How to find the stdev and mean for yolov5 model optimised using tensorrt optimizer?

Autonomous Vehicles DRIVE AGX Orin DRIVE AGX Orin General
,
1

Please provide the following info (tick the boxes after creating this topic):
Software Version
DRIVE OS 6.0.4 SDK
other

Target Operating System
Linux
QNX
other

Hardware Platform
DRIVE AGX Orin Developer Kit (940-63710-0010-D00)
DRIVE AGX Orin Developer Kit (940-63710-0010-C00)
DRIVE AGX Orin Developer Kit (not sure its number)
other

SDK Manager Version
1.8.2.10409
other

Host Machine Version
native Ubuntu Linux 20.04 Host installed with SDK Manager
native Ubuntu Linux 20.04 Host installed with DRIVE OS Docker Containers
native Ubuntu Linux 18.04 Host installed with DRIVE OS Docker Containers
other

Does incorrect parameters lead to jittery detections ? as when we implemented yolov5 with tensorrt outside driveworks the detections were smooth, we could narrow down the problem to the .json file and its parameters? Anyone know the right parameters or is there any other reason for the jitter.

{
  "dataConditionerParams": {
    "meanValue": [103.530, 116.280, 123.675],
    "stdev": [1.0, 1.0, 1.0],
    "splitPlanes": true,
    "pixelScaleCoefficient": -0.0039215697906911373,
    "ignoreAspectRatio": true,
    "doPerPlaneMeanNormalization": true
  },
  "tonemapType": "none",
  "__comment": "tonemapType can be one of {none, agtm}"
}

We use the above parameters to currently to load the model we got it from here.

2

Dear @nehal,
Could you share the yolov5 preprocessing operation code(which worked with TRT alone). This helps to check parameters needed for json file in DW

3

This is the tensorrt pre processing script that I used I found it here.


#include "yolov5_detector_internal.hpp"

#if __has_include(<opencv2/cudaarithm.hpp>) && __has_include(<opencv2/cudawarping.hpp>)

#define YOLOV5_OPENCV_HAS_CUDA 1

#include <opencv2/cudaarithm.hpp>   /*  cv::cuda::split(...)    */
#include <opencv2/cudawarping.hpp>  /*  cv::cuda::resize(...)   */
#include <opencv2/core/cuda_stream_accessor.hpp>
#endif

#include <cuda_runtime_api.h>

namespace yolov5
{

namespace internal
{

int32_t dimsVolume(const nvinfer1::Dims& dims) noexcept
{
    int32_t r = 0;
    if(dims.nbDims > 0)
    {
        r = 1;
    }

    for(int32_t i = 0; i < dims.nbDims; ++i)
    {
        r = r * dims.d[i];
    }
    return r;
}


bool dimsToString(const nvinfer1::Dims& dims, std::string* out) noexcept
{
    try
    {
        *out = "(";
        for(int32_t i = 0; i < dims.nbDims; ++i)
        {
            *out += std::to_string(dims.d[i]);
            if(i < dims.nbDims - 1)
            {
                out->push_back(',');
            }
        }
        out->push_back(')');
    }
    catch(const std::exception& e)
    {
        return false;
    }
    return true;
}

EngineBinding::EngineBinding() noexcept
{
}

EngineBinding::~EngineBinding() noexcept
{
}

void EngineBinding::swap(EngineBinding& other) noexcept
{
    std::swap(_index, other._index);
    std::swap(_name, other._name);
    std::swap(_volume, other._volume);

    nvinfer1::Dims tmp;
    std::memcpy(&tmp, &_dims, sizeof(nvinfer1::Dims));
    std::memcpy(&_dims, &other._dims, sizeof(nvinfer1::Dims));
    std::memcpy(&other._dims, &tmp, sizeof(nvinfer1::Dims));

    std::swap(_isInput, other._isInput);
}

const int& EngineBinding::index() const noexcept
{
    return _index;
}

const std::string& EngineBinding::name() const noexcept
{
    return _name;
}

const nvinfer1::Dims& EngineBinding::dims() const noexcept
{
    return _dims;
}

const int& EngineBinding::volume() const noexcept
{
    return _volume;
}

bool EngineBinding::isDynamic() const noexcept
{
    for(int i = 0; i < _dims.nbDims; ++i)
    {
        if(_dims.d[i] == -1)
        {
            return true;
        }
    }
    return false;
}

