path: root/src/armnn/test/optimizations/FuseBatchNormTests.cpp

//
// Copyright © 2020 Arm Ltd and Contributors. All rights reserved.
// SPDX-License-Identifier: MIT
//

#include "LayersFwd.hpp"

#include <Network.hpp>
#include <ResolveType.hpp>
#include <armnn/INetwork.hpp>
#include <TestUtils.hpp>

#include <doctest/doctest.h>

using namespace armnn;

TEST_SUITE("Optimizer")
{
namespace
{

class Conv2dTest
{
public:
    using ConvDescriptorType            = armnn::Convolution2dDescriptor;
    using ConvLayerType                 = armnn::Convolution2dLayer;

    static IConnectableLayer *AddConvolution(INetwork *network,
                                             const Convolution2dDescriptor &descriptor,
                                             const ConstTensor &weights,
                                             const Optional<ConstTensor> &biases,
                                             const char *name)
    {
        ARMNN_NO_DEPRECATE_WARN_BEGIN
        return network->AddConvolution2dLayer(descriptor, weights, biases, name);
        ARMNN_NO_DEPRECATE_WARN_END
    }

    static std::vector<IConnectableLayer*> AddConstantLayers(INetwork *network,
                                                             const Convolution2dDescriptor &descriptor,
                                                             const ConstTensor &weights,
                                                             const Optional<ConstTensor> &biases)
    {
        IgnoreUnused(network);
        IgnoreUnused(descriptor);
        IgnoreUnused(weights);
        IgnoreUnused(biases);

        return {};
    }
};

class DepthwiseConv2dTest
{
public:
    using ConvDescriptorType            = armnn::DepthwiseConvolution2dDescriptor;
    using ConvLayerType                 = armnn::DepthwiseConvolution2dLayer;

    static IConnectableLayer* AddConvolution(INetwork* network,
                                             const DepthwiseConvolution2dDescriptor& descriptor,
                                             const ConstTensor& weights,
                                             const Optional<ConstTensor>& biases,
                                             const char* name)
    {
        IgnoreUnused(weights);
        IgnoreUnused(biases);

        return network->AddDepthwiseConvolution2dLayer(descriptor, name);
    }

    static std::vector<IConnectableLayer*> AddConstantLayers(INetwork *network,
                                                             const DepthwiseConvolution2dDescriptor &descriptor,
                                                             const ConstTensor &weights,
                                                             const Optional<ConstTensor> &biases)
    {
        auto weightsLayer = network->AddConstantLayer(weights, "Weights");
        weightsLayer->GetOutputSlot(0).SetTensorInfo(weights.GetInfo());
        std::vector<IConnectableLayer*> layers = {weightsLayer};

        if (descriptor.m_BiasEnabled)
        {
            auto biasLayer = network->AddConstantLayer(biases.value(), "Bias");
            biasLayer->GetOutputSlot(0).SetTensorInfo(biases.value().GetInfo());
            layers.emplace_back(biasLayer);
        }

        return layers;
    }
};

template<typename T>
std::vector<T> GetVector(unsigned int size, float initial, float increment)
{
    std::vector<float> typeVector(size, initial);
    std::vector<T> vector(size);

    if (size > 1)
    {
        for (unsigned int i = 0; i < size; ++i)
        {
            vector[i] = T(initial + (increment * static_cast<float>(i)));
        }
    }
    return vector;
}

} // namespace

template <typename Conv2dTest,
          armnn::DataType ArmnnType,
          typename ConvDescriptorType = typename Conv2dTest::ConvDescriptorType,
          typename T = armnn::ResolveType<ArmnnType>>
INetworkPtr CreateNetwork(bool depthwise, bool preventFusing)
{
    // Define layers information
    ConvDescriptorType convolution2dDescriptor;
    convolution2dDescriptor.m_BiasEnabled = false;
    convolution2dDescriptor.m_DataLayout = DataLayout::NHWC;
    convolution2dDescriptor.m_StrideX = 1;
    convolution2dDescriptor.m_StrideY = 1;
    BatchNormalizationDescriptor batchNormDescriptor;
    batchNormDescriptor.m_DataLayout = DataLayout::NHWC;

    const unsigned int inputDimensionSizes[] = {1, 4, 4, 3};  // NHWCin
    unsigned int weightsDimensionSizes[]     = {4, 2, 2, 3};  // CoutHWCin
    unsigned int outputDimensionSizes[]      = {1, 3, 3, 4};  // NHWCout

    if (depthwise)
    {
        // [1, H, W, Cout]
        weightsDimensionSizes[0] = 1;
        weightsDimensionSizes[1] = 2;
        weightsDimensionSizes[2] = 2;
        weightsDimensionSizes[3] = 12;
        outputDimensionSizes[3]  = weightsDimensionSizes[3];
    }
    const unsigned int outputChannelSize[]   = {outputDimensionSizes[3]};  // Cout

