February 27, 2024: I was dealing with COVID and a cross-country move for the past couple of months. I'm back now. Sorry for the slow response on some issues.

%matplotlib inline
import matplotlib.pyplot as plt
import torch.utils.data
import torch.nn
from random import randrange
import os
os.environ["WDS_VERBOSE_CACHE"] = "1"
os.environ["GOPEN_VERBOSE"] = "0"

The WebDataset Format

WebDataset format files are tar files, with two conventions:

• within each tar file, files that belong together and make up a training sample share the same basename when stripped of all filename extensions
• the shards of a tar file are numbered like something-000000.tar to something-012345.tar, usually specified using brace notation something-{000000..012345}.tar

WebDataset can read files from local disk or from any pipe, which allows it to access files using common cloud object stores. WebDataset can also read concatenated MsgPack and CBORs sources.

The WebDataset representation allows writing purely sequential I/O pipelines for large scale deep learning. This is important for achieving high I/O rates from local storage (3x-10x for local drives compared to random access) and for using object stores and cloud storage for training.

The WebDataset format represents images, movies, audio, etc. in their native file formats, making the creation of WebDataset format data as easy as just creating a tar archive. Because of the way data is aligned, WebDataset works well with block deduplication as well and aligns data on predictable boundaries.

Standard tools can be used for accessing and processing WebDataset-format files.

bucket = "https://storage.googleapis.com/webdataset/testdata/"
dataset = "publaynet-train-{000000..000009}.tar"

url = bucket + dataset
!curl -s {url} | tar tf - | sed 10q

PMC4991227_00003.json
PMC4991227_00003.png
PMC4537884_00002.json
PMC4537884_00002.png
PMC4323233_00003.json
PMC4323233_00003.png
PMC5429906_00004.json
PMC5429906_00004.png
PMC5592712_00002.json
PMC5592712_00002.png
tar: stdout: write error

Note that in these .tar files, we have pairs of .json and .png files; each such pair makes up a training sample.

WebDataset Libraries

There are several libraries supporting the WebDataset format:

• webdataset for Python3 (includes the wids library), this repository
• Webdataset.jl a Julia implementation
• tarp, a Golang implementation and command line tool
• Ray Data sources and sinks

The webdataset library can be used with PyTorch, Tensorflow, and Jax.

The webdataset Library

The webdataset library is an implementation of PyTorch IterableDataset (or a mock implementation thereof if you aren't using PyTorch). It implements as form of stream processing. Some of its features are:

• large scale parallel data access through sharding
• high performance disk I/O due to purely sequential reads
• latency insensitive due to big fat pipes
• no local storage required
• instant startup for training jobs
• only requires reading from file descriptors/network streams, no special APIs
• its API encourages high performance I/O pipelines
• scalable from tiny desktop datasets to petascale datasets
• provides local caching if desired
• requires no dataset metadata; any collection of shards can be read and used instantly

The main limitations people run into are related to the fact that IterableDataset is less commonly used in PyTorch and some existing code may not support it as well, and that achieving an exactly balanced number of training samples across many compute nodes for a fixed epoch size is tricky; for multinode training, webdataset is usually used with shard resampling.

There are two interfaces, the concise "fluid" interface and a longer "pipeline" interface. We'll show examples using the fluid interface, which is usually what you want.

import webdataset as wds
pil_dataset = wds.WebDataset(url).shuffle(1000).decode("pil").to_tuple("png", "json")

The resulting datasets are standard PyTorch IterableDataset instances.

isinstance(pil_dataset, torch.utils.data.IterableDataset)

True

for image, json in pil_dataset:
    break
plt.imshow(image)

<matplotlib.image.AxesImage at 0x7f73806db970>

We can add onto the existing pipeline for augmentation and data preparation.

import torchvision.transforms as transforms
from PIL import Image

preproc = transforms.Compose([
    transforms.Resize((224, 224)),
    transforms.ToTensor(),
    lambda x: 1-x,
])

def preprocess(sample):
    image, json = sample
    try:
        label = json["annotations"][0]["category_id"]
    except:
        label = 0
    return preproc(image), label

dataset = pil_dataset.map(preprocess)

for image, label in dataset:
    break
plt.imshow(image.numpy().transpose(1, 2, 0))

<matplotlib.image.AxesImage at 0x7f7375fc2230>

WebDataset is just an instance of a standard IterableDataset. It's a single-threaded way of iterating over a dataset. Since image decompression and data augmentation can be compute intensive, PyTorch usually uses the DataLoader class to parallelize data loading and preprocessing. WebDataset is fully compatible with the standard DataLoader.

