diff --git a/deploy/cpp/readme.md b/deploy/cpp/readme.md index a8af240c9..72ce8dacc 100644 --- a/deploy/cpp/readme.md +++ b/deploy/cpp/readme.md @@ -215,9 +215,9 @@ cp ../configs/inference_cls.yaml tools/ #### 2.3.2 执行 ```shell -./build/clas_system -c inference_cls.yaml +./build/clas_system -c tools/inference_cls.yaml # or -./build/clas_system -config inference_cls.yaml +./build/clas_system -config tools/inference_cls.yaml ``` 最终屏幕上会输出结果，如下图所示。 diff --git a/deploy/cpp_shitu/src/feature_extracter.cpp b/deploy/cpp_shitu/src/feature_extracter.cpp index f442b51ce..37c7590ff 100644 --- a/deploy/cpp_shitu/src/feature_extracter.cpp +++ b/deploy/cpp_shitu/src/feature_extracter.cpp @@ -18,102 +18,102 @@ namespace Feature { -void FeatureExtracter::LoadModel(const std::string &model_path, - const std::string ¶ms_path) { - paddle_infer::Config config; - config.SetModel(model_path, params_path); + void FeatureExtracter::LoadModel(const std::string &model_path, + const std::string ¶ms_path) { + paddle_infer::Config config; + config.SetModel(model_path, params_path); - if (this->use_gpu_) { - config.EnableUseGpu(this->gpu_mem_, this->gpu_id_); - if (this->use_tensorrt_) { - config.EnableTensorRtEngine( - 1 << 20, 1, 3, - this->use_fp16_ ? paddle_infer::Config::Precision::kHalf - : paddle_infer::Config::Precision::kFloat32, - false, false); + if (this->use_gpu_) { + config.EnableUseGpu(this->gpu_mem_, this->gpu_id_); + if (this->use_tensorrt_) { + config.EnableTensorRtEngine( + 1 << 20, 1, 3, + this->use_fp16_ ? paddle_infer::Config::Precision::kHalf + : paddle_infer::Config::Precision::kFloat32, + false, false); + } + } else { + config.DisableGpu(); + if (this->use_mkldnn_) { + config.EnableMKLDNN(); + // cache 10 different shapes for mkldnn to avoid memory leak + config.SetMkldnnCacheCapacity(10); + } + config.SetCpuMathLibraryNumThreads(this->cpu_math_library_num_threads_); + } + + config.SwitchUseFeedFetchOps(false); + // true for multiple input + config.SwitchSpecifyInputNames(true); + + config.SwitchIrOptim(true); + + config.EnableMemoryOptim(); + config.DisableGlogInfo(); + + this->predictor_ = CreatePredictor(config); } - } else { - config.DisableGpu(); - if (this->use_mkldnn_) { - config.EnableMKLDNN(); - // cache 10 different shapes for mkldnn to avoid memory leak - config.SetMkldnnCacheCapacity(10); + + void FeatureExtracter::Run(cv::Mat &img, std::vector &out_data, + std::vector ×) { + cv::Mat resize_img; + std::vector time; + + auto preprocess_start = std::chrono::steady_clock::now(); + this->resize_op_.Run(img, resize_img, this->resize_short_, + this->resize_size_); + + this->normalize_op_.Run(&resize_img, this->mean_, this->std_, this->scale_); + std::vector input(1 * 3 * resize_img.rows * resize_img.cols, 0.0f); + this->permute_op_.Run(&resize_img, input.data()); + + auto input_names = this->predictor_->GetInputNames(); + auto input_t = this->predictor_->GetInputHandle(input_names[0]); + input_t->Reshape({1, 3, resize_img.rows, resize_img.cols}); + auto preprocess_end = std::chrono::steady_clock::now(); + + auto infer_start = std::chrono::steady_clock::now(); + input_t->CopyFromCpu(input.data()); + this->predictor_->Run(); + + auto output_names = this->predictor_->GetOutputNames(); + auto output_t = this->predictor_->GetOutputHandle(output_names[0]); + std::vector output_shape = output_t->shape(); + int out_num = std::accumulate(output_shape.begin(), output_shape.end(), 1, + std::multiplies()); + + out_data.resize(out_num); + output_t->CopyToCpu(out_data.data()); + auto