SIZE(outputs.size(), 1);
+
+    armnn::TensorInfo inputTensorInfo  = ToTensorInfo(inputs[0]);
+    armnn::TensorInfo filterTensorInfo = ToTensorInfo(inputs[1]);
+
+    // TfLite uses NHWC tensors
+    const unsigned int inputHeight = inputTensorInfo.GetShape()[1];
+    const unsigned int inputWidth  = inputTensorInfo.GetShape()[2];
+
+    const unsigned int filterHeight = filterTensorInfo.GetShape()[1];
+    const unsigned int filterWidth  = filterTensorInfo.GetShape()[2];
+
+    CalcPadding(inputHeight,
+                filterHeight,
+                desc.m_StrideY,
+                1, // DilationY
+                desc.m_PadTop,
+                desc.m_PadBottom,
+                options->padding);
+
+    CalcPadding(inputWidth,
+                filterWidth,
+                desc.m_StrideX,
+                1, // DilationX
+                desc.m_PadLeft,
+                desc.m_PadRight,
+                options->padding);
+
+    auto filterTensorAndData = CreateConstTensor(inputs[1],
+                                                 filterTensorInfo,
+                                                 armnn::Optional<armnn::PermutationVector&>());
+
+    armnn::IConnectableLayer* layer = nullptr;
+    auto layerName = boost::str(boost::format("TransposeConv:%1%:%2%") % subgraphIndex % operatorIndex);
+
+    if (inputs.size() == 3)
+    {
+        desc.m_BiasEnabled = true;
+        armnn::TensorInfo biasTensorInfo = ToTensorInfo(inputs[2]);
+        auto biasTensorAndData = CreateConstTensor(inputs[2],
+                                                   biasTensorInfo,
+                                                   armnn::Optional<armnn::PermutationVector&>());
+        layer = m_Network->AddTransposeConvolution2dLayer(desc,
+                                                          filterTensorAndData.first,
+                                                          Optional<ConstTensor>(biasTensorAndData.first),
+                                                          layerName.c_str());
+    }
+    else
+    {
+        layer = m_Network->AddTransposeConvolution2dLayer(desc,
+                                                          filterTensorAndData.first,
+                                                          EmptyOptional(),
+                                                          layerName.c_str());
+    }
+
+    BOOST_ASSERT(layer != nullptr);
+
+    armnn::TensorInfo outputTensorInfo = ToTensorInfo(outputs[0]);
+    layer->GetOutputSlot(0).SetTensorInfo(outputTensorInfo);
+
+    // only the tensors for the inputs are relevant, exclude the const (filter) tensor
+    auto inputTensorIndexes = AsUnsignedVector(GetInputTensorIds(m_Model, subgraphIndex, operatorIndex));
+    RegisterInputSlots(subgraphIndex, operatorIndex, layer, {inputTensorIndexes[0]});
+
+    auto outputTensorIndexes = AsUnsignedVector(GetOutputTensorIds(m_Model, subgraphIndex, operatorIndex));
+    RegisterOutputSlots(subgraphIndex, operatorIndex, layer, {outputTensorIndexes[0]});
+}
+
 void TfLiteParser::ParseAveragePool2D(size_t subgraphIndex, size_t operatorIndex)
 {
     ParsePool(subgraphIndex, operatorIndex, PoolingAlgorithm::Average);
@@ -1776,7 +1862,7 @@ void TfLiteParser::ParseFullyConnected(size_t subgraphIndex, size_t operatorInde
     auto filterTensorAndData = CreateConstTensor(inputs[1],
                                                  filterTensorInfo,
                                                  armnn::Optional<armnn::PermutationVector&>());
-    armnn::IConnectableLayer* layer;
+    armnn::IConnectableLayer* layer = nullptr;
     auto layerName = boost::str(boost::format("FullyConnected:%1%:%2%") % subgraphIndex % operatorIndex);
 
     if (inputs.size() == 3)
diff --git a/src/armnnTfLiteParser/TfLiteParser.hpp b/src/armnnTfLiteParser/TfLiteParser.hpp
index 90b800d56d..dc0d6344f8 100644
--- a/src/armnnTfLiteParser/TfLiteParser.hpp
+++ b/src/armnnTfLiteParser/TfLiteParser.hpp
@@ -120,6 +120,7 @@ private:
     void ParseStridedSlice(size_t subgraphIndex, size_t operatorIndex);
     void ParseSub(size_t subgraphIndex, size_t operatorIndex);
     void ParseTanH(size_t subgraphIndex, size_t operatorIndex);
+    void ParseTransposeConv(size_t subgraphIndex, size_t operatorIndex);
     void ParseUnpack(size_t subgraphIndex, size_t operatorIndex);
 
     void RegisterProducerOfTensor(size_t subgraphIndex, size_t tensorIndex, armnn::IOutputSlot* slot);
diff --git a/src/armnnTfLiteParser/test/TransposeConv.cpp b/src/armnnTfLiteParser/test/TransposeConv.cpp
new file mode 100644
index 0000000000..05b48ecf0b
--- /dev/null
+++ b/src/armnnTfLiteParser/test/TransposeConv.cpp
@@ -0,0 +1,177 @@
+//
+// Copyright © 2017 Arm Ltd. All rights reserved.
