summaryrefslogtreecommitdiff
path: root/tests/YoloInferenceTest.hpp
blob: edc4808939afb4a3e84745df7824fdc304c1fbfa (plain)
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
//
// Copyright © 2017 Arm Ltd. All rights reserved.
// See LICENSE file in the project root for full license information.
//
#pragma once

#include "InferenceTest.hpp"
#include "YoloDatabase.hpp"

#include <algorithm>
#include <array>
#include <utility>

#include <boost/assert.hpp>
#include <boost/multi_array.hpp>
#include <boost/test/tools/floating_point_comparison.hpp>

constexpr size_t YoloOutputSize = 1470;

template <typename Model>
class YoloTestCase : public InferenceModelTestCase<Model>
{
public:
    YoloTestCase(Model& model,
        unsigned int testCaseId,
        YoloTestCaseData& testCaseData)
     : InferenceModelTestCase<Model>(model, testCaseId, std::move(testCaseData.m_InputImage), YoloOutputSize)
     , m_FloatComparer(boost::math::fpc::percent_tolerance(1.0f))
     , m_TopObjectDetections(std::move(testCaseData.m_TopObjectDetections))
    {
    }

    virtual TestCaseResult ProcessResult(const InferenceTestOptions& options) override
    {
        using Boost3dArray = boost::multi_array<float, 3>;

        const std::vector<float>& output = this->GetOutput();
        BOOST_ASSERT(output.size() == YoloOutputSize);

        constexpr Boost3dArray::index gridSize = 7;
        constexpr Boost3dArray::index numClasses = 20;
        constexpr Boost3dArray::index numScales = 2;

        const float* outputPtr =  output.data();

        // Range 0-980. Class probabilities. 7x7x20
        Boost3dArray classProbabilities(boost::extents[gridSize][gridSize][numClasses]);
        for (Boost3dArray::index y = 0; y < gridSize; ++y)
        {
            for (Boost3dArray::index x = 0; x < gridSize; ++x)
            {
                for (Boost3dArray::index c = 0; c < numClasses; ++c)
                {
                    classProbabilities[y][x][c] = *outputPtr++;
                }
            }
        }

        // Range 980-1078. Scales. 7x7x2
        Boost3dArray scales(boost::extents[gridSize][gridSize][numScales]);
        for (Boost3dArray::index y = 0; y < gridSize; ++y)
        {
            for (Boost3dArray::index x = 0; x < gridSize; ++x)
            {
                for (Boost3dArray::index s = 0; s < numScales; ++s)
                {
                    scales[y][x][s] = *outputPtr++;
                }
            }
        }

        // Range 1078-1469. Bounding boxes. 7x7x2x4
        constexpr float imageWidthAsFloat = static_cast<float>(YoloImageWidth);
        constexpr float imageHeightAsFloat = static_cast<float>(YoloImageHeight);

        boost::multi_array<float, 4> boxes(boost::extents[gridSize][gridSize][numScales][4]);
        for (Boost3dArray::index y = 0; y < gridSize; ++y)
        {
            for (Boost3dArray::index x = 0; x < gridSize; ++x)
            {
                for (Boost3dArray::index s = 0; s < numScales; ++s)
                {
                    float bx = *outputPtr++;
                    float by = *outputPtr++;
                    float bw = *outputPtr++;
                    float bh = *outputPtr++;

                    boxes[y][x][s][0] = ((bx + static_cast<float>(x)) / 7.0f) * imageWidthAsFloat;
                    boxes[y][x][s][1] = ((by + static_cast<float>(y)) / 7.0f) * imageHeightAsFloat;
                    boxes[y][x][s][2] = bw * bw * static_cast<float>(imageWidthAsFloat);
                    boxes[y][x][s][3] = bh * bh * static_cast<float>(imageHeightAsFloat);
                }
            }
        }
        BOOST_ASSERT(output.data() + YoloOutputSize == outputPtr);

        std::vector<YoloDetectedObject> detectedObjects;
        detectedObjects.reserve(gridSize * gridSize * numScales * numClasses);

        for (Boost3dArray::index y = 0; y < gridSize; ++y)
        {
            for (Boost3dArray::index x = 0; x < gridSize; ++x)
            {
                for (Boost3dArray::index s = 0; s < numScales; ++s)
                {
                    for (Boost3dArray::index c = 0; c < numClasses; ++c)
                    {
                        // Resolved confidence: Class probabilities * scales
                        const float confidence = classProbabilities[y][x][c] * scales[y][x][s];