Here are a number of notebooks showing how to use WebDataset for image classification and LLM training:

• train-resnet50-wds -- simple, single GPU training from Imagenet
• train-resnet50-multiray-wds -- multinode training using webdataset
• generate-text-dataset -- initial dataset generation
• tesseract-wds -- shard-to-shard transformations, here for OCR running over large datasets
• train-ocr-errors-hf -- an example of LLM fine tuning using a dataset in webdataset format

The wds-notes notebook contains some additional documentation and information about the library.

The webdataset Pipeline API

The wds.WebDataset fluid interface is just a convenient shorthand for writing down pipelines. The underlying pipeline is an instance of the wds.DataPipeline class, and you can construct data pipelines explicitly, similar to the way you use nn.Sequential inside models.

dataset = wds.DataPipeline(
    wds.SimpleShardList(url),

    # at this point we have an iterator over all the shards
    wds.shuffle(100),

    # add wds.split_by_node here if you are using multiple nodes
    wds.split_by_worker,

    # at this point, we have an iterator over the shards assigned to each worker
    wds.tarfile_to_samples(),

    # this shuffles the samples in memory
    wds.shuffle(1000),

    # this decodes the images and json
    wds.decode("pil"),
    wds.to_tuple("png", "json"),
    wds.map(preprocess),
    wds.shuffle(1000),
    wds.batched(16)
)

batch = next(iter(dataset))
batch[0].shape, batch[1].shape

(torch.Size([16, 3, 224, 224]), (16,))

The wids Library for Indexed WebDatasets

Installing the webdataset library installs a second library called wids. This library provides fully indexed/random access to the same datasets that webdataset accesses using iterators/streaming.

Like the webdataset library, wids is high scalable and provides efficient access to very large datasets. Being indexed, it is easily backwards compatible with existing data pipelines based on indexed dataset, including precise epochs for multinode training. The library comes with its own ChunkedSampler and DistributedChunkedSampler classes, which provided shuffling accross nodes while still preserving enough locality of reference for efficient training.

Internally, the library uses a mmap-based tar file reader implementation; this allows very fast access without precomputed indexes, and it also means that shard and the equivalet of "shuffle buffers" are shared in memory between workers on the same machine.

This additional power comes at some cost: the library requires a small metadata file that lists all the shards in a dataset and the number of samples contained in each, the library requires local storage for as many shards as there are I/O workers on a node, it uses shared memory and mmap, and the availability of indexing makes it easy to accidentally use inefficient access patterns.

Generally, the recommendation is to use webdataset for all data generation, data transformation, and training code, and to use wids only if you need fully random access to datasets (e.g., for browing or sparse sampling), need an indexed-based sampler, or are converting tricky legacy code.

import wids

train_url = "https://storage.googleapis.com/webdataset/fake-imagenet/imagenet-train.json"

dataset = wids.ShardListDataset(train_url)

sample = dataset[1900]

print(sample.keys())
print(sample[".txt"])
plt.imshow(sample[".jpg"])

dict_keys(['.cls', '.jpg', '.txt', '__key__', '__dataset__', '__index__', '__shard__', '__shardindex__'])
a high quality color photograph of a dog


https://storage.googleapis.com/webdataset/fake-ima base: https://storage.googleapis.com/webdataset/fake-imagenet name: imagenet-train nfiles: 1282 nbytes: 31242280960 samples: 128200 cache: /tmp/_wids_cache





<matplotlib.image.AxesImage at 0x7f7373669e70>

There are several examples of how to use wids in the examples directory.

• train-resnet50-wids shows how to train a ResNet-50 model on ImageNet using wids
• train-resnet50-multiray-wids shows how to train a ResNet-50 model on ImageNet using multiple nodes

Note that the APIs between webdataset and wids are not fully consistent:

• wids keeps the extension's "." in the keys, while webdataset removes it (".txt" vs "txt")
• wids doesn't have a fully fluid interface, and add_transformation just adds to a list of transformations
• webdataset currently can't read the wids JSON specifications

Installation and Documentation

$ pip install webdataset

For the Github version:

$ pip install git+https://github.com/tmbdev/webdataset.git

Here are some videos talking about WebDataset and large scale deep learning:

• Introduction to Large Scale Deep Learning
• Loading Training Data with WebDataset
• Creating Datasets in WebDataset Format
• Tools for Working with Large Datasets

Dependencies

The WebDataset library only requires PyTorch, NumPy, and a small library called braceexpand.

WebDataset loads a few additional libraries dynamically only when they are actually needed and only in the decoder:

• PIL/Pillow for image decoding
• torchvision, torchvideo, torchaudio for image/video/audio decoding
• msgpack for MessagePack decoding
• the curl command line tool for accessing HTTP servers
• the Google/Amazon/Azure command line tools for accessing cloud storage buckets

Loading of one of these libraries is triggered by configuring a decoder that attempts to decode content in the given format and encountering a file in that format during decoding. (Eventually, the torch... dependencies will be refactored into those libraries.)