const bool& EngineBinding::isInput() const noexcept
{
    return _isInput;
}

void EngineBinding::toString(std::string* out) const noexcept
{
    std::string dimsStr;
    if(!dimsToString(_dims, &dimsStr))
    {
        return;
    }

    try
    {
        *out = "name: '" + _name + "'"
                    + " ;  dims: " + dimsStr 
                    + " ;  isInput: " + (_isInput ? "true" : "false")
                    + " ;  dynamic: " + (isDynamic() ? "true" : "false");
    }
    catch(const std::exception& e)
    {
    }
}

bool EngineBinding::setup(const std::unique_ptr<nvinfer1::ICudaEngine>& engine,
                    const std::string& name, EngineBinding* binding) noexcept
{
    try
    {
        binding->_name = name;
    }
    catch(const std::exception& e)
    {
        return false;
    }

    binding->_index = engine->getBindingIndex(name.c_str());
    if(binding->_index == -1)
    {
        return false;
    }

    binding->_dims = engine->getBindingDimensions(binding->_index);
    binding->_volume = dimsVolume(binding->_dims);

    binding->_isInput = engine->bindingIsInput(binding->_index);

    return true;
}

bool EngineBinding::setup(const std::unique_ptr<nvinfer1::ICudaEngine>& engine,
                    const int& index, EngineBinding* binding) noexcept
{
    binding->_index = index;
    const char* name = engine->getBindingName(index);
    if(name == nullptr)
    {
        return false;
    }

    try
    {
        binding->_name = std::string(name);
    }
    catch(const std::exception& e)
    {
        return false;
    }

    binding->_dims = engine->getBindingDimensions(binding->_index);
    binding->_volume = dimsVolume(binding->_dims);

    binding->_isInput = engine->bindingIsInput(binding->_index);

    return true;
}

DeviceMemory::DeviceMemory() noexcept
{
}

DeviceMemory::~DeviceMemory() noexcept
{
    for(unsigned int i = 0; i < _memory.size(); ++i)
    {
        if(_memory[i])
        {
            cudaFree(_memory[i]);
        }
    }
    _memory.clear();
}

void DeviceMemory::swap(DeviceMemory& other) noexcept
{
    std::swap(_memory, other._memory);
}

void** DeviceMemory::begin() const noexcept
{
    return (void**)_memory.data();
}

void* DeviceMemory::at(const int& index) const noexcept
{
    return _memory[index];
}

Result DeviceMemory::setup(const std::shared_ptr<Logger>& logger, 
                    std::unique_ptr<nvinfer1::ICudaEngine>& engine,
                    DeviceMemory* output) noexcept
{
    const int32_t nbBindings = engine->getNbBindings();
    for(int i = 0; i < nbBindings; ++i)
    {
        const nvinfer1::Dims dims = engine->getBindingDimensions(i); 
        const int volume = dimsVolume(dims);

        try
        {
            output->_memory.push_back(nullptr);
        }
        catch(const std::exception& e)
        {
            logger->logf(LOGGING_ERROR, "[DeviceMemory] setup() failure: "
                        "exception: %s", e.what()); 
            return RESULT_FAILURE_ALLOC;
        }
        void** ptr = &output->_memory.back();

        auto r = cudaMalloc(ptr, volume * sizeof(float));
        if(r != 0 || *ptr == nullptr)
        {
            logger->logf(LOGGING_ERROR, "[DeviceMemory] setup() failure: "
                        "could not allocate device memory: %s", 
                        cudaGetErrorString(r));
            return RESULT_FAILURE_CUDA_ERROR;
        }

    }
    return RESULT_SUCCESS;
}

bool opencvHasCuda() noexcept
{
    int r = 0;
    try
    {
        r = cv::cuda::getCudaEnabledDeviceCount();
    }
    catch(const std::exception& e)
    {
        return 0;
    }
	return (r > 0);
}

PreprocessorTransform::PreprocessorTransform() noexcept
    : _inputSize(0, 0), _f(1), _leftWidth(0), _topHeight(0)
{
}

