WebDataset is a PyTorch Dataset (IterableDataset) implementation providing efficient access to datasets stored in POSIX tar archives and uses only sequential/streaming data access. This brings substantial performance advantage in many compute environments, and it is essential for very large scale training.
While WebDataset scales to very large problems, it also works well with smaller datasets and simplifies creation, management, and distribution of training data for deep learning.
WebDataset implements standard PyTorch
IterableDataset interface and works with the PyTorch
Access to datasets is as simple as:
import webdataset as wds dataset = wds.WebDataset(url).shuffle(1000).decode("torchrgb").to_tuple("jpg;png", "json") dataloader = torch.utils.data.DataLoader(dataset, num_workers=4, batch_size=16) for inputs, outputs in dataloader: ...
In that code snippet,
url can refer to a local file, a local HTTP server, a cloud storage object, an object
on an object store, or even the output of arbitrary command pipelines.
WebDataset fulfills a similar function to Tensorflow's TFRecord/tf.Example classes, but it is much easier to adopt because it does not actually require any kind of data conversion: data is stored in exactly the same format inside tar files as it is on disk, and all preprocessing and data augmentation code remains unchanged.
$ 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
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
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