对齐PyTorch,一文详解OneFlow的DataLoader实现

Posted OneFlow深度学习框架

tags:

篇首语:本文由小常识网(cha138.com)小编为大家整理,主要介绍了对齐PyTorch,一文详解OneFlow的DataLoader实现相关的知识,希望对你有一定的参考价值。

撰文 | 赵露阳

在最新的OneFlow v0.5.0版本中,我们增加了许多新特性,比如:

  • 新增动态图特性:OneFlow 默认以动态图模式(eager)运行,与静态图模式(graph)相比,更容易搭建网络、调试和验证算法。

  • 面向对象式的动态图接口 nn.Module,熟悉 PyTorch 的用户可以轻松上手。

  • “一行代码转换 OneFlow 与 PyTorch 网络”:与 PyTorch 对齐的算子数目增加至200+。在 ResNet50、AlexNet 等 十几个常用网络 上已通过 import oneflow as torch 和 import torch as flow 验证。注意:此特性是为方便用户由 PyTorch 迁移至 OneFlow 而设计,并不是承诺完全兼容 PyTorch。

  • 面向对象式的静态图接口:新增面向对象的静态图接口 nn.Graph。保留了 OneFlow 静态图性能优势的同时,让静态图的编程门槛与动态图接近,期待更多的算法工程师把 OneFlow 的高性能优势玩起来。这是一个用 nn.Graph 搭建 ResNet50 示例

  • 易用高效的分布式训练:分布式训练是大势所趋,OneFlow 本版本新增的 Consistent Tensor,让用户可以像操作单机单卡一样,操作整个集群,并立即看到效果。新增的 launch 模块、DDP 模块 配合 OneFlow 的一致性视角 让用户轻松启动分布式训练,无论是 数据并行、模型并行、还是流水并行,OneFlow 均原生支持,易用高效。

其中,最重要的新特性之一,就是OneFlow的动态图做到了几乎和PyTorch一致,从Tensor、nn.Module、到autograd、functional api等,其中也包括和torch几乎对齐的DataLoader/Dataset设计,笔者有幸开发了OneFlow中的这一模块。

https://github.com/Oneflow-Inc/oneflow/pull/5406
https://github.com/Oneflow-Inc/oneflow/pull/5500
https://github.com/Oneflow-Inc/oneflow/pull/5644
https://github.com/Oneflow-Inc/oneflow/pull/6280

本文将对OneFlow/PyTorch中的DataLoader原理、工作流程进行梳理:

  • dataloader简介

  • dataloader原理

  • dataloader工作流程

  • multiprocessing dataloader工作原理

1

简介

简单来说,DataLoader是深度学习中必不可少的,用于处理Dataset产生每个iter过程中批量数据和label的一种数据加载器。正如PyTorch文档中的描述:DataLoader,结合了Sampler、Dataset,提供了对某个dataset可迭代的数据集合。DataLoader支持单进程、多进程的加载数据集合。

2

dataloader原理


核心组建

  • Dataloader

  • Dataset

  • Sampler

  • Fetcher

DataLoader工作原理的简单总结:

1.Dataloader是负责数据加载的核心;DataLoaderIter是具体执行单位。dataloader进入到每一次iter中都会通过DataloaderIter来处理具体的数据加载过程;

2.Dataset是数据集的基类,任何自定义数据集都需要继承它并通过重写getitem方法来定义取数据的方式;

3.Sampler是负责index相关的采样器、每个iter迭代都会通过Sampler生成要采样的数据集的index;

4.Fetcher更像是数据的收集器。根据Sampler产生的batch个index去数据集中fetch对应的数据、并通过相应的collate_fn方法将获取的数据收集打包成最终可用的形式,返回给DataLoader。

使用示例

1.MNIST

下面用PyTorch官方examples的一个简单例子,用MNIST数据集训练分类网络来说明DataLoader的用法:

transform=transforms.Compose([
        transforms.ToTensor(),
        transforms.Normalize((0.1307,), (0.3081,))
        ])
dataset1 = datasets.MNIST('../data', train=True, download=True,
                          transform=transform)
dataset2 = datasets.MNIST('../data', train=False,
                          transform=transform)
train_loader = torch.utils.data.DataLoader(dataset1,**train_kwargs)
test_loader = torch.utils.data.DataLoader(dataset2, **test_kwargs)

