Skip to content

TorchData 0.6.0 Release Notes
Compare
Choose a tag to compare
@NivekT NivekT released this 15 Mar 19:38
· 10 commits to release/0.6 since this release
v0.6.0
5bbcd77

TorchData 0.6.0 Beta Release Notes

Highlights

We are excited to announce the release of TorchData 0.6.0. This release is composed of about 130 commits since 0.5.0, made by 27 contributors. We want to sincerely thank our community for continuously improving TorchData.

TorchData 0.6.0 updates are primarily focused on DataLoader2. We graduate some of its APIs from the prototype stage and introduce additional features. Highlights include:

  • Graduation of MultiProcessingReadingService from prototype to beta
    • This is the default ReadingService that we expect most users to use; it closely aligns with the functionalities of old DataLoader with improvements
    • With this graduation, we expect the APIs and behaviors to be mostly stable going forward. We will continue to add new features as they become ready.
  • Introduction of Sequential ReadingService
    • Enables the usage of multiple ReadingServices at the same time
  • Adding comprehensive tutorial of DataLoader2 and its subcomponents

Backwards Incompatible Change

DataLoader2

  • Officially graduate PrototypeMultiProcessingReadingService to MultiProcessingReadingService (#1009)
    • The APIs of MultiProcessingReadingService as well as the internal implementation have changed. Overall, this should provide a better user experience.
    • Please refer to our documentation for details.

0.5.00.6.0
It previously took the following arguments: 
MultiProcessingReadingService(
    num_workers: int = 0,
    pin_memory: bool = False,
    timeout: float = 0,
    worker_init_fn: Optional[Callable[[int], None]] = None,
    multiprocessing_context=None,
    prefetch_factor: Optional[int] = None,
    persistent_workers: bool = False,
)
The new version takes these arguments: 
MultiProcessingReadingService(
    num_workers: int = 0,
    multiprocessing_context: Optional[str] = None,
    worker_prefetch_cnt: int = 10,
    main_prefetch_cnt: int = 10,
    worker_init_fn: Optional[Callable[[DataPipe, WorkerInfo], DataPipe]] = None,
    worker_reset_fn: Optional[Callable[[DataPipe, WorkerInfo, SeedGenerator], DataPipe]] = None,
)

  • Deep copy ReadingService during DataLoader2 initialization (#746)
    • Within DataLoader2, a deep copy of the passed-in ReadingService object is created during initialization and will be subsequently used.
    • This prevents multiple DataLoader2s from accidentally sharing states when the same ReadingService object is passed into them.

0.5.00.6.0
Previously, a ReadingService object that is used in multiple DataLoader2 shared state among them. 
>>> dp = IterableWrapper([0, 1, 2, 3, 4])
>>> rs = MultiProcessingReadingService(num_workers=2)
>>> dl1 = DataLoader2(dp, reading_service=rs)
>>> dl2 = DataLoader2(dp, reading_service=rs)
>>> next(iter(dl1))
>>> print(f"Number of processes that exist in `dl1`'s RS after initializing `dl1`: {len(dl1.reading_service._worker_processes)}")
# Number of processes that exist in `dl1`'s RS after initializing `dl1`: 2
>>> next(iter(dl2))
# Note that we are still examining `dl1.read_service` below
>>> print(f"Number of processes that exist in `dl1`'s RS after initializing `dl2`: {len(dl1.reading_service._worker_processes)}")
# Number of processes that exist in `dl1`'s RS after initializing `dl1`: 4
DataLoader2 now deep copies the ReadingService object during initialization and the ReadingService state is no longer shared. 
>>> dp = IterableWrapper([0, 1, 2, 3, 4])
>>> rs = MultiProcessingReadingService(num_workers=2)
>>> dl1 = DataLoader2(dp, reading_service=rs)
>>> dl2 = DataLoader2(dp, reading_service=rs)
>>> next(iter(dl1))
>>> print(f"Number of processes that exist in `dl1`'s RS after initializing `dl1`: {len(dl1.reading_service._worker_processes)}")
# Number of processes that exist in `dl1`'s RS after initializing `dl1`: 2
>>> next(iter(dl2))
# Note that we are still examining `dl1.read_service` below
>>> print(f"Number of processes that exist in `dl1`'s RS after initializing `dl2`: {len(dl1.reading_service._worker_processes)}")
# Number of processes that exist in `dl1`'s RS after initializing `dl1`: 2

Deprecations

DataPipe

In PyTorch Core

  • Remove previously deprecated FileLoaderDataPipe (#89794)
  • Mark imports from torch.utils.data.datapipes.iter.grouping as deprecated (#94527)

TorchData

  • Remove certain deprecated functional APIs as previously scheduled (#890)

Releng

  • Drop support for Python 3.7 as aligned with PyTorch core library (#974)