infer_end = std::chrono::steady_clock::now(); + + auto postprocess_start = std::chrono::steady_clock::now(); + if (this->feature_norm) + FeatureNorm(out_data); + auto postprocess_end = std::chrono::steady_clock::now(); + + std::chrono::duration preprocess_diff = + preprocess_end - preprocess_start; + time.push_back(double(preprocess_diff.count()) * 1000); + std::chrono::duration inference_diff = infer_end - infer_start; + double inference_cost_time = double(inference_diff.count() * 1000); + time.push_back(inference_cost_time); + // std::chrono::duration postprocess_diff = + // postprocess_end - postprocess_start; + time.push_back(0); + + // std::cout << "result: " << std::endl; + // std::cout << "\tclass id: " << maxPosition << std::endl; + // std::cout << std::fixed << std::setprecision(10) + // << "\tscore: " << double(out_data[maxPosition]) << std::endl; + times[0] += time[0]; + times[1] += time[1]; + times[2] += time[2]; } - config.SetCpuMathLibraryNumThreads(this->cpu_math_library_num_threads_); - } - config.SwitchUseFeedFetchOps(false); - // true for multiple input - config.SwitchSpecifyInputNames(true); - - config.SwitchIrOptim(true); - - config.EnableMemoryOptim(); - config.DisableGlogInfo(); - - this->predictor_ = CreatePredictor(config); -} - -void FeatureExtracter::Run(cv::Mat &img, std::vector &out_data, - std::vector ×) { - cv::Mat resize_img; - std::vector time; - - auto preprocess_start = std::chrono::steady_clock::now(); - this->resize_op_.Run(img, resize_img, this->resize_short_, - this->resize_size_); - - this->normalize_op_.Run(&resize_img, this->mean_, this->std_, this->scale_); - std::vector input(1 * 3 * resize_img.rows * resize_img.cols, 0.0f); - this->permute_op_.Run(&resize_img, input.data()); - - auto input_names = this->predictor_->GetInputNames(); - auto input_t = this->predictor_->GetInputHandle(input_names[0]); - input_t->Reshape({1, 3, resize_img.rows, resize_img.cols}); - auto preprocess_end = std::chrono::steady_clock::now(); - - auto infer_start = std::chrono::steady_clock::now(); - input_t->CopyFromCpu(input.data()); - this->predictor_->Run(); - - auto output_names = this->predictor_->GetOutputNames(); - auto output_t = this->predictor_->GetOutputHandle(output_names[0]); - std::vector output_shape = output_t->shape(); - int out_num = std::accumulate(output_shape.begin(), output_shape.end(), 1, - std::multiplies()); - - out_data.resize(out_num); - output_t->CopyToCpu(out_data.data()); - auto infer_end = std::chrono::steady_clock::now(); - - auto postprocess_start = std::chrono::steady_clock::now(); - if (this->feature_norm) - FeatureNorm(out_data); - auto postprocess_end = std::chrono::steady_clock::now(); - - std::chrono::duration preprocess_diff = - preprocess_end - preprocess_start; - time.push_back(double(preprocess_diff.count()) * 1000); - std::chrono::duration inference_diff = infer_end - infer_start; - double inference_cost_time = double(inference_diff.count() * 1000); - time.push_back(inference_cost_time); - // std::chrono::duration postprocess_diff = - // postprocess_end - postprocess_start; - time.push_back(0); - - // std::cout << "result: " << std::endl; - // std::cout << "\tclass id: " << maxPosition << std::endl; - // std::cout << std::fixed << std::setprecision(10) - // << "\tscore: " << double(out_data[maxPosition]) << std::endl; - times[0] += time[0]; - times[1] += time[1]; - times[2] += time[2]; -} - -void FeatureExtracter::FeatureNorm(std::vector &featuer) { - float featuer_sqrt = std::sqrt(std::inner_product( - featuer.begin(), featuer.end(), featuer.begin(), 0.0f)); - for (int i = 0; i < featuer.size(); ++i) - featuer[i] /= featuer_sqrt; -} + void FeatureExtracter::FeatureNorm(std::vector &featuer) { + float featuer_sqrt = std::sqrt(std::inner_product( + featuer.begin(), featuer.end(), featuer.begin(), 0.0f)); + for (int i = 0; i < featuer.size(); ++i) + featuer[i] /= featuer_sqrt; + } } // namespace Feature diff --git a/deploy/cpp_shitu/src/main.cpp b/deploy/cpp_shitu/src/main.cpp index 700d9aa3e..be37d3afd 100644 --- a/deploy/cpp_shitu/src/main.cpp +++ b/deploy/cpp_shitu/src/main.cpp @@ -37,304 +37,306 @@ using namespace std; using namespace cv; -DEFINE_string(config, "", "Path of yaml file"); -DEFINE_string(c, "", "Path of yaml file"); +DEFINE_string(config, +"", "Path of yaml file"); +DEFINE_string(c, +"", "Path of yaml file"); -void DetPredictImage(const std::vector &batch_imgs, - const std::vector &all_img_paths, +void DetPredictImage(const std::vector &batch_imgs, + const std::vector &all_img_paths, const int batch_size, Detection::ObjectDetector *det, - std::vector &im_result, + std::vector &im_result, std::vector &im_bbox_num, std::vector &det_t, const bool visual_det = false, const bool run_benchmark = false, const std::string &output_dir = "output") { - int steps = ceil(float(all_img_paths.size()) / batch_size); - // printf("total images = %d, batch_size = %d, total steps = %d\n", - // all_img_paths.size(), batch_size, steps); - for (int idx = 0; idx < steps; idx++) { - int left_image_cnt = all_img_paths.size() - idx * batch_size; - if (left_image_cnt > batch_size) { - left_image_cnt = batch_size; - } - // for (int bs = 0; bs < left_image_cnt; bs++) { - // std::string image_file_path = all_img_paths.at(idx * batch_size+bs); - // cv::Mat im = cv::imread(image_file_path, 1); - // batch_imgs.insert(batch_imgs.end(), im); - // } + int steps = ceil(float(all_img_paths.size()) / batch_size); + // printf("total images = %d, batch_size = %d, total steps = %d\n", + // all_img_paths.size(), batch_size, steps); + for (int idx = 0; idx < steps; idx++) { + int left_image_cnt = all_img_paths.size() - idx * batch_size; + if (left_image_cnt > batch_size) { + left_image_cnt = batch_size; + } + // for (int bs = 0; bs < left_image_cnt; bs++) { + // std::string image_file_path = all_img_paths.at(idx * batch_size+bs); + // cv::Mat im = cv::imread(image_file_path, 1); + // batch_imgs.insert(batch_imgs.end(), im); + // } - // Store all detected result - std::vector result; - std::vector bbox_num; - std::vector det_times; - bool is_rbox = false; - if (run_benchmark) { - det->Predict(batch_imgs, 10, 10, &result, &bbox_num, &det_times); - } else { - det->Predict(batch_imgs, 0, 1, &result, &bbox_num, &det_times); - // get labels and colormap - auto labels = det->GetLabelList(); - auto colormap = Detection::GenerateColorMap(labels.size()); + // Store all detected result + std::vector result; + std::vector bbox_num; + std::vector det_times; + bool is_rbox = false; + if (run_benchmark) { + det->Predict(batch_imgs, 10, 10, &result, &bbox_num, &det_times); + } else { + det->Predict(batch_imgs, 0, 1, &result, &bbox_num, &det_times); + // get labels and colormap + auto labels = det->GetLabelList(); + auto colormap = Detection::GenerateColorMap(labels.size()); - int item_start_idx = 0; - for (int i = 0; i < left_image_cnt; i++) { - cv::Mat im = batch_imgs[i]; - int detect_num = 0; + int