aleju/imgaug ? reverse-engineered prompt

Build me a Python image augmentation library for machine learning that can take a set of input images and turn them into a much larger set of varied training examples.

I want it to handle not just images, but also heatmaps, segmentation maps, keypoints, bounding boxes, polygons, and line strings, and keep everything aligned when transformations are applied. Please include common augmentations like flips, crops, padding, affine and perspective transforms, blur, noise, contrast changes, hue and saturation shifts, and dropout style effects. It should be easy to combine multiple augmentations, apply them in random order, and sometimes only apply them to some images.

Make it feel polished with clear examples, a simple way to use it, and solid tests so I can trust that images and annotations stay in sync. If you need to check current docs or behavior for compatibility, look them up online.