背景:
最近由于项目原因,需要用C++做一些目标检测的任务,就捣鼓一下YOLOv5,发现部署确实很方便,将YOLOv5模型转为onnx模型后,可以用OpenCV的dnn.readNetFromONNX读取该模型,接着就是输入预处理和输出结果解析的事情。
然而,当我将tf15训练得到的FasterRCNN模型并利用tf2onnx成功转为onnx模型后,却不能用OpenCV读取,报出以下错误,而onnxruntime可以成功调用该模型。
cv2.error: OpenCV(4.5.4) D:\a\opencv-python\opencv-python\opencv\modules\dnn\src\onnx\onnx_graph_simplifier.cpp:692:
error: (-210:Unsupported format or combination of formats)
Unsupported data type: BOOL in function 'cv::dnn::dnn4_v20211004::getMatFromTensor'
大概意思可能是:不支持的数据类型从而导致不支持该操作吧
程序:
所以,只好采用onnxruntime的C++接口进行模型调用,废话不多说,直接上代码:
//FRCNN.h
#pragma once
#include<iostream>
#include<fstream>
#include<numeric>
#include<opencv.hpp>
#include"../commonStruct.h"
#include"../BaseShipDetectionModel.h"
#include <onnxruntime_cxx_api.h>
#
class FRCNN:public BaseShipDetectionModel
{
public:
FRCNN();
~FRCNN();
bool readModel(std::string &netPath, bool isCuda=false);
bool DetectShip(cv::Mat &SrcImg, std::vector<Output> &output);
void drawPredShip(cv::Mat &img, std::vector<Output>& result);
private:
enum OutputFlag
{
//NOTHING,
BOXES,
SCORES,
CLSIDS
};
Ort::Env *OnnxEnv;
Ort::SessionOptions OnnxSessionOp;
Ort::Session* OnnxSession;
Ort::AllocatorWithDefaultOptions allocator;
Ort::MemoryInfo *memory_info;
cv::Size2f factor;
const int netWidth = 1067;
const int netHeight = 600;
float nmsThreshold = 0.45;
float boxThreshold = 0.31;
float classThreshold = 0.25;
size_t num_input_nodes, num_output_nodes;
std::vector<const char*> input_node_names, output_node_names;
std::vector<OutputFlag> output_node_namesFlag;
//Ort::Value *input_tensor;
std::vector<int64_t> input_node_dims = { netHeight, netWidth,3 };
size_t input_tensor_size = 3 * netHeight * netWidth;
void parseOnnxOutput(std::vector<Ort::Value>&inputTensors, std::vector<Output> &results);
};
//FRCNN.cpp
#include "FRCNN.h"
using namespace std;
using namespace cv;
using namespace dnn;
#if 1
FRCNN::FRCNN()
{
num_input_nodes = 0;
num_output_nodes = 0;
//Ort::Env env(ORT_LOGGING_LEVEL_VERBOSE, "test");
OnnxEnv = new Ort::Env(ORT_LOGGING_LEVEL_WARNING, "test");
//Ort::MemoryInfo memory_info = Ort::MemoryInfo::CreateCpu(OrtArenaAllocator, OrtMemTypeDefault);
memory_info=new Ort::MemoryInfo(Ort::MemoryInfo::CreateCpu(OrtArenaAllocator, OrtMemTypeDefault));
//memory_info =new Ort::MemoryInfo::CreateCpu(OrtArenaAllocator, OrtMemTypeDefault);
if (OnnxEnv == nullptr) {
std::cout << "new Error for " << VNAME(OnnxEnv) << std::endl;
throw OnnxEnv;
}
if (memory_info == nullptr) {
std::cout << "new Error for " << VNAME(memory_info) << std::endl;
throw memory_info;
}
OnnxSessionOp.SetIntraOpNumThreads(5);
OnnxSessionOp.SetGraphOptimizationLevel(GraphOptimizationLevel::ORT_ENABLE_ALL);
flag_onnx = true;
/*cout << OnnxEnv << '\t' << *OnnxEnv << endl;
cout << &env << '\t' << env << endl;*/
//test(*OnnxEnv, env, OnnxEnv);
//std::cout<<env.operator const OrtEnv *
}
FRCNN::~FRCNN()
{
if(OnnxEnv!=nullptr)
delete OnnxEnv;
if (OnnxSession != nullptr)
delete OnnxSession;
}
bool FRCNN::readModel(string &netPath, bool isCuda) {
try
{
std::ifstream f(netPath.c_str());
std::cout << f.good() << std::endl;
std::wstring wnetPath = std::wstring(netPath.begin(), netPath.end());
OnnxSession=new Ort::Session((*OnnxEnv), wnetPath.c_str(), OnnxSessionOp);
if (OnnxSession == nullptr) {
std::cout << "new Error for " << VNAME(OnnxSession) << std::endl;
throw OnnxSession;
}
// print model input layer (node names, types, shape etc.)
