airsplay/webdataset

A high-performance Python-based I/O system for large (and small) deep learning problems, with strong support for PyTorch.

%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"

The WebDataset Format

WebDataset reads dataset that are stored as tar files, with the simple convention that files that belong together and make up a training sample share the same basename. 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.

%%bash
curl -s http://storage.googleapis.com/nvdata-openimages/openimages-train-000000.tar | tar tf - | sed 10q

e39871fd9fd74f55.jpg
e39871fd9fd74f55.json
f18b91585c4d3f3e.jpg
f18b91585c4d3f3e.json
ede6e66b2fb59aab.jpg
ede6e66b2fb59aab.json
ed600d57fcee4f94.jpg
ed600d57fcee4f94.json
ff47e649b23f446d.jpg
ff47e649b23f446d.json

Related Projects

• the new torchdata library in PyTorch will add native (built-in) support for WebDataset
• the AIStore server provides high-speed storage, caching, and data transformation for WebDataset data
• WebDataset training can be carried out directly against S3, GCS, and other cloud storage buckets
• NVIDIA's DALI library supports reading WebDataset format data directly
• there is a companion project to read WebDataset data in Julia
• the tarp command line program can be used for quick and easy dataset transformations of WebDataset data

WebDataset

WebDataset makes it easy to write I/O pipelines for large datasets. Datasets can be stored locally or in the cloud.

When your data is stored in the cloud, by default, local copies are downloaded and cached when you open the remote dataset.

import webdataset as wds

url = "http://storage.googleapis.com/nvdata-publaynet/publaynet-train-{000000..000009}.tar"
dataset = wds.WebDataset(url).shuffle(1000).decode("rgb").to_tuple("png", "json")

The resulting datasets are standard PyTorch IterableDataset instances.

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

True

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

<matplotlib.image.AxesImage at 0x7f40849aa5b0>

Let's add some code to transform the input data.

def preprocess(sample):
    image, json = sample
    try:
        label = json["annotations"][0]["category_id"]
    except:
        label = 0
    return 1-image, label

dataset = dataset.map(preprocess)

for image, label in dataset:
    break
plt.imshow(image)
print(label)

1

Note that this uses the fluid interface to WebDataset, a convenient shorthand for a lot of training loops. We'll see later what this expands to.

Expanding Samples

Let's add another processing pipeline stage; this one expands a single large input sample into multiple smaller samples. We shuffle the newly generated sub-samples further to mix up sub-samples from different images in the stream.

This uses the .compose method, which takes a function that maps an interator over samples into another iterator over samples.

def get_patches(src):
    for sample in src:
        image, label = sample
        h, w = image.shape[:2]
        for i in range(16):
            y, x = randrange(h-256), randrange(w-265)
            patch = image[y:y+256, x:x+256]
            yield (patch, label)
        
dataset = dataset.compose(get_patches).shuffle(10000)

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

<matplotlib.image.AxesImage at 0x7f40848c18e0>

DataLoader

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.

loader = torch.utils.data.DataLoader(dataset, num_workers=4, batch_size=8)

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

(torch.Size([8, 256, 256, 3]), torch.Size([8]))

The webdataset library contains a small wrapper that adds a fluid interface to the DataLoader (and is otherwise identical).

This comes in handy if you want to shuffle across dataset instances and allows you to change batch size dynamically.

loader = wds.WebLoader(dataset, num_workers=4, batch_size=8)
loader = loader.unbatched().shuffle(1000).batched(12)

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

(torch.Size([12, 256, 256, 3]), torch.Size([12]))

It is generally most efficient to avoid batching in the DataLoader altogether; instead, batch in the dataset and then rebatch after the loader.

