Derek Lamberti | d6cb30e | 2020-04-28 13:31:29 +0100 | [diff] [blame] | 1 | // |
| 2 | // Copyright © 2020 Arm Ltd. All rights reserved. |
| 3 | // SPDX-License-Identifier: MIT |
| 4 | // |
| 5 | //#include "../InferenceTest.hpp" |
| 6 | //#include "../ImagePreprocessor.hpp" |
| 7 | #include "armnnTfLiteParser/ITfLiteParser.hpp" |
| 8 | |
| 9 | #include "NMS.hpp" |
| 10 | |
| 11 | #include <stb/stb_image.h> |
| 12 | |
| 13 | #include <armnn/INetwork.hpp> |
| 14 | #include <armnn/IRuntime.hpp> |
| 15 | #include <armnn/Logging.hpp> |
| 16 | #include <armnn/utility/IgnoreUnused.hpp> |
| 17 | |
| 18 | #include <chrono> |
| 19 | #include <iostream> |
| 20 | #include <fstream> |
| 21 | |
| 22 | using namespace armnnTfLiteParser; |
| 23 | using namespace armnn; |
| 24 | |
| 25 | static const int OPEN_FILE_ERROR = -2; |
| 26 | static const int OPTIMIZE_NETWORK_ERROR = -3; |
| 27 | static const int LOAD_NETWORK_ERROR = -4; |
| 28 | static const int LOAD_IMAGE_ERROR = -5; |
| 29 | static const int GENERAL_ERROR = -100; |
| 30 | |
| 31 | #define CHECK_OK(v) \ |
| 32 | do { \ |
| 33 | try { \ |
| 34 | auto r_local = v; \ |
| 35 | if (r_local != 0) { return r_local;} \ |
| 36 | } \ |
| 37 | catch(armnn::Exception e) \ |
| 38 | { \ |
| 39 | ARMNN_LOG(error) << "Oops: " << e.what(); \ |
| 40 | return GENERAL_ERROR; \ |
| 41 | } \ |
| 42 | } while(0) |
| 43 | |
| 44 | |
| 45 | |
| 46 | template<typename TContainer> |
| 47 | inline armnn::InputTensors MakeInputTensors(const std::vector<armnn::BindingPointInfo>& inputBindings, |
| 48 | const std::vector<TContainer>& inputDataContainers) |
| 49 | { |
| 50 | armnn::InputTensors inputTensors; |
| 51 | |
| 52 | const size_t numInputs = inputBindings.size(); |
| 53 | if (numInputs != inputDataContainers.size()) |
| 54 | { |
| 55 | throw armnn::Exception("Mismatching vectors"); |
| 56 | } |
| 57 | |
| 58 | for (size_t i = 0; i < numInputs; i++) |
| 59 | { |
| 60 | const armnn::BindingPointInfo& inputBinding = inputBindings[i]; |
| 61 | const TContainer& inputData = inputDataContainers[i]; |
| 62 | |
| 63 | armnn::ConstTensor inputTensor(inputBinding.second, inputData.data()); |
| 64 | inputTensors.push_back(std::make_pair(inputBinding.first, inputTensor)); |
| 65 | } |
| 66 | |
| 67 | return inputTensors; |
| 68 | } |
| 69 | |
| 70 | template<typename TContainer> |
| 71 | inline armnn::OutputTensors MakeOutputTensors(const std::vector<armnn::BindingPointInfo>& outputBindings, |
| 72 | const std::vector<TContainer>& outputDataContainers) |
| 73 | { |
| 74 | armnn::OutputTensors outputTensors; |
| 75 | |
| 76 | const size_t numOutputs = outputBindings.size(); |
| 77 | if (numOutputs != outputDataContainers.size()) |
| 78 | { |
| 79 | throw armnn::Exception("Mismatching vectors"); |
| 80 | } |
| 81 | |
| 82 | for (size_t i = 0; i < numOutputs; i++) |
| 83 | { |
| 84 | const armnn::BindingPointInfo& outputBinding = outputBindings[i]; |
| 85 | const TContainer& outputData = outputDataContainers[i]; |
| 86 | |
| 87 | armnn::Tensor outputTensor(outputBinding.second, const_cast<float*>(outputData.data())); |
| 88 | outputTensors.push_back(std::make_pair(outputBinding.first, outputTensor)); |
| 89 | } |
| 90 | |
| 91 | return outputTensors; |