New Features

DataLoader2

  • Add graph function to list DataPipes from DataPipe graphs (#888)
  • Add functions to set seeds to DataPipe graphs (#894)
  • Add worker_init_fn and worker_reset_fn to MultiProcessingReadingService (#907)
  • Add round robin sharding to support non-replicable DataPipe for MultiProcessing (#919)
    • Guarantee that DataPipes execute reset_iterator when all loops have received reset request in the dispatching process (#994)
  • Add initial support for randomness control within DataLoader2 (#801)
  • Add support for Sequential ReadingService (commit)
  • Enable SequentialReadingService to support MultiProcessing + Distributed (#985)
  • Add limit, pause, resume operations to halt DataPipes in DataLoader2 (#879)

DataPipe

  • Add ShardExpander IterDataPipe (#405)
  • Add RoundRobinDemux IterDataPipe (#903)
  • Implement PinMemory IterDataPipe (#1014)

Releng

  • Add conda Python 3.11 builds (#1010)
  • Enable Python 3.11 conda builds for Mac/Windows (#1026)
  • Update C++ standard to 17 (#1051)

Improvements

DataLoader2

In PyTorch Core

  • Fix apply_sharding to accept one sharding_filter per branch (#90769)

TorchData

  • Consolidate checkpoint contract with checkpoint component (#867)
  • Update load_state_dict() signature to align with TorchSnapshot (#887)
  • Apply sharding based on priority and combine DistInfo and ExtraInfo (used to store distributed metadata) (#916)
  • Prevent reset iteration message from being sent to workers twice (#917)
  • Add support to keep non-replicable DataPipe in the main process (#950)
  • Safeguard DataLoader2Iterator's __getattr__ method (#1004)
  • Forward worker exceptions and have DataLoader2 exit with them (#1003)
  • Attach traceback to Exception and test dispatching process (#1036)

DataPipe

In PyTorch Core

  • Add auto-completion to DataPipes in REPLs (e.g. Jupyter notebook) (#86960)
  • Add group support to sharding_filter (#88424)
  • Add keep_key option to Grouper (#92532)

TorchData

  • Add a masks option to filter files in S3 DataPipe (#880)
  • Make HeaderIterDataPipe with limit=None a no-op (#908)
  • Update fsspec DataPipe to be compatible with the latest version of fsspec (#957)
  • Expand the possible input options for HuggingFace DataPipe (#952)
  • Improve exception handling/skipping in online DataPipes (#968)
  • Allow the option to place key in output in MapKeyZipper (#1042)
  • Allow single key option for Slicer (#1041)

Releng

  • Add pure Python platform-agnostic wheel (#988)

Bug Fixes

DataLoader2

In PyTorch Core

  • Change serialization wrapper implementation to be an iterator (#87459)

DataPipe

In PyTorch Core

  • Fix type checking to accept both Iter and Map DataPipe (#87285)
  • Fix: Make __len__ of datapipes dynamic (#88302)
  • Properly cleanup unclosed files within generator function (#89973)
  • Remove iterator depletion in Zipper (#89974)

TorchData

  • Fix to_graph DataPipeGraph visualization function (#872)
  • Make lengths of DataPipe dynamic (#873)
  • Fix max_token_bucketize to accept incomparable data (#883)
  • Fix S3FileLoader local file clobbering (#895)
  • Fix fsspec DataPipe for paths starting with az:// (#849)
  • Properly cleanup unclosed files within generator function (#910)

Performance

DataLoader2

  • Add minimal, reproducible AWS S3 benchmark (#847)

Docs

DataLoader2

  • Add Distributed ReadingService DataLoader2 training loop example (#863)
  • Update README and documentation with latest changes (#954)
  • Update Colab example with DataLoader2 content (#979)
  • Add initial DataLoader2 Tutorial (#980)
  • Add LAION-5B Example with DataLoader2 (#1034)
  • Add Round Robin Sharding documentation (#1050)

DataPipe

  • Add pin_memory to documentation (#1046)

Releng

  • Fix links in README (#995)
  • Fix links in contribution guide (#995)

Devs

DataPipe

In PyTorch Core

  • Add container template for _Fork and _Demux (#89216)
  • Refactor sharding data pipe into a separate file (#94095)
  • Fix interface generation in setup.py (#87081)

TorchData

  • Add tests to validate iteration over combining DataPipe with infinite input (#912)

Releng

  • Update GHA version to utilize Node16 (#830)
  • Enable usage of sphinx doctest (#850)
  • Update submodule (#955)
  • Make portalocker optional dependency (#1007)

Future Plans

For DataLoader2, we are actively developing new features such as the checkpointing and the ability to execute part of the DataPipe graph on a single process before dispatching the outputs to worker processes. You may begin to see some of these features in nightly builds. We expect them to be part of the next release.

We welcome feedback and feature requests (let us know your use cases!). We always welcome potential contributors.

Beta Usage Note

This library is currently in the Beta stage and currently does not have a fully stable release. The API may change based on user feedback or performance. We are committed to bring this library to stable release, but future changes may not be completely backward compatible. If you install from source or use the nightly version of this library, use it along with the PyTorch nightly binaries. If you have suggestions on the API or use cases you'd like to be covered, please open a GitHub issue. We'd love to hear thoughts and feedback.

Assets 2
Loading