YOLO系列 --- YOLOV7算法（六）：YOLO V7算法onnx模型部署

YOLO系列 — YOLOV7算法（六）：YOLO V7算法onnx模型部署

有很多人来问我，系列基于YOLO v7算法训练出来一个权重文件，算算法如何进行部署。模型所以特地写一篇部署的部署blog~
一般，我们基于pytorch深度学习框架训练出来的系列权重文件是pt格式的，我们可以用python来直接调用这个文件。算算法但是模型实际工业中，一般都是部署c++去调用权重文件的，所以我们需要将pt权重文件转换为能用c++去调用的系列格式。一般来说，算算法我习惯用以下方式：

使用libtorch进行转换，模型将pt转换为torchscript.pt格式的部署权重文件，然后直接用官方提供的系列libtorch来调用
先将pt转换为onnx格式的权重，onnx是算算法一种开放神经网络交换格式。然后用opencv里面的模型api去调用
同样是先转换为onnx格式的，然后用onnx runtime去调用权重文件（本篇blog使用的方法）
先将pt权重文件转换为tensort格式，然后用tensor去调用

ps：当然，还有很多很多支持c++调用深度学习权重文件的，这里我只是列举了我个人比较喜欢用的几种调用方式。

一、环境配置

本篇blog使用是用onnx runtime去调用onnx权重文件，然后基于visual studio来配置运行环境。我们先配置visual studio的环境，这里我们主要要配置两个外部库，一个是opencv（用于图片的读取和写入），另外一个就是onnx runtime（用于调用权重文件）。网上有很多关于该部分的讲解，我找了两个写的还不错的直接分享给大家吧：

VisualStudio2019配置OpenCV4.1.0（opencv的版本可以随意选择，不过最好选择大于3.4.x以上的版本）
VS2019 快速配置Onnxruntime环境

二、转换权重文件

YOLO V7项目下载路径：YOLO V7
这里值得注意，一定一定一定要下载最新的项目，我第一次下载YOLO v7的时候作者还没有解决模型export.py中的bug，导出的onnx模型没法被调用。我重新下载了最新的代码，才跑通。
简单说下export.py的几个需要修改的参数：

parser = argparse.ArgumentParser()    parser.add_argument('--weights', type=str, default='', help='weights path') #YOLO V7训练得到的pt权重文件    parser.add_argument('--img-size', nargs='+', type=int, default=[640, 640], help='image size')  # 图片长宽，保持跟训练时候一致即可    parser.add_argument('--batch-size', type=int, default=1, help='batch size') #默认为1    parser.add_argument('--dynamic', action='store_true', help='dynamic ONNX axes')    parser.add_argument('--grid', action='store_true', help='export Detect() layer grid') #这个参数一定要加上去，如果不加的话默认导出onnx模型是不带最后一层Detect层的，最后结果是没办法解析出来的    parser.add_argument('--end2end', action='store_true', help='export end2end onnx')    parser.add_argument('--max-wh', type=int, default=None, help='None for tensorrt nms, int value for onnx-runtime nms')    parser.add_argument('--topk-all', type=int, default=100, help='topk objects for every images')    parser.add_argument('--iou-thres', type=float, default=0.45, help='iou threshold for NMS')     parser.add_argument('--conf-thres', type=float, default=0.25, help='conf threshold for NMS')    parser.add_argument('--device', default='cpu', help='cuda device, i.e. 0 or 0,1,2,3 or cpu')    parser.add_argument('--simplify', action='store_true', default=True, help='simplify onnx model') #在导出onnx模型的时候，是否做模型剪枝操作，建议加上，如果不加，opencv去调用onnx模型可能会出错    parser.add_argument('--include-nms', action='store_true', help='export end2end onnx')    opt = parser.parse_args()