可以看到,dataset1、dataset2分别是表示数据集的训练集、测试集。在PyTorch中是通过torchvision.datasets.MNIST定义的。MNIST继承自VisionDataset,而VisionDataset则继承自torch.utils.data.Dataset。在MNIST中,实现了数据集最重要的getitem方法,用于根据index取对应数据:

def __getitem__(self, index: int) -> Tuple[Any, Any]:
    """
        Args:
            index (int): Index
        Returns:
            tuple: (image, target) where target is index of the target class.
        """
    img, target = self.data[index], int(self.targets[index])

    # doing this so that it is consistent with all other datasets
    # to return a PIL Image
    img = Image.fromarray(img.numpy(), mode='L')

    if self.transform is not None:
        img = self.transform(img)

        if self.target_transform is not None:
            target = self.target_transform(target)

            return img, target

在OneFlow中,oneflow.utils.data对应torch.utils.data;flowvision对应torchvision,使用方式几乎完全一致。例如:对应MNIST数据集,即可直接通过flowvision.datasets.MNIST使用。

dataset1、dataset2定义完成后,传入分别用于训练、验证的dataloader(train_loader、test_loader)。之后,在train/test的循环中,即可迭代dataloader获取每个iter的数据和label:

def train(args, model, device, train_loader, optimizer, epoch):
    model.train()
    for batch_idx, (data, target) in enumerate(train_loader):
        data, target = data.to(device), target.to(device)
        optimizer.zero_grad()
        output = model(data)
        ....

2.ImageNet

这里还是用PyTorch官方examples里ImageNet数据集的训练为例:

train_dataset = datasets.ImageFolder(
        traindir,
        transforms.Compose([
            transforms.RandomResizedCrop(224),
            transforms.RandomHorizontalFlip(),
            transforms.ToTensor(),
            normalize,
        ]))

    if args.distributed:
        train_sampler = torch.utils.data.distributed.DistributedSampler(train_dataset)
    else:
        train_sampler = None

    train_loader = torch.utils.data.DataLoader(
        train_dataset, batch_size=args.batch_size, shuffle=(train_sampler is None),
        num_workers=args.workers, pin_memory=True, sampler=train_sampler)

    val_loader = torch.utils.data.DataLoader(
        datasets.ImageFolder(valdir, transforms.Compose([
            transforms.Resize(256),
            transforms.CenterCrop(224),
            transforms.ToTensor(),
            normalize,
        ])),
        batch_size=args.batch_size, shuffle=False,
        num_workers=args.workers, pin_memory=True)

可以看见,大体流程和上面的MNIST差不多:

1.先是构造Dataset,这里为通过datasets.ImageFolder构造。ImageFolder是用于读取/处理以文件夹形式存放的图片数据集:

class ImageFolder(DatasetFolder):
    r"""A generic data loader where the images are arranged in this way by default:
    .. code-block:: shell 
        root/dog/xxx.png
        root/dog/xxy.png
        root/dog/[...]/xxz.png
        root/cat/123.png
        root/cat/nsdf3.png
        root/cat/[...]/asd932_.png
    This class inherits from :class:`~vision.datasets.DatasetFolder` so
    the same methods can be overridden to customize the dataset.
    Args:
        root (string): Root directory path.
        transform (callable, optional): A function/transform that  takes in an PIL image
            and returns a transformed version. E.g, ``transforms.RandomCrop``
        target_transform (callable, optional): A function/transform that takes in the
            target and transforms it.
        loader (callable, optional): A function to load an image given its path.
        is_valid_file (callable, optional): A function that takes path of an Image file
            and check if the file is a valid file (used to check of corrupt files)
     Attributes:
        classes (list): List of the class names sorted alphabetically.
        class_to_idx (dict): Dict with items (class_name, class_index).
        imgs (list): List of (image path, class_index) tuples
    """

    def __init__(
        self,
        root: str,
        transform: Optional[Callable] = None,
        target_transform: Optional[Callable] = None,
        loader: Callable[[str], Any] = default_loader,
        is_valid_file: Optional[Callable[[str], bool]] = None,
    ):
        super(ImageFolder, self).__init__(
            root,
            loader,
            IMG_EXTENSIONS if is_valid_file is None else None,
            transform=transform,
            target_transform=target_transform,
            is_valid_file=is_valid_file,
        )
        self.imgs = self.samples

