608 lines
21 KiB
C++
608 lines
21 KiB
C++
|
#include "ANSOPENVINOOD.h"
|
||
|
|
#include "Utility.h"
|
||
|
|
namespace ANSCENTER
|
||
|
|
{
|
||
|
|
bool OPENVINOOD::OptimizeModel(bool fp16, std::string& optimizedModelFolder) {
|
||
|
|
std::lock_guard<std::recursive_mutex> lock(_mutex);
|
||
|
|
if (!ANSODBase::OptimizeModel(fp16, optimizedModelFolder)) {
|
||
|
|
return false;
|
||
|
|
}
|
||
|
|
if (FileExist(_modelFilePath)) {
|
||
|
|
std::string modelName = GetFileNameWithoutExtension(_modelFilePath);
|
||
|
|
std::string binaryModelName = modelName + ".bin";
|
||
|
|
std::string modelFolder = GetParentFolder(_modelFilePath);
|
||
|
|
std::string optimizedModelPath = CreateFilePath(modelFolder, binaryModelName);
|
||
|
|
if (FileExist(optimizedModelPath)) {
|
||
|
|
this->_logger.LogDebug("OPENVINOOD::OptimizeModel", "This model is optimized. No need other optimization.", __FILE__, __LINE__);
|
||
|
|
optimizedModelFolder = modelFolder;
|
||
|
|
return true;
|
||
|
|
}
|
||
|
|
else {
|
||
|
|
this->_logger.LogFatal("OPENVINOOD::OptimizeModel", "This model can not be optimized.", __FILE__, __LINE__);
|
||
|
|
optimizedModelFolder = modelFolder;
|
||
|
|
return false;
|
||
|
|
}
|
||
|
|
}
|
||
|
|
else {
|
||
|
|
this->_logger.LogFatal("OPENVINOOD::OptimizeModel", "This model is not exist. Please check the model path again.", __FILE__, __LINE__);
|
||
|
|
optimizedModelFolder = "";
|
||
|
|
return false;
|
||
|
|
}
|
||
|
|
}
|
||
|
|
bool OPENVINOOD::LoadModel(const std::string& modelZipFilePath, const std::string& modelZipPassword) {
|
||
|
|
std::lock_guard<std::recursive_mutex> lock(_mutex);
|
||
|
|
try {
|
||
|
|
bool result = ANSODBase::LoadModel(modelZipFilePath, modelZipPassword);
|
||
|
|
if (!result) return false;
|
||
|
|
// 0. Check if the configuration file exist
|
||
|
|
if (FileExist(_modelConfigFile)) {
|
||
|
|
ModelType modelType;
|
||
|
|
std::vector<int> inputShape;
|
||
|
|
_classes = ANSUtilityHelper::GetConfigFileContent(_modelConfigFile, modelType, inputShape);
|
||
|
|
if (inputShape.size() == 2) {
|
||
|
|
if (inputShape[0] > 0)_modelConfig.inpHeight = inputShape[0];
|
||
|
|
if (inputShape[1] > 0)_modelConfig.inpWidth = inputShape[1];
|
||
|
|
}
|
||
|
|
}
|
||
|
|
else {// This is old version of model zip file
|
||
|
|
std::string onnxfile = CreateFilePath(_modelFolder, "train_last.xml");//yolov8n.xml
|
||
|
|
if (std::filesystem::exists(onnxfile)) {
|
||
|
|
_modelFilePath = onnxfile;
|
||
|
|
_classFilePath = CreateFilePath(_modelFolder, "classes.names");
|
||
|
|
this->_logger.LogDebug("OPENVINOOD::Initialize. Loading OpenVINO weight", _modelFilePath, __FILE__, __LINE__);
|
||
|
|
}
|
||
|
|
else {
|
||
|
|
this->_logger.LogError("OPENVINOOD::Initialize. Model file is not exist", _modelFilePath, __FILE__, __LINE__);
|
||
|
|
return false;
|
||
|
|
}
|
||
|
|
std::ifstream isValidFileName(_classFilePath);
|
||
|
|
if (!isValidFileName)
|
||
|
|
{
|
||
|
|
this->_logger.LogDebug("OPENVINOOD::Initialize. Load classes from string", _classFilePath, __FILE__, __LINE__);
|
||
|
|
LoadClassesFromString();
|
||
|
|
}
|
||
|
|
else {
|
||
|
|
this->_logger.LogDebug("OPENVINOOD::Initialize. Load classes from file", _classFilePath, __FILE__, __LINE__);
|
||
|
|
LoadClassesFromFile();
|
||
|
|
}
|
||
|
|
}
|
||
|
|
|
||
|
|
// Load Model from Here
|
||
|
|
InitialModel();
|
||
|
|
_isInitialized = true;
|
||
|
|
return true;
|
||
|
|
}
|
||
|
|
catch (std::exception& e) {
|
||
|
|
this->_logger.LogFatal("OPENVINOOD::LoadModel", e.what(), __FILE__, __LINE__);
|
||
|
|
return false;
|
||
|
|
}
|
||
|
|
}
|
||
|
|
bool OPENVINOOD::LoadModelFromFolder(std::string licenseKey, ModelConfig modelConfig, std::string modelName, std::string className, const std::string& modelFolder, std::string& labelMap) {
|
||
|
|
std::lock_guard<std::recursive_mutex> lock(_mutex);
|
||
|
|
try {
|
||
|
|
bool result = ANSODBase::LoadModelFromFolder(licenseKey, modelConfig, modelName, className, modelFolder, labelMap);
|
||
|
|
if (!result) return false;
|
||
|
|
std::string _modelName = modelName;
|
||
|
|
if (_modelName.empty()) {
|
||
|
|
_modelName = "train_last";
|
||
|
|
}
|
||
|
|
std::string modelFullName = _modelName + ".xml";
|
||
|
|
_modelConfig = modelConfig;
|
||
|
|
_modelConfig.detectionType = ANSCENTER::DetectionType::DETECTION;
|
||
|
|
_modelConfig.modelType = ModelType::OPENVINO;
|
||
|
|
_modelConfig.inpHeight = 640;
|
||
|
|
_modelConfig.inpWidth = 640;
|
||
|
|
|
||
|
|
if (_modelConfig.modelMNSThreshold < 0.2)
|
||
|
|
_modelConfig.modelMNSThreshold = 0.5;
|
||
|
|
if (_modelConfig.modelConfThreshold < 0.2)
|
||
|
|
_modelConfig.modelConfThreshold = 0.5;
|
||
|
|
// 0. Check if the configuration file exist
|
||
|
|
if (FileExist(_modelConfigFile)) {
|
||
|
|
ModelType modelType;
|
||
|
|
std::vector<int> inputShape;
|
||
|
|
_classes = ANSUtilityHelper::GetConfigFileContent(_modelConfigFile, modelType, inputShape);
|
||
|
|
if (inputShape.size() == 2) {
|
||
|
|
if (inputShape[0] > 0)_modelConfig.inpHeight = inputShape[0];
|
||
|
|
if (inputShape[1] > 0)_modelConfig.inpWidth = inputShape[1];
|
||
|
|
}
|
||
|
|
}
|
||
|
|
else {// This is old version of model zip file
|
||
|
|
std::string onnxfile = CreateFilePath(_modelFolder, modelFullName);//yolov8n.xml
|
||
|
|
if (std::filesystem::exists(onnxfile)) {
|
||
|
|
_modelFilePath = onnxfile;
|
||
|
|
_classFilePath = CreateFilePath(_modelFolder, className);
|
||
|
|
this->_logger.LogDebug("OPENVINOOD::Initialize. Loading OpenVINO weight", _modelFilePath, __FILE__, __LINE__);
|
||
|
|
}
|
||
|
|
else {
|
||
|
|
this->_logger.LogError("OPENVINOOD::Initialize. Model file is not exist", _modelFilePath, __FILE__, __LINE__);
|
||
|
|
return false;
|
||
|
|
}
|
||
|
|
std::ifstream isValidFileName(_classFilePath);
|
||
|
|
if (!isValidFileName)
|
||
|
|
{
|
||
|
|
this->_logger.LogDebug("OPENVINOOD::Initialize. Load classes from string", _classFilePath, __FILE__, __LINE__);
|
||
|
|
LoadClassesFromString();
|
||
|
|
}
|
||
|
|
else {
|
||
|
|
this->_logger.LogDebug("OPENVINOOD::Initialize. Load classes from file", _classFilePath, __FILE__, __LINE__);
|
||
|
|
LoadClassesFromFile();
|
||
|
|
}
|
||
|
|
}
|
||
|
|
|
||
|
|
// 1. Load labelMap and engine
|
||
|
|
labelMap.clear();
|
||
|
|
if (!_classes.empty())
|
||
|
|
labelMap = VectorToCommaSeparatedString(_classes);
|
||
|
|
|
||
|
|
// Load Model from Here
|
||
|
|
InitialModel();
|
||
|
|
_isInitialized = true;
|
||
|
|
return true;
|
||
|
|
}
|
||
|
|
catch (std::exception& e) {
|
||
|
|
this->_logger.LogFatal("OPENVINOOD::LoadModel", e.what(), __FILE__, __LINE__);
|
||
|
|
return false;
|
||
