【图像分类】学习率更新策略/优化器

项目地址：https://github.com/Fafa-DL/Awesome-Backbones

项目说明：开箱即用，涵盖主流模型的图像分类|主干网络学习/对比/魔改项目

在models下的配置文件中，学习率更新方式以及参数均在lr_config中，在上一期讲网络如何被搭建时说了type是我们所需的调用关键字。在core/optimizers/lr_update.py中可以看到目前所支持的学习率更新策略。

常用的为StepLrUpdater与CosineAnnealingLrUpdater，均继承了LrUpdater类，先来看看它

class LrUpdater(object):
    """LR Scheduler in MMCV.

    Args:
        by_epoch (bool): LR changes epoch by epoch
        warmup (string): Type of warmup used. It can be None(use no warmup),
            'constant', 'linear' or 'exp'
        warmup_iters (int): The number of iterations or epochs that warmup
            lasts
        warmup_ratio (float): LR used at the beginning of warmup equals to
            warmup_ratio * initial_lr
        warmup_by_epoch (bool): When warmup_by_epoch == True, warmup_iters
            means the number of epochs that warmup lasts, otherwise means the
            number of iteration that warmup lasts
    """

    def __init__(self,
                 by_epoch=True,
                 warmup=None,
                 warmup_iters=0,
                 warmup_ratio=0.1,
                 warmup_by_epoch=False):
        # validate the "warmup" argument
        if warmup is not None:
            if warmup not in ['constant', 'linear', 'exp']:
                raise ValueError(
                    f'"{warmup}" is not a supported type for warming up, valid'
                    ' types are "constant" and "linear"')
        if warmup is not None:
            assert warmup_iters > 0, \
                '"warmup_iters" must be a positive integer'
            assert 0 < warmup_ratio <= 1.0, \
                '"warmup_ratio" must be in range (0,1]'

        self.by_epoch = by_epoch
        self.warmup = warmup
        self.warmup_iters = warmup_iters
        self.warmup_ratio = warmup_ratio
        self.warmup_by_epoch = warmup_by_epoch

        if self.warmup_by_epoch:
            self.warmup_epochs = self.warmup_iters
            self.warmup_iters = None
        else:
            self.warmup_epochs = None

        self.base_lr = []  # initial lr for all param groups
        self.regular_lr = []  # expected lr if no warming up is performed
        
    # 给optimizer设置新lr
    def _set_lr(self, runner, lr_groups):

        for param_group, lr in zip(runner.get("optimizer").param_groups,
                                    lr_groups):
            param_group['lr'] = lr

    def get_lr(self, runner, base_lr):
        raise NotImplementedError
    
    # 获取新lr
    def get_regular_lr(self, runner):

        return [self.get_lr(runner, _base_lr) for _base_lr in self.base_lr]

    # 获取warmup学习率更新策略的lr
    def get_warmup_lr(self, cur_iters):

        def _get_warmup_lr(cur_iters, regular_lr):
            if self.warmup == 'constant':
                warmup_lr = [_lr * self.warmup_ratio for _lr in regular_lr]
            elif self.warmup == 'linear':
                k = (1 - cur_iters / self.warmup_iters) * (1 -
                                                           self.warmup_ratio)
                warmup_lr = [_lr * (1 - k) for _lr in regular_lr]
            elif self.warmup == 'exp':
                k = self.warmup_ratio**(1 - cur_iters / self.warmup_iters)
                warmup_lr = [_lr * k for _lr in regular_lr]
            return warmup_lr

        if isinstance(self.regular_lr, dict):
            lr_groups = {}
            for key, regular_lr in self.regular_lr.items():
                lr_groups[key] = _get_warmup_lr(cur_iters, regular_lr)
            return lr_groups
        else:
            return _get_warmup_lr(cur_iters, self.regular_lr)
        
