cv::MAT如何转换为NHCW格式?

How cv::MAT convert to NHCW format?

在用户Guide.html中,tensorRT的Input/output需要使用NCHW格式。
NCHW 格式是什么?
如何将 cv::MAT 转换为 NCHW 格式?

我 运行 像下面的代码一样使用 TensorRT 进行推理。
没有错误。但是,这不是正确的输出结果。

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);

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

NHWC:对于每个像素,其 3 种颜色按 BGR 顺序存储在一起。

TensorRT 要求您的图像数据采用 NCHW 顺序。但是 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 读取到一个缓冲区,您可以在缓冲区中按 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 顺序并减去平均值,这两个步骤都在一个步骤中完成。您可以修改它以满足您的目的。

此代码片段按照 Ashwin 的解释进行转换

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

// 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));

这是另一个基于 OpenCV 的简单答案:

cv::resize(img, img, input_size);
img.convertTo(img, CV_32FC3, 1.f / 255.f);
vector<float> chw_buffer(img.elemSize() * img.total());
std::vector<Mat> chw;
for (size_t n = 0; n < channels; ++n)
    chw.emplace_back(Mat(input_size, CV_32FC1, chw_buffer.data() + n * input_width * input_height));
cv::split(img, chw);
cudaMemcpyAsync(gpu_buffers[inputIndex], chw_buffer.size(), cudaMemcpyHostToDevice, stream);

这是 CPU 上 NHWC 到 NCHW 的简短直接解决方案:

static void hwc_to_chw(cv::InputArray src, cv::OutputArray dst) {
  std::vector<cv::Mat> channels;
  cv::split(src, channels);

  // Stretch one-channel images to vector
  for (auto &img : channels) {
    img = img.reshape(1, 1);
  }

  // Concatenate three vectors to one
  cv::hconcat( channels, dst );
}