Python 中的模板类

Ary*_*thy 6 python templates metaclass python-3.x python-typing

我正在尝试用 Python 创建代数,但发现创建参数化类很困难。

作为一个例子,考虑这个ProductWeight类。它包含另外两个Weight对象,具有强制类型(至少静态地,通过mypy)。

这将失败,因为(根据设计)我无法访问类W1W2调用它们的类方法,例如zero. (它们没有指定;ProductWeight没有模板化。)ProductWeight当我创建实例时,不知道要绑定到它的类型。

from typing import Generic, TypeVar
W1 = TypeVar("W1")
W2 = TypeVar("W2")

class ProductWeight(Generic[W1, W2]):
    def __init__(self, value1: W1, value2: W2):
        self.value1_ = value1
        self.value2_ = value2

    @classmethod
    def zero(cls):
        return cls(W1.zero(), W2.zero())  # Will fail - no access to W1 and W2.
    
Run Code Online (Sandbox Code Playgroud)

相比之下,这在 C++ 中很简单:因为类型是参数化的,所以它能够查找W1::Zero.

template<typename W1, typename W2>
public:
    ProductWeight(W1 w1, W2 w2) : value1_(w1), value2_(w2) {}
    static ProductWeight Zero() {
        return ProductWeight(W1::Zero(), W2::Zero());
    }
private:
    W1 value1_;
    W2 value2_;
};
Run Code Online (Sandbox Code Playgroud)

Python 中有解决这个问题的方法吗?要么创建一个内部类,要么以其他方式向类(而不是类实例)提供类型?

为了获得最小的可重现示例,您可以使用另一种Weight类型的实现。

class SimpleWeight:
    def __init__(self, value):
        assert value in {True, False}
        self.value_ = value

    @classmethod
    def zero(cls):
        return cls(False)

weight1 = SimpleWeight(False)
weight2 = SimpleWeight(True)
product = ProductWeight(weight1, weight2)

print(SimpleWeight.zero())  # So far, so good.
print(ProductWeight.zero())  # Oof.
# Predictably, it failed because `ProductWeight` is not specialized.
Run Code Online (Sandbox Code Playgroud)

这是可预测的错误消息:

Traceback (most recent call last):
  File "garbage.py", line 32, in <module>
    print(ProductWeight.zero())  # Oof.
  File "garbage.py", line 14, in zero
    return cls(W1.zero(), W2.zero())  # Will fail - no access to W1 and W2.
AttributeError: 'TypeVar' object has no attribute 'zero'
Run Code Online (Sandbox Code Playgroud)

理想情况下,可以创建如下所示的参数类型:

product = ProductWeight[SimpleWeight, SimpleWeight](weight1, weight2)
# And similarly:
print(ProductWeight[SimpleWeight, SimpleWeight].zero())
Run Code Online (Sandbox Code Playgroud)

san*_*ser 8

为了解决这个问题,最重要的是要认识到 C++ 中的泛型会创建多个类,而在 Python 中,无论其构造函数中的参数是什么类型,您始终只有一个类。

换句话说,在C++中vector<int>, 和vector<string>是两个类。如果将它们绑定到 Python 解释器,则必须为它们分配两个不同的名称,例如VectorIntVectorString

代码

def ProductWeight(value1, value2):
    def init(self, value1, value2):
        self.value1 = value1
        self.value2 = value2

    def zero(cls):
        return cls(cls.W1.zero(), cls.W2.zero())

    W1 = type(value1)
    W2 = type(value2)
    name = f'ProductWeight{W1}{W2}'

    try:
        return ProductWeight.types[name](value1, value2)
    except KeyError:
        pass

    cls = type(name, (), {'__init__': init})
    cls.W1 = W1
    cls.W2 = W2
    cls.zero = classmethod(zero)
    ProductWeight.types[name] = cls

    return cls(value1, value2)

ProductWeight.types = {}

class SimpleWeight:
    def __init__(self, value):
        assert value in {True, False}
        self.value_ = value

    @classmethod
    def zero(cls):
        return cls(False)

weight1 = SimpleWeight(False)
weight2 = SimpleWeight(True)
product = ProductWeight(weight1, weight2)

print(SimpleWeight.zero())
print(type(product).zero())
Run Code Online (Sandbox Code Playgroud)

该代码尝试尊重您的原始 API,您可以ProductWeight使用通用参数创建类的实例。但是,如果要调用实例的类方法,则必须访问实例的基础类型(请注意最后一行从到的zero更改)。ProductWeighttype(product)

为了方便起见,您可以将此引用保存到变量中。

该函数ProductWeight充当通用工厂。每次调用它时,它都会根据参数的类型为新类创建一个名称。如果这样的类已经存在,它只返回它的一个新实例。否则,它使用该函数创建新类type,然后返回一个新实例。

