import torch
from ._BaseOptimiser import BaseOptimiser
[docs]
class ADADELTA(BaseOptimiser):
"""
The adadelta optimiser. A first order method and therefore does not use information on second gradients, i.e. the hessian matrix. Hence, does not require a lot of memory.
Args:
learning_rate (float, optional): The learning rate of the optimiser. Must be positive. Defaults to 0.001.
rho (float, optional): Determines how long the previous gradients affect the current step direction. Must be in range [0, 1). Defaults to 0.9.
weight_decay (float, optional): Determines if regularisation should be applied to the weights. Must be in range [0, 1). Defaults to 0.
"""
def __init__(self, learning_rate=0.001, rho=0.9, weight_decay=0):
if not isinstance(learning_rate, int | float) or learning_rate <= 0:
raise ValueError("learning_rate must be a positive real number.")
if not isinstance(rho, int | float) or rho < 0 or rho >= 1:
raise ValueError("rho must be in range [0, 1).")
if not isinstance(weight_decay, int | float) or weight_decay < 0 or weight_decay >= 1:
raise ValueError("weight_decay must be in range [0, 1).")
self.learning_rate = learning_rate
self.rho = rho
self.weight_decay = weight_decay
[docs]
def initialise_parameters(self, model_parameters):
"""
Initialises the optimiser with the parameters that need to be optimised.
Args:
model_parameters (list[torch.Tensor]): The parameters that will be optimised. Must be a list or a tuple of torch tensors.
"""
if not isinstance(model_parameters, list | tuple):
raise TypeError("model_parameters must be a list or a tuple of torch tensors.")
self.model_parameters = model_parameters
self.square_avg = [torch.zeros_like(param) for param in model_parameters]
self.accumulate_variables = [torch.zeros_like(param) for param in model_parameters]
self.t = 0
[docs]
def update_parameters(self):
"""
Takes a step towards the optimum for each parameter.
"""
self.t += 1
for i, parameter in enumerate(self.model_parameters):
if self.weight_decay > 0: parameter.grad += self.weight_decay * parameter
self.square_avg[i] = self.square_avg[i] * self.rho + parameter.grad ** 2 * (1 - self.rho)
delta_x = parameter.grad * torch.sqrt(self.accumulate_variables[i] + 1e-10) / torch.sqrt(self.square_avg[i] + 1e-10)
self.accumulate_variables[i] = self.accumulate_variables[i] * self.rho + delta_x ** 2 * (1 - self.rho)
parameter -= self.learning_rate * delta_x