Comments (34)
SHOUshou0426
commented on July 4, 2024
from tensorrt-for-yolo-series.
Linaom1214
commented on July 4, 2024
您好，作者，我使用环境为ubuntu18.04 cuda10.2 cudnn8.1.1 tensorrt7.2.3.4，在使用nrom/yolo.cpp 推理时，使用官方v8s模型，转换tensorrt 模型，推理后的图片出现乱框情况，是否可以给个解答呢
提供一下您导出的命令
from tensorrt-for-yolo-series.
SHOUshou0426
commented on July 4, 2024
您好，我是自己编写的onnx-tensorrt 请您查看
// onnx to yolov8
void Yolo::onnxYoloEngine(const nvinfer1::DataType dataType){
if(fileExists(m_EnginePath))return;
// load onnx model
std::ifstream onnxFile(m_WtsFilePath.c_str(), std::ios::binary | std::ios::in);
if (!onnxFile.is_open()) {
std::cerr << "Error: Failed to open ONNX file." << std::endl;
}
// onnx create
auto builder = nvinfer1::createInferBuilder(gLogger);
const auto explicitBatch = 1U << static_cast<uint32_t>(NetworkDefinitionCreationFlag::kEXPLICIT_BATCH);
auto network = builder->createNetworkV2(explicitBatch);
nvonnxparser::IParser* parser = nvonnxparser::createParser(*network, gLogger);
parser->parseFromFile(m_WtsFilePath.c_str(), static_cast<int>(Logger::Severity::kINFO));
for (int i = 0; i < parser->getNbErrors(); ++i)
{
std::cout << parser->getError(i)->desc() << std::endl;
}
std::cout << "successfully load the onnx model" << std::endl;
// build the engine
builder->setMaxBatchSize(batchSize);
IBuilderConfig* config = builder->createBuilderConfig();
config->setMaxWorkspaceSize(1 << 24); // 16MB
//precision int8 fp16 fp32
//if (dataType == nvinfer1::DataType::kINT8)
//{
// config->setFlag(nvinfer1::BuilderFlag::kINT8);
// config->setInt8Calibrator(calibrator);
//}
//else if (dataType == nvinfer1::DataType::kHALF)
//{
// config->setFlag(nvinfer1::BuilderFlag::kFP16);
//}
if(dataType == nvinfer1::DataType::kHALF)
{
config->setFlag(nvinfer1::BuilderFlag::kFP16);
}
engine = builder->buildEngineWithConfig(*network, *config);
std::cout << "Building the TensorRT Engine..." << std::endl;
//Serialize the engine
BuildOnnxEngine();
// onnx destroy
// engine->destroy();
network->destroy();
builder->destroy();
parser->destroy();
}
void Yolo::BuildOnnxEngine() {
std::cout << "Serializing the TensorRT Engine..." << std::endl;
assert(engine && "Invalid TensorRT Engine");
trtModelStream = engine->serialize();
assert(trtModelStream != nullptr);
assert(trtModelStream && "Unable to serialize engine");
assert(!m_EnginePath.empty() && "Enginepath is empty");
// write data to output file
std::stringstream gieModelStream;
gieModelStream.seekg(0, gieModelStream.beg);
gieModelStream.write(static_cast<const char*>(trtModelStream->data()), trtModelStream->size());
std::ofstream outFile;
outFile.open(m_EnginePath, std::ios::binary);
outFile << gieModelStream.rdbuf();
outFile.close();
std::cout << "Serialized plan file cached at location : " << m_EnginePath << std::endl;
}
from tensorrt-for-yolo-series.