A complete pipeline then looks like this.

url = "http://storage.googleapis.com/nvdata-publaynet/publaynet-train-{000000..000009}.tar"
dataset = wds.WebDataset(url).shuffle(1000).decode("rgb").to_tuple("png", "json").map(preprocess)
dataset = dataset.compose(get_patches)
dataset = dataset.batched(16)

loader = wds.WebLoader(dataset, num_workers=4, batch_size=None)
loader = loader.unbatched().shuffle(1000).batched(12)

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

(torch.Size([12, 256, 256, 3]), torch.Size([12]))

Pipeline Interface

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),
    wds.split_by_worker,
    # at this point, we have an iterator over the shards assigned to each worker
    wds.tarfile_to_samples(),
    wds.shuffle(1000),
    wds.decode("torchrgb"),
    # at this point, we have an list of decompressed training samples from each shard in this worker in sequence
    get_patches,  # note that can put iterator->iterator functions into the pipeline directly
    wds.shuffle(10000),
    wds.to_tuple("big.jpg", "json"),
    wds.batched(16)
)

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

(torch.Size([12, 256, 256, 3]), torch.Size([12]))

Multinode Training

Multinode training in PyTorch and other frameworks is complex. It depends on how exactly you distribute training across nodes, whether you want to keep "exact epochs" (exactly and only one sample from the dataset per epoch), and whether your training framework can deal with unequal number of samples per node.

The simplest solution for multinode training is to use a resampling strategy for the shards, generating an infinite stream of samples. You then set the epoch length explicitly with the .with_epoch method.

dataset = wds.WebDataset(url, resampled=True).shuffle(1000).decode("rgb").to_tuple("png", "json").map(preprocess).with_epoch(10000)
sample = next(iter(dataset))

Inside a pipeline, you can do the same thing using the ResampledShards generator. Shuffling and splitting by worker are then not needed.

dataset = wds.DataPipeline(
    wds.ResampledShards(url),
    # at this point we have an iterator over all the shards
    wds.tarfile_to_samples(),
    wds.shuffle(1000),
    wds.decode("torchrgb"),
    # at this point, we have an list of decompressed training samples from each shard in this worker in sequence
    get_patches,  # note that can put iterator->iterator functions into the pipeline directly
    wds.shuffle(10000),
    wds.to_tuple("big.jpg", "json"),
    wds.batched(16)
).with_epoch(10000)

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

(torch.Size([12, 256, 256, 3]), torch.Size([12]))

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

Examples: (NB: some of these are for older versions of WebDataset, but the differences should be small)

• loading videos
• splitting raw videos into clips for training
• converting the Falling Things dataset

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.)

Data Decoding

Data decoding is a special kind of transformations of samples. You could simply write a decoding function like this:

def my_sample_decoder(sample):
    result = dict(__key__=sample["__key__"])
    for key, value in sample.items():
        if key == "png" or key.endswith(".png"):
            result[key] = mageio.imread(io.BytesIO(value))
        elif ...:
            ...
    return result

dataset = wds.Processor(wds.map, my_sample_decoder)(dataset)

This gets tedious, though, and it also unnecessarily hardcodes the sample's keys into the processing pipeline. To help with this, there is a helper class that simplifies this kind of code. The primary use of Decoder is for decoding compressed image, video, and audio formats, as well as unzipping .gz files.

Here is an example of automatically decoding .png images with imread and using the default torch_video and torch_audio decoders for video and audio:

def my_png_decoder(key, value):
    if not key.endswith(".png"):
        return None
    assert isinstance(value, bytes)
    return imageio.imread(io.BytesIO(value))

dataset = wds.Decoder(my_png_decoder, wds.torch_video, wds.torch_audio)(dataset)

You can use whatever criteria you like for deciding how to decode values in samples. When used with standard WebDataset format files, the keys are the full extensions of the file names inside a .tar file. For consistency, it's recommended that you primarily rely on the extensions (e.g., .png, .mp4) to decide which decoders to use. There is a special helper function that simplifies this:

def my_decoder(value):
    return imageio.imread(io.BytesIO(value))
    
dataset = wds.Decoder(wds.handle_extension(".png", my_decoder))(dataset)

Alternative Representation: CBOR

An alternative representation of collections of samples is based on the IETF CBOR standard, an efficient, binary representation of data structures. CBOR files are particularly useful for large collections of very small samples (data tuples, short strings, etc.)

import cbor
import numpy as np

Writing CBOR files is very easy:

with open("test.cbors", "wb") as stream:
    for i in np.linspace(-5.0, 5.0, 1000):
        cbor.dump((i, str(i)[-5:]), stream)

Of course, you can these files directly:

with open("test.cbors", "rb") as stream:
    for i in range(3):
        print(cbor.load(stream))