    # 记录初始lr
    def before_run(self, runner):
        # NOTE: when resuming from a checkpoint, if 'initial_lr' is not saved,
        # it will be set according to the optimizer params
        
        for group in runner.get("optimizer").param_groups:
            group.setdefault('initial_lr', group['lr'])
        self.base_lr = [
            group['initial_lr'] for group in runner.get("optimizer").param_groups
        ]
        
    # 在周期更新前获取新lr并在optimizer中更新它
    def before_train_epoch(self, runner):
        if self.warmup_iters is None: # 即self.warmup_by_epoch为True，warmup_epochs = warmup_iters
            epoch_len = len(runner.get("train_loader"))
            self.warmup_iters = self.warmup_epochs * epoch_len # 按周期更新则warmup iters = warmup_epochs * datasets//batch size
        # 不按周期更新则没必要在此进行lr更新，在下一步before_train_iter中更新
        if not self.by_epoch:
            return

        self.regular_lr = self.get_regular_lr(runner)
        self._set_lr(runner, self.regular_lr)
        
    # 首先判断是否按周期更新lr，若按迭代次数更新即by_epoch为False，大于等于warmup_iters使用正常lr更新方式，小于则用warmup方式更新lr
    def before_train_iter(self, runner):
        cur_iter = runner.get("iter")
        if not self.by_epoch:
            self.regular_lr = self.get_regular_lr(runner)
            if self.warmup is None or cur_iter >= self.warmup_iters:
                self._set_lr(runner, self.regular_lr)
            else:
                warmup_lr = self.get_warmup_lr(cur_iter)
                self._set_lr(runner, warmup_lr)
        elif self.by_epoch:
            # 按周期更新lr，若当前迭代步数大于warmup迭代步数阈值，直接返回，使用before_train_epoch中的lr
            if self.warmup is None or cur_iter > self.warmup_iters:
                return
            elif cur_iter == self.warmup_iters: # 等于则用常规方式
                self._set_lr(runner, self.regular_lr)
            else:
                warmup_lr = self.get_warmup_lr(cur_iter) # 小于则用warmup方式
                self._set_lr(runner, warmup_lr)

LrUpdater执行的顺序为：

1. __init__，初始化，主要针对warmup，需注意的是若warmup_by_epoch=True则warmup_iters代表的是warmup作用的epoch数，若warmup_by_epoch=False，则warmup_iters代表的是warmup作用的iters数，1 epoch = total images // batch size iters
2. before_run，在这里记录了初始学习率，通过获取配置文件optimizer_cfg中的lr定义初始学习率
3. before_train_epoch，在每个epoch更新前获取新lr并在optimizer中更新学习率。若by_epoch=False，大于等于warmup_iters使用常规lr更新方式，小于则用warmup方式更新lr
   1. get_regular_lr，调用配置文件中type指定的学习率更新方式，如StepLrUpdater与CosineAnnealingLrUpdater，通过get_lr获取当前epoch/iter下的学习率
   2. _set_lr给优化器设置最新的学习率
4. before_train_iter，在每个iter更新前获取新lr并在optimizer中更新学习率。首先判断是否按epoch更新lr
   1. 若by_epoch = False，即按iter更新，大于等于warmup_iters使用正常lr更新方式，小于则用warmup方式更新lr
   2. 若by_epoch = True，按epoch更新lr，若当前迭代步数大于warmup迭代步数阈值，直接返回，使用before_train_epoch中的lr，等于则用常规方式，小于则用warmup方式
   3. get_warmup_lr获取warmup学习率更新策略的lr

说完了核心LrUpdater类，再说说StepLrUpdater与CosineAnnealingLrUpdater，我们知道主要功能是根据当前iter或epoch计算最新学习率并返回，先看看StepLrUpdater

class StepLrUpdater(LrUpdater):
    """Step LR scheduler with min_lr clipping.

