Optimizers manage weight update starting from gradient values. They may have complex internal states to better move on the loss multi-dimensional surface. Please use the fixed signature __init__(hyperparameters: Namespace, named_parameters: Generator) -> None for all the subclasses.
This optimizer is same as AdamW but for a small fix to the moving average update mechanism. Original implementation can be found here.