2 // Copyright © 2017 Arm Ltd. All rights reserved.
3 // SPDX-License-Identifier: MIT
5 #include "InferenceTest.hpp"
7 #include <armnn/utility/Assert.hpp>
8 #include <boost/numeric/conversion/cast.hpp>
9 #include <boost/program_options.hpp>
10 #include <fmt/format.h>
19 using namespace std::chrono;
20 using namespace armnn::test;
27 using TContainer = boost::variant<std::vector<float>, std::vector<int>, std::vector<unsigned char>>;
29 template <typename TTestCaseDatabase, typename TModel>
30 ClassifierTestCase<TTestCaseDatabase, TModel>::ClassifierTestCase(
31 int& numInferencesRef,
32 int& numCorrectInferencesRef,
33 const std::vector<unsigned int>& validationPredictions,
34 std::vector<unsigned int>* validationPredictionsOut,
36 unsigned int testCaseId,
38 std::vector<typename TModel::DataType> modelInput)
39 : InferenceModelTestCase<TModel>(
40 model, testCaseId, std::vector<TContainer>{ modelInput }, { model.GetOutputSize() })
42 , m_QuantizationParams(model.GetQuantizationParams())
43 , m_NumInferencesRef(numInferencesRef)
44 , m_NumCorrectInferencesRef(numCorrectInferencesRef)
45 , m_ValidationPredictions(validationPredictions)
46 , m_ValidationPredictionsOut(validationPredictionsOut)
50 struct ClassifierResultProcessor : public boost::static_visitor<>
52 using ResultMap = std::map<float,int>;
54 ClassifierResultProcessor(float scale, int offset)
59 void operator()(const std::vector<float>& values)
61 SortPredictions(values, [](float value)
67 void operator()(const std::vector<uint8_t>& values)
69 auto& scale = m_Scale;
70 auto& offset = m_Offset;
71 SortPredictions(values, [&scale, &offset](uint8_t value)
73 return armnn::Dequantize(value, scale, offset);
77 void operator()(const std::vector<int>& values)
80 ARMNN_ASSERT_MSG(false, "Non-float predictions output not supported.");
83 ResultMap& GetResultMap() { return m_ResultMap; }
86 template<typename Container, typename Delegate>
87 void SortPredictions(const Container& c, Delegate delegate)
90 for (const auto& value : c)
92 int classification = index++;
93 // Take the first class with each probability
94 // This avoids strange results when looping over batched results produced
95 // with identical test data.
96 ResultMap::iterator lb = m_ResultMap.lower_bound(value);
98 if (lb == m_ResultMap.end() || !m_ResultMap.key_comp()(value, lb->first))
100 // If the key is not already in the map, insert it.
101 m_ResultMap.insert(lb, ResultMap::value_type(delegate(value), classification));
106 ResultMap m_ResultMap;
112 template <typename TTestCaseDatabase, typename TModel>
113 TestCaseResult ClassifierTestCase<TTestCaseDatabase, TModel>::ProcessResult(const InferenceTestOptions& params)
115 auto& output = this->GetOutputs()[0];
116 const auto testCaseId = this->GetTestCaseId();
118 ClassifierResultProcessor resultProcessor(m_QuantizationParams.first, m_QuantizationParams.second);
119 boost::apply_visitor(resultProcessor, output);
121 ARMNN_LOG(info) << "= Prediction values for test #" << testCaseId;
122 auto it = resultProcessor.GetResultMap().rbegin();
123 for (int i=0; i<5 && it != resultProcessor.GetResultMap().rend(); ++i)
125 ARMNN_LOG(info) << "Top(" << (i+1) << ") prediction is " << it->second <<
126 " with value: " << (it->first);
130 unsigned int prediction = 0;
131 boost::apply_visitor([&](auto&& value)
133 prediction = boost::numeric_cast<unsigned int>(
134 std::distance(value.begin(), std::max_element(value.begin(), value.end())));
138 // If we're just running the defaultTestCaseIds, each one must be classified correctly.
139 if (params.m_IterationCount == 0 && prediction != m_Label)
141 ARMNN_LOG(error) << "Prediction for test case " << testCaseId << " (" << prediction << ")" <<
142 " is incorrect (should be " << m_Label << ")";
143 return TestCaseResult::Failed;
146 // If a validation file was provided as input, it checks that the prediction matches.
