diff options
author | Matthew Sloyan <matthew.sloyan@arm.com> | 2021-07-13 19:46:11 +0100 |
---|---|---|
committer | Matthew Sloyan <matthew.sloyan@arm.com> | 2021-08-06 09:25:26 +0000 |
commit | 81beae3a870004795275e9266bc43d845b9f78db (patch) | |
tree | 70af86f3c36c8e330c72770e6f1419ca7b2a4bb8 /src/backends/backendsCommon/test | |
parent | 95e9efc28ce70a8cda93e722f5ce90ebc96bdd95 (diff) | |
download | armnn-81beae3a870004795275e9266bc43d845b9f78db.tar.gz |
IVGCVSW-6119 ConstTensorsAsInput: FullyConnected
* Constant weights and biases are now stored as Constant layers.
* Updated Serializer, Deserializer and unit tests to reflect this.
* Updated TfLiteDelegate, TfLiteParser and OnnxParser.
* Updated Schema with IsConstant and ConstantTensorsAsInputs.
* Updated Ref backend to handle constant weights and
bias as inputs rather than reading from member variables.
* Added dynamic or constant input EndToEnd tests.
!android-nn-driver:5959
Signed-off-by: Matthew Sloyan <matthew.sloyan@arm.com>
Change-Id: Ibf3cf437df1100e4b322b0d303c575c6339f9696
Diffstat (limited to 'src/backends/backendsCommon/test')
-rw-r--r-- | src/backends/backendsCommon/test/FullyConnectedEndToEndTestImpl.hpp | 170 | ||||
-rw-r--r-- | src/backends/backendsCommon/test/layerTests/FullyConnectedTestImpl.cpp | 112 |
2 files changed, 194 insertions, 88 deletions
diff --git a/src/backends/backendsCommon/test/FullyConnectedEndToEndTestImpl.hpp b/src/backends/backendsCommon/test/FullyConnectedEndToEndTestImpl.hpp index 923d6f3641..af6b56852a 100644 --- a/src/backends/backendsCommon/test/FullyConnectedEndToEndTestImpl.hpp +++ b/src/backends/backendsCommon/test/FullyConnectedEndToEndTestImpl.hpp @@ -28,10 +28,7 @@ armnn::INetworkPtr CreateFullyConnectedNetworkNonConstWeights(const armnn::Tenso armnn::IConnectableLayer* inputLayer = network->AddInputLayer(0, "Input"); armnn::IConnectableLayer* weightsInputLayer = network->AddInputLayer(1, "Weights_Input"); - armnn::IConnectableLayer* fullyConnectedLayer = network->AddFullyConnectedLayer(descriptor, - armnn::EmptyOptional(), - armnn::EmptyOptional(), - "Fully_Connected"); + armnn::IConnectableLayer* fullyConnectedLayer = network->AddFullyConnectedLayer(descriptor, "Fully_Connected"); armnn::IConnectableLayer* outputLayer = network->AddOutputLayer(0, "Output"); Connect(inputLayer, fullyConnectedLayer, inputTensorInfo, 0, 0); @@ -41,6 +38,52 @@ armnn::INetworkPtr CreateFullyConnectedNetworkNonConstWeights(const armnn::Tenso return network; } +armnn::INetworkPtr CreateFullyConnectedNetworkNonConstWeightsConstBias(const armnn::TensorInfo& inputTensorInfo, + const armnn::TensorInfo& outputTensorInfo, + const armnn::TensorInfo& weightsTensorInfo, + const armnn::TensorInfo& biasTensorInfo, + const armnn::ConstTensor& biasConstantTensor, + armnn::FullyConnectedDescriptor descriptor) +{ + armnn::INetworkPtr network(armnn::INetwork::Create()); + + armnn::IConnectableLayer* inputLayer = network->AddInputLayer(0, "Input"); + armnn::IConnectableLayer* weightsInputLayer = network->AddInputLayer(1, "Weights_Input"); + armnn::IConnectableLayer* biasLayer = network->AddConstantLayer(biasConstantTensor, "Weights"); + armnn::IConnectableLayer* fullyConnectedLayer = network->AddFullyConnectedLayer(descriptor, "Fully_Connected"); + armnn::IConnectableLayer* outputLayer = network->AddOutputLayer(0, "Output"); + + Connect(inputLayer, fullyConnectedLayer, inputTensorInfo, 0, 0); + Connect(weightsInputLayer, fullyConnectedLayer, weightsTensorInfo, 0, 1); + Connect(biasLayer, fullyConnectedLayer, biasTensorInfo, 