ProductWeight它本身也是一个单例对象,它具有已创建类型的字典。

结论

您可能会注意到,此解决方案使用的内存比其 C++ 对应方案多得多。然而,考虑到您选择使用 Python 而不是 C++,您可能不会太担心这一点。

更重要的是,您必须决定这是否是您要走的正确道路。请记住,Python 中没有泛型,只有“动态”。因此,从长远来看,你的思维方式带来的障碍比它消除的障碍还要多。

让 Python 听起来像 C++

这部分回答了您问题的编辑:

理想情况下,可以创建如下所示的参数类型:

product = ProductWeight[SimpleWeight, SimpleWeight](weight1, weight2)
# And similarly:
print(ProductWeight[SimpleWeight, SimpleWeight].zero())
Run Code Online (Sandbox Code Playgroud)

不要害怕,因为这实际上是可能的。以下代码使用__class_getitem__Python 3.7,但是通过解决方法,它也可以在旧版本上运行,请参阅有关静态getitem方法的问题。

代码重写

# SimpleWeight didn't change
from simple_weight import SimpleWeight

class ProductWeight:
    types = {}

    def __class_getitem__(cls, key):
        try:
            W1, W2 = key
        except ValueError:
            raise Exception('ProductWeight[] takes exactly two arguments.')

        name = f'{ProductWeight.__name__}<{W1.__name__}, {W2.__name__}>'

        try:
            return cls.types[name]
        except KeyError:
            pass

        new_type = type(name, (), {'__init__': cls.init})
        new_type.W1 = W1
        new_type.W2 = W2
        new_type.zero = classmethod(cls.zero)
        cls.types[name] = new_type

        return new_type

    def __init__(self):
        raise Exception('ProductWeight is a static class and cannot be instantiated.')

    def init(self, value1, value2):
        self.value1 = value1
        self.value2 = value2

    def zero(cls):
        return cls(cls.W1.zero(), cls.W2.zero())

weight1 = SimpleWeight(False)
weight2 = SimpleWeight(True)
product = ProductWeight[SimpleWeight, SimpleWeight](weight1, weight2)

print(SimpleWeight.zero())
print(ProductWeight[SimpleWeight, SimpleWeight].zero())
Run Code Online (Sandbox Code Playgroud)

这利用了括号运算符__getitem__接受任意数量的参数并将它们打包到一个元组中的事实,该元组是其第一个参数。您可以解压元组并获取所有类型。这可以扩展到考虑任意数量的类型,甚至是在运行时选择的数量。

推断类型

上一个版本失去了从传递给构造函数的参数推断类型的能力。通过创建抽象工厂,我们可以恢复此功能。

# SimpleWeight didn't change
from simple_weight import SimpleWeight

class ProductWeightAbstractFactory:
    def __call__(self, value1, value2):
        return self[type(value1), type(value2)](value1, value2)

    def __getitem__(self, types):
        W1, W2 = types
        name = f'ProductWeight<{W1.__name__}, {W2.__name__}>'

        try:
            return self.types[name]
        except KeyError:
            pass

        cls = type(self)
        new_type = type(name, (), {'__init__': cls.init})
        new_type.W1 = W1
        new_type.W2 = W2
        new_type.zero = classmethod(cls.zero)
        self.types[name] = new_type

        return new_type

    def __init__(self):
        self.types = {}

    def init(self, value1, value2):
        self.value1 = value1
        self.value2 = value2

    def zero(cls):
        return cls(cls.W1.zero(), cls.W2.zero())

ProductWeight = ProductWeightAbstractFactory()

weight1 = SimpleWeight(False)
weight2 = SimpleWeight(True)
product = ProductWeight[SimpleWeight, SimpleWeight](weight1, weight2)
inferred_product = ProductWeight(weight1, weight2)

print(SimpleWeight.zero())
print(ProductWeight[SimpleWeight, SimpleWeight].zero())
print(type(inferred_product).zero())
Run Code Online (Sandbox Code Playgroud)

请注意,您必须在使用工厂之前创建一个实例:

ProductWeight = ProductWeightAbstractFactory()
Run Code Online (Sandbox Code Playgroud)

现在您可以使用括号创建具有显式类型的对象:

product = ProductWeight[SimpleWeight, SimpleWeight](weight1, weight2)
Run Code Online (Sandbox Code Playgroud)

或者您可以推断类型以使代码简洁:

product = ProductWeight(weight1, weight2)
Run Code Online (Sandbox Code Playgroud)

现在它比以往任何时候都更接近 C++ 语法。

类型检查

为了在开发时提供更多安全性,您还可以在构造函数中引入类型检查。

def init(self, value1, value2):
    def check_types(objects, required_types):
        for index, (obj, t) in enumerate(zip(objects, required_types)):
            if not issubclass(type(obj), t):
                raise Exception(f'Parameter {index + 1} is not a subclass of its required type {t}.')

    cls = type(self)
    check_types((value1, value2), (cls.W1, cls.W2))

    self.value1 = value1
    self.value2 = value2
Run Code Online (Sandbox Code Playgroud)

此代码将失败:

product = ProductWeight[SimpleWeight, SimpleWeight](1, 2)
Run Code Online (Sandbox Code Playgroud)