PreprocessorTransform::PreprocessorTransform(const cv::Size& inputSize, 
                        const double& f, const int& leftWidth, 
                        const int& topHeight) noexcept
    : _inputSize(inputSize), _f(f), 
        _leftWidth(leftWidth), _topHeight(topHeight)
{
}

PreprocessorTransform::~PreprocessorTransform() noexcept
{
}

cv::Rect PreprocessorTransform::transformBbox(
                                const cv::Rect& input) const noexcept
{
    cv::Rect r;
    r.x = (input.x - _leftWidth) / _f;
    r.x = MAX(0, MIN(r.x, _inputSize.width-1));

    r.y = (input.y - _topHeight) / _f;
    r.y = MAX(0, MIN(r.y, _inputSize.height-1));

    r.width = input.width / _f;
    if(r.x + r.width > _inputSize.width)
    {
        r.width = _inputSize.width - r.x;
    }
    r.height = input.height / _f;
    if(r.y + r.height > _inputSize.height)
    {
        r.height = _inputSize.height - r.y;
    }
    return r;
}


Preprocessor::Preprocessor() noexcept
{
}

Preprocessor::~Preprocessor() noexcept
{
}

void Preprocessor::setLogger(std::shared_ptr<Logger> logger) noexcept
{
    _logger = logger;
}

bool Preprocessor::process(const int& index, const cv::Mat& input,
                        const bool& last) noexcept
{
    YOLOV5_UNUSED(input);
    YOLOV5_UNUSED(last);

    if(_transforms.size() < (unsigned int)index+1)
    {
        try
        {
            _transforms.resize(index + 1);
        }
        catch(const std::exception& e)
        {
            _logger->logf(LOGGING_ERROR, "[Preprocessor] process() "
                            "failure: got exception setting up transforms: "
                            "%s", e.what());
            return false;
        }
    }
    return true;
}

bool Preprocessor::process(const int& index, 
            const cv::cuda::GpuMat& input, const bool& last) noexcept
{
    YOLOV5_UNUSED(input);
    YOLOV5_UNUSED(last);

    if(_transforms.size() < (unsigned int)index+1)
    {
        try
        {
            _transforms.resize(index + 1);
        }
        catch(const std::exception& e)
        {
            _logger->logf(LOGGING_ERROR, "[Preprocessor] process() "
                            "failure: got exception setting up transforms: "
                            "%s", e.what());
            return false;
        }
    }
    return true;
}

cv::Rect Preprocessor::transformBbox(const int& index, 
                    const cv::Rect& bbox) const noexcept
{
    return _transforms[index].transformBbox(bbox);
}

template <typename T>
static void setupChannels(const cv::Size& size, 
                        const Preprocessor::InputType& inputType, 
                        float* inputPtr,
                        std::vector<T>& channels)
{
    const int channelSize = size.area();
    if(inputType == Preprocessor::INPUTTYPE_BGR)  /*  INPUT_BGR   */
    {
        /*  B channel will go here  */
        channels.push_back(T(size, CV_32FC1, inputPtr + 2 * channelSize));
        /*  G channel will go here  */
        channels.push_back(T(size, CV_32FC1, inputPtr + 1 * channelSize));
        /*  R channel will go here  */
        channels.push_back(T(size, CV_32FC1, inputPtr));
    }
    else    /*  INPUTTYPE_RGB   */
    {
        /*  R channel will go here  */
        channels.push_back(T(size, CV_32FC1, inputPtr));
        /*  G channel will go here  */
        channels.push_back(T(size, CV_32FC1, inputPtr + 1 * channelSize));
        /*  B channel will go here  */
        channels.push_back(T(size, CV_32FC1, inputPtr + 2 * channelSize));
    }
}

CvCpuPreprocessor::CvCpuPreprocessor() noexcept 
    : _cudaStream(nullptr), _lastType((InputType)-1), _lastBatchSize(-1), 
    _networkCols(0), _networkRows(0)
{
}

