TorchData

Why TorchData? | Install guide | What are DataPipes? | Beta Usage and Feedback | Contributing | Future Plans

This library is currently in the Beta stage and new features are under active development. The API may change based on user feedback or performance. We are committed to bring this library to stable release, but a few 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.

torchdata is a library of common modular data loading primitives for easily constructing flexible and performant data pipelines.

This library introduces composable Iterable-style and Map-style building blocks called DataPipes that work well out of the box with the PyTorch's DataLoader. These built-in DataPipes have the necessary functionalities to reproduce many different datasets in TorchVision and TorchText, namely loading files (from local or cloud), parsing, caching, transforming, filtering, and many more utilities. To understand the basic structure of DataPipes, please see What are DataPipes? below, and to see how DataPipes can be practically composed together into datasets, please see our examples.

On top of DataPipes, this library provides a new DataLoader2 that allows the execution of these data pipelines in various settings and execution backends (ReadingService). You can learn more about the new version of DataLoader2 in our full DataLoader2 documentation. Additional features are work in progres, such as checkpointing and advanced control of randomness and determinism.

Note that because many features of the original DataLoader have been modularized into DataPipes, their source codes live as standard DataPipes in pytorch/pytorch rather than torchdata to preserve backward-compatibility support and functional parity within torch. Regardless, you can to them by importing them from torchdata.

Why composable data loading?

Over many years of feedback and organic community usage of the PyTorch DataLoader and Dataset, we've found that:

1. The original DataLoader bundled too many features together, making them difficult to extend, manipulate, or replace. This has created a proliferation of use-case specific DataLoader variants in the community rather than an ecosystem of interoperable elements.
2. Many libraries, including each of the PyTorch domain libraries, have rewritten the same data loading utilities over and over again. We can save OSS maintainers time and effort rewriting, debugging, and maintaining these commonly used elements.

These reasons inspired the creation of DataPipe and DataLoader2, with a goal to make data loading components more flexible and reusable.

Installation

Version Compatibility

The following is the corresponding torchdata versions and supported Python versions.

torch	torchdata	python
master / nightly	main / nightly	>=3.8, <=3.11
1.13.1	0.5.1	>=3.7, <=3.10
1.12.1	0.4.1	>=3.7, <=3.10
1.12.0	0.4.0	>=3.7, <=3.10
1.11.0	0.3.0	>=3.7, <=3.10

Colab

Follow the instructions in this Colab notebook. The notebook also contains a simple usage example.

Local pip or conda

First, set up an environment. We will be installing a PyTorch binary as well as torchdata. If you're using conda, create a conda environment:

conda create --name torchdata
conda activate torchdata

If you wish to use venv instead:

python -m venv torchdata-env
source torchdata-env/bin/activate

Install torchdata:

Using pip:

pip install torchdata

Using conda:

conda install -c pytorch torchdata

You can then proceed to run our examples, such as the IMDb one.

From source

pip install .

If you'd like to include the S3 IO datapipes and aws-sdk-cpp, you may also follow the instructions here

In case building TorchData from source fails, install the nightly version of PyTorch following the linked guide on the contributing page.

From nightly

The nightly version of TorchData is also provided and updated daily from main branch.

Using pip:

pip install --pre torchdata --extra-index-url https://download.pytorch.org/whl/nightly/cpu

Using conda:

conda install torchdata -c pytorch-nightly

What are DataPipes?

Early on, we observed widespread confusion between the PyTorch Dataset which represented reusable loading tooling (e.g. TorchVision's ImageFolder), and those that represented pre-built iterators/accessors over actual data corpora (e.g. TorchVision's ImageNet). This led to an unfortunate pattern of siloed inheritance of data tooling rather than composition.