147 if (!m_ValidationPredictions.empty() && prediction != m_ValidationPredictions[testCaseId])
149 ARMNN_LOG(error) << "Prediction for test case " << testCaseId << " (" << prediction << ")" <<
150 " doesn't match the prediction in the validation file (" << m_ValidationPredictions[testCaseId] << ")";
151 return TestCaseResult::Failed;
154 // If a validation file was requested as output, it stores the predictions.
155 if (m_ValidationPredictionsOut)
157 m_ValidationPredictionsOut->push_back(prediction);
160 // Updates accuracy stats.
161 m_NumInferencesRef++;
162 if (prediction == m_Label)
164 m_NumCorrectInferencesRef++;
167 return TestCaseResult::Ok;
170 template <typename TDatabase, typename InferenceModel>
171 template <typename TConstructDatabaseCallable, typename TConstructModelCallable>
172 ClassifierTestCaseProvider<TDatabase, InferenceModel>::ClassifierTestCaseProvider(
173 TConstructDatabaseCallable constructDatabase, TConstructModelCallable constructModel)
174 : m_ConstructModel(constructModel)
175 , m_ConstructDatabase(constructDatabase)
177 , m_NumCorrectInferences(0)
181 template <typename TDatabase, typename InferenceModel>
182 void ClassifierTestCaseProvider<TDatabase, InferenceModel>::AddCommandLineOptions(
183 boost::program_options::options_description& options)
185 namespace po = boost::program_options;
187 options.add_options()
188 ("validation-file-in", po::value<std::string>(&m_ValidationFileIn)->default_value(""),
189 "Reads expected predictions from the given file and confirms they match the actual predictions.")
190 ("validation-file-out", po::value<std::string>(&m_ValidationFileOut)->default_value(""),
191 "Predictions are saved to the given file for later use via --validation-file-in.")
192 ("data-dir,d", po::value<std::string>(&m_DataDir)->required(),
193 "Path to directory containing test data");
195 InferenceModel::AddCommandLineOptions(options, m_ModelCommandLineOptions);
198 template <typename TDatabase, typename InferenceModel>
199 bool ClassifierTestCaseProvider<TDatabase, InferenceModel>::ProcessCommandLineOptions(
200 const InferenceTestOptions& commonOptions)
202 if (!ValidateDirectory(m_DataDir))
209 m_Model = m_ConstructModel(commonOptions, m_ModelCommandLineOptions);
215 m_Database = std::make_unique<TDatabase>(m_ConstructDatabase(m_DataDir.c_str(), *m_Model));
224 template <typename TDatabase, typename InferenceModel>
225 std::unique_ptr<IInferenceTestCase>
226 ClassifierTestCaseProvider<TDatabase, InferenceModel>::GetTestCase(unsigned int testCaseId)
228 std::unique_ptr<typename TDatabase::TTestCaseData> testCaseData = m_Database->GetTestCaseData(testCaseId);
229 if (testCaseData == nullptr)
234 return std::make_unique<ClassifierTestCase<TDatabase, InferenceModel>>(
236 m_NumCorrectInferences,
237 m_ValidationPredictions,
238 m_ValidationFileOut.empty() ? nullptr : &m_ValidationPredictionsOut,
241 testCaseData->m_Label,
242 std::move(testCaseData->m_InputImage));
245 template <typename TDatabase, typename InferenceModel>
246 bool ClassifierTestCaseProvider<TDatabase, InferenceModel>::OnInferenceTestFinished()
248 const double accuracy = boost::numeric_cast<double>(m_NumCorrectInferences) /
249 boost::numeric_cast<double>(m_NumInferences);
250 ARMNN_LOG(info) << std::fixed << std::setprecision(3) << "Overall accuracy: " << accuracy;
252 // If a validation file was requested as output, the predictions are saved to it.
