首页 > 其他分享> > [E2E_L8_1]segmentation_demo道路分割例子和GOMFCTemplate的初步融合

[E2E_L8_1]segmentation_demo道路分割例子和GOMFCTemplate的初步融合

2019-08-09 11:54:06 作者：互联网

一、来源

模型例子自己带来副图像

二、简化 #include <algorithm> #include <fstream> #include <iomanip> #include <vector> #include <string> #include <chrono> #include <memory> #include <utility> #include <format_reader_ptr.h> #include <inference_engine.hpp> #include <ext_list.hpp> #include <samples/slog.hpp> #include <samples/args_helper.hpp> #include <samples/ocv_common.hpp> #include <format_reader_ptr.h> #include "segmentation_demo.h" using namespace InferenceEngine; using namespace std; using namespace cv; //-i E:/OpenVINO_modelZoo/road.png -m E:/OpenVINO_modelZoo/road-segmentation-adas-0001.xml void main() { std::vector<std::string> images; string imageNames = "E:/OpenVINO_modelZoo/road.png"; images.push_back(imageNames); // --------------------------- 1.为IE准备插件------------------------------------- InferencePlugin plugin(PluginDispatcher().getSuitablePlugin(TargetDevice::eCPU)); printPluginVersion(plugin, std::cout);//正确回显表示成功 plugin.AddExtension(std::make_shared<Extensions::Cpu::CpuExtensions>());//Extension，useful // --------------------------- 2.读取IR模型(xml和bin)--------------------------------- CNNNetReader networkReader; networkReader.ReadNetwork("E:/OpenVINO_modelZoo/road-segmentation-adas-0001.xml"); networkReader.ReadWeights("E:/OpenVINO_modelZoo/road-segmentation-adas-0001.bin"); CNNNetwork network = networkReader.getNetwork(); // --------------------------- 3. 准备输入输出的------------------------------------------ InputsDataMap inputInfo(network.getInputsInfo());//获得输入信息 BlobMap inputBlobs; //保持所有输入的blob数据 if (inputInfo.size() != 1) throw std::logic_error("错误，该模型应该为单输入"); auto inputInfoItem = *inputInfo.begin();//开始读入 std::vector<std::shared_ptr<unsigned char>> imagesData; for (auto & i : images) { FormatReader::ReaderPtr reader(i.c_str()); //使用FormatReader来读取图片数据,这里的images是一个vector，注意对于批量数据的读取 if (reader.get() == nullptr) { slog::warn << "Image " + i + " 无法读取!" << slog::endl; continue; } /** 获得图片数据 **/ std::shared_ptr<unsigned char> data( reader->getData(inputInfoItem.second->getTensorDesc().getDims()[3], inputInfoItem.second->getTensorDesc().getDims()[2])); if (data.get() != nullptr) { imagesData.push_back(data); } } if (imagesData.empty()) throw std::logic_error("错误的格式，请检查!"); network.setBatchSize(imagesData.size()); slog::info << "Batch size is " << std::to_string(networkReader.getNetwork().getBatchSize()) << slog::endl; inputInfoItem.second->setPrecision(Precision::U8); //准备输出数据 OutputsDataMap outputInfo(network.getOutputsInfo()); std::string firstOutputName; for (auto & item : outputInfo) { if (firstOutputName.empty()) { firstOutputName = item.first; } DataPtr outputData = item.second; if (!outputData) { throw std::logic_error("错误的格式，请检查！"); } item.second->setPrecision(Precision::FP32); } // --------------------------- 4. 