def get_lr_schedulers()

in main.py [0:0]

• 41 lines of code
• 6 McCabe index (conditional complexity)

def get_lr_schedulers(args, datasets, is_trnsf):

    if args.train.scheduling_mode == "decay":
        gamma = get_exponential_decay_gamma(
            args.train.scheduling_factor, args.train.trnsf_kwargs.max_epochs
        )

        lr_scheduler = [
            LRScheduler(torch.optim.lr_scheduler.ExponentialLR, gamma=gamma)
        ]
    elif args.train.scheduling_mode == "plateau":
        lr_scheduler = [
            LRScheduler(
                torch.optim.lr_scheduler.ReduceLROnPlateau,
                monitor="valid_loss",
                factor=0.2,
            )
        ]
    elif args.train.scheduling_mode == "biplateau":
        lr_scheduler = [
            LRScheduler(
                torch.optim.lr_scheduler.ReduceLROnPlateau,
                monitor="valid_loss",
                factor=0.2,  # 0.1
                patience=3,  # 0.5
                verbose=True,
                threshold=0.01,
                min_lr=1e-5,
            ),
            # increase lr but max at 0.5
            LRScheduler(SetLR, lr_lambda=lambda _, lr, __: min(lr * 1.3, 0.5)),
            # dirty way for not increasing lr in case loss didn't improve
            LRScheduler(
                torch.optim.lr_scheduler.ReduceLROnPlateau,
                monitor="valid_loss",
                factor=1 / 1.3,
                patience=1,
            ),
        ]
    elif args.train.scheduling_mode is None:
        lr_scheduler = []
    else:
        raise ValueError(
            f"Unkown scheduling_mode={args.train.scheduling_mode}")

    return lr_scheduler