tests/YoloInferenceTest.hpp

   1 //
   2 // Copyright © 2017 Arm Ltd. All rights reserved.
   3 // SPDX-License-Identifier: MIT
   4 //
   5 #pragma once
   6
   7 #include "InferenceTest.hpp"
   8 #include "YoloDatabase.hpp"
   9
  10 #include <algorithm>
  11 #include <array>
  12 #include <utility>
  13
  14 #include <boost/assert.hpp>
  15 #include <boost/multi_array.hpp>
  16 #include <boost/test/tools/floating_point_comparison.hpp>
  17
  18 constexpr size_t YoloOutputSize = 1470;
  19
  20 template <typename Model>
  21 class YoloTestCase : public InferenceModelTestCase<Model>
  22 {
  23 public:
  24     YoloTestCase(Model& model,
  25         unsigned int testCaseId,
  26         YoloTestCaseData& testCaseData)
  27      : InferenceModelTestCase<Model>(model, testCaseId, { std::move(testCaseData.m_InputImage) }, { YoloOutputSize })
  28      , m_FloatComparer(boost::math::fpc::percent_tolerance(1.0f))
  29      , m_TopObjectDetections(std::move(testCaseData.m_TopObjectDetections))
  30     {
  31     }
  32
  33     virtual TestCaseResult ProcessResult(const InferenceTestOptions& options) override
  34     {
  35         using Boost3dArray = boost::multi_array<float, 3>;
  36
  37         const std::vector<float>& output = boost::get<std::vector<float>>(this->GetOutputs()[0]);
  38         BOOST_ASSERT(output.size() == YoloOutputSize);
  39
  40         constexpr Boost3dArray::index gridSize = 7;
  41         constexpr Boost3dArray::index numClasses = 20;
  42         constexpr Boost3dArray::index numScales = 2;
  43
  44         const float* outputPtr =  output.data();
  45
  46         // Range 0-980. Class probabilities. 7x7x20
  47         Boost3dArray classProbabilities(boost::extents[gridSize][gridSize][numClasses]);
  48         for (Boost3dArray::index y = 0; y < gridSize; ++y)
  49         {
  50             for (Boost3dArray::index x = 0; x < gridSize; ++x)
  51             {
  52                 for (Boost3dArray::index c = 0; c < numClasses; ++c)
  53                 {
  54                     classProbabilities[y][x][c] = *outputPtr++;
  55                 }
  56             }
  57         }
  58
  59         // Range 980-1078. Scales. 7x7x2
  60         Boost3dArray scales(boost::extents[gridSize][gridSize][numScales]);
  61         for (Boost3dArray::index y = 0; y < gridSize; ++y)
  62         {
  63             for (Boost3dArray::index x = 0; x < gridSize; ++x)
  64             {
  65                 for (Boost3dArray::index s = 0; s < numScales; ++s)
  66                 {
  67                     scales[y][x][s] = *outputPtr++;
  68                 }
  69             }
  70         }
  71
  72         // Range 1078-1469. Bounding boxes. 7x7x2x4
  73         constexpr float imageWidthAsFloat = static_cast<float>(YoloImageWidth);
  74         constexpr float imageHeightAsFloat = static_cast<float>(YoloImageHeight);
  75
  76         boost::multi_array<float, 4> boxes(boost::extents[gridSize][gridSize][numScales][4]);
  77         for (Boost3dArray::index y = 0; y < gridSize; ++y)
  78         {
  79             for (Boost3dArray::index x = 0; x < gridSize; ++x)
  80             {
  81                 for (Boost3dArray::index s = 0; s < numScales; ++s)
  82                 {
  83                     float bx = *outputPtr++;
  84                     float by = *outputPtr++;
  85                     float bw = *outputPtr++;
  86                     float bh = *outputPtr++;
  87
  88                     boxes[y][x][s][0] = ((bx + static_cast<float>(x)) / 7.0f) * imageWidthAsFloat;
  89                     boxes[y][x][s][1] = ((by + static_cast<float>(y)) / 7.0f) * imageHeightAsFloat;
  90                     boxes[y][x][s][2] = bw * bw * static_cast<float>(imageWidthAsFloat);
  91                     boxes[y][x][s][3] = bh * bh * static_cast<float>(imageHeightAsFloat);
  92                 }
  93             }
  94         }
  95         BOOST_ASSERT(output.data() + YoloOutputSize == outputPtr);
  96
  97         std::vector<YoloDetectedObject> detectedObjects;
  98         detectedObjects.reserve(gridSize * gridSize * numScales * numClasses);
  99
 100         for (Boost3dArray::index y = 0; y < gridSize; ++y)
 101         {
 102             for (Boost3dArray::index x = 0; x < gridSize; ++x)
 103             {
 104                 for (Boost3dArray::index s = 0; s < numScales; ++s)
 105                 {
 106                     for (Boost3dArray::index c = 0; c < numClasses; ++c)
 107                     {
 108                         // Resolved confidence: class probabilities * scales.
 109                         const float confidence = classProbabilities[y][x][c] * scales[y][x][s];
 110
 111                         // Resolves bounding box and stores.
