Alternative that doesn't violate the Liskov substitution principle

Question

Considering the following structure of classes:

from abc import ABC, abstractmethod


class ModelSettings(ABC):
    ...

class BlueModelSettings(ModelSettings):
    ...

class RedModelSettings(ModelSettings):
    ...


class Model(ABC):
    @abstractmethod
    def compute(self, settings: ModelSettings):
        ...

class BlueModel(Model):
    def compute(self, settings: BlueModelSettings):
        ...

class RedModel(Model):
    def compute(self, settings: RedModelSettings):
        ...

MyPy is complaining that the implementations of compute in BlueModel and RedModel violate the Liskov substitution principle. I understand this and also understand why it makes sense.

My question is, instead of the above structure, what would be another approach that would satisfy the following requirements:

• base model has a contract that compute should receive settings, not apples or oranges
• but each specific model should accept only settings relevant to its own model kind

In a way, I essentially want that a subclass' method argument is stricter than what its base class stipulates, which goes directly against the Liskov principle. Which is what's telling me there might be a more suitable approach.

Note:

• Model is library code so it can't know of e.g. BlueModelSettings which is client code

Answer 1

You need your Model to be generic in the type of settings.

T = TypeVar('T', bound=ModelSettings)

class Model(ABC, Generic[T]):
    @abstractmethod
    def compute(self, settings: T):
        ...

class BlueModel(Model[BlueModelSettings]):
    def compute(self, settings: BlueModelSettings):
        ...

class RedModel(Model[RedModelSettings]):
    def compute(self, settings: RedModelSettings):
        ...