0, 2); + Connect(fullyConnectedLayer, outputLayer, outputTensorInfo, 0, 0); + + return network; +} + +armnn::INetworkPtr CreateFullyConnectedNetworkConstWeightsNonConstBias(const armnn::TensorInfo& inputTensorInfo, + const armnn::TensorInfo& outputTensorInfo, + const armnn::TensorInfo& weightsTensorInfo, + const armnn::TensorInfo& biasTensorInfo, + const armnn::ConstTensor& weightsConstantTensor, + armnn::FullyConnectedDescriptor descriptor) +{ + armnn::INetworkPtr network(armnn::INetwork::Create()); + + armnn::IConnectableLayer* inputLayer = network->AddInputLayer(0, "Input"); + armnn::IConnectableLayer* weightsLayer = network->AddConstantLayer(weightsConstantTensor, "Weights"); + armnn::IConnectableLayer* biasLayer = network->AddInputLayer(2, "Bias_Input"); + armnn::IConnectableLayer* fullyConnectedLayer = network->AddFullyConnectedLayer(descriptor, "Fully_Connected"); + armnn::IConnectableLayer* outputLayer = network->AddOutputLayer(0, "Output"); + + Connect(inputLayer, fullyConnectedLayer, inputTensorInfo, 0, 0); + Connect(weightsLayer, fullyConnectedLayer, weightsTensorInfo, 0, 1); + Connect(biasLayer, fullyConnectedLayer, biasTensorInfo, 0, 2); + Connect(fullyConnectedLayer, outputLayer, outputTensorInfo, 0, 0); + + return network; +} + template<armnn::DataType ArmnnType, typename T = armnn::ResolveType<ArmnnType>> void FullyConnectedWithDynamicWeightsEndToEnd(const std::vector<armnn::BackendId>& backends) { @@ -94,4 +137,123 @@ void FullyConnectedWithDynamicWeightsEndToEnd(const std::vector<armnn::BackendId backends, 1.0f); } + +template<armnn::DataType ArmnnType, typename T = armnn::ResolveType<ArmnnType>> +void FullyConnectedWithDynamicOrConstantInputsEndToEnd(const std::vector<armnn::BackendId>& backends, + const bool transposeWeights, + const bool constantWeightsOrBias) +{ + unsigned int inputWidth = 1; + unsigned int inputHeight = 1; + unsigned int inputChannels = 5; + unsigned int inputNum = 2; + + unsigned int outputChannels = 3; + unsigned int outputNum = 2; + + unsigned int inputShape[] = { inputNum, inputChannels, inputHeight, inputWidth }; + unsigned int outputShape[] = { outputNum, outputChannels }; + unsigned int weightsShape[] = { inputChannels, outputChannels }; + + if (transposeWeights) + { + std::swap(weightsShape[0], weightsShape[1]); + } + + unsigned int biasShape[] = { outputChannels }; + + armnn::TensorInfo inputTensorInfo = armnn::TensorInfo(4, inputShape, armnn::DataType::Float32); + armnn::TensorInfo outputTensorInfo = armnn::TensorInfo(2, outputShape, armnn::DataType::Float32); + armnn::TensorInfo weightsDesc = armnn::TensorInfo(2, weightsShape, armnn::DataType::Float32); + armnn::TensorInfo biasesDesc = armnn::TensorInfo(1, biasShape, armnn::DataType::Float32); + + std::vector<float> input = + { + 1.0f, 2.0f, 3.0f, 4.0f, 5.0f, + 5.0f, 4.0f, 3.0f, 2.0f, 1.0f + }; + + std::vector<float> weights = + { + .5f, 2.f, .5f, + .5f, 2.f, 1.f, + .5f, 2.f, 2.f, + .5f, 2.f, 3.f, + .5f, 2.f, 4.f + }; + + if (transposeWeights) + { + weights = + { + .5f, .5f, .5f, .5f, .5f, + 2.f, 2.f, 2.f, 2.f, 2.f, + .5f, 1.f, 2.f, 3.f, 4.f + }; + } + + std::vector<float> biasValues = std::vector<float>({10.f, 20.f, 30.f}); + + std::vector<float> expectedOutput = + { + 0.5f + 1.0f + 1.5f + 2.0f + 2.5f + biasValues[0], + 2.0f + 4.0f + 6.0f + 8.0f + 10.f + biasValues[1], + 0.5f + 2.0f + 6.0f + 12.f + 20.f + biasValues[2], + + 2.5f + 2.0f + 1.5f + 1.0f + 0.5f + biasValues[0], + 10.0f + 8.0f + 6.0f + 4.0f + 2.f + biasValues[1], + 2.5f + 4.0f + 6.0f + 