    TensorInfo inputInfo(4, inputDimensionSizes, ArmnnType);
    TensorInfo outputInfo(4, outputDimensionSizes, ArmnnType);

    std::vector<int> weightsIntVector = { 1,  2,  3,  4,   5,  6,  7,  8,   9, 10, 11, 12,
                                         11, 12, 13, 14,  15, 16, 17, 18,  19, 20, 21, 22,
                                         21, 22, 23, 24,  25, 26, 27, 28,  29, 30, 31, 32,
                                         31, 32, 33, 34,  35, 36, 37, 38,  39, 40, 41, 42};
    std::vector<T> weightsVector(begin(weightsIntVector), end(weightsIntVector));
    TensorInfo weightsInfo(4, weightsDimensionSizes, ArmnnType, 0.0f, 0, true);
    ConstTensor weights(weightsInfo, weightsVector);

    std::vector<T> betaVector     = GetVector<T>(outputDimensionSizes[3], 0.0f, 0.2f);
    std::vector<T> gammaVector    = GetVector<T>(outputDimensionSizes[3], 0.5f, 0.1f);
    std::vector<T> meanVector     = GetVector<T>(outputDimensionSizes[3], 0.1f, 0.1f);
    std::vector<T> varianceVector = GetVector<T>(outputDimensionSizes[3], 1.0f, 0.1f);

    ConstTensor beta    (TensorInfo(1, outputChannelSize, ArmnnType, 0.0f, 0, true), betaVector);
    ConstTensor gamma   (TensorInfo(1, outputChannelSize, ArmnnType, 0.0f, 0, true), gammaVector);
    ConstTensor mean    (TensorInfo(1, outputChannelSize, ArmnnType, 0.0f, 0, true), meanVector);
    ConstTensor variance(TensorInfo(1, outputChannelSize, ArmnnType, 0.0f, 0, true), varianceVector);

    // Create a network
    INetworkPtr network = INetwork::Create();

    IConnectableLayer* inputLayer     = network->AddInputLayer(0);

    IConnectableLayer* convLayer      = Conv2dTest::AddConvolution(network.get(),
                                                                   convolution2dDescriptor,
                                                                   weights,
                                                                   Optional<ConstTensor>(),
                                                                   "convolution");

    IConnectableLayer* batchNormLayer = network->AddBatchNormalizationLayer(batchNormDescriptor,
                                                                            mean,
                                                                            variance,
                                                                            beta,
                                                                            gamma,
                                                                            "batchNorm");

    IConnectableLayer* outputLayer    = network->AddOutputLayer(0);
    IConnectableLayer* output2Layer   = nullptr;

    if (preventFusing)
    {
        output2Layer                  = network->AddOutputLayer(1);
    }

    std::vector<IConnectableLayer*> constantLayers = Conv2dTest::AddConstantLayers(network.get(),
                                                                                   convolution2dDescriptor,
                                                                                   weights,
                                                                                   Optional<ConstTensor>());

    // Connect constant layers to receiverLayer.
    for (unsigned int i = 0; i < constantLayers.size(); ++i)
    {
        constantLayers[i]->GetOutputSlot(0).Connect(convLayer->GetInputSlot(i + 1));
    }

    // Set layer information
    inputLayer    ->GetOutputSlot(0).SetTensorInfo(inputInfo);
    convLayer     ->GetOutputSlot(0).SetTensorInfo(outputInfo);
    batchNormLayer->GetOutputSlot(0).SetTensorInfo(outputInfo);

    // Connect layers
    inputLayer    ->GetOutputSlot(0).Connect(convLayer->GetInputSlot(0));
    convLayer     ->GetOutputSlot(0).Connect(batchNormLayer->GetInputSlot(0));
    batchNormLayer->GetOutputSlot(0).Connect(outputLayer->GetInputSlot(0));

    if (preventFusing)
    {
        convLayer ->GetOutputSlot(0).Connect(output2Layer->GetInputSlot(0));
    }

    return network;
}

template <typename Conv2dTest,
          armnn::DataType ArmnnType,
          typename ConvDescriptorType = typename Conv2dTest::ConvDescriptorType,
          typename ConvLayerType = typename Conv2dTest::ConvLayerType,
          typename T = armnn::ResolveType<ArmnnType>>
void FuseBatchNormIntoConvTest(bool depthwise, float tolerance, armnn::Compute backendId)
{
    // FIRST NETWORK: Fused
    // Construct ArmNN network
    INetworkPtr networkFused = CreateNetwork<Conv2dTest, ArmnnType>(depthwise, false);

    // Create ArmNN runtime
    IRuntimePtr run = IRuntime::Create(IRuntime::CreationOptions()); // default options

    // Optimise ArmNN network
    IOptimizedNetworkPtr optNetFused = Optimize(*networkFused, {backendId}, run->GetDeviceSpec());

    Graph& graphFused = GetGraphForTesting(optNetFused.get());

    auto checkFusedConv2d = [ ](const armnn::Layer* const layer) -> bool
    {
        return IsLayerOfType<ConvLayerType>(layer) &&
               (layer->GetNameStr() == "fused-batchNorm-into-convolution");
    };

    CHECK(5 == graphFused.GetNumLayers());
    CHECK(CheckSequence(graphFused.cbegin(),
                        graphFused.cend(),
                        &IsLayerOfType<InputLayer>,
                        &IsLayerOfType<ConstantLayer>,
                        &IsLayerOfType<ConstantLayer>,
                        checkFusedConv2d,
                        &IsLayerOfType<OutputLayer>));