| 92 | } |
| 93 | |
| 94 | int LoadModel(const char* filename, |
| 95 | ITfLiteParser& parser, |
| 96 | IRuntime& runtime, |
| 97 | NetworkId& networkId, |
| 98 | const std::vector<BackendId>& backendPreferences) |
| 99 | { |
| 100 | std::ifstream stream(filename, std::ios::in | std::ios::binary); |
| 101 | if (!stream.is_open()) |
| 102 | { |
| 103 | ARMNN_LOG(error) << "Could not open model: " << filename; |
| 104 | return OPEN_FILE_ERROR; |
| 105 | } |
| 106 | |
| 107 | std::vector<uint8_t> contents((std::istreambuf_iterator<char>(stream)), std::istreambuf_iterator<char>()); |
| 108 | stream.close(); |
| 109 | |
| 110 | auto model = parser.CreateNetworkFromBinary(contents); |
| 111 | contents.clear(); |
| 112 | ARMNN_LOG(debug) << "Model loaded ok: " << filename; |
| 113 | |
| 114 | // Optimize backbone model |
| 115 | auto optimizedModel = Optimize(*model, backendPreferences, runtime.GetDeviceSpec()); |
| 116 | if (!optimizedModel) |
| 117 | { |
| 118 | ARMNN_LOG(fatal) << "Could not optimize the model:" << filename; |
| 119 | return OPTIMIZE_NETWORK_ERROR; |
| 120 | } |
| 121 | |
| 122 | // Load backbone model into runtime |
| 123 | { |
| 124 | std::string errorMessage; |
| 125 | INetworkProperties modelProps; |
| 126 | Status status = runtime.LoadNetwork(networkId, std::move(optimizedModel), errorMessage, modelProps); |
| 127 | if (status != Status::Success) |
| 128 | { |
| 129 | ARMNN_LOG(fatal) << "Could not load " << filename << " model into runtime: " << errorMessage; |
| 130 | return LOAD_NETWORK_ERROR; |
| 131 | } |
| 132 | } |
| 133 | |
| 134 | return 0; |
| 135 | } |
| 136 | |
| 137 | std::vector<float> LoadImage(const char* filename) |
| 138 | { |
| 139 | struct Memory |
| 140 | { |
| 141 | ~Memory() {stbi_image_free(m_Data);} |
| 142 | bool IsLoaded() const { return m_Data != nullptr;} |
| 143 | |
| 144 | unsigned char* m_Data; |
| 145 | }; |
| 146 | |
| 147 | std::vector<float> image; |
| 148 | |
| 149 | int width; |
| 150 | int height; |
| 151 | int channels; |
| 152 | |
| 153 | Memory mem = {stbi_load(filename, &width, &height, &channels, 3)}; |
| 154 | if (!mem.IsLoaded()) |
| 155 | { |
| 156 | ARMNN_LOG(error) << "Could not load input image file: " << filename; |
| 157 | return image; |
| 158 | } |
| 159 | |
| 160 | if (width != 1920 || height != 1080 || channels != 3) |
| 161 | { |
| 162 | ARMNN_LOG(error) << "Input image has wong dimension: " << width << "x" << height << "x" << channels << ". " |
| 163 | " Expected 1920x1080x3."; |
| 164 | return image; |
| 165 | } |
| 166 | |
| 167 | image.resize(1920*1080*3); |
| 168 | |
| 169 | // Expand to float. Does this need de-gamma? |
| 170 | for (unsigned int idx=0; idx <= 1920*1080*3; idx++) |
| 171 | { |
| 172 | image[idx] = static_cast<float>(mem.m_Data[idx]) /255.0f; |
| 173 | } |
| 174 | |
| 175 | return image; |
| 176 | } |
| 177 | |
| 178 | int main(int argc, char* argv[]) |
| 179 | { |
| 180 | if (argc != 3) |
| 181 | { |
| 182 | ARMNN_LOG(error) << "Expected arguments: {PathToModels} {PathToData}"; |
| 183 | } |
| 184 | std::string modelsPath(argv[1]); |
| 185 | std::string imagePath(argv[2]); |
| 186 | |
| 187 | std::string backboneModelFile = modelsPath + "yolov3_1080_1920_backbone_int8.tflite"; |
