From 6e36a64e26520e3f169bb2a92972a24e1be915a7 Mon Sep 17 00:00:00 2001 From: Sadik Armagan Date: Tue, 10 Nov 2020 21:18:41 +0000 Subject: IVGCVSW-5389 'TfLiteDelegate: Implement the FullyConnected operator' * Added FullyConnected operator support to delegate Signed-off-by: Sadik Armagan Change-Id: Iae9c0980a4bfd6aa4d90f107f329dfa782baeefe --- delegate/src/test/FullyConnectedTestHelper.hpp | 232 +++++++++++++++++++++++++ 1 file changed, 232 insertions(+) create mode 100644 delegate/src/test/FullyConnectedTestHelper.hpp (limited to 'delegate/src/test/FullyConnectedTestHelper.hpp') diff --git a/delegate/src/test/FullyConnectedTestHelper.hpp b/delegate/src/test/FullyConnectedTestHelper.hpp new file mode 100644 index 0000000000..4eed9580f1 --- /dev/null +++ b/delegate/src/test/FullyConnectedTestHelper.hpp @@ -0,0 +1,232 @@ +// +// Copyright © 2020 Arm Ltd and Contributors. All rights reserved. +// SPDX-License-Identifier: MIT +// + +#pragma once + +#include + +#include +#include +#include +#include +#include +#include + +#include + +namespace +{ + +template +std::vector CreateFullyConnectedTfLiteModel(tflite::TensorType tensorType, + tflite::ActivationFunctionType activationType, + const std::vector & inputTensorShape, + const std::vector & weightsTensorShape, + const std::vector & biasTensorShape, + const std::vector & outputTensorShape, + const std::vector & weightsData, + float quantScale = 1.0f, + int quantOffset = 0, + float outputQuantScale = 2.0f, + int outputQuantOffset = 0) +{ + using namespace tflite; + flatbuffers::FlatBufferBuilder flatBufferBuilder; + std::array, 3> buffers; + buffers[0] = CreateBuffer(flatBufferBuilder, flatBufferBuilder.CreateVector({})); + buffers[1] = CreateBuffer(flatBufferBuilder, + flatBufferBuilder.CreateVector(reinterpret_cast(weightsData.data()), + sizeof(T) * weightsData.size())); + + auto biasTensorType = ::tflite::TensorType_FLOAT32; + if (tensorType == ::tflite::TensorType_UINT8) + { + biasTensorType = ::tflite::TensorType_INT32; + std::vector biasData = { 10 }; + buffers[2] = CreateBuffer(flatBufferBuilder, + flatBufferBuilder.CreateVector(reinterpret_cast(biasData.data()), + sizeof(int32_t) * biasData.size())); + + } + else + { + std::vector biasData = { 10 }; + buffers[2] = CreateBuffer(flatBufferBuilder, + flatBufferBuilder.CreateVector(reinterpret_cast(biasData.data()), + sizeof(float) * biasData.size())); + } + + auto quantizationParameters = + CreateQuantizationParameters(flatBufferBuilder, + 0, + 0, + flatBufferBuilder.CreateVector({ quantScale }), + flatBufferBuilder.CreateVector({ quantOffset })); + + auto outputQuantizationParameters = + CreateQuantizationParameters(flatBufferBuilder, + 0, + 0, + flatBufferBuilder.CreateVector({ outputQuantScale }), + flatBufferBuilder.CreateVector({ outputQuantOffset })); + + std::array, 4> tensors; + tensors[0] = CreateTensor(flatBufferBuilder, + flatBufferBuilder.CreateVector(inputTensorShape.data(), + inputTensorShape.size()), + tensorType, + 0, + flatBufferBuilder.CreateString("input_0"), + quantizationParameters); + tensors[1] = CreateTensor(flatBufferBuilder, + flatBufferBuilder.CreateVector(weightsTensorShape.data(), + weightsTensorShape.size()), + tensorType, + 1, + flatBufferBuilder.CreateString("weights"), + quantizationParameters); + tensors[2] = CreateTensor(flatBufferBuilder, + flatBufferBuilder.CreateVector(biasTensorShape.data(), + biasTensorShape.size()), + biasTensorType, + 2, + flatBufferBuilder.CreateString("bias"), + quantizationParameters); + + tensors[3] = CreateTensor(flatBufferBuilder, + flatBufferBuilder.CreateVector(outputTensorShape.data(), + outputTensorShape.size()), + tensorType, + 0, + flatBufferBuilder.CreateString("output"), + outputQuantizationParameters); + + + // create operator + tflite::BuiltinOptions operatorBuiltinOptionsType = BuiltinOptions_FullyConnectedOptions; + flatbuffers::Offset operatorBuiltinOptions = + CreateFullyConnectedOptions(flatBufferBuilder, + activationType, + FullyConnectedOptionsWeightsFormat_DEFAULT, false).Union(); + + const std::vector operatorInputs{ {0, 1, 2} }; + const std::vector operatorOutputs{ {3} }; + flatbuffers::Offset fullyConnectedOperator = + CreateOperator(flatBufferBuilder, + 0, + flatBufferBuilder.CreateVector(operatorInputs.data(), operatorInputs.size()), + flatBufferBuilder.CreateVector(operatorOutputs.data(), operatorOutputs.size()), + operatorBuiltinOptionsType, operatorBuiltinOptions); + + const