Linaom1214
commented on July 4, 2024
您好，我是自己编写的onnx-tensorrt 请您查看
// onnx to yolov8 void Yolo::onnxYoloEngine(const nvinfer1::DataType dataType){ if(fileExists(m_EnginePath))return; // load onnx model std::ifstream onnxFile(m_WtsFilePath.c_str(), std::ios::binary | std::ios::in); if (!onnxFile.is_open()) { std::cerr << "Error: Failed to open ONNX file." << std::endl; } // onnx create auto builder = nvinfer1::createInferBuilder(gLogger); const auto explicitBatch = 1U << static_cast<uint32_t>(NetworkDefinitionCreationFlag::kEXPLICIT_BATCH); auto network = builder->createNetworkV2(explicitBatch); nvonnxparser::IParser* parser = nvonnxparser::createParser(*network, gLogger); parser->parseFromFile(m_WtsFilePath.c_str(), static_cast<int>(Logger::Severity::kINFO)); for (int i = 0; i < parser->getNbErrors(); ++i) { std::cout << parser->getError(i)->desc() << std::endl; } std::cout << "successfully load the onnx model" << std::endl; // build the engine builder->setMaxBatchSize(batchSize); IBuilderConfig* config = builder->createBuilderConfig(); config->setMaxWorkspaceSize(1 << 24); // 16MB //precision int8 fp16 fp32 //if (dataType == nvinfer1::DataType::kINT8) //{ // config->setFlag(nvinfer1::BuilderFlag::kINT8); // config->setInt8Calibrator(calibrator); //} //else if (dataType == nvinfer1::DataType::kHALF) //{ // config->setFlag(nvinfer1::BuilderFlag::kFP16); //} if(dataType == nvinfer1::DataType::kHALF) { config->setFlag(nvinfer1::BuilderFlag::kFP16); } engine = builder->buildEngineWithConfig(*network, *config); std::cout << "Building the TensorRT Engine..." << std::endl; //Serialize the engine BuildOnnxEngine(); // onnx destroy // engine->destroy(); network->destroy(); builder->destroy(); parser->destroy(); } void Yolo::BuildOnnxEngine() { std::cout << "Serializing the TensorRT Engine..." << std::endl; assert(engine && "Invalid TensorRT Engine"); trtModelStream = engine->serialize(); assert(trtModelStream != nullptr); assert(trtModelStream && "Unable to serialize engine"); assert(!m_EnginePath.empty() && "Enginepath is empty"); // write data to output file std::stringstream gieModelStream; gieModelStream.seekg(0, gieModelStream.beg); gieModelStream.write(static_cast<const char*>(trtModelStream->data()), trtModelStream->size()); std::ofstream outFile; outFile.open(m_EnginePath, std::ios::binary); outFile << gieModelStream.rdbuf(); outFile.close(); std::cout << "Serialized plan file cached at location : " << m_EnginePath << std::endl; }
先验证一下，直接用python推理结果对不对，可能模型本身就有问题
from tensorrt-for-yolo-series.
SHOUshou0426
commented on July 4, 2024
您那边使用官方的模型，转换tensorrt 并没有出现乱框的情况，是吗
from tensorrt-for-yolo-series.
Linaom1214
commented on July 4, 2024
您那边使用官方的模型，转换tensorrt 并没有出现乱框的情况，是吗
是的，不过要确保，转出的时候是全精度的onnx模型
from tensorrt-for-yolo-series.
SHOUshou0426
commented on July 4, 2024
您那边使用官方的模型，转换tensorrt 并没有出现乱框的情况，是吗
是的，不过要确保，转出的时候是全精度的onnx模型
FP32 精度是吗
from tensorrt-for-yolo-series.
Linaom1214
commented on July 4, 2024
您那边使用官方的模型，转换tensorrt 并没有出现乱框的情况，是吗
是的，不过要确保，转出的时候是全精度的onnx模型
FP32 精度是吗
是的
from tensorrt-for-yolo-series.
SHOUshou0426
commented on July 4, 2024
您那边使用官方的模型，转换tensorrt 并没有出现乱框的情况，是吗
是的，不过要确保，转出的时候是全精度的onnx模型
FP32 精度是吗
是的
好的，感谢您的回答，我验证一下，是否是模型问题，后续有问题，请多多指教
from tensorrt-for-yolo-series.
SHOUshou0426
commented on July 4, 2024
您那边使用官方的模型，转换tensorrt 并没有出现乱框的情况，是吗
是的，不过要确保，转出的时候是全精度的onnx模型
FP32 精度是吗
是的
您好，我尝试了使用trtexec 转换进行推理，还是出现上述情况
trtexec --onnx=/workspace/robot/ultralytics/yolov8s.onnx --saveEngine=yolov8s.engine --explicitBatch
from tensorrt-for-yolo-series.
