C++/Python Type Mapping

dict

We map Python's dict to C++'s dict:

#ifndef CUSTOM_DICT
using dict = std::shared_ptr<std::unordered_map<std::string, ipipe::any>>;
#else
// todo :  custom dict with bidirectional binding between C++ and Python
#endif

C++ to Python

Basic Type Conversion

cpp	python	Remarks
std::string	bytes
cv::Mat	numpy.ndarray
at::Tensor	torch.Tensor
int	int
float	float
double	float
unsigned int	int
char	str	Inherited from Pybind11.
unsigned char	int
bool	bool

Container Conversion

For containers or nested containers, the conversion rules are as follows (where T is the basic type from the above table):

cpp	python	Remarks
std::unordered_map<std::string, T>	dict[str, T]
std::unordered_set<T>	set[T]	T must be hashable
std::vector<T>	list[T]	T is not an algebraic type
std::vector<T>	numpy.ndarray[T]	T is an algebraic type and not bool
std::vector<std::vector<T>>	list[numpy.ndarray[T]]	T is an algebraic type and not bool
std::map<std::string, T>	Not supported
std::set<T>	Not supported

Nested Types

The conversion of any to Python types also supports the following nested types:

cpp	python	Remarks
std::vector<any>	list[Any]	Can be nested infinitely
std::unordered_map<std::string, any>	dict[str, Any]	Can be nested infinitely, 'data' key will be ignored

C++ Wrapper

For non-containers that fail to convert using the above rules, a custom type will be returned to Python:

cpp	python	Remarks
any	torchpipe.any

Python to C++

Basic Type Conversion

python	cpp	Remarks
str	std::string
bytes	std::string
numpy.ndarray	Not supported
torch.Tensor	at::Tensor
bool	bool
float	float	precision loss
int	int

Container Conversion

For containers or nested containers, the conversion table is as follows (where T is the basic type from the above table):

python	cpp	Remarks
dict[str, T]	std::unordered_map<std::string, T>
set[T]	std::unordered_set<T>	T must be hashable, i.e. T != torch.Tensor
list[T]	std::vector<T>

When the container on the Python side is empty, the C++ side cannot perceive the specific type. In this case, the framework will automatically handle such situations in a compatible way, so there is no need to pay extra attention to this scenario when using it.