Work out data augmentation multiplier instantly with clear inputs, formula shown and shareable results.
Offline augmentation writes every transformed copy to disk, so the dataset multiplies in both example count and storage. Keeping the originals adds one more copy per image. On-the-fly augmentation in the data loader avoids the storage cost entirely and gives a different random transform each epoch, which usually regularises better — the offline route is mainly for reproducibility or slow pipelines.
Augmentation volume
total = base x copies (+ base if originals kept); bytes = total x average size
No. It encodes invariances you believe hold, such as a flipped cat still being a cat. It reduces overfitting but cannot substitute for genuinely new examples.
No, apart from the deterministic resize and normalisation used at inference. Augmenting validation data makes the score measure the wrong distribution.