| 188 | std::string detectorModelFile = modelsPath + "yolov3_1080_1920_detector_fp32.tflite"; |
| 189 | std::string imageFile = imagePath + "1080_1920.jpg"; |
| 190 | |
| 191 | // Configure the logging |
| 192 | SetAllLoggingSinks(true, true, true); |
| 193 | SetLogFilter(LogSeverity::Trace); |
| 194 | |
| 195 | |
| 196 | // Create runtime |
| 197 | IRuntime::CreationOptions runtimeOptions; // default |
| 198 | auto runtime = IRuntime::Create(runtimeOptions); |
| 199 | if (!runtime) |
| 200 | { |
| 201 | ARMNN_LOG(fatal) << "Could not create runtime."; |
| 202 | return -1; |
| 203 | } |
| 204 | |
| 205 | // Create TfLite Parsers |
| 206 | ITfLiteParser::TfLiteParserOptions parserOptions; |
| 207 | auto parser = ITfLiteParser::Create(parserOptions); |
| 208 | |
| 209 | // Load backbone model |
| 210 | ARMNN_LOG(info) << "Loading backbone..."; |
| 211 | NetworkId backboneId; |
| 212 | CHECK_OK(LoadModel(backboneModelFile.c_str(), *parser, *runtime, backboneId, {"GpuAcc", "CpuRef"})); |
| 213 | auto inputId = parser->GetNetworkInputBindingInfo(0, "inputs"); |
| 214 | auto bbOut0Id = parser->GetNetworkOutputBindingInfo(0, "input_to_detector_1"); |
| 215 | auto bbOut1Id = parser->GetNetworkOutputBindingInfo(0, "input_to_detector_2"); |
| 216 | auto bbOut2Id = parser->GetNetworkOutputBindingInfo(0, "input_to_detector_3"); |
| 217 | auto backboneProfile = runtime->GetProfiler(backboneId); |
| 218 | backboneProfile->EnableProfiling(true); |
| 219 | |
| 220 | // Load detector model |
| 221 | ARMNN_LOG(info) << "Loading detector..."; |
| 222 | NetworkId detectorId; |
| 223 | CHECK_OK(LoadModel(detectorModelFile.c_str(), *parser, *runtime, detectorId, {"CpuAcc", "CpuRef"})); |
| 224 | auto detectIn0Id = parser->GetNetworkInputBindingInfo(0, "input_to_detector_1"); |
| 225 | auto detectIn1Id = parser->GetNetworkInputBindingInfo(0, "input_to_detector_2"); |
| 226 | auto detectIn2Id = parser->GetNetworkInputBindingInfo(0, "input_to_detector_3"); |
| 227 | auto outputBoxesId = parser->GetNetworkOutputBindingInfo(0, "output_boxes"); |
| 228 | auto detectorProfile = runtime->GetProfiler(detectorId); |
| 229 | |
| 230 | // Load input from file |
| 231 | ARMNN_LOG(info) << "Loading test image..."; |
| 232 | auto image = LoadImage(imageFile.c_str()); |
| 233 | if (image.empty()) |
| 234 | { |
| 235 | return LOAD_IMAGE_ERROR; |
| 236 | } |
| 237 | |
| 238 | |
| 239 | // Allocate the intermediate tensors |
| 240 | std::vector<float> intermediateMem0(bbOut0Id.second.GetNumElements()); |
| 241 | std::vector<float> intermediateMem1(bbOut1Id.second.GetNumElements()); |
| 242 | std::vector<float> intermediateMem2(bbOut2Id.second.GetNumElements()); |
| 243 | std::vector<float> intermediateMem3(outputBoxesId.second.GetNumElements()); |
| 244 | |
| 245 | // Setup inputs and outputs |
| 246 | using BindingInfos = std::vector<armnn::BindingPointInfo>; |
| 247 | using FloatTensors = std::vector<std::vector<float>>; |
| 248 | |
| 249 | InputTensors bbInputTensors = MakeInputTensors(BindingInfos{inputId}, |
| 250 | FloatTensors{std::move(image)}); |
| 251 | OutputTensors bbOutputTensors = MakeOutputTensors(BindingInfos{bbOut0Id, bbOut1Id, bbOut2Id}, |
| 252 | FloatTensors{intermediateMem0, |