DataPipe is simply a renaming and repurposing of the PyTorch Dataset for composed usage. A DataPipe takes in some access function over Python data structures, __iter__ for IterDataPipes and __getitem__ for MapDataPipes, and returns a new access function with a slight transformation applied. For example, take a look at this JsonParser, which accepts an IterDataPipe over file names and raw streams, and produces a new iterator over the filenames and deserialized data:

import json

class JsonParserIterDataPipe(IterDataPipe):
    def __init__(self, source_datapipe, **kwargs) -> None:
        self.source_datapipe = source_datapipe
        self.kwargs = kwargs

    def __iter__(self):
        for file_name, stream in self.source_datapipe:
            data = stream.read()
            yield file_name, json.loads(data, **self.kwargs)

    def __len__(self):
        return len(self.source_datapipe)

You can see in this example how DataPipes can be easily chained together to compose graphs of transformations that reproduce sophisticated data pipelines, with streamed operation as a first-class citizen.

Under this naming convention, Dataset simply refers to a graph of DataPipes, and a dataset module like ImageNet can be rebuilt as a factory function returning the requisite composed DataPipes. Note that the vast majority of built-in features are implemented as IterDataPipes, we encourage the usage of built-in IterDataPipe as much as possible and convert them to MapDataPipe only when necessary.

DataLoader2

A new, light-weight DataLoader2 is introduced to decouple the overloaded data-manipulation functionalities from torch.utils.data.DataLoader to DataPipe operations. Besides, certain features can only be achieved with DataLoader2, such as like checkpointing/snapshotting and switching backend services to perform high-performant operations.

Please read the full documentation here.

Tutorial

A tutorial of this library is available here on the documentation site. It covers four topics: using DataPipes, working with DataLoader, implementing DataPipes, and working with Cloud Storage Providers.

There is also a tutorial available on how to work with the new DataLoader2.

Usage Examples

We provide a simple usage example in this Colab notebook. It can also be downloaded and executed locally as a Jupyter notebook.

In addition, there are several data loading implementations of popular datasets across different research domains that use DataPipes. You can find a few selected examples here.

Frequently Asked Questions (FAQ)

What should I do if the existing set of DataPipes does not do what I need?

You can implement your own custom DataPipe. If you believe your use case is common enough such that the community can benefit from having your custom DataPipe added to this library, feel free to open a GitHub issue. We will be happy to discuss!

What happens when the Shuffler DataPipe is used with DataLoader?

In order to enable shuffling, you need to add a Shuffler to your DataPipe line. Then, by default, shuffling will happen at the point where you specified as long as you do not set shuffle=False within DataLoader.

What happens when the Batcher DataPipe is used with DataLoader?

If you choose to use Batcher while setting batch_size > 1 for DataLoader, your samples will be batched more than once. You should choose one or the other.

Why are there fewer built-in MapDataPipes than IterDataPipes?

By design, there are fewer MapDataPipes than IterDataPipes to avoid duplicate implementations of the same functionalities as MapDataPipe. We encourage users to use the built-in IterDataPipe for various functionalities, and convert it to MapDataPipe as needed.

How is multiprocessing handled with DataPipes?

Multi-process data loading is still handled by the DataLoader, see the DataLoader documentation for more details. As of PyTorch version >= 1.12.0 (TorchData version >= 0.4.0), data sharding is automatically done for DataPipes within the DataLoader as long as a ShardingFilter DataPipe exists in your pipeline. Please see the tutorial for an example.

What is the upcoming plan for DataLoader?

DataLoader2 is in the prototype phase and more features are actively being developed. Please see the README file in torchdata/dataloader2. If you would like to experiment with it (or other prototype features), we encourage you to install the nightly version of this library.

Why is there an Error saying the specified DLL could not be found at the time of importing portalocker?

It only happens for people who runs torchdata on Windows OS as a common problem with pywin32. And, you can find the reason and the solution for it in the link.

Contributing

We welcome PRs! See the CONTRIBUTING file.

Beta Usage and Feedback

We'd love to hear from and work with early adopters to shape our designs. Please reach out by raising an issue if you're interested in using this tooling for your project.

License

TorchData is BSD licensed, as found in the LICENSE file.