An unopinionated replacement for PyTorch's Dataset and ImageFolder classes, for datasets stored as uncompressed Tar archives.
Just Tar it: No particular structure is enforced in the Tar archive. This means that you can just archive your files with no modification, and handle any data/meta-data with your dataset code.
Why? Storing a dataset as millions of small files makes access inefficient, and can create other difficulties in large-scale scenarios (e.g. running out of inodes, inneficient operations in distributed filesystems which are optimised for fewer large files). A Tar file is a simple and uncompressed archive format for which numerous utilities exist, and it allows fast random access into a single archive file.
The default TarDataset simply loads all PNG, JPG and JPEG images from a Tar file, and allows you to iterate them.
Images are returned as Tensor. Here some RGB values are printed.
from tardataset import TarDataset
dataset = TarDataset('example-data/colors.tar')
for (idx, image) in enumerate(dataset):
print(f"Image #{idx}, color: {image[:,0,0]}")For image classification datasets, where images are usually stored in one folder per class (e.g. ImageNet), TarImageFolder is a drop-in replacement for torchvision.dataset.ImageFolder.
For more complex scenarios -- say, you store some data in one or more JSON files, or you have folders with video frames in specific formats -- you can subclass TarDataset, and read the data in any format you like.
There is a more comprehensive set of examples as a Jupyter notebook in example.ipynb.
A large-scale data loading example is given in imagenet-example.py. Only the section of code responsible for data loading was modified from the official PyTorch ImageNet example.
First, ensure that the data is in the expected format for the original example to work, in a folder named ILSVRC12. Then, create a Tar archive from it (tar cf ILSVRC12.tar ILSVRC12 on Linux or a utility like 7-Zip on Windows). Finally, run our modified imagenet-example.py, passing it the path to the Tar archive instead.
João Henriques, Visual Geometry Group (VGG), University of Oxford