std::vector subgraphInputs{ {0, 1, 2} }; + const std::vector subgraphOutputs{ {3} }; + flatbuffers::Offset subgraph = + CreateSubGraph(flatBufferBuilder, + flatBufferBuilder.CreateVector(tensors.data(), tensors.size()), + flatBufferBuilder.CreateVector(subgraphInputs.data(), subgraphInputs.size()), + flatBufferBuilder.CreateVector(subgraphOutputs.data(), subgraphOutputs.size()), + flatBufferBuilder.CreateVector(&fullyConnectedOperator, 1)); + + flatbuffers::Offset modelDescription = + flatBufferBuilder.CreateString("ArmnnDelegate: FullyConnected Operator Model"); + flatbuffers::Offset operatorCode = CreateOperatorCode(flatBufferBuilder, + tflite::BuiltinOperator_FULLY_CONNECTED); + + flatbuffers::Offset flatbufferModel = + CreateModel(flatBufferBuilder, + TFLITE_SCHEMA_VERSION, + flatBufferBuilder.CreateVector(&operatorCode, 1), + flatBufferBuilder.CreateVector(&subgraph, 1), + modelDescription, + flatBufferBuilder.CreateVector(buffers.data(), buffers.size())); + + flatBufferBuilder.Finish(flatbufferModel); + + return std::vector(flatBufferBuilder.GetBufferPointer(), + flatBufferBuilder.GetBufferPointer() + flatBufferBuilder.GetSize()); +} + +template +void FullyConnectedTest(std::vector& backends, + tflite::TensorType tensorType, + tflite::ActivationFunctionType activationType, + const std::vector & inputTensorShape, + const std::vector & weightsTensorShape, + const std::vector & biasTensorShape, + const std::vector & outputTensorShape, + const std::vector & inputValues, + const std::vector & expectedOutputValues, + const std::vector & weightsData, + float quantScale = 1.0f, + int quantOffset = 0) +{ + using namespace tflite; + + std::vector modelBuffer = CreateFullyConnectedTfLiteModel(tensorType, + activationType, + inputTensorShape, + weightsTensorShape, + biasTensorShape, + outputTensorShape, + weightsData, + quantScale, + quantOffset); + + const Model* tfLiteModel = GetModel(modelBuffer.data()); + // Create TfLite Interpreters + std::unique_ptr armnnDelegateInterpreter; + CHECK(InterpreterBuilder(tfLiteModel, ::tflite::ops::builtin::BuiltinOpResolver()) + (&armnnDelegateInterpreter) == kTfLiteOk); + CHECK(armnnDelegateInterpreter != nullptr); + CHECK(armnnDelegateInterpreter->AllocateTensors() == kTfLiteOk); + + std::unique_ptr tfLiteInterpreter; + CHECK(InterpreterBuilder(tfLiteModel, ::tflite::ops::builtin::BuiltinOpResolver()) + (&tfLiteInterpreter) == kTfLiteOk); + CHECK(tfLiteInterpreter != nullptr); + CHECK(tfLiteInterpreter->AllocateTensors() == kTfLiteOk); + + // Create the ArmNN Delegate + armnnDelegate::DelegateOptions delegateOptions(backends); + std::unique_ptr + theArmnnDelegate(armnnDelegate::TfLiteArmnnDelegateCreate(delegateOptions), + armnnDelegate::TfLiteArmnnDelegateDelete); + CHECK(theArmnnDelegate != nullptr); + // Modify armnnDelegateInterpreter to use armnnDelegate + CHECK(armnnDelegateInterpreter->ModifyGraphWithDelegate(theArmnnDelegate.get()) == kTfLiteOk); + + // Set input data + auto tfLiteDelegateInputId = tfLiteInterpreter->inputs()[0]; + auto tfLiteDelageInputData = tfLiteInterpreter->typed_tensor(tfLiteDelegateInputId); + for (unsigned int i = 0; i < inputValues.size(); ++i) + { + tfLiteDelageInputData[i] = inputValues[i]; + } + + auto armnnDelegateInputId = armnnDelegateInterpreter->inputs()[0]; + auto armnnDelegateInputData = armnnDelegateInterpreter->typed_tensor(armnnDelegateInputId); + for (unsigned int i = 0; i < inputValues.size(); ++i) + { + armnnDelegateInputData[i] = inputValues[i]; + } + + // Run EnqueWorkload + CHECK(tfLiteInterpreter->Invoke() == kTfLiteOk); + CHECK(armnnDelegateInterpreter->Invoke() == kTfLiteOk); + + // Compare output data + auto tfLiteDelegateOutputId = tfLiteInterpreter->outputs()[0]; + auto tfLiteDelageOutputData = tfLiteInterpreter->typed_tensor(tfLiteDelegateOutputId); + auto armnnDelegateOutputId = armnnDelegateInterpreter->outputs()[0]; + auto armnnDelegateOutputData = armnnDelegateInterpreter->typed_tensor(armnnDelegateOutputId); + for (size_t i = 0; i < expectedOutputValues.size(); i++) + { + CHECK(expectedOutputValues[i] == tfLiteDelageOutputData[i]); + CHECK(expectedOutputValues[i] == armnnDelegateOutputData[i]); + CHECK(tfLiteDelageOutputData[i] == armnnDelegateOutputData[i]); + } +} + +} // anonymous namespace \ No newline at end of file -- cgit v1.2.1