Enhance API Design

[DapengFeng]

Hi @andrewssobral , inspired by your excellent repo:), I have implemented a similar Adapter Class. Do you have any suggestions for API design enhancement?

template <typename T,
          uint32_t channels_ = 1,
          typename = std::enable_if_t<std::is_arithmetic_v<T>>>
class Adapter {
 public:
  using Tensor = torch::Tensor;
  using Mat = cv::Mat;
  using MatrixColMajor = Eigen::Matrix<T, Eigen::Dynamic, Eigen::Dynamic>;
  using MatrixRowMajor =
      Eigen::Matrix<T, Eigen::Dynamic, Eigen::Dynamic, Eigen::RowMajor>;
  explicit Adapter(const Tensor& tensor) {
    Tensor cpu_tensor;
    if constexpr (channels_ == 1) {
      cpu_tensor = tensor.detach().to(torch::kCPU, true).contiguous();
    } else {
      cpu_tensor = tensor.detach().to(torch::kCPU, true).permute({1, 2, 0}).contiguous();
    }
    data_ptr_ = cpu_tensor.data_ptr();
    rows_ = cpu_tensor.size(0);
    cols_ = cpu_tensor.size(1);
  }

  explicit Adapter(const Mat& mat) {
    data_ptr_ = mat.data;
    rows_ = mat.rows;
    cols_ = mat.cols;
  }

  explicit Adapter(const MatrixColMajor& mat) {
    data_ptr_ = const_cast<T*>(mat.data());
    rows_ = mat.cols();
    cols_ = mat.rows();
    is_raw_ = false;
  }

  inline Tensor toTensor(const bool copy = true) {
    Tensor tensor;
    if constexpr (channels_ == 1) {
      tensor = torch::from_blob(data_ptr_, {rows_, cols_}, torch::TensorOptions(torch::CppTypeToScalarType<T>()));
    } else {
      tensor = torch::from_blob(data_ptr_, {rows_, cols_, channels_},
                                torch::TensorOptions(torch::CppTypeToScalarType<T>()))
                   .permute({2, 0, 1})
                   .contiguous();
    }
    if (!is_raw_) {
      tensor = tensor.mT();
    }
    if (copy) {
      return tensor.clone();
    } else {
      return tensor;
    }
  }

  template <typename = std::enable_if_t<std::is_floating_point_v<T> && sizeof(T) == sizeof(float)>>
  inline Mat toCvMat(const bool copy = true) {
    Mat mat(rows_, cols_, CV_32FC(channels_), data_ptr_);
    if (!is_raw_) {
      mat = mat.t();
    }
    if (copy) {
      return mat.clone();
    } else {
      return mat;
    }
  }

  template <typename = std::enable_if_t<channels_ == 1>>
  inline MatrixColMajor toEigenMatrix() {
    if (!is_raw_) {
      return Eigen::Map<MatrixColMajor>(reinterpret_cast<T*>(data_ptr_), cols_, rows_);
    }
    return Eigen::Map<MatrixRowMajor>(reinterpret_cast<T*>(data_ptr_), rows_, cols_);
  }

 private:
  void* data_ptr_;
  uint32_t rows_;
  uint32_t cols_;
  bool is_raw_ = true;
};

Note

• Shadow copy of Eigen Matrix.
• High dimensional tensor.
• Host (cpu) and Device (gpu, npu, and others).