|
|
}
|
||
|
|
}
|
||
|
|
cv::Mat OPENVINOOD::PreProcessing(const cv::Mat& source) {
|
||
|
|
try {
|
||
|
|
if (source.empty()) {
|
||
|
|
this->_logger.LogFatal("OPENVINOOD::PreProcessing", "Empty image provided", __FILE__, __LINE__);
|
||
|
|
return cv::Mat();
|
||
|
|
}
|
||
|
|
|
||
|
|
// Convert grayscale to 3-channel BGR if needed
|
||
|
|
cv::Mat processedImage;
|
||
|
|
if (source.channels() == 1) {
|
||
|
|
cv::cvtColor(source, processedImage, cv::COLOR_GRAY2BGR);
|
||
|
|
}
|
||
|
|
else {
|
||
|
|
processedImage = source;
|
||
|
|
}
|
||
|
|
|
||
|
|
int col = processedImage.cols;
|
||
|
|
int row = processedImage.rows;
|
||
|
|
int maxSize = std::max(col, row);
|
||
|
|
|
||
|
|
// Create a square padded image with a black background
|
||
|
|
cv::Mat result = cv::Mat::zeros(maxSize, maxSize, CV_8UC3);
|
||
|
|
|
||
|
|
// Copy the original image to the top-left corner of the square matrix
|
||
|
|
processedImage.copyTo(result(cv::Rect(0, 0, col, row)));
|
||
|
|
|
||
|
|
return result;
|
||
|
|
}
|
||
|
|
catch (const std::exception& e) {
|
||
|
|
this->_logger.LogFatal("OPENVINOOD::PreProcessing", e.what(), __FILE__, __LINE__);
|
||
|
|
return cv::Mat();
|
||
|
|
}
|
||
|
|
}
|
||
|
|
bool OPENVINOOD::Initialize(std::string licenseKey, ModelConfig modelConfig, const std::string& modelZipFilePath, const std::string& modelZipPassword, std::string& labelMap) {
|
||
|
|
std::lock_guard<std::recursive_mutex> lock(_mutex);
|
||
|
|
try {
|
||
|
|
bool result = ANSODBase::Initialize(licenseKey, modelConfig, modelZipFilePath, modelZipPassword, labelMap);
|
||
|
|
if (!result) return false;
|
||
|
|
// Parsing for YOLO only here
|
||
|
|
_modelConfig = modelConfig;
|
||
|
|
_modelConfig.detectionType = ANSCENTER::DetectionType::DETECTION;
|
||
|
|
_modelConfig.modelType = ModelType::OPENVINO;
|
||
|
|
_modelConfig.inpHeight = 640;
|
||
|
|
_modelConfig.inpWidth = 640;
|
||
|
|
if (_modelConfig.modelMNSThreshold < 0.2)
|
||
|
|
_modelConfig.modelMNSThreshold = 0.5;
|
||
|
|
if (_modelConfig.modelConfThreshold < 0.2)
|
||
|
|
_modelConfig.modelConfThreshold = 0.5;
|
||
|
|
model_input_shape_ = cv::Size2f(_modelConfig.inpWidth, _modelConfig.inpHeight);
|
||
|
|
// 0. Check if the configuration file exist
|
||
|
|
if (FileExist(_modelConfigFile)) {
|
||
|
|
ModelType modelType;
|
||
|
|
std::vector<int> inputShape;
|
||
|
|
_classes = ANSUtilityHelper::GetConfigFileContent(_modelConfigFile, modelType, inputShape);
|
||
|
|
if (inputShape.size() == 2) {
|
||
|
|
if (inputShape[0] > 0)_modelConfig.inpHeight = inputShape[0];
|
||
|
|
if (inputShape[1] > 0)_modelConfig.inpWidth = inputShape[1];
|
||
|
|
}
|
||
|
|
}
|
||
|
|
else {// This is old version of model zip file
|
||
|
|
std::string onnxfile = CreateFilePath(_modelFolder, "train_last.xml");//yolov8n.xml
|
||
|
|
if (std::filesystem::exists(onnxfile)) {
|
||
|
|
_modelFilePath = onnxfile;
|
||
|
|
_classFilePath = CreateFilePath(_modelFolder, "classes.names");
|
||
|
|
this->_logger.LogDebug("OPENVINOOD::Initialize. Loading OpenVINO weight", _modelFilePath, __FILE__, __LINE__);
|
||
|
|
}
|
||
|
|
else {
|
||
|
|
this->_logger.LogError("OPENVINOOD::Initialize. Model file is not exist", _modelFilePath, __FILE__, __LINE__);
|
||
|
|
return false;
|
||
|
|
}
|
||
|
|
std::ifstream isValidFileName(_classFilePath);