SHOUshou0426
commented on July 4, 2024
您好，我是自己编写的onnx-tensorrt 请您查看
// onnx to yolov8 void Yolo::onnxYoloEngine(const nvinfer1::DataType dataType){ if(fileExists(m_EnginePath))return; // load onnx model std::ifstream onnxFile(m_WtsFilePath.c_str(), std::ios::binary | std::ios::in); if (!onnxFile.is_open()) { std::cerr << "Error: Failed to open ONNX file." << std::endl; } // onnx create auto builder = nvinfer1::createInferBuilder(gLogger); const auto explicitBatch = 1U << static_cast<uint32_t>(NetworkDefinitionCreationFlag::kEXPLICIT_BATCH); auto network = builder->createNetworkV2(explicitBatch); nvonnxparser::IParser* parser = nvonnxparser::createParser(*network, gLogger); parser->parseFromFile(m_WtsFilePath.c_str(), static_cast<int>(Logger::Severity::kINFO)); for (int i = 0; i < parser->getNbErrors(); ++i) { std::cout << parser->getError(i)->desc() << std::endl; } std::cout << "successfully load the onnx model" << std::endl; // build the engine builder->setMaxBatchSize(batchSize); IBuilderConfig* config = builder->createBuilderConfig(); config->setMaxWorkspaceSize(1 << 24); // 16MB //precision int8 fp16 fp32 //if (dataType == nvinfer1::DataType::kINT8) //{ // config->setFlag(nvinfer1::BuilderFlag::kINT8); // config->setInt8Calibrator(calibrator); //} //else if (dataType == nvinfer1::DataType::kHALF) //{ // config->setFlag(nvinfer1::BuilderFlag::kFP16); //} if(dataType == nvinfer1::DataType::kHALF) { config->setFlag(nvinfer1::BuilderFlag::kFP16); } engine = builder->buildEngineWithConfig(*network, *config); std::cout << "Building the TensorRT Engine..." << std::endl; //Serialize the engine BuildOnnxEngine(); // onnx destroy // engine->destroy(); network->destroy(); builder->destroy(); parser->destroy(); } void Yolo::BuildOnnxEngine() { std::cout << "Serializing the TensorRT Engine..." << std::endl; assert(engine && "Invalid TensorRT Engine"); trtModelStream = engine->serialize(); assert(trtModelStream != nullptr); assert(trtModelStream && "Unable to serialize engine"); assert(!m_EnginePath.empty() && "Enginepath is empty"); // write data to output file std::stringstream gieModelStream; gieModelStream.seekg(0, gieModelStream.beg); gieModelStream.write(static_cast<const char*>(trtModelStream->data()), trtModelStream->size()); std::ofstream outFile; outFile.open(m_EnginePath, std::ios::binary); outFile << gieModelStream.rdbuf(); outFile.close(); std::cout << "Serialized plan file cached at location : " << m_EnginePath << std::endl; }先验证一下，直接用python推理结果对不对，可能模型本身就有问题
您可以将您的模型转换命令告诉一下吗，我可以尝试一下您的转换方式，并验证
from tensorrt-for-yolo-series.
Linaom1214
commented on July 4, 2024
您好，我是自己编写的onnx-tensorrt 请您查看
// onnx to yolov8 void Yolo::onnxYoloEngine(const nvinfer1::DataType dataType){ if(fileExists(m_EnginePath))return; // load onnx model std::ifstream onnxFile(m_WtsFilePath.c_str(), std::ios::binary | std::ios::in); if (!onnxFile.is_open()) { std::cerr << "Error: Failed to open ONNX file." << std::endl; } // onnx create auto builder = nvinfer1::createInferBuilder(gLogger); const auto explicitBatch = 1U << static_cast<uint32_t>(NetworkDefinitionCreationFlag::kEXPLICIT_BATCH); auto network = builder->createNetworkV2(explicitBatch); nvonnxparser::IParser* parser = nvonnxparser::createParser(*network, gLogger); parser->parseFromFile(m_WtsFilePath.c_str(), static_cast<int>(Logger::Severity::kINFO)); for (int i = 0; i < parser->getNbErrors(); ++i) { std::cout << parser->getError(i)->desc() << std::endl; } std::cout << "successfully load the onnx model" << std::endl; // build the engine builder->setMaxBatchSize(batchSize); IBuilderConfig* config = builder->createBuilderConfig(); config->setMaxWorkspaceSize(1 << 24); // 16MB //precision int8 fp16 fp32 //if (dataType == nvinfer1::DataType::kINT8) //{ // config->setFlag(nvinfer1::BuilderFlag::kINT8); // config->setInt8Calibrator(calibrator); //} //else if (dataType == nvinfer1::DataType::kHALF) //{ // config->setFlag(nvinfer1::BuilderFlag::kFP16); //} if(dataType == nvinfer1::DataType::kHALF) { config->setFlag(nvinfer1::BuilderFlag::kFP16); } engine = builder->buildEngineWithConfig(*network, *config); std::cout << "Building the TensorRT Engine..." << std::endl; //Serialize the engine BuildOnnxEngine(); // onnx destroy // engine->destroy(); network->destroy(); builder->destroy(); parser->destroy(); } void Yolo::BuildOnnxEngine() { std::cout << "Serializing the TensorRT Engine..." << std::endl; assert(engine && "Invalid TensorRT Engine"); trtModelStream = engine->serialize(); assert(trtModelStream != nullptr); assert(trtModelStream && "Unable to serialize engine"); assert(!m_EnginePath.empty() && "Enginepath is empty"); // write data to output file std::stringstream gieModelStream; gieModelStream.seekg(0, gieModelStream.beg); gieModelStream.write(static_cast<const char*>(trtModelStream->data()), trtModelStream->size()); std::ofstream outFile; outFile.open(m_EnginePath, std::ios::binary); outFile << gieModelStream.rdbuf(); outFile.close(); std::cout << "Serialized plan file cached at location : " << m_EnginePath << std::endl; }先验证一下，直接用python推理结果对不对，可能模型本身就有问题
您可以将您的模型转换命令告诉一下吗，我可以尝试一下您的转换方式，并验证
参考本仓库提供的就行，或者直接采用trtexec
from tensorrt-for-yolo-series.
SHOUshou0426
commented on July 4, 2024
您好，我是自己编写的onnx-tensorrt 请您查看
// onnx to yolov8 void Yolo::onnxYoloEngine(const nvinfer1::DataType dataType){ if(fileExists(m_EnginePath))return; // load onnx model std::ifstream onnxFile(m_WtsFilePath.c_str(), std::ios::binary | std::ios::in); if (!onnxFile.is_open()) { std::cerr << "Error: Failed to open ONNX file." << std::endl; } // onnx create auto builder = nvinfer1::createInferBuilder(gLogger); const auto explicitBatch = 1U << static_cast<uint32_t>(NetworkDefinitionCreationFlag::kEXPLICIT_BATCH); auto network = builder->createNetworkV2(explicitBatch); nvonnxparser::IParser* parser = nvonnxparser::createParser(*network, gLogger); parser->parseFromFile(m_WtsFilePath.c_str(), static_cast<int>(Logger::Severity::kINFO)); for (int i = 0; i < parser->getNbErrors(); ++i) { std::cout << parser->getError(i)->desc() << std::endl; } std::cout << "successfully load the onnx model" << std::endl; // build the engine builder->setMaxBatchSize(batchSize); IBuilderConfig* config = builder->createBuilderConfig(); config->setMaxWorkspaceSize(1 << 24); // 16MB //precision int8 fp16 fp32 //if (dataType == nvinfer1::DataType::kINT8) //{ // config->setFlag(nvinfer1::BuilderFlag::kINT8); // config->setInt8Calibrator(calibrator); //} //else if (dataType == nvinfer1::DataType::kHALF) //{ // config->setFlag(nvinfer1::BuilderFlag::kFP16); //} if(dataType == nvinfer1::DataType::kHALF) { config->setFlag(nvinfer1::BuilderFlag::kFP16); } engine = builder->buildEngineWithConfig(*network, *config); std::cout << "Building the TensorRT Engine..." << std::endl; //Serialize the engine BuildOnnxEngine(); // onnx destroy // engine->destroy(); network->destroy(); builder->destroy(); parser->destroy(); } void Yolo::BuildOnnxEngine() { std::cout << "Serializing the TensorRT Engine..." << std::endl; assert(engine && "Invalid TensorRT Engine"); trtModelStream = engine->serialize(); assert(trtModelStream != nullptr); assert(trtModelStream && "Unable to serialize engine"); assert(!m_EnginePath.empty() && "Enginepath is empty"); // write data to output file std::stringstream gieModelStream; gieModelStream.seekg(0, gieModelStream.beg); gieModelStream.write(static_cast<const char*>(trtModelStream->data()), trtModelStream->size()); std::ofstream outFile; outFile.open(m_EnginePath, std::ios::binary); outFile << gieModelStream.rdbuf(); outFile.close(); std::cout << "Serialized plan file cached at location : " << m_EnginePath << std::endl; }先验证一下，直接用python推理结果对不对，可能模型本身就有问题
您可以将您的模型转换命令告诉一下吗，我可以尝试一下您的转换方式，并验证
参考本仓库提供的就行，或者直接采用trtexec
我使用的则是trtexec trtexec --onnx=/workspace/robot/ultralytics/yolov8s.onnx --saveEngine=yolov8s.engine --explicitBatch
from tensorrt-for-yolo-series.