    Args:
        step (int | list[int]): Step to decay the LR. If an int value is given,
            regard it as the decay interval. If a list is given, decay LR at
            these steps.
        gamma (float, optional): Decay LR ratio. Default: 0.1.
        min_lr (float, optional): Minimum LR value to keep. If LR after decay
            is lower than `min_lr`, it will be clipped to this value. If None
            is given, we don't perform lr clipping. Default: None.
    """

    def __init__(self, step, gamma=0.1, min_lr=None, **kwargs):

        self.step = step
        self.gamma = gamma
        self.min_lr = min_lr
        super(StepLrUpdater, self).__init__(**kwargs)

    def get_lr(self, runner, base_lr):
        progress = runner.get('epoch') if self.by_epoch else runner.get('iter')

        # calculate exponential term
        if isinstance(self.step, int):
            exp = progress // self.step
        else:
            exp = len(self.step)
            for i, s in enumerate(self.step):
                if progress < s:
                    exp = i
                    break

        lr = base_lr * (self.gamma**exp)
        if self.min_lr is not None:
            # clip to a minimum value
            lr = max(lr, self.min_lr)
        return lr

step：步骤衰减LR，有两种形式，一是int，另一个是list。举例：

• step=2，即每2个iter/epoch更新一次学习率
• step=[20,60,80]，即0-20，21-60，61-80，80-X间的iter/epoch更新一次学习率

gamma：学习率衰减比率，对应公式lr = base_lr * (self.gamma**exp)

min_lr：最小学习率，若设置了该值，则当学习率小于该值时使用min_lr

再看看CosineAnnealingLrUpdater

class CosineAnnealingLrUpdater(LrUpdater):

    def __init__(self, min_lr=None, min_lr_ratio=None, **kwargs):
        assert (min_lr is None) ^ (min_lr_ratio is None)
        self.min_lr = min_lr
        self.min_lr_ratio = min_lr_ratio
        super(CosineAnnealingLrUpdater, self).__init__(**kwargs)

    def get_lr(self, runner, base_lr):
        if self.by_epoch:
            progress = runner.get('epoch')
            max_progress = runner.get('max_epochs')
        else:
            progress = runner.get('iter')
            max_progress = runner.get('max_iters')

        if self.min_lr_ratio is not None:
            target_lr = base_lr * self.min_lr_ratio
        else:
            target_lr = self.min_lr
        return annealing_cos(base_lr, target_lr, progress / max_progress)

min_lr：最小学习率

min_lr_ratio：最小学习率比率，min_lr_ratio与min_lr不能同时存在，因为这两个参数仅是用于计算最终学习率，当设置了min_lr，则最终学习率为min_lr；若设置了min_lr_ratio，则最终学习率为base_lr * min_lr_ratio

annealing_cos，学习率更新计算方法，具体如下

def annealing_cos(start, end, factor, weight=1):
    """Calculate annealing cos learning rate.

    Cosine anneal from `weight * start + (1 - weight) * end` to `end` as
    percentage goes from 0.0 to 1.0.

    Args:
        start (float): The starting learning rate of the cosine annealing.
        end (float): The ending learing rate of the cosine annealing.
        factor (float): The coefficient of `pi` when calculating the current
            percentage. Range from 0.0 to 1.0.
        weight (float, optional): The combination factor of `start` and `end`
            when calculating the actual starting learning rate. Default to 1.
    """
    cos_out = cos(pi * factor) + 1
    return end + 0.5 * weight * (start - end) * cos_out

至此学习率更新介绍完毕，最后说下优化器是如何调用，对应配置文件中的optimizer_cfg，在上一期讲网络如何被搭建时说了type是我们所需的调用关键字，此处type与torch.optim中的优化器一一对应，也就是不需要额外重写优化器类。

在tools/train.py中可看到import torch.optim as optim，对应构建optimizer的语句为

optimizer = eval('optim.' + optimizer_cfg.pop('type'))(params=model.parameters(),**optimizer_cfg) 

还是借助eval，所以如果有修改优化器的需要，大家可以查阅pytorch手册自行修改。

以上就是本次学习率更新策略/优化器的全部内容，在后续可视化教程中会绘制学习率曲线