可以看到其继承自DatasetFolder、初始化时主要参数有:

  • root:图片文件夹路径

  • transform:对经过loader读取到的PIL图片,经过哪些transform处理,如上述的Resize、CenterCrop等

  • loader:一个用于根据path加载图片的图像加载器,通常默认的loader是PIL

DatasetFolder中实现了Dataset中最重要的getitem方法:

def __getitem__(self, index: int) -> Tuple[Any, Any]:
    """
        Args:
            index (int): Index
        Returns:
            tuple: (sample, target) where target is class_index of the target class.
        """
    path, target = self.samples[index]
    sample = self.loader(path)
    if self.transform is not None:
        sample = self.transform(sample)
        if self.target_transform is not None:
            target = self.target_transform(target)

            return sample, target

通过getitem定义了如何根据index取到相应数据的方式。

2.其次如果是多机分布式训练,则Sampler需要使用专门为分布式训练设计的DistributedSampler类(否则不用特殊设置,用默认的即可);这里还有个细节,训练集和验证集上,对dataset做了不同的transform,训练集用了RandomResizedCrop、RandomHorizontalFlip;验证集则是Resize、CenterCrop,经过transform后,最终通过ToTensor方法转化成Tensor。

3.构造用于训练、验证的Dataloader(train_loader、val_loader),后面的使用方式就很简单了,在train/eval的loop中直接使用即可:

for i, (images, target) in enumerate(train_loader):
        # measure data loading time
        data_time.update(time.time() - end)

        if args.gpu is not None:
            images = images.cuda(args.gpu, non_blocking=True)
        if torch.cuda.is_available():
            target = target.cuda(args.gpu, non_blocking=True)
        .....


3

dataloader工作流程

下面结合代码看一下主要流程:

Dataset 

任何自定义数据集,必须继承Dataset类并实现_getitem__方法,用于定义根据传入的index获取数据的方式。同时,自定义数据集也可选重写len方法,用于判断数据集的size。

class Dataset(Generic[T_co]):
    r"""An abstract class representing a :class:`Dataset`.
    All datasets that represent a map from keys to data samples should subclass
    it. All subclasses should overwrite :meth:`__getitem__`, supporting fetching a
    data sample for a given key. Subclasses could also optionally overwrite
    :meth:`__len__`, which is expected to return the size of the dataset by many
    :class:`~flow.utils.data.Sampler` implementations and the default options
    of :class:`~flow.utils.data.DataLoader`.
    .. note::
      :class:`~flow.utils.data.DataLoader` by default constructs a index
      sampler that yields integral indices.  To make it work with a map-style
      dataset with non-integral indices/keys, a custom sampler must be provided.
    """

    def __getitem__(self, index) -> T_co:
        raise NotImplementedError

    def __add__(self, other: "Dataset[T_co]") -> "ConcatDataset[T_co]":
        return ConcatDataset([self, other])

DataLoader 

DataLoader是整个数据处理过程的核心。

class DataLoader(Generic[T_co]):
    def __init__(
            self,
            dataset: Dataset[T_co],
            batch_size: Optional[int] = 1,
            shuffle: bool = False,
            sampler: Optional[Sampler[int]] = None,
            batch_sampler: Optional[Sampler[Sequence[int]]] = None,
            num_workers: int = 0,
            collate_fn: Optional[_collate_fn_t] = None,
            drop_last: bool = False,
            timeout: float = 0,
            worker_init_fn: Optional[_worker_init_fn_t] = None,
            multiprocessing_context=None,
            generator=None,
            *,
            prefetch_factor: int = 2,
            persistent_workers: bool = False
        ):

    ...
    ...
# We quote '_BaseDataLoaderIter' since it isn't defined yet and the definition can't be moved up
    # since '_BaseDataLoaderIter' references 'DataLoader'.
    def __iter__(self) -> "_BaseDataLoaderIter":
        # When using a single worker the returned iterator should be
        # created everytime to avoid reseting its state
        # However, in the case of a multiple workers iterator
        # the iterator is only created once in the lifetime of the
        # DataLoader object so that workers can be reused
        if self.persistent_workers and self.num_workers > 0:
            if self._iterator is None:
                self._iterator = self._get_iterator()
            else:
                self._iterator._reset(self)
            return self._iterator
        else:
            return self._get_iterator()

def _get_iterator(self) -> "_BaseDataLoaderIter":
        if self.num_workers == 0 or self.num_workers == 1:
            return _SingleProcessDataLoaderIter(self)
        else:
            self.check_worker_number_rationality()
            return _MultiProcessingDataLoaderIter(self)