|
||
|
|
if (!isValidFileName)
|
||
|
|
{
|
||
|
|
this->_logger.LogDebug("OPENVINOOD::Initialize. Load classes from string", _classFilePath, __FILE__, __LINE__);
|
||
|
|
LoadClassesFromString();
|
||
|
|
}
|
||
|
|
else {
|
||
|
|
this->_logger.LogDebug("OPENVINOOD::Initialize. Load classes from file", _classFilePath, __FILE__, __LINE__);
|
||
|
|
LoadClassesFromFile();
|
||
|
|
}
|
||
|
|
}
|
||
|
|
|
||
|
|
// 1. Load labelMap and engine
|
||
|
|
labelMap.clear();
|
||
|
|
if (!_classes.empty())
|
||
|
|
labelMap = VectorToCommaSeparatedString(_classes);
|
||
|
|
|
||
|
|
// Load Model from Here
|
||
|
|
InitialModel();
|
||
|
|
_isInitialized = true;
|
||
|
|
return true;
|
||
|
|
|
||
|
|
}
|
||
|
|
catch (std::exception& e) {
|
||
|
|
this->_logger.LogFatal("OPENVINOOD::Initialize", e.what(), __FILE__, __LINE__);
|
||
|
|
return false;
|
||
|
|
}
|
||
|
|
}
|
||
|
|
std::vector<Object> OPENVINOOD::RunInference(const cv::Mat& input) {
|
||
|
|
return RunInference(input, "CustomCam");
|
||
|
|
}
|
||
|
|
std::vector<Object> OPENVINOOD::RunInference(const cv::Mat& input, const std::string& camera_id)
|
||
|
|
{
|
||
|
|
std::lock_guard<std::recursive_mutex> lock(_mutex);
|
||
|
|
|
||
|
|
// Early validation
|
||
|
|
if (!_licenseValid) {
|
||
|
|
_logger.LogError("OPENVINOOD::RunInference", "Invalid License",
|
||
|
|
__FILE__, __LINE__);
|
||
|
|
return {};
|
||
|
|
}
|
||
|
|
|
||
|
|
if (!_isInitialized) {
|
||
|
|
_logger.LogError("OPENVINOOD::RunInference", "Model is not initialized",
|
||
|
|
__FILE__, __LINE__);
|
||
|
|
return {};
|
||
|
|
}
|
||
|
|
|
||
|
|
if (input.empty() || input.cols < 10 || input.rows < 10) {
|
||
|
|
_logger.LogError("OPENVINOOD::RunInference", "Invalid input image",
|
||
|
|
__FILE__, __LINE__);
|
||
|
|
return {};
|
||
|
|
}
|
||
|
|
|
||
|
|
try {
|
||
|
|
// Step 1: Preprocessing
|
||
|
|
cv::Mat letterbox_img = PreProcessing(input);
|
||
|
|
if (letterbox_img.empty()) {
|
||
|
|
_logger.LogError("OPENVINOOD::RunInference", "PreProcessing failed",
|
||
|
|
__FILE__, __LINE__);
|
||
|
|
return {};
|
||
|
|
}
|
||
|
|
|
||
|
|
// Step 2: Create blob
|
||
|
|
constexpr int imageSize = 640;
|
||
|
|
const float scale = static_cast<float>(letterbox_img.rows) / imageSize;
|
||
|
|
|
||
|
|
cv::Mat blob = cv::dnn::blobFromImage(letterbox_img, 1.0 / 255.0,
|
||
|
|
cv::Size(imageSize, imageSize),
|
||
|
|
cv::Scalar(), true);
|
||
|
|
|
||
|
|
// Step 3: Run inference
|
||
|
|
auto input_port = compiled_model_.input();
|
||
|
|
ov::Tensor input_tensor(input_port.get_element_type(),
|
||
|
|
input_port.get_shape(), blob.ptr(0));
|
||
|
|
inference_request_.set_input_tensor(input_tensor);
|
||
|
|
inference_request_.infer();
|
||
|
|
|
||
|
|
// Step 4: Get output
|
||
|
|
auto output = inference_request_.get_output_tensor(0);
|
||
|
|
auto output_shape = output.get_shape();
|
||
|
|
|
||
|
|
if (output_shape.size() != 3) {
|
||
|
|
_logger.LogError("OPENVINOOD::RunInference", "Unexpected output shape",
|
||
|
|
__FILE__, __LINE__);
|
||
|
|
return {};
|
||
|
|
}
|
||
|
|
|
||
|
|
const int rows = static_cast<int>(output_shape[2]); // 8400
|
||
|
|
const int dimensions = static_cast<int>(output_shape[1]); // 84
|
||
|
|
float* data = output.data<float>();
|
||
|
|
|
||
|
|
// Step 5: Parse detections (avoid transpose!)