                        // Resolve bounding box and store
                        YoloBoundingBox box;
                        box.m_X = boxes[y][x][s][0];
                        box.m_Y = boxes[y][x][s][1];
                        box.m_W = boxes[y][x][s][2];
                        box.m_H = boxes[y][x][s][3];

                        detectedObjects.emplace_back(c, box, confidence);
                    }
                }
            }
        }

        // Sort detected objects by confidence
        std::sort(detectedObjects.begin(), detectedObjects.end(),
            [](const YoloDetectedObject& a, const YoloDetectedObject& b)
            {
                // Sort by largest confidence first, then by class
                return a.m_Confidence > b.m_Confidence
                    || (a.m_Confidence == b.m_Confidence && a.m_Class > b.m_Class);
            });

        // Check the top N detections
        auto outputIt  = detectedObjects.begin();
        auto outputEnd = detectedObjects.end();

        for (const YoloDetectedObject& expectedDetection : m_TopObjectDetections)
        {
            if (outputIt == outputEnd)
            {
                // Somehow expected more things to check than detections found by the model
                return TestCaseResult::Abort;
            }

            const YoloDetectedObject& detectedObject = *outputIt;
            if (detectedObject.m_Class != expectedDetection.m_Class)
            {
                BOOST_LOG_TRIVIAL(error) << "Prediction for test case " << this->GetTestCaseId() <<
                    " (" << detectedObject.m_Class << ")" <<
                    " is incorrect (should be " << expectedDetection.m_Class << ")";
                return TestCaseResult::Failed;
            }

            if (!m_FloatComparer(detectedObject.m_Box.m_X, expectedDetection.m_Box.m_X) ||
                !m_FloatComparer(detectedObject.m_Box.m_Y, expectedDetection.m_Box.m_Y) ||
                !m_FloatComparer(detectedObject.m_Box.m_W, expectedDetection.m_Box.m_W) ||
                !m_FloatComparer(detectedObject.m_Box.m_H, expectedDetection.m_Box.m_H) ||
                !m_FloatComparer(detectedObject.m_Confidence, expectedDetection.m_Confidence))
            {
                BOOST_LOG_TRIVIAL(error) << "Detected bounding box for test case " << this->GetTestCaseId() <<
                    " is incorrect";
                return TestCaseResult::Failed;
            }

            ++outputIt;
        }

        return TestCaseResult::Ok;
    }

private:
    boost::math::fpc::close_at_tolerance<float> m_FloatComparer;
    std::vector<YoloDetectedObject> m_TopObjectDetections;
};

template <typename Model>
class YoloTestCaseProvider : public IInferenceTestCaseProvider
{
public:
    template <typename TConstructModelCallable>
    YoloTestCaseProvider(TConstructModelCallable constructModel)
        : m_ConstructModel(constructModel)
    {
    }

    virtual void AddCommandLineOptions(boost::program_options::options_description& options) override
    {
        namespace po = boost::program_options;

        options.add_options()
            ("data-dir,d", po::value<std::string>(&m_DataDir)->required(),
                "Path to directory containing test data");

        Model::AddCommandLineOptions(options, m_ModelCommandLineOptions);
    }

    virtual bool ProcessCommandLineOptions() override
    {
        if (!ValidateDirectory(m_DataDir))
        {
            return false;
        }

        m_Model = m_ConstructModel(m_ModelCommandLineOptions);
        if (!m_Model)
        {
            return false;
        }

        m_Database = std::make_unique<YoloDatabase>(m_DataDir.c_str());
        if (!m_Database)
        {
            return false;
        }

        return true;
    }

    virtual std::unique_ptr<IInferenceTestCase> GetTestCase(unsigned int testCaseId) override
    {
        std::unique_ptr<YoloTestCaseData> testCaseData = m_Database->GetTestCaseData(testCaseId);
        if (!testCaseData)
        {
            return nullptr;
        }

        return std::make_unique<YoloTestCase<Model>>(*m_Model, testCaseId, *testCaseData);
    }

private:
    typename Model::CommandLineOptions m_ModelCommandLineOptions;
    std::function<std::unique_ptr<Model>(typename Model::CommandLineOptions)> m_ConstructModel;
    std::unique_ptr<Model> m_Model;

    std::string m_DataDir;
    std::unique_ptr<YoloDatabase> m_Database;
};