from collections.abc import Callable
from typing import Any
from typing_extensions import TypeAlias

import numpy as np
from tensorflow import Tensor
from tensorflow._aliases import FloatArray, FloatDataSequence, FloatTensorCompatible, Integer

# The implementation uses isinstance so it must be dict and not any Mapping.
_Activation: TypeAlias = str | None | Callable[[Tensor], Tensor] | dict[str, Any]
# Ints are not allowed.
_ActivationInput: TypeAlias = Tensor | FloatDataSequence | FloatArray | np.number[Any] | float

def deserialize(config: dict[str, Any], custom_objects: dict[str, Callable[..., Any]] | None = None) -> Callable[..., Any]: ...
def elu(x: _ActivationInput, alpha: FloatTensorCompatible | FloatDataSequence = 1.0) -> Tensor: ...
def exponential(x: _ActivationInput) -> Tensor: ...
def gelu(x: _ActivationInput, approximate: bool = False) -> Tensor: ...
def get(identifier: _Activation) -> Callable[[Tensor], Tensor]: ...
def hard_sigmoid(x: _ActivationInput) -> Tensor: ...
def linear(x: _ActivationInput) -> Tensor: ...
def mish(x: _ActivationInput) -> Tensor: ...
def relu(
    x: _ActivationInput,
    negative_slope: FloatTensorCompatible = 0.0,
    max_value: FloatTensorCompatible | FloatDataSequence | None = None,
    threshold: FloatTensorCompatible | FloatDataSequence = 0.0,
) -> Tensor: ...
def selu(x: _ActivationInput) -> Tensor: ...
def serialize(activation: Callable[..., Any]) -> str | dict[str, Any]: ...
def sigmoid(x: _ActivationInput) -> Tensor: ...
def softmax(x: Tensor, axis: Integer = -1) -> Tensor: ...
def softplus(x: _ActivationInput) -> Tensor: ...
def softsign(x: _ActivationInput) -> Tensor: ...
def swish(x: _ActivationInput) -> Tensor: ...
def tanh(x: _ActivationInput) -> Tensor: ...