|
||
|
|
std::vector<int> class_ids;
|
||
|
|
std::vector<float> class_scores;
|
||
|
|
std::vector<cv::Rect> boxes;
|
||
|
|
|
||
|
|
// Pre-allocate for efficiency
|
||
|
|
class_ids.reserve(rows / 10); // Estimate ~10% pass threshold
|
||
|
|
class_scores.reserve(rows / 10);
|
||
|
|
boxes.reserve(rows / 10);
|
||
|
|
|
||
|
|
const cv::Rect imageBounds(0, 0, input.cols, input.rows);
|
||
|
|
|
||
|
|
// Process detections without transpose
|
||
|
|
for (int i = 0; i < rows; i++) {
|
||
|
|
// Get class scores starting at index 4
|
||
|
|
float max_score = -1.0f;
|
||
|
|
int max_class_id = 0;
|
||
|
|
|
||
|
|
for (int j = 4; j < dimensions; j++) {
|
||
|
|
const float score = data[j * rows + i]; // Column-major access
|
||
|
|
if (score > max_score) {
|
||
|
|
max_score = score;
|
||
|
|
max_class_id = j - 4;
|
||
|
|
}
|
||
|
|
}
|
||
|
|
|
||
|
|
if (max_score > _modelConfig.detectionScoreThreshold) {
|
||
|
|
// Extract box coordinates
|
||
|
|
const float cx = data[0 * rows + i];
|
||
|
|
const float cy = data[1 * rows + i];
|
||
|
|
const float w = data[2 * rows + i];
|
||
|
|
const float h = data[3 * rows + i];
|
||
|
|
|
||
|
|
// Convert to pixel coordinates and clamp
|
||
|
|
cv::Rect box(
|
||
|
|
static_cast<int>((cx - 0.5f * w) * scale),
|
||
|
|
static_cast<int>((cy - 0.5f * h) * scale),
|
||
|
|
static_cast<int>(w * scale),
|
||
|
|
static_cast<int>(h * scale)
|
||
|
|
);
|
||
|
|
|
||
|
|
box &= imageBounds; // Clamp to image bounds
|
||
|
|
|
||
|
|
if (box.area() > 0) {
|
||
|
|
class_ids.push_back(max_class_id);
|
||
|
|
class_scores.push_back(max_score);
|
||
|
|
boxes.push_back(box);
|
||
|
|
}
|
||
|
|
}
|
||
|
|
}
|
||
|
|
|
||
|
|
// Step 6: Apply NMS
|
||
|
|
if (boxes.empty()) {
|
||
|
|
return {};
|
||
|
|
}
|
||
|
|
|
||
|
|
std::vector<int> indices;
|
||
|
|
cv::dnn::NMSBoxes(boxes, class_scores,
|
||
|
|
_modelConfig.detectionScoreThreshold,
|
||
|
|
_modelConfig.modelMNSThreshold, indices);
|
||
|
|
|
||
|
|
// Step 7: Build output objects
|
||
|
|
std::vector<Object> outputs;
|
||
|
|
outputs.reserve(indices.size());
|
||
|
|
|
||
|
|
const int classNameSize = static_cast<int>(_classes.size());
|
||
|
|
|
||
|
|
for (int idx : indices) {
|
||
|
|
Object result;
|
||
|
|
result.classId = class_ids[idx];
|
||
|
|
result.confidence = class_scores[idx];
|
||
|
|
result.box = boxes[idx];
|
||
|
|
result.cameraId = camera_id;
|
||
|
|
|
||
|
|
// Set class name
|
||
|
|
if (!_classes.empty()) {
|
||
|
|
result.className = (result.classId < classNameSize)
|
||
|
|
? _classes[result.classId]
|
||
|
|
: _classes.back();
|
||
|
|
}
|
||
|
|
else {
|
||
|
|
result.className = "Unknown";
|
||
|
|
}
|
||
|
|
|
||
|
|
// Set polygon
|
||
|
|
result.polygon = ANSUtilityHelper::RectToNormalizedPolygon(
|
||
|
|
result.box, input.cols, input.rows
|
||
|
|
);
|
||
|
|
|
||
|
|
outputs.push_back(std::move(result));
|
||
|
|
}
|
||
|
|
|
||
|
|
if (_trackerEnabled) {
|
||
|
|