item_start_idx = 0; + for (int i = 0; i < left_image_cnt; i++) { + cv::Mat im = batch_imgs[i]; + int detect_num = 0; - for (int j = 0; j < bbox_num[i]; j++) { - Detection::ObjectResult item = result[item_start_idx + j]; - if (item.confidence < det->GetThreshold() || item.class_id == -1) { - continue; - } - detect_num += 1; - im_result.push_back(item); - if (visual_det) { - if (item.rect.size() > 6) { - is_rbox = true; - printf( - "class=%d confidence=%.4f rect=[%d %d %d %d %d %d %d %d]\n", - item.class_id, item.confidence, item.rect[0], item.rect[1], - item.rect[2], item.rect[3], item.rect[4], item.rect[5], - item.rect[6], item.rect[7]); - } else { - printf("class=%d confidence=%.4f rect=[%d %d %d %d]\n", - item.class_id, item.confidence, item.rect[0], item.rect[1], - item.rect[2], item.rect[3]); + for (int j = 0; j < bbox_num[i]; j++) { + Detection::ObjectResult item = result[item_start_idx + j]; + if (item.confidence < det->GetThreshold() || item.class_id == -1) { + continue; + } + detect_num += 1; + im_result.push_back(item); + if (visual_det) { + if (item.rect.size() > 6) { + is_rbox = true; + printf( + "class=%d confidence=%.4f rect=[%d %d %d %d %d %d %d %d]\n", + item.class_id, item.confidence, item.rect[0], item.rect[1], + item.rect[2], item.rect[3], item.rect[4], item.rect[5], + item.rect[6], item.rect[7]); + } else { + printf("class=%d confidence=%.4f rect=[%d %d %d %d]\n", + item.class_id, item.confidence, item.rect[0], item.rect[1], + item.rect[2], item.rect[3]); + } + } + } + im_bbox_num.push_back(detect_num); + item_start_idx = item_start_idx + bbox_num[i]; + + // Visualization result + if (visual_det) { + std::cout << all_img_paths.at(idx * batch_size + i) + << " The number of detected box: " << detect_num + << std::endl; + cv::Mat vis_img = Detection::VisualizeResult(im, im_result, labels, + colormap, is_rbox); + std::vector compression_params; + compression_params.push_back(CV_IMWRITE_JPEG_QUALITY); + compression_params.push_back(95); + std::string output_path(output_dir); + if (output_dir.rfind(OS_PATH_SEP) != output_dir.size() - 1) { + output_path += OS_PATH_SEP; + } + std::string image_file_path = all_img_paths.at(idx * batch_size + i); + output_path += + image_file_path.substr(image_file_path.find_last_of('/') + 1); + cv::imwrite(output_path, vis_img, compression_params); + printf("Visualized output saved as %s\n", output_path.c_str()); + } } - } } - im_bbox_num.push_back(detect_num); - item_start_idx = item_start_idx + bbox_num[i]; - - // Visualization result - if (visual_det) { - std::cout << all_img_paths.at(idx * batch_size + i) - << " The number of detected box: " << detect_num - << std::endl; - cv::Mat vis_img = Detection::VisualizeResult(im, im_result, labels, - colormap, is_rbox); - std::vector compression_params; - compression_params.push_back(CV_IMWRITE_JPEG_QUALITY); - compression_params.push_back(95); - std::string output_path(output_dir); - if (output_dir.rfind(OS_PATH_SEP) != output_dir.size() - 1) { - output_path += OS_PATH_SEP; - } - std::string image_file_path = all_img_paths.at(idx * batch_size + i); - output_path += - image_file_path.substr(image_file_path.find_last_of('/') + 1); - cv::imwrite(output_path, vis_img, compression_params); - printf("Visualized output saved as %s\n", output_path.c_str()); - } - } + det_t[0] += det_times[0]; + det_t[1] += det_times[1]; + det_t[2] += det_times[2]; } - det_t[0] += det_times[0]; - det_t[1] += det_times[1]; - det_t[2] += det_times[2]; - } } void