6.f + 4.f + biasValues[2] + }; + + FullyConnectedDescriptor descriptor; + descriptor.m_BiasEnabled = true; + descriptor.m_TransposeWeightMatrix = transposeWeights; + descriptor.m_ConstantWeights = constantWeightsOrBias; + + if (!constantWeightsOrBias) + { + // Tests non constant weights and constant bias. + ConstTensor biasConstantTensor(biasesDesc, biasValues.data()); + + armnn::INetworkPtr network = CreateFullyConnectedNetworkNonConstWeightsConstBias(inputTensorInfo, + outputTensorInfo, + weightsDesc, + biasesDesc, + biasConstantTensor, + descriptor); + CHECK(network); + + std::map<int, std::vector<T>> inputTensorData = {{ 0, input }, {1, weights}}; + std::map<int, std::vector<T>> expectedOutputTensorData = {{ 0, expectedOutput }}; + + EndToEndLayerTestImpl<ArmnnType, ArmnnType>(move(network), + inputTensorData, + expectedOutputTensorData, + backends, + 1.0f); + } + else + { + // Tests constant weights and non constant bias. + ConstTensor weightsConstantTensor(weightsDesc, weights.data()); + + armnn::INetworkPtr network = CreateFullyConnectedNetworkConstWeightsNonConstBias(inputTensorInfo, + outputTensorInfo, + weightsDesc, + biasesDesc, + weightsConstantTensor, + descriptor); + CHECK(network); + + std::map<int, std::vector<T>> inputTensorData = {{ 0, input }, {2, biasValues}}; + std::map<int, std::vector<T>> expectedOutputTensorData = {{ 0, expectedOutput }}; + + EndToEndLayerTestImpl<ArmnnType, ArmnnType>(move(network), + inputTensorData, + expectedOutputTensorData, + backends, + 1.0f); + } +} + } // anonymous namespace diff --git a/src/backends/backendsCommon/test/layerTests/FullyConnectedTestImpl.cpp b/src/backends/backendsCommon/test/layerTests/FullyConnectedTestImpl.cpp index c47048e566..dcf87fe92b 100644 --- a/src/backends/backendsCommon/test/layerTests/FullyConnectedTestImpl.cpp +++ b/src/backends/backendsCommon/test/layerTests/FullyConnectedTestImpl.cpp @@ -22,56 +22,6 @@ template<typename T, typename B> LayerTestResult<T, 2> SimpleFullyConnectedTestImpl( - armnn::IWorkloadFactory& workloadFactory, - const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager, - const armnn::ITensorHandleFactory& tensorHandleFactory, - armnn::TensorInfo inputTensorInfo, - armnn::TensorInfo outputTensorInfo, - armnn::TensorInfo weightsDesc, - armnn::TensorInfo biasesDesc, - std::vector<T>& weights, - std::vector<B>& bias, - std::vector<T>& input, - bool biasEnabled, - bool transposeWeights) -{ - std::unique_ptr<armnn::ITensorHandle> inputHandle = tensorHandleFactory.CreateTensorHandle(inputTensorInfo); - std::unique_ptr<armnn::ITensorHandle> outputHandle = tensorHandleFactory.CreateTensorHandle(outputTensorInfo); - - armnn::FullyConnectedQueueDescriptor data; - armnn::WorkloadInfo info; - armnn::ScopedTensorHandle weightsTensor(weightsDesc); - armnn::ScopedTensorHandle biasTensor(biasesDesc); - - std::vector<T> actualOutput(outputTensorInfo.GetNumElements()); - - AllocateAndCopyDataToITensorHandle(&weightsTensor, weights.data()); - AllocateAndCopyDataToITensorHandle(&biasTensor, bias.data()); - - AddInputToWorkload(data, info, inputTensorInfo, inputHandle.get()); - AddOutputToWorkload(data, info, outputTensorInfo, outputHandle.get()); - data.m_Weight = &weightsTensor; - data.m_Bias = &biasTensor; - data.m_Parameters.m_BiasEnabled = biasEnabled; - data.m_Parameters.m_TransposeWeightMatrix = transposeWeights; - - std::unique_ptr<armnn::IWorkload> workload = workloadFactory.CreateFullyConnected(data, info); - LayerTestResult<T, 