 112                         YoloBoundingBox box;
 113                         box.m_X = boxes[y][x][s][0];
 114                         box.m_Y = boxes[y][x][s][1];
 115                         box.m_W = boxes[y][x][s][2];
 116                         box.m_H = boxes[y][x][s][3];
 117
 118                         detectedObjects.emplace_back(c, box, confidence);
 119                     }
 120                 }
 121             }
 122         }
 123
 124         // Sorts detected objects by confidence.
 125         std::sort(detectedObjects.begin(), detectedObjects.end(),
 126             [](const YoloDetectedObject& a, const YoloDetectedObject& b)
 127             {
 128                 // Sorts by largest confidence first, then by class.
 129                 return a.m_Confidence > b.m_Confidence
 130                     || (a.m_Confidence == b.m_Confidence && a.m_Class > b.m_Class);
 131             });
 132
 133         // Checks the top N detections.
 134         auto outputIt  = detectedObjects.begin();
 135         auto outputEnd = detectedObjects.end();
 136
 137         for (const YoloDetectedObject& expectedDetection : m_TopObjectDetections)
 138         {
 139             if (outputIt == outputEnd)
 140             {
 141                 // Somehow expected more things to check than detections found by the model.
 142                 return TestCaseResult::Abort;
 143             }
 144
 145             const YoloDetectedObject& detectedObject = *outputIt;
 146             if (detectedObject.m_Class != expectedDetection.m_Class)
 147             {
 148                 BOOST_LOG_TRIVIAL(error) << "Prediction for test case " << this->GetTestCaseId() <<
 149                     " is incorrect: Expected (" << expectedDetection.m_Class << ")" <<
 150                     " but predicted (" << detectedObject.m_Class << ")";
 151                 return TestCaseResult::Failed;
 152             }
 153
 154             if (!m_FloatComparer(detectedObject.m_Box.m_X, expectedDetection.m_Box.m_X) ||
 155                 !m_FloatComparer(detectedObject.m_Box.m_Y, expectedDetection.m_Box.m_Y) ||
 156                 !m_FloatComparer(detectedObject.m_Box.m_W, expectedDetection.m_Box.m_W) ||
 157                 !m_FloatComparer(detectedObject.m_Box.m_H, expectedDetection.m_Box.m_H) ||
 158                 !m_FloatComparer(detectedObject.m_Confidence, expectedDetection.m_Confidence))
 159             {
 160                 BOOST_LOG_TRIVIAL(error) << "Detected bounding box for test case " << this->GetTestCaseId() <<
 161                     " is incorrect";
 162                 return TestCaseResult::Failed;
 163             }
 164
 165             ++outputIt;
 166         }
 167
 168         return TestCaseResult::Ok;
 169     }
 170
 171 private:
 172     boost::math::fpc::close_at_tolerance<float> m_FloatComparer;
 173     std::vector<YoloDetectedObject> m_TopObjectDetections;
 174 };
 175
 176 template <typename Model>
 177 class YoloTestCaseProvider : public IInferenceTestCaseProvider
 178 {
 179 public:
 180     template <typename TConstructModelCallable>
 181     explicit YoloTestCaseProvider(TConstructModelCallable constructModel)
 182         : m_ConstructModel(constructModel)
 183     {
 184     }
 185
 186     virtual void AddCommandLineOptions(boost::program_options::options_description& options) override
 187     {
 188         namespace po = boost::program_options;
 189
 190         options.add_options()
 191             ("data-dir,d", po::value<std::string>(&m_DataDir)->required(),
 192                 "Path to directory containing test data");
 193
 194         Model::AddCommandLineOptions(options, m_ModelCommandLineOptions);
 195     }
 196
 197     virtual bool ProcessCommandLineOptions(const InferenceTestOptions &commonOptions) override
 198     {
 199         if (!ValidateDirectory(m_DataDir))
 200         {
 201             return false;
 202         }
 203
 204         m_Model = m_ConstructModel(commonOptions, m_ModelCommandLineOptions);
 205         if (!m_Model)
 206         {
 207             return false;
 208         }
 209
 210         m_Database = std::make_unique<YoloDatabase>(m_DataDir.c_str());
 211         if (!m_Database)
 212         {
 213             return false;
 214         }
 215
 216         return true;
 217     }
 218
 219     virtual std::unique_ptr<IInferenceTestCase> GetTestCase(unsigned int testCaseId) override
 220     {
 221         std::unique_ptr<YoloTestCaseData> testCaseData = m_Database->GetTestCaseData(testCaseId);
 222         if (!testCaseData)
 223         {
 224             return nullptr;
 225         }
 226
 227         return std::make_unique<YoloTestCase<Model>>(*m_Model, testCaseId, *testCaseData);
 228     }
 229
 230 private:
 231     typename Model::CommandLineOptions m_ModelCommandLineOptions;
 232     std::function<std::unique_ptr<Model>(const InferenceTestOptions&,
 233                                          typename Model::CommandLineOptions)> m_ConstructModel;
 234     std::unique_ptr<Model> m_Model;
 235
 236     std::string m_DataDir;
 237     std::unique_ptr<YoloDatabase> m_Database;
 238 };