SHOUshou0426
commented on July 4, 2024
您好，我是自己编写的onnx-tensorrt 请您查看
// onnx to yolov8 void Yolo::onnxYoloEngine(const nvinfer1::DataType dataType){ if(fileExists(m_EnginePath))return; // load onnx model std::ifstream onnxFile(m_WtsFilePath.c_str(), std::ios::binary | std::ios::in); if (!onnxFile.is_open()) { std::cerr << "Error: Failed to open ONNX file." << std::endl; } // onnx create auto builder = nvinfer1::createInferBuilder(gLogger); const auto explicitBatch = 1U << static_cast<uint32_t>(NetworkDefinitionCreationFlag::kEXPLICIT_BATCH); auto network = builder->createNetworkV2(explicitBatch); nvonnxparser::IParser* parser = nvonnxparser::createParser(*network, gLogger); parser->parseFromFile(m_WtsFilePath.c_str(), static_cast<int>(Logger::Severity::kINFO)); for (int i = 0; i < parser->getNbErrors(); ++i) { std::cout << parser->getError(i)->desc() << std::endl; } std::cout << "successfully load the onnx model" << std::endl; // build the engine builder->setMaxBatchSize(batchSize); IBuilderConfig* config = builder->createBuilderConfig(); config->setMaxWorkspaceSize(1 << 24); // 16MB //precision int8 fp16 fp32 //if (dataType == nvinfer1::DataType::kINT8) //{ // config->setFlag(nvinfer1::BuilderFlag::kINT8); // config->setInt8Calibrator(calibrator); //} //else if (dataType == nvinfer1::DataType::kHALF) //{ // config->setFlag(nvinfer1::BuilderFlag::kFP16); //} if(dataType == nvinfer1::DataType::kHALF) { config->setFlag(nvinfer1::BuilderFlag::kFP16); } engine = builder->buildEngineWithConfig(*network, *config); std::cout << "Building the TensorRT Engine..." << std::endl; //Serialize the engine BuildOnnxEngine(); // onnx destroy // engine->destroy(); network->destroy(); builder->destroy(); parser->destroy(); } void Yolo::BuildOnnxEngine() { std::cout << "Serializing the TensorRT Engine..." << std::endl; assert(engine && "Invalid TensorRT Engine"); trtModelStream = engine->serialize(); assert(trtModelStream != nullptr); assert(trtModelStream && "Unable to serialize engine"); assert(!m_EnginePath.empty() && "Enginepath is empty"); // write data to output file std::stringstream gieModelStream; gieModelStream.seekg(0, gieModelStream.beg); gieModelStream.write(static_cast<const char*>(trtModelStream->data()), trtModelStream->size()); std::ofstream outFile; outFile.open(m_EnginePath, std::ios::binary); outFile << gieModelStream.rdbuf(); outFile.close(); std::cout << "Serialized plan file cached at location : " << m_EnginePath << std::endl; }先验证一下，直接用python推理结果对不对，可能模型本身就有问题
您可以将您的模型转换命令告诉一下吗，我可以尝试一下您的转换方式，并验证
参考本仓库提供的就行，或者直接采用trtexec
我使用的则是trtexec trtexec --onnx=/workspace/robot/ultralytics/yolov8s.onnx --saveEngine=yolov8s.engine --explicitBatch
将模型重新转换 同样出现了上述乱框情况
from tensorrt-for-yolo-series.
