在C++中将pytorch张量转换为opencv mat，反之亦然

Question

我想在 C++ 中将 pytorch 张量转换为 opencv mat，反之亦然。 我有这两个功能：

cv::Mat TensorToCVMat(torch::Tensor tensor)
{
    std::cout << "converting tensor to cvmat\n";
    tensor = tensor.squeeze().detach().permute({1, 2, 0});
    tensor = tensor.mul(255).clamp(0, 255).to(torch::kU8);
    tensor = tensor.to(torch::kCPU);
    int64_t height = tensor.size(0);
    int64_t width = tensor.size(1);
    cv::Mat mat(width, height, CV_8UC3);
    std::memcpy((void *)mat.data, tensor.data_ptr(), sizeof(torch::kU8) * tensor.numel());
    return mat.clone();
}

torch::Tensor CVMatToTensor(cv::Mat mat)
{
    std::cout << "converting cvmat to tensor\n";
    cv::cvtColor(mat, mat, cv::COLOR_BGR2RGB);
    cv::Mat matFloat;
    mat.convertTo(matFloat, CV_32F, 1.0 / 255);
    auto size = matFloat.size();
    auto nChannels = matFloat.channels();
    auto tensor = torch::from_blob(matFloat.data, {1, size.height, size.width, nChannels});
    return tensor.permute({0, 3, 1, 2});
}

在我的代码中，我加载了两个图像（ image1和image2 ），我想将它们转换为 pytorch 张量，然后返回到 opencv mat 以检查它是否有效。 问题是我在第一次调用TensorToCVMat内存访问错误，我无法弄清楚出了什么问题，因为我对 C++ 编程没有太多经验。

cv::Mat image1;
image1 = cv::imread(argv[1]);
if (!image1.data)
{
    std::cout << "no image data\n";
    return -1;
}
cv::Mat image2;
image2 = cv::imread(argv[2]);
if (!image2.data)
{
    std::cout << "no image data\n";
    return -1;
}

torch::Tensor tensor1 = CVMatToTensor(image1);
cv::Mat new_image1 = TensorToCVMat(tensor1); // <<< this is where the memory access error is thrown
torch::Tensor tensor2 = CVMatToTensor(image2);
cv::Mat new_image2 = TensorToCVMat(tensor2);

如果你能给我提示或解释来解决这个问题，那就太好了。

Answer 1

不确定错误是否发生在 memcpy 步骤。 但是您可以使用Mat构造函数的void* data变体

Mat (int rows, int cols, int type, void *data, size_t step=AUTO_STEP)

你可以跳过 memcpy 步骤

tensor = uint8_tensor //shape: (h, w, 3)
cv::Mat mat = cv::Mat(height, width, CV_8UC3, tensor.data_ptr());
return mat;

Answer 2

我正在使用火炬>=1.7.0。

对于dtype=float和大小[1, 3, height, width]的张量，这对我有用

cv::Mat torchTensortoCVMat(torch::Tensor& tensor)
    {
        tensor = tensor.squeeze().detach();
        tensor = tensor.permute({1, 2, 0}).contiguous();
        tensor = tensor.mul(255).clamp(0, 255).to(torch::kU8);
        tensor = tensor.to(torch::kCPU);
        int64_t height = tensor.size(0);
        int64_t width = tensor.size(1);
        cv::Mat mat = cv::Mat(cv::Size(width, height), CV_8UC3, tensor.data_ptr<uchar>());
        return mat.clone();
    }

Answer 3

我的张量形状是500x500x3 ，我必须添加tensor.reshape({width * height * 3})以获得实际图像

cv::Mat TensorToCVMat(torch::Tensor tensor)
{
    // torch.squeeze(input, dim=None, *, out=None) → Tensor
    // Returns a tensor with all the dimensions of input of size 1 removed.
    // tensor.detach
    // Returns a new Tensor, detached from the current graph.
    // permute dimension, 3x700x700 => 700x700x3
    tensor = tensor.detach().permute({1, 2, 0});
    // float to 255 range
    tensor = tensor.mul(255).clamp(0, 255).to(torch::kU8);
    // GPU to CPU?, may not needed
    tensor = tensor.to(torch::kCPU);
    // shape of tensor
    int64_t height = tensor.size(0);
    int64_t width = tensor.size(1);

    // Mat takes data form like {0,0,255,0,0,255,...} ({B,G,R,B,G,R,...})
    // so we must reshape tensor, otherwise we get a 3x3 grid
    tensor = tensor.reshape({width * height * 3});
    // CV_8UC3 is an 8-bit unsigned integer matrix/image with 3 channels
    cv::Mat imgbin(cv::Size(width, height), CV_8UC3, tensor.data_ptr());

    return imgbin;
}