Albumentations frames augmentation as invariance bet

This is the right frame because most augmentation failures are symmetry mistakes, not missing hyperparameters. The piece is strongest when it turns augmentation from folklore into an explicit contract between data, labels, and model capacity.

• –Flips, crops, rotations, and color jitter are all bets about label-preserving variation, and those bets change by class, task, and signal strength.
• –The nastiest failure mode is not "unrealistic" augmentation; it's plausible augmentation that quietly washes out the feature the model actually uses.
• –The operational advice is solid: baseline first, then one transform family at a time, with visual checks and per-class ablations before long runs.
• –Augmentation should be budgeted alongside dropout, label smoothing, and model size, or it becomes a recipe for underfitting.
• –In some problems, the right move is more representative data, not a stronger synthetic policy.