The AugLy library provides tools for multimodal data augmentation and adversarial robustness benchmarking. It supports image, text, and audio processing, allowing developers to create reproducible pipelines for machine learning models.
The framework includes a wide range of image transformations, such as pixelization, rotation, and overlaying text or emojis. It features metadata tracking, which records intensity scores and dimensions for each applied transformation.
AugLy supports bounding box-aware augmentations, ensuring that object coordinates are automatically adjusted during operations like cropping, flipping, or rotation. Users can also define custom transformation classes by subclassing the library's base modules.
The library allows for deterministic pipelines using seeded random number generators. This ensures identical outputs for complex chains of operations, such as multi-round image re-encoding or sequential filtering.