outputs = ApplyTracking(outputs, camera_id);
|
||
|
|
if (_stabilizationEnabled) outputs = StabilizeDetections(outputs, camera_id);
|
||
|
|
}
|
||
|
|
return outputs;
|
||
|
|
|
||
|
|
}
|
||
|
|
catch (const std::exception& e) {
|
||
|
|
_logger.LogFatal("OPENVINOOD::RunInference", e.what(), __FILE__, __LINE__);
|
||
|
|
return {};
|
||
|
|
}
|
||
|
|
catch (...) {
|
||
|
|
_logger.LogFatal("OPENVINOOD::RunInference", "Unknown error",
|
||
|
|
__FILE__, __LINE__);
|
||
|
|
return {};
|
||
|
|
}
|
||
|
|
}
|
||
|
|
OPENVINOOD::~OPENVINOOD() {
|
||
|
|
try {
|
||
|
|
if (FolderExist(_modelFolder)) {
|
||
|
|
if (!DeleteFolder(_modelFolder)) {
|
||
|
|
this->_logger.LogError("OPENVINOOD::~OPENVINOOD", "Failed to delete OpenVINO Models", __FILE__, __LINE__);
|
||
|
|
}
|
||
|
|
}
|
||
|
|
}
|
||
|
|
catch (std::exception& e) {
|
||
|
|
this->_logger.LogError("OPENVINOOD::~OPENVINOOD()", "Failed to release OPENVINO Models", __FILE__, __LINE__);
|
||
|
|
}
|
||
|
|
}
|
||
|
|
bool OPENVINOOD::Destroy() {
|
||
|
|
try {
|
||
|
|
if (FolderExist(_modelFolder)) {
|
||
|
|
DeleteFolder(_modelFolder);
|
||
|
|
}
|
||
|
|
return true;
|
||
|
|
}
|
||
|
|
catch (std::exception& e) {
|
||
|
|
this->_logger.LogError("OPENVINOOD::Destroy()", "Failed to release OPENVINO Models", __FILE__, __LINE__);
|
||
|
|
return false;
|
||
|
|
}
|
||
|
|
}
|
||
|
|
//private
|
||
|
|
void OPENVINOOD::InitialModel() {
|
||
|
|
try {
|
||
|
|
// Step 1: Initialize OpenVINO Runtime Core
|
||
|
|
ov::Core core;
|
||
|
|
// Step 2: Get Available Devices and Log
|
||
|
|
std::vector<std::string> available_devices = core.get_available_devices();
|
||
|
|
// Define device priority: NPU > GPU > CPU
|
||
|
|
std::vector<std::string> priority_devices = { "NPU", "GPU" };
|
||
|
|
bool device_found = false;
|
||
|
|
|
||
|
|
// Iterate over prioritized devices
|
||
|
|
for (const auto& device : priority_devices) {
|
||
|
|
if (std::find(available_devices.begin(), available_devices.end(), device) != available_devices.end()) {
|
||
|
|
if (device == "NPU") {
|
||
|
|
core.set_property("NPU", ov::hint::performance_mode(ov::hint::PerformanceMode::CUMULATIVE_THROUGHPUT));
|
||
|
|
core.set_property("GPU", ov::hint::performance_mode(ov::hint::PerformanceMode::CUMULATIVE_THROUGHPUT));
|
||
|
|
compiled_model_ = core.compile_model(_modelFilePath, "MULTI:NPU, GPU");
|
||
|
|
}
|
||
|
|
else {
|
||
|
|
// Configure and compile for individual device
|
||
|
|
//core.set_property(device, ov::hint::performance_mode(ov::hint::PerformanceMode::THROUGHPUT));
|
||
|
|
compiled_model_ = core.compile_model(_modelFilePath, device);
|
||
|
|
}
|
||
|
|
device_found = true;
|
||
|
|
break;
|
||
|
|
}
|
||
|
|
}
|
||
|
|
// Fallback: Default to CPU if no devices found
|
||
|
|
if (!device_found) {
|
||
|
|
//core.set_property("CPU", ov::hint::performance_mode(ov::hint::PerformanceMode::THROUGHPUT));
|
||
|
|
compiled_model_ = core.compile_model(_modelFilePath, "CPU");
|
||
|
|
}
