From 0d35a93d68e321e8c4b16baa8b9754b98cc9faf3 Mon Sep 17 00:00:00 2001 From: Matthew Sloyan Date: Mon, 9 Nov 2020 12:25:05 +0000 Subject: IVGCVSW-5398 TfLiteDelegate: Implement the Quantization operators * Enabled quantization operators DEQUANTIZE and QUANTIZE. * Implemented unit tests for quantization operators. * Added utils function for checking if affine quantization. Signed-off-by: Matthew Sloyan Change-Id: I84b5c75bda629d9234f5ed198b04f527705a54aa --- delegate/src/test/QuantizationTestHelper.hpp | 197 +++++++++++++++++++++++++++ 1 file changed, 197 insertions(+) create mode 100644 delegate/src/test/QuantizationTestHelper.hpp (limited to 'delegate/src/test/QuantizationTestHelper.hpp') diff --git a/delegate/src/test/QuantizationTestHelper.hpp b/delegate/src/test/QuantizationTestHelper.hpp new file mode 100644 index 0000000000..2843e43233 --- /dev/null +++ b/delegate/src/test/QuantizationTestHelper.hpp @@ -0,0 +1,197 @@ +// +// Copyright © 2020 Arm Ltd and Contributors. All rights reserved. +// SPDX-License-Identifier: MIT +// + +#pragma once + +#include + +#include +#include +#include +#include +#include +#include + +#include + +namespace +{ + +std::vector CreateQuantizationTfLiteModel(tflite::BuiltinOperator quantizationOperatorCode, + tflite::TensorType inputTensorType, + tflite::TensorType outputTensorType, + const std::vector & inputTensorShape, + const std::vector & outputTensorShape, + float quantScale = 1.0f, + int quantOffset = 0) +{ + using namespace tflite; + flatbuffers::FlatBufferBuilder flatBufferBuilder; + + std::vector> buffers; + buffers.push_back(CreateBuffer(flatBufferBuilder, flatBufferBuilder.CreateVector({}))); + + auto quantizationParameters = + CreateQuantizationParameters(flatBufferBuilder, + 0, + 0, + flatBufferBuilder.CreateVector({ quantScale }), + flatBufferBuilder.CreateVector({ quantOffset }), + QuantizationDetails_CustomQuantization); + + std::array, 2> tensors; + tensors[0] = CreateTensor(flatBufferBuilder, + flatBufferBuilder.CreateVector(inputTensorShape.data(), + inputTensorShape.size()), + inputTensorType, + 0, + flatBufferBuilder.CreateString("input"), + quantizationParameters); + tensors[1] = CreateTensor(flatBufferBuilder, + flatBufferBuilder.CreateVector(outputTensorShape.data(), + outputTensorShape.size()), + outputTensorType, + 0, + flatBufferBuilder.CreateString("output"), + quantizationParameters); + + // create operator + tflite::BuiltinOptions operatorBuiltinOptionsType = tflite::BuiltinOptions_NONE; + flatbuffers::Offset operatorBuiltinOptions = 0; + switch (quantizationOperatorCode) + { + case BuiltinOperator_QUANTIZE: + { + operatorBuiltinOptionsType = BuiltinOptions_QuantizeOptions; + operatorBuiltinOptions = CreateQuantizeOptions(flatBufferBuilder).Union(); + break; + } + case BuiltinOperator_DEQUANTIZE: + { + operatorBuiltinOptionsType = BuiltinOptions_DequantizeOptions; + operatorBuiltinOptions = CreateDequantizeOptions(flatBufferBuilder).Union(); + break; + } + default: + break; + } + + const std::vector operatorInputs{ {0} }; + const std::vector operatorOutputs{{1}}; + flatbuffers::Offset quantizationOperator = + CreateOperator(flatBufferBuilder, + 0, + flatBufferBuilder.CreateVector(operatorInputs.data(), operatorInputs.size()), + flatBufferBuilder.CreateVector(operatorOutputs.data(), operatorOutputs.size()), + operatorBuiltinOptionsType, + operatorBuiltinOptions); + + const std::vector subgraphInputs{ {0} }; + const std::vector subgraphOutputs{{1}}; + 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(&quantizationOperator, 1)); + + flatbuffers::Offset modelDescription = + flatBufferBuilder.CreateString("ArmnnDelegate: Quantization Operator Model"); + flatbuffers::Offset operatorCode = CreateOperatorCode(flatBufferBuilder, quantizationOperatorCode); + + 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 QuantizationTest(tflite::BuiltinOperator quantizeOperatorCode, + tflite::TensorType inputTensorType, + tflite::TensorType outputTensorType, + std::vector& backends, + std::vector& inputShape, + std::vector& outputShape, + std::vector& inputValues, + std::vector& expectedOutputValues, + float quantScale = 1.0f, + int quantOffset = 0) +{ + using namespace tflite; + std::vector modelBuffer = CreateQuantizationTfLiteModel(quantizeOperatorCode, + inputTensorType, + outputTensorType, + inputShape, + outputShape, + 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] == armnnDelegateOutputData[i]); + CHECK(tfLiteDelageOutputData[i] == expectedOutputValues[i]); + CHECK(tfLiteDelageOutputData[i] == armnnDelegateOutputData[i]); + } +} + +} // anonymous namespace \ No newline at end of file -- cgit v1.2.1