读取模型 ------------------------------------------（后面这些操作应该可以合并了） ExecutableNetwork executableNetwork = plugin.LoadNetwork(network, {}); // --------------------------- 5. 创建推断 ------------------------------------------------- InferRequest infer_request = executableNetwork.CreateInferRequest(); // --------------------------- 6. 将数据塞入模型 ------------------------------------------------- for (const auto & item : inputInfo) { /** 创建输入BLOB **/ Blob::Ptr input = infer_request.GetBlob(item.first); /** 3 通道塞数据 **/ size_t num_channels = input->getTensorDesc().getDims()[1]; size_t image_size = input->getTensorDesc().getDims()[3] * input->getTensorDesc().getDims()[2]; auto data = input->buffer().as<PrecisionTrait<Precision::U8>::value_type*>(); for (size_t image_id = 0; image_id < imagesData.size(); ++image_id) { for (size_t pid = 0; pid < image_size; pid++) { for (size_t ch = 0; ch < num_channels; ++ch) { data[image_id * image_size * num_channels + ch * image_size + pid] = imagesData.at(image_id).get()[pid*num_channels + ch]; } } } } // --------------------------- 7. 推断结果 ------------------------------------------------- for (size_t iter = 0; iter < images.size(); ++iter) { infer_request.Infer();//多张图片多次推断 } // --------------------------- 8. 处理结果------------------------------------------------------- slog::info << "输出结果" << slog::endl; const Blob::Ptr output_blob = infer_request.GetBlob(firstOutputName); const auto output_data = output_blob->buffer().as<float*>(); size_t N = output_blob->getTensorDesc().getDims().at(0); size_t C = output_blob->getTensorDesc().getDims().at(1); size_t H = output_blob->getTensorDesc().getDims().at(2); size_t W = output_blob->getTensorDesc().getDims().at(3); size_t image_stride = W * H * C; for (size_t image = 0; image < N; ++image) { std::vector<std::vector<size_t>> outArrayClasses(H, std::vector<size_t>(W, 0)); std::vector<std::vector<float>> outArrayProb(H, std::vector<float>(W, 0.)); for (size_t w = 0; w < W; ++w) { for (size_t h = 0; h < H; ++h) { if (C == 1) { outArrayClasses[h][w] = static_cast<size_t>(output_data[image_stride * image + W * h + w]); } else { for (size_t ch = 0; ch < C; ++ch) { auto data = output_data[image_stride * image + W * H * ch + W * h + w]; if (data > outArrayProb[h][w]) { outArrayClasses[h][w] = ch; outArrayProb[h][w] = data; } } } } } std::string fileName = "out_" + std::to_string(image) + ".bmp"; std::ofstream outFile(fileName, std::ofstream::binary); if (!outFile.is_open()) { throw std::logic_error("Can't open file : " + fileName); } writeOutputBmp(outArrayClasses, C, outFile); //输出的代码 slog::info << "File : " << fileName << " was created" << slog::endl; } // ----------------------------------------------------------------------------------------------------- } 在改写的过程中有几点注意 1、添加lib和对应的dll文件，主要就是用于文件读取的；