| 253 | intermediateMem1, |
| 254 | intermediateMem2}); |
| 255 | InputTensors detectInputTensors = MakeInputTensors(BindingInfos{detectIn0Id, |
| 256 | detectIn1Id, |
| 257 | detectIn2Id}, |
| 258 | FloatTensors{intermediateMem0, |
| 259 | intermediateMem1, |
| 260 | intermediateMem2}); |
| 261 | OutputTensors detectOutputTensors = MakeOutputTensors(BindingInfos{outputBoxesId}, |
| 262 | FloatTensors{intermediateMem3}); |
| 263 | |
| 264 | static const int numIterations=2; |
| 265 | using DurationUS = std::chrono::duration<double, std::micro>; |
| 266 | std::vector<DurationUS> nmsDurations(0); |
| 267 | nmsDurations.reserve(numIterations); |
| 268 | for (int i=0; i < numIterations; i++) |
| 269 | { |
| 270 | // Execute backbone |
| 271 | ARMNN_LOG(info) << "Running backbone..."; |
| 272 | runtime->EnqueueWorkload(backboneId, bbInputTensors, bbOutputTensors); |
| 273 | |
| 274 | // Execute detector |
| 275 | ARMNN_LOG(info) << "Running detector..."; |
| 276 | runtime->EnqueueWorkload(detectorId, detectInputTensors, detectOutputTensors); |
| 277 | |
| 278 | // Execute NMS |
| 279 | ARMNN_LOG(info) << "Running nms..."; |
| 280 | using clock = std::chrono::steady_clock; |
| 281 | auto nmsStartTime = clock::now(); |
| 282 | yolov3::NMSConfig config; |
| 283 | config.num_boxes = 127800; |
| 284 | config.num_classes = 80; |
| 285 | config.confidence_threshold = 0.9f; |
| 286 | config.iou_threshold = 0.5f; |
| 287 | auto filtered_boxes = yolov3::nms(config, intermediateMem3); |
| 288 | auto nmsEndTime = clock::now(); |
| 289 | |
| 290 | // Enable the profiling after the warm-up run |
| 291 | if (i>0) |
| 292 | { |
| 293 | print_detection(std::cout, filtered_boxes); |
| 294 | |
| 295 | const auto nmsDuration = DurationUS(nmsStartTime - nmsEndTime); |
| 296 | nmsDurations.push_back(nmsDuration); |
| 297 | } |
| 298 | backboneProfile->EnableProfiling(true); |
| 299 | detectorProfile->EnableProfiling(true); |
| 300 | } |
| 301 | // Log timings to file |
| 302 | std::ofstream backboneProfileStream("backbone.json"); |
| 303 | backboneProfile->Print(backboneProfileStream); |
| 304 | backboneProfileStream.close(); |
| 305 | |
| 306 | std::ofstream detectorProfileStream("detector.json"); |
| 307 | detectorProfile->Print(detectorProfileStream); |
| 308 | detectorProfileStream.close(); |
| 309 | |
| 310 | // Manually construct the json output |
| 311 | std::ofstream nmsProfileStream("nms.json"); |
| 312 | nmsProfileStream << "{" << "\n"; |
| 313 | nmsProfileStream << R"( "NmsTimings": {)" << "\n"; |
| 314 | nmsProfileStream << R"( "raw": [)" << "\n"; |
| 315 | bool isFirst = true; |
| 316 | for (auto duration : nmsDurations) |
| 317 | { |
| 318 | if (!isFirst) |
| 319 | { |
| 320 | nmsProfileStream << ",\n"; |
| 321 | } |
| 322 | |
| 323 | nmsProfileStream << " " << duration.count(); |
| 324 | isFirst = false; |
| 325 | } |
| 326 | nmsProfileStream << "\n"; |
| 327 | nmsProfileStream << R"( "units": "us")" << "\n"; |
| 328 | nmsProfileStream << " ]" << "\n"; |
| 329 | nmsProfileStream << " }" << "\n"; |
| 330 | nmsProfileStream << "}" << "\n"; |
| 331 | nmsProfileStream.close(); |
| 332 | |
| 333 | ARMNN_LOG(info) << "Run completed"; |
| 334 | return 0; |
| 335 | } |