    // Load network into runtime
    NetworkId networkIdentifier;
    CHECK(run->LoadNetwork(networkIdentifier, std::move(optNetFused)) == Status::Success);

    //Creates structures for inputs and outputs.
    std::vector<T> inputDataFused = GetVector<T>(48, 1.0f, 0.1f);

    std::vector<T> outputDataFused(36);

    if (depthwise)
    {
        outputDataFused.resize(108);
    }

    TensorInfo inputTensorInfo = run->GetInputTensorInfo(networkIdentifier, 0);
    inputTensorInfo.SetConstant(true);
    InputTensors inputTensorsFused {
            {0, ConstTensor(inputTensorInfo, inputDataFused.data())}};
    OutputTensors outputTensorsFused{
            {0, Tensor(run->GetOutputTensorInfo(networkIdentifier, 0), outputDataFused.data())}};

    // Execute network
    run->EnqueueWorkload(networkIdentifier, inputTensorsFused, outputTensorsFused);

    // SECOND NETWORK: NotFused
    // Construct ArmNN network
    INetworkPtr networkNotFused = CreateNetwork<Conv2dTest, ArmnnType>(depthwise, true);

    // Create ArmNN runtime
    IRuntimePtr runNotFused = IRuntime::Create(IRuntime::CreationOptions()); // default options

    // Optimise ArmNN network
    IOptimizedNetworkPtr optNetNotFused = Optimize(*networkNotFused, { backendId }, runNotFused->GetDeviceSpec());

    Graph& graphNotFused = GetGraphForTesting(optNetNotFused.get());

    CHECK(6 == graphNotFused.GetNumLayers());
    CHECK(CheckSequence(graphNotFused.cbegin(),
                        graphNotFused.cend(),
                        &IsLayerOfType<armnn::InputLayer>,
                        &IsLayerOfType<armnn::ConstantLayer>,
                        &IsLayerOfType<ConvLayerType>,
                        &IsLayerOfType<armnn::BatchNormalizationLayer>,
                        &IsLayerOfType<armnn::OutputLayer>,
                        &IsLayerOfType<armnn::OutputLayer>));

    // Load network into runtime
    NetworkId networkIdentifierNotFused;
    CHECK(runNotFused->LoadNetwork(networkIdentifierNotFused, std::move(optNetNotFused)) == Status::Success);

    //Creates structures for inputs and outputs.
    std::vector<T> inputDataNotFused = GetVector<T>(48, 1.0f, 0.1f);

    std::vector<T> outputDataNotFused(36);
    std::vector<T> outputData2NotFused(36);

    if (depthwise)
    {
        outputDataNotFused.resize(108);
        outputData2NotFused.resize(108);
    }

    TensorInfo inputTensorInfo2 = runNotFused->GetInputTensorInfo(networkIdentifierNotFused, 0);
    inputTensorInfo2.SetConstant(true);
    InputTensors inputTensorsNotFused{
        { 0, ConstTensor(inputTensorInfo2, inputDataNotFused.data()) } };
    OutputTensors outputTensorsNotFused{
        { 0, Tensor(runNotFused->GetOutputTensorInfo(networkIdentifierNotFused, 0), outputDataNotFused.data()) },
        { 1, Tensor(runNotFused->GetOutputTensorInfo(networkIdentifierNotFused, 1), outputData2NotFused.data()) } };

    // Execute network
    runNotFused->EnqueueWorkload(networkIdentifierNotFused, inputTensorsNotFused, outputTensorsNotFused);

    // Check the output of the fused-convolution matches with the output of the batchNormm in the "NotFused" network
    auto epsilon = T(tolerance);
    for (unsigned int n = 0; n < outputDataFused.size(); ++n)
    {
        CHECK_EQ(outputDataFused[n], doctest::Approx(outputDataNotFused[n]).epsilon(epsilon));
    }
}

// This unit test needs the reference backend, it's not available if the reference backend is not built
#if defined(ARMNNREF_ENABLED)
TEST_CASE("FuseBatchNormIntoConv2DFloat32Test")
{
    FuseBatchNormIntoConvTest<Conv2dTest, DataType::Float32>(false, 0.0001f, armnn::Compute::CpuRef);
}

TEST_CASE("FuseBatchNormIntoConv2DFloat16Test")
{
    FuseBatchNormIntoConvTest<Conv2dTest, DataType::Float16>(false, 0.1f, armnn::Compute::CpuRef);
}

TEST_CASE("FuseBatchNormIntoDepthwiseConv2DFloat32Test")
{
    FuseBatchNormIntoConvTest<DepthwiseConv2dTest, DataType::Float32>(true, 0.0001f,armnn::Compute::CpuRef);
}

TEST_CASE("FuseBatchNormIntoDepthwiseConv2DFloat16Test")
{
    FuseBatchNormIntoConvTest<DepthwiseConv2dTest, DataType::Float16>(true, 0.2f,armnn::Compute::CpuRef);
}
#endif

}