AugLy is a data augmentations library that supports audio, image, text, and video modalities with over 100 augmentations.
The library addresses the need to improve model robustness through data augmentation and evaluation. It provides both function-based and class-based transforms with composition operators, and can return metadata about applied transformations including their intensity. The tool is designed around augmentations that reflect real-world user behavior on internet platforms, such as converting images to memes, overlaying text and emojis on images and videos, and reposting screenshots from social media.
AugLy is particularly suited for problems where these internet-native augmentations are relevant, such as copy detection, hate speech detection, and copyright infringement. The library also includes more generic augmentations for general-purpose use. Developers can choose to install dependencies for specific modalities rather than the full library, and the tool works with Python 3.6 and later.
The project maintains active test coverage through continuous integration workflows. Contributions are welcomed according to the repository guidelines, and the work has been formally documented through academic citation.