|
||
|
|
// Step 3: Create Inference Request
|
||
|
|
inference_request_ = compiled_model_.create_infer_request();
|
||
|
|
}
|
||
|
|
catch (const std::exception& e) {
|
||
|
|
// Log any errors
|
||
|
|
this->_logger.LogFatal("OPENVINOOD::InitialModel", e.what(), __FILE__, __LINE__);
|
||
|
|
}
|
||
|
|
}
|
||
|
|
}
|
||
|
|
|
||
|
|
//std::vector<Object> OPENVINOOD::RunInference(const cv::Mat& input, const std::string& camera_id) {
|
||
|
|
// std::lock_guard<std::recursive_mutex> lock(_mutex);
|
||
|
|
// std::vector<Object> outputs;
|
||
|
|
// if (!_licenseValid) {
|
||
|
|
// _logger.LogError("OPENVINOOD::RunInference", "Invalid License", __FILE__, __LINE__);
|
||
|
|
// return outputs;
|
||
|
|
// }
|
||
|
|
// if (!_isInitialized) {
|
||
|
|
// _logger.LogError("OPENVINOOD::RunInference", "Model is not initialized", __FILE__, __LINE__);
|
||
|
|
// return outputs;
|
||
|
|
// }
|
||
|
|
// if (input.empty()) {
|
||
|
|
// _logger.LogError("OPENVINOOD::RunInference", "Input image is empty", __FILE__, __LINE__);
|
||
|
|
// return outputs;
|
||
|
|
// }
|
||
|
|
// try {
|
||
|
|
// // Step 0: Prepare input
|
||
|
|
// if (input.empty()) return outputs;
|
||
|
|
// if ((input.cols < 10) || (input.rows < 10)) return outputs;
|
||
|
|
// cv::Mat letterbox_img = PreProcessing(input);
|
||
|
|
// if (letterbox_img.empty()) {
|
||
|
|
// _logger.LogError("OPENVINOOD::RunInference", "PreProcessing failed", __FILE__, __LINE__);
|
||
|
|
// return outputs;
|
||
|
|
// }
|
||
|
|
// int imageSize = 640;
|
||
|
|
// int maxImageSize = std::max(letterbox_img.cols, letterbox_img.rows);
|
||
|
|
// //if (maxImageSize < imageSize)imageSize = maxImageSize;
|
||
|
|
// float scale = static_cast<float>(letterbox_img.rows) / imageSize;
|
||
|
|
// cv::Mat blob = cv::dnn::blobFromImage(letterbox_img, 1.0 / 255.0,
|
||
|
|
// cv::Size(imageSize, imageSize),
|
||
|
|
// cv::Scalar(), true);
|
||
|
|
|
||
|
|
// // Step 1: Feed blob to the network
|
||
|
|
// auto input_port = compiled_model_.input();
|
||
|
|
// ov::Tensor input_tensor(input_port.get_element_type(), input_port.get_shape(), blob.ptr(0));
|
||
|
|
// inference_request_.set_input_tensor(input_tensor);
|
||
|
|
// inference_request_.infer();
|
||
|
|
|
||
|
|
// // Step 4: Get output
|
||
|
|
// auto output = inference_request_.get_output_tensor(0);
|
||
|
|
// auto output_shape = output.get_shape();
|
||
|
|
|
||
|
|
// if (output_shape.size() != 3) {
|
||
|
|
// _logger.LogError("OPENVINOOD::RunInference", "Unexpected output shape", __FILE__, __LINE__);
|
||
|
|
// return outputs;
|
||
|
|
// }
|
||
|
|
|
||
|
|
// int rows = output_shape[2]; // 8400
|
||
|
|
// int dimensions = output_shape[1]; // 84: box[cx, cy, w, h]+80 class scores
|
||
|
|
// float* data = output.data<float>();
|
||
|
|
// cv::Mat output_buffer(dimensions, rows, CV_32F, data);
|
||
|
|
// transpose(output_buffer, output_buffer); // [8400, 84]
|
||
|
|
|
||
|
|
// std::vector<int> class_ids;
