TorchData (🚨 Warning: Unstable Prototype 🚨)

Why torchdata? | Install guide | What are DataPipes? | Prototype Usage and Feedback | Contributing | Future Plans

This is a prototype library currently under heavy development. It does not currently have stable releases, and as such will likely be modified significantly in BC-breaking ways until beta release (targeting early 2022), and can only be used 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 prototype library of common modular data loading primitives for easily constructing flexible and performant data pipelines.

It aims to provide composable iter-style and map-style building blocks called DataPipes that work well out of the box with the PyTorch DataLoader. Right now it only contains basic functionality to reproduce several datasets in TorchVision and TorchText, namely including loading, parsing, caching, and several other utilities (e.g. hash checking). We plan to expand and harden this set considerably over the coming months.

To understand the basic structure of DataPipes, please see What are DataPipes? below, and to see how DataPipes can be practically composed into datasets, please see our examples/ directory.

Note that because many features of the original DataLoader have been modularized into DataPipes, some now live as standard DataPipes in pytorch/pytorch rather than torchdata to preserve BC functional parity within torch.

Why composable data loading?

Over many years of feedback and organic community usage of the PyTorch DataLoader and DataSets, 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 table-stakes elements.

Installation

Colab

Follow the instructions in this Colab notebook

Local pip or conda

First, set up an environment. We will be installing a nightly 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

Next, install one of the following following PyTorch nightly binaries.

# For CUDA 10.2
pip install --pre torch -f https://download.pytorch.org/whl/nightly/cu102/torch_nightly.html
# For CUDA 11.1
pip install --pre torch -f https://download.pytorch.org/whl/nightly/cu111/torch_nightly.html
# For CPU-only build
pip install --pre torch -f https://download.pytorch.org/whl/nightly/cpu/torch_nightly.html

If you already have a nightly of PyTorch installed and wanted to upgrade it (recommended!), append --upgrade to one of those commands.

Install torchdata:

pip install --user "git+https://github.com/pytorch/data.git"

Run a quick sanity check in python:

from torchdata.datapipes.iter import HttpReader
URL = "https://raw.githubusercontent.com/mhjabreel/CharCnn_Keras/master/data/ag_news_csv/train.csv"
ag_news_train = HttpReader([URL]).parse_csv().map(lambda t: (int(t[0]), " ".join(t[1:])))
agn_batches = ag_news_train.batch(2).map(lambda batch: {'labels': [sample[0] for sample in batch],\
                                      'text': [sample[1].split() for sample in batch]})
batch = next(iter(agn_batches))
assert batch['text'][0][0:8] == ['Wall', 'St.', 'Bears', 'Claw', 'Back', 'Into', 'the', 'Black']

From source

$ pip install -e git+https://github.com/pytorch/data#egg=torchdata

What are DataPipes?

Early on, we observed widespread confusion between the PyTorch DataSets 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 inheritence 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):
        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)

    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 sohpisticated 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 initial support is focused on IterDataPipes, while more MapDataPipes support will come later.

Implementing DataPipes

As a guiding example, let's implement an IterDataPipe that applies a callable to the input iterator. For MapDataPipes, take a look at the map folder for examples, and follow the steps below for the __getitem__ method instead of __iter__.

Naming

The naming convention for DataPipes is "Operation"-er, followed by IterDataPipe or MapDataPipe, as each DataPipe is essentially a container to apply an operation to data yielded from a source DataPipe. For succintness, we alias to just "Operation-er" in init files. For our IterDataPipe example, we'll name the module MapperIterDataPipe and alias it as iter.Mapper under datapipes.

Constructor

DataSets are now generally constructed as stacks of DataPipes, so each DataPipe typically takes a source DataPipe as its first argument.

class MapperIterDataPipe(IterDataPipe):
    def __init__(self, dp, fn):
        super().__init__()
        self.dp = dp
        self.fn = fn

Note:

• Avoid loading data from the source DataPipe in __init__ function, in order to support lazy data loading and save memory.
• If IterDataPipe instance holds data in memory, please be ware of the in-place modification of data. When second iterator is created from the instance, the data may have already changed. Please take IterableWrapper class as reference to deepcopy data for each iterator.

Iterator

For IterDataPipes, an __iter__ function is needed to consume data from the source IterDataPipe then apply the operation over the data before yield.

class MapperIterDataPipe(IterDataPipe):
    ...

    def __iter__(self):
        for d in self.dp:
            yield self.fn(d)

Length

In many cases, as in our MapperIterDataPipe example, the __len__ method of a DataPipe returns the length of the source DataPipe.

class MapperIterDataPipe(IterDataPipe):
    ...

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

However, note that __len__ is optional for IterDataPipe and often inadvisable. For CSVParserIterDataPipe in the using DataPipes section below, __len__ is not implemented because the number of rows in each file is unknown before loading it. In some special cases, __len__ can be made to either return an integer or raise an Error depending on the input. In those cases, the Error must be a TypeError to support Python's build-in functions like list(dp).

Registering DataPipes with the functional API

Each DataPipe can be registered to support functional invocation using the decorator functional_datapipe.

@functional_datapipe("map")
class MapperIterDataPipe(IterDataPipe):
    ...

The stack of DataPipes can then be constructed in functional form:

>>> import torch.utils.data.datapipes as dp
>>> datapipes1 = dp.iter.FileLoader(['a.file', 'b.file']).map(fn=decoder).shuffle().batch(2)

>>> datapipes2 = dp.iter.FileLoader(['a.file', 'b.file'])
>>> datapipes2 = dp.iter.Mapper(datapipes2)
>>> datapipes2 = dp.iter.Shuffler(datapipes2)
>>> datapipes2 = dp.iter.Batcher(datapipes2, 2)

In the above example, datapipes1 and datapipes2 represent the exact same stack of IterDataPipes.

Using DataPipes

For a complete example, suppose we want to load data from CSV files with the following steps:

• List all csv files in a directory
• Load csv files
• Parse csv file and yield rows

To support the above pipeline, CSVParser is registered as parse_csv_files to consume file streams and expand them as rows.

@functional_datapipe("parse_csv_files")
class CSVParserIterDataPipe(IterDataPipe):
    def __init__(self, dp, **fmtparams):
        self.dp = dp
        self.fmtparams = fmtparams

    def __iter__(self):
        for filename, stream in self.dp:
            reader = csv.reader(stream, **self.fmtparams)
            for row in reader:
                yield filename, row

Then, the pipeline can be assembled as follows:

>>> import torch.utils.data.datapipes as dp

>>> FOLDER = 'path/2/csv/folder'
>>> datapipe = dp.iter.FileLister([FOLDER]).filter(fn=lambda filename: filename.endswith('.csv'))
>>> datapipe = dp.iter.FileLoader(datapipe, mode='rt')
>>> datapipe = datapipe.parse_csv_files(delimiter=' ')

>>> for d in datapipe: # Start loading data
...     pass

Contributing

We welcome PRs! See the CONTRIBUTING file.

Prototype 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.

Future Plans

We hope to sufficiently expand the library, harden APIs, and gather feedback to enable a beta release at the time of the PyTorch 1.11 release (early 2022).

License

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