CvCpuPreprocessor::~CvCpuPreprocessor() noexcept
{
}

bool CvCpuPreprocessor::setup(const nvinfer1::Dims& inputDims,
                            const int& flags, const int& batchSize,
                            float* inputMemory) noexcept
{
    if(!_cudaStream)
    {
        auto r = cudaStreamCreate(&_cudaStream);
        if(r != 0)
        {
            _logger->logf(LOGGING_ERROR, "[CvCpuPreprocessor] setup() "
                        "failure: could not create cuda stream: %s", 
                        cudaGetErrorString(r));
            return false;
        }
    }

    if((flags & INPUT_RGB) && (flags & INPUT_BGR))
    {
        _logger->log(LOGGING_ERROR, "[CvCpuPreprocessor] setup() "
                    "failure: both INPUT_RGB and INPUT_BGR flags specified"); 
        return false;
    }
    InputType inputType = INPUTTYPE_BGR;
    if(flags & INPUT_RGB)
    {
        inputType = INPUTTYPE_RGB;
    }
    
    if(_lastType == inputType && _lastBatchSize == batchSize)
    {
        return true;
    }
    _lastType = inputType;
    _lastBatchSize = batchSize;

    _networkRows = inputDims.d[2];
    _networkCols = inputDims.d[3];
    const cv::Size networkSize(_networkCols, _networkRows);

    _deviceInputMemory = inputMemory;

    _inputChannels.clear();
    try
    {
        /*  Set up host input memory    */
        _hostInputMemory.resize(dimsVolume(inputDims));

        _inputChannels.resize(batchSize);
        for(unsigned int i = 0; i < _inputChannels.size(); ++i)
        {
            float* inputPtr = _hostInputMemory.data() 
                        + i * networkSize.area() * 3;

            std::vector<cv::Mat>& channels = _inputChannels[i];
            setupChannels<cv::Mat>(networkSize, inputType, inputPtr, channels);
        }
    }
    catch(const std::exception& e)
    {
        _logger->logf(LOGGING_ERROR, "[CvCpuPreprocessor] setup() failure: "
                    "got exception while trying to set up input channels: %s",
                    e.what());
        return false;
    }
    return true;
}

void CvCpuPreprocessor::reset() noexcept
{
    /*  this will trigger setup() to take effect next time  */
    _lastType = (InputType)-1;
}

bool CvCpuPreprocessor::process(const int& index, 
                                const cv::Mat& input,
                                const bool& last) noexcept
{
    if(!Preprocessor::process(index, input, last))
    {
        return false;
    }

    PreprocessorTransform& transform = _transforms[index];
    try
    {
        if(input.rows == _networkRows && input.cols == _networkCols)
        {
            input.convertTo(_buffer3, CV_32FC3, 1.0f / 255.0f);
            transform = PreprocessorTransform(input.size(), 1.0, 0, 0);
        }
        else
        {
            const double f =
                    MIN((double)_networkRows / (double)input.rows,
                        (double)_networkCols / (double)input.cols);
            const cv::Size boxSize = cv::Size(input.cols * f,
                                            input.rows * f);

            const int dr = _networkRows - boxSize.height;
            const int dc = _networkCols - boxSize.width;
            const int topHeight = std::floor(dr / 2.0);
            const int bottomHeight = std::ceil(dr / 2.0);
            const int leftWidth = std::floor(dc / 2.0);
            const int rightWidth = std::ceil(dc / 2.0);

            transform = PreprocessorTransform(input.size(), 
                                            f, leftWidth, topHeight);

            cv::resize(input, _buffer1, boxSize, 0, 0, cv::INTER_LINEAR);
            cv::copyMakeBorder(_buffer1, _buffer2, 
                            topHeight, bottomHeight, leftWidth, rightWidth,
                            cv::BORDER_CONSTANT, cv::Scalar(0, 0, 0));
            _buffer2.convertTo(_buffer3, CV_32FC3, 1.0f / 255.0f);
        }
        cv::split(_buffer3, _inputChannels[index]);
    }
    catch(const std::exception& e)
    {
        _logger->logf(LOGGING_ERROR, "[CvCpuPreprocessor] process() "
                            "failure: got exception setting up input: "
                            "%s", e.what());
        return false;
    }

