def start_of_epoch_hook()

in banding_removal/fastmri/learning_rate_mixin.py [0:0]

• 40 lines of code
• 16 McCabe index (conditional complexity)

    def start_of_epoch_hook(self, epoch):
        args = self.args
        momentum = args.momentum
        decay = args.decay

        if args.lr_reduction == "default":
            lr = args.lr * (0.1 ** (epoch // 75))
        elif args.lr_reduction == "none":
            lr = args.lr
        elif args.lr_reduction == "60-90":
            lr = args.lr * (0.1 ** (epoch // 60)) * (0.1 ** (epoch // 90))
        elif args.lr_reduction == "58":
            lr = args.lr * (0.1 ** (epoch // 58))
        elif args.lr_reduction == "80-120":
            lr = args.lr * (0.1 ** (epoch // 80)) * (0.1 ** (epoch // 120))
        elif args.lr_reduction == "150-225":
            lr = args.lr * (0.1 ** (epoch // 150)) * (0.1 ** (epoch // 225))
        elif args.lr_reduction == "40-50-55":
            lr = args.lr * (0.1 ** (epoch // 40)) * (0.1 ** (epoch // 50)) * (0.1 ** (epoch // 55))
        elif args.lr_reduction == "50-55-58":
            lr = args.lr * (0.1 ** (epoch // 50)) * (0.1 ** (epoch // 55)) * (0.1 ** (epoch // 58))
        elif args.lr_reduction == "linear":
            lr = args.lr * (0.985 ** epoch)
        elif args.lr_reduction == "every20":
            lr = args.lr * (0.1 ** (epoch // 20))
        elif args.lr_reduction == "every40":
            lr = args.lr * (0.1 ** (epoch // 40))
        elif args.lr_reduction == "rampupto":
            if epoch == 0:
                lr = args.lr*0.1
            else:
                lr = args.lr
        else:
            raise Exception(f"Lr scheme not recognised: {args.lr_reduction}")

        for i, param_group in enumerate(self.optimizer.param_groups):
            if 'group_scaling' not in param_group:
                param_group['group_scaling'] = 1.0

            group_lr = lr*param_group['group_scaling']
            param_group['lr'] = group_lr
            #param_group['momentum'] = momentum
            #param_group['weight_decay'] = decay
            logging.info(f"Group {i} Learning rate: {group_lr} momentum: {momentum} decay: {decay}")

        super().start_of_epoch_hook(epoch)