[-5.0, '-5.0']
[-4.98998998998999, '98999']
[-4.97997997997998, '97998']

And CBOR files/shards integrate fully into DataPipeline with the cbors_to_samples function.

dataset = wds.DataPipeline(
    wds.SimpleShardList("test.cbors"),
    wds.cbors_to_samples(),
)
data = list(iter(dataset))
len(data), data[0]

(1000, [-5.0, '-5.0'])

"Smaller" Datasets and Desktop Computing

WebDataset is an ideal solution for training on petascale datasets kept on high performance distributed data stores like AIStore, AWS/S3, and Google Cloud. Compared to data center GPU servers, desktop machines have much slower network connections, but training jobs on desktop machines often also use much smaller datasets. WebDataset also is very useful for such smaller datasets, and it can easily be used for developing and testing on small datasets and then scaling up to large datasets by simply using more shards.

Here are different usage scenarios:

• desktop deep learning, smaller datasets
  • copy all shards to local disk manually
  • use automatic shard caching
• prototyping, development, testing of jobs for large scale training
  • copy a small subset of shards to local disk
  • use automatic shard caching with a small subrange of shards
• cloud training against cloud buckets
  • use WebDataset directly with remote URLs
• on premises training with high performance store (e.g., AIStore) and fast networks
  • use WebDataset directly with remote URLs
• on premises training with slower object stores and/or slower networks
  • use automatic shard caching

Location Independence, Caching, Etc.

WebDataset makes it easy to use a single specification for your datasets and run your code without change in different environments.

Loadable Dataset Specifications

If you write your input pipelines such that they are defined by a dataset specification in some language, you can most easily retarget your training pipelines to different datasets. You can do this either by dynamically loading the Python code that constructs the pipeline or by using a YAML/JSON dataset specification.

A YAML dataset specification looks like this:

dataset:
  - shards: gs://nvdata-ocropus-tess/ia1-{000000..000033}.tar
    scaleprob: 0.3
  - shards: gs://nvdata-ocropus-tess/cdipsub-{000000..000022}.tar
    scale: [1.0, 3.0]
  - shards: gs://nvdata-ocropus-tess/gsub-{000000..000167}.tar
    scale: [0.4, 1.0]
  - shards: gs://nvdata-ocropus-tess/bin-gsub-{000000..000167}.tar
    extensions: nrm.jpg
    scale: [0.3, 1.0]
  - shards: gs://nvdata-ocropus/rendered.tar
    scaleprob: 1.0

Note that datasets can be composed from different shard collections, mixed in different proportions.

The dataset specification reader will be integrated in the next minor version update.

AIStore Proxy

If you want to use an AISTore server as a cache, you can tell any WebDataset pipeline to replace direct accesses to your URLs to proxied accesses via the AIStore server. To do that, you need to set a couple of environment variables.

export AIS_ENDPOINT=http://nix:51080
export USE_AIS_FOR="gs"

Now, any accesses to Google Cloud Storage (gs:// urls) will be routed through the AIS server.

URL Rewriting

You can rewrite URLs using regular expressions via an environment variable; the syntax is WDS_REWRITE=regex=regex;regex=regex.

For example, to replace gs:// accesses with local file accesses, use

export WDS_REWRITE="gs://=/shared/data/"

To access Google cloud data via ssh, you might use something like:

export WDS_REWRITE="gs://=pipe:ssh proxyhost gsutil cat "

Use the Caching Mechanism

If you use the built-in caching mechanism, you can simply download shards to a local directory and specify that directory as the cache directory. The shards in that directory will override the shards that are being downloaded. Shards in the cache are mapped based on the pathname and file name of your shard names.

Direct Copying of Shards

Let's take the OpenImages dataset as an example; it's half a terabyte large. For development and testing, you may not want to download the entire dataset, but you may also not want to use the dataset remotely. With WebDataset, you can just download a small number of shards and use them during development.