DataLoader在每一个iter迭代过程中,最重要的就是通过上面的__iter__方法完成取数据和label。__iter__里通过_get_iterator方法获取相应的DataLoaderIter实例。

  • 在单进程下,即_SingleProcessDataLoaderIter

  • 多进程下,即_MultiProcessingDataLoaderIter,他们都继承自_BaseDataLoaderIter

DataLoaderIter 

DataLoaderIter负责DataLoader在每个迭代中具体事务的处理。

class _BaseDataLoaderIter(object):
    def __init__(self, loader: DataLoader) -> None:
        self._dataset = loader.dataset
        self._dataset_kind = loader._dataset_kind
        self._IterableDataset_len_called = loader._IterableDataset_len_called
        self._auto_collation = loader._auto_collation
        self._drop_last = loader.drop_last
        self._index_sampler = loader._index_sampler
        self._num_workers = loader.num_workers
        self._prefetch_factor = loader.prefetch_factor
        self._pin_memory = False
        self._timeout = loader.timeout
        self._collate_fn = loader.collate_fn
        self._sampler_iter = iter(self._index_sampler)
        self._base_seed = flow.tensor([0], dtype=flow.int64).uniform_().numpy().item()
        # TODO: flow.empty()
        # self._base_seed = flow.empty((), dtype=flow.int64).random_(generator=loader.generator).item()
        self._persistent_workers = loader.persistent_workers
        self._num_yielded = 0
        self._profile_name = "enumerate(DataLoader)#{}.__next__".format(
            self.__class__.__name__
        )

    def __iter__(self) -> "_BaseDataLoaderIter":
        return self

    def _reset(self, loader, first_iter=False):
        self._sampler_iter = iter(self._index_sampler)
        self._num_yielded = 0
        self._IterableDataset_len_called = loader._IterableDataset_len_called

    def _next_index(self):
        return next(self._sampler_iter)  # may raise StopIteration

    def _next_data(self):
        raise NotImplementedError

    def __next__(self) -> Any:
        if self._sampler_iter is None:
            self._reset()
        data = self._next_data()
        self._num_yielded += 1
        if (
            self._dataset_kind == _DatasetKind.Iterable
            and self._IterableDataset_len_called is not None
            and self._num_yielded > self._IterableDataset_len_called
        ):
            warn_msg = (
                "Length of IterableDataset {} was reported to be {} (when accessing len(dataloader)), but {} "
                "samples have been fetched. "
            ).format(self._dataset, self._IterableDataset_len_called, self._num_yielded)
            if self._num_workers > 1:
                warn_msg += "Multiprocessing dataloader is not support yet!"
            warnings.warn(warn_msg)
        return data

    def __len__(self) -> int:
        return len(self._index_sampler)

    def __getstate__(self):
        raise NotImplementedError("{} cannot be pickled", self.__class__.__name__)


class _SingleProcessDataLoaderIter(_BaseDataLoaderIter):
    def __init__(self, loader):
        super(_SingleProcessDataLoaderIter, self).__init__(loader)
        assert self._timeout == 0
        assert 0 <= self._num_workers <= 1

        self._dataset_fetcher = _DatasetKind.create_fetcher(
            self._dataset_kind,
            self._dataset,
            self._auto_collation,
            self._collate_fn,
            self._drop_last,
        )

    def _next_data(self):
        index = self._next_index()  # may raise StopIteration
        if self._pin_memory:
            raise NotImplementedError("Dataloader pin memory is not support yet!")
        return self._dataset_fetcher.fetch(index)

在每一个iter迭代时,会调用_BaseDataLoaderIter的__next__方法,进而调用自类实现的_next_data方法获取数据。以_SingleProcessDataLoaderIter为例:

  • index = self._next_index()通过Sampler获取此次迭代的数据集索引;

