Augmentor

Why we included this project

Teams training image-based models often run into a small, one-note dataset, and Augmentor exists to stretch those samples further. You build a stochastic pipeline out of building blocks: rotations, shears, crops, perspective skews, elastic distortions, each firing per image with a probability you set and parameters drawn from ranges you define. The library is framework independent and driven by a plain Python API, so it slots into a preprocessing step ahead of PyTorch or TensorFlow without tying you to one training stack. That makes it a practical pick for researchers and small teams who want reproducible, fine-grained control over how much synthetic variation reaches the model, and it scripts easily into an existing data prep workflow. If your bottleneck is dataset diversity rather than model architecture, this is a low-friction way to generate far more training examples from the images you already have.