!curl -L -s http://storage.googleapis.com/nvdata-openimages/openimages-train-000000.tar > /tmp/openimages-train-000000.tar

dataset = wds.DataPipeline(
    wds.SimpleShardList("/tmp/openimages-train-000000.tar"),
    wds.tarfile_to_samples(),
)
repr(next(iter(dataset)))[:200]

"{'__key__': 'e39871fd9fd74f55', '__url__': '/tmp/openimages-train-000000.tar', 'jpg': b'\\xff\\xd8\\xff\\xe0\\x00\\x10JFIF\\x00\\x01\\x01\\x01\\x01:\\x01:\\x00\\x00\\xff\\xdb\\x00C\\x00\\x06\\x04\\x05\\x06\\x05\\x04\\x06\\x06\\"

Note that the WebDataset class works the same way on local files as it does on remote files. Furthermore, unlike other kinds of dataset formats and archive formats, downloaded datasets are immediately useful and don't need to be unpacked.

Creating a WebDataset

Using tar

Since WebDatasets are just regular tar files, you can usually create them by just using the tar command. All you have to do is to arrange for any files that should be in the same sample to share the same basename. Many datasets already come that way. For those, you can simply create a WebDataset with

$ tar --sort=name -cf dataset.tar dataset/

If your dataset has some other directory layout, you may need a different file name in the archive from the name on disk. You can use the --transform argument to GNU tar to transform file names. You can also use the -T argument to read the files from a text file and embed other options in that text file.

The tarp create Command

The tarp command is a little utility for manipulating tar archives. Its create subcommand makes it particularly simple to construct tar archives from files. The tarp create command takes a recipe for building a tar archive that contains lines of the form:

archive-name-1 source-name-1
archive-name-2 source-name-2
...

The source name can either be a file, "text:something", or "pipe:something".

Programmatically in Python

You can also create a WebDataset with library functions in this library:

• webdataset.TarWriter takes dictionaries containing key value pairs and writes them to disk
• webdataset.ShardWriter takes dictionaries containing key value pairs and writes them to disk as a series of shards

Here is a quick way of converting an existing dataset into a WebDataset; this will store all tensors as Python pickles:

sink = wds.TarWriter("dest.tar")
dataset = open_my_dataset()
for index, (input, output) in dataset:
    sink.write({
        "__key__": "sample%06d" % index,
        "input.pyd": input,
        "output.pyd": output,
    })
sink.close()

Storing data as Python pickles allows most common Python datatypes to be stored, it is lossless, and the format is fast to decode. However, it is uncompressed and cannot be read by non-Python programs. It's often better to choose other storage formats, e.g., taking advantage of common image compression formats.

If you know that the input is an image and the output is an integer class, you can also write something like this:

sink = wds.TarWriter("dest.tar")
dataset = open_my_dataset()
for index, (input, output) in dataset:
    assert input.ndim == 3 and input.shape[2] == 3
    assert input.dtype = np.float32 and np.amin(input) >= 0 and np.amax(input) <= 1
    assert type(output) == int
    sink.write({
        "__key__": "sample%06d" % index,
        "input.jpg": input,
        "output.cls": output,
    })
sink.close()

The assert statements in that loop are not necessary, but they document and illustrate the expectations for this particular dataset. Generally, the ".jpg" encoder can actually encode a wide variety of array types as images. The ".cls" encoder always requires an integer for encoding.

Here is how you can use TarWriter for writing a dataset without using an encoder:

sink = wds.TarWriter("dest.tar", encoder=False)
for basename in basenames:
    with open(f"{basename}.png", "rb") as stream):
        image = stream.read()
    cls = lookup_cls(basename)
    sample = {
        "__key__": basename,
        "input.png": image,
        "target.cls": cls
    }
    sink.write(sample)
sink.close()

Since no encoder is used, if you want to be able to read this data with the default decoder, image must contain a byte string corresponding to a PNG image (as indicated by the ".png" extension on its dictionary key), and cls must contain an integer encoded in ASCII (as indicated by the ".cls" extension on its dictionary key).