  • self._dataset_fetcher.fetch(index)Fetcher根据index索引取相应的数据。

Fetcher 

Fetcher作为数据收集器,会根据Sampler产生的batch的index,来从数据集中切分、收集、打包成完整可用的一个batch的数据,并返回给DataLoader使用。

class _MapDatasetFetcher(_BaseDatasetFetcher):
    def __init__(self, dataset, auto_collation, collate_fn, drop_last):
        super(_MapDatasetFetcher, self).__init__(
            dataset, auto_collation, collate_fn, drop_last
        )

    def fetch(self, possibly_batched_index):
        if self.auto_collation:
            data = [self.dataset[idx] for idx in possibly_batched_index]
        else:
            data = self.dataset[possibly_batched_index]
        return self.collate_fn(data)

Fetcher这里和DataLoaderIter(BaseDataLoaderIter)_类似,_都有一个基类的实现BaseDatasetFetcher。根据不同的数据类型,进入到不同的子类实现中,这里以常用的_MapDatasetFetcher的子类实现为例,看一下Fetcher的主要工作。

可以看见,主要就是:

  • data = [self.dataset[idx] for idx in possibly_batched_index]

  • return self.collate_fn(data)

1.根据传入的batch个index列表,去dataset中去切分相应的数据,返回的是取出后的batch个数据的列表;

2.根据传入的或自定义的collate_fn方法,收集处理这batch个数据,并打包成训练/验证时可直接使用的Tensor。

4

multiprocessing dataloader工作原理

原理

普通的单进程DataLoader在处理每个iter的数据处理是iter-by-iter且同步的,受制于Python没有实际上的多线程执行,所以单进程的DataLoader通常是比较慢的。多进程DataLoader,即通过Python的multiprocessing开启多个Python的worker进程,譬如开启4个worker进程后,理论上每单位时间可以处理4个iter的数据集,加速数据处理/加载的过程。

单进程DataLoader下,由于数据处理是iter-by-iter的,下一个iter的处理需要等待当前iter完成后才可开始;多进程DataLoader和单进程DataLoader的主要区别就在于可以通过Python的multiprocessing模块,启动多个worker进程加速这个过程。

这里以4进程的DataLoader为例:

DataLoader的主线程将当前iter的任务下发给worker1之后,再下发下一个iter的任务给worker2....直至下发第4个iter的处理任务给worker4。这一步骤主要在dataloader.py的L1024-L1026中实现:

# prime the prefetch loop
for _ in range(self._prefetch_factor * self._num_workers):
    self._try_put_index()

陆续发送完index后,这4个worker可以并行的工作,陆续完成自己iter的处理任务后,将结果塞入一个Queue队列中,DataLoader的主线程从队列中取数据即可。

具体到每个worker的工作流程,其实和单进程的DataLoader工作流程是类似的,下面主要介绍下多进程和单进程DataLoader的区别,以及多个worker之间是如何协同工作的。

工作流程

_MultiProcessingDataLoaderIter

def _next_data(self):
  # DataLoaderIter通过此方法获取每个iter的数据,主要调用_get_data实现

def _get_data(self):
  # _get_data方法中,主要通过调用_try_get_data()获取数据

def _try_get_data(self, timeout=_utils.MP_STATUS_CHECK_INTERVAL):
  # 从主进程的_data_queue中获取数据
  ...
  try:
      data = self._data_queue.get(timeout=timeout)
      return (True, data)
  except Exception as e:
     ...

def _process_data(self, data):
  # 主要工作即:1.通过_try_put_index()来将下一个iter的index放入一个活跃的worker进程中
  # 2.同时标记_rcvd_idx,使其增加1。
  self._rcvd_idx += 1
  self._try_put_index()
  if isinstance(data, ExceptionWrapper):
    data.reraise()
    return data

def _try_put_index(self):
  # 主要工作即遍历所有workers,找到第一个活跃的worker(worker_queue_idx标识)
  # 将index和_send_idx信息放入此worker的index_queue中
  # 每个worker拥有独立的index_queue,收到index_queue的信息后即开始工作
  assert self._tasks_outstanding < self._prefetch_factor * self._num_workers
  try:
    index = self._next_index()
    except StopIteration:
      return
    for _ in range(self._num_workers):  # find the next active worker, if any
      worker_queue_idx = next(self._worker_queue_idx_cycle)
      if self._workers_status[worker_queue_idx]:
        break
    else:
      # not found (i.e., didn't break)
      return

    self._index_queues[worker_queue_idx].put((self._send_idx, index))
    self._task_info[self._send_idx] = (worker_queue_idx,)      
    self._tasks_outstanding += 1
    self._send_idx += 1

