Source code for torch_uncertainty.post_processing.calibration.temperature_scaler

from typing import Literal

import torch
from torch import Tensor, nn

from .scaler import Scaler




class TemperatureScaler(Scaler):
    def __init__(
        self,
        model: nn.Module | None = None,
        init_val: float = 1,
        lr: float = 0.1,
        max_iter: int = 100,
        eps: float = 1e-8,
        device: Literal["cpu", "cuda"] | torch.device | None = None,
    ) -> None:
        """Temperature scaling post-processing for calibrated probabilities.

        Args:
            model (nn.Module): Model to calibrate.
            init_val (float, optional): Initial value for the temperature. Defaults to ``1``.
            lr (float, optional): Learning rate for the optimizer. Defaults to ``0.1``.
            max_iter (int, optional): Maximum number of iterations for the optimizer. Defaults to ``100``.
            eps (float): Small value for stability. Defaults to ``1e-8``.
            device (Optional[Literal["cpu", "cuda"]], optional): Device to use for optimization. Defaults to ``None``.

        References:
            [1] `On calibration of modern neural networks. In ICML 2017
            <https://arxiv.org/abs/1706.04599>`_.
        """
        super().__init__(model=model, lr=lr, max_iter=max_iter, eps=eps, device=device)

        if init_val <= 0:
            raise ValueError(f"Initial temperature value must be positive. Got {init_val}")

        self.set_temperature(init_val)



    def set_temperature(self, val: float) -> None:
        """Set the temperature to a fixed value.

        Args:
            val (float): Temperature value.
        """
        if val <= 0:
            raise ValueError(f"Temperature value must be positive. Got {val}")

        self.temp = nn.Parameter(torch.ones(1, device=self.device) * val, requires_grad=True)


    def _scale(self, logits: Tensor) -> Tensor:
        return logits / self.temperature[0]

    @property
    def temperature(self) -> list:
        return [self.temp]