PrintResult(std::string &img_path, - std::vector &det_result, + std::vector &det_result, std::vector &indeices, VectorSearch &vector_search, SearchResult &search_result) { - printf("%s:\n", img_path.c_str()); - for (int i = 0; i < indeices.size(); ++i) { - int t = indeices[i]; - printf("\tresult%d: bbox[%d, %d, %d, %d], score: %f, label: %s\n", i, - det_result[t].rect[0], det_result[t].rect[1], det_result[t].rect[2], - det_result[t].rect[3], det_result[t].confidence, - vector_search.GetLabel(search_result.I[search_result.return_k * t]) - .c_str()); - } + printf("%s:\n", img_path.c_str()); + for (int i = 0; i < indeices.size(); ++i) { + int t = indeices[i]; + printf("\tresult%d: bbox[%d, %d, %d, %d], score: %f, label: %s\n", i, + det_result[t].rect[0], det_result[t].rect[1], det_result[t].rect[2], + det_result[t].rect[3], det_result[t].confidence, + vector_search.GetLabel(search_result.I[search_result.return_k * t]) + .c_str()); + } } int main(int argc, char **argv) { - google::ParseCommandLineFlags(&argc, &argv, true); - std::string yaml_path = ""; - if (FLAGS_config == "" && FLAGS_c == "") { - std::cerr << "[ERROR] usage: " << std::endl - << argv[0] << " -c $yaml_path" << std::endl - << "or:" << std::endl - << argv[0] << " -config $yaml_path" << std::endl; - exit(1); - } else if (FLAGS_config != "") { - yaml_path = FLAGS_config; - } else { - yaml_path = FLAGS_c; - } - - YamlConfig config(yaml_path); - config.PrintConfigInfo(); - - // initialize detector, rec_Model, vector_search - Feature::FeatureExtracter feature_extracter(config.config_file); - Detection::ObjectDetector detector(config.config_file); - VectorSearch searcher(config.config_file); - - // config - const int batch_size = config.config_file["Global"]["batch_size"].as(); - bool visual_det = false; - if (config.config_file["Global"]["visual_det"].IsDefined()) { - visual_det = config.config_file["Global"]["visual_det"].as(); - } - bool benchmark = false; - if (config.config_file["Global"]["benchmark"].IsDefined()) { - benchmark = config.config_file["Global"]["benchmark"].as(); - } - int max_det_results = 5; - if (config.config_file["Global"]["max_det_results"].IsDefined()) { - max_det_results = config.config_file["Global"]["max_det_results"].as(); - } - float rec_nms_thresold = 0.05; - if (config.config_file["Global"]["rec_nms_thresold"].IsDefined()) { - rec_nms_thresold = - config.config_file["Global"]["rec_nms_thresold"].as(); - } - - // load image_file_path - std::string path = - config.config_file["Global"]["infer_imgs"].as(); - std::vector img_files_list; - if (cv::utils::fs::isDirectory(path)) { - std::vector filenames; - cv::glob(path, filenames); - for (auto f : filenames) { - img_files_list.push_back(f); - } - } else { - img_files_list.push_back(path); - } - std::cout << "img_file_list length: " << img_files_list.size() << std::endl; - // for time log - std::vector cls_times = {0, 0, 0}; - std::vector det_times = {0, 0, 0}; - std::vector search_times = {0, 0, 0}; - int instance_num = 0; - // for read images - std::vector batch_imgs; - std::vector img_paths; - // for detection - std::vector det_result; - std::vector det_bbox_num; - // for vector search - std::vector features; - std::vector feature; - // for nms - std::vector indeices; - - int warmup_iter = img_files_list.size() > 5 ? 