SHOUshou0426
commented on July 4, 2024
您好，我是自己编写的onnx-tensorrt 请您查看
// onnx to yolov8 void Yolo::onnxYoloEngine(const nvinfer1::DataType dataType){ if(fileExists(m_EnginePath))return; // load onnx model std::ifstream onnxFile(m_WtsFilePath.c_str(), std::ios::binary | std::ios::in); if (!onnxFile.is_open()) { std::cerr << "Error: Failed to open ONNX file." << std::endl; } // onnx create auto builder = nvinfer1::createInferBuilder(gLogger); const auto explicitBatch = 1U << static_cast<uint32_t>(NetworkDefinitionCreationFlag::kEXPLICIT_BATCH); auto network = builder->createNetworkV2(explicitBatch); nvonnxparser::IParser* parser = nvonnxparser::createParser(*network, gLogger); parser->parseFromFile(m_WtsFilePath.c_str(), static_cast<int>(Logger::Severity::kINFO)); for (int i = 0; i < parser->getNbErrors(); ++i) { std::cout << parser->getError(i)->desc() << std::endl; } std::cout << "successfully load the onnx model" << std::endl; // build the engine builder->setMaxBatchSize(batchSize); IBuilderConfig* config = builder->createBuilderConfig(); config->setMaxWorkspaceSize(1 << 24); // 16MB //precision int8 fp16 fp32 //if (dataType == nvinfer1::DataType::kINT8) //{ // config->setFlag(nvinfer1::BuilderFlag::kINT8); // config->setInt8Calibrator(calibrator); //} //else if (dataType == nvinfer1::DataType::kHALF) //{ // config->setFlag(nvinfer1::BuilderFlag::kFP16); //} if(dataType == nvinfer1::DataType::kHALF) { config->setFlag(nvinfer1::BuilderFlag::kFP16); } engine = builder->buildEngineWithConfig(*network, *config); std::cout << "Building the TensorRT Engine..." << std::endl; //Serialize the engine BuildOnnxEngine(); // onnx destroy // engine->destroy(); network->destroy(); builder->destroy(); parser->destroy(); } void Yolo::BuildOnnxEngine() { std::cout << "Serializing the TensorRT Engine..." << std::endl; assert(engine && "Invalid TensorRT Engine"); trtModelStream = engine->serialize(); assert(trtModelStream != nullptr); assert(trtModelStream && "Unable to serialize engine"); assert(!m_EnginePath.empty() && "Enginepath is empty"); // write data to output file std::stringstream gieModelStream; gieModelStream.seekg(0, gieModelStream.beg); gieModelStream.write(static_cast<const char*>(trtModelStream->data()), trtModelStream->size()); std::ofstream outFile; outFile.open(m_EnginePath, std::ios::binary); outFile << gieModelStream.rdbuf(); outFile.close(); std::cout << "Serialized plan file cached at location : " << m_EnginePath << std::endl; }先验证一下，直接用python推理结果对不对，可能模型本身就有问题
您可以将您的模型转换命令告诉一下吗，我可以尝试一下您的转换方式，并验证
参考本仓库提供的就行，或者直接采用trtexec
使用仓库的转换也是出现乱框情况
from tensorrt-for-yolo-series.
Linaom1214
commented on July 4, 2024
@SHOUshou0426 这个仓库v8 只支持end2end
from tensorrt-for-yolo-series.
SHOUshou0426
commented on July 4, 2024
@SHOUshou0426 这个仓库v8 只支持end2end
我使用的是支持nms 的库，nrom/yolo.cpp 那个不是end2end 的
from tensorrt-for-yolo-series.
SHOUshou0426
commented on July 4, 2024
@SHOUshou0426 这个仓库v8 只支持end2end
我查看了cpp 库两个文件夹，end2end 没有nms 操作，norm 有nms 操作，然后我使用的是norm 库里的，
from tensorrt-for-yolo-series.
Linaom1214
commented on July 4, 2024
@SHOUshou0426 这个仓库v8 只支持end2end
我查看了cpp 库两个文件夹，end2end 没有nms 操作，norm 有nms 操作，然后我使用的是norm 库里的，
好的，我之后找个时间测试一下
from tensorrt-for-yolo-series.
SHOUshou0426
commented on July 4, 2024
@SHOUshou0426 这个仓库v8 只支持end2end
我查看了cpp 库两个文件夹，end2end 没有nms 操作，norm 有nms 操作，然后我使用的是norm 库里的，
好的，我之后找个时间测试一下
我晚上查看一下后处理吧，如果修改了，我告诉您一下
from tensorrt-for-yolo-series.
Linaom1214
commented on July 4, 2024
@SHOUshou0426 这个仓库v8 只支持end2end
我查看了cpp 库两个文件夹，end2end 没有nms 操作，norm 有nms 操作，然后我使用的是norm 库里的，
好的，我之后找个时间测试一下
我晚上查看一下后处理吧，如果修改了，我告诉您一下
好的
from tensorrt-for-yolo-series.
