Question

在User Guide.html中，tensorRT的输入/输出需要使用NCHW格式什么是NCHW fomat？
如何将cv :: MAT转换为NCHW格式？

我使用TensorRT进行推理，如下面的代码没什么错误。但是，这不是输出的结果。

int batchSize = 1;
int size_of_single_input = 256 * 256 * 3 * sizeof(float);
int size_of_single_output = 100 * 1 * 1 * sizeof(float); 

IBuilder* builder = createInferBuilder(gLogger);

INetworkDefinition* network = builder->createNetwork();

CaffeParser parser;
auto blob_name_to_tensor = parser.parse(“deploy.prototxt”,
                                        "sample.caffemodel",
                                        *network,
                                        DataType::kFLOAT);

network->markOutput(*blob_name_to_tensor->find("prob"));

builder->setMaxBatchSize(1);
builder->setMaxWorkspaceSize(1 << 30); 
ICudaEngine* engine = builder->buildCudaEngine(*network);

IExecutionContext *context = engine->createExecutionContext();

int inputIndex = engine->getBindingIndex(INPUT_LAYER_NAME),
int outputIndex = engine->getBindingIndex(OUTPUT_LAYER_NAME);

cv::Mat input;
input = imread("./sample.jpg");
cvtColor(input, input, CV_BGR2RGB);
cv::resize(input, input, cv::Size(256, 256));

float output[OUTPUTSIZE];

void* buffers = malloc(engine->getNbBindings() * sizeof(void*));
cudaMalloc(&buffers[inputIndex], batchSize * size_of_single_input);
cudaMalloc(&buffers[outputIndex], batchSize * size_of_single_output);

cudaStream_t stream;
cudaStreamCreate(&stream);

cudaMemcpyAsync(buffers[inputIndex], (float *)input, 
                batchSize * size_of_single_input, 
                cudaMemcpyHostToDevice, stream);

context.enqueue(batchSize, buffers, stream, nullptr);


cudaMemcpyAsync(output, buffers[outputIndex], 
                batchSize * size_of_single_output, 
                cudaMemcpyDeviceToHost, stream));

cudaStreamSynchronize(stream);

Answer 1

NCHW：对于3通道图像，比如BGR，首先存储B通道的像素，然后存储G通道，最后存储R通道。

NHWC：对于每个像素，它的3种颜色以BGR顺序存储在一起。

TensorRT要求您的图像数据为NCHW顺序。但OpenCV以NHWC顺序读取它。您可以编写一个简单的函数来将数据从NHWC读取到缓冲区，然后以NCHW顺序存储它们。将此缓冲区复制到设备内存并传递给TensorRT。

您可以在TensorRT安装的samples/sampleFasterRCNN/sampleFasterRCNN.cpp文件中找到此操作的示例。它读取PPM文件，该文件也是NHWC顺序，然后将其转换为NCHW顺序，并在一个步骤中减去平均值。您可以根据自己的目的进行修改。

Answer 2

此代码段按照Ashwin的说明进行了转换

bool SampleUffSSD::processInput(const samplesCommon::BufferManager& buffers)
     const int batchSize = mParams.batchSize;
 
     // Available images


     std::vector<std::string> imageList = {"test.jpeg"};
     mPPMs.resize(batchSize);
     assert(mPPMs.size() <= imageList.size());
     for (int i = 0; i < batchSize; ++i)
     {

        readImage(locateFile(imageList[i], mParams.dataDirs), image);
     }
 
     float* hostDataBuffer = static_cast<float*>(buffers.getHostBuffer(mParams.inputTensorNames[0]));
     // Host memory for input buffer

    for (int i = 0, volImg = inputH * inputW; i < mParams.batchSize; ++i)
     {

        for (unsigned j = 0, volChl = inputH * inputW; j < inputH; ++j)
         {

                   for( unsigned k = 0; k < inputW; ++ k)
                       {
                cv::Vec3b bgr = image.at<cv::Vec3b>(j,k);
                hostDataBuffer[i * volImg + 0 * volChl + j * inputW + k] = (2.0 / 255.0) * float(bgr[2]) - 1.0;
                hostDataBuffer[i * volImg + 1 * volChl + j * inputW + k] = (2.0 / 255.0) * float(bgr[1]) - 1.0;
                hostDataBuffer[i * volImg + 2 * volChl + j * inputW + k] = (2.0 / 255.0) * float(bgr[0]) - 1.0;
             }
         }
     }

来源：https://forums.developer.nvidia.com/t/custom-trained-ssd-inception-model-in-tensorrt-c-version/143048/14

Answer 3

// suppose all data types are int.
// size of mat is 256*256*3.

cv::Mat NCHW,NHWC;

std::vector<cv::Mat> channels;
split(NHWC, channels);

memcpy(NCHW.data,channels[0].data,256*256*sizeof(int));
memcpy(NCHW.data+256*256,channels[1].data,256*256*sizeof(int));
memcpy(NCHW.data+2*256*256,channels[2].data,256*256*sizeof(int));

cv :: MAT如何转换为NHCW格式？

3 个答案: