YOCO is a simple method/strategy of performing augmentations, which enjoys the properties of parameter-free, easy usage, and boosting almost all augmentations for free (negligible computation & memory cost). We hope our study will attract the community’s attention in revisiting how to perform data augmentations.
You Only Cut Once: Boosting Data Augmentation with a Single Cut
Junlin Han, Pengfei Fang, Weihao Li, Jie Hong, Ali Armin, Ian Reid, Lars Petersson, Hongdong Li
DATA61-CSIRO and Australian National University and University of Adelaide
Preprint
@inproceedings{han2022yoco,
title={You Only Cut Once: Boosting Data Augmentation with a Single Cut},
author={Junlin Han and Pengfei Fang and Weihao Li and Jie Hong and Mohammad Ali Armin and and Ian Reid and Lars Petersson and Hongdong Li},
booktitle={arXiv preprint arXiv:2201.12078},
year={2022}
}
YOCO cuts one image into two equal pieces, either in the height or the width dimension. The same data augmentations are performed independently within each piece. Augmented pieces are then concatenated together to form one single augmented image.
Overall, YOCO benefits almost all augmentations in multiple vision tasks (classification, contrastive learning, object detection, instance segmentation, image deraining, image super-resolution). Please see our paper for more.
Applying YOCO is quite easy, here is a demo code of performing YOCO at the batch level.
***
images: images to be augmented, here is tensor with (b,c,h,w) shape
aug: composed augmentation operations
h: height of images
w: width of images
***
def YOCO(images, aug, h, w):
images = torch.cat((aug(images[:, :, :, 0:int(w/2)]), aug(images[:, :, :, int(w/2):w])), dim=3) if \
torch.rand(1) > 0.5 else torch.cat((aug(images[:, :, 0:int(h/2), :]), aug(images[:, :, int(h/2):h, :])), dim=2)
return images
for i, (images, target) in enumerate(train_loader):
aug = torch.nn.Sequential(
transforms.RandomHorizontalFlip(), )
_, _, h, w = images.shape
# perform augmentations with YOCO
images = YOCO(images, aug, h, w)
This repo aims to be minimal modifications on official PyTorch ImageNet training code and MoCo. Following their instructions to install the environments and prepare the datasets.
timm is also required for ImageNet classification, simply run
pip install timm
For each quadruplet, we show the original input image, augmented image from image-level augmentation, and two images from different cut dimensions produced by YOCO.
junlin.han@data61.csiro.au or junlinhcv@gmail.com
If you tried YOCO in other tasks/datasets/augmentations, please feel free to let me know the results. They will be collected and presented in this repo, regardless of positive or negative. Many thanks!
Our code is developed based on official PyTorch ImageNet training code and MoCo.