使用nptyping标注含NDArray的元组返回函数时遇错误求解决

错误根源是nptyping的NDArray不支持直接用Python原生的|运算符表示联合数据类型，它要求数据类型参数为字符串描述，或使用nptyping提供的Union包裹多个类型。以下是两种可行的修正方案：

方案1：使用nptyping的Union #

从nptyping导入Union，用它来包裹多个数据类型，替代Int32 | Float32的写法：

from nptyping import NDArray, UInt8, Int32, Float32, Shape, Bool, Union
import numpy as np
import rasterio
from typing import Generator, Any, Final, Literal


def _scale_and_round(
    self, arr: NDArray[Any, Float32]
) -> tuple[NDArray[Any, Union[Int32, Float32]], dict[str, Any]]:
    
    array: NDArray[Any, Any] = arr * self.scale_factor
    if self.scale_factor == 1000:
        array = array.astype(np.int32)
    return array, self.metadata

def ndvi(
    self, red_src: Any, nir_src: Any
) -> tuple[NDArray[Any, Union[Int32, Float32]], dict[str, Any]]:
    
    redB: NDArray[Any, Any] = red_src.read()
    nirB: NDArray[Any, Any] = nir_src.read()
    np.seterr(divide="ignore", invalid="ignore")
    ndvi: NDArray[Any, Float32] = (
        nirB.astype(np.float32) - redB.astype(np.float32)
    ) / (nirB.astype(np.float32) + redB.astype(np.float32))
    # replace nan with 0
    where_are_NaNs: NDArray[Any, Bool] = np.isnan(ndvi)
    ndvi[where_are_NaNs] = 0

    return self._scale_and_round(ndvi)

方案2：使用字符串形式的联合类型 #

直接用字符串描述联合类型，无需额外导入Union：

from nptyping import NDArray, UInt8, Int32, Float32, Shape, Bool
import numpy as np
import rasterio
from typing import Generator, Any, Final, Literal


def _scale_and_round(
    self, arr: NDArray[Any, Float32]
) -> tuple[NDArray[Any, "Int32 | Float32"], dict[str, Any]]:
    
    array: NDArray[Any, Any] = arr * self.scale_factor
    if self.scale_factor == 1000:
        array = array.astype(np.int32)
    return array, self.metadata

def ndvi(
    self, red_src: Any, nir_src: Any
) -> tuple[NDArray[Any, "Int32 | Float32"], dict[str, Any]]:
    
    redB: NDArray[Any, Any] = red_src.read()
    nirB: NDArray[Any, Any] = nir_src.read()
    np.seterr(divide="ignore", invalid="ignore")
    ndvi: NDArray[Any, Float32] = (
        nirB.astype(np.float32) - redB.astype(np.float32)
    ) / (nirB.astype(np.float32) + redB.astype(np.float32))
    # replace nan with 0
    where_are_NaNs: NDArray[Any, Bool] = np.isnan(ndvi)
    ndvi[where_are_NaNs] = 0

    return self._scale_and_round(ndvi)

注意：原代码缺少import numpy as np，需补充后才能正常运行np.int32、np.isnan等调用。

内容的提问来源于stack exchange，提问作者diegus