|
||
|
|
// std::vector<float> class_scores;
|
||
|
|
// std::vector<cv::Rect> boxes;
|
||
|
|
|
||
|
|
// // Step 5: Post-processing
|
||
|
|
// for (int i = 0; i < output_buffer.rows; i++) {
|
||
|
|
// cv::Mat classes_scores = output_buffer.row(i).colRange(4, dimensions);
|
||
|
|
// cv::Point class_id;
|
||
|
|
// double max_class_score;
|
||
|
|
// minMaxLoc(classes_scores, nullptr, &max_class_score, nullptr, &class_id);
|
||
|
|
|
||
|
|
// if (max_class_score > _modelConfig.detectionScoreThreshold) {
|
||
|
|
// class_scores.push_back(static_cast<float>(max_class_score));
|
||
|
|
// class_ids.push_back(class_id.x);
|
||
|
|
|
||
|
|
// float cx = output_buffer.at<float>(i, 0);
|
||
|
|
// float cy = output_buffer.at<float>(i, 1);
|
||
|
|
// float w = output_buffer.at<float>(i, 2);
|
||
|
|
// float h = output_buffer.at<float>(i, 3);
|
||
|
|
|
||
|
|
// int left = static_cast<int>((cx - 0.5f * w) * scale);
|
||
|
|
// int top = static_cast<int>((cy - 0.5f * h) * scale);
|
||
|
|
// int width = static_cast<int>(w * scale);
|
||
|
|
// int height = static_cast<int>(h * scale);
|
||
|
|
|
||
|
|
// left = std::max(0, left);
|
||
|
|
// top = std::max(0, top);
|
||
|
|
// width = std::min(input.cols - left, width);
|
||
|
|
// height = std::min(input.rows - top, height);
|
||
|
|
|
||
|
|
// boxes.emplace_back(left, top, width, height);
|
||
|
|
// }
|
||
|
|
// }
|
||
|
|
// // NMS
|
||
|
|
// int classNameSize = static_cast<int>(_classes.size());
|
||
|
|
// std::vector<int> indices;
|
||
|
|
// cv::dnn::NMSBoxes(boxes, class_scores, _modelConfig.detectionScoreThreshold, _modelConfig.modelMNSThreshold, indices);
|
||
|
|
// for (int id : indices) {
|
||
|
|
// if (class_scores[id] >= _modelConfig.detectionScoreThreshold) {
|
||
|
|
// Object result;
|
||
|
|
// int class_id = class_ids[id];
|
||
|
|
// result.classId = class_id;
|
||
|
|
// if (!_classes.empty()) {
|
||
|
|
// if (result.classId < classNameSize) {
|
||
|
|
// result.className = _classes[result.classId];
|
||
|
|
// }
|
||
|
|
// else {
|
||
|
|
// result.className = _classes[classNameSize - 1]; // Use last valid class name if out of range
|
||
|
|
// }
|
||
|
|
// }
|
||
|
|
// else {
|
||
|
|
// result.className = "Unknown"; // Fallback if _classes is empty
|
||
|
|
// }
|
||
|
|
|
||
|
|
// result.confidence = class_scores[id];
|
||
|
|
// result.box = boxes[id];
|
||
|
|
// result.polygon = ANSUtilityHelper::RectToNormalizedPolygon(result.box, input.cols, input.rows);
|
||
|
|
// result.cameraId = camera_id;
|
||
|
|
// outputs.push_back(result);
|
||
|
|
// }
|
||
|
|
// }
|
||
|
|
// }
|
||
|
|
// catch (const std::exception& e) {
|
||
|
|
// _logger.LogFatal("OPENVINOOD::RunInference", e.what(), __FILE__, __LINE__);
|
||
|
|
// }
|
||
|
|
// catch (...) {
|
||
|
|
// _logger.LogFatal("OPENVINOOD::RunInference", "Unknown error occurred", __FILE__, __LINE__);
|
||
|
|
// }
|
||
|
|
// return outputs; //
|
||
|
|
//}
|