Leoyed
commented on July 4, 2024
您好，请问这个问题解决了吗？我用norm里的推理也是出现您这个情况，都是乱框的，请问下 怎么处理呢
from tensorrt-for-yolo-series.
SHOUshou0426
commented on July 4, 2024
您好，请问这个问题解决了吗？我用norm里的推理也是出现您这个情况，都是乱框的，请问下 怎么处理呢
我把修改后的重新发送给您
from tensorrt-for-yolo-series.
SHOUshou0426
commented on July 4, 2024
@Leoyed 给我一个邮箱我发送给您
from tensorrt-for-yolo-series.
Leoyed
commented on July 4, 2024
@Leoyed 给我一个邮箱我发送给您
谢谢老板您，这是我的邮箱：[email protected]
from tensorrt-for-yolo-series.
Leoyed
commented on July 4, 2024
您好，请问这个问题解决了吗？我用norm里的推理也是出现您这个情况，都是乱框的，请问下 怎么处理呢
我把修改后的重新发送给您
大佬，请问下这个是哪块有问题的，这两天看了下代码感觉后处理有点不太对，不知道是不是这样
from tensorrt-for-yolo-series.
Linaom1214
commented on July 4, 2024
您好，请问这个问题解决了吗？我用norm里的推理也是出现您这个情况，都是乱框的，请问下 怎么处理呢
我把修改后的重新发送给您
您可以提一个pr，我这边修复一下
from tensorrt-for-yolo-series.
SHOUshou0426
commented on July 4, 2024
@Linaom1214 pr 让 @Leoyed 提一下把，我这边有一些自己的工作
from tensorrt-for-yolo-series.
wyq-aki
commented on July 4, 2024
您好，请问这个问题解决了吗？我用norm里的推理也是出现您这个情况，都是乱框的，请问下 怎么处理呢
我把修改后的重新发送给您
大佬，请问下这个是哪块有问题的，这两天看了下代码感觉后处理有点不太对，不知道是不是这样
你好，请问这块是后处理的问题吗
from tensorrt-for-yolo-series.
leayz-888
commented on July 4, 2024
您好，请问一下，norm里v8的后处理中，generate_yolo_proposals函数里面有段代码是这样的：
float box_objectness = feat_blob[basic_pos+4];
// std::cout<<feat_blob<<std::endl;
for (int class_idx = 0; class_idx < num_class; class_idx++)
{
float box_cls_score = feat_blob[basic_pos + 5 + class_idx];
float box_prob = box_objectness * box_cls_score;
if (box_prob > prob_threshold){......}
这里用feat_blob[basic_pos+4]取置信度，但是v8中是没有回归置信度的，这样取到的是第一个类别的分类概率，这样后面再根据box_objectnessbox_cls_score过滤低置信度的检测框会不会有问题了，可能这是导致v8 norm出现乱框的原因哦
from tensorrt-for-yolo-series.
wyq-aki
commented on July 4, 2024
您好，请问一下，norm里v8的后处理中，generate_yolo_proposals函数里面有段代码是这样的： float box_objectness = feat_blob[basic_pos+4]; // std::cout<<_feat_blob<<std::endl; for (int class_idx = 0; class_idx < num_class; class_idx++) { float box_cls_score = feat_blob[basic_pos + 5 + class_idx]; float box_prob = box_objectness * box_cls_score; if (box_prob > prob_threshold){......} 这里用feat_blob[basic_pos+4]取置信度，但是v8中是没有回归置信度的，这样取到的是第一个类别的分类概率，这样后面再根据box_objectness_box_cls_score过滤低置信度的检测框会不会有问题了，可能这是导致v8 norm出现乱框的原因哦
是的，作者trt.py里如果end2end=False的话，在utils.py里对模型推理出来的output进行reshape时出错了，我修改成predictions=np.reshape(data,(int(4+self.n_classes), -1)).T，以及postprecess()里scores=predictions[:,4:]，以及vis()里那行*ratio的去掉，预测框就正常了
from tensorrt-for-yolo-series.
Leoyed
commented on July 4, 2024
您好，请问这个问题解决了吗？我用norm里的推理也是出现您这个情况，都是乱框的，请问下 怎么处理呢
我把修改后的重新发送给您
大佬，请问下这个是哪块有问题的，这两天看了下代码感觉后处理有点不太对，不知道是不是这样
你好，请问这块是后处理的问题吗
对的，这后处理代码方法不对。不过这里的前处理是用CPU处理，耗时也很长。
from tensorrt-for-yolo-series.