_next_data()

⬇️

_get_data() ➡️ _try_get_data()

⬇️

_process_data() ➡️ _try_put_index()

每个worker独立工作,主要代码在oneflow/python/oneflow/utils/data/_utils/worker.py的_worker_loop()方法中:

while watchdog.is_alive():
            try:
                r = index_queue.get(timeout=MP_STATUS_CHECK_INTERVAL)
            except queue.Empty:
                continue
            if isinstance(r, _ResumeIteration):
                # Acknowledge the main process
                data_queue.put((r, None))
                iteration_end = False
                # Recreate the fetcher for worker-reuse policy
                fetcher = _DatasetKind.create_fetcher(
                    dataset_kind, dataset, auto_collation, collate_fn, drop_last
                )
                continue
            elif r is None:
                # Received the final signal
                assert done_event.is_set() or iteration_end
                break
            elif done_event.is_set() or iteration_end:
                # `done_event` is set. But I haven't received the final signal
                # (None) yet. I will keep continuing until get it, and skip the
                # processing steps.
                continue
            idx, index = r
            data: Union[_IterableDatasetStopIteration, ExceptionWrapper]

            if init_exception is not None:
                data = init_exception
                init_exception = None
            else:
                try:
                    data = fetcher.fetch(index)
                except Exception as e:
                    if (
                        isinstance(e, StopIteration)
                        and dataset_kind == _DatasetKind.Iterable
                    ):
                        data = _IterableDatasetStopIteration(worker_id)
                        # Set `iteration_end`
                        #   (1) to save future `next(...)` calls, and
                        #   (2) to avoid sending multiple `_IterableDatasetStopIteration`s.
                        iteration_end = True
                    else:
                        # It is important that we don't store exc_info in a variable.
                        # `ExceptionWrapper` does the correct thing.
                        # See NOTE [ Python Traceback Reference Cycle Problem ]
                        data = ExceptionWrapper(
                            where="in DataLoader worker process {}".format(worker_id)
                        )
            data_queue.put((idx, data))
            del data, idx, index, r  # save memory
    except KeyboardInterrupt:
        # Main process will raise KeyboardInterrupt anyways.
        pass

每个worker在自己的worker loop中,一旦

r = index_queue.get(timeout=MP_STATUS_CHECK_INTERVAL)获取index_queue中的index数据,就会开始工作:

idx, index = r >> data = fetcher.fetch(index) 这部分内容和之前描述的单进程DataLoader的工作流程没有区别。

当获取到处理完成的数据data后,会将其放入到data loader main线程的data_queue中: data_queue.put((idx, data)) 等待DataLoader主线程从queue中获取结果。

以上即为多进程DataLoader的主要工作流程。

5

结语

本文梳理总结了DataLoader/Dataset,希望能对大家了解OneFlow/PyTorch动态图模式下的DataLoader/Dataset工作原理有所帮助。 

对齐PyTorch的DataLoader/Dataset只是第一步,后续仍然面临着效率瓶颈等问题,因为即使使用了multiprocess的DataLoader,在某些情况下,图像解码、Python下调用C++ op执行各种transform时仍可能遭遇性能问题,造成训练过程中GPU打不满/等待CPU数据处理等情况,后续需要考虑更高效的解决方案(如Dali等)。

其他人都在看

点击“阅读原文”,欢迎下载体验OneFlow新一代开源深度学习框架

以上是关于对齐PyTorch,一文详解OneFlow的DataLoader实现的主要内容,如果未能解决你的问题,请参考以下文章

深度学习框架如何优雅的做算子对齐任务?

深度学习框架如何优雅的做算子对齐任务?

兼容PyTorch,25倍性能加速,OneFlow“超速”了

以OneFlow为例梳理深度学习框架的那些插值方法

OneFlow如何做静态图的算子对齐任务

适配PyTorch FX,OneFlow让量化感知训练更简单