5 : img_files_list.size(); - if (benchmark) { - img_files_list.insert(img_files_list.begin(), img_files_list.begin(), - img_files_list.begin() + warmup_iter); - } - - for (int idx = 0; idx < img_files_list.size(); ++idx) { - std::string img_path = img_files_list[idx]; - cv::Mat srcimg = cv::imread(img_path, cv::IMREAD_COLOR); - if (!srcimg.data) { - std::cerr << "[ERROR] image read failed! image path: " << img_path - << "\n"; - exit(-1); - } - cv::cvtColor(srcimg, srcimg, cv::COLOR_BGR2RGB); - - batch_imgs.push_back(srcimg); - img_paths.push_back(img_path); - - // step1: get all detection results - DetPredictImage(batch_imgs, img_paths, batch_size, &detector, det_result, - det_bbox_num, det_times, visual_det, false); - - // select max_det_results bbox - if (det_result.size() > max_det_results) { - det_result.resize(max_det_results); - } - instance_num += det_result.size(); - - // step2: add the whole image for recognition to improve recall - Detection::ObjectResult result_whole_img = { - {0, 0, srcimg.cols - 1, srcimg.rows - 1}, 0, 1.0}; - det_result.push_back(result_whole_img); - det_bbox_num[0] = det_result.size() + 1; - - // step3: extract feature for all boxes in an inmage - SearchResult search_result; - for (int j = 0; j < det_result.size(); ++j) { - int w = det_result[j].rect[2] - det_result[j].rect[0]; - int h = det_result[j].rect[3] - det_result[j].rect[1]; - cv::Rect rect(det_result[j].rect[0], det_result[j].rect[1], w, h); - cv::Mat crop_img = srcimg(rect); - feature_extracter.Run(crop_img, feature, cls_times); - features.insert(features.end(), feature.begin(), feature.end()); + google::ParseCommandLineFlags(&argc, &argv, true); + std::string yaml_path = ""; + if (FLAGS_config == "" && FLAGS_c == "") { + std::cerr << "[ERROR] usage: " << std::endl + << argv[0] << " -c $yaml_path" << std::endl + << "or:" << std::endl + << argv[0] << " -config $yaml_path" << std::endl; + exit(1); + } else if (FLAGS_config != "") { + yaml_path = FLAGS_config; + } else { + yaml_path = FLAGS_c; } - // step4: get search result - auto search_start = std::chrono::steady_clock::now(); - search_result = searcher.Search(features.data(), det_result.size()); - auto search_end = std::chrono::steady_clock::now(); + YamlConfig config(yaml_path); + config.PrintConfigInfo(); - // nms for search result - for (int i = 0; i < det_result.size(); ++i) { - det_result[i].confidence = search_result.D[search_result.return_k * i]; + // initialize detector, rec_Model, vector_search + Feature::FeatureExtracter feature_extracter(config.config_file); + Detection::ObjectDetector detector(config.config_file); + VectorSearch searcher(config.config_file); + + // config + const int batch_size = config.config_file["Global"]["batch_size"].as(); + bool visual_det = false; + if (config.config_file["Global"]["visual_det"].IsDefined()) { + visual_det = config.config_file["Global"]["visual_det"].as(); } - NMSBoxes(det_result, searcher.GetThreshold(), rec_nms_thresold, indeices); - auto nms_end = std::chrono::steady_clock::now(); - std::chrono::duration search_diff = search_end - search_start; - search_times[1] += double(search_diff.count() * 1000); - - std::chrono::duration nms_diff = nms_end - search_end; - search_times[2] += double(nms_diff.count() * 1000); - - // print result - if (not benchmark or (benchmark and idx >= warmup_iter)) - PrintResult(img_path, det_result, indeices, searcher, search_result); - - // for postprocess - batch_imgs.clear(); - img_paths.clear(); - det_bbox_num.clear(); - det_result.clear(); - feature.clear(); - features.clear(); - indeices.clear(); - if (benchmark and warmup_iter == idx + 1) { - det_times = {0, 0, 0}; - cls_times = {0, 0, 0}; - search_times = {0, 0, 0}; - instance_num = 0; + bool benchmark = false; + if (config.config_file["Global"]["benchmark"].IsDefined()) { + benchmark = config.config_file["Global"]["benchmark"].as(); + } + int max_det_results = 5; + if (config.config_file["Global"]["max_det_results"].IsDefined()) { + max_det_results = config.config_file["Global"]["max_det_results"].as(); + } + float rec_nms_thresold = 0.05; + if (config.config_file["Global"]["rec_nms_thresold"].IsDefined()) { + rec_nms_thresold = + config.config_file["Global"]["rec_nms_thresold"].as(); } - } - if (benchmark) { - std::string presion = "fp32"; - if (config.config_file["Global"]["use_fp16"].IsDefined() and - config.config_file["Global"]["use_fp16"].as()) - presion = "fp16"; - bool use_gpu = config.config_file["Global"]["use_gpu"].as(); - bool use_tensorrt = config.config_file["Global"]["use_tensorrt"].as(); - bool enable_mkldnn = - config.config_file["Global"]["enable_mkldnn"].as(); - int cpu_num_threads = - config.config_file["Global"]["cpu_num_threads"].as(); - int batch_size = config.config_file["Global"]["batch_size"].as(); - std::vector shape = - config.config_file["Global"]["image_shape"].as>(); - std::string det_shape = std::to_string(shape[0]); - for (int i = 1; i < shape.size(); ++i) - det_shape = det_shape + ", " + std::to_string(shape[i]); + // load image_file_path + std::string path = + config.config_file["Global"]["infer_imgs"].as(); + std::vector img_files_list; + if (cv::utils::fs::isDirectory(path)) { + std::vector filenames; + cv::glob(path, filenames); + for (auto f : filenames) { + img_files_list.push_back(f); + } + } else { + img_files_list.push_back(path); + } + std::cout << "img_file_list length: " << img_files_list.size() << std::endl; + // for time log + std::vector cls_times = {0, 0, 0}; + std::vector det_times = {0, 0, 0}; + std::vector search_times = {0, 0, 0}; + int instance_num = 0; + // for read images + std::vector batch_imgs; + std::vector img_paths; + // for detection + std::vector det_result; + std::vector det_bbox_num; + // for vector search + std::vector features; + std::vector feature; + // for nms + std::vector indeices; - AutoLogger autolog_det("Det", use_gpu, use_tensorrt, enable_mkldnn, - cpu_num_threads, batch_size, det_shape, presion, - det_times, img_files_list.size() - warmup_iter); - autolog_det.report(); - AutoLogger autolog_rec("Rec", use_gpu, use_tensorrt, enable_mkldnn, - cpu_num_threads, batch_size, "3, 224, 224", presion, - cls_times, instance_num); - autolog_rec.report(); - AutoLogger autolog_search("Search", false, use_tensorrt, enable_mkldnn, - cpu_num_threads, batch_size, "dynamic", presion, - search_times, instance_num); - autolog_search.report(); - } - return 0; + int warmup_iter = img_files_list.size() > 5 ? 5 : img_files_list.size(); + if (benchmark) { + img_files_list.insert(img_files_list.begin(), img_files_list.begin(), + img_files_list.begin() + warmup_iter); + } + + for (int idx = 0; idx < img_files_list.size(); ++idx) { + std::string img_path = img_files_list[idx]; + cv::Mat srcimg = cv::imread(img_path, cv::IMREAD_COLOR); + if (!srcimg.data) { + std::cerr << "[ERROR] image read failed! image path: " << img_path + << "\n"; + exit(-1); + } + cv::cvtColor(srcimg, srcimg, cv::COLOR_BGR2RGB); + + batch_imgs.push_back(srcimg); + img_paths.push_back(img_path); + + // step1: get all detection results + DetPredictImage(batch_imgs, img_paths, batch_size, &detector, det_result, + det_bbox_num, det_times, visual_det, false); + + // select