Linaom1214
commented on July 4, 2024
您好，请问一下，norm里v8的后处理中，generate_yolo_proposals函数里面有段代码是这样的： float box_objectness = feat_blob[basic_pos+4]; // std::cout<<_feat_blob<<std::endl; for (int class_idx = 0; class_idx < num_class; class_idx++) { float box_cls_score = feat_blob[basic_pos + 5 + class_idx]; float box_prob = box_objectness * box_cls_score; if (box_prob > prob_threshold){......} 这里用feat_blob[basic_pos+4]取置信度，但是v8中是没有回归置信度的，这样取到的是第一个类别的分类概率，这样后面再根据box_objectness_box_cls_score过滤低置信度的检测框会不会有问题了，可能这是导致v8 norm出现乱框的原因哦
是的，作者trt.py里如果end2end=False的话，在utils.py里对模型推理出来的output进行reshape时出错了，我修改成
predictions=np.reshape(data,(int(4+self.n_classes), -1)).T，以及postprecess()里scores=predictions[:,4:]，以及vis()里那行*ratio的去掉，预测框就正常了
目前的代码v8只支持end2end， c++ 也是一样的，之前我调试的过程中onnx是全精度的情况下，结果是没有问题的。 @Leoyed 您有别的解决方案吗？
from tensorrt-for-yolo-series.
Linaom1214
commented on July 4, 2024
@Leoyed @leayz-888 @wyq-aki @SHOUshou0426 精度损失可能是预处理的问题，目前已经修复了这个问题，目前正在适配v10，v9已经完成适配，预计今天更新
from tensorrt-for-yolo-series.
Related Issues (20)
- c++ endtoend 关于预测的置信度绘制 HOT 4
- memory leak: Destroy function does not work
- Detection duplicates with fp16 on Jetson Nano (TensorRT v8.2.1.8) HOT 2
- Support for windows?
- License? HOT 4
- TensorRT Conversion Issue "TypeError: pybind11::init(): factory function returned nullptr" HOT 2
- yolox 自己训练的模型 trt推理 位置不对 HOT 2
- int8量化的时候，输入是多个，怎么修改呢？ calib_shape = [calib_batch_size] + list(inputs[0].shape[1:])不对吧 HOT 4
- 怎么将这个项目与RealSense深度相机结合起来？ HOT 3
- yolov9能支持么？ HOT 3
- Jetson nano yolov8部署 onnx ->> trt HOT 2
- Is the TRT conversion I used in YOLOV7 not available in YOLOV9?
- Can I use this code for exporting the RTDETR Model
- 有v8obb的部分嘛
- calibration数据集制作 HOT 1
- yolox和yolov7的官方代码似乎已经做了nms onnx推理 HOT 1
- 你好，我导出的yolov5s int8模型进行推理时，没有检测到任何目标，det的shape为【0,6】，请问如何解决
- No module named 'cuda' HOT 1
- Does this support dynamic batch input convert? HOT 4
Recommend Projects
-
React
A declarative, efficient, and flexible JavaScript library for building user interfaces.
-
Vue.js
🖖 Vue.js is a progressive, incrementally-adoptable JavaScript framework for building UI on the web.
-
Typescript
TypeScript is a superset of JavaScript that compiles to clean JavaScript output.
-
TensorFlow
An Open Source Machine Learning Framework for Everyone
-
Django
The Web framework for perfectionists with deadlines.
-
Laravel
A PHP framework for web artisans
-
D3
Bring data to life with SVG, Canvas and HTML. 📊📈🎉
-
Recommend Topics
-
javascript
JavaScript (JS) is a lightweight interpreted programming language with first-class functions.
-
web
Some thing interesting about web. New door for the world.
-
server
A server is a program made to process requests and deliver data to clients.
-
Machine learning
Machine learning is a way of modeling and interpreting data that allows a piece of software to respond intelligently.
-
Visualization
Some thing interesting about visualization, use data art
-
Game
Some thing interesting about game, make everyone happy.
Recommend Org
-
Facebook
We are working to build community through open source technology. NB: members must have two-factor auth.
-
Microsoft
Open source projects and samples from Microsoft.
-
Google
Google ❤️ Open Source for everyone.
-
Alibaba
Alibaba Open Source for everyone
-
D3
Data-Driven Documents codes.
-
Tencent
China tencent open source team.
from tensorrt-for-yolo-series.