Writing Filters and Offline Augmentation

Webdataset can be used for filters and offline augmentation of datasets. Here is a complete example that pre-augments a shard and extracts class labels.

from torchvision import transforms
from itertools import islice

def extract_class(data):
    # mock implementation
    return 0

def preproc(image):
    image = transforms.ToTensor()(image)
    # more preprocessing here
    return image

def augment_wds(input, output, maxcount=999999999):
    src = wds.DataPipeline(
        wds.SimpleShardList(input),
        wds.tarfile_to_samples(),
        wds.decode("pil"),
        wds.to_tuple("__key__", "jpg;png", "json"),
        wds.map_tuple(None, preproc, None),
    )
    with wds.TarWriter(output) as dst:
        for key, image, data in islice(src, 0, maxcount):
            print(key)
            image = image.numpy().transpose(1, 2, 0)
            image -= np.amin(image)
            image /= np.amax(image)
            sample = {
                "__key__": key,
                "png": image,
                "cls": extract_class(data)
            }
            dst.write(sample)

Now run the augmentation pipeline:

url = "http://storage.googleapis.com/nvdata-openimages/openimages-train-000000.tar"
url = f"pipe:curl -L -s {url} || true"
augment_wds(url, "_temp.tar", maxcount=5)

e39871fd9fd74f55
f18b91585c4d3f3e
ede6e66b2fb59aab
ed600d57fcee4f94
ff47e649b23f446d

To verify that things worked correctly, let's look at the output file:

%%bash
tar tf _temp.tar

e39871fd9fd74f55.cls
e39871fd9fd74f55.png
f18b91585c4d3f3e.cls
f18b91585c4d3f3e.png
ede6e66b2fb59aab.cls
ede6e66b2fb59aab.png
ed600d57fcee4f94.cls
ed600d57fcee4f94.png
ff47e649b23f446d.cls
ff47e649b23f446d.png

If you want to preprocess the entire OpenImages dataset with a process like this, you can use your favorite job queueing or worflow system.

For example, using Dask, you could process all 554 shards in parallel using code like this:

shards = braceexpand.braceexpand("{000000..000554}")
inputs = [f"gs://bucket/openimages-{shard}.tar" for shard in shards]
outputs = [f"gs://bucket2/openimages-augmented-{shard}.tar" for shard in shards]
results = [dask.delayed(augment_wds)(args) for args in zip(inputs, outputs)]
dask.compute(*results)

Note that the data is streaming from and to Google Cloud Storage buckets, so very little local storage is required on each worker.

For very large scale processing, it's easiest to submit separate jobs to a Kubernetes cluster using the Kubernetes Job template, or using a workflow engine like Argo.

Whether you prefer WebDataset or Dataset is a matter of style.

Syntax for URL Sources

The SimpleShardList and ResampledShards take either a string or a list of URLs as an argument. If it is given a string, the string is expanded using the braceexpand library. So, the following are equivalent:

ShardList("dataset-{000..001}.tar")
ShardList(["dataset-000.tar", "dataset-001.tar"])

The url strings in a shard list are handled by default by the webdataset.url_opener filter. It recognizes three simple kinds of strings: "-", "/path/to/file", and "pipe:command":

• the string "-", referring to stdin
• a UNIX path, opened as a regular file
• a URL-like string with the schema "pipe:"; such URLs are opened with subprocess.Popen. For example:
  • pipe:curl -s -L http://server/file accesses a file via HTTP
  • pipe:gsutil cat gs://bucket/file accesses a file on GCS
  • pipe:az cp --container bucket --name file --file /dev/stdout accesses a file on Azure
  • pipe:ssh host cat file accesses a file via ssh

It might seem at first glance to be "more efficient" to use built-in Python libraries for accessing object stores rather than subprocesses, but efficient object store access from Python really requires spawning a separate process anyway, so this approach to accessing object stores is not only convenient, it also is as efficient as we can make it in Python.

Related Libraries and Software

The AIStore server provides an efficient backend for WebDataset; it functions like a combination of web server, content distribution network, P2P network, and distributed file system. Together, AIStore and WebDataset can serve input data from rotational drives distributed across many servers at the speed of local SSDs to many GPUs, at a fraction of the cost. We can easily achieve hundreds of MBytes/s of I/O per GPU even in large, distributed training jobs.

The tarproc utilities provide command line manipulation and processing of webdatasets and other tar files, including splitting, concatenation, and xargs-like functionality.

The tensorcom library provides fast three-tiered I/O; it can be inserted between AIStore and WebDataset to permit distributed data augmentation and I/O. It is particularly useful when data augmentation requires more CPU than the GPU server has available.

You can find the full PyTorch ImageNet sample code converted to WebDataset at tmbdev/pytorch-imagenet-wds