max_det_results bbox + if (det_result.size() > max_det_results) { + det_result.resize(max_det_results); + } + instance_num += det_result.size(); + + // step2: add the whole image for recognition to improve recall + Detection::ObjectResult result_whole_img = { + {0, 0, srcimg.cols - 1, srcimg.rows - 1}, 0, 1.0}; + det_result.push_back(result_whole_img); + det_bbox_num[0] = det_result.size() + 1; + + // step3: extract feature for all boxes in an inmage + SearchResult search_result; + for (int j = 0; j < det_result.size(); ++j) { + int w = det_result[j].rect[2] - det_result[j].rect[0]; + int h = det_result[j].rect[3] - det_result[j].rect[1]; + cv::Rect rect(det_result[j].rect[0], det_result[j].rect[1], w, h); + cv::Mat crop_img = srcimg(rect); + feature_extracter.Run(crop_img, feature, cls_times); + features.insert(features.end(), feature.begin(), feature.end()); + } + + // step4: get search result + auto search_start = std::chrono::steady_clock::now(); + search_result = searcher.Search(features.data(), det_result.size()); + auto search_end = std::chrono::steady_clock::now(); + + // nms for search result + for (int i = 0; i < det_result.size(); ++i) { + det_result[i].confidence = search_result.D[search_result.return_k * i]; + } + NMSBoxes(det_result, searcher.GetThreshold(), rec_nms_thresold, indeices); + auto nms_end = std::chrono::steady_clock::now(); + std::chrono::duration search_diff = search_end - search_start; + search_times[1] += double(search_diff.count() * 1000); + + std::chrono::duration nms_diff = nms_end - search_end; + search_times[2] += double(nms_diff.count() * 1000); + + // print result + if (not benchmark or (benchmark and idx >= warmup_iter)) + PrintResult(img_path, det_result, indeices, searcher, search_result); + + // for postprocess + batch_imgs.clear(); + img_paths.clear(); + det_bbox_num.clear(); + det_result.clear(); + feature.clear(); + features.clear(); + indeices.clear(); + if (benchmark and warmup_iter == idx + 1) { + det_times = {0, 0, 0}; + cls_times = {0, 0, 0}; + search_times = {0, 0, 0}; + instance_num = 0; + } + } + + if (benchmark) { + std::string presion = "fp32"; + if (config.config_file["Global"]["use_fp16"].IsDefined() and + config.config_file["Global"]["use_fp16"].as()) + presion = "fp16"; + bool use_gpu = config.config_file["Global"]["use_gpu"].as(); + bool use_tensorrt = config.config_file["Global"]["use_tensorrt"].as(); + bool enable_mkldnn = + config.config_file["Global"]["enable_mkldnn"].as(); + int cpu_num_threads = + config.config_file["Global"]["cpu_num_threads"].as(); + int batch_size = config.config_file["Global"]["batch_size"].as(); + std::vector shape = + config.config_file["Global"]["image_shape"].as < std::vector < int >> (); + std::string det_shape = std::to_string(shape[0]); + for (int i = 1; i < shape.size(); ++i) + det_shape = det_shape + ", " + std::to_string(shape[i]); + + AutoLogger autolog_det("Det", use_gpu, use_tensorrt, enable_mkldnn, + cpu_num_threads, batch_size, det_shape, presion, + det_times, img_files_list.size() - warmup_iter); + autolog_det.report(); + AutoLogger autolog_rec("Rec", use_gpu, use_tensorrt, enable_mkldnn, + cpu_num_threads, batch_size, "3, 224, 224", presion, + cls_times, instance_num); + autolog_rec.report(); + AutoLogger autolog_search("Search", false, use_tensorrt, enable_mkldnn, + cpu_num_threads, batch_size, "dynamic", presion, + search_times, instance_num); + autolog_search.report(); + } + return 0; }