Cosine_LR_Annealing.py

"""
This is an implementation of the cosine annealing learning rate schedule. The idea of this algorithm is from the paper https://arxiv.org/abs/1608.03983.
As my knowledge, in pytorch, they have the implementation for the cosine annealing without the restart step, which is the main idea of the above paper. 
"""
class CosineWithRestarts(_LRScheduler):
    """
    Cosine annealing with restarts.

    Parameters
    ----------
    optimizer : torch.optim.Optimizer

    T_max : int
        The maximum number of iterations within the first cycle.

    eta_min : float, optional (default: 0)
        The minimum learning rate.

    last_epoch : int, optional (default: -1)
        The index of the last epoch.

    """

    def __init__(self,
                 optimizer: torch.optim.Optimizer,
                 T_max: int,
                 eta_min: float = 0.,
                 last_epoch: int = -1,
                 factor: float = 1.) -> None:
        # pylint: disable=invalid-name
        self.T_max = T_max
        self.eta_min = eta_min
        self.factor = factor
        self._last_restart: int = 0
        self._cycle_counter: int = 0
        self._cycle_factor: float = 1.
        self._updated_cycle_len: int = T_max
        self._initialized: bool = False
        super(CosineWithRestarts, self).__init__(optimizer, last_epoch)

    def get_lr(self):
        """Get updated learning rate."""
        # HACK: We need to check if this is the first time get_lr() was called, since
        # we want to start with step = 0, but _LRScheduler calls get_lr with
        # last_epoch + 1 when initialized.
        if not self._initialized:
            self._initialized = True
            return self.base_lrs

        step = self.last_epoch + 1
        self._cycle_counter = step - self._last_restart

        lrs = [
            (
                self.eta_min + ((lr - self.eta_min) / 2) *
                (
                    np.cos(
                        np.pi *
                        ((self._cycle_counter) % self._updated_cycle_len) /
                        self._updated_cycle_len
                    ) + 1
                )
            ) for lr in self.base_lrs
        ]

        if self._cycle_counter % self._updated_cycle_len == 0:
            # Adjust the cycle length.
            self._cycle_factor *= self.factor
            self._cycle_counter = 0
            self._updated_cycle_len = int(self._cycle_factor * self.T_max)
            self._last_restart = step

        return lrs