+// SPDX-License-Identifier: MIT
+//
+
+#include <boost/test/unit_test.hpp>
+#include "ParserFlatbuffersFixture.hpp"
+#include "../TfLiteParser.hpp"
+
+BOOST_AUTO_TEST_SUITE(TensorflowLiteParser)
+
+struct TransposeConvFixture : public ParserFlatbuffersFixture
+{
+    explicit TransposeConvFixture(const std::string& inputShape,
+                                  const std::string& outputShape,
+                                  const std::string& filterShape,
+                                  const std::string& filterData,
+                                  const bool biasEnabled,
+                                  const std::string& biasShape,
+                                  const std::string& biasData,
+                                  const std::string& strideX,
+                                  const std::string& strideY,
+                                  const std::string& dataType)
+    {
+        std::string biasString;
+        if (biasEnabled)
+        {
+            biasString = R"(
+                        {
+                            "shape": )" + biasShape + R"(,
+                            "type": "INT32",
+                            "buffer": 2,
+                            "name": "biasTensor",
+                            "quantization": {
+                                "min": [ 0.0 ],
+                                "max": [ 255.0 ],
+                                "scale": [ 1.0 ],
+                                "zero_point": [ 0 ],
+                            }
+                        },)";
+
+        }
+        m_JsonString = R"(
+            {
+                "version": 3,
+                "operator_codes": [ { "builtin_code": "TRANSPOSE_CONV" } ],
+                "subgraphs": [ {
+                    "tensors": [
+                        {
+                            "shape": )" + inputShape + R"(,
+                            "type": ")" + dataType + R"(",
+                            "buffer": 0,
+                            "name": "inputTensor",
+                            "quantization": {
+                                "min": [ 0.0 ],
+                                "max": [ 255.0 ],
+                                "scale": [ 1.0 ],
+                                "zero_point": [ 0 ],
+                            }
+                        },
+                        {
+                            "shape": )" + filterShape + R"(,
+                            "type": ")" + dataType + R"(",
+                            "buffer": 1,
+                            "name": "filterTensor",
+                            "quantization": {
+                                "min": [ 0.0 ],
+                                "max": [ 255.0 ],
+                                "scale": [ 1.0 ],
+                                "zero_point": [ 0 ],
+                            }
+                        },)" + biasString + R"(
+                        {
+                            "shape": )" + outputShape + R"(,
+                            "type": ")" + dataType + R"(",
+                            "buffer": )" + (biasEnabled ? "3" : "2") + R"(,
+                            "name": "outputTensor",
+                            "quantization": {
+                                "min": [ 0.0 ],
+                                "max": [ 255.0 ],
+                                "scale": [ 1.0 ],
+                                "zero_point": [ 0 ],
+                            }
+                        }
+                    ],
+                    "inputs": [ 0 ],
+                    "outputs": [ )" + (biasEnabled ? "3" : "2") + R"( ],
+                    "operators": [
+                        {
+                            "opcode_index": 0,
+                            "inputs": [ 0, 1)" + (biasEnabled ? ", 2" : "") + R"( ],
+                            "outputs": [ )" + (biasEnabled ? "3" : "2") + R"( ],
+                            "builtin_options_type": "TransposeConvOptions",
+                            "builtin_options": {
+                                "padding": "SAME",
+                                "stride_w": )" + strideX + R"(,
+                                "stride_h": )" + strideY + R"(
+                            },
+                            "custom_options_format": "FLEXBUFFERS"
+                        }
+                    ],
+                } ],
+                "buffers" : [
+                    { },
+                    { "data": )" + filterData + R"( },
+                    { )" + (biasEnabled ? (R"("data": )" + biasData) : "") + R"( },
+                    { }
+                ]
+            }
+        )";
+        SetupSingleInputSingleOutput("inputTensor", "outputTensor");
+    }
+};
+
+struct SimpleTransposeConvFixture : TransposeConvFixture
+{
+    SimpleTransposeConvFixture()
+    : TransposeConvFixture("[ 1, 2, 2, 1 ]",  // inputShape
+                           "[ 1, 3, 3, 1 ]",  // outputShape
+                           "[ 1, 2, 2, 1 ]",  // filterShape
+                           "[ 0, 1, 2, 4 ]",  // filterData
+                           false,             // biasEnabled
+                           "",                // biasShape
+                           "",                // biasData
+                           "1",               // strideX
+                           "1",               // strideY
+                           "UINT8")           // dataType
+    {}
+};
+
+BOOST_FIXTURE_TEST_CASE( ParseSimpleTransposeConv, SimpleTransposeConvFixture )
+{
+    RunTest<4, armnn::DataType::QuantisedAsymm8>(
+        0,
+        {
+            1, 2,
+            3, 4
+        },
+        {
+            0, 1,  2,
+            2, 11, 12,
+            6, 20, 16
+        });
+}
+
+struct TransposeConvWithBiasFixture : TransposeConvFixture
+{
+    TransposeConvWithBiasFixture()
+    : TransposeConvFixture("[ 1, 2, 2, 1 ]",  // inputShape
+                           "[ 1, 3, 3, 1 ]",  // outputShape
+                           "[ 1, 2, 2, 1 ]",  // filterShape
+                           "[ 0, 1, 2, 4 ]",  // filterData
+                           true,              // biasEnabled
+                           "[ 1 ]",           // biasShape
+                           "[ 2, 0, 0, 0 ]",  // biasData
+                           "1",               // strideX
+                           "1",               // strideY
+                           "UINT8")           // dataType
+    {}
+};
+
+BOOST_FIXTURE_TEST_CASE( ParseTransposeConvWithBias, TransposeConvWithBiasFixture )
+{
+    RunTest<4, armnn::DataType::QuantisedAsymm8>(
+        0,
+        {
+            1, 2,
+            3, 4
+        },
+        {
+            2, 3,  5,
+            4, 13, 14,
+            8, 22, 18
+        });
+}
+
+BOOST_AUTO_TEST_SUITE_END()