Record sequential storage for deep learning.
Project description
Paddle-WebDataset, This is a Paddle version of WebDataset
%matplotlib inline
import matplotlib.pyplot as plt
import paddle.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
tosomething-012345.tar
, usually specified using brace notationsomething-{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.
PP-WebDataset Libraries
There are several libraries supporting the WebDataset format:
ppwebdataset
for Python3 (includes thewids
library), this repository- Webdataset.jl a Julia implementation
- tarp, a Golang implementation and command line tool
- Ray Data sources and sinks
The ppwebdataset
library can be used with Paddle.
The ppwebdataset
Library
The ppwebdataset
library is an implementation of Paddle IterableDataset
(or a mock implementation thereof if you aren't using Paddle). 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 Paddle 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 ppwebdataset as wds
pil_dataset = wds.WebDataset(url).shuffle(1000).decode("pil").to_tuple("png", "json")
The resulting datasets are standard Paddle IterableDataset
instances.
isinstance(pil_dataset, paddle.io.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 paddle.vision.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, Paddle 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
([16, 3, 224, 224], (16,))
The ppwids
Library for Indexed WebDatasets
Installing the ppwebdataset
library installs a second library called ppwids
. This library provides fully indexed/random access to the same datasets that ppwebdataset
accesses using iterators/streaming.
Like the ppebdataset
library, ppwids
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 ppwids
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 ppwids
train_url = "https://storage.googleapis.com/webdataset/fake-imagenet/imagenet-train.json"
dataset = ppwids.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 ppwids
in the examples directory.
- train-resnet50-wids shows how to train a ResNet-50 model on ImageNet using
ppwids
- 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:
ppwids
keeps the extension's "." in the keys, whilewebdataset
removes it (".txt" vs "txt")ppwids
doesn't have a fully fluid interface, andadd_transformation
just adds to a list of transformationsppwebdataset
currently can't read theppwids
JSON specifications
Installation and Documentation
$ pip install ppwebdataset
For the Github version:
$ pip install git+https://github.com/JunnYu/ppwebdataset.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 Paddle, 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
paddle.vision
for image decodingmsgpack
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 paddle... dependencies will be refactored into those libraries.)
Project details
Release history Release notifications | RSS feed
Download files
Download the file for your platform. If you're not sure which to choose, learn more about installing packages.
Source Distribution
Built Distribution
File details
Details for the file ppwebdataset-0.2.86.tar.gz
.
File metadata
- Download URL: ppwebdataset-0.2.86.tar.gz
- Upload date:
- Size: 66.4 kB
- Tags: Source
- Uploaded using Trusted Publishing? No
- Uploaded via: twine/4.0.2 CPython/3.8.10
File hashes
Algorithm | Hash digest | |
---|---|---|
SHA256 | b033a49e6e46ced66a67d509cc52ebeef371144432bff7c22faf7ea92c8031c8 |
|
MD5 | a6c4fac79c684c6f2b442b6e820eb873 |
|
BLAKE2b-256 | d8a9acb880766eb869786ea38529a3f620cf0f48f0ad8a4c2515c37bf5d6dc83 |
File details
Details for the file ppwebdataset-0.2.86-py3-none-any.whl
.
File metadata
- Download URL: ppwebdataset-0.2.86-py3-none-any.whl
- Upload date:
- Size: 74.5 kB
- Tags: Python 3
- Uploaded using Trusted Publishing? No
- Uploaded via: twine/4.0.2 CPython/3.8.10
File hashes
Algorithm | Hash digest | |
---|---|---|
SHA256 | 295bd55a8100bc3263509cd95baa96fa3ce6d6d758fa5aac7ca4522a80561eb2 |
|
MD5 | ea3c2dadf25d6a78502eb9ba3c355a58 |
|
BLAKE2b-256 | ee81f70981e1b75df1dff91a6e83efa1886e4f458c9810be318a3690691fa80d |