以及format_reader.dll 文件放到目录下面； 2、头文件修改正确

三、改写 这个代码里面使用的是format_reader，使用起来颇为不方便，修改为OpenCV负责输入输出。 #include <algorithm> #include <fstream> #include <iomanip> #include <vector> #include <string> #include <chrono> #include <memory> #include <utility> #include <format_reader_ptr.h> #include <inference_engine.hpp> #include <ext_list.hpp> #include <samples/slog.hpp> #include <samples/args_helper.hpp> #include <samples/ocv_common.hpp> #include <format_reader_ptr.h> #include "segmentation_demo.h" using namespace InferenceEngine; using namespace std; using namespace cv; //-i E:/OpenVINO_modelZoo/road.png -m E:/OpenVINO_modelZoo/road-segmentation-adas-0001.xml void main() { std::vector<std::string> images; string imageNames = "E:/OpenVINO_modelZoo/road.png"; images.push_back(imageNames); // --------------------------- 1.为IE准备插件------------------------------------- InferencePlugin plugin(PluginDispatcher().getSuitablePlugin(TargetDevice::eCPU)); printPluginVersion(plugin, std::cout);//正确回显表示成功 plugin.AddExtension(std::make_shared<Extensions::Cpu::CpuExtensions>());//Extension，useful // --------------------------- 2.读取IR模型(xml和bin)--------------------------------- CNNNetReader networkReader; networkReader.ReadNetwork("E:/OpenVINO_modelZoo/road-segmentation-adas-0001.xml"); networkReader.ReadWeights("E:/OpenVINO_modelZoo/road-segmentation-adas-0001.bin"); CNNNetwork network = networkReader.getNetwork(); // --------------------------- 3. 准备输入输出的------------------------------------------ InputsDataMap inputInfo(network.getInputsInfo());//获得输入信息 BlobMap inputBlobs; //保持所有输入的blob数据 if (inputInfo.size() != 1) throw std::logic_error("错误，该模型应该为单输入"); //auto lrInputInfoItem = *inputInfo.begin();//开始读入 //int w = static_cast<int>(lrInputInfoItem.second->getTensorDesc().getDims()[3]); //这种写法也是可以的,它的first就是data //int h = static_cast<int>(lrInputInfoItem.second->getTensorDesc().getDims()[2]); auto lrInputInfoItem = inputInfo["data"]; //开始读入 int w = static_cast<int>(lrInputInfoItem->getTensorDesc().getDims()[3]); //模型要求的输入大小 int h = static_cast<int>(lrInputInfoItem->getTensorDesc().getDims()[2]); Mat src = imread(imageNames); if (src.empty()) return; network.setBatchSize(1);//只有1副图片，故BatchSize = 1 //准备输出数据 OutputsDataMap outputInfo(network.getOutputsInfo());//获得输出信息 std::string firstOutputName; for (auto &item : outputInfo) { if (firstOutputName.empty()) { firstOutputName = item.first; } DataPtr outputData = item.second; if (!outputData) { throw std::logic_error("错误的格式，请检查！"); } item.second->setPrecision(Precision::FP32); } // --------------------------- 4. 读取模型 ------------------------------------------（后面这些操作应该可以合并了） ExecutableNetwork executableNetwork = plugin.LoadNetwork(network, {}); // --------------------------- 5. 创建推断 ------------------------------------------------- InferRequest infer_request = executableNetwork.CreateInferRequest(); // --------------------------- 6. 将数据塞入模型 ------------------------------------------------- Blob::Ptr lrInputBlob = infer_request.GetBlob("data"); //data这个名字是我看出来的，实际上这里可以更统一一些 matU8ToBlob<float_t>(src, lrInputBlob, 0);//重要的转换函数，第3个参数是batchSize,应该是自己+1的 // --------------------------- 7. 推断结果 ------------------------------------------------- infer_request.Infer();//多张图片多次推断 // --------------------------- 8. 处理结果------------------------------------------------------- const Blob::Ptr outputBlob = infer_request.GetBlob(firstOutputName); const auto outputData = outputBlob->buffer().as<PrecisionTrait<Precision::FP32>::value_type*>(); size_t numOfImages = outputBlob->getTensorDesc().getDims()[0]; size_t numOfChannels = outputBlob->getTensorDesc().getDims()[1]; h = outputBlob->getTensorDesc().getDims()[2]; w = outputBlob->getTensorDesc().getDims()[3]; size_t nunOfPixels = w * h; //写在内存里的结果，还是要拼出来的 std::vector<cv::Mat> imgPlanes{ cv::Mat(h, w, CV_32FC1, &(outputData[0])), cv::Mat(h, w, CV_32FC1, &(outputData[nunOfPixels])), cv::Mat(h, w, CV_32FC1, &(outputData[nunOfPixels * 2])) }; for (auto & img : imgPlanes) //本来是平的 img.convertTo(img, CV_8UC1, 255); cv::Mat resultImg; cv::merge(imgPlanes, resultImg); cv::imshow("result", resultImg); cv::waitKey(); } 这里需要注意的一点是在读取图片的大小的时候，我这里使用了 Blob::Ptr lrInputBlob = infer_request.GetBlob("data"); //data这个名字是我看出来的，实际上这里可以更统一一些其前提是我知道这里叫做 data，这里可以改成更统一的方式。从结果来看，我认为OpenCV转换后的结果更好。当然差别只是在着色而已。