Skip to main content

optimizers

This module contains the optimizers used in the FPM reconstruction process. The optimizers are used as arguments in the reconstruct function in the algorithm.py module. Optimizers update the object and pupil in each iteration of the reconstruction process.

All Optimizer functions can only accept one positional argument of type: OptimizerInputs. This structure is created and passed in by the reconstruction algorithm.

Class: OptimizerInputs​

fpm_py.optimizers.OptimizerInputs:
    ...

Dataclass for the inputs to the optimizer functions. All optimizer functions must accept this dataclass as the only positional argument.

Attributes: object (torch.Tensor): The object array. pupil (torch.Tensor): The pupil array. wave_fourier (torch.Tensor): The current Fourier domain of the wave. wave_fourier_new (torch.Tensor): The new Fourier domain of the wave. x (int): Bottom row where pupil overlaps with object. y (int): Leftmost column where pupil overlaps with object.

Function: quasi_second_order​

fpm_py.optimizers.quasi_second_order(
    inputs: OptimizerInputs,
    alpha_o: float,
    mu_o: float,
    alpha_p: float,
    mu_p: float
):
    ...

Simple gradient descent optimizer with learning rate and regularization hyperparams for object and pupil.

Args:​

  • inputs (OptimizerInputs): The inputs to the optimizer function.
  • alpha_o (float): Regularization hyperparameter for object.
  • mu_o (float): Learning rate hyperparameter for object.
  • alpha_p (float): Regularization hyperparameter for pupil.
  • mu_p (float): Learning rate hyperparameter for pupil.

Returns:​

  • tuple: The updated object and pupil.

Function: tomas​

fpm_py.optimizers.tomas(
    inputs: OptimizerInputs,
    alpha: float,
    beta: float
):
    ...

Tomas' optimizer with learning rate and regularization hyperparams for object and pupil.

Args:​

  • inputs (OptimizerInputs): The inputs to the optimizer function.
  • alpha (float): The learning rate for the object.
  • beta (float): The learning rate for the pupil.

Returns:​

  • tuple: The updated object and pupil.