Beacon-aug

• A cross-library image augmentation module for deep learning training

Check our website for user instructions

Collaborative projects in Adobe Research

• Hide and Seek Augmentation (@ Krishna kumar Singh)
• TextFlow: Text synthesis (@Zhifei zhang )
• CAI: Image Fingerprinting for Network Robustness (@ Maksym Andriushchenko)
• Effect of Anti-aliasing resize function to FID (@Richard Zhang)
• GAN ensembling (@Lucy Chai @ Richard Zhang)
• PS Research team: mmcv, Resize 
• human-data-texture-augmentation

Contributors

• Module building: Rebecca Li, Yannick Hold-Geoffroy, Geoffrey Oxholm
• Customized functions and advanced properties contributing: Richard Zhang, Maksym Andriushchenko, Krishna Kumar Singh, Zhifei Zhang

Why Beacon-aug

• Incorporate the largest number of image augmentation operations(> 300) from 8 popular libraries
• Seamless cross library exchanging over different libraries
• Adobe featured customized functions of both parametric and GAN based transformations designed by Adobe Researchers
• Run external AI function inferencing as easy as general augmentation
• Advanced transformation pipelines for complex tasks (e.g. segmentation, detection, GAN training, network robustness)
• Support various input formats : np.array,PIL , Torch.tensor
• Extend the high-level attributes from Albumentations to other libraries by dynamic loading
• Differentiable check, anti-aliasing for operators
• Fast-visualization of the augmentation pipeline
• Easy to add customized functions for public contributors
• Image augmentation module for Adobe's GAN showcase package beacon ( 🆕 available now!)

Install

How to install

pip install git+https://github.com/adobe-research/beacon-aug

How to develop

git clone https://github.com/adobe-research/beacon-aug.git

conda create -y -n aug python=3.7
conda activate aug

python setup.py develop   # for developer 

import beacon_aug will now directly link to the package in the cloned directory, easing development.

How to use

import beacon_aug as BA
aug = BA.HorizontalFlip(p=1,  library="albumentations")
image_auged = aug(image=image)["image"]

Check tutorials for more examples:

Operator Overview

Basic Operators

Customized Operators

• check this instructions to add customized functions
• Contributors:
  • TextFlow: Zhifei Zhang @zzhang
  • HidePatch: Krishna kumar Singh @krishsin

BA.{CustomizedOperator}(library="custom")

Standard Operators:

Beacon_aug support all the operators in all standard libraries

BA.{StandardOperator}(library={library})

• Check this list for the operators supporting one library
• Check the chart for the operators supporting multiple libraries

Advanced Operators

• Check here for user instruction

Trouble Shooting

• Check here
• If you have troubles not solved, try the prune version by simply reinstall beacon-aug replacing requirements.txt with requirements_prune.txt in a new virtual environment

Citation

@misc{beacon-aug2021,
  author = {Li, Xiaoyang Rebecca
            and Hold-Geoffroy, Yannick
            and Oxholm Geoffrey
            and Singh, Krishna Kumar
            and Zhang, Zhifei Zhang
            and Zhang, Richard
            and Andriushchenko, Maksym
            and others},
  title = {Beacon-aug: A cross-library image augmentation toolbox.},
  howpublished = {\url{https://github.com/adobe-research/beacon-aug}},
  year = {2021},
  note = {Online; accessed Jan-11-2022}
}