最后，导出onnx模型，发现权重文件大小较原先pt文件减少了一倍。

三、onnx runtime调用onnx模型

代码链接：onnx runtime调用onnx模型

#include #include #include #include #include //#include #include using namespace std;using namespace cv;using namespace Ort;struct Net_config{ 	float confThreshold; // Confidence threshold	float nmsThreshold;  // Non-maximum suppression threshold	string modelpath;};typedef struct BoxInfo{ 	float x1;	float y1;	float x2;	float y2;	float score;	int label;} BoxInfo;class YOLOV7{ public:	YOLOV7(Net_config config);	void detect(Mat& frame);private:	int inpWidth;	int inpHeight;	int nout;	int num_proposal;	vectorclass_names;	int num_class;	float confThreshold;	float nmsThreshold;	vectorinput_image_;	void normalize_(Mat img);	void nms(vector& input_boxes);	Env env = Env(ORT_LOGGING_LEVEL_ERROR, "YOLOV7");	Ort::Session* ort_session = nullptr;	SessionOptions sessionOptions = SessionOptions();	vectorinput_names;	vectoroutput_names;	vector>input_node_dims; // >=1 outputs	vector>output_node_dims; // >=1 outputs};YOLOV7::YOLOV7(Net_config config){ 	this->confThreshold = config.confThreshold;	this->nmsThreshold = config.nmsThreshold;	string classesFile = ""; #coco.names路径	string model_path = config.modelpath;	std::wstring widestr = std::wstring(model_path.begin(), model_path.end());	//OrtStatus* status = OrtSessionOptionsAppendExecutionProvider_CUDA(sessionOptions, 0);	sessionOptions.SetGraphOptimizationLevel(ORT_ENABLE_BASIC);	ort_session = new Session(env, widestr.c_str(), sessionOptions);	size_t numInputNodes = ort_session->GetInputCount();	size_t numOutputNodes = ort_session->GetOutputCount();	AllocatorWithDefaultOptions allocator;	for (int i = 0; i < numInputNodes; i++)	{ 		input_names.push_back(ort_session->GetInputName(i, allocator));		Ort::TypeInfo input_type_info = ort_session->GetInputTypeInfo(i);		auto input_tensor_info = input_type_info.GetTensorTypeAndShapeInfo();		auto input_dims = input_tensor_info.GetShape();		input_node_dims.push_back(input_dims);	}	for (int i = 0; i < numOutputNodes; i++)	{ 		output_names.push_back(ort_session->GetOutputName(i, allocator));		Ort::TypeInfo output_type_info = ort_session->GetOutputTypeInfo(i);		auto output_tensor_info = output_type_info.GetTensorTypeAndShapeInfo();		auto output_dims = output_tensor_info.GetShape();		output_node_dims.push_back(output_dims);	}	this->inpHeight = input_node_dims[0][2];	this->inpWidth = input_node_dims[0][3];	this->nout = output_node_dims[0][2];	this->num_proposal = output_node_dims[0][1];	ifstream ifs(classesFile.c_str());	string line;	while (getline(ifs, line)) this->class_names.push_back(line);	this->num_class = class_names.size();}void YOLOV7::normalize_(Mat img){ 	//    img.convertTo(img, CV_32F);	int row = img.rows;	int col = img.cols;	this->input_image_.resize(row * col * img.channels());	for (int c = 0; c < 3; c++)	{ 		for (int i = 0; i < row; i++)		{ 			for (int j = 0; j < col; j++)			{ 				float pix = img.ptr(i)[j * 3 + 2 - c];				this->input_image_[c * row * col + i * col + j] = pix / 255.0;			}		}	}}void YOLOV7::nms(vector& input_boxes){ 	sort(input_boxes.begin(), input_boxes.end(), [](BoxInfo a, BoxInfo b) {  return a.score >b.score; });	vectorvArea(input_boxes.size());	for (int i = 0; i < int(input_boxes.size()); ++i)	{ 		vArea[i] = (input_boxes.at(i).x2 - input_boxes.at(i).x1 + 1)			* (input_boxes.at(i).y2 - input_boxes.at(i).y1 + 1);	}	vectorisSuppressed(input_boxes.size(), false);	for (int i = 0; i < int(input_boxes.size()); ++i)	{ 		if (isSuppressed[i]) {  continue; }		for (int j = i + 1; j < int(input_boxes.size()); ++j)		{ 			if (isSuppressed[j]) {  continue; }			float xx1 = (max)(input_boxes[i].x1, input_boxes[j].x1);			float yy1 = (max)(input_boxes[i].y1, input_boxes[j].y1);			float xx2 = (min)(input_boxes[i].x2, input_boxes[j].x2);			float yy2 = (min)(input_boxes[i].y2, input_boxes[j].y2);			float w = (max)(float(0), xx2 - xx1 + 1);			float h = (max)(float(0), yy2 - yy1 + 1);			float inter = w * h;			float ovr = inter / (vArea[i] + vArea[j] - inter);			if (ovr >= this->nmsThreshold)			{ 				isSuppressed[j] = true;			}		}	}	// return post_nms;	int idx_t = 0;	input_boxes.erase(remove_if(input_boxes.begin(), input_boxes.end(), [&idx_t, &isSuppressed](const BoxInfo& f) {  return isSuppressed[idx_t++]; }), input_boxes.end());}void YOLOV7::detect(Mat& frame){ 	Mat dstimg;	resize(frame, dstimg, Size(this->inpWidth, this->inpHeight));	this->normalize_(dstimg);	arrayinput_shape_{  1, 3, this->inpHeight, this->inpWidth };	auto allocator_info = MemoryInfo::CreateCpu(OrtDeviceAllocator, OrtMemTypeCPU);	Value input_tensor_ = Value::CreateTensor(allocator_info, input_image_.data(), input_image_.size(), input_shape_.data(), input_shape_.size());	// 开始推理	vectorort_outputs = ort_session->Run(RunOptions{  nullptr }, &input_names[0], &input_tensor_, 1, output_names.data(), output_names.size());   	/generate proposals	vectorgenerate_boxes;	float ratioh = (float)frame.rows / this->inpHeight, ratiow = (float)frame.cols / this->inpWidth;	int n = 0, k = 0; ///cx,cy,w,h,box_score, class_score	const float* pdata = ort_outputs[0].GetTensorMutableData();	for (n = 0; n < this->num_proposal; n++)   	{ 		float box_score = pdata[4];		if (box_score >this->confThreshold)		{ 			int max_ind = 0;			float max_class_socre = 0;			for (k = 0; k < num_class; k++)			{ 				if (pdata[k + 5] >max_class_socre)				{ 					max_class_socre = pdata[k + 5];					max_ind = k;				}			}			max_class_socre *= box_score;			if (max_class_socre >this->confThreshold)			{ 				float cx = pdata[0] * ratiow;  				float cy = pdata[1] * ratioh;   				float w = pdata[2] * ratiow;   				float h = pdata[3] * ratioh;  				float xmin = cx - 0.5 * w;				float ymin = cy - 0.5 * h;				float xmax = cx + 0.5 * w;				float ymax = cy + 0.5 * h;				generate_boxes.push_back(BoxInfo{  xmin, ymin, xmax, ymax, max_class_socre, max_ind });			}		}		pdata += nout;	}	// Perform non maximum suppression to eliminate redundant overlapping boxes with	// lower confidences	nms(generate_boxes);	for (size_t i = 0; i < generate_boxes.size(); ++i)	{ 		int xmin = int(generate_boxes[i].x1);		int ymin = int(generate_boxes[i].y1);		rectangle(frame, Point(xmin, ymin), Point(int(generate_boxes[i].x2), int(generate_boxes[i].y2)), Scalar(0, 0, 255), 2);		string label = format("%.2f", generate_boxes[i].score);		label = this->class_names[generate_boxes[i].label] + ":" + label;		putText(frame, label, Point(xmin, ymin - 5), FONT_HERSHEY_SIMPLEX, 0.75, Scalar(0, 255, 0), 1);	}}int main(){ 	Net_config YOLOV7_nets = {  0.3, 0.5, "E:/work/People_Detect/yolov7-main/models/yolov7_640x640.onnx" };   choices=["models/yolov7_640x640.onnx", "models/yolov7-tiny_640x640.onnx", "models/yolov7_736x1280.onnx", "models/yolov7-tiny_384x640.onnx", "models/yolov7_480x640.onnx", "models/yolov7_384x640.onnx", "models/yolov7-tiny_256x480.onnx", "models/yolov7-tiny_256x320.onnx", "models/yolov7_256x320.onnx", "models/yolov7-tiny_256x640.onnx", "models/yolov7_256x640.onnx", "models/yolov7-tiny_480x640.onnx", "models/yolov7-tiny_736x1280.onnx", "models/yolov7_256x480.onnx"] # onnx权重文件路径，这里只能使用上述这种命名方式，因为中间需要解析出模型的测试图片大小	YOLOV7 net(YOLOV7_nets);	string imgpath = ""; #测试图片路径	Mat srcimg = imread(imgpath);	net.detect(srcimg);	static const string kWinName = "Deep learning object detection in ONNXRuntime";	namedWindow(kWinName, WINDOW_NORMAL);	imshow(kWinName, srcimg);	waitKey(0);	destroyAllWindows();}