    /*  Copy from host to device    */
    if(last)
    {
        const int volume = _inputChannels.size()    /*  batch size  */
                * 3 * _buffer3.size().area();   /*  channels * rows*cols   */

        auto r = cudaMemcpyAsync(_deviceInputMemory, 
                (void*)_hostInputMemory.data(), 
                (int)(volume * sizeof(float)),
                cudaMemcpyHostToDevice, _cudaStream);
        if(r != 0)
        {
            _logger->logf(LOGGING_ERROR, "[CvCpuPreprocessor] process() "
                        "failure: could not set up host-to-device transfer "
                        "for input: %s", cudaGetErrorString(r));
            return false;
        }
    }
    return true;
}

bool CvCpuPreprocessor::process(const int& index, 
                                const cv::cuda::GpuMat& input,
                                const bool& last) noexcept
{
    YOLOV5_UNUSED(index);
    YOLOV5_UNUSED(input);
    YOLOV5_UNUSED(last);
    _logger->log(LOGGING_ERROR, "[CvCpuPreprocessor] process() failure: CUDA "
                    "method is not supported. Use CPU variant instead");
    /*  not supported   */
    return false;
}

cudaStream_t CvCpuPreprocessor::cudaStream() const noexcept
{
    return _cudaStream;
}

bool CvCpuPreprocessor::synchronizeCudaStream() noexcept
{
    auto r = cudaStreamSynchronize(_cudaStream); 
    if(r != 0)
    {
        _logger->logf(LOGGING_ERROR, "[CvCpuPreprocessor] "
                "synchronizeCudaStream() failure: %s", cudaGetErrorString(r));
        return false;
    }
    return true;
}

CvCudaPreprocessor::CvCudaPreprocessor() noexcept 
    : _lastType((InputType)-1), _lastBatchSize(-1), 
    _networkCols(0), _networkRows(0)
{
}

CvCudaPreprocessor::~CvCudaPreprocessor() noexcept
{
}

bool CvCudaPreprocessor::setup(const nvinfer1::Dims& inputDims,
                            const int& flags, const int& batchSize,
                            float* inputMemory) noexcept
{
#ifdef YOLOV5_OPENCV_HAS_CUDA
    if((flags & INPUT_RGB) && (flags & INPUT_BGR))
    {
        _logger->log(LOGGING_ERROR, "[CvCudaPreprocessor] setup() "
                    "failure: both INPUT_RGB and INPUT_BGR flags specified"); 
        return false;
    }
    InputType inputType = INPUTTYPE_BGR;
    if(flags & INPUT_RGB)
    {
        inputType = INPUTTYPE_RGB;
    }

    if(_lastType == inputType && _lastBatchSize == batchSize)
    {
        return true;
    }
    _lastType = inputType;
    _lastBatchSize = batchSize;

    _networkRows = inputDims.d[2];
    _networkCols = inputDims.d[3];
    const cv::Size networkSize(_networkCols, _networkRows);

    _inputChannels.clear();
    try
    {
        _inputChannels.resize(batchSize);
        for(unsigned int i = 0; i < _inputChannels.size(); ++i)
        {
            float* inputPtr = inputMemory + i * networkSize.area() * 3;

            std::vector<cv::cuda::GpuMat>& channels = _inputChannels[i];
            setupChannels<cv::cuda::GpuMat>(networkSize, inputType, inputPtr, 
                                            channels);
        }
    }
    catch(const std::exception& e)
    {
        _logger->logf(LOGGING_ERROR, "[CvCudaPreprocessor] setup() failure: "
                    "got exception while trying to set up input channels: %s",
                    e.what());
        return false;
    }
    return true;
#else
    YOLOV5_UNUSED(inputDims);
    YOLOV5_UNUSED(flags);
    YOLOV5_UNUSED(batchSize);
    YOLOV5_UNUSED(inputMemory);
    _logger->log(LOGGING_ERROR, "[CvCudaPreprocessor] setup() failure: "
                    "OpenCV without CUDA support");
    return false;
#endif
}

void CvCudaPreprocessor::reset() noexcept
{
    /*  this will trigger setup() to take effect next time  */
    _lastType = (InputType)-1;
}