253 if (!m_ValidationFileOut.empty())
255 std::ofstream validationFileOut(m_ValidationFileOut.c_str(), std::ios_base::trunc | std::ios_base::out);
256 if (validationFileOut.good())
258 for (const unsigned int prediction : m_ValidationPredictionsOut)
260 validationFileOut << prediction << std::endl;
265 ARMNN_LOG(error) << "Failed to open output validation file: " << m_ValidationFileOut;
273 template <typename TDatabase, typename InferenceModel>
274 void ClassifierTestCaseProvider<TDatabase, InferenceModel>::ReadPredictions()
276 // Reads the expected predictions from the input validation file (if provided).
277 if (!m_ValidationFileIn.empty())
279 std::ifstream validationFileIn(m_ValidationFileIn.c_str(), std::ios_base::in);
280 if (validationFileIn.good())
282 while (!validationFileIn.eof())
285 validationFileIn >> i;
286 m_ValidationPredictions.emplace_back(i);
291 throw armnn::Exception(fmt::format("Failed to open input validation file: {}"
292 , m_ValidationFileIn));
297 template<typename TConstructTestCaseProvider>
298 int InferenceTestMain(int argc,
300 const std::vector<unsigned int>& defaultTestCaseIds,
301 TConstructTestCaseProvider constructTestCaseProvider)
303 // Configures logging for both the ARMNN library and this test program.
305 armnn::LogSeverity level = armnn::LogSeverity::Info;
307 armnn::LogSeverity level = armnn::LogSeverity::Debug;
309 armnn::ConfigureLogging(true, true, level);
313 std::unique_ptr<IInferenceTestCaseProvider> testCaseProvider = constructTestCaseProvider();
314 if (!testCaseProvider)
319 InferenceTestOptions inferenceTestOptions;
320 if (!ParseCommandLine(argc, argv, *testCaseProvider, inferenceTestOptions))
325 const bool success = InferenceTest(inferenceTestOptions, defaultTestCaseIds, *testCaseProvider);
326 return success ? 0 : 1;
328 catch (armnn::Exception const& e)
330 ARMNN_LOG(fatal) << "Armnn Error: " << e.what();
336 // This function allows us to create a classifier inference test based on:
337 // - a model file name
338 // - which can be a binary or a text file for protobuf formats
339 // - an input tensor name
340 // - an output tensor name
341 // - a set of test case ids
342 // - a callback method which creates an object that can return images
343 // called 'Database' in these tests
344 // - and an input tensor shape
346 template<typename TDatabase,
348 typename TConstructDatabaseCallable>
349 int ClassifierInferenceTestMain(int argc,
351 const char* modelFilename,
353 const char* inputBindingName,
354 const char* outputBindingName,
355 const std::vector<unsigned int>& defaultTestCaseIds,
356 TConstructDatabaseCallable constructDatabase,
357 const armnn::TensorShape* inputTensorShape)
360 ARMNN_ASSERT(modelFilename);
361 ARMNN_ASSERT(inputBindingName);
362 ARMNN_ASSERT(outputBindingName);
364 return InferenceTestMain(argc, argv, defaultTestCaseIds,
368 using InferenceModel = InferenceModel<TParser, typename TDatabase::DataType>;
369 using TestCaseProvider = ClassifierTestCaseProvider<TDatabase, InferenceModel>;
371 return make_unique<TestCaseProvider>(constructDatabase,
373 (const InferenceTestOptions &commonOptions,
374 typename InferenceModel::CommandLineOptions modelOptions)
376 if (!ValidateDirectory(modelOptions.m_ModelDir))
378 return std::unique_ptr<InferenceModel>();
381 typename InferenceModel::Params modelParams;
382 modelParams.m_ModelPath = modelOptions.m_ModelDir + modelFilename;
383 modelParams.m_InputBindings = { inputBindingName };
384 modelParams.m_OutputBindings = { outputBindingName };
386 if (inputTensorShape)
388 modelParams.m_InputShapes.push_back(*inputTensorShape);
391 modelParams.m_IsModelBinary = isModelBinary;
392 modelParams.m_ComputeDevices = modelOptions.GetComputeDevicesAsBackendIds();
393 modelParams.m_VisualizePostOptimizationModel = modelOptions.m_VisualizePostOptimizationModel;
394 modelParams.m_EnableFp16TurboMode = modelOptions.m_EnableFp16TurboMode;
396 return std::make_unique<InferenceModel>(modelParams,
397 commonOptions.m_EnableProfiling,
398 commonOptions.m_DynamicBackendsPath);