yolov5 tensorrt

目录

python tensorrt:激活函数：hard_sigmoid

Config

c++ diou_nms

---

这个感觉靠谱，还没试：

tensorrtx/yolov5 at master · wang-xinyu/tensorrtx · GitHub

这个不错，c++调通的版本：

yolov5转tensorrt c++_jacke121的专栏-CSDN博客

yolov5 tensorrt in8训练开源项目：

https://github.com/maggiez0138/yolov5_quant_sample

python tensorrt:激活函数：hard_sigmoid

https://github.com/TrojanXu/yolov5-tensorrt

https://github.com/wang-xinyu/tensorrtx

The Pytorch implementation is ultralytics/yolov5.

Currently, we support yolov5 v1.0(yolov5s only), v2.0, v3.0 and v3.1.

• For yolov5 v3.1, please visit yolov5 release v3.1, and use the latest commit of this repo.
• For yolov5 v3.0, please visit yolov5 release v3.0, and use the latest commit of this repo.
• For yolov5 v2.0, please visit yolov5 release v2.0, and checkout commit '7cd092d' of this repo.
• For yolov5 v1.0, please visit yolov5 release v1.0, and checkout commit '0504551' of this repo.

Config

• Choose the model s/m/l/x by NET macro in yolov5.cpp
• Input shape defined in yololayer.h
• Number of classes defined in yololayer.h, DO NOT FORGET TO ADAPT THIS, If using your own model
• FP16/FP32 can be selected by the macro in yolov5.cpp
• GPU id can be selected by the macro in yolov5.cpp
• NMS thresh in yolov5.cpp
• BBox confidence thresh in yolov5.cpp
• Batch size in yolov5.cpp

v2也是leakyrelu。

2.0:这个是匹配3.0的版本，用的leakyrelu,可以检测，v3.0自己训练的精度比较低

GitHub - BaofengZan/yolov5_2.0-TensorRt: U版yolov5 2.0的tensorrt加速

https://github.com/AIpakchoi/yolov5_tensorrt/blob/110762eea4a7a53a91bbce35f94239136db157d4/yolov5l/common.hpp

GitHub - baituhuangyu/yolov5-tensorrt: yolov5 tensorrt inference

https://github.com/Thinker-or-Dreamer/UAV-And-RobotArm/tree/master/yolov5

HardSwishLayer_TRT

https://github.com/hlld/tensorrt-yolov5

linux的：

2020.10.23 17天以前更新的，激活函数：kHARD_SIGMOID

https://github.com/wang-xinyu/tensorrtx/tree/master/yolov5

yololayer.h

修改自己类别个数：

    static constexpr int CLASS_NUM = 1;

下面是阈值参数，nms阈值参数：

#define USE_FP16  // comment out this if want to use FP32
#define DEVICE 0  // GPU id
#define NMS_THRESH 0.4
#define CONF_THRESH 0.3


std::cerr << "yolov5_rt.exe -s s // serialize model to plan file" << std::endl;
std::cerr << "yolov5_rt.exe -e s -c 0 // detect cam" << std::endl;
std::cerr << "yolov5_rt.exe -e s -d samples  // deserialize plan file and run inference" << std::endl;

后面3行是调用demo，分两步，编译和执行

问题：原版网络检测出来的框没问题

自己训练的，tensorrt检测的与pytorch检测出来的有偏差，原因还未找到。

可能是anchors的原因，但是没找到证据。

c++ diou_nms

struct BBox
{
    float x1, y1, x2, y2;
};

struct BBoxInfo
{
    BBox box;
    int label;
    int classId; // For coco benchmarking
    float prob;
};

