C++：onnxruntime调用FasterRCNN模型

1
2
3
4
cv2.error: OpenCV(4.5.4) D:aopencv-pythonopencv-pythonopencvmodulesdnnsrconnxonnx_graph_simplifier.cpp:692:
error: (-210:Unsupported format or combination of formats)
Unsupported data type: BOOL in function 'cv::dnn::dnn4_v20211004::getMatFromTensor'

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
//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:
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);
typedef struct Output {
int ClsId;//
float confidence;//
cv::Rect box;//
Output() :ClsId(), confidence(), box() {}
Output(int a, float b, cv::Rect c) {
this->ClsId = a;
this->confidence = b;
this->box=c;
}
}Output ;
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);
};

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
//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)
{
int left, top, baseLine, color_num;
for (int i = 0; i < result.size(); i++) {
left = result[i].box.x;
top = result[i].box.y;
color_num = i;
rectangle(img, result[i].box, cv::Scalar(0,0,255), 2, 8);
std::string label = std::to_string(result[i].ClsId) + ":" + std::to_string(result[i].confidence);
cv::Size labelSize = cv::getTextSize(label, cv::FONT_HERSHEY_SIMPLEX, 0.5, 1, &baseLine);
top = std::max(top, labelSize.height);
putText(img, label, cv::Point(left, top), cv::FONT_HERSHEY_SIMPLEX, 1, cv::Scalar(0,0,255), 2);
}
}
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