pytorch DistributedDataParallel 多卡训练结果变差的问题分析

DDP 数据shuffle 的设置

使用DDP要给dataloader传入sampler参数（torch.utils.data.distributed.DistributedSampler(dataset, num_replicas=None, rank=None, shuffle=True, seed=0, drop_last=False)） 。 默认shuffle=True，但按照pytorch DistributedSampler的实现：

    def __iter__(self) -> Iterator[T_co]:
        if self.shuffle:
            # deterministically shuffle based on epoch and seed
            g = torch.Generator()
            g.manual_seed(self.seed + self.epoch)
            indices = torch.randperm(len(self.dataset), generator=g).tolist()  # type: ignore
        else:
            indices = list(range(len(self.dataset)))  # type: ignore

产生随机indix的种子是和当前的epoch有关，所以需要在训练的时候手动set epoch的值来实现真正的shuffle：

for epoch in range(start_epoch, n_epochs):
    if is_distributed:
        sampler.set_epoch(epoch)
    train(loader)

DDP 增大batchsize 效果变差的问题

large batchsize：
理论上的优点：

• 数据中的噪声影响可能会变小，可能容易接近最优点；

缺点和问题：

• 降低了梯度的variance；(理论上，对于凸优化问题，低的梯度variance可以得到更好的优化效果; 但是实际上Keskar et al验证了增大batchsize会导致差的泛化能力);
• 对于非凸优化问题，损失函数包含多个局部最优点，小的batchsize有噪声的干扰可能容易跳出局部最优点，而大的batchsize有可能停在局部最优点跳不出来。

解决方法：

• 增大learning_rate，但是可能出现问题，在训练开始就用很大的learning_rate 可能导致模型不收敛 (https://arxiv.org/abs/1609.04836)
• 使用warming up (https://arxiv.org/abs/1706.02677)

warmup

在训练初期就用很大的learning_rate可能会导致训练不收敛的问题，warmup的思想是在训练初期用小的学习率，随着训练慢慢变大学习率，直到base learning_rate，再使用其他decay（CosineAnnealingLR）的方式训练.

# copy from https://github.com/ildoonet/pytorch-gradual-warmup-lr/blob/master/warmup_scheduler/scheduler.py
from torch.optim.lr_scheduler import _LRScheduler
from torch.optim.lr_scheduler import ReduceLROnPlateau


class GradualWarmupScheduler(_LRScheduler):
    """ Gradually warm-up(increasing) learning rate in optimizer.
    Proposed in 'Accurate, Large Minibatch SGD: Training ImageNet in 1 Hour'.
    Args:
        optimizer (Optimizer): Wrapped optimizer.
        multiplier: target learning rate = base lr * multiplier if multiplier > 1.0. if multiplier = 1.0, lr starts from 0 and ends up with the base_lr.
        total_epoch: target learning rate is reached at total_epoch, gradually
        after_scheduler: after target_epoch, use this scheduler(eg. ReduceLROnPlateau)
    """

    def __init__(self, optimizer, multiplier, total_epoch, after_scheduler=None):
        self.multiplier = multiplier
        if self.multiplier < 1.:
            raise ValueError('multiplier should be greater thant or equal to 1.')
        self.total_epoch = total_epoch
        self.after_scheduler = after_scheduler
        self.finished = False
        super(GradualWarmupScheduler, self).__init__(optimizer)

    def get_lr(self):
        if self.last_epoch > self.total_epoch:
            if self.after_scheduler:
                if not self.finished:
                    self.after_scheduler.base_lrs = [base_lr * self.multiplier for base_lr in self.base_lrs]
                    self.finished = True
                return self.after_scheduler.get_last_lr()
            return [base_lr * self.multiplier for base_lr in self.base_lrs]

        if self.multiplier == 1.0:
            return [base_lr * (float(self.last_epoch) / self.total_epoch) for base_lr in self.base_lrs]
        else:
            return [base_lr * ((self.multiplier - 1.) * self.last_epoch / self.total_epoch + 1.) for base_lr in self.base_lrs]

    def step_ReduceLROnPlateau(self, metrics, epoch=None):
        if epoch is None:
            epoch = self.last_epoch + 1
        self.last_epoch = epoch if epoch != 0 else 1  # ReduceLROnPlateau is called at the end of epoch, whereas others are called at beginning
        if self.last_epoch <= self.total_epoch:
            warmup_lr = [base_lr * ((self.multiplier - 1.) * self.last_epoch / self.total_epoch + 1.) for base_lr in self.base_lrs]
            for param_group, lr in zip(self.optimizer.param_groups, warmup_lr):
                param_group['lr'] = lr
        else:
            if epoch is None:
                self.after_scheduler.step(metrics, None)
            else:
                self.after_scheduler.step(metrics, epoch - self.total_epoch)

    def step(self, epoch=None, metrics=None):
        if type(self.after_scheduler) != ReduceLROnPlateau:
            if self.finished and self.after_scheduler:
                if epoch is None:
                    self.after_scheduler.step(None)
                else:
                    self.after_scheduler.step(epoch - self.total_epoch)
                self._last_lr = self.after_scheduler.get_last_lr()
            else:
                return super(GradualWarmupScheduler, self).step(epoch)
        else:
            self.step_ReduceLROnPlateau(metrics, epoch)

参考：
https://github.com/ildoonet/pytorch-gradual-warmup-lr
https://aws.amazon.com/blogs/machine-learning/the-importance-of-hyperparameter-tuning-for-scaling-deep-learning-training-to-multiple-gpus/