2> result(outputTensorInfo); - - inputHandle->Allocate(); - outputHandle->Allocate(); - CopyDataToITensorHandle(inputHandle.get(), input.data()); - - ExecuteWorkload(*workload, memoryManager); - - CopyDataFromITensorHandle(actualOutput.data(), outputHandle.get()); - result.m_ActualData = actualOutput; - - return result; -} - -template<typename T, typename B> -LayerTestResult<T, 2> SimpleFullyConnectedTestWeightsAsInputsImpl( armnn::IWorkloadFactory& workloadFactory, const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager, const armnn::ITensorHandleFactory& tensorHandleFactory, @@ -83,7 +33,8 @@ LayerTestResult<T, 2> SimpleFullyConnectedTestWeightsAsInputsImpl( std::vector<B>& bias, std::vector<T>& input, bool biasEnabled, - bool transposeWeights) + bool transposeWeights, + bool constantWeights) { std::unique_ptr<armnn::ITensorHandle> input0Handle = tensorHandleFactory.CreateTensorHandle(inputTensorInfo); std::unique_ptr<armnn::ITensorHandle> input1Handle = tensorHandleFactory.CreateTensorHandle(weightsTensorInfo); @@ -93,13 +44,23 @@ LayerTestResult<T, 2> SimpleFullyConnectedTestWeightsAsInputsImpl( armnn::FullyConnectedQueueDescriptor data; armnn::WorkloadInfo info; + armnn::ScopedTensorHandle weightsTensor(weightsTensorInfo); + armnn::ScopedTensorHandle biasTensor(biasesTensorInfo); + + AllocateAndCopyDataToITensorHandle(&weightsTensor, weights.data()); + AllocateAndCopyDataToITensorHandle(&biasTensor, bias.data()); AddInputToWorkload(data, info, inputTensorInfo, input0Handle.get()); AddInputToWorkload(data, info, weightsTensorInfo, input1Handle.get()); AddOutputToWorkload(data, info, outputTensorInfo, outputHandle.get()); + + // Need to set as layer members will be null when creating the workload because the optimization hasn't been run. + data.m_Weight = &weightsTensor; + data.m_Bias = &biasTensor; + data.m_Parameters.m_BiasEnabled = biasEnabled; data.m_Parameters.m_TransposeWeightMatrix = transposeWeights; - data.m_Parameters.m_ConstantWeights = false; + data.m_Parameters.m_ConstantWeights = constantWeights; std::unique_ptr<armnn::ITensorHandle> input2Handle = nullptr; if (biasEnabled) @@ -180,36 +141,19 @@ LayerTestResult<T, 2> FullyConnectedTest( std::vector<int32_t> bias = {9250, 67500}; - if (constantWeights) - { - result = SimpleFullyConnectedTestImpl<T>(workloadFactory, - memoryManager, - tensorHandleFactory, - inputTensorInfo, - outputTensorInfo, - weightsDesc, - biasesDesc, - weights, - bias, - input, - biasEnabled, - true); - } - else - { - result = SimpleFullyConnectedTestWeightsAsInputsImpl<T>(workloadFactory, - memoryManager, - tensorHandleFactory, - inputTensorInfo, - outputTensorInfo, - weightsDesc, - biasesDesc, - weights, - bias, - input, - biasEnabled, - true); - } + result = SimpleFullyConnectedTestImpl<T>(workloadFactory, + memoryManager, + tensorHandleFactory, + inputTensorInfo, + outputTensorInfo, + weightsDesc, + biasesDesc, + weights, + bias, + input, + biasEnabled, + true, + constantWeights); if (biasEnabled) { @@ -299,7 +243,7 @@ LayerTestResult<T, 2> FullyConnectedLargeTestCommon( inputTensorInfo, outputTensorInfo, weightsDesc, biasesDesc, weights, biasValues, input, - true, transposeWeights + true, transposeWeights, true ); result.m_ExpectedData = armnnUtils::QuantizedVector<T>({ 965432.0f }, qScale, qOffset); @@ -408,7 +352,7 @@ LayerTestResult<float, 2> FullyConnectedFloat32Test( inputTensorInfo, outputTensorInfo, weightsDesc, biasesDesc, weights, biasValues, input, - biasEnabled, transposeWeights + biasEnabled, transposeWeights, true ); std::vector<float> expectedOutput = |