可变学习率

from bisect import bisect_right

import torch


# FIXME ideally this would be achieved with a CombinedLRScheduler,
# separating MultiStepLR with WarmupLR
# but the current LRScheduler design doesn't allow it
class WarmupMultiStepLR(torch.optim.lr_scheduler._LRScheduler):
    def __init__(
        self,
        optimizer,
        milestones,
        gamma=0.1,
        warmup_factor=1.0 / 3,
        warmup_iters=500,
        warmup_method="linear",
        last_epoch=-1,
    ):
        if not list(milestones) == sorted(milestones):
            raise ValueError(
                "Milestones should be a list of" " increasing integers. Got {}",
                milestones,
            )

        if warmup_method not in ("constant", "linear"):
            raise ValueError(
                "Only 'constant' or 'linear' warmup_method accepted"
                "got {}".format(warmup_method)
            )
        self.milestones = milestones
        self.gamma = gamma
        self.warmup_factor = warmup_factor
        self.warmup_iters = warmup_iters
        self.warmup_method = warmup_method
        super(WarmupMultiStepLR, self).__init__(optimizer, last_epoch)

    def get_lr(self):
        warmup_factor = 1
        if self.last_epoch < self.warmup_iters:
            if self.warmup_method == "constant":
                warmup_factor = self.warmup_factor
            elif self.warmup_method == "linear":
                alpha = float(self.last_epoch) / self.warmup_iters
                warmup_factor = self.warmup_factor * (1 - alpha) + alpha
        return [
            base_lr
            * warmup_factor
            * self.gamma ** bisect_right(self.milestones, self.last_epoch)
            for base_lr in self.base_lrs
        ]





def make_optimizer(cfg, model):
    logger = logging.getLogger("fcos_core.trainer")
    params = []
    for key, value in model.named_parameters():
        if not value.requires_grad:
            continue
        lr = 0.001
        weight_decay = 0.0005
        if "bias" in key:
            lr = 0.001 * 2
            weight_decay = 0
        if key.endswith(".offset.weight") or key.endswith(".offset.bias"):
            logger.info("set lr factor of {} as {}".format(
                key, 1.0
            ))
            lr *= 1.0
        params += [{"params": [value], "lr": lr, "weight_decay": weight_decay}]

    optimizer = torch.optim.SGD(params, lr, momentum=0.9)
    return optimizer


def make_lr_scheduler(cfg, optimizer):
    return WarmupMultiStepLR(
        optimizer,
        (30000,80000),
        0.1,
        warmup_factor = 1.0 / 3,
        warmup_iters= 500,
        warmup_method= "linear",
    )