向下转换为 pybind11 派生类

Downcast to pybind11 derived class

本文关键字:派生 pybind11 转换      更新时间:2023-10-16

我正在使用pybind11的"覆盖Python中的虚拟函数"功能来创建从C++抽象类继承的Python类。我有一个C++类State它在 Python 中被子类化为MyState.在这种情况下,我有一些MyState对象丢失了其类型信息,Python 认为这是一个State。我需要将其向下转换为 Python 代码中的MyState,但我不知道这样做的好方法。

下面是C++示例代码:

#include <memory>
#include <pybind11/pybind11.h>
namespace py = pybind11;
// ========== State ==========
class State {
public:
virtual ~State() = default;
virtual void dump() = 0;
};
using StatePtr = std::shared_ptr<State>;
class PyState : public State {
public:
using State::State;
void dump() override {
PYBIND11_OVERLOAD_PURE(void, State, dump);
}
};
// ========== Machine ==========
class Machine {
public:
virtual ~Machine() = default;
virtual StatePtr begin() = 0;
virtual StatePtr step(const StatePtr&) = 0;
};
using MachinePtr = std::shared_ptr<Machine>;
class PyMachine : public Machine {
public:
using Machine::Machine;
StatePtr begin() override {
PYBIND11_OVERLOAD_PURE(StatePtr, Machine, begin);
}
StatePtr step(const StatePtr& state) override {
PYBIND11_OVERLOAD_PURE(StatePtr, Machine, step, state);
}
};
// ========== run ==========
void run(const MachinePtr& machine) {
StatePtr state = machine->begin();
for (int i = 0; i < 5; ++i) {
state = machine->step(state);
state->dump();
}
}
// ========== pybind11 ==========
PYBIND11_MODULE(example, m) {
py::class_<State, StatePtr, PyState>(m, "State").def(py::init<>());
py::class_<Machine, MachinePtr, PyMachine>(m, "Machine")
.def(py::init<>())
.def("begin", &Machine::begin)
.def("step", &Machine::step);
m.def("run", &run, "Run the machine");
}

还有 Python 代码:

#!/usr/bin/env python3
from example import Machine, State, run

class MyState(State):
def __init__(self, x):
State.__init__(self)
self.x = x
def dump(self):
print(self.x)

class MyMachine(Machine):
def __init__(self):
Machine.__init__(self)
def begin(self):
return MyState(0)
def step(self, state):
# problem: when called from C++, `state` is an `example.State`
# instead of `MyState`. In order to access `state.x` we need
# some way to downcast it...
return MyState(state.x + 1)

machine = MyMachine()
print("running machine with python")
state = machine.begin()
for _ in range(5):
state = machine.step(state)
state.dump()
print("running machine with C++")
run(machine)  # error

错误信息:

running machine with python
1
2
3
4
5
running machine with C++
Traceback (most recent call last):
File "<string>", line 38, in <module>
File "<string>", line 36, in __run
File "/usr/local/fbcode/platform007/lib/python3.6/runpy.py", line 193, in _run_module_as_main
"__main__", mod_spec)
File "/usr/local/fbcode/platform007/lib/python3.6/runpy.py", line 85, in _run_code
exec(code, run_globals)
File "/data/users/jcai/fbsource/fbcode/buck-out/dev/gen/experimental/jcai/pybind/run_example#link-tree/run_example.py", line 38, in <module>
run(machine)  # error
File "/data/users/jcai/fbsource/fbcode/buck-out/dev/gen/experimental/jcai/pybind/run_example#link-tree/run_example.py", line 26, in step
return MyState(state.x + 1)
AttributeError: 'example.State' object has no attribute 'x'

我确实有一个笨拙的解决方法,我基本上保留了一个"向下的地图"std::unordered_map<State*, py::object>并注册每个创建的MyState。但我宁愿不诉诸这样的事情。

我认为您可能遇到了这一系列问题:

https://github.com/pybind/pybind11/issues/1774

最终,因为你只是直接返回MyState门,然后直接进入C++,Python解释器会失去对你的实例的跟踪,并继续垃圾回收对象的Python部分,这就是为什么你的对象最终被切片的原因。

潜在的解决方案:

  • 存储对返回MyState的引用,至少足够长的时间让 Python 解释器再次获得引用。
    • 例如,将return MyState(...)更改为self._stashed_state = MyState(...); return self._stashed_state
  • 看看你是否可以以某种方式incref你的类的 Python 版本C++(哎呀,但它会起作用)
  • 查看上述问题中列出的解决方法(不记得所有解决方法)
  • 使用我们的pybind11分支,它可以处理这个问题,但也拖入其他内容:RobotLocomotion/pybind11 概述

您可能还想发布现有问题之一,提及您也遇到了此问题(只是为了可以跟踪它)。