import torch.optim as optim from graphmetro.config import cfg def create_optimizer(params, lr=None): r"""Creates a config-driven optimizer.""" params = filter(lambda p: p.requires_grad, params) if cfg.optim.optimizer == 'adam': optimizer = optim.Adam(params, lr=cfg.optim.standard_lr if lr is None else lr, weight_decay=cfg.optim.weight_decay) elif cfg.optim.optimizer == 'sgd': optimizer = optim.SGD(params, lr=cfg.optim.standard_lr if lr is None else lr, momentum=cfg.optim.momentum, weight_decay=cfg.optim.weight_decay) else: raise ValueError('Optimizer {} not supported'.format( cfg.optim.optimizer)) return optimizer def create_scheduler(optimizer): r"""Creates a config-driven learning rate scheduler.""" if cfg.optim.scheduler == 'none': scheduler = optim.lr_scheduler.StepLR(optimizer, step_size=cfg.optim.max_epoch + 1, gamma=0.1) elif cfg.optim.scheduler == 'step': scheduler = optim.lr_scheduler.StepLR(optimizer, step_size=cfg.optim.max_epoch // 3, gamma=0.1) elif cfg.optim.scheduler == 'multistep': scheduler = optim.lr_scheduler.MultiStepLR(optimizer, milestones=cfg.optim.steps, gamma=cfg.optim.lr_decay) elif cfg.optim.scheduler == 'cos': scheduler = optim.lr_scheduler.CosineAnnealingLR( optimizer, T_max=cfg.optim.max_epoch) else: raise ValueError('Scheduler {} not supported'.format( cfg.optim.scheduler)) return scheduler