bool CvCudaPreprocessor::process(const int& index, 
                        const cv::Mat& input, const bool& last) noexcept
{
#ifdef YOLOV5_OPENCV_HAS_CUDA
    try
    {
        _buffer0.upload(input);
    }
    catch(const std::exception& e)
    {
        _logger->logf(LOGGING_ERROR, "[CvCudaPreprocessor] process() "
                            "failure: got exception trying to upload input "
                            "to CUDA device: %s", e.what());
        return false;
    }
    return process(index, _buffer0, last);
#else
    YOLOV5_UNUSED(index);
    YOLOV5_UNUSED(input);
    YOLOV5_UNUSED(last);
    _logger->log(LOGGING_ERROR, "[CvCudaPreprocessor] process() failure: "
                    "OpenCV without CUDA support");
    return false;
#endif
}

bool CvCudaPreprocessor::process(const int& index, 
                        const cv::cuda::GpuMat& input,
                        const bool& last) noexcept
{
#ifdef YOLOV5_OPENCV_HAS_CUDA
    if(!Preprocessor::process(index, input, last))
    {
        return false;
    }

    if(index >= 1)
    {
        /*  If we're processing a batch, first process the previous image
            before changing the internal buffers    */
        if(!synchronizeCudaStream())
        {
            return false;
        }
    }

    PreprocessorTransform& transform = _transforms[index];
    try
    {
        if(input.rows == _networkRows && input.cols == _networkCols)
        {
            input.convertTo(_buffer3, CV_32FC3, 1.0f / 255.0f, _cudaStream);
            transform = PreprocessorTransform(input.size(), 1.0, 0, 0);
        }
        else
        {
            const double f =
                    MIN((double)_networkRows / (double)input.rows,
                        (double)_networkCols / (double)input.cols);
            const cv::Size boxSize = cv::Size(input.cols * f,
                                            input.rows * f);

            const int dr = _networkRows - boxSize.height;
            const int dc = _networkCols - boxSize.width;
            const int topHeight = std::floor(dr / 2.0);
            const int bottomHeight = std::ceil(dr / 2.0);
            const int leftWidth = std::floor(dc / 2.0);
            const int rightWidth = std::ceil(dc / 2.0);

            transform = PreprocessorTransform(input.size(), 
                                        f, leftWidth, topHeight);

            cv::cuda::resize(input, _buffer1, boxSize, 0, 0, cv::INTER_LINEAR,
                            _cudaStream);
            cv::cuda::copyMakeBorder(_buffer1, _buffer2, 
                            topHeight, bottomHeight, leftWidth, rightWidth,
                            cv::BORDER_CONSTANT, cv::Scalar(0, 0, 0), 
                            _cudaStream);
            _buffer2.convertTo(_buffer3, CV_32FC3, 1.0f / 255.0f, _cudaStream);
        }
        cv::cuda::split(_buffer3, _inputChannels[index], _cudaStream);
    }
    catch(const std::exception& e)
    {
        _logger->logf(LOGGING_ERROR, "[CvCudaPreprocessor] process() "
                            "failure: got exception setting up input: "
                            "%s", e.what());
        return false;
    }
    return true;
#else
    YOLOV5_UNUSED(index);
    YOLOV5_UNUSED(input);
    YOLOV5_UNUSED(last);
    _logger->log(LOGGING_ERROR, "[CvCudaPreprocessor] process() failure: "
                    "OpenCV without CUDA support");
    return false;
#endif
}

cudaStream_t CvCudaPreprocessor::cudaStream() const noexcept
{
#ifdef YOLOV5_OPENCV_HAS_CUDA
    return cv::cuda::StreamAccessor::getStream(_cudaStream);
#else
    return nullptr;
#endif
}

bool CvCudaPreprocessor::synchronizeCudaStream() noexcept
{
    try
    {
        _cudaStream.waitForCompletion();
    }
    catch(const std::exception& e)
    {
        _logger->logf(LOGGING_ERROR, "[CvCudaPreprocessor] "
                    "synchronizeCudaStream() failure: %s", e.what());
        return false;
    }
    return true;
}


}   /*  namespace internal  */

}   /*  namespace yolov5    */

@SivaRamaKrishnaNV

4

There is no update from you for a period, assuming this is not an issue any more.
Hence we are closing this topic. If need further support, please open a new one.
Thanks

Dear @nehal,
Just want to confirm if you are using DRIVE AGX Orin or DRIVE AGX Xavier platfrom? It is not clear from the description about the platform.

7