// print number of model input nodes
num_input_nodes = OnnxSession->GetInputCount();
num_output_nodes = OnnxSession->GetOutputCount();
for (int i = 0; i < num_input_nodes; ++i) {
input_node_names.push_back(OnnxSession->GetInputName(0, allocator));
//= { "image:0"};
std::cout << input_node_names[i] << std::endl;
}
for (int i = 0; i < num_output_nodes; ++i) {
char* name = OnnxSession->GetOutputName(i, allocator);
output_node_names.push_back(name);
if (strstr(name, "boxes") != nullptr) {
output_node_namesFlag.push_back(BOXES);
}
else if (strstr(name, "scores") != nullptr) {
output_node_namesFlag.push_back(SCORES);
}
else if (strstr(name, "labels") != nullptr) {
output_node_namesFlag.push_back(CLSIDS);
}
else {
//output_node_namesFlag.push_back(NOTHING);
throw(name);
}
//= { "output/boxes:0", "output/scores:0","output/labels:0"};
std::cout << output_node_names[i] << std::endl;
}
}
catch (const std::exception& e) {
return false;
}
return true;
}
bool FRCNN::DetectShip(cv::Mat &SrcImg, std::vector<Output> &results) {
if (SrcImg.empty()) {
std::cout << "empty image error!" << std::endl;
return false;
}
int col = SrcImg.cols;
int row = SrcImg.rows;
int i, j;
results.clear();
Mat netInputImg,Img;
std::vector<int> indices;
Output result;
//netInputImg.create(SrcImg.size,SrcImg.depth());
//SrcImg.copyTo(netInputImg);
cv::resize(SrcImg, Img, cv::Size(netWidth, netHeight), 0.0, 0.0, cv::INTER_LINEAR);
factor= cv::Size2f((float)SrcImg.cols / netWidth, (float)SrcImg.rows / netHeight);
try {
netInputImg.create(cv::Size(netWidth, netHeight), CV_32FC3);//allocate the continuous Mat
Img.convertTo(netInputImg, CV_32F);
assert(netInputImg.isContinuous());//
Ort::Value input_tensor = Ort::Value::CreateTensor<float>(*memory_info, (float*)netInputImg.data, input_tensor_size, input_node_dims.data(), 3);
assert(input_tensor.IsTensor());
std::vector<Ort::Value> ort_inputs;
ort_inputs.push_back(std::move(input_tensor));
//Run the Detection
std::vector<Ort::Value> output_tensors = OnnxSession->Run(Ort::RunOptions{ nullptr }, input_node_names.data(), ort_inputs.data(), ort_inputs.size(), output_node_names.data(), 3);
parseOnnxOutput(output_tensors, results);
}
catch (...) {
std::cout << "prediction error!" << std::endl;
return false;
}
if (results.size())
return true;
else
return false;
}
void FRCNN::drawPredShip(cv::Mat & img, std::vector<Output>& result)
{
BaseShipDetectionModel::drawPredShip(img, result);
}
void FRCNN::parseOnnxOutput(std::vector<Ort::Value>& inputTensors, std::vector<Output>& results)
{
std::vector<int64_t> classIds;