std::vector<BBoxInfo> diou_nms(const float nmsThresh, std::vector<BBoxInfo> binfo)
{
	auto overlap1D = [](float x1min, float x1max, float x2min, float x2max) -> float
	{
		if (x1min > x2min)
		{
			std::swap(x1min, x2min);
			std::swap(x1max, x2max);
		}
		return x1max < x2min ? 0 : std::min(x1max, x2max) - x2min;
	};
	auto computeIoU = [&overlap1D](BBox& bbox1, BBox& bbox2) -> float
	{
		float overlapX = overlap1D(bbox1.x1, bbox1.x2, bbox2.x1, bbox2.x2);
		float overlapY = overlap1D(bbox1.y1, bbox1.y2, bbox2.y1, bbox2.y2);
		float area1 = (bbox1.x2 - bbox1.x1) * (bbox1.y2 - bbox1.y1);
		float area2 = (bbox2.x2 - bbox2.x1) * (bbox2.y2 - bbox2.y1);
		float overlap2D = overlapX * overlapY;
		float u = area1 + area2 - overlap2D;
		return u == 0 ? 0 : overlap2D / u;
	};

	//https://arxiv.org/pdf/1911.08287.pdf
	auto R = [](BBox &bbox1,BBox &bbox2) ->float
	{
		float center1_x = (bbox1.x1 + bbox1.x2) / 2.f;
		float center1_y = (bbox1.y1 + bbox1.y2) / 2.f;
		float center2_x = (bbox2.x1 + bbox2.x2) / 2.f;
		float center2_y = (bbox2.y1 + bbox2.y2) / 2.f;

		float d_center = (center1_x - center2_x)* (center1_x - center2_x) 
						+ (center1_y - center2_y)*(center1_y - center2_y);
		//smallest_enclosing box
		float box_x1 = std::min({ bbox1.x1, bbox1.x2, bbox2.x1, bbox2.x2 });
		float box_y1 = std::min({ bbox1.y1, bbox1.y2, bbox2.y1, bbox2.y2 });
		float box_x2 = std::max({ bbox1.x1, bbox1.x2, bbox2.x1, bbox2.x2 });
		float box_y2 = std::max({ bbox1.y1, bbox1.y2, bbox2.y1, bbox2.y2 });

		float d_diagonal = (box_x1 - box_x2) * (box_x1 - box_x2) +
						   (box_y1 - box_y2) * (box_y1 - box_y2);

		return d_center / d_diagonal;
	};
	std::stable_sort(binfo.begin(), binfo.end(),
		[](const BBoxInfo& b1, const BBoxInfo& b2) { return b1.prob > b2.prob; });
	std::vector<BBoxInfo> out;
	for (auto& i : binfo)
	{
		bool keep = true;
		for (auto& j : out)
		{
			if (keep)
			{
				float overlap = computeIoU(i.box, j.box);
				float r = R(i.box, j.box);
				keep = (overlap-r) <= nmsThresh;
			}
			else
				break;
		}
		if (keep) out.push_back(i);
	}
	return out;
}

设置阈值：

文件路径：class_yolo_detector.hpp

	void parse_config()
	{
		_yolo_info.networkType = _vec_net_type[_config.net_type];
		_yolo_info.configFilePath = _config.file_model_cfg;
		_yolo_info.wtsFilePath = _config.file_model_weights;
		_yolo_info.precision = _vec_precision[_config.inference_precison];
		_yolo_info.deviceType = "kGPU";
		auto npos = _yolo_info.wtsFilePath.find(".weights");
		assert(npos != std::string::npos
			&& "wts file file not recognised. File needs to be of '.weights' format");
		_yolo_info.data_path = _yolo_info.wtsFilePath.substr(0, npos);
		_yolo_info.calibrationTablePath = _yolo_info.data_path + "-calibration.table";
		_yolo_info.inputBlobName = "data";

		_infer_param.printPerfInfo = false;
		_infer_param.printPredictionInfo = false;
		_infer_param.calibImages = _config.calibration_image_list_file_txt;
		_infer_param.calibImagesPath = "";
		_infer_param.probThresh = _config.detect_thresh;
		_infer_param.nmsThresh = 0.3;
	}