std::vector<float> confidences;
std::vector<cv::Rect> boxes;
int i, j;
std::vector<int64_t> shape;
size_t eleCount;
size_t DimCount;
int xmin, xmax, ymin, ymax;
for (i = 0; i < num_output_nodes; ++i) {
Ort::TensorTypeAndShapeInfo Info = inputTensors[i].GetTensorTypeAndShapeInfo();
//std::cout << ":GetDimensionsCount:" << Info.GetDimensionsCount() << '\t';
shape = Info.GetShape();
DimCount = shape.size();
//std::cout << i << "shape:";
//for (int j = 0; j < shape.size(); ++j) {
// std::cout << shape[j] << '\t';
//}
eleCount = Info.GetElementCount();
//std::cout << ":GetElementCount:" << eleCount << '\t';
ONNXTensorElementDataType onnxType = Info.GetElementType();
void* ptr = nullptr;
//std::cout << "GetElementType:" << onnxType << '\t' << std::endl;
switch (onnxType) {
case ONNX_TENSOR_ELEMENT_DATA_TYPE_FLOAT:// maps to c type float
{
ptr = inputTensors[i].GetTensorMutableData<float>();
}
break;
case ONNX_TENSOR_ELEMENT_DATA_TYPE_INT64:// maps to c type int64_t
{
ptr = inputTensors[i].GetTensorMutableData<int64_t>();
}
break;
default:
throw("Unknown DataType!");
break;
}
//std::cout << sizeof(ptr) << sizeof((float*)ptr) << sizeof((int64_t*)ptr)<<sizeof((uint8_t*)ptr)<< sizeof((uint16_t*)ptr)<< sizeof((int32_t*)ptr) << std::endl;
//= { "output/boxes:0", "output/scores:0","output/labels:0"};
/*
output[0]//(44->(11,4))
output[1]//(11->(11))
output[2]//(11->(11))
*/
switch (output_node_namesFlag[i]) {
case BOXES://xmin,ymin,xmax,ymax
{
float* p_boxes = (float*)ptr;
for (j = 0; j < eleCount; j += 4) {
xmin = p_boxes[j] *factor.width;
ymin = p_boxes[j + 1] * factor.height;
xmax = p_boxes[j + 2] *factor.width;
ymax = p_boxes[j + 3] *factor.height;
boxes.push_back(cv::Rect(xmin, ymin, xmax - xmin, ymax - ymin));
}
break;
}
case SCORES:
{
float* p_scores = (float*)ptr;
for (j = 0; j < eleCount; j++) {
confidences.push_back(p_scores[j]);
}
break;
}
case CLSIDS:
{
int64_t* p_clsids = (int64_t*)ptr;
for (j = 0; j < eleCount; j++) {
classIds.push_back(p_clsids[j]);
}
break;
}
}
}
Output result;
assert((boxes.size() == classIds.size())&&(boxes.size()==confidences.size()));
for (i = 0; i < boxes.size(); ++i) {
//j = indices[i];
if (confidences[i] > boxThreshold) {
result.ClsId = classIds[i]-1;//except background
result.confidence = confidences[i];
result.box = boxes[i];
results.push_back(result);
}
}
}
#endif
代码调用顺序:
1.readModel//读取模型
2.DetectShip//检测目标,函数名根据需要修改
3.drawPredShip//画图,函数名根据需要修改
该代码对应FasterRCNN模型下载链接: fasterRCNN.model-深度学习文档类资源-CSDN文库
后记:
本文仅为onnxruntime的C++调用作个笔记,